Root/mm/page_alloc.c

1/*
2 * linux/mm/page_alloc.c
3 *
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
6 *
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
15 */
16
17#include <linux/stddef.h>
18#include <linux/mm.h>
19#include <linux/swap.h>
20#include <linux/interrupt.h>
21#include <linux/pagemap.h>
22#include <linux/jiffies.h>
23#include <linux/bootmem.h>
24#include <linux/memblock.h>
25#include <linux/compiler.h>
26#include <linux/kernel.h>
27#include <linux/kmemcheck.h>
28#include <linux/module.h>
29#include <linux/suspend.h>
30#include <linux/pagevec.h>
31#include <linux/blkdev.h>
32#include <linux/slab.h>
33#include <linux/ratelimit.h>
34#include <linux/oom.h>
35#include <linux/notifier.h>
36#include <linux/topology.h>
37#include <linux/sysctl.h>
38#include <linux/cpu.h>
39#include <linux/cpuset.h>
40#include <linux/memory_hotplug.h>
41#include <linux/nodemask.h>
42#include <linux/vmalloc.h>
43#include <linux/vmstat.h>
44#include <linux/mempolicy.h>
45#include <linux/stop_machine.h>
46#include <linux/sort.h>
47#include <linux/pfn.h>
48#include <linux/backing-dev.h>
49#include <linux/fault-inject.h>
50#include <linux/page-isolation.h>
51#include <linux/page_cgroup.h>
52#include <linux/debugobjects.h>
53#include <linux/kmemleak.h>
54#include <linux/memory.h>
55#include <linux/compaction.h>
56#include <trace/events/kmem.h>
57#include <linux/ftrace_event.h>
58#include <linux/memcontrol.h>
59#include <linux/prefetch.h>
60
61#include <asm/tlbflush.h>
62#include <asm/div64.h>
63#include "internal.h"
64
65#ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
66DEFINE_PER_CPU(int, numa_node);
67EXPORT_PER_CPU_SYMBOL(numa_node);
68#endif
69
70#ifdef CONFIG_HAVE_MEMORYLESS_NODES
71/*
72 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
73 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
74 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
75 * defined in <linux/topology.h>.
76 */
77DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
78EXPORT_PER_CPU_SYMBOL(_numa_mem_);
79#endif
80
81/*
82 * Array of node states.
83 */
84nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
85    [N_POSSIBLE] = NODE_MASK_ALL,
86    [N_ONLINE] = { { [0] = 1UL } },
87#ifndef CONFIG_NUMA
88    [N_NORMAL_MEMORY] = { { [0] = 1UL } },
89#ifdef CONFIG_HIGHMEM
90    [N_HIGH_MEMORY] = { { [0] = 1UL } },
91#endif
92    [N_CPU] = { { [0] = 1UL } },
93#endif /* NUMA */
94};
95EXPORT_SYMBOL(node_states);
96
97unsigned long totalram_pages __read_mostly;
98unsigned long totalreserve_pages __read_mostly;
99int percpu_pagelist_fraction;
100gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
101
102#ifdef CONFIG_PM_SLEEP
103/*
104 * The following functions are used by the suspend/hibernate code to temporarily
105 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
106 * while devices are suspended. To avoid races with the suspend/hibernate code,
107 * they should always be called with pm_mutex held (gfp_allowed_mask also should
108 * only be modified with pm_mutex held, unless the suspend/hibernate code is
109 * guaranteed not to run in parallel with that modification).
110 */
111
112static gfp_t saved_gfp_mask;
113
114void pm_restore_gfp_mask(void)
115{
116    WARN_ON(!mutex_is_locked(&pm_mutex));
117    if (saved_gfp_mask) {
118        gfp_allowed_mask = saved_gfp_mask;
119        saved_gfp_mask = 0;
120    }
121}
122
123void pm_restrict_gfp_mask(void)
124{
125    WARN_ON(!mutex_is_locked(&pm_mutex));
126    WARN_ON(saved_gfp_mask);
127    saved_gfp_mask = gfp_allowed_mask;
128    gfp_allowed_mask &= ~GFP_IOFS;
129}
130#endif /* CONFIG_PM_SLEEP */
131
132#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
133int pageblock_order __read_mostly;
134#endif
135
136static void __free_pages_ok(struct page *page, unsigned int order);
137
138/*
139 * results with 256, 32 in the lowmem_reserve sysctl:
140 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
141 * 1G machine -> (16M dma, 784M normal, 224M high)
142 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
143 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
144 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
145 *
146 * TBD: should special case ZONE_DMA32 machines here - in those we normally
147 * don't need any ZONE_NORMAL reservation
148 */
149int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
150#ifdef CONFIG_ZONE_DMA
151     256,
152#endif
153#ifdef CONFIG_ZONE_DMA32
154     256,
155#endif
156#ifdef CONFIG_HIGHMEM
157     32,
158#endif
159     32,
160};
161
162EXPORT_SYMBOL(totalram_pages);
163
164static char * const zone_names[MAX_NR_ZONES] = {
165#ifdef CONFIG_ZONE_DMA
166     "DMA",
167#endif
168#ifdef CONFIG_ZONE_DMA32
169     "DMA32",
170#endif
171     "Normal",
172#ifdef CONFIG_HIGHMEM
173     "HighMem",
174#endif
175     "Movable",
176};
177
178int min_free_kbytes = 1024;
179
180static unsigned long __meminitdata nr_kernel_pages;
181static unsigned long __meminitdata nr_all_pages;
182static unsigned long __meminitdata dma_reserve;
183
184#ifdef CONFIG_ARCH_POPULATES_NODE_MAP
185  /*
186   * MAX_ACTIVE_REGIONS determines the maximum number of distinct
187   * ranges of memory (RAM) that may be registered with add_active_range().
188   * Ranges passed to add_active_range() will be merged if possible
189   * so the number of times add_active_range() can be called is
190   * related to the number of nodes and the number of holes
191   */
192  #ifdef CONFIG_MAX_ACTIVE_REGIONS
193    /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */
194    #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS
195  #else
196    #if MAX_NUMNODES >= 32
197      /* If there can be many nodes, allow up to 50 holes per node */
198      #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50)
199    #else
200      /* By default, allow up to 256 distinct regions */
201      #define MAX_ACTIVE_REGIONS 256
202    #endif
203  #endif
204
205  static struct node_active_region __meminitdata early_node_map[MAX_ACTIVE_REGIONS];
206  static int __meminitdata nr_nodemap_entries;
207  static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
208  static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
209  static unsigned long __initdata required_kernelcore;
210  static unsigned long __initdata required_movablecore;
211  static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
212
213  /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
214  int movable_zone;
215  EXPORT_SYMBOL(movable_zone);
216#endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
217
218#if MAX_NUMNODES > 1
219int nr_node_ids __read_mostly = MAX_NUMNODES;
220int nr_online_nodes __read_mostly = 1;
221EXPORT_SYMBOL(nr_node_ids);
222EXPORT_SYMBOL(nr_online_nodes);
223#endif
224
225int page_group_by_mobility_disabled __read_mostly;
226
227static void set_pageblock_migratetype(struct page *page, int migratetype)
228{
229
230    if (unlikely(page_group_by_mobility_disabled))
231        migratetype = MIGRATE_UNMOVABLE;
232
233    set_pageblock_flags_group(page, (unsigned long)migratetype,
234                    PB_migrate, PB_migrate_end);
235}
236
237bool oom_killer_disabled __read_mostly;
238
239#ifdef CONFIG_DEBUG_VM
240static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
241{
242    int ret = 0;
243    unsigned seq;
244    unsigned long pfn = page_to_pfn(page);
245
246    do {
247        seq = zone_span_seqbegin(zone);
248        if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
249            ret = 1;
250        else if (pfn < zone->zone_start_pfn)
251            ret = 1;
252    } while (zone_span_seqretry(zone, seq));
253
254    return ret;
255}
256
257static int page_is_consistent(struct zone *zone, struct page *page)
258{
259    if (!pfn_valid_within(page_to_pfn(page)))
260        return 0;
261    if (zone != page_zone(page))
262        return 0;
263
264    return 1;
265}
266/*
267 * Temporary debugging check for pages not lying within a given zone.
268 */
269static int bad_range(struct zone *zone, struct page *page)
270{
271    if (page_outside_zone_boundaries(zone, page))
272        return 1;
273    if (!page_is_consistent(zone, page))
274        return 1;
275
276    return 0;
277}
278#else
279static inline int bad_range(struct zone *zone, struct page *page)
280{
281    return 0;
282}
283#endif
284
285static void bad_page(struct page *page)
286{
287    static unsigned long resume;
288    static unsigned long nr_shown;
289    static unsigned long nr_unshown;
290
291    /* Don't complain about poisoned pages */
292    if (PageHWPoison(page)) {
293        reset_page_mapcount(page); /* remove PageBuddy */
294        return;
295    }
296
297    /*
298     * Allow a burst of 60 reports, then keep quiet for that minute;
299     * or allow a steady drip of one report per second.
300     */
301    if (nr_shown == 60) {
302        if (time_before(jiffies, resume)) {
303            nr_unshown++;
304            goto out;
305        }
306        if (nr_unshown) {
307            printk(KERN_ALERT
308                  "BUG: Bad page state: %lu messages suppressed\n",
309                nr_unshown);
310            nr_unshown = 0;
311        }
312        nr_shown = 0;
313    }
314    if (nr_shown++ == 0)
315        resume = jiffies + 60 * HZ;
316
317    printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
318        current->comm, page_to_pfn(page));
319    dump_page(page);
320
321    dump_stack();
322out:
323    /* Leave bad fields for debug, except PageBuddy could make trouble */
324    reset_page_mapcount(page); /* remove PageBuddy */
325    add_taint(TAINT_BAD_PAGE);
326}
327
328/*
329 * Higher-order pages are called "compound pages". They are structured thusly:
330 *
331 * The first PAGE_SIZE page is called the "head page".
332 *
333 * The remaining PAGE_SIZE pages are called "tail pages".
334 *
335 * All pages have PG_compound set. All pages have their ->private pointing at
336 * the head page (even the head page has this).
337 *
338 * The first tail page's ->lru.next holds the address of the compound page's
339 * put_page() function. Its ->lru.prev holds the order of allocation.
340 * This usage means that zero-order pages may not be compound.
341 */
342
343static void free_compound_page(struct page *page)
344{
345    __free_pages_ok(page, compound_order(page));
346}
347
348void prep_compound_page(struct page *page, unsigned long order)
349{
350    int i;
351    int nr_pages = 1 << order;
352
353    set_compound_page_dtor(page, free_compound_page);
354    set_compound_order(page, order);
355    __SetPageHead(page);
356    for (i = 1; i < nr_pages; i++) {
357        struct page *p = page + i;
358
359        __SetPageTail(p);
360        p->first_page = page;
361    }
362}
363
364/* update __split_huge_page_refcount if you change this function */
365static int destroy_compound_page(struct page *page, unsigned long order)
366{
367    int i;
368    int nr_pages = 1 << order;
369    int bad = 0;
370
371    if (unlikely(compound_order(page) != order) ||
372        unlikely(!PageHead(page))) {
373        bad_page(page);
374        bad++;
375    }
376
377    __ClearPageHead(page);
378
379    for (i = 1; i < nr_pages; i++) {
380        struct page *p = page + i;
381
382        if (unlikely(!PageTail(p) || (p->first_page != page))) {
383            bad_page(page);
384            bad++;
385        }
386        __ClearPageTail(p);
387    }
388
389    return bad;
390}
391
392static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
393{
394    int i;
395
396    /*
397     * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
398     * and __GFP_HIGHMEM from hard or soft interrupt context.
399     */
400    VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
401    for (i = 0; i < (1 << order); i++)
402        clear_highpage(page + i);
403}
404
405static inline void set_page_order(struct page *page, int order)
406{
407    set_page_private(page, order);
408    __SetPageBuddy(page);
409}
410
411static inline void rmv_page_order(struct page *page)
412{
413    __ClearPageBuddy(page);
414    set_page_private(page, 0);
415}
416
417/*
418 * Locate the struct page for both the matching buddy in our
419 * pair (buddy1) and the combined O(n+1) page they form (page).
420 *
421 * 1) Any buddy B1 will have an order O twin B2 which satisfies
422 * the following equation:
423 * B2 = B1 ^ (1 << O)
424 * For example, if the starting buddy (buddy2) is #8 its order
425 * 1 buddy is #10:
426 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
427 *
428 * 2) Any buddy B will have an order O+1 parent P which
429 * satisfies the following equation:
430 * P = B & ~(1 << O)
431 *
432 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
433 */
434static inline unsigned long
435__find_buddy_index(unsigned long page_idx, unsigned int order)
436{
437    return page_idx ^ (1 << order);
438}
439
440/*
441 * This function checks whether a page is free && is the buddy
442 * we can do coalesce a page and its buddy if
443 * (a) the buddy is not in a hole &&
444 * (b) the buddy is in the buddy system &&
445 * (c) a page and its buddy have the same order &&
446 * (d) a page and its buddy are in the same zone.
447 *
448 * For recording whether a page is in the buddy system, we set ->_mapcount -2.
449 * Setting, clearing, and testing _mapcount -2 is serialized by zone->lock.
450 *
451 * For recording page's order, we use page_private(page).
452 */
453static inline int page_is_buddy(struct page *page, struct page *buddy,
454                                int order)
455{
456    if (!pfn_valid_within(page_to_pfn(buddy)))
457        return 0;
458
459    if (page_zone_id(page) != page_zone_id(buddy))
460        return 0;
461
462    if (PageBuddy(buddy) && page_order(buddy) == order) {
463        VM_BUG_ON(page_count(buddy) != 0);
464        return 1;
465    }
466    return 0;
467}
468
469/*
470 * Freeing function for a buddy system allocator.
471 *
472 * The concept of a buddy system is to maintain direct-mapped table
473 * (containing bit values) for memory blocks of various "orders".
474 * The bottom level table contains the map for the smallest allocatable
475 * units of memory (here, pages), and each level above it describes
476 * pairs of units from the levels below, hence, "buddies".
477 * At a high level, all that happens here is marking the table entry
478 * at the bottom level available, and propagating the changes upward
479 * as necessary, plus some accounting needed to play nicely with other
480 * parts of the VM system.
481 * At each level, we keep a list of pages, which are heads of continuous
482 * free pages of length of (1 << order) and marked with _mapcount -2. Page's
483 * order is recorded in page_private(page) field.
484 * So when we are allocating or freeing one, we can derive the state of the
485 * other. That is, if we allocate a small block, and both were
486 * free, the remainder of the region must be split into blocks.
487 * If a block is freed, and its buddy is also free, then this
488 * triggers coalescing into a block of larger size.
489 *
490 * -- wli
491 */
492
493static inline void __free_one_page(struct page *page,
494        struct zone *zone, unsigned int order,
495        int migratetype)
496{
497    unsigned long page_idx;
498    unsigned long combined_idx;
499    unsigned long uninitialized_var(buddy_idx);
500    struct page *buddy;
501
502    if (unlikely(PageCompound(page)))
503        if (unlikely(destroy_compound_page(page, order)))
504            return;
505
506    VM_BUG_ON(migratetype == -1);
507
508    page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
509
510    VM_BUG_ON(page_idx & ((1 << order) - 1));
511    VM_BUG_ON(bad_range(zone, page));
512
513    while (order < MAX_ORDER-1) {
514        buddy_idx = __find_buddy_index(page_idx, order);
515        buddy = page + (buddy_idx - page_idx);
516        if (!page_is_buddy(page, buddy, order))
517            break;
518
519        /* Our buddy is free, merge with it and move up one order. */
520        list_del(&buddy->lru);
521        zone->free_area[order].nr_free--;
522        rmv_page_order(buddy);
523        combined_idx = buddy_idx & page_idx;
524        page = page + (combined_idx - page_idx);
525        page_idx = combined_idx;
526        order++;
527    }
528    set_page_order(page, order);
529
530    /*
531     * If this is not the largest possible page, check if the buddy
532     * of the next-highest order is free. If it is, it's possible
533     * that pages are being freed that will coalesce soon. In case,
534     * that is happening, add the free page to the tail of the list
535     * so it's less likely to be used soon and more likely to be merged
536     * as a higher order page
537     */
538    if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
539        struct page *higher_page, *higher_buddy;
540        combined_idx = buddy_idx & page_idx;
541        higher_page = page + (combined_idx - page_idx);
542        buddy_idx = __find_buddy_index(combined_idx, order + 1);
543        higher_buddy = page + (buddy_idx - combined_idx);
544        if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
545            list_add_tail(&page->lru,
546                &zone->free_area[order].free_list[migratetype]);
547            goto out;
548        }
549    }
550
551    list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
552out:
553    zone->free_area[order].nr_free++;
554}
555
556/*
557 * free_page_mlock() -- clean up attempts to free and mlocked() page.
558 * Page should not be on lru, so no need to fix that up.
559 * free_pages_check() will verify...
560 */
561static inline void free_page_mlock(struct page *page)
562{
563    __dec_zone_page_state(page, NR_MLOCK);
564    __count_vm_event(UNEVICTABLE_MLOCKFREED);
565}
566
567static inline int free_pages_check(struct page *page)
568{
569    if (unlikely(page_mapcount(page) |
570        (page->mapping != NULL) |
571        (atomic_read(&page->_count) != 0) |
572        (page->flags & PAGE_FLAGS_CHECK_AT_FREE) |
573        (mem_cgroup_bad_page_check(page)))) {
574        bad_page(page);
575        return 1;
576    }
577    if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
578        page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
579    return 0;
580}
581
582/*
583 * Frees a number of pages from the PCP lists
584 * Assumes all pages on list are in same zone, and of same order.
585 * count is the number of pages to free.
586 *
587 * If the zone was previously in an "all pages pinned" state then look to
588 * see if this freeing clears that state.
589 *
590 * And clear the zone's pages_scanned counter, to hold off the "all pages are
591 * pinned" detection logic.
592 */
593static void free_pcppages_bulk(struct zone *zone, int count,
594                    struct per_cpu_pages *pcp)
595{
596    int migratetype = 0;
597    int batch_free = 0;
598    int to_free = count;
599
600    spin_lock(&zone->lock);
601    zone->all_unreclaimable = 0;
602    zone->pages_scanned = 0;
603
604    while (to_free) {
605        struct page *page;
606        struct list_head *list;
607
608        /*
609         * Remove pages from lists in a round-robin fashion. A
610         * batch_free count is maintained that is incremented when an
611         * empty list is encountered. This is so more pages are freed
612         * off fuller lists instead of spinning excessively around empty
613         * lists
614         */
615        do {
616            batch_free++;
617            if (++migratetype == MIGRATE_PCPTYPES)
618                migratetype = 0;
619            list = &pcp->lists[migratetype];
620        } while (list_empty(list));
621
622        /* This is the only non-empty list. Free them all. */
623        if (batch_free == MIGRATE_PCPTYPES)
624            batch_free = to_free;
625
626        do {
627            page = list_entry(list->prev, struct page, lru);
628            /* must delete as __free_one_page list manipulates */
629            list_del(&page->lru);
630            /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
631            __free_one_page(page, zone, 0, page_private(page));
632            trace_mm_page_pcpu_drain(page, 0, page_private(page));
633        } while (--to_free && --batch_free && !list_empty(list));
634    }
635    __mod_zone_page_state(zone, NR_FREE_PAGES, count);
636    spin_unlock(&zone->lock);
637}
638
639static void free_one_page(struct zone *zone, struct page *page, int order,
640                int migratetype)
641{
642    spin_lock(&zone->lock);
643    zone->all_unreclaimable = 0;
644    zone->pages_scanned = 0;
645
646    __free_one_page(page, zone, order, migratetype);
647    __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order);
648    spin_unlock(&zone->lock);
649}
650
651static bool free_pages_prepare(struct page *page, unsigned int order)
652{
653    int i;
654    int bad = 0;
655
656    trace_mm_page_free_direct(page, order);
657    kmemcheck_free_shadow(page, order);
658
659    if (PageAnon(page))
660        page->mapping = NULL;
661    for (i = 0; i < (1 << order); i++)
662        bad += free_pages_check(page + i);
663    if (bad)
664        return false;
665
666    if (!PageHighMem(page)) {
667        debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
668        debug_check_no_obj_freed(page_address(page),
669                       PAGE_SIZE << order);
670    }
671    arch_free_page(page, order);
672    kernel_map_pages(page, 1 << order, 0);
673
674    return true;
675}
676
677static void __free_pages_ok(struct page *page, unsigned int order)
678{
679    unsigned long flags;
680    int wasMlocked = __TestClearPageMlocked(page);
681
682    if (!free_pages_prepare(page, order))
683        return;
684
685    local_irq_save(flags);
686    if (unlikely(wasMlocked))
687        free_page_mlock(page);
688    __count_vm_events(PGFREE, 1 << order);
689    free_one_page(page_zone(page), page, order,
690                    get_pageblock_migratetype(page));
691    local_irq_restore(flags);
692}
693
694/*
695 * permit the bootmem allocator to evade page validation on high-order frees
696 */
697void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
698{
699    if (order == 0) {
700        __ClearPageReserved(page);
701        set_page_count(page, 0);
702        set_page_refcounted(page);
703        __free_page(page);
704    } else {
705        int loop;
706
707        prefetchw(page);
708        for (loop = 0; loop < BITS_PER_LONG; loop++) {
709            struct page *p = &page[loop];
710
711            if (loop + 1 < BITS_PER_LONG)
712                prefetchw(p + 1);
713            __ClearPageReserved(p);
714            set_page_count(p, 0);
715        }
716
717        set_page_refcounted(page);
718        __free_pages(page, order);
719    }
720}
721
722
723/*
724 * The order of subdivision here is critical for the IO subsystem.
725 * Please do not alter this order without good reasons and regression
726 * testing. Specifically, as large blocks of memory are subdivided,
727 * the order in which smaller blocks are delivered depends on the order
728 * they're subdivided in this function. This is the primary factor
729 * influencing the order in which pages are delivered to the IO
730 * subsystem according to empirical testing, and this is also justified
731 * by considering the behavior of a buddy system containing a single
732 * large block of memory acted on by a series of small allocations.
733 * This behavior is a critical factor in sglist merging's success.
734 *
735 * -- wli
736 */
737static inline void expand(struct zone *zone, struct page *page,
738    int low, int high, struct free_area *area,
739    int migratetype)
740{
741    unsigned long size = 1 << high;
742
743    while (high > low) {
744        area--;
745        high--;
746        size >>= 1;
747        VM_BUG_ON(bad_range(zone, &page[size]));
748        list_add(&page[size].lru, &area->free_list[migratetype]);
749        area->nr_free++;
750        set_page_order(&page[size], high);
751    }
752}
753
754/*
755 * This page is about to be returned from the page allocator
756 */
757static inline int check_new_page(struct page *page)
758{
759    if (unlikely(page_mapcount(page) |
760        (page->mapping != NULL) |
761        (atomic_read(&page->_count) != 0) |
762        (page->flags & PAGE_FLAGS_CHECK_AT_PREP) |
763        (mem_cgroup_bad_page_check(page)))) {
764        bad_page(page);
765        return 1;
766    }
767    return 0;
768}
769
770static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
771{
772    int i;
773
774    for (i = 0; i < (1 << order); i++) {
775        struct page *p = page + i;
776        if (unlikely(check_new_page(p)))
777            return 1;
778    }
779
780    set_page_private(page, 0);
781    set_page_refcounted(page);
782
783    arch_alloc_page(page, order);
784    kernel_map_pages(page, 1 << order, 1);
785
786    if (gfp_flags & __GFP_ZERO)
787        prep_zero_page(page, order, gfp_flags);
788
789    if (order && (gfp_flags & __GFP_COMP))
790        prep_compound_page(page, order);
791
792    return 0;
793}
794
795/*
796 * Go through the free lists for the given migratetype and remove
797 * the smallest available page from the freelists
798 */
799static inline
800struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
801                        int migratetype)
802{
803    unsigned int current_order;
804    struct free_area * area;
805    struct page *page;
806
807    /* Find a page of the appropriate size in the preferred list */
808    for (current_order = order; current_order < MAX_ORDER; ++current_order) {
809        area = &(zone->free_area[current_order]);
810        if (list_empty(&area->free_list[migratetype]))
811            continue;
812
813        page = list_entry(area->free_list[migratetype].next,
814                            struct page, lru);
815        list_del(&page->lru);
816        rmv_page_order(page);
817        area->nr_free--;
818        expand(zone, page, order, current_order, area, migratetype);
819        return page;
820    }
821
822    return NULL;
823}
824
825
826/*
827 * This array describes the order lists are fallen back to when
828 * the free lists for the desirable migrate type are depleted
829 */
830static int fallbacks[MIGRATE_TYPES][MIGRATE_TYPES-1] = {
831    [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
832    [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
833    [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
834    [MIGRATE_RESERVE] = { MIGRATE_RESERVE, MIGRATE_RESERVE, MIGRATE_RESERVE }, /* Never used */
835};
836
837/*
838 * Move the free pages in a range to the free lists of the requested type.
839 * Note that start_page and end_pages are not aligned on a pageblock
840 * boundary. If alignment is required, use move_freepages_block()
841 */
842static int move_freepages(struct zone *zone,
843              struct page *start_page, struct page *end_page,
844              int migratetype)
845{
846    struct page *page;
847    unsigned long order;
848    int pages_moved = 0;
849
850#ifndef CONFIG_HOLES_IN_ZONE
851    /*
852     * page_zone is not safe to call in this context when
853     * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
854     * anyway as we check zone boundaries in move_freepages_block().
855     * Remove at a later date when no bug reports exist related to
856     * grouping pages by mobility
857     */
858    BUG_ON(page_zone(start_page) != page_zone(end_page));
859#endif
860
861    for (page = start_page; page <= end_page;) {
862        /* Make sure we are not inadvertently changing nodes */
863        VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
864
865        if (!pfn_valid_within(page_to_pfn(page))) {
866            page++;
867            continue;
868        }
869
870        if (!PageBuddy(page)) {
871            page++;
872            continue;
873        }
874
875        order = page_order(page);
876        list_move(&page->lru,
877              &zone->free_area[order].free_list[migratetype]);
878        page += 1 << order;
879        pages_moved += 1 << order;
880    }
881
882    return pages_moved;
883}
884
885static int move_freepages_block(struct zone *zone, struct page *page,
886                int migratetype)
887{
888    unsigned long start_pfn, end_pfn;
889    struct page *start_page, *end_page;
890
891    start_pfn = page_to_pfn(page);
892    start_pfn = start_pfn & ~(pageblock_nr_pages-1);
893    start_page = pfn_to_page(start_pfn);
894    end_page = start_page + pageblock_nr_pages - 1;
895    end_pfn = start_pfn + pageblock_nr_pages - 1;
896
897    /* Do not cross zone boundaries */
898    if (start_pfn < zone->zone_start_pfn)
899        start_page = page;
900    if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
901        return 0;
902
903    return move_freepages(zone, start_page, end_page, migratetype);
904}
905
906static void change_pageblock_range(struct page *pageblock_page,
907                    int start_order, int migratetype)
908{
909    int nr_pageblocks = 1 << (start_order - pageblock_order);
910
911    while (nr_pageblocks--) {
912        set_pageblock_migratetype(pageblock_page, migratetype);
913        pageblock_page += pageblock_nr_pages;
914    }
915}
916
917/* Remove an element from the buddy allocator from the fallback list */
918static inline struct page *
919__rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
920{
921    struct free_area * area;
922    int current_order;
923    struct page *page;
924    int migratetype, i;
925
926    /* Find the largest possible block of pages in the other list */
927    for (current_order = MAX_ORDER-1; current_order >= order;
928                        --current_order) {
929        for (i = 0; i < MIGRATE_TYPES - 1; i++) {
930            migratetype = fallbacks[start_migratetype][i];
931
932            /* MIGRATE_RESERVE handled later if necessary */
933            if (migratetype == MIGRATE_RESERVE)
934                continue;
935
936            area = &(zone->free_area[current_order]);
937            if (list_empty(&area->free_list[migratetype]))
938                continue;
939
940            page = list_entry(area->free_list[migratetype].next,
941                    struct page, lru);
942            area->nr_free--;
943
944            /*
945             * If breaking a large block of pages, move all free
946             * pages to the preferred allocation list. If falling
947             * back for a reclaimable kernel allocation, be more
948             * aggressive about taking ownership of free pages
949             */
950            if (unlikely(current_order >= (pageblock_order >> 1)) ||
951                    start_migratetype == MIGRATE_RECLAIMABLE ||
952                    page_group_by_mobility_disabled) {
953                unsigned long pages;
954                pages = move_freepages_block(zone, page,
955                                start_migratetype);
956
957                /* Claim the whole block if over half of it is free */
958                if (pages >= (1 << (pageblock_order-1)) ||
959                        page_group_by_mobility_disabled)
960                    set_pageblock_migratetype(page,
961                                start_migratetype);
962
963                migratetype = start_migratetype;
964            }
965
966            /* Remove the page from the freelists */
967            list_del(&page->lru);
968            rmv_page_order(page);
969
970            /* Take ownership for orders >= pageblock_order */
971            if (current_order >= pageblock_order)
972                change_pageblock_range(page, current_order,
973                            start_migratetype);
974
975            expand(zone, page, order, current_order, area, migratetype);
976
977            trace_mm_page_alloc_extfrag(page, order, current_order,
978                start_migratetype, migratetype);
979
980            return page;
981        }
982    }
983
984    return NULL;
985}
986
987/*
988 * Do the hard work of removing an element from the buddy allocator.
989 * Call me with the zone->lock already held.
990 */
991static struct page *__rmqueue(struct zone *zone, unsigned int order,
992                        int migratetype)
993{
994    struct page *page;
995
996retry_reserve:
997    page = __rmqueue_smallest(zone, order, migratetype);
998
999    if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1000        page = __rmqueue_fallback(zone, order, migratetype);
1001
1002        /*
1003         * Use MIGRATE_RESERVE rather than fail an allocation. goto
1004         * is used because __rmqueue_smallest is an inline function
1005         * and we want just one call site
1006         */
1007        if (!page) {
1008            migratetype = MIGRATE_RESERVE;
1009            goto retry_reserve;
1010        }
1011    }
1012
1013    trace_mm_page_alloc_zone_locked(page, order, migratetype);
1014    return page;
1015}
1016
1017/*
1018 * Obtain a specified number of elements from the buddy allocator, all under
1019 * a single hold of the lock, for efficiency. Add them to the supplied list.
1020 * Returns the number of new pages which were placed at *list.
1021 */
1022static int rmqueue_bulk(struct zone *zone, unsigned int order,
1023            unsigned long count, struct list_head *list,
1024            int migratetype, int cold)
1025{
1026    int i;
1027    
1028    spin_lock(&zone->lock);
1029    for (i = 0; i < count; ++i) {
1030        struct page *page = __rmqueue(zone, order, migratetype);
1031        if (unlikely(page == NULL))
1032            break;
1033
1034        /*
1035         * Split buddy pages returned by expand() are received here
1036         * in physical page order. The page is added to the callers and
1037         * list and the list head then moves forward. From the callers
1038         * perspective, the linked list is ordered by page number in
1039         * some conditions. This is useful for IO devices that can
1040         * merge IO requests if the physical pages are ordered
1041         * properly.
1042         */
1043        if (likely(cold == 0))
1044            list_add(&page->lru, list);
1045        else
1046            list_add_tail(&page->lru, list);
1047        set_page_private(page, migratetype);
1048        list = &page->lru;
1049    }
1050    __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1051    spin_unlock(&zone->lock);
1052    return i;
1053}
1054
1055#ifdef CONFIG_NUMA
1056/*
1057 * Called from the vmstat counter updater to drain pagesets of this
1058 * currently executing processor on remote nodes after they have
1059 * expired.
1060 *
1061 * Note that this function must be called with the thread pinned to
1062 * a single processor.
1063 */
1064void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1065{
1066    unsigned long flags;
1067    int to_drain;
1068
1069    local_irq_save(flags);
1070    if (pcp->count >= pcp->batch)
1071        to_drain = pcp->batch;
1072    else
1073        to_drain = pcp->count;
1074    free_pcppages_bulk(zone, to_drain, pcp);
1075    pcp->count -= to_drain;
1076    local_irq_restore(flags);
1077}
1078#endif
1079
1080/*
1081 * Drain pages of the indicated processor.
1082 *
1083 * The processor must either be the current processor and the
1084 * thread pinned to the current processor or a processor that
1085 * is not online.
1086 */
1087static void drain_pages(unsigned int cpu)
1088{
1089    unsigned long flags;
1090    struct zone *zone;
1091
1092    for_each_populated_zone(zone) {
1093        struct per_cpu_pageset *pset;
1094        struct per_cpu_pages *pcp;
1095
1096        local_irq_save(flags);
1097        pset = per_cpu_ptr(zone->pageset, cpu);
1098
1099        pcp = &pset->pcp;
1100        if (pcp->count) {
1101            free_pcppages_bulk(zone, pcp->count, pcp);
1102            pcp->count = 0;
1103        }
1104        local_irq_restore(flags);
1105    }
1106}
1107
1108/*
1109 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1110 */
1111void drain_local_pages(void *arg)
1112{
1113    drain_pages(smp_processor_id());
1114}
1115
1116/*
1117 * Spill all the per-cpu pages from all CPUs back into the buddy allocator
1118 */
1119void drain_all_pages(void)
1120{
1121    on_each_cpu(drain_local_pages, NULL, 1);
1122}
1123
1124#ifdef CONFIG_HIBERNATION
1125
1126void mark_free_pages(struct zone *zone)
1127{
1128    unsigned long pfn, max_zone_pfn;
1129    unsigned long flags;
1130    int order, t;
1131    struct list_head *curr;
1132
1133    if (!zone->spanned_pages)
1134        return;
1135
1136    spin_lock_irqsave(&zone->lock, flags);
1137
1138    max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
1139    for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1140        if (pfn_valid(pfn)) {
1141            struct page *page = pfn_to_page(pfn);
1142
1143            if (!swsusp_page_is_forbidden(page))
1144                swsusp_unset_page_free(page);
1145        }
1146
1147    for_each_migratetype_order(order, t) {
1148        list_for_each(curr, &zone->free_area[order].free_list[t]) {
1149            unsigned long i;
1150
1151            pfn = page_to_pfn(list_entry(curr, struct page, lru));
1152            for (i = 0; i < (1UL << order); i++)
1153                swsusp_set_page_free(pfn_to_page(pfn + i));
1154        }
1155    }
1156    spin_unlock_irqrestore(&zone->lock, flags);
1157}
1158#endif /* CONFIG_PM */
1159
1160/*
1161 * Free a 0-order page
1162 * cold == 1 ? free a cold page : free a hot page
1163 */
1164void free_hot_cold_page(struct page *page, int cold)
1165{
1166    struct zone *zone = page_zone(page);
1167    struct per_cpu_pages *pcp;
1168    unsigned long flags;
1169    int migratetype;
1170    int wasMlocked = __TestClearPageMlocked(page);
1171
1172    if (!free_pages_prepare(page, 0))
1173        return;
1174
1175    migratetype = get_pageblock_migratetype(page);
1176    set_page_private(page, migratetype);
1177    local_irq_save(flags);
1178    if (unlikely(wasMlocked))
1179        free_page_mlock(page);
1180    __count_vm_event(PGFREE);
1181
1182    /*
1183     * We only track unmovable, reclaimable and movable on pcp lists.
1184     * Free ISOLATE pages back to the allocator because they are being
1185     * offlined but treat RESERVE as movable pages so we can get those
1186     * areas back if necessary. Otherwise, we may have to free
1187     * excessively into the page allocator
1188     */
1189    if (migratetype >= MIGRATE_PCPTYPES) {
1190        if (unlikely(migratetype == MIGRATE_ISOLATE)) {
1191            free_one_page(zone, page, 0, migratetype);
1192            goto out;
1193        }
1194        migratetype = MIGRATE_MOVABLE;
1195    }
1196
1197    pcp = &this_cpu_ptr(zone->pageset)->pcp;
1198    if (cold)
1199        list_add_tail(&page->lru, &pcp->lists[migratetype]);
1200    else
1201        list_add(&page->lru, &pcp->lists[migratetype]);
1202    pcp->count++;
1203    if (pcp->count >= pcp->high) {
1204        free_pcppages_bulk(zone, pcp->batch, pcp);
1205        pcp->count -= pcp->batch;
1206    }
1207
1208out:
1209    local_irq_restore(flags);
1210}
1211
1212/*
1213 * split_page takes a non-compound higher-order page, and splits it into
1214 * n (1<<order) sub-pages: page[0..n]
1215 * Each sub-page must be freed individually.
1216 *
1217 * Note: this is probably too low level an operation for use in drivers.
1218 * Please consult with lkml before using this in your driver.
1219 */
1220void split_page(struct page *page, unsigned int order)
1221{
1222    int i;
1223
1224    VM_BUG_ON(PageCompound(page));
1225    VM_BUG_ON(!page_count(page));
1226
1227#ifdef CONFIG_KMEMCHECK
1228    /*
1229     * Split shadow pages too, because free(page[0]) would
1230     * otherwise free the whole shadow.
1231     */
1232    if (kmemcheck_page_is_tracked(page))
1233        split_page(virt_to_page(page[0].shadow), order);
1234#endif
1235
1236    for (i = 1; i < (1 << order); i++)
1237        set_page_refcounted(page + i);
1238}
1239
1240/*
1241 * Similar to split_page except the page is already free. As this is only
1242 * being used for migration, the migratetype of the block also changes.
1243 * As this is called with interrupts disabled, the caller is responsible
1244 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1245 * are enabled.
1246 *
1247 * Note: this is probably too low level an operation for use in drivers.
1248 * Please consult with lkml before using this in your driver.
1249 */
1250int split_free_page(struct page *page)
1251{
1252    unsigned int order;
1253    unsigned long watermark;
1254    struct zone *zone;
1255
1256    BUG_ON(!PageBuddy(page));
1257
1258    zone = page_zone(page);
1259    order = page_order(page);
1260
1261    /* Obey watermarks as if the page was being allocated */
1262    watermark = low_wmark_pages(zone) + (1 << order);
1263    if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1264        return 0;
1265
1266    /* Remove page from free list */
1267    list_del(&page->lru);
1268    zone->free_area[order].nr_free--;
1269    rmv_page_order(page);
1270    __mod_zone_page_state(zone, NR_FREE_PAGES, -(1UL << order));
1271
1272    /* Split into individual pages */
1273    set_page_refcounted(page);
1274    split_page(page, order);
1275
1276    if (order >= pageblock_order - 1) {
1277        struct page *endpage = page + (1 << order) - 1;
1278        for (; page < endpage; page += pageblock_nr_pages)
1279            set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1280    }
1281
1282    return 1 << order;
1283}
1284
1285/*
1286 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1287 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1288 * or two.
1289 */
1290static inline
1291struct page *buffered_rmqueue(struct zone *preferred_zone,
1292            struct zone *zone, int order, gfp_t gfp_flags,
1293            int migratetype)
1294{
1295    unsigned long flags;
1296    struct page *page;
1297    int cold = !!(gfp_flags & __GFP_COLD);
1298
1299again:
1300    if (likely(order == 0)) {
1301        struct per_cpu_pages *pcp;
1302        struct list_head *list;
1303
1304        local_irq_save(flags);
1305        pcp = &this_cpu_ptr(zone->pageset)->pcp;
1306        list = &pcp->lists[migratetype];
1307        if (list_empty(list)) {
1308            pcp->count += rmqueue_bulk(zone, 0,
1309                    pcp->batch, list,
1310                    migratetype, cold);
1311            if (unlikely(list_empty(list)))
1312                goto failed;
1313        }
1314
1315        if (cold)
1316            page = list_entry(list->prev, struct page, lru);
1317        else
1318            page = list_entry(list->next, struct page, lru);
1319
1320        list_del(&page->lru);
1321        pcp->count--;
1322    } else {
1323        if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1324            /*
1325             * __GFP_NOFAIL is not to be used in new code.
1326             *
1327             * All __GFP_NOFAIL callers should be fixed so that they
1328             * properly detect and handle allocation failures.
1329             *
1330             * We most definitely don't want callers attempting to
1331             * allocate greater than order-1 page units with
1332             * __GFP_NOFAIL.
1333             */
1334            WARN_ON_ONCE(order > 1);
1335        }
1336        spin_lock_irqsave(&zone->lock, flags);
1337        page = __rmqueue(zone, order, migratetype);
1338        spin_unlock(&zone->lock);
1339        if (!page)
1340            goto failed;
1341        __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << order));
1342    }
1343
1344    __count_zone_vm_events(PGALLOC, zone, 1 << order);
1345    zone_statistics(preferred_zone, zone, gfp_flags);
1346    local_irq_restore(flags);
1347
1348    VM_BUG_ON(bad_range(zone, page));
1349    if (prep_new_page(page, order, gfp_flags))
1350        goto again;
1351    return page;
1352
1353failed:
1354    local_irq_restore(flags);
1355    return NULL;
1356}
1357
1358/* The ALLOC_WMARK bits are used as an index to zone->watermark */
1359#define ALLOC_WMARK_MIN WMARK_MIN
1360#define ALLOC_WMARK_LOW WMARK_LOW
1361#define ALLOC_WMARK_HIGH WMARK_HIGH
1362#define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */
1363
1364/* Mask to get the watermark bits */
1365#define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1)
1366
1367#define ALLOC_HARDER 0x10 /* try to alloc harder */
1368#define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
1369#define ALLOC_CPUSET 0x40 /* check for correct cpuset */
1370
1371#ifdef CONFIG_FAIL_PAGE_ALLOC
1372
1373static struct fail_page_alloc_attr {
1374    struct fault_attr attr;
1375
1376    u32 ignore_gfp_highmem;
1377    u32 ignore_gfp_wait;
1378    u32 min_order;
1379
1380#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1381
1382    struct dentry *ignore_gfp_highmem_file;
1383    struct dentry *ignore_gfp_wait_file;
1384    struct dentry *min_order_file;
1385
1386#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1387
1388} fail_page_alloc = {
1389    .attr = FAULT_ATTR_INITIALIZER,
1390    .ignore_gfp_wait = 1,
1391    .ignore_gfp_highmem = 1,
1392    .min_order = 1,
1393};
1394
1395static int __init setup_fail_page_alloc(char *str)
1396{
1397    return setup_fault_attr(&fail_page_alloc.attr, str);
1398}
1399__setup("fail_page_alloc=", setup_fail_page_alloc);
1400
1401static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1402{
1403    if (order < fail_page_alloc.min_order)
1404        return 0;
1405    if (gfp_mask & __GFP_NOFAIL)
1406        return 0;
1407    if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1408        return 0;
1409    if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1410        return 0;
1411
1412    return should_fail(&fail_page_alloc.attr, 1 << order);
1413}
1414
1415#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1416
1417static int __init fail_page_alloc_debugfs(void)
1418{
1419    mode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1420    struct dentry *dir;
1421    int err;
1422
1423    err = init_fault_attr_dentries(&fail_page_alloc.attr,
1424                       "fail_page_alloc");
1425    if (err)
1426        return err;
1427    dir = fail_page_alloc.attr.dentries.dir;
1428
1429    fail_page_alloc.ignore_gfp_wait_file =
1430        debugfs_create_bool("ignore-gfp-wait", mode, dir,
1431                      &fail_page_alloc.ignore_gfp_wait);
1432
1433    fail_page_alloc.ignore_gfp_highmem_file =
1434        debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1435                      &fail_page_alloc.ignore_gfp_highmem);
1436    fail_page_alloc.min_order_file =
1437        debugfs_create_u32("min-order", mode, dir,
1438                   &fail_page_alloc.min_order);
1439
1440    if (!fail_page_alloc.ignore_gfp_wait_file ||
1441            !fail_page_alloc.ignore_gfp_highmem_file ||
1442            !fail_page_alloc.min_order_file) {
1443        err = -ENOMEM;
1444        debugfs_remove(fail_page_alloc.ignore_gfp_wait_file);
1445        debugfs_remove(fail_page_alloc.ignore_gfp_highmem_file);
1446        debugfs_remove(fail_page_alloc.min_order_file);
1447        cleanup_fault_attr_dentries(&fail_page_alloc.attr);
1448    }
1449
1450    return err;
1451}
1452
1453late_initcall(fail_page_alloc_debugfs);
1454
1455#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1456
1457#else /* CONFIG_FAIL_PAGE_ALLOC */
1458
1459static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1460{
1461    return 0;
1462}
1463
1464#endif /* CONFIG_FAIL_PAGE_ALLOC */
1465
1466/*
1467 * Return true if free pages are above 'mark'. This takes into account the order
1468 * of the allocation.
1469 */
1470static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1471              int classzone_idx, int alloc_flags, long free_pages)
1472{
1473    /* free_pages my go negative - that's OK */
1474    long min = mark;
1475    int o;
1476
1477    free_pages -= (1 << order) + 1;
1478    if (alloc_flags & ALLOC_HIGH)
1479        min -= min / 2;
1480    if (alloc_flags & ALLOC_HARDER)
1481        min -= min / 4;
1482
1483    if (free_pages <= min + z->lowmem_reserve[classzone_idx])
1484        return false;
1485    for (o = 0; o < order; o++) {
1486        /* At the next order, this order's pages become unavailable */
1487        free_pages -= z->free_area[o].nr_free << o;
1488
1489        /* Require fewer higher order pages to be free */
1490        min >>= 1;
1491
1492        if (free_pages <= min)
1493            return false;
1494    }
1495    return true;
1496}
1497
1498bool zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1499              int classzone_idx, int alloc_flags)
1500{
1501    return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1502                    zone_page_state(z, NR_FREE_PAGES));
1503}
1504
1505bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark,
1506              int classzone_idx, int alloc_flags)
1507{
1508    long free_pages = zone_page_state(z, NR_FREE_PAGES);
1509
1510    if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1511        free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1512
1513    return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1514                                free_pages);
1515}
1516
1517#ifdef CONFIG_NUMA
1518/*
1519 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1520 * skip over zones that are not allowed by the cpuset, or that have
1521 * been recently (in last second) found to be nearly full. See further
1522 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1523 * that have to skip over a lot of full or unallowed zones.
1524 *
1525 * If the zonelist cache is present in the passed in zonelist, then
1526 * returns a pointer to the allowed node mask (either the current
1527 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1528 *
1529 * If the zonelist cache is not available for this zonelist, does
1530 * nothing and returns NULL.
1531 *
1532 * If the fullzones BITMAP in the zonelist cache is stale (more than
1533 * a second since last zap'd) then we zap it out (clear its bits.)
1534 *
1535 * We hold off even calling zlc_setup, until after we've checked the
1536 * first zone in the zonelist, on the theory that most allocations will
1537 * be satisfied from that first zone, so best to examine that zone as
1538 * quickly as we can.
1539 */
1540static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1541{
1542    struct zonelist_cache *zlc; /* cached zonelist speedup info */
1543    nodemask_t *allowednodes; /* zonelist_cache approximation */
1544
1545    zlc = zonelist->zlcache_ptr;
1546    if (!zlc)
1547        return NULL;
1548
1549    if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1550        bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1551        zlc->last_full_zap = jiffies;
1552    }
1553
1554    allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1555                    &cpuset_current_mems_allowed :
1556                    &node_states[N_HIGH_MEMORY];
1557    return allowednodes;
1558}
1559
1560/*
1561 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1562 * if it is worth looking at further for free memory:
1563 * 1) Check that the zone isn't thought to be full (doesn't have its
1564 * bit set in the zonelist_cache fullzones BITMAP).
1565 * 2) Check that the zones node (obtained from the zonelist_cache
1566 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1567 * Return true (non-zero) if zone is worth looking at further, or
1568 * else return false (zero) if it is not.
1569 *
1570 * This check -ignores- the distinction between various watermarks,
1571 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1572 * found to be full for any variation of these watermarks, it will
1573 * be considered full for up to one second by all requests, unless
1574 * we are so low on memory on all allowed nodes that we are forced
1575 * into the second scan of the zonelist.
1576 *
1577 * In the second scan we ignore this zonelist cache and exactly
1578 * apply the watermarks to all zones, even it is slower to do so.
1579 * We are low on memory in the second scan, and should leave no stone
1580 * unturned looking for a free page.
1581 */
1582static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1583                        nodemask_t *allowednodes)
1584{
1585    struct zonelist_cache *zlc; /* cached zonelist speedup info */
1586    int i; /* index of *z in zonelist zones */
1587    int n; /* node that zone *z is on */
1588
1589    zlc = zonelist->zlcache_ptr;
1590    if (!zlc)
1591        return 1;
1592
1593    i = z - zonelist->_zonerefs;
1594    n = zlc->z_to_n[i];
1595
1596    /* This zone is worth trying if it is allowed but not full */
1597    return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1598}
1599
1600/*
1601 * Given 'z' scanning a zonelist, set the corresponding bit in
1602 * zlc->fullzones, so that subsequent attempts to allocate a page
1603 * from that zone don't waste time re-examining it.
1604 */
1605static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1606{
1607    struct zonelist_cache *zlc; /* cached zonelist speedup info */
1608    int i; /* index of *z in zonelist zones */
1609
1610    zlc = zonelist->zlcache_ptr;
1611    if (!zlc)
1612        return;
1613
1614    i = z - zonelist->_zonerefs;
1615
1616    set_bit(i, zlc->fullzones);
1617}
1618
1619#else /* CONFIG_NUMA */
1620
1621static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1622{
1623    return NULL;
1624}
1625
1626static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1627                nodemask_t *allowednodes)
1628{
1629    return 1;
1630}
1631
1632static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1633{
1634}
1635#endif /* CONFIG_NUMA */
1636
1637/*
1638 * get_page_from_freelist goes through the zonelist trying to allocate
1639 * a page.
1640 */
1641static struct page *
1642get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1643        struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1644        struct zone *preferred_zone, int migratetype)
1645{
1646    struct zoneref *z;
1647    struct page *page = NULL;
1648    int classzone_idx;
1649    struct zone *zone;
1650    nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1651    int zlc_active = 0; /* set if using zonelist_cache */
1652    int did_zlc_setup = 0; /* just call zlc_setup() one time */
1653
1654    classzone_idx = zone_idx(preferred_zone);
1655zonelist_scan:
1656    /*
1657     * Scan zonelist, looking for a zone with enough free.
1658     * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1659     */
1660    for_each_zone_zonelist_nodemask(zone, z, zonelist,
1661                        high_zoneidx, nodemask) {
1662        if (NUMA_BUILD && zlc_active &&
1663            !zlc_zone_worth_trying(zonelist, z, allowednodes))
1664                continue;
1665        if ((alloc_flags & ALLOC_CPUSET) &&
1666            !cpuset_zone_allowed_softwall(zone, gfp_mask))
1667                goto try_next_zone;
1668
1669        BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1670        if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1671            unsigned long mark;
1672            int ret;
1673
1674            mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1675            if (zone_watermark_ok(zone, order, mark,
1676                    classzone_idx, alloc_flags))
1677                goto try_this_zone;
1678
1679            if (zone_reclaim_mode == 0)
1680                goto this_zone_full;
1681
1682            ret = zone_reclaim(zone, gfp_mask, order);
1683            switch (ret) {
1684            case ZONE_RECLAIM_NOSCAN:
1685                /* did not scan */
1686                goto try_next_zone;
1687            case ZONE_RECLAIM_FULL:
1688                /* scanned but unreclaimable */
1689                goto this_zone_full;
1690            default:
1691                /* did we reclaim enough */
1692                if (!zone_watermark_ok(zone, order, mark,
1693                        classzone_idx, alloc_flags))
1694                    goto this_zone_full;
1695            }
1696        }
1697
1698try_this_zone:
1699        page = buffered_rmqueue(preferred_zone, zone, order,
1700                        gfp_mask, migratetype);
1701        if (page)
1702            break;
1703this_zone_full:
1704        if (NUMA_BUILD)
1705            zlc_mark_zone_full(zonelist, z);
1706try_next_zone:
1707        if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) {
1708            /*
1709             * we do zlc_setup after the first zone is tried but only
1710             * if there are multiple nodes make it worthwhile
1711             */
1712            allowednodes = zlc_setup(zonelist, alloc_flags);
1713            zlc_active = 1;
1714            did_zlc_setup = 1;
1715        }
1716    }
1717
1718    if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) {
1719        /* Disable zlc cache for second zonelist scan */
1720        zlc_active = 0;
1721        goto zonelist_scan;
1722    }
1723    return page;
1724}
1725
1726/*
1727 * Large machines with many possible nodes should not always dump per-node
1728 * meminfo in irq context.
1729 */
1730static inline bool should_suppress_show_mem(void)
1731{
1732    bool ret = false;
1733
1734#if NODES_SHIFT > 8
1735    ret = in_interrupt();
1736#endif
1737    return ret;
1738}
1739
1740static DEFINE_RATELIMIT_STATE(nopage_rs,
1741        DEFAULT_RATELIMIT_INTERVAL,
1742        DEFAULT_RATELIMIT_BURST);
1743
1744void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
1745{
1746    va_list args;
1747    unsigned int filter = SHOW_MEM_FILTER_NODES;
1748
1749    if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
1750        return;
1751
1752    /*
1753     * This documents exceptions given to allocations in certain
1754     * contexts that are allowed to allocate outside current's set
1755     * of allowed nodes.
1756     */
1757    if (!(gfp_mask & __GFP_NOMEMALLOC))
1758        if (test_thread_flag(TIF_MEMDIE) ||
1759            (current->flags & (PF_MEMALLOC | PF_EXITING)))
1760            filter &= ~SHOW_MEM_FILTER_NODES;
1761    if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
1762        filter &= ~SHOW_MEM_FILTER_NODES;
1763
1764    if (fmt) {
1765        printk(KERN_WARNING);
1766        va_start(args, fmt);
1767        vprintk(fmt, args);
1768        va_end(args);
1769    }
1770
1771    pr_warning("%s: page allocation failure: order:%d, mode:0x%x\n",
1772           current->comm, order, gfp_mask);
1773
1774    dump_stack();
1775    if (!should_suppress_show_mem())
1776        show_mem(filter);
1777}
1778
1779static inline int
1780should_alloc_retry(gfp_t gfp_mask, unsigned int order,
1781                unsigned long pages_reclaimed)
1782{
1783    /* Do not loop if specifically requested */
1784    if (gfp_mask & __GFP_NORETRY)
1785        return 0;
1786
1787    /*
1788     * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
1789     * means __GFP_NOFAIL, but that may not be true in other
1790     * implementations.
1791     */
1792    if (order <= PAGE_ALLOC_COSTLY_ORDER)
1793        return 1;
1794
1795    /*
1796     * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
1797     * specified, then we retry until we no longer reclaim any pages
1798     * (above), or we've reclaimed an order of pages at least as
1799     * large as the allocation's order. In both cases, if the
1800     * allocation still fails, we stop retrying.
1801     */
1802    if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
1803        return 1;
1804
1805    /*
1806     * Don't let big-order allocations loop unless the caller
1807     * explicitly requests that.
1808     */
1809    if (gfp_mask & __GFP_NOFAIL)
1810        return 1;
1811
1812    return 0;
1813}
1814
1815static inline struct page *
1816__alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
1817    struct zonelist *zonelist, enum zone_type high_zoneidx,
1818    nodemask_t *nodemask, struct zone *preferred_zone,
1819    int migratetype)
1820{
1821    struct page *page;
1822
1823    /* Acquire the OOM killer lock for the zones in zonelist */
1824    if (!try_set_zonelist_oom(zonelist, gfp_mask)) {
1825        schedule_timeout_uninterruptible(1);
1826        return NULL;
1827    }
1828
1829    /*
1830     * Go through the zonelist yet one more time, keep very high watermark
1831     * here, this is only to catch a parallel oom killing, we must fail if
1832     * we're still under heavy pressure.
1833     */
1834    page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
1835        order, zonelist, high_zoneidx,
1836        ALLOC_WMARK_HIGH|ALLOC_CPUSET,
1837        preferred_zone, migratetype);
1838    if (page)
1839        goto out;
1840
1841    if (!(gfp_mask & __GFP_NOFAIL)) {
1842        /* The OOM killer will not help higher order allocs */
1843        if (order > PAGE_ALLOC_COSTLY_ORDER)
1844            goto out;
1845        /* The OOM killer does not needlessly kill tasks for lowmem */
1846        if (high_zoneidx < ZONE_NORMAL)
1847            goto out;
1848        /*
1849         * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
1850         * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
1851         * The caller should handle page allocation failure by itself if
1852         * it specifies __GFP_THISNODE.
1853         * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
1854         */
1855        if (gfp_mask & __GFP_THISNODE)
1856            goto out;
1857    }
1858    /* Exhausted what can be done so it's blamo time */
1859    out_of_memory(zonelist, gfp_mask, order, nodemask);
1860
1861out:
1862    clear_zonelist_oom(zonelist, gfp_mask);
1863    return page;
1864}
1865
1866#ifdef CONFIG_COMPACTION
1867/* Try memory compaction for high-order allocations before reclaim */
1868static struct page *
1869__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
1870    struct zonelist *zonelist, enum zone_type high_zoneidx,
1871    nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1872    int migratetype, unsigned long *did_some_progress,
1873    bool sync_migration)
1874{
1875    struct page *page;
1876
1877    if (!order || compaction_deferred(preferred_zone))
1878        return NULL;
1879
1880    current->flags |= PF_MEMALLOC;
1881    *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
1882                        nodemask, sync_migration);
1883    current->flags &= ~PF_MEMALLOC;
1884    if (*did_some_progress != COMPACT_SKIPPED) {
1885
1886        /* Page migration frees to the PCP lists but we want merging */
1887        drain_pages(get_cpu());
1888        put_cpu();
1889
1890        page = get_page_from_freelist(gfp_mask, nodemask,
1891                order, zonelist, high_zoneidx,
1892                alloc_flags, preferred_zone,
1893                migratetype);
1894        if (page) {
1895            preferred_zone->compact_considered = 0;
1896            preferred_zone->compact_defer_shift = 0;
1897            count_vm_event(COMPACTSUCCESS);
1898            return page;
1899        }
1900
1901        /*
1902         * It's bad if compaction run occurs and fails.
1903         * The most likely reason is that pages exist,
1904         * but not enough to satisfy watermarks.
1905         */
1906        count_vm_event(COMPACTFAIL);
1907        defer_compaction(preferred_zone);
1908
1909        cond_resched();
1910    }
1911
1912    return NULL;
1913}
1914#else
1915static inline struct page *
1916__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
1917    struct zonelist *zonelist, enum zone_type high_zoneidx,
1918    nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1919    int migratetype, unsigned long *did_some_progress,
1920    bool sync_migration)
1921{
1922    return NULL;
1923}
1924#endif /* CONFIG_COMPACTION */
1925
1926/* The really slow allocator path where we enter direct reclaim */
1927static inline struct page *
1928__alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
1929    struct zonelist *zonelist, enum zone_type high_zoneidx,
1930    nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1931    int migratetype, unsigned long *did_some_progress)
1932{
1933    struct page *page = NULL;
1934    struct reclaim_state reclaim_state;
1935    bool drained = false;
1936
1937    cond_resched();
1938
1939    /* We now go into synchronous reclaim */
1940    cpuset_memory_pressure_bump();
1941    current->flags |= PF_MEMALLOC;
1942    lockdep_set_current_reclaim_state(gfp_mask);
1943    reclaim_state.reclaimed_slab = 0;
1944    current->reclaim_state = &reclaim_state;
1945
1946    *did_some_progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
1947
1948    current->reclaim_state = NULL;
1949    lockdep_clear_current_reclaim_state();
1950    current->flags &= ~PF_MEMALLOC;
1951
1952    cond_resched();
1953
1954    if (unlikely(!(*did_some_progress)))
1955        return NULL;
1956
1957retry:
1958    page = get_page_from_freelist(gfp_mask, nodemask, order,
1959                    zonelist, high_zoneidx,
1960                    alloc_flags, preferred_zone,
1961                    migratetype);
1962
1963    /*
1964     * If an allocation failed after direct reclaim, it could be because
1965     * pages are pinned on the per-cpu lists. Drain them and try again
1966     */
1967    if (!page && !drained) {
1968        drain_all_pages();
1969        drained = true;
1970        goto retry;
1971    }
1972
1973    return page;
1974}
1975
1976/*
1977 * This is called in the allocator slow-path if the allocation request is of
1978 * sufficient urgency to ignore watermarks and take other desperate measures
1979 */
1980static inline struct page *
1981__alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
1982    struct zonelist *zonelist, enum zone_type high_zoneidx,
1983    nodemask_t *nodemask, struct zone *preferred_zone,
1984    int migratetype)
1985{
1986    struct page *page;
1987
1988    do {
1989        page = get_page_from_freelist(gfp_mask, nodemask, order,
1990            zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
1991            preferred_zone, migratetype);
1992
1993        if (!page && gfp_mask & __GFP_NOFAIL)
1994            wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
1995    } while (!page && (gfp_mask & __GFP_NOFAIL));
1996
1997    return page;
1998}
1999
2000static inline
2001void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
2002                        enum zone_type high_zoneidx,
2003                        enum zone_type classzone_idx)
2004{
2005    struct zoneref *z;
2006    struct zone *zone;
2007
2008    for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
2009        wakeup_kswapd(zone, order, classzone_idx);
2010}
2011
2012static inline int
2013gfp_to_alloc_flags(gfp_t gfp_mask)
2014{
2015    int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2016    const gfp_t wait = gfp_mask & __GFP_WAIT;
2017
2018    /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2019    BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2020
2021    /*
2022     * The caller may dip into page reserves a bit more if the caller
2023     * cannot run direct reclaim, or if the caller has realtime scheduling
2024     * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2025     * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2026     */
2027    alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2028
2029    if (!wait) {
2030        /*
2031         * Not worth trying to allocate harder for
2032         * __GFP_NOMEMALLOC even if it can't schedule.
2033         */
2034        if (!(gfp_mask & __GFP_NOMEMALLOC))
2035            alloc_flags |= ALLOC_HARDER;
2036        /*
2037         * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2038         * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
2039         */
2040        alloc_flags &= ~ALLOC_CPUSET;
2041    } else if (unlikely(rt_task(current)) && !in_interrupt())
2042        alloc_flags |= ALLOC_HARDER;
2043
2044    if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2045        if (!in_interrupt() &&
2046            ((current->flags & PF_MEMALLOC) ||
2047             unlikely(test_thread_flag(TIF_MEMDIE))))
2048            alloc_flags |= ALLOC_NO_WATERMARKS;
2049    }
2050
2051    return alloc_flags;
2052}
2053
2054static inline struct page *
2055__alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2056    struct zonelist *zonelist, enum zone_type high_zoneidx,
2057    nodemask_t *nodemask, struct zone *preferred_zone,
2058    int migratetype)
2059{
2060    const gfp_t wait = gfp_mask & __GFP_WAIT;
2061    struct page *page = NULL;
2062    int alloc_flags;
2063    unsigned long pages_reclaimed = 0;
2064    unsigned long did_some_progress;
2065    bool sync_migration = false;
2066
2067    /*
2068     * In the slowpath, we sanity check order to avoid ever trying to
2069     * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2070     * be using allocators in order of preference for an area that is
2071     * too large.
2072     */
2073    if (order >= MAX_ORDER) {
2074        WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2075        return NULL;
2076    }
2077
2078    /*
2079     * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2080     * __GFP_NOWARN set) should not cause reclaim since the subsystem
2081     * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2082     * using a larger set of nodes after it has established that the
2083     * allowed per node queues are empty and that nodes are
2084     * over allocated.
2085     */
2086    if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2087        goto nopage;
2088
2089restart:
2090    if (!(gfp_mask & __GFP_NO_KSWAPD))
2091        wake_all_kswapd(order, zonelist, high_zoneidx,
2092                        zone_idx(preferred_zone));
2093
2094    /*
2095     * OK, we're below the kswapd watermark and have kicked background
2096     * reclaim. Now things get more complex, so set up alloc_flags according
2097     * to how we want to proceed.
2098     */
2099    alloc_flags = gfp_to_alloc_flags(gfp_mask);
2100
2101    /*
2102     * Find the true preferred zone if the allocation is unconstrained by
2103     * cpusets.
2104     */
2105    if (!(alloc_flags & ALLOC_CPUSET) && !nodemask)
2106        first_zones_zonelist(zonelist, high_zoneidx, NULL,
2107                    &preferred_zone);
2108
2109rebalance:
2110    /* This is the last chance, in general, before the goto nopage. */
2111    page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2112            high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2113            preferred_zone, migratetype);
2114    if (page)
2115        goto got_pg;
2116
2117    /* Allocate without watermarks if the context allows */
2118    if (alloc_flags & ALLOC_NO_WATERMARKS) {
2119        page = __alloc_pages_high_priority(gfp_mask, order,
2120                zonelist, high_zoneidx, nodemask,
2121                preferred_zone, migratetype);
2122        if (page)
2123            goto got_pg;
2124    }
2125
2126    /* Atomic allocations - we can't balance anything */
2127    if (!wait)
2128        goto nopage;
2129
2130    /* Avoid recursion of direct reclaim */
2131    if (current->flags & PF_MEMALLOC)
2132        goto nopage;
2133
2134    /* Avoid allocations with no watermarks from looping endlessly */
2135    if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2136        goto nopage;
2137
2138    /*
2139     * Try direct compaction. The first pass is asynchronous. Subsequent
2140     * attempts after direct reclaim are synchronous
2141     */
2142    page = __alloc_pages_direct_compact(gfp_mask, order,
2143                    zonelist, high_zoneidx,
2144                    nodemask,
2145                    alloc_flags, preferred_zone,
2146                    migratetype, &did_some_progress,
2147                    sync_migration);
2148    if (page)
2149        goto got_pg;
2150    sync_migration = true;
2151
2152    /* Try direct reclaim and then allocating */
2153    page = __alloc_pages_direct_reclaim(gfp_mask, order,
2154                    zonelist, high_zoneidx,
2155                    nodemask,
2156                    alloc_flags, preferred_zone,
2157                    migratetype, &did_some_progress);
2158    if (page)
2159        goto got_pg;
2160
2161    /*
2162     * If we failed to make any progress reclaiming, then we are
2163     * running out of options and have to consider going OOM
2164     */
2165    if (!did_some_progress) {
2166        if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
2167            if (oom_killer_disabled)
2168                goto nopage;
2169            page = __alloc_pages_may_oom(gfp_mask, order,
2170                    zonelist, high_zoneidx,
2171                    nodemask, preferred_zone,
2172                    migratetype);
2173            if (page)
2174                goto got_pg;
2175
2176            if (!(gfp_mask & __GFP_NOFAIL)) {
2177                /*
2178                 * The oom killer is not called for high-order
2179                 * allocations that may fail, so if no progress
2180                 * is being made, there are no other options and
2181                 * retrying is unlikely to help.
2182                 */
2183                if (order > PAGE_ALLOC_COSTLY_ORDER)
2184                    goto nopage;
2185                /*
2186                 * The oom killer is not called for lowmem
2187                 * allocations to prevent needlessly killing
2188                 * innocent tasks.
2189                 */
2190                if (high_zoneidx < ZONE_NORMAL)
2191                    goto nopage;
2192            }
2193
2194            goto restart;
2195        }
2196    }
2197
2198    /* Check if we should retry the allocation */
2199    pages_reclaimed += did_some_progress;
2200    if (should_alloc_retry(gfp_mask, order, pages_reclaimed)) {
2201        /* Wait for some write requests to complete then retry */
2202        wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2203        goto rebalance;
2204    } else {
2205        /*
2206         * High-order allocations do not necessarily loop after
2207         * direct reclaim and reclaim/compaction depends on compaction
2208         * being called after reclaim so call directly if necessary
2209         */
2210        page = __alloc_pages_direct_compact(gfp_mask, order,
2211                    zonelist, high_zoneidx,
2212                    nodemask,
2213                    alloc_flags, preferred_zone,
2214                    migratetype, &did_some_progress,
2215                    sync_migration);
2216        if (page)
2217            goto got_pg;
2218    }
2219
2220nopage:
2221    warn_alloc_failed(gfp_mask, order, NULL);
2222    return page;
2223got_pg:
2224    if (kmemcheck_enabled)
2225        kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2226    return page;
2227
2228}
2229
2230/*
2231 * This is the 'heart' of the zoned buddy allocator.
2232 */
2233struct page *
2234__alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2235            struct zonelist *zonelist, nodemask_t *nodemask)
2236{
2237    enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2238    struct zone *preferred_zone;
2239    struct page *page;
2240    int migratetype = allocflags_to_migratetype(gfp_mask);
2241
2242    gfp_mask &= gfp_allowed_mask;
2243
2244    lockdep_trace_alloc(gfp_mask);
2245
2246    might_sleep_if(gfp_mask & __GFP_WAIT);
2247
2248    if (should_fail_alloc_page(gfp_mask, order))
2249        return NULL;
2250
2251    /*
2252     * Check the zones suitable for the gfp_mask contain at least one
2253     * valid zone. It's possible to have an empty zonelist as a result
2254     * of GFP_THISNODE and a memoryless node
2255     */
2256    if (unlikely(!zonelist->_zonerefs->zone))
2257        return NULL;
2258
2259    get_mems_allowed();
2260    /* The preferred zone is used for statistics later */
2261    first_zones_zonelist(zonelist, high_zoneidx,
2262                nodemask ? : &cpuset_current_mems_allowed,
2263                &preferred_zone);
2264    if (!preferred_zone) {
2265        put_mems_allowed();
2266        return NULL;
2267    }
2268
2269    /* First allocation attempt */
2270    page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2271            zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET,
2272            preferred_zone, migratetype);
2273    if (unlikely(!page))
2274        page = __alloc_pages_slowpath(gfp_mask, order,
2275                zonelist, high_zoneidx, nodemask,
2276                preferred_zone, migratetype);
2277    put_mems_allowed();
2278
2279    trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2280    return page;
2281}
2282EXPORT_SYMBOL(__alloc_pages_nodemask);
2283
2284/*
2285 * Common helper functions.
2286 */
2287unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2288{
2289    struct page *page;
2290
2291    /*
2292     * __get_free_pages() returns a 32-bit address, which cannot represent
2293     * a highmem page
2294     */
2295    VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2296
2297    page = alloc_pages(gfp_mask, order);
2298    if (!page)
2299        return 0;
2300    return (unsigned long) page_address(page);
2301}
2302EXPORT_SYMBOL(__get_free_pages);
2303
2304unsigned long get_zeroed_page(gfp_t gfp_mask)
2305{
2306    return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2307}
2308EXPORT_SYMBOL(get_zeroed_page);
2309
2310void __pagevec_free(struct pagevec *pvec)
2311{
2312    int i = pagevec_count(pvec);
2313
2314    while (--i >= 0) {
2315        trace_mm_pagevec_free(pvec->pages[i], pvec->cold);
2316        free_hot_cold_page(pvec->pages[i], pvec->cold);
2317    }
2318}
2319
2320void __free_pages(struct page *page, unsigned int order)
2321{
2322    if (put_page_testzero(page)) {
2323        if (order == 0)
2324            free_hot_cold_page(page, 0);
2325        else
2326            __free_pages_ok(page, order);
2327    }
2328}
2329
2330EXPORT_SYMBOL(__free_pages);
2331
2332void free_pages(unsigned long addr, unsigned int order)
2333{
2334    if (addr != 0) {
2335        VM_BUG_ON(!virt_addr_valid((void *)addr));
2336        __free_pages(virt_to_page((void *)addr), order);
2337    }
2338}
2339
2340EXPORT_SYMBOL(free_pages);
2341
2342static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
2343{
2344    if (addr) {
2345        unsigned long alloc_end = addr + (PAGE_SIZE << order);
2346        unsigned long used = addr + PAGE_ALIGN(size);
2347
2348        split_page(virt_to_page((void *)addr), order);
2349        while (used < alloc_end) {
2350            free_page(used);
2351            used += PAGE_SIZE;
2352        }
2353    }
2354    return (void *)addr;
2355}
2356
2357/**
2358 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2359 * @size: the number of bytes to allocate
2360 * @gfp_mask: GFP flags for the allocation
2361 *
2362 * This function is similar to alloc_pages(), except that it allocates the
2363 * minimum number of pages to satisfy the request. alloc_pages() can only
2364 * allocate memory in power-of-two pages.
2365 *
2366 * This function is also limited by MAX_ORDER.
2367 *
2368 * Memory allocated by this function must be released by free_pages_exact().
2369 */
2370void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2371{
2372    unsigned int order = get_order(size);
2373    unsigned long addr;
2374
2375    addr = __get_free_pages(gfp_mask, order);
2376    return make_alloc_exact(addr, order, size);
2377}
2378EXPORT_SYMBOL(alloc_pages_exact);
2379
2380/**
2381 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2382 * pages on a node.
2383 * @nid: the preferred node ID where memory should be allocated
2384 * @size: the number of bytes to allocate
2385 * @gfp_mask: GFP flags for the allocation
2386 *
2387 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2388 * back.
2389 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2390 * but is not exact.
2391 */
2392void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
2393{
2394    unsigned order = get_order(size);
2395    struct page *p = alloc_pages_node(nid, gfp_mask, order);
2396    if (!p)
2397        return NULL;
2398    return make_alloc_exact((unsigned long)page_address(p), order, size);
2399}
2400EXPORT_SYMBOL(alloc_pages_exact_nid);
2401
2402/**
2403 * free_pages_exact - release memory allocated via alloc_pages_exact()
2404 * @virt: the value returned by alloc_pages_exact.
2405 * @size: size of allocation, same value as passed to alloc_pages_exact().
2406 *
2407 * Release the memory allocated by a previous call to alloc_pages_exact.
2408 */
2409void free_pages_exact(void *virt, size_t size)
2410{
2411    unsigned long addr = (unsigned long)virt;
2412    unsigned long end = addr + PAGE_ALIGN(size);
2413
2414    while (addr < end) {
2415        free_page(addr);
2416        addr += PAGE_SIZE;
2417    }
2418}
2419EXPORT_SYMBOL(free_pages_exact);
2420
2421static unsigned int nr_free_zone_pages(int offset)
2422{
2423    struct zoneref *z;
2424    struct zone *zone;
2425
2426    /* Just pick one node, since fallback list is circular */
2427    unsigned int sum = 0;
2428
2429    struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2430
2431    for_each_zone_zonelist(zone, z, zonelist, offset) {
2432        unsigned long size = zone->present_pages;
2433        unsigned long high = high_wmark_pages(zone);
2434        if (size > high)
2435            sum += size - high;
2436    }
2437
2438    return sum;
2439}
2440
2441/*
2442 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2443 */
2444unsigned int nr_free_buffer_pages(void)
2445{
2446    return nr_free_zone_pages(gfp_zone(GFP_USER));
2447}
2448EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2449
2450/*
2451 * Amount of free RAM allocatable within all zones
2452 */
2453unsigned int nr_free_pagecache_pages(void)
2454{
2455    return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2456}
2457
2458static inline void show_node(struct zone *zone)
2459{
2460    if (NUMA_BUILD)
2461        printk("Node %d ", zone_to_nid(zone));
2462}
2463
2464void si_meminfo(struct sysinfo *val)
2465{
2466    val->totalram = totalram_pages;
2467    val->sharedram = 0;
2468    val->freeram = global_page_state(NR_FREE_PAGES);
2469    val->bufferram = nr_blockdev_pages();
2470    val->totalhigh = totalhigh_pages;
2471    val->freehigh = nr_free_highpages();
2472    val->mem_unit = PAGE_SIZE;
2473}
2474
2475EXPORT_SYMBOL(si_meminfo);
2476
2477#ifdef CONFIG_NUMA
2478void si_meminfo_node(struct sysinfo *val, int nid)
2479{
2480    pg_data_t *pgdat = NODE_DATA(nid);
2481
2482    val->totalram = pgdat->node_present_pages;
2483    val->freeram = node_page_state(nid, NR_FREE_PAGES);
2484#ifdef CONFIG_HIGHMEM
2485    val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
2486    val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2487            NR_FREE_PAGES);
2488#else
2489    val->totalhigh = 0;
2490    val->freehigh = 0;
2491#endif
2492    val->mem_unit = PAGE_SIZE;
2493}
2494#endif
2495
2496/*
2497 * Determine whether the node should be displayed or not, depending on whether
2498 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
2499 */
2500bool skip_free_areas_node(unsigned int flags, int nid)
2501{
2502    bool ret = false;
2503
2504    if (!(flags & SHOW_MEM_FILTER_NODES))
2505        goto out;
2506
2507    get_mems_allowed();
2508    ret = !node_isset(nid, cpuset_current_mems_allowed);
2509    put_mems_allowed();
2510out:
2511    return ret;
2512}
2513
2514#define K(x) ((x) << (PAGE_SHIFT-10))
2515
2516/*
2517 * Show free area list (used inside shift_scroll-lock stuff)
2518 * We also calculate the percentage fragmentation. We do this by counting the
2519 * memory on each free list with the exception of the first item on the list.
2520 * Suppresses nodes that are not allowed by current's cpuset if
2521 * SHOW_MEM_FILTER_NODES is passed.
2522 */
2523void show_free_areas(unsigned int filter)
2524{
2525    int cpu;
2526    struct zone *zone;
2527
2528    for_each_populated_zone(zone) {
2529        if (skip_free_areas_node(filter, zone_to_nid(zone)))
2530            continue;
2531        show_node(zone);
2532        printk("%s per-cpu:\n", zone->name);
2533
2534        for_each_online_cpu(cpu) {
2535            struct per_cpu_pageset *pageset;
2536
2537            pageset = per_cpu_ptr(zone->pageset, cpu);
2538
2539            printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2540                   cpu, pageset->pcp.high,
2541                   pageset->pcp.batch, pageset->pcp.count);
2542        }
2543    }
2544
2545    printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2546        " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2547        " unevictable:%lu"
2548        " dirty:%lu writeback:%lu unstable:%lu\n"
2549        " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2550        " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n",
2551        global_page_state(NR_ACTIVE_ANON),
2552        global_page_state(NR_INACTIVE_ANON),
2553        global_page_state(NR_ISOLATED_ANON),
2554        global_page_state(NR_ACTIVE_FILE),
2555        global_page_state(NR_INACTIVE_FILE),
2556        global_page_state(NR_ISOLATED_FILE),
2557        global_page_state(NR_UNEVICTABLE),
2558        global_page_state(NR_FILE_DIRTY),
2559        global_page_state(NR_WRITEBACK),
2560        global_page_state(NR_UNSTABLE_NFS),
2561        global_page_state(NR_FREE_PAGES),
2562        global_page_state(NR_SLAB_RECLAIMABLE),
2563        global_page_state(NR_SLAB_UNRECLAIMABLE),
2564        global_page_state(NR_FILE_MAPPED),
2565        global_page_state(NR_SHMEM),
2566        global_page_state(NR_PAGETABLE),
2567        global_page_state(NR_BOUNCE));
2568
2569    for_each_populated_zone(zone) {
2570        int i;
2571
2572        if (skip_free_areas_node(filter, zone_to_nid(zone)))
2573            continue;
2574        show_node(zone);
2575        printk("%s"
2576            " free:%lukB"
2577            " min:%lukB"
2578            " low:%lukB"
2579            " high:%lukB"
2580            " active_anon:%lukB"
2581            " inactive_anon:%lukB"
2582            " active_file:%lukB"
2583            " inactive_file:%lukB"
2584            " unevictable:%lukB"
2585            " isolated(anon):%lukB"
2586            " isolated(file):%lukB"
2587            " present:%lukB"
2588            " mlocked:%lukB"
2589            " dirty:%lukB"
2590            " writeback:%lukB"
2591            " mapped:%lukB"
2592            " shmem:%lukB"
2593            " slab_reclaimable:%lukB"
2594            " slab_unreclaimable:%lukB"
2595            " kernel_stack:%lukB"
2596            " pagetables:%lukB"
2597            " unstable:%lukB"
2598            " bounce:%lukB"
2599            " writeback_tmp:%lukB"
2600            " pages_scanned:%lu"
2601            " all_unreclaimable? %s"
2602            "\n",
2603            zone->name,
2604            K(zone_page_state(zone, NR_FREE_PAGES)),
2605            K(min_wmark_pages(zone)),
2606            K(low_wmark_pages(zone)),
2607            K(high_wmark_pages(zone)),
2608            K(zone_page_state(zone, NR_ACTIVE_ANON)),
2609            K(zone_page_state(zone, NR_INACTIVE_ANON)),
2610            K(zone_page_state(zone, NR_ACTIVE_FILE)),
2611            K(zone_page_state(zone, NR_INACTIVE_FILE)),
2612            K(zone_page_state(zone, NR_UNEVICTABLE)),
2613            K(zone_page_state(zone, NR_ISOLATED_ANON)),
2614            K(zone_page_state(zone, NR_ISOLATED_FILE)),
2615            K(zone->present_pages),
2616            K(zone_page_state(zone, NR_MLOCK)),
2617            K(zone_page_state(zone, NR_FILE_DIRTY)),
2618            K(zone_page_state(zone, NR_WRITEBACK)),
2619            K(zone_page_state(zone, NR_FILE_MAPPED)),
2620            K(zone_page_state(zone, NR_SHMEM)),
2621            K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
2622            K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
2623            zone_page_state(zone, NR_KERNEL_STACK) *
2624                THREAD_SIZE / 1024,
2625            K(zone_page_state(zone, NR_PAGETABLE)),
2626            K(zone_page_state(zone, NR_UNSTABLE_NFS)),
2627            K(zone_page_state(zone, NR_BOUNCE)),
2628            K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
2629            zone->pages_scanned,
2630            (zone->all_unreclaimable ? "yes" : "no")
2631            );
2632        printk("lowmem_reserve[]:");
2633        for (i = 0; i < MAX_NR_ZONES; i++)
2634            printk(" %lu", zone->lowmem_reserve[i]);
2635        printk("\n");
2636    }
2637
2638    for_each_populated_zone(zone) {
2639         unsigned long nr[MAX_ORDER], flags, order, total = 0;
2640
2641        if (skip_free_areas_node(filter, zone_to_nid(zone)))
2642            continue;
2643        show_node(zone);
2644        printk("%s: ", zone->name);
2645
2646        spin_lock_irqsave(&zone->lock, flags);
2647        for (order = 0; order < MAX_ORDER; order++) {
2648            nr[order] = zone->free_area[order].nr_free;
2649            total += nr[order] << order;
2650        }
2651        spin_unlock_irqrestore(&zone->lock, flags);
2652        for (order = 0; order < MAX_ORDER; order++)
2653            printk("%lu*%lukB ", nr[order], K(1UL) << order);
2654        printk("= %lukB\n", K(total));
2655    }
2656
2657    printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
2658
2659    show_swap_cache_info();
2660}
2661
2662static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
2663{
2664    zoneref->zone = zone;
2665    zoneref->zone_idx = zone_idx(zone);
2666}
2667
2668/*
2669 * Builds allocation fallback zone lists.
2670 *
2671 * Add all populated zones of a node to the zonelist.
2672 */
2673static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
2674                int nr_zones, enum zone_type zone_type)
2675{
2676    struct zone *zone;
2677
2678    BUG_ON(zone_type >= MAX_NR_ZONES);
2679    zone_type++;
2680
2681    do {
2682        zone_type--;
2683        zone = pgdat->node_zones + zone_type;
2684        if (populated_zone(zone)) {
2685            zoneref_set_zone(zone,
2686                &zonelist->_zonerefs[nr_zones++]);
2687            check_highest_zone(zone_type);
2688        }
2689
2690    } while (zone_type);
2691    return nr_zones;
2692}
2693
2694
2695/*
2696 * zonelist_order:
2697 * 0 = automatic detection of better ordering.
2698 * 1 = order by ([node] distance, -zonetype)
2699 * 2 = order by (-zonetype, [node] distance)
2700 *
2701 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
2702 * the same zonelist. So only NUMA can configure this param.
2703 */
2704#define ZONELIST_ORDER_DEFAULT 0
2705#define ZONELIST_ORDER_NODE 1
2706#define ZONELIST_ORDER_ZONE 2
2707
2708/* zonelist order in the kernel.
2709 * set_zonelist_order() will set this to NODE or ZONE.
2710 */
2711static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
2712static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
2713
2714
2715#ifdef CONFIG_NUMA
2716/* The value user specified ....changed by config */
2717static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2718/* string for sysctl */
2719#define NUMA_ZONELIST_ORDER_LEN 16
2720char numa_zonelist_order[16] = "default";
2721
2722/*
2723 * interface for configure zonelist ordering.
2724 * command line option "numa_zonelist_order"
2725 * = "[dD]efault - default, automatic configuration.
2726 * = "[nN]ode - order by node locality, then by zone within node
2727 * = "[zZ]one - order by zone, then by locality within zone
2728 */
2729
2730static int __parse_numa_zonelist_order(char *s)
2731{
2732    if (*s == 'd' || *s == 'D') {
2733        user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2734    } else if (*s == 'n' || *s == 'N') {
2735        user_zonelist_order = ZONELIST_ORDER_NODE;
2736    } else if (*s == 'z' || *s == 'Z') {
2737        user_zonelist_order = ZONELIST_ORDER_ZONE;
2738    } else {
2739        printk(KERN_WARNING
2740            "Ignoring invalid numa_zonelist_order value: "
2741            "%s\n", s);
2742        return -EINVAL;
2743    }
2744    return 0;
2745}
2746
2747static __init int setup_numa_zonelist_order(char *s)
2748{
2749    int ret;
2750
2751    if (!s)
2752        return 0;
2753
2754    ret = __parse_numa_zonelist_order(s);
2755    if (ret == 0)
2756        strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
2757
2758    return ret;
2759}
2760early_param("numa_zonelist_order", setup_numa_zonelist_order);
2761
2762/*
2763 * sysctl handler for numa_zonelist_order
2764 */
2765int numa_zonelist_order_handler(ctl_table *table, int write,
2766        void __user *buffer, size_t *length,
2767        loff_t *ppos)
2768{
2769    char saved_string[NUMA_ZONELIST_ORDER_LEN];
2770    int ret;
2771    static DEFINE_MUTEX(zl_order_mutex);
2772
2773    mutex_lock(&zl_order_mutex);
2774    if (write)
2775        strcpy(saved_string, (char*)table->data);
2776    ret = proc_dostring(table, write, buffer, length, ppos);
2777    if (ret)
2778        goto out;
2779    if (write) {
2780        int oldval = user_zonelist_order;
2781        if (__parse_numa_zonelist_order((char*)table->data)) {
2782            /*
2783             * bogus value. restore saved string
2784             */
2785            strncpy((char*)table->data, saved_string,
2786                NUMA_ZONELIST_ORDER_LEN);
2787            user_zonelist_order = oldval;
2788        } else if (oldval != user_zonelist_order) {
2789            mutex_lock(&zonelists_mutex);
2790            build_all_zonelists(NULL);
2791            mutex_unlock(&zonelists_mutex);
2792        }
2793    }
2794out:
2795    mutex_unlock(&zl_order_mutex);
2796    return ret;
2797}
2798
2799
2800#define MAX_NODE_LOAD (nr_online_nodes)
2801static int node_load[MAX_NUMNODES];
2802
2803/**
2804 * find_next_best_node - find the next node that should appear in a given node's fallback list
2805 * @node: node whose fallback list we're appending
2806 * @used_node_mask: nodemask_t of already used nodes
2807 *
2808 * We use a number of factors to determine which is the next node that should
2809 * appear on a given node's fallback list. The node should not have appeared
2810 * already in @node's fallback list, and it should be the next closest node
2811 * according to the distance array (which contains arbitrary distance values
2812 * from each node to each node in the system), and should also prefer nodes
2813 * with no CPUs, since presumably they'll have very little allocation pressure
2814 * on them otherwise.
2815 * It returns -1 if no node is found.
2816 */
2817static int find_next_best_node(int node, nodemask_t *used_node_mask)
2818{
2819    int n, val;
2820    int min_val = INT_MAX;
2821    int best_node = -1;
2822    const struct cpumask *tmp = cpumask_of_node(0);
2823
2824    /* Use the local node if we haven't already */
2825    if (!node_isset(node, *used_node_mask)) {
2826        node_set(node, *used_node_mask);
2827        return node;
2828    }
2829
2830    for_each_node_state(n, N_HIGH_MEMORY) {
2831
2832        /* Don't want a node to appear more than once */
2833        if (node_isset(n, *used_node_mask))
2834            continue;
2835
2836        /* Use the distance array to find the distance */
2837        val = node_distance(node, n);
2838
2839        /* Penalize nodes under us ("prefer the next node") */
2840        val += (n < node);
2841
2842        /* Give preference to headless and unused nodes */
2843        tmp = cpumask_of_node(n);
2844        if (!cpumask_empty(tmp))
2845            val += PENALTY_FOR_NODE_WITH_CPUS;
2846
2847        /* Slight preference for less loaded node */
2848        val *= (MAX_NODE_LOAD*MAX_NUMNODES);
2849        val += node_load[n];
2850
2851        if (val < min_val) {
2852            min_val = val;
2853            best_node = n;
2854        }
2855    }
2856
2857    if (best_node >= 0)
2858        node_set(best_node, *used_node_mask);
2859
2860    return best_node;
2861}
2862
2863
2864/*
2865 * Build zonelists ordered by node and zones within node.
2866 * This results in maximum locality--normal zone overflows into local
2867 * DMA zone, if any--but risks exhausting DMA zone.
2868 */
2869static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
2870{
2871    int j;
2872    struct zonelist *zonelist;
2873
2874    zonelist = &pgdat->node_zonelists[0];
2875    for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
2876        ;
2877    j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2878                            MAX_NR_ZONES - 1);
2879    zonelist->_zonerefs[j].zone = NULL;
2880    zonelist->_zonerefs[j].zone_idx = 0;
2881}
2882
2883/*
2884 * Build gfp_thisnode zonelists
2885 */
2886static void build_thisnode_zonelists(pg_data_t *pgdat)
2887{
2888    int j;
2889    struct zonelist *zonelist;
2890
2891    zonelist = &pgdat->node_zonelists[1];
2892    j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2893    zonelist->_zonerefs[j].zone = NULL;
2894    zonelist->_zonerefs[j].zone_idx = 0;
2895}
2896
2897/*
2898 * Build zonelists ordered by zone and nodes within zones.
2899 * This results in conserving DMA zone[s] until all Normal memory is
2900 * exhausted, but results in overflowing to remote node while memory
2901 * may still exist in local DMA zone.
2902 */
2903static int node_order[MAX_NUMNODES];
2904
2905static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
2906{
2907    int pos, j, node;
2908    int zone_type; /* needs to be signed */
2909    struct zone *z;
2910    struct zonelist *zonelist;
2911
2912    zonelist = &pgdat->node_zonelists[0];
2913    pos = 0;
2914    for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
2915        for (j = 0; j < nr_nodes; j++) {
2916            node = node_order[j];
2917            z = &NODE_DATA(node)->node_zones[zone_type];
2918            if (populated_zone(z)) {
2919                zoneref_set_zone(z,
2920                    &zonelist->_zonerefs[pos++]);
2921                check_highest_zone(zone_type);
2922            }
2923        }
2924    }
2925    zonelist->_zonerefs[pos].zone = NULL;
2926    zonelist->_zonerefs[pos].zone_idx = 0;
2927}
2928
2929static int default_zonelist_order(void)
2930{
2931    int nid, zone_type;
2932    unsigned long low_kmem_size,total_size;
2933    struct zone *z;
2934    int average_size;
2935    /*
2936         * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
2937     * If they are really small and used heavily, the system can fall
2938     * into OOM very easily.
2939     * This function detect ZONE_DMA/DMA32 size and configures zone order.
2940     */
2941    /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
2942    low_kmem_size = 0;
2943    total_size = 0;
2944    for_each_online_node(nid) {
2945        for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2946            z = &NODE_DATA(nid)->node_zones[zone_type];
2947            if (populated_zone(z)) {
2948                if (zone_type < ZONE_NORMAL)
2949                    low_kmem_size += z->present_pages;
2950                total_size += z->present_pages;
2951            } else if (zone_type == ZONE_NORMAL) {
2952                /*
2953                 * If any node has only lowmem, then node order
2954                 * is preferred to allow kernel allocations
2955                 * locally; otherwise, they can easily infringe
2956                 * on other nodes when there is an abundance of
2957                 * lowmem available to allocate from.
2958                 */
2959                return ZONELIST_ORDER_NODE;
2960            }
2961        }
2962    }
2963    if (!low_kmem_size || /* there are no DMA area. */
2964        low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
2965        return ZONELIST_ORDER_NODE;
2966    /*
2967     * look into each node's config.
2968       * If there is a node whose DMA/DMA32 memory is very big area on
2969      * local memory, NODE_ORDER may be suitable.
2970         */
2971    average_size = total_size /
2972                (nodes_weight(node_states[N_HIGH_MEMORY]) + 1);
2973    for_each_online_node(nid) {
2974        low_kmem_size = 0;
2975        total_size = 0;
2976        for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2977            z = &NODE_DATA(nid)->node_zones[zone_type];
2978            if (populated_zone(z)) {
2979                if (zone_type < ZONE_NORMAL)
2980                    low_kmem_size += z->present_pages;
2981                total_size += z->present_pages;
2982            }
2983        }
2984        if (low_kmem_size &&
2985            total_size > average_size && /* ignore small node */
2986            low_kmem_size > total_size * 70/100)
2987            return ZONELIST_ORDER_NODE;
2988    }
2989    return ZONELIST_ORDER_ZONE;
2990}
2991
2992static void set_zonelist_order(void)
2993{
2994    if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
2995        current_zonelist_order = default_zonelist_order();
2996    else
2997        current_zonelist_order = user_zonelist_order;
2998}
2999
3000static void build_zonelists(pg_data_t *pgdat)
3001{
3002    int j, node, load;
3003    enum zone_type i;
3004    nodemask_t used_mask;
3005    int local_node, prev_node;
3006    struct zonelist *zonelist;
3007    int order = current_zonelist_order;
3008
3009    /* initialize zonelists */
3010    for (i = 0; i < MAX_ZONELISTS; i++) {
3011        zonelist = pgdat->node_zonelists + i;
3012        zonelist->_zonerefs[0].zone = NULL;
3013        zonelist->_zonerefs[0].zone_idx = 0;
3014    }
3015
3016    /* NUMA-aware ordering of nodes */
3017    local_node = pgdat->node_id;
3018    load = nr_online_nodes;
3019    prev_node = local_node;
3020    nodes_clear(used_mask);
3021
3022    memset(node_order, 0, sizeof(node_order));
3023    j = 0;
3024
3025    while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3026        int distance = node_distance(local_node, node);
3027
3028        /*
3029         * If another node is sufficiently far away then it is better
3030         * to reclaim pages in a zone before going off node.
3031         */
3032        if (distance > RECLAIM_DISTANCE)
3033            zone_reclaim_mode = 1;
3034
3035        /*
3036         * We don't want to pressure a particular node.
3037         * So adding penalty to the first node in same
3038         * distance group to make it round-robin.
3039         */
3040        if (distance != node_distance(local_node, prev_node))
3041            node_load[node] = load;
3042
3043        prev_node = node;
3044        load--;
3045        if (order == ZONELIST_ORDER_NODE)
3046            build_zonelists_in_node_order(pgdat, node);
3047        else
3048            node_order[j++] = node; /* remember order */
3049    }
3050
3051    if (order == ZONELIST_ORDER_ZONE) {
3052        /* calculate node order -- i.e., DMA last! */
3053        build_zonelists_in_zone_order(pgdat, j);
3054    }
3055
3056    build_thisnode_zonelists(pgdat);
3057}
3058
3059/* Construct the zonelist performance cache - see further mmzone.h */
3060static void build_zonelist_cache(pg_data_t *pgdat)
3061{
3062    struct zonelist *zonelist;
3063    struct zonelist_cache *zlc;
3064    struct zoneref *z;
3065
3066    zonelist = &pgdat->node_zonelists[0];
3067    zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3068    bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3069    for (z = zonelist->_zonerefs; z->zone; z++)
3070        zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3071}
3072
3073#ifdef CONFIG_HAVE_MEMORYLESS_NODES
3074/*
3075 * Return node id of node used for "local" allocations.
3076 * I.e., first node id of first zone in arg node's generic zonelist.
3077 * Used for initializing percpu 'numa_mem', which is used primarily
3078 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3079 */
3080int local_memory_node(int node)
3081{
3082    struct zone *zone;
3083
3084    (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3085                   gfp_zone(GFP_KERNEL),
3086                   NULL,
3087                   &zone);
3088    return zone->node;
3089}
3090#endif
3091
3092#else /* CONFIG_NUMA */
3093
3094static void set_zonelist_order(void)
3095{
3096    current_zonelist_order = ZONELIST_ORDER_ZONE;
3097}
3098
3099static void build_zonelists(pg_data_t *pgdat)
3100{
3101    int node, local_node;
3102    enum zone_type j;
3103    struct zonelist *zonelist;
3104
3105    local_node = pgdat->node_id;
3106
3107    zonelist = &pgdat->node_zonelists[0];
3108    j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3109
3110    /*
3111     * Now we build the zonelist so that it contains the zones
3112     * of all the other nodes.
3113     * We don't want to pressure a particular node, so when
3114     * building the zones for node N, we make sure that the
3115     * zones coming right after the local ones are those from
3116     * node N+1 (modulo N)
3117     */
3118    for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3119        if (!node_online(node))
3120            continue;
3121        j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3122                            MAX_NR_ZONES - 1);
3123    }
3124    for (node = 0; node < local_node; node++) {
3125        if (!node_online(node))
3126            continue;
3127        j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3128                            MAX_NR_ZONES - 1);
3129    }
3130
3131    zonelist->_zonerefs[j].zone = NULL;
3132    zonelist->_zonerefs[j].zone_idx = 0;
3133}
3134
3135/* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3136static void build_zonelist_cache(pg_data_t *pgdat)
3137{
3138    pgdat->node_zonelists[0].zlcache_ptr = NULL;
3139}
3140
3141#endif /* CONFIG_NUMA */
3142
3143/*
3144 * Boot pageset table. One per cpu which is going to be used for all
3145 * zones and all nodes. The parameters will be set in such a way
3146 * that an item put on a list will immediately be handed over to
3147 * the buddy list. This is safe since pageset manipulation is done
3148 * with interrupts disabled.
3149 *
3150 * The boot_pagesets must be kept even after bootup is complete for
3151 * unused processors and/or zones. They do play a role for bootstrapping
3152 * hotplugged processors.
3153 *
3154 * zoneinfo_show() and maybe other functions do
3155 * not check if the processor is online before following the pageset pointer.
3156 * Other parts of the kernel may not check if the zone is available.
3157 */
3158static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3159static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3160static void setup_zone_pageset(struct zone *zone);
3161
3162/*
3163 * Global mutex to protect against size modification of zonelists
3164 * as well as to serialize pageset setup for the new populated zone.
3165 */
3166DEFINE_MUTEX(zonelists_mutex);
3167
3168/* return values int ....just for stop_machine() */
3169static __init_refok int __build_all_zonelists(void *data)
3170{
3171    int nid;
3172    int cpu;
3173
3174#ifdef CONFIG_NUMA
3175    memset(node_load, 0, sizeof(node_load));
3176#endif
3177    for_each_online_node(nid) {
3178        pg_data_t *pgdat = NODE_DATA(nid);
3179
3180        build_zonelists(pgdat);
3181        build_zonelist_cache(pgdat);
3182    }
3183
3184    /*
3185     * Initialize the boot_pagesets that are going to be used
3186     * for bootstrapping processors. The real pagesets for
3187     * each zone will be allocated later when the per cpu
3188     * allocator is available.
3189     *
3190     * boot_pagesets are used also for bootstrapping offline
3191     * cpus if the system is already booted because the pagesets
3192     * are needed to initialize allocators on a specific cpu too.
3193     * F.e. the percpu allocator needs the page allocator which
3194     * needs the percpu allocator in order to allocate its pagesets
3195     * (a chicken-egg dilemma).
3196     */
3197    for_each_possible_cpu(cpu) {
3198        setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3199
3200#ifdef CONFIG_HAVE_MEMORYLESS_NODES
3201        /*
3202         * We now know the "local memory node" for each node--
3203         * i.e., the node of the first zone in the generic zonelist.
3204         * Set up numa_mem percpu variable for on-line cpus. During
3205         * boot, only the boot cpu should be on-line; we'll init the
3206         * secondary cpus' numa_mem as they come on-line. During
3207         * node/memory hotplug, we'll fixup all on-line cpus.
3208         */
3209        if (cpu_online(cpu))
3210            set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3211#endif
3212    }
3213
3214    return 0;
3215}
3216
3217/*
3218 * Called with zonelists_mutex held always
3219 * unless system_state == SYSTEM_BOOTING.
3220 */
3221void __ref build_all_zonelists(void *data)
3222{
3223    set_zonelist_order();
3224
3225    if (system_state == SYSTEM_BOOTING) {
3226        __build_all_zonelists(NULL);
3227        mminit_verify_zonelist();
3228        cpuset_init_current_mems_allowed();
3229    } else {
3230        /* we have to stop all cpus to guarantee there is no user
3231           of zonelist */
3232#ifdef CONFIG_MEMORY_HOTPLUG
3233        if (data)
3234            setup_zone_pageset((struct zone *)data);
3235#endif
3236        stop_machine(__build_all_zonelists, NULL, NULL);
3237        /* cpuset refresh routine should be here */
3238    }
3239    vm_total_pages = nr_free_pagecache_pages();
3240    /*
3241     * Disable grouping by mobility if the number of pages in the
3242     * system is too low to allow the mechanism to work. It would be
3243     * more accurate, but expensive to check per-zone. This check is
3244     * made on memory-hotadd so a system can start with mobility
3245     * disabled and enable it later
3246     */
3247    if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3248        page_group_by_mobility_disabled = 1;
3249    else
3250        page_group_by_mobility_disabled = 0;
3251
3252    printk("Built %i zonelists in %s order, mobility grouping %s. "
3253        "Total pages: %ld\n",
3254            nr_online_nodes,
3255            zonelist_order_name[current_zonelist_order],
3256            page_group_by_mobility_disabled ? "off" : "on",
3257            vm_total_pages);
3258#ifdef CONFIG_NUMA
3259    printk("Policy zone: %s\n", zone_names[policy_zone]);
3260#endif
3261}
3262
3263/*
3264 * Helper functions to size the waitqueue hash table.
3265 * Essentially these want to choose hash table sizes sufficiently
3266 * large so that collisions trying to wait on pages are rare.
3267 * But in fact, the number of active page waitqueues on typical
3268 * systems is ridiculously low, less than 200. So this is even
3269 * conservative, even though it seems large.
3270 *
3271 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3272 * waitqueues, i.e. the size of the waitq table given the number of pages.
3273 */
3274#define PAGES_PER_WAITQUEUE 256
3275
3276#ifndef CONFIG_MEMORY_HOTPLUG
3277static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3278{
3279    unsigned long size = 1;
3280
3281    pages /= PAGES_PER_WAITQUEUE;
3282
3283    while (size < pages)
3284        size <<= 1;
3285
3286    /*
3287     * Once we have dozens or even hundreds of threads sleeping
3288     * on IO we've got bigger problems than wait queue collision.
3289     * Limit the size of the wait table to a reasonable size.
3290     */
3291    size = min(size, 4096UL);
3292
3293    return max(size, 4UL);
3294}
3295#else
3296/*
3297 * A zone's size might be changed by hot-add, so it is not possible to determine
3298 * a suitable size for its wait_table. So we use the maximum size now.
3299 *
3300 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3301 *
3302 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3303 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3304 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3305 *
3306 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3307 * or more by the traditional way. (See above). It equals:
3308 *
3309 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3310 * ia64(16K page size) : = ( 8G + 4M)byte.
3311 * powerpc (64K page size) : = (32G +16M)byte.
3312 */
3313static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3314{
3315    return 4096UL;
3316}
3317#endif
3318
3319/*
3320 * This is an integer logarithm so that shifts can be used later
3321 * to extract the more random high bits from the multiplicative
3322 * hash function before the remainder is taken.
3323 */
3324static inline unsigned long wait_table_bits(unsigned long size)
3325{
3326    return ffz(~size);
3327}
3328
3329#define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3330
3331/*
3332 * Check if a pageblock contains reserved pages
3333 */
3334static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
3335{
3336    unsigned long pfn;
3337
3338    for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3339        if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
3340            return 1;
3341    }
3342    return 0;
3343}
3344
3345/*
3346 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3347 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3348 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3349 * higher will lead to a bigger reserve which will get freed as contiguous
3350 * blocks as reclaim kicks in
3351 */
3352static void setup_zone_migrate_reserve(struct zone *zone)
3353{
3354    unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
3355    struct page *page;
3356    unsigned long block_migratetype;
3357    int reserve;
3358
3359    /* Get the start pfn, end pfn and the number of blocks to reserve */
3360    start_pfn = zone->zone_start_pfn;
3361    end_pfn = start_pfn + zone->spanned_pages;
3362    reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3363                            pageblock_order;
3364
3365    /*
3366     * Reserve blocks are generally in place to help high-order atomic
3367     * allocations that are short-lived. A min_free_kbytes value that
3368     * would result in more than 2 reserve blocks for atomic allocations
3369     * is assumed to be in place to help anti-fragmentation for the
3370     * future allocation of hugepages at runtime.
3371     */
3372    reserve = min(2, reserve);
3373
3374    for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
3375        if (!pfn_valid(pfn))
3376            continue;
3377        page = pfn_to_page(pfn);
3378
3379        /* Watch out for overlapping nodes */
3380        if (page_to_nid(page) != zone_to_nid(zone))
3381            continue;
3382
3383        /* Blocks with reserved pages will never free, skip them. */
3384        block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
3385        if (pageblock_is_reserved(pfn, block_end_pfn))
3386            continue;
3387
3388        block_migratetype = get_pageblock_migratetype(page);
3389
3390        /* If this block is reserved, account for it */
3391        if (reserve > 0 && block_migratetype == MIGRATE_RESERVE) {
3392            reserve--;
3393            continue;
3394        }
3395
3396        /* Suitable for reserving if this block is movable */
3397        if (reserve > 0 && block_migratetype == MIGRATE_MOVABLE) {
3398            set_pageblock_migratetype(page, MIGRATE_RESERVE);
3399            move_freepages_block(zone, page, MIGRATE_RESERVE);
3400            reserve--;
3401            continue;
3402        }
3403
3404        /*
3405         * If the reserve is met and this is a previous reserved block,
3406         * take it back
3407         */
3408        if (block_migratetype == MIGRATE_RESERVE) {
3409            set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3410            move_freepages_block(zone, page, MIGRATE_MOVABLE);
3411        }
3412    }
3413}
3414
3415/*
3416 * Initially all pages are reserved - free ones are freed
3417 * up by free_all_bootmem() once the early boot process is
3418 * done. Non-atomic initialization, single-pass.
3419 */
3420void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
3421        unsigned long start_pfn, enum memmap_context context)
3422{
3423    struct page *page;
3424    unsigned long end_pfn = start_pfn + size;
3425    unsigned long pfn;
3426    struct zone *z;
3427
3428    if (highest_memmap_pfn < end_pfn - 1)
3429        highest_memmap_pfn = end_pfn - 1;
3430
3431    z = &NODE_DATA(nid)->node_zones[zone];
3432    for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3433        /*
3434         * There can be holes in boot-time mem_map[]s
3435         * handed to this function. They do not
3436         * exist on hotplugged memory.
3437         */
3438        if (context == MEMMAP_EARLY) {
3439            if (!early_pfn_valid(pfn))
3440                continue;
3441            if (!early_pfn_in_nid(pfn, nid))
3442                continue;
3443        }
3444        page = pfn_to_page(pfn);
3445        set_page_links(page, zone, nid, pfn);
3446        mminit_verify_page_links(page, zone, nid, pfn);
3447        init_page_count(page);
3448        reset_page_mapcount(page);
3449        SetPageReserved(page);
3450        /*
3451         * Mark the block movable so that blocks are reserved for
3452         * movable at startup. This will force kernel allocations
3453         * to reserve their blocks rather than leaking throughout
3454         * the address space during boot when many long-lived
3455         * kernel allocations are made. Later some blocks near
3456         * the start are marked MIGRATE_RESERVE by
3457         * setup_zone_migrate_reserve()
3458         *
3459         * bitmap is created for zone's valid pfn range. but memmap
3460         * can be created for invalid pages (for alignment)
3461         * check here not to call set_pageblock_migratetype() against
3462         * pfn out of zone.
3463         */
3464        if ((z->zone_start_pfn <= pfn)
3465            && (pfn < z->zone_start_pfn + z->spanned_pages)
3466            && !(pfn & (pageblock_nr_pages - 1)))
3467            set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3468
3469        INIT_LIST_HEAD(&page->lru);
3470#ifdef WANT_PAGE_VIRTUAL
3471        /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3472        if (!is_highmem_idx(zone))
3473            set_page_address(page, __va(pfn << PAGE_SHIFT));
3474#endif
3475    }
3476}
3477
3478static void __meminit zone_init_free_lists(struct zone *zone)
3479{
3480    int order, t;
3481    for_each_migratetype_order(order, t) {
3482        INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
3483        zone->free_area[order].nr_free = 0;
3484    }
3485}
3486
3487#ifndef __HAVE_ARCH_MEMMAP_INIT
3488#define memmap_init(size, nid, zone, start_pfn) \
3489    memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3490#endif
3491
3492static int zone_batchsize(struct zone *zone)
3493{
3494#ifdef CONFIG_MMU
3495    int batch;
3496
3497    /*
3498     * The per-cpu-pages pools are set to around 1000th of the
3499     * size of the zone. But no more than 1/2 of a meg.
3500     *
3501     * OK, so we don't know how big the cache is. So guess.
3502     */
3503    batch = zone->present_pages / 1024;
3504    if (batch * PAGE_SIZE > 512 * 1024)
3505        batch = (512 * 1024) / PAGE_SIZE;
3506    batch /= 4; /* We effectively *= 4 below */
3507    if (batch < 1)
3508        batch = 1;
3509
3510    /*
3511     * Clamp the batch to a 2^n - 1 value. Having a power
3512     * of 2 value was found to be more likely to have
3513     * suboptimal cache aliasing properties in some cases.
3514     *
3515     * For example if 2 tasks are alternately allocating
3516     * batches of pages, one task can end up with a lot
3517     * of pages of one half of the possible page colors
3518     * and the other with pages of the other colors.
3519     */
3520    batch = rounddown_pow_of_two(batch + batch/2) - 1;
3521
3522    return batch;
3523
3524#else
3525    /* The deferral and batching of frees should be suppressed under NOMMU
3526     * conditions.
3527     *
3528     * The problem is that NOMMU needs to be able to allocate large chunks
3529     * of contiguous memory as there's no hardware page translation to
3530     * assemble apparent contiguous memory from discontiguous pages.
3531     *
3532     * Queueing large contiguous runs of pages for batching, however,
3533     * causes the pages to actually be freed in smaller chunks. As there
3534     * can be a significant delay between the individual batches being
3535     * recycled, this leads to the once large chunks of space being
3536     * fragmented and becoming unavailable for high-order allocations.
3537     */
3538    return 0;
3539#endif
3540}
3541
3542static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
3543{
3544    struct per_cpu_pages *pcp;
3545    int migratetype;
3546
3547    memset(p, 0, sizeof(*p));
3548
3549    pcp = &p->pcp;
3550    pcp->count = 0;
3551    pcp->high = 6 * batch;
3552    pcp->batch = max(1UL, 1 * batch);
3553    for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
3554        INIT_LIST_HEAD(&pcp->lists[migratetype]);
3555}
3556
3557/*
3558 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3559 * to the value high for the pageset p.
3560 */
3561
3562static void setup_pagelist_highmark(struct per_cpu_pageset *p,
3563                unsigned long high)
3564{
3565    struct per_cpu_pages *pcp;
3566
3567    pcp = &p->pcp;
3568    pcp->high = high;
3569    pcp->batch = max(1UL, high/4);
3570    if ((high/4) > (PAGE_SHIFT * 8))
3571        pcp->batch = PAGE_SHIFT * 8;
3572}
3573
3574static void setup_zone_pageset(struct zone *zone)
3575{
3576    int cpu;
3577
3578    zone->pageset = alloc_percpu(struct per_cpu_pageset);
3579
3580    for_each_possible_cpu(cpu) {
3581        struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
3582
3583        setup_pageset(pcp, zone_batchsize(zone));
3584
3585        if (percpu_pagelist_fraction)
3586            setup_pagelist_highmark(pcp,
3587                (zone->present_pages /
3588                    percpu_pagelist_fraction));
3589    }
3590}
3591
3592/*
3593 * Allocate per cpu pagesets and initialize them.
3594 * Before this call only boot pagesets were available.
3595 */
3596void __init setup_per_cpu_pageset(void)
3597{
3598    struct zone *zone;
3599
3600    for_each_populated_zone(zone)
3601        setup_zone_pageset(zone);
3602}
3603
3604static noinline __init_refok
3605int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
3606{
3607    int i;
3608    struct pglist_data *pgdat = zone->zone_pgdat;
3609    size_t alloc_size;
3610
3611    /*
3612     * The per-page waitqueue mechanism uses hashed waitqueues
3613     * per zone.
3614     */
3615    zone->wait_table_hash_nr_entries =
3616         wait_table_hash_nr_entries(zone_size_pages);
3617    zone->wait_table_bits =
3618        wait_table_bits(zone->wait_table_hash_nr_entries);
3619    alloc_size = zone->wait_table_hash_nr_entries
3620                    * sizeof(wait_queue_head_t);
3621
3622    if (!slab_is_available()) {
3623        zone->wait_table = (wait_queue_head_t *)
3624            alloc_bootmem_node_nopanic(pgdat, alloc_size);
3625    } else {
3626        /*
3627         * This case means that a zone whose size was 0 gets new memory
3628         * via memory hot-add.
3629         * But it may be the case that a new node was hot-added. In
3630         * this case vmalloc() will not be able to use this new node's
3631         * memory - this wait_table must be initialized to use this new
3632         * node itself as well.
3633         * To use this new node's memory, further consideration will be
3634         * necessary.
3635         */
3636        zone->wait_table = vmalloc(alloc_size);
3637    }
3638    if (!zone->wait_table)
3639        return -ENOMEM;
3640
3641    for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
3642        init_waitqueue_head(zone->wait_table + i);
3643
3644    return 0;
3645}
3646
3647static int __zone_pcp_update(void *data)
3648{
3649    struct zone *zone = data;
3650    int cpu;
3651    unsigned long batch = zone_batchsize(zone), flags;
3652
3653    for_each_possible_cpu(cpu) {
3654        struct per_cpu_pageset *pset;
3655        struct per_cpu_pages *pcp;
3656
3657        pset = per_cpu_ptr(zone->pageset, cpu);
3658        pcp = &pset->pcp;
3659
3660        local_irq_save(flags);
3661        free_pcppages_bulk(zone, pcp->count, pcp);
3662        setup_pageset(pset, batch);
3663        local_irq_restore(flags);
3664    }
3665    return 0;
3666}
3667
3668void zone_pcp_update(struct zone *zone)
3669{
3670    stop_machine(__zone_pcp_update, zone, NULL);
3671}
3672
3673static __meminit void zone_pcp_init(struct zone *zone)
3674{
3675    /*
3676     * per cpu subsystem is not up at this point. The following code
3677     * relies on the ability of the linker to provide the
3678     * offset of a (static) per cpu variable into the per cpu area.
3679     */
3680    zone->pageset = &boot_pageset;
3681
3682    if (zone->present_pages)
3683        printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
3684            zone->name, zone->present_pages,
3685                     zone_batchsize(zone));
3686}
3687
3688__meminit int init_currently_empty_zone(struct zone *zone,
3689                    unsigned long zone_start_pfn,
3690                    unsigned long size,
3691                    enum memmap_context context)
3692{
3693    struct pglist_data *pgdat = zone->zone_pgdat;
3694    int ret;
3695    ret = zone_wait_table_init(zone, size);
3696    if (ret)
3697        return ret;
3698    pgdat->nr_zones = zone_idx(zone) + 1;
3699
3700    zone->zone_start_pfn = zone_start_pfn;
3701
3702    mminit_dprintk(MMINIT_TRACE, "memmap_init",
3703            "Initialising map node %d zone %lu pfns %lu -> %lu\n",
3704            pgdat->node_id,
3705            (unsigned long)zone_idx(zone),
3706            zone_start_pfn, (zone_start_pfn + size));
3707
3708    zone_init_free_lists(zone);
3709
3710    return 0;
3711}
3712
3713#ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3714/*
3715 * Basic iterator support. Return the first range of PFNs for a node
3716 * Note: nid == MAX_NUMNODES returns first region regardless of node
3717 */
3718static int __meminit first_active_region_index_in_nid(int nid)
3719{
3720    int i;
3721
3722    for (i = 0; i < nr_nodemap_entries; i++)
3723        if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
3724            return i;
3725
3726    return -1;
3727}
3728
3729/*
3730 * Basic iterator support. Return the next active range of PFNs for a node
3731 * Note: nid == MAX_NUMNODES returns next region regardless of node
3732 */
3733static int __meminit next_active_region_index_in_nid(int index, int nid)
3734{
3735    for (index = index + 1; index < nr_nodemap_entries; index++)
3736        if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
3737            return index;
3738
3739    return -1;
3740}
3741
3742#ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
3743/*
3744 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
3745 * Architectures may implement their own version but if add_active_range()
3746 * was used and there are no special requirements, this is a convenient
3747 * alternative
3748 */
3749int __meminit __early_pfn_to_nid(unsigned long pfn)
3750{
3751    int i;
3752
3753    for (i = 0; i < nr_nodemap_entries; i++) {
3754        unsigned long start_pfn = early_node_map[i].start_pfn;
3755        unsigned long end_pfn = early_node_map[i].end_pfn;
3756
3757        if (start_pfn <= pfn && pfn < end_pfn)
3758            return early_node_map[i].nid;
3759    }
3760    /* This is a memory hole */
3761    return -1;
3762}
3763#endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
3764
3765int __meminit early_pfn_to_nid(unsigned long pfn)
3766{
3767    int nid;
3768
3769    nid = __early_pfn_to_nid(pfn);
3770    if (nid >= 0)
3771        return nid;
3772    /* just returns 0 */
3773    return 0;
3774}
3775
3776#ifdef CONFIG_NODES_SPAN_OTHER_NODES
3777bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
3778{
3779    int nid;
3780
3781    nid = __early_pfn_to_nid(pfn);
3782    if (nid >= 0 && nid != node)
3783        return false;
3784    return true;
3785}
3786#endif
3787
3788/* Basic iterator support to walk early_node_map[] */
3789#define for_each_active_range_index_in_nid(i, nid) \
3790    for (i = first_active_region_index_in_nid(nid); i != -1; \
3791                i = next_active_region_index_in_nid(i, nid))
3792
3793/**
3794 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
3795 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
3796 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
3797 *
3798 * If an architecture guarantees that all ranges registered with
3799 * add_active_ranges() contain no holes and may be freed, this
3800 * this function may be used instead of calling free_bootmem() manually.
3801 */
3802void __init free_bootmem_with_active_regions(int nid,
3803                        unsigned long max_low_pfn)
3804{
3805    int i;
3806
3807    for_each_active_range_index_in_nid(i, nid) {
3808        unsigned long size_pages = 0;
3809        unsigned long end_pfn = early_node_map[i].end_pfn;
3810
3811        if (early_node_map[i].start_pfn >= max_low_pfn)
3812            continue;
3813
3814        if (end_pfn > max_low_pfn)
3815            end_pfn = max_low_pfn;
3816
3817        size_pages = end_pfn - early_node_map[i].start_pfn;
3818        free_bootmem_node(NODE_DATA(early_node_map[i].nid),
3819                PFN_PHYS(early_node_map[i].start_pfn),
3820                size_pages << PAGE_SHIFT);
3821    }
3822}
3823
3824#ifdef CONFIG_HAVE_MEMBLOCK
3825/*
3826 * Basic iterator support. Return the last range of PFNs for a node
3827 * Note: nid == MAX_NUMNODES returns last region regardless of node
3828 */
3829static int __meminit last_active_region_index_in_nid(int nid)
3830{
3831    int i;
3832
3833    for (i = nr_nodemap_entries - 1; i >= 0; i--)
3834        if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
3835            return i;
3836
3837    return -1;
3838}
3839
3840/*
3841 * Basic iterator support. Return the previous active range of PFNs for a node
3842 * Note: nid == MAX_NUMNODES returns next region regardless of node
3843 */
3844static int __meminit previous_active_region_index_in_nid(int index, int nid)
3845{
3846    for (index = index - 1; index >= 0; index--)
3847        if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
3848            return index;
3849
3850    return -1;
3851}
3852
3853#define for_each_active_range_index_in_nid_reverse(i, nid) \
3854    for (i = last_active_region_index_in_nid(nid); i != -1; \
3855                i = previous_active_region_index_in_nid(i, nid))
3856
3857u64 __init find_memory_core_early(int nid, u64 size, u64 align,
3858                    u64 goal, u64 limit)
3859{
3860    int i;
3861
3862    /* Need to go over early_node_map to find out good range for node */
3863    for_each_active_range_index_in_nid_reverse(i, nid) {
3864        u64 addr;
3865        u64 ei_start, ei_last;
3866        u64 final_start, final_end;
3867
3868        ei_last = early_node_map[i].end_pfn;
3869        ei_last <<= PAGE_SHIFT;
3870        ei_start = early_node_map[i].start_pfn;
3871        ei_start <<= PAGE_SHIFT;
3872
3873        final_start = max(ei_start, goal);
3874        final_end = min(ei_last, limit);
3875
3876        if (final_start >= final_end)
3877            continue;
3878
3879        addr = memblock_find_in_range(final_start, final_end, size, align);
3880
3881        if (addr == MEMBLOCK_ERROR)
3882            continue;
3883
3884        return addr;
3885    }
3886
3887    return MEMBLOCK_ERROR;
3888}
3889#endif
3890
3891int __init add_from_early_node_map(struct range *range, int az,
3892                   int nr_range, int nid)
3893{
3894    int i;
3895    u64 start, end;
3896
3897    /* need to go over early_node_map to find out good range for node */
3898    for_each_active_range_index_in_nid(i, nid) {
3899        start = early_node_map[i].start_pfn;
3900        end = early_node_map[i].end_pfn;
3901        nr_range = add_range(range, az, nr_range, start, end);
3902    }
3903    return nr_range;
3904}
3905
3906void __init work_with_active_regions(int nid, work_fn_t work_fn, void *data)
3907{
3908    int i;
3909    int ret;
3910
3911    for_each_active_range_index_in_nid(i, nid) {
3912        ret = work_fn(early_node_map[i].start_pfn,
3913                  early_node_map[i].end_pfn, data);
3914        if (ret)
3915            break;
3916    }
3917}
3918/**
3919 * sparse_memory_present_with_active_regions - Call memory_present for each active range
3920 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
3921 *
3922 * If an architecture guarantees that all ranges registered with
3923 * add_active_ranges() contain no holes and may be freed, this
3924 * function may be used instead of calling memory_present() manually.
3925 */
3926void __init sparse_memory_present_with_active_regions(int nid)
3927{
3928    int i;
3929
3930    for_each_active_range_index_in_nid(i, nid)
3931        memory_present(early_node_map[i].nid,
3932                early_node_map[i].start_pfn,
3933                early_node_map[i].end_pfn);
3934}
3935
3936/**
3937 * get_pfn_range_for_nid - Return the start and end page frames for a node
3938 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
3939 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
3940 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
3941 *
3942 * It returns the start and end page frame of a node based on information
3943 * provided by an arch calling add_active_range(). If called for a node
3944 * with no available memory, a warning is printed and the start and end
3945 * PFNs will be 0.
3946 */
3947void __meminit get_pfn_range_for_nid(unsigned int nid,
3948            unsigned long *start_pfn, unsigned long *end_pfn)
3949{
3950    int i;
3951    *start_pfn = -1UL;
3952    *end_pfn = 0;
3953
3954    for_each_active_range_index_in_nid(i, nid) {
3955        *start_pfn = min(*start_pfn, early_node_map[i].start_pfn);
3956        *end_pfn = max(*end_pfn, early_node_map[i].end_pfn);
3957    }
3958
3959    if (*start_pfn == -1UL)
3960        *start_pfn = 0;
3961}
3962
3963/*
3964 * This finds a zone that can be used for ZONE_MOVABLE pages. The
3965 * assumption is made that zones within a node are ordered in monotonic
3966 * increasing memory addresses so that the "highest" populated zone is used
3967 */
3968static void __init find_usable_zone_for_movable(void)
3969{
3970    int zone_index;
3971    for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
3972        if (zone_index == ZONE_MOVABLE)
3973            continue;
3974
3975        if (arch_zone_highest_possible_pfn[zone_index] >
3976                arch_zone_lowest_possible_pfn[zone_index])
3977            break;
3978    }
3979
3980    VM_BUG_ON(zone_index == -1);
3981    movable_zone = zone_index;
3982}
3983
3984/*
3985 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
3986 * because it is sized independent of architecture. Unlike the other zones,
3987 * the starting point for ZONE_MOVABLE is not fixed. It may be different
3988 * in each node depending on the size of each node and how evenly kernelcore
3989 * is distributed. This helper function adjusts the zone ranges
3990 * provided by the architecture for a given node by using the end of the
3991 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
3992 * zones within a node are in order of monotonic increases memory addresses
3993 */
3994static void __meminit adjust_zone_range_for_zone_movable(int nid,
3995                    unsigned long zone_type,
3996                    unsigned long node_start_pfn,
3997                    unsigned long node_end_pfn,
3998                    unsigned long *zone_start_pfn,
3999                    unsigned long *zone_end_pfn)
4000{
4001    /* Only adjust if ZONE_MOVABLE is on this node */
4002    if (zone_movable_pfn[nid]) {
4003        /* Size ZONE_MOVABLE */
4004        if (zone_type == ZONE_MOVABLE) {
4005            *zone_start_pfn = zone_movable_pfn[nid];
4006            *zone_end_pfn = min(node_end_pfn,
4007                arch_zone_highest_possible_pfn[movable_zone]);
4008
4009        /* Adjust for ZONE_MOVABLE starting within this range */
4010        } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4011                *zone_end_pfn > zone_movable_pfn[nid]) {
4012            *zone_end_pfn = zone_movable_pfn[nid];
4013
4014        /* Check if this whole range is within ZONE_MOVABLE */
4015        } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4016            *zone_start_pfn = *zone_end_pfn;
4017    }
4018}
4019
4020/*
4021 * Return the number of pages a zone spans in a node, including holes
4022 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4023 */
4024static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4025                    unsigned long zone_type,
4026                    unsigned long *ignored)
4027{
4028    unsigned long node_start_pfn, node_end_pfn;
4029    unsigned long zone_start_pfn, zone_end_pfn;
4030
4031    /* Get the start and end of the node and zone */
4032    get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4033    zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4034    zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4035    adjust_zone_range_for_zone_movable(nid, zone_type,
4036                node_start_pfn, node_end_pfn,
4037                &zone_start_pfn, &zone_end_pfn);
4038
4039    /* Check that this node has pages within the zone's required range */
4040    if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4041        return 0;
4042
4043    /* Move the zone boundaries inside the node if necessary */
4044    zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4045    zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4046
4047    /* Return the spanned pages */
4048    return zone_end_pfn - zone_start_pfn;
4049}
4050
4051/*
4052 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4053 * then all holes in the requested range will be accounted for.
4054 */
4055unsigned long __meminit __absent_pages_in_range(int nid,
4056                unsigned long range_start_pfn,
4057                unsigned long range_end_pfn)
4058{
4059    int i = 0;
4060    unsigned long prev_end_pfn = 0, hole_pages = 0;
4061    unsigned long start_pfn;
4062
4063    /* Find the end_pfn of the first active range of pfns in the node */
4064    i = first_active_region_index_in_nid(nid);
4065    if (i == -1)
4066        return 0;
4067
4068    prev_end_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
4069
4070    /* Account for ranges before physical memory on this node */
4071    if (early_node_map[i].start_pfn > range_start_pfn)
4072        hole_pages = prev_end_pfn - range_start_pfn;
4073
4074    /* Find all holes for the zone within the node */
4075    for (; i != -1; i = next_active_region_index_in_nid(i, nid)) {
4076
4077        /* No need to continue if prev_end_pfn is outside the zone */
4078        if (prev_end_pfn >= range_end_pfn)
4079            break;
4080
4081        /* Make sure the end of the zone is not within the hole */
4082        start_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
4083        prev_end_pfn = max(prev_end_pfn, range_start_pfn);
4084
4085        /* Update the hole size cound and move on */
4086        if (start_pfn > range_start_pfn) {
4087            BUG_ON(prev_end_pfn > start_pfn);
4088            hole_pages += start_pfn - prev_end_pfn;
4089        }
4090        prev_end_pfn = early_node_map[i].end_pfn;
4091    }
4092
4093    /* Account for ranges past physical memory on this node */
4094    if (range_end_pfn > prev_end_pfn)
4095        hole_pages += range_end_pfn -
4096                max(range_start_pfn, prev_end_pfn);
4097
4098    return hole_pages;
4099}
4100
4101/**
4102 * absent_pages_in_range - Return number of page frames in holes within a range
4103 * @start_pfn: The start PFN to start searching for holes
4104 * @end_pfn: The end PFN to stop searching for holes
4105 *
4106 * It returns the number of pages frames in memory holes within a range.
4107 */
4108unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4109                            unsigned long end_pfn)
4110{
4111    return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4112}
4113
4114/* Return the number of page frames in holes in a zone on a node */
4115static unsigned long __meminit zone_absent_pages_in_node(int nid,
4116                    unsigned long zone_type,
4117                    unsigned long *ignored)
4118{
4119    unsigned long node_start_pfn, node_end_pfn;
4120    unsigned long zone_start_pfn, zone_end_pfn;
4121
4122    get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4123    zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type],
4124                            node_start_pfn);
4125    zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type],
4126                            node_end_pfn);
4127
4128    adjust_zone_range_for_zone_movable(nid, zone_type,
4129            node_start_pfn, node_end_pfn,
4130            &zone_start_pfn, &zone_end_pfn);
4131    return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4132}
4133
4134#else
4135static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4136                    unsigned long zone_type,
4137                    unsigned long *zones_size)
4138{
4139    return zones_size[zone_type];
4140}
4141
4142static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4143                        unsigned long zone_type,
4144                        unsigned long *zholes_size)
4145{
4146    if (!zholes_size)
4147        return 0;
4148
4149    return zholes_size[zone_type];
4150}
4151
4152#endif
4153
4154static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4155        unsigned long *zones_size, unsigned long *zholes_size)
4156{
4157    unsigned long realtotalpages, totalpages = 0;
4158    enum zone_type i;
4159
4160    for (i = 0; i < MAX_NR_ZONES; i++)
4161        totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4162                                zones_size);
4163    pgdat->node_spanned_pages = totalpages;
4164
4165    realtotalpages = totalpages;
4166    for (i = 0; i < MAX_NR_ZONES; i++)
4167        realtotalpages -=
4168            zone_absent_pages_in_node(pgdat->node_id, i,
4169                                zholes_size);
4170    pgdat->node_present_pages = realtotalpages;
4171    printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4172                            realtotalpages);
4173}
4174
4175#ifndef CONFIG_SPARSEMEM
4176/*
4177 * Calculate the size of the zone->blockflags rounded to an unsigned long
4178 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4179 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4180 * round what is now in bits to nearest long in bits, then return it in
4181 * bytes.
4182 */
4183static unsigned long __init usemap_size(unsigned long zonesize)
4184{
4185    unsigned long usemapsize;
4186
4187    usemapsize = roundup(zonesize, pageblock_nr_pages);
4188    usemapsize = usemapsize >> pageblock_order;
4189    usemapsize *= NR_PAGEBLOCK_BITS;
4190    usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4191
4192    return usemapsize / 8;
4193}
4194
4195static void __init setup_usemap(struct pglist_data *pgdat,
4196                struct zone *zone, unsigned long zonesize)
4197{
4198    unsigned long usemapsize = usemap_size(zonesize);
4199    zone->pageblock_flags = NULL;
4200    if (usemapsize)
4201        zone->pageblock_flags = alloc_bootmem_node_nopanic(pgdat,
4202                                   usemapsize);
4203}
4204#else
4205static inline void setup_usemap(struct pglist_data *pgdat,
4206                struct zone *zone, unsigned long zonesize) {}
4207#endif /* CONFIG_SPARSEMEM */
4208
4209#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4210
4211/* Return a sensible default order for the pageblock size. */
4212static inline int pageblock_default_order(void)
4213{
4214    if (HPAGE_SHIFT > PAGE_SHIFT)
4215        return HUGETLB_PAGE_ORDER;
4216
4217    return MAX_ORDER-1;
4218}
4219
4220/* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4221static inline void __init set_pageblock_order(unsigned int order)
4222{
4223    /* Check that pageblock_nr_pages has not already been setup */
4224    if (pageblock_order)
4225        return;
4226
4227    /*
4228     * Assume the largest contiguous order of interest is a huge page.
4229     * This value may be variable depending on boot parameters on IA64
4230     */
4231    pageblock_order = order;
4232}
4233#else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4234
4235/*
4236 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4237 * and pageblock_default_order() are unused as pageblock_order is set
4238 * at compile-time. See include/linux/pageblock-flags.h for the values of
4239 * pageblock_order based on the kernel config
4240 */
4241static inline int pageblock_default_order(unsigned int order)
4242{
4243    return MAX_ORDER-1;
4244}
4245#define set_pageblock_order(x) do {} while (0)
4246
4247#endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4248
4249/*
4250 * Set up the zone data structures:
4251 * - mark all pages reserved
4252 * - mark all memory queues empty
4253 * - clear the memory bitmaps
4254 */
4255static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4256        unsigned long *zones_size, unsigned long *zholes_size)
4257{
4258    enum zone_type j;
4259    int nid = pgdat->node_id;
4260    unsigned long zone_start_pfn = pgdat->node_start_pfn;
4261    int ret;
4262
4263    pgdat_resize_init(pgdat);
4264    pgdat->nr_zones = 0;
4265    init_waitqueue_head(&pgdat->kswapd_wait);
4266    pgdat->kswapd_max_order = 0;
4267    pgdat_page_cgroup_init(pgdat);
4268    
4269    for (j = 0; j < MAX_NR_ZONES; j++) {
4270        struct zone *zone = pgdat->node_zones + j;
4271        unsigned long size, realsize, memmap_pages;
4272        enum lru_list l;
4273
4274        size = zone_spanned_pages_in_node(nid, j, zones_size);
4275        realsize = size - zone_absent_pages_in_node(nid, j,
4276                                zholes_size);
4277
4278        /*
4279         * Adjust realsize so that it accounts for how much memory
4280         * is used by this zone for memmap. This affects the watermark
4281         * and per-cpu initialisations
4282         */
4283        memmap_pages =
4284            PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT;
4285        if (realsize >= memmap_pages) {
4286            realsize -= memmap_pages;
4287            if (memmap_pages)
4288                printk(KERN_DEBUG
4289                       " %s zone: %lu pages used for memmap\n",
4290                       zone_names[j], memmap_pages);
4291        } else
4292            printk(KERN_WARNING
4293                " %s zone: %lu pages exceeds realsize %lu\n",
4294                zone_names[j], memmap_pages, realsize);
4295
4296        /* Account for reserved pages */
4297        if (j == 0 && realsize > dma_reserve) {
4298            realsize -= dma_reserve;
4299            printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4300                    zone_names[0], dma_reserve);
4301        }
4302
4303        if (!is_highmem_idx(j))
4304            nr_kernel_pages += realsize;
4305        nr_all_pages += realsize;
4306
4307        zone->spanned_pages = size;
4308        zone->present_pages = realsize;
4309#ifdef CONFIG_NUMA
4310        zone->node = nid;
4311        zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
4312                        / 100;
4313        zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
4314#endif
4315        zone->name = zone_names[j];
4316        spin_lock_init(&zone->lock);
4317        spin_lock_init(&zone->lru_lock);
4318        zone_seqlock_init(zone);
4319        zone->zone_pgdat = pgdat;
4320
4321        zone_pcp_init(zone);
4322        for_each_lru(l)
4323            INIT_LIST_HEAD(&zone->lru[l].list);
4324        zone->reclaim_stat.recent_rotated[0] = 0;
4325        zone->reclaim_stat.recent_rotated[1] = 0;
4326        zone->reclaim_stat.recent_scanned[0] = 0;
4327        zone->reclaim_stat.recent_scanned[1] = 0;
4328        zap_zone_vm_stats(zone);
4329        zone->flags = 0;
4330        if (!size)
4331            continue;
4332
4333        set_pageblock_order(pageblock_default_order());
4334        setup_usemap(pgdat, zone, size);
4335        ret = init_currently_empty_zone(zone, zone_start_pfn,
4336                        size, MEMMAP_EARLY);
4337        BUG_ON(ret);
4338        memmap_init(size, nid, j, zone_start_pfn);
4339        zone_start_pfn += size;
4340    }
4341}
4342
4343static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4344{
4345    /* Skip empty nodes */
4346    if (!pgdat->node_spanned_pages)
4347        return;
4348
4349#ifdef CONFIG_FLAT_NODE_MEM_MAP
4350    /* ia64 gets its own node_mem_map, before this, without bootmem */
4351    if (!pgdat->node_mem_map) {
4352        unsigned long size, start, end;
4353        struct page *map;
4354
4355        /*
4356         * The zone's endpoints aren't required to be MAX_ORDER
4357         * aligned but the node_mem_map endpoints must be in order
4358         * for the buddy allocator to function correctly.
4359         */
4360        start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4361        end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
4362        end = ALIGN(end, MAX_ORDER_NR_PAGES);
4363        size = (end - start) * sizeof(struct page);
4364        map = alloc_remap(pgdat->node_id, size);
4365        if (!map)
4366            map = alloc_bootmem_node_nopanic(pgdat, size);
4367        pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4368    }
4369#ifndef CONFIG_NEED_MULTIPLE_NODES
4370    /*
4371     * With no DISCONTIG, the global mem_map is just set as node 0's
4372     */
4373    if (pgdat == NODE_DATA(0)) {
4374        mem_map = NODE_DATA(0)->node_mem_map;
4375#ifdef CONFIG_ARCH_POPULATES_NODE_MAP
4376        if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4377            mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4378#endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4379    }
4380#endif
4381#endif /* CONFIG_FLAT_NODE_MEM_MAP */
4382}
4383
4384void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4385        unsigned long node_start_pfn, unsigned long *zholes_size)
4386{
4387    pg_data_t *pgdat = NODE_DATA(nid);
4388
4389    pgdat->node_id = nid;
4390    pgdat->node_start_pfn = node_start_pfn;
4391    calculate_node_totalpages(pgdat, zones_size, zholes_size);
4392
4393    alloc_node_mem_map(pgdat);
4394#ifdef CONFIG_FLAT_NODE_MEM_MAP
4395    printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4396        nid, (unsigned long)pgdat,
4397        (unsigned long)pgdat->node_mem_map);
4398#endif
4399
4400    free_area_init_core(pgdat, zones_size, zholes_size);
4401}
4402
4403#ifdef CONFIG_ARCH_POPULATES_NODE_MAP
4404
4405#if MAX_NUMNODES > 1
4406/*
4407 * Figure out the number of possible node ids.
4408 */
4409static void __init setup_nr_node_ids(void)
4410{
4411    unsigned int node;
4412    unsigned int highest = 0;
4413
4414    for_each_node_mask(node, node_possible_map)
4415        highest = node;
4416    nr_node_ids = highest + 1;
4417}
4418#else
4419static inline void setup_nr_node_ids(void)
4420{
4421}
4422#endif
4423
4424/**
4425 * add_active_range - Register a range of PFNs backed by physical memory
4426 * @nid: The node ID the range resides on
4427 * @start_pfn: The start PFN of the available physical memory
4428 * @end_pfn: The end PFN of the available physical memory
4429 *
4430 * These ranges are stored in an early_node_map[] and later used by
4431 * free_area_init_nodes() to calculate zone sizes and holes. If the
4432 * range spans a memory hole, it is up to the architecture to ensure
4433 * the memory is not freed by the bootmem allocator. If possible
4434 * the range being registered will be merged with existing ranges.
4435 */
4436void __init add_active_range(unsigned int nid, unsigned long start_pfn,
4437                        unsigned long end_pfn)
4438{
4439    int i;
4440
4441    mminit_dprintk(MMINIT_TRACE, "memory_register",
4442            "Entering add_active_range(%d, %#lx, %#lx) "
4443            "%d entries of %d used\n",
4444            nid, start_pfn, end_pfn,
4445            nr_nodemap_entries, MAX_ACTIVE_REGIONS);
4446
4447    mminit_validate_memmodel_limits(&start_pfn, &end_pfn);
4448
4449    /* Merge with existing active regions if possible */
4450    for (i = 0; i < nr_nodemap_entries; i++) {
4451        if (early_node_map[i].nid != nid)
4452            continue;
4453
4454        /* Skip if an existing region covers this new one */
4455        if (start_pfn >= early_node_map[i].start_pfn &&
4456                end_pfn <= early_node_map[i].end_pfn)
4457            return;
4458
4459        /* Merge forward if suitable */
4460        if (start_pfn <= early_node_map[i].end_pfn &&
4461                end_pfn > early_node_map[i].end_pfn) {
4462            early_node_map[i].end_pfn = end_pfn;
4463            return;
4464        }
4465
4466        /* Merge backward if suitable */
4467        if (start_pfn < early_node_map[i].start_pfn &&
4468                end_pfn >= early_node_map[i].start_pfn) {
4469            early_node_map[i].start_pfn = start_pfn;
4470            return;
4471        }
4472    }
4473
4474    /* Check that early_node_map is large enough */
4475    if (i >= MAX_ACTIVE_REGIONS) {
4476        printk(KERN_CRIT "More than %d memory regions, truncating\n",
4477                            MAX_ACTIVE_REGIONS);
4478        return;
4479    }
4480
4481    early_node_map[i].nid = nid;
4482    early_node_map[i].start_pfn = start_pfn;
4483    early_node_map[i].end_pfn = end_pfn;
4484    nr_nodemap_entries = i + 1;
4485}
4486
4487/**
4488 * remove_active_range - Shrink an existing registered range of PFNs
4489 * @nid: The node id the range is on that should be shrunk
4490 * @start_pfn: The new PFN of the range
4491 * @end_pfn: The new PFN of the range
4492 *
4493 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
4494 * The map is kept near the end physical page range that has already been
4495 * registered. This function allows an arch to shrink an existing registered
4496 * range.
4497 */
4498void __init remove_active_range(unsigned int nid, unsigned long start_pfn,
4499                unsigned long end_pfn)
4500{
4501    int i, j;
4502    int removed = 0;
4503
4504    printk(KERN_DEBUG "remove_active_range (%d, %lu, %lu)\n",
4505              nid, start_pfn, end_pfn);
4506
4507    /* Find the old active region end and shrink */
4508    for_each_active_range_index_in_nid(i, nid) {
4509        if (early_node_map[i].start_pfn >= start_pfn &&
4510            early_node_map[i].end_pfn <= end_pfn) {
4511            /* clear it */
4512            early_node_map[i].start_pfn = 0;
4513            early_node_map[i].end_pfn = 0;
4514            removed = 1;
4515            continue;
4516        }
4517        if (early_node_map[i].start_pfn < start_pfn &&
4518            early_node_map[i].end_pfn > start_pfn) {
4519            unsigned long temp_end_pfn = early_node_map[i].end_pfn;
4520            early_node_map[i].end_pfn = start_pfn;
4521            if (temp_end_pfn > end_pfn)
4522                add_active_range(nid, end_pfn, temp_end_pfn);
4523            continue;
4524        }
4525        if (early_node_map[i].start_pfn >= start_pfn &&
4526            early_node_map[i].end_pfn > end_pfn &&
4527            early_node_map[i].start_pfn < end_pfn) {
4528            early_node_map[i].start_pfn = end_pfn;
4529            continue;
4530        }
4531    }
4532
4533    if (!removed)
4534        return;
4535
4536    /* remove the blank ones */
4537    for (i = nr_nodemap_entries - 1; i > 0; i--) {
4538        if (early_node_map[i].nid != nid)
4539            continue;
4540        if (early_node_map[i].end_pfn)
4541            continue;
4542        /* we found it, get rid of it */
4543        for (j = i; j < nr_nodemap_entries - 1; j++)
4544            memcpy(&early_node_map[j], &early_node_map[j+1],
4545                sizeof(early_node_map[j]));
4546        j = nr_nodemap_entries - 1;
4547        memset(&early_node_map[j], 0, sizeof(early_node_map[j]));
4548        nr_nodemap_entries--;
4549    }
4550}
4551
4552/**
4553 * remove_all_active_ranges - Remove all currently registered regions
4554 *
4555 * During discovery, it may be found that a table like SRAT is invalid
4556 * and an alternative discovery method must be used. This function removes
4557 * all currently registered regions.
4558 */
4559void __init remove_all_active_ranges(void)
4560{
4561    memset(early_node_map, 0, sizeof(early_node_map));
4562    nr_nodemap_entries = 0;
4563}
4564
4565/* Compare two active node_active_regions */
4566static int __init cmp_node_active_region(const void *a, const void *b)
4567{
4568    struct node_active_region *arange = (struct node_active_region *)a;
4569    struct node_active_region *brange = (struct node_active_region *)b;
4570
4571    /* Done this way to avoid overflows */
4572    if (arange->start_pfn > brange->start_pfn)
4573        return 1;
4574    if (arange->start_pfn < brange->start_pfn)
4575        return -1;
4576
4577    return 0;
4578}
4579
4580/* sort the node_map by start_pfn */
4581void __init sort_node_map(void)
4582{
4583    sort(early_node_map, (size_t)nr_nodemap_entries,
4584            sizeof(struct node_active_region),
4585            cmp_node_active_region, NULL);
4586}
4587
4588/* Find the lowest pfn for a node */
4589static unsigned long __init find_min_pfn_for_node(int nid)
4590{
4591    int i;
4592    unsigned long min_pfn = ULONG_MAX;
4593
4594    /* Assuming a sorted map, the first range found has the starting pfn */
4595    for_each_active_range_index_in_nid(i, nid)
4596        min_pfn = min(min_pfn, early_node_map[i].start_pfn);
4597
4598    if (min_pfn == ULONG_MAX) {
4599        printk(KERN_WARNING
4600            "Could not find start_pfn for node %d\n", nid);
4601        return 0;
4602    }
4603
4604    return min_pfn;
4605}
4606
4607/**
4608 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4609 *
4610 * It returns the minimum PFN based on information provided via
4611 * add_active_range().
4612 */
4613unsigned long __init find_min_pfn_with_active_regions(void)
4614{
4615    return find_min_pfn_for_node(MAX_NUMNODES);
4616}
4617
4618/*
4619 * early_calculate_totalpages()
4620 * Sum pages in active regions for movable zone.
4621 * Populate N_HIGH_MEMORY for calculating usable_nodes.
4622 */
4623static unsigned long __init early_calculate_totalpages(void)
4624{
4625    int i;
4626    unsigned long totalpages = 0;
4627
4628    for (i = 0; i < nr_nodemap_entries; i++) {
4629        unsigned long pages = early_node_map[i].end_pfn -
4630                        early_node_map[i].start_pfn;
4631        totalpages += pages;
4632        if (pages)
4633            node_set_state(early_node_map[i].nid, N_HIGH_MEMORY);
4634    }
4635      return totalpages;
4636}
4637
4638/*
4639 * Find the PFN the Movable zone begins in each node. Kernel memory
4640 * is spread evenly between nodes as long as the nodes have enough
4641 * memory. When they don't, some nodes will have more kernelcore than
4642 * others
4643 */
4644static void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn)
4645{
4646    int i, nid;
4647    unsigned long usable_startpfn;
4648    unsigned long kernelcore_node, kernelcore_remaining;
4649    /* save the state before borrow the nodemask */
4650    nodemask_t saved_node_state = node_states[N_HIGH_MEMORY];
4651    unsigned long totalpages = early_calculate_totalpages();
4652    int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
4653
4654    /*
4655     * If movablecore was specified, calculate what size of
4656     * kernelcore that corresponds so that memory usable for
4657     * any allocation type is evenly spread. If both kernelcore
4658     * and movablecore are specified, then the value of kernelcore
4659     * will be used for required_kernelcore if it's greater than
4660     * what movablecore would have allowed.
4661     */
4662    if (required_movablecore) {
4663        unsigned long corepages;
4664
4665        /*
4666         * Round-up so that ZONE_MOVABLE is at least as large as what
4667         * was requested by the user
4668         */
4669        required_movablecore =
4670            roundup(required_movablecore, MAX_ORDER_NR_PAGES);
4671        corepages = totalpages - required_movablecore;
4672
4673        required_kernelcore = max(required_kernelcore, corepages);
4674    }
4675
4676    /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4677    if (!required_kernelcore)
4678        goto out;
4679
4680    /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4681    find_usable_zone_for_movable();
4682    usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
4683
4684restart:
4685    /* Spread kernelcore memory as evenly as possible throughout nodes */
4686    kernelcore_node = required_kernelcore / usable_nodes;
4687    for_each_node_state(nid, N_HIGH_MEMORY) {
4688        /*
4689         * Recalculate kernelcore_node if the division per node
4690         * now exceeds what is necessary to satisfy the requested
4691         * amount of memory for the kernel
4692         */
4693        if (required_kernelcore < kernelcore_node)
4694            kernelcore_node = required_kernelcore / usable_nodes;
4695
4696        /*
4697         * As the map is walked, we track how much memory is usable
4698         * by the kernel using kernelcore_remaining. When it is
4699         * 0, the rest of the node is usable by ZONE_MOVABLE
4700         */
4701        kernelcore_remaining = kernelcore_node;
4702
4703        /* Go through each range of PFNs within this node */
4704        for_each_active_range_index_in_nid(i, nid) {
4705            unsigned long start_pfn, end_pfn;
4706            unsigned long size_pages;
4707
4708            start_pfn = max(early_node_map[i].start_pfn,
4709                        zone_movable_pfn[nid]);
4710            end_pfn = early_node_map[i].end_pfn;
4711            if (start_pfn >= end_pfn)
4712                continue;
4713
4714            /* Account for what is only usable for kernelcore */
4715            if (start_pfn < usable_startpfn) {
4716                unsigned long kernel_pages;
4717                kernel_pages = min(end_pfn, usable_startpfn)
4718                                - start_pfn;
4719
4720                kernelcore_remaining -= min(kernel_pages,
4721                            kernelcore_remaining);
4722                required_kernelcore -= min(kernel_pages,
4723                            required_kernelcore);
4724
4725                /* Continue if range is now fully accounted */
4726                if (end_pfn <= usable_startpfn) {
4727
4728                    /*
4729                     * Push zone_movable_pfn to the end so
4730                     * that if we have to rebalance
4731                     * kernelcore across nodes, we will
4732                     * not double account here
4733                     */
4734                    zone_movable_pfn[nid] = end_pfn;
4735                    continue;
4736                }
4737                start_pfn = usable_startpfn;
4738            }
4739
4740            /*
4741             * The usable PFN range for ZONE_MOVABLE is from
4742             * start_pfn->end_pfn. Calculate size_pages as the
4743             * number of pages used as kernelcore
4744             */
4745            size_pages = end_pfn - start_pfn;
4746            if (size_pages > kernelcore_remaining)
4747                size_pages = kernelcore_remaining;
4748            zone_movable_pfn[nid] = start_pfn + size_pages;
4749
4750            /*
4751             * Some kernelcore has been met, update counts and
4752             * break if the kernelcore for this node has been
4753             * satisified
4754             */
4755            required_kernelcore -= min(required_kernelcore,
4756                                size_pages);
4757            kernelcore_remaining -= size_pages;
4758            if (!kernelcore_remaining)
4759                break;
4760        }
4761    }
4762
4763    /*
4764     * If there is still required_kernelcore, we do another pass with one
4765     * less node in the count. This will push zone_movable_pfn[nid] further
4766     * along on the nodes that still have memory until kernelcore is
4767     * satisified
4768     */
4769    usable_nodes--;
4770    if (usable_nodes && required_kernelcore > usable_nodes)
4771        goto restart;
4772
4773    /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4774    for (nid = 0; nid < MAX_NUMNODES; nid++)
4775        zone_movable_pfn[nid] =
4776            roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
4777
4778out:
4779    /* restore the node_state */
4780    node_states[N_HIGH_MEMORY] = saved_node_state;
4781}
4782
4783/* Any regular memory on that node ? */
4784static void check_for_regular_memory(pg_data_t *pgdat)
4785{
4786#ifdef CONFIG_HIGHMEM
4787    enum zone_type zone_type;
4788
4789    for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) {
4790        struct zone *zone = &pgdat->node_zones[zone_type];
4791        if (zone->present_pages)
4792            node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY);
4793    }
4794#endif
4795}
4796
4797/**
4798 * free_area_init_nodes - Initialise all pg_data_t and zone data
4799 * @max_zone_pfn: an array of max PFNs for each zone
4800 *
4801 * This will call free_area_init_node() for each active node in the system.
4802 * Using the page ranges provided by add_active_range(), the size of each
4803 * zone in each node and their holes is calculated. If the maximum PFN
4804 * between two adjacent zones match, it is assumed that the zone is empty.
4805 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4806 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4807 * starts where the previous one ended. For example, ZONE_DMA32 starts
4808 * at arch_max_dma_pfn.
4809 */
4810void __init free_area_init_nodes(unsigned long *max_zone_pfn)
4811{
4812    unsigned long nid;
4813    int i;
4814
4815    /* Sort early_node_map as initialisation assumes it is sorted */
4816    sort_node_map();
4817
4818    /* Record where the zone boundaries are */
4819    memset(arch_zone_lowest_possible_pfn, 0,
4820                sizeof(arch_zone_lowest_possible_pfn));
4821    memset(arch_zone_highest_possible_pfn, 0,
4822                sizeof(arch_zone_highest_possible_pfn));
4823    arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
4824    arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
4825    for (i = 1; i < MAX_NR_ZONES; i++) {
4826        if (i == ZONE_MOVABLE)
4827            continue;
4828        arch_zone_lowest_possible_pfn[i] =
4829            arch_zone_highest_possible_pfn[i-1];
4830        arch_zone_highest_possible_pfn[i] =
4831            max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
4832    }
4833    arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
4834    arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
4835
4836    /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4837    memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
4838    find_zone_movable_pfns_for_nodes(zone_movable_pfn);
4839
4840    /* Print out the zone ranges */
4841    printk("Zone PFN ranges:\n");
4842    for (i = 0; i < MAX_NR_ZONES; i++) {
4843        if (i == ZONE_MOVABLE)
4844            continue;
4845        printk(" %-8s ", zone_names[i]);
4846        if (arch_zone_lowest_possible_pfn[i] ==
4847                arch_zone_highest_possible_pfn[i])
4848            printk("empty\n");
4849        else
4850            printk("%0#10lx -> %0#10lx\n",
4851                arch_zone_lowest_possible_pfn[i],
4852                arch_zone_highest_possible_pfn[i]);
4853    }
4854
4855    /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4856    printk("Movable zone start PFN for each node\n");
4857    for (i = 0; i < MAX_NUMNODES; i++) {
4858        if (zone_movable_pfn[i])
4859            printk(" Node %d: %lu\n", i, zone_movable_pfn[i]);
4860    }
4861
4862    /* Print out the early_node_map[] */
4863    printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries);
4864    for (i = 0; i < nr_nodemap_entries; i++)
4865        printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map[i].nid,
4866                        early_node_map[i].start_pfn,
4867                        early_node_map[i].end_pfn);
4868
4869    /* Initialise every node */
4870    mminit_verify_pageflags_layout();
4871    setup_nr_node_ids();
4872    for_each_online_node(nid) {
4873        pg_data_t *pgdat = NODE_DATA(nid);
4874        free_area_init_node(nid, NULL,
4875                find_min_pfn_for_node(nid), NULL);
4876
4877        /* Any memory on that node */
4878        if (pgdat->node_present_pages)
4879            node_set_state(nid, N_HIGH_MEMORY);
4880        check_for_regular_memory(pgdat);
4881    }
4882}
4883
4884static int __init cmdline_parse_core(char *p, unsigned long *core)
4885{
4886    unsigned long long coremem;
4887    if (!p)
4888        return -EINVAL;
4889
4890    coremem = memparse(p, &p);
4891    *core = coremem >> PAGE_SHIFT;
4892
4893    /* Paranoid check that UL is enough for the coremem value */
4894    WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
4895
4896    return 0;
4897}
4898
4899/*
4900 * kernelcore=size sets the amount of memory for use for allocations that
4901 * cannot be reclaimed or migrated.
4902 */
4903static int __init cmdline_parse_kernelcore(char *p)
4904{
4905    return cmdline_parse_core(p, &required_kernelcore);
4906}
4907
4908/*
4909 * movablecore=size sets the amount of memory for use for allocations that
4910 * can be reclaimed or migrated.
4911 */
4912static int __init cmdline_parse_movablecore(char *p)
4913{
4914    return cmdline_parse_core(p, &required_movablecore);
4915}
4916
4917early_param("kernelcore", cmdline_parse_kernelcore);
4918early_param("movablecore", cmdline_parse_movablecore);
4919
4920#endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4921
4922/**
4923 * set_dma_reserve - set the specified number of pages reserved in the first zone
4924 * @new_dma_reserve: The number of pages to mark reserved
4925 *
4926 * The per-cpu batchsize and zone watermarks are determined by present_pages.
4927 * In the DMA zone, a significant percentage may be consumed by kernel image
4928 * and other unfreeable allocations which can skew the watermarks badly. This
4929 * function may optionally be used to account for unfreeable pages in the
4930 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
4931 * smaller per-cpu batchsize.
4932 */
4933void __init set_dma_reserve(unsigned long new_dma_reserve)
4934{
4935    dma_reserve = new_dma_reserve;
4936}
4937
4938void __init free_area_init(unsigned long *zones_size)
4939{
4940    free_area_init_node(0, zones_size,
4941            __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
4942}
4943
4944static int page_alloc_cpu_notify(struct notifier_block *self,
4945                 unsigned long action, void *hcpu)
4946{
4947    int cpu = (unsigned long)hcpu;
4948
4949    if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
4950        drain_pages(cpu);
4951
4952        /*
4953         * Spill the event counters of the dead processor
4954         * into the current processors event counters.
4955         * This artificially elevates the count of the current
4956         * processor.
4957         */
4958        vm_events_fold_cpu(cpu);
4959
4960        /*
4961         * Zero the differential counters of the dead processor
4962         * so that the vm statistics are consistent.
4963         *
4964         * This is only okay since the processor is dead and cannot
4965         * race with what we are doing.
4966         */
4967        refresh_cpu_vm_stats(cpu);
4968    }
4969    return NOTIFY_OK;
4970}
4971
4972void __init page_alloc_init(void)
4973{
4974    hotcpu_notifier(page_alloc_cpu_notify, 0);
4975}
4976
4977/*
4978 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
4979 * or min_free_kbytes changes.
4980 */
4981static void calculate_totalreserve_pages(void)
4982{
4983    struct pglist_data *pgdat;
4984    unsigned long reserve_pages = 0;
4985    enum zone_type i, j;
4986
4987    for_each_online_pgdat(pgdat) {
4988        for (i = 0; i < MAX_NR_ZONES; i++) {
4989            struct zone *zone = pgdat->node_zones + i;
4990            unsigned long max = 0;
4991
4992            /* Find valid and maximum lowmem_reserve in the zone */
4993            for (j = i; j < MAX_NR_ZONES; j++) {
4994                if (zone->lowmem_reserve[j] > max)
4995                    max = zone->lowmem_reserve[j];
4996            }
4997
4998            /* we treat the high watermark as reserved pages. */
4999            max += high_wmark_pages(zone);
5000
5001            if (max > zone->present_pages)
5002                max = zone->present_pages;
5003            reserve_pages += max;
5004        }
5005    }
5006    totalreserve_pages = reserve_pages;
5007}
5008
5009/*
5010 * setup_per_zone_lowmem_reserve - called whenever
5011 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5012 * has a correct pages reserved value, so an adequate number of
5013 * pages are left in the zone after a successful __alloc_pages().
5014 */
5015static void setup_per_zone_lowmem_reserve(void)
5016{
5017    struct pglist_data *pgdat;
5018    enum zone_type j, idx;
5019
5020    for_each_online_pgdat(pgdat) {
5021        for (j = 0; j < MAX_NR_ZONES; j++) {
5022            struct zone *zone = pgdat->node_zones + j;
5023            unsigned long present_pages = zone->present_pages;
5024
5025            zone->lowmem_reserve[j] = 0;
5026
5027            idx = j;
5028            while (idx) {
5029                struct zone *lower_zone;
5030
5031                idx--;
5032
5033                if (sysctl_lowmem_reserve_ratio[idx] < 1)
5034                    sysctl_lowmem_reserve_ratio[idx] = 1;
5035
5036                lower_zone = pgdat->node_zones + idx;
5037                lower_zone->lowmem_reserve[j] = present_pages /
5038                    sysctl_lowmem_reserve_ratio[idx];
5039                present_pages += lower_zone->present_pages;
5040            }
5041        }
5042    }
5043
5044    /* update totalreserve_pages */
5045    calculate_totalreserve_pages();
5046}
5047
5048/**
5049 * setup_per_zone_wmarks - called when min_free_kbytes changes
5050 * or when memory is hot-{added|removed}
5051 *
5052 * Ensures that the watermark[min,low,high] values for each zone are set
5053 * correctly with respect to min_free_kbytes.
5054 */
5055void setup_per_zone_wmarks(void)
5056{
5057    unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5058    unsigned long lowmem_pages = 0;
5059    struct zone *zone;
5060    unsigned long flags;
5061
5062    /* Calculate total number of !ZONE_HIGHMEM pages */
5063    for_each_zone(zone) {
5064        if (!is_highmem(zone))
5065            lowmem_pages += zone->present_pages;
5066    }
5067
5068    for_each_zone(zone) {
5069        u64 tmp;
5070
5071        spin_lock_irqsave(&zone->lock, flags);
5072        tmp = (u64)pages_min * zone->present_pages;
5073        do_div(tmp, lowmem_pages);
5074        if (is_highmem(zone)) {
5075            /*
5076             * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5077             * need highmem pages, so cap pages_min to a small
5078             * value here.
5079             *
5080             * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5081             * deltas controls asynch page reclaim, and so should
5082             * not be capped for highmem.
5083             */
5084            int min_pages;
5085
5086            min_pages = zone->present_pages / 1024;
5087            if (min_pages < SWAP_CLUSTER_MAX)
5088                min_pages = SWAP_CLUSTER_MAX;
5089            if (min_pages > 128)
5090                min_pages = 128;
5091            zone->watermark[WMARK_MIN] = min_pages;
5092        } else {
5093            /*
5094             * If it's a lowmem zone, reserve a number of pages
5095             * proportionate to the zone's size.
5096             */
5097            zone->watermark[WMARK_MIN] = tmp;
5098        }
5099
5100        zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
5101        zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5102        setup_zone_migrate_reserve(zone);
5103        spin_unlock_irqrestore(&zone->lock, flags);
5104    }
5105
5106    /* update totalreserve_pages */
5107    calculate_totalreserve_pages();
5108}
5109
5110/*
5111 * The inactive anon list should be small enough that the VM never has to
5112 * do too much work, but large enough that each inactive page has a chance
5113 * to be referenced again before it is swapped out.
5114 *
5115 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5116 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5117 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5118 * the anonymous pages are kept on the inactive list.
5119 *
5120 * total target max
5121 * memory ratio inactive anon
5122 * -------------------------------------
5123 * 10MB 1 5MB
5124 * 100MB 1 50MB
5125 * 1GB 3 250MB
5126 * 10GB 10 0.9GB
5127 * 100GB 31 3GB
5128 * 1TB 101 10GB
5129 * 10TB 320 32GB
5130 */
5131static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5132{
5133    unsigned int gb, ratio;
5134
5135    /* Zone size in gigabytes */
5136    gb = zone->present_pages >> (30 - PAGE_SHIFT);
5137    if (gb)
5138        ratio = int_sqrt(10 * gb);
5139    else
5140        ratio = 1;
5141
5142    zone->inactive_ratio = ratio;
5143}
5144
5145static void __meminit setup_per_zone_inactive_ratio(void)
5146{
5147    struct zone *zone;
5148
5149    for_each_zone(zone)
5150        calculate_zone_inactive_ratio(zone);
5151}
5152
5153/*
5154 * Initialise min_free_kbytes.
5155 *
5156 * For small machines we want it small (128k min). For large machines
5157 * we want it large (64MB max). But it is not linear, because network
5158 * bandwidth does not increase linearly with machine size. We use
5159 *
5160 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5161 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5162 *
5163 * which yields
5164 *
5165 * 16MB: 512k
5166 * 32MB: 724k
5167 * 64MB: 1024k
5168 * 128MB: 1448k
5169 * 256MB: 2048k
5170 * 512MB: 2896k
5171 * 1024MB: 4096k
5172 * 2048MB: 5792k
5173 * 4096MB: 8192k
5174 * 8192MB: 11584k
5175 * 16384MB: 16384k
5176 */
5177int __meminit init_per_zone_wmark_min(void)
5178{
5179    unsigned long lowmem_kbytes;
5180
5181    lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5182
5183    min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5184    if (min_free_kbytes < 128)
5185        min_free_kbytes = 128;
5186    if (min_free_kbytes > 65536)
5187        min_free_kbytes = 65536;
5188    setup_per_zone_wmarks();
5189    refresh_zone_stat_thresholds();
5190    setup_per_zone_lowmem_reserve();
5191    setup_per_zone_inactive_ratio();
5192    return 0;
5193}
5194module_init(init_per_zone_wmark_min)
5195
5196/*
5197 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5198 * that we can call two helper functions whenever min_free_kbytes
5199 * changes.
5200 */
5201int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
5202    void __user *buffer, size_t *length, loff_t *ppos)
5203{
5204    proc_dointvec(table, write, buffer, length, ppos);
5205    if (write)
5206        setup_per_zone_wmarks();
5207    return 0;
5208}
5209
5210#ifdef CONFIG_NUMA
5211int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
5212    void __user *buffer, size_t *length, loff_t *ppos)
5213{
5214    struct zone *zone;
5215    int rc;
5216
5217    rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5218    if (rc)
5219        return rc;
5220
5221    for_each_zone(zone)
5222        zone->min_unmapped_pages = (zone->present_pages *
5223                sysctl_min_unmapped_ratio) / 100;
5224    return 0;
5225}
5226
5227int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
5228    void __user *buffer, size_t *length, loff_t *ppos)
5229{
5230    struct zone *zone;
5231    int rc;
5232
5233    rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5234    if (rc)
5235        return rc;
5236
5237    for_each_zone(zone)
5238        zone->min_slab_pages = (zone->present_pages *
5239                sysctl_min_slab_ratio) / 100;
5240    return 0;
5241}
5242#endif
5243
5244/*
5245 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5246 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5247 * whenever sysctl_lowmem_reserve_ratio changes.
5248 *
5249 * The reserve ratio obviously has absolutely no relation with the
5250 * minimum watermarks. The lowmem reserve ratio can only make sense
5251 * if in function of the boot time zone sizes.
5252 */
5253int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
5254    void __user *buffer, size_t *length, loff_t *ppos)
5255{
5256    proc_dointvec_minmax(table, write, buffer, length, ppos);
5257    setup_per_zone_lowmem_reserve();
5258    return 0;
5259}
5260
5261/*
5262 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5263 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
5264 * can have before it gets flushed back to buddy allocator.
5265 */
5266
5267int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
5268    void __user *buffer, size_t *length, loff_t *ppos)
5269{
5270    struct zone *zone;
5271    unsigned int cpu;
5272    int ret;
5273
5274    ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5275    if (!write || (ret == -EINVAL))
5276        return ret;
5277    for_each_populated_zone(zone) {
5278        for_each_possible_cpu(cpu) {
5279            unsigned long high;
5280            high = zone->present_pages / percpu_pagelist_fraction;
5281            setup_pagelist_highmark(
5282                per_cpu_ptr(zone->pageset, cpu), high);
5283        }
5284    }
5285    return 0;
5286}
5287
5288int hashdist = HASHDIST_DEFAULT;
5289
5290#ifdef CONFIG_NUMA
5291static int __init set_hashdist(char *str)
5292{
5293    if (!str)
5294        return 0;
5295    hashdist = simple_strtoul(str, &str, 0);
5296    return 1;
5297}
5298__setup("hashdist=", set_hashdist);
5299#endif
5300
5301/*
5302 * allocate a large system hash table from bootmem
5303 * - it is assumed that the hash table must contain an exact power-of-2
5304 * quantity of entries
5305 * - limit is the number of hash buckets, not the total allocation size
5306 */
5307void *__init alloc_large_system_hash(const char *tablename,
5308                     unsigned long bucketsize,
5309                     unsigned long numentries,
5310                     int scale,
5311                     int flags,
5312                     unsigned int *_hash_shift,
5313                     unsigned int *_hash_mask,
5314                     unsigned long limit)
5315{
5316    unsigned long long max = limit;
5317    unsigned long log2qty, size;
5318    void *table = NULL;
5319
5320    /* allow the kernel cmdline to have a say */
5321    if (!numentries) {
5322        /* round applicable memory size up to nearest megabyte */
5323        numentries = nr_kernel_pages;
5324        numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
5325        numentries >>= 20 - PAGE_SHIFT;
5326        numentries <<= 20 - PAGE_SHIFT;
5327
5328        /* limit to 1 bucket per 2^scale bytes of low memory */
5329        if (scale > PAGE_SHIFT)
5330            numentries >>= (scale - PAGE_SHIFT);
5331        else
5332            numentries <<= (PAGE_SHIFT - scale);
5333
5334        /* Make sure we've got at least a 0-order allocation.. */
5335        if (unlikely(flags & HASH_SMALL)) {
5336            /* Makes no sense without HASH_EARLY */
5337            WARN_ON(!(flags & HASH_EARLY));
5338            if (!(numentries >> *_hash_shift)) {
5339                numentries = 1UL << *_hash_shift;
5340                BUG_ON(!numentries);
5341            }
5342        } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
5343            numentries = PAGE_SIZE / bucketsize;
5344    }
5345    numentries = roundup_pow_of_two(numentries);
5346
5347    /* limit allocation size to 1/16 total memory by default */
5348    if (max == 0) {
5349        max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5350        do_div(max, bucketsize);
5351    }
5352
5353    if (numentries > max)
5354        numentries = max;
5355
5356    log2qty = ilog2(numentries);
5357
5358    do {
5359        size = bucketsize << log2qty;
5360        if (flags & HASH_EARLY)
5361            table = alloc_bootmem_nopanic(size);
5362        else if (hashdist)
5363            table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5364        else {
5365            /*
5366             * If bucketsize is not a power-of-two, we may free
5367             * some pages at the end of hash table which
5368             * alloc_pages_exact() automatically does
5369             */
5370            if (get_order(size) < MAX_ORDER) {
5371                table = alloc_pages_exact(size, GFP_ATOMIC);
5372                kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5373            }
5374        }
5375    } while (!table && size > PAGE_SIZE && --log2qty);
5376
5377    if (!table)
5378        panic("Failed to allocate %s hash table\n", tablename);
5379
5380    printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
5381           tablename,
5382           (1UL << log2qty),
5383           ilog2(size) - PAGE_SHIFT,
5384           size);
5385
5386    if (_hash_shift)
5387        *_hash_shift = log2qty;
5388    if (_hash_mask)
5389        *_hash_mask = (1 << log2qty) - 1;
5390
5391    return table;
5392}
5393
5394/* Return a pointer to the bitmap storing bits affecting a block of pages */
5395static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
5396                            unsigned long pfn)
5397{
5398#ifdef CONFIG_SPARSEMEM
5399    return __pfn_to_section(pfn)->pageblock_flags;
5400#else
5401    return zone->pageblock_flags;
5402#endif /* CONFIG_SPARSEMEM */
5403}
5404
5405static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
5406{
5407#ifdef CONFIG_SPARSEMEM
5408    pfn &= (PAGES_PER_SECTION-1);
5409    return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5410#else
5411    pfn = pfn - zone->zone_start_pfn;
5412    return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5413#endif /* CONFIG_SPARSEMEM */
5414}
5415
5416/**
5417 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5418 * @page: The page within the block of interest
5419 * @start_bitidx: The first bit of interest to retrieve
5420 * @end_bitidx: The last bit of interest
5421 * returns pageblock_bits flags
5422 */
5423unsigned long get_pageblock_flags_group(struct page *page,
5424                    int start_bitidx, int end_bitidx)
5425{
5426    struct zone *zone;
5427    unsigned long *bitmap;
5428    unsigned long pfn, bitidx;
5429    unsigned long flags = 0;
5430    unsigned long value = 1;
5431
5432    zone = page_zone(page);
5433    pfn = page_to_pfn(page);
5434    bitmap = get_pageblock_bitmap(zone, pfn);
5435    bitidx = pfn_to_bitidx(zone, pfn);
5436
5437    for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5438        if (test_bit(bitidx + start_bitidx, bitmap))
5439            flags |= value;
5440
5441    return flags;
5442}
5443
5444/**
5445 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5446 * @page: The page within the block of interest
5447 * @start_bitidx: The first bit of interest
5448 * @end_bitidx: The last bit of interest
5449 * @flags: The flags to set
5450 */
5451void set_pageblock_flags_group(struct page *page, unsigned long flags,
5452                    int start_bitidx, int end_bitidx)
5453{
5454    struct zone *zone;
5455    unsigned long *bitmap;
5456    unsigned long pfn, bitidx;
5457    unsigned long value = 1;
5458
5459    zone = page_zone(page);
5460    pfn = page_to_pfn(page);
5461    bitmap = get_pageblock_bitmap(zone, pfn);
5462    bitidx = pfn_to_bitidx(zone, pfn);
5463    VM_BUG_ON(pfn < zone->zone_start_pfn);
5464    VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages);
5465
5466    for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5467        if (flags & value)
5468            __set_bit(bitidx + start_bitidx, bitmap);
5469        else
5470            __clear_bit(bitidx + start_bitidx, bitmap);
5471}
5472
5473/*
5474 * This is designed as sub function...plz see page_isolation.c also.
5475 * set/clear page block's type to be ISOLATE.
5476 * page allocater never alloc memory from ISOLATE block.
5477 */
5478
5479static int
5480__count_immobile_pages(struct zone *zone, struct page *page, int count)
5481{
5482    unsigned long pfn, iter, found;
5483    /*
5484     * For avoiding noise data, lru_add_drain_all() should be called
5485     * If ZONE_MOVABLE, the zone never contains immobile pages
5486     */
5487    if (zone_idx(zone) == ZONE_MOVABLE)
5488        return true;
5489
5490    if (get_pageblock_migratetype(page) == MIGRATE_MOVABLE)
5491        return true;
5492
5493    pfn = page_to_pfn(page);
5494    for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
5495        unsigned long check = pfn + iter;
5496
5497        if (!pfn_valid_within(check))
5498            continue;
5499
5500        page = pfn_to_page(check);
5501        if (!page_count(page)) {
5502            if (PageBuddy(page))
5503                iter += (1 << page_order(page)) - 1;
5504            continue;
5505        }
5506        if (!PageLRU(page))
5507            found++;
5508        /*
5509         * If there are RECLAIMABLE pages, we need to check it.
5510         * But now, memory offline itself doesn't call shrink_slab()
5511         * and it still to be fixed.
5512         */
5513        /*
5514         * If the page is not RAM, page_count()should be 0.
5515         * we don't need more check. This is an _used_ not-movable page.
5516         *
5517         * The problematic thing here is PG_reserved pages. PG_reserved
5518         * is set to both of a memory hole page and a _used_ kernel
5519         * page at boot.
5520         */
5521        if (found > count)
5522            return false;
5523    }
5524    return true;
5525}
5526
5527bool is_pageblock_removable_nolock(struct page *page)
5528{
5529    struct zone *zone = page_zone(page);
5530    return __count_immobile_pages(zone, page, 0);
5531}
5532
5533int set_migratetype_isolate(struct page *page)
5534{
5535    struct zone *zone;
5536    unsigned long flags, pfn;
5537    struct memory_isolate_notify arg;
5538    int notifier_ret;
5539    int ret = -EBUSY;
5540
5541    zone = page_zone(page);
5542
5543    spin_lock_irqsave(&zone->lock, flags);
5544
5545    pfn = page_to_pfn(page);
5546    arg.start_pfn = pfn;
5547    arg.nr_pages = pageblock_nr_pages;
5548    arg.pages_found = 0;
5549
5550    /*
5551     * It may be possible to isolate a pageblock even if the
5552     * migratetype is not MIGRATE_MOVABLE. The memory isolation
5553     * notifier chain is used by balloon drivers to return the
5554     * number of pages in a range that are held by the balloon
5555     * driver to shrink memory. If all the pages are accounted for
5556     * by balloons, are free, or on the LRU, isolation can continue.
5557     * Later, for example, when memory hotplug notifier runs, these
5558     * pages reported as "can be isolated" should be isolated(freed)
5559     * by the balloon driver through the memory notifier chain.
5560     */
5561    notifier_ret = memory_isolate_notify(MEM_ISOLATE_COUNT, &arg);
5562    notifier_ret = notifier_to_errno(notifier_ret);
5563    if (notifier_ret)
5564        goto out;
5565    /*
5566     * FIXME: Now, memory hotplug doesn't call shrink_slab() by itself.
5567     * We just check MOVABLE pages.
5568     */
5569    if (__count_immobile_pages(zone, page, arg.pages_found))
5570        ret = 0;
5571
5572    /*
5573     * immobile means "not-on-lru" paes. If immobile is larger than
5574     * removable-by-driver pages reported by notifier, we'll fail.
5575     */
5576
5577out:
5578    if (!ret) {
5579        set_pageblock_migratetype(page, MIGRATE_ISOLATE);
5580        move_freepages_block(zone, page, MIGRATE_ISOLATE);
5581    }
5582
5583    spin_unlock_irqrestore(&zone->lock, flags);
5584    if (!ret)
5585        drain_all_pages();
5586    return ret;
5587}
5588
5589void unset_migratetype_isolate(struct page *page)
5590{
5591    struct zone *zone;
5592    unsigned long flags;
5593    zone = page_zone(page);
5594    spin_lock_irqsave(&zone->lock, flags);
5595    if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
5596        goto out;
5597    set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5598    move_freepages_block(zone, page, MIGRATE_MOVABLE);
5599out:
5600    spin_unlock_irqrestore(&zone->lock, flags);
5601}
5602
5603#ifdef CONFIG_MEMORY_HOTREMOVE
5604/*
5605 * All pages in the range must be isolated before calling this.
5606 */
5607void
5608__offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
5609{
5610    struct page *page;
5611    struct zone *zone;
5612    int order, i;
5613    unsigned long pfn;
5614    unsigned long flags;
5615    /* find the first valid pfn */
5616    for (pfn = start_pfn; pfn < end_pfn; pfn++)
5617        if (pfn_valid(pfn))
5618            break;
5619    if (pfn == end_pfn)
5620        return;
5621    zone = page_zone(pfn_to_page(pfn));
5622    spin_lock_irqsave(&zone->lock, flags);
5623    pfn = start_pfn;
5624    while (pfn < end_pfn) {
5625        if (!pfn_valid(pfn)) {
5626            pfn++;
5627            continue;
5628        }
5629        page = pfn_to_page(pfn);
5630        BUG_ON(page_count(page));
5631        BUG_ON(!PageBuddy(page));
5632        order = page_order(page);
5633#ifdef CONFIG_DEBUG_VM
5634        printk(KERN_INFO "remove from free list %lx %d %lx\n",
5635               pfn, 1 << order, end_pfn);
5636#endif
5637        list_del(&page->lru);
5638        rmv_page_order(page);
5639        zone->free_area[order].nr_free--;
5640        __mod_zone_page_state(zone, NR_FREE_PAGES,
5641                      - (1UL << order));
5642        for (i = 0; i < (1 << order); i++)
5643            SetPageReserved((page+i));
5644        pfn += (1 << order);
5645    }
5646    spin_unlock_irqrestore(&zone->lock, flags);
5647}
5648#endif
5649
5650#ifdef CONFIG_MEMORY_FAILURE
5651bool is_free_buddy_page(struct page *page)
5652{
5653    struct zone *zone = page_zone(page);
5654    unsigned long pfn = page_to_pfn(page);
5655    unsigned long flags;
5656    int order;
5657
5658    spin_lock_irqsave(&zone->lock, flags);
5659    for (order = 0; order < MAX_ORDER; order++) {
5660        struct page *page_head = page - (pfn & ((1 << order) - 1));
5661
5662        if (PageBuddy(page_head) && page_order(page_head) >= order)
5663            break;
5664    }
5665    spin_unlock_irqrestore(&zone->lock, flags);
5666
5667    return order < MAX_ORDER;
5668}
5669#endif
5670
5671static struct trace_print_flags pageflag_names[] = {
5672    {1UL << PG_locked, "locked" },
5673    {1UL << PG_error, "error" },
5674    {1UL << PG_referenced, "referenced" },
5675    {1UL << PG_uptodate, "uptodate" },
5676    {1UL << PG_dirty, "dirty" },
5677    {1UL << PG_lru, "lru" },
5678    {1UL << PG_active, "active" },
5679    {1UL << PG_slab, "slab" },
5680    {1UL << PG_owner_priv_1, "owner_priv_1" },
5681    {1UL << PG_arch_1, "arch_1" },
5682    {1UL << PG_reserved, "reserved" },
5683    {1UL << PG_private, "private" },
5684    {1UL << PG_private_2, "private_2" },
5685    {1UL << PG_writeback, "writeback" },
5686#ifdef CONFIG_PAGEFLAGS_EXTENDED
5687    {1UL << PG_head, "head" },
5688    {1UL << PG_tail, "tail" },
5689#else
5690    {1UL << PG_compound, "compound" },
5691#endif
5692    {1UL << PG_swapcache, "swapcache" },
5693    {1UL << PG_mappedtodisk, "mappedtodisk" },
5694    {1UL << PG_reclaim, "reclaim" },
5695    {1UL << PG_swapbacked, "swapbacked" },
5696    {1UL << PG_unevictable, "unevictable" },
5697#ifdef CONFIG_MMU
5698    {1UL << PG_mlocked, "mlocked" },
5699#endif
5700#ifdef CONFIG_ARCH_USES_PG_UNCACHED
5701    {1UL << PG_uncached, "uncached" },
5702#endif
5703#ifdef CONFIG_MEMORY_FAILURE
5704    {1UL << PG_hwpoison, "hwpoison" },
5705#endif
5706    {-1UL, NULL },
5707};
5708
5709static void dump_page_flags(unsigned long flags)
5710{
5711    const char *delim = "";
5712    unsigned long mask;
5713    int i;
5714
5715    printk(KERN_ALERT "page flags: %#lx(", flags);
5716
5717    /* remove zone id */
5718    flags &= (1UL << NR_PAGEFLAGS) - 1;
5719
5720    for (i = 0; pageflag_names[i].name && flags; i++) {
5721
5722        mask = pageflag_names[i].mask;
5723        if ((flags & mask) != mask)
5724            continue;
5725
5726        flags &= ~mask;
5727        printk("%s%s", delim, pageflag_names[i].name);
5728        delim = "|";
5729    }
5730
5731    /* check for left over flags */
5732    if (flags)
5733        printk("%s%#lx", delim, flags);
5734
5735    printk(")\n");
5736}
5737
5738void dump_page(struct page *page)
5739{
5740    printk(KERN_ALERT
5741           "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
5742        page, atomic_read(&page->_count), page_mapcount(page),
5743        page->mapping, page->index);
5744    dump_page_flags(page->flags);
5745    mem_cgroup_print_bad_page(page);
5746}
5747

Archive Download this file



interactive