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/compaction.h>
55#include <trace/events/kmem.h>
56#include <linux/ftrace_event.h>
57#include <linux/memcontrol.h>
58#include <linux/prefetch.h>
59#include <linux/migrate.h>
60#include <linux/page-debug-flags.h>
61
62#include <asm/tlbflush.h>
63#include <asm/div64.h>
64#include "internal.h"
65
66#ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
67DEFINE_PER_CPU(int, numa_node);
68EXPORT_PER_CPU_SYMBOL(numa_node);
69#endif
70
71#ifdef CONFIG_HAVE_MEMORYLESS_NODES
72/*
73 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
74 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
75 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
76 * defined in <linux/topology.h>.
77 */
78DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
79EXPORT_PER_CPU_SYMBOL(_numa_mem_);
80#endif
81
82/*
83 * Array of node states.
84 */
85nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
86    [N_POSSIBLE] = NODE_MASK_ALL,
87    [N_ONLINE] = { { [0] = 1UL } },
88#ifndef CONFIG_NUMA
89    [N_NORMAL_MEMORY] = { { [0] = 1UL } },
90#ifdef CONFIG_HIGHMEM
91    [N_HIGH_MEMORY] = { { [0] = 1UL } },
92#endif
93    [N_CPU] = { { [0] = 1UL } },
94#endif /* NUMA */
95};
96EXPORT_SYMBOL(node_states);
97
98unsigned long totalram_pages __read_mostly;
99unsigned long totalreserve_pages __read_mostly;
100/*
101 * When calculating the number of globally allowed dirty pages, there
102 * is a certain number of per-zone reserves that should not be
103 * considered dirtyable memory. This is the sum of those reserves
104 * over all existing zones that contribute dirtyable memory.
105 */
106unsigned long dirty_balance_reserve __read_mostly;
107
108int percpu_pagelist_fraction;
109gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
110
111#ifdef CONFIG_PM_SLEEP
112/*
113 * The following functions are used by the suspend/hibernate code to temporarily
114 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
115 * while devices are suspended. To avoid races with the suspend/hibernate code,
116 * they should always be called with pm_mutex held (gfp_allowed_mask also should
117 * only be modified with pm_mutex held, unless the suspend/hibernate code is
118 * guaranteed not to run in parallel with that modification).
119 */
120
121static gfp_t saved_gfp_mask;
122
123void pm_restore_gfp_mask(void)
124{
125    WARN_ON(!mutex_is_locked(&pm_mutex));
126    if (saved_gfp_mask) {
127        gfp_allowed_mask = saved_gfp_mask;
128        saved_gfp_mask = 0;
129    }
130}
131
132void pm_restrict_gfp_mask(void)
133{
134    WARN_ON(!mutex_is_locked(&pm_mutex));
135    WARN_ON(saved_gfp_mask);
136    saved_gfp_mask = gfp_allowed_mask;
137    gfp_allowed_mask &= ~GFP_IOFS;
138}
139
140bool pm_suspended_storage(void)
141{
142    if ((gfp_allowed_mask & GFP_IOFS) == GFP_IOFS)
143        return false;
144    return true;
145}
146#endif /* CONFIG_PM_SLEEP */
147
148#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
149int pageblock_order __read_mostly;
150#endif
151
152static void __free_pages_ok(struct page *page, unsigned int order);
153
154/*
155 * results with 256, 32 in the lowmem_reserve sysctl:
156 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
157 * 1G machine -> (16M dma, 784M normal, 224M high)
158 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
159 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
160 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
161 *
162 * TBD: should special case ZONE_DMA32 machines here - in those we normally
163 * don't need any ZONE_NORMAL reservation
164 */
165int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
166#ifdef CONFIG_ZONE_DMA
167     256,
168#endif
169#ifdef CONFIG_ZONE_DMA32
170     256,
171#endif
172#ifdef CONFIG_HIGHMEM
173     32,
174#endif
175     32,
176};
177
178EXPORT_SYMBOL(totalram_pages);
179
180static char * const zone_names[MAX_NR_ZONES] = {
181#ifdef CONFIG_ZONE_DMA
182     "DMA",
183#endif
184#ifdef CONFIG_ZONE_DMA32
185     "DMA32",
186#endif
187     "Normal",
188#ifdef CONFIG_HIGHMEM
189     "HighMem",
190#endif
191     "Movable",
192};
193
194int min_free_kbytes = 1024;
195
196static unsigned long __meminitdata nr_kernel_pages;
197static unsigned long __meminitdata nr_all_pages;
198static unsigned long __meminitdata dma_reserve;
199
200#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
201static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
202static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
203static unsigned long __initdata required_kernelcore;
204static unsigned long __initdata required_movablecore;
205static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
206
207/* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
208int movable_zone;
209EXPORT_SYMBOL(movable_zone);
210#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
211
212#if MAX_NUMNODES > 1
213int nr_node_ids __read_mostly = MAX_NUMNODES;
214int nr_online_nodes __read_mostly = 1;
215EXPORT_SYMBOL(nr_node_ids);
216EXPORT_SYMBOL(nr_online_nodes);
217#endif
218
219int page_group_by_mobility_disabled __read_mostly;
220
221/*
222 * NOTE:
223 * Don't use set_pageblock_migratetype(page, MIGRATE_ISOLATE) directly.
224 * Instead, use {un}set_pageblock_isolate.
225 */
226void set_pageblock_migratetype(struct page *page, int migratetype)
227{
228
229    if (unlikely(page_group_by_mobility_disabled))
230        migratetype = MIGRATE_UNMOVABLE;
231
232    set_pageblock_flags_group(page, (unsigned long)migratetype,
233                    PB_migrate, PB_migrate_end);
234}
235
236bool oom_killer_disabled __read_mostly;
237
238#ifdef CONFIG_DEBUG_VM
239static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
240{
241    int ret = 0;
242    unsigned seq;
243    unsigned long pfn = page_to_pfn(page);
244
245    do {
246        seq = zone_span_seqbegin(zone);
247        if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
248            ret = 1;
249        else if (pfn < zone->zone_start_pfn)
250            ret = 1;
251    } while (zone_span_seqretry(zone, seq));
252
253    return ret;
254}
255
256static int page_is_consistent(struct zone *zone, struct page *page)
257{
258    if (!pfn_valid_within(page_to_pfn(page)))
259        return 0;
260    if (zone != page_zone(page))
261        return 0;
262
263    return 1;
264}
265/*
266 * Temporary debugging check for pages not lying within a given zone.
267 */
268static int bad_range(struct zone *zone, struct page *page)
269{
270    if (page_outside_zone_boundaries(zone, page))
271        return 1;
272    if (!page_is_consistent(zone, page))
273        return 1;
274
275    return 0;
276}
277#else
278static inline int bad_range(struct zone *zone, struct page *page)
279{
280    return 0;
281}
282#endif
283
284static void bad_page(struct page *page)
285{
286    static unsigned long resume;
287    static unsigned long nr_shown;
288    static unsigned long nr_unshown;
289
290    /* Don't complain about poisoned pages */
291    if (PageHWPoison(page)) {
292        reset_page_mapcount(page); /* remove PageBuddy */
293        return;
294    }
295
296    /*
297     * Allow a burst of 60 reports, then keep quiet for that minute;
298     * or allow a steady drip of one report per second.
299     */
300    if (nr_shown == 60) {
301        if (time_before(jiffies, resume)) {
302            nr_unshown++;
303            goto out;
304        }
305        if (nr_unshown) {
306            printk(KERN_ALERT
307                  "BUG: Bad page state: %lu messages suppressed\n",
308                nr_unshown);
309            nr_unshown = 0;
310        }
311        nr_shown = 0;
312    }
313    if (nr_shown++ == 0)
314        resume = jiffies + 60 * HZ;
315
316    printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
317        current->comm, page_to_pfn(page));
318    dump_page(page);
319
320    print_modules();
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 tail pages have their ->first_page
336 * pointing at the head page.
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        __SetPageTail(p);
359        set_page_count(p, 0);
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
405#ifdef CONFIG_DEBUG_PAGEALLOC
406unsigned int _debug_guardpage_minorder;
407
408static int __init debug_guardpage_minorder_setup(char *buf)
409{
410    unsigned long res;
411
412    if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
413        printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
414        return 0;
415    }
416    _debug_guardpage_minorder = res;
417    printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
418    return 0;
419}
420__setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
421
422static inline void set_page_guard_flag(struct page *page)
423{
424    __set_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
425}
426
427static inline void clear_page_guard_flag(struct page *page)
428{
429    __clear_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
430}
431#else
432static inline void set_page_guard_flag(struct page *page) { }
433static inline void clear_page_guard_flag(struct page *page) { }
434#endif
435
436static inline void set_page_order(struct page *page, int order)
437{
438    set_page_private(page, order);
439    __SetPageBuddy(page);
440}
441
442static inline void rmv_page_order(struct page *page)
443{
444    __ClearPageBuddy(page);
445    set_page_private(page, 0);
446}
447
448/*
449 * Locate the struct page for both the matching buddy in our
450 * pair (buddy1) and the combined O(n+1) page they form (page).
451 *
452 * 1) Any buddy B1 will have an order O twin B2 which satisfies
453 * the following equation:
454 * B2 = B1 ^ (1 << O)
455 * For example, if the starting buddy (buddy2) is #8 its order
456 * 1 buddy is #10:
457 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
458 *
459 * 2) Any buddy B will have an order O+1 parent P which
460 * satisfies the following equation:
461 * P = B & ~(1 << O)
462 *
463 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
464 */
465static inline unsigned long
466__find_buddy_index(unsigned long page_idx, unsigned int order)
467{
468    return page_idx ^ (1 << order);
469}
470
471/*
472 * This function checks whether a page is free && is the buddy
473 * we can do coalesce a page and its buddy if
474 * (a) the buddy is not in a hole &&
475 * (b) the buddy is in the buddy system &&
476 * (c) a page and its buddy have the same order &&
477 * (d) a page and its buddy are in the same zone.
478 *
479 * For recording whether a page is in the buddy system, we set ->_mapcount -2.
480 * Setting, clearing, and testing _mapcount -2 is serialized by zone->lock.
481 *
482 * For recording page's order, we use page_private(page).
483 */
484static inline int page_is_buddy(struct page *page, struct page *buddy,
485                                int order)
486{
487    if (!pfn_valid_within(page_to_pfn(buddy)))
488        return 0;
489
490    if (page_zone_id(page) != page_zone_id(buddy))
491        return 0;
492
493    if (page_is_guard(buddy) && page_order(buddy) == order) {
494        VM_BUG_ON(page_count(buddy) != 0);
495        return 1;
496    }
497
498    if (PageBuddy(buddy) && page_order(buddy) == order) {
499        VM_BUG_ON(page_count(buddy) != 0);
500        return 1;
501    }
502    return 0;
503}
504
505/*
506 * Freeing function for a buddy system allocator.
507 *
508 * The concept of a buddy system is to maintain direct-mapped table
509 * (containing bit values) for memory blocks of various "orders".
510 * The bottom level table contains the map for the smallest allocatable
511 * units of memory (here, pages), and each level above it describes
512 * pairs of units from the levels below, hence, "buddies".
513 * At a high level, all that happens here is marking the table entry
514 * at the bottom level available, and propagating the changes upward
515 * as necessary, plus some accounting needed to play nicely with other
516 * parts of the VM system.
517 * At each level, we keep a list of pages, which are heads of continuous
518 * free pages of length of (1 << order) and marked with _mapcount -2. Page's
519 * order is recorded in page_private(page) field.
520 * So when we are allocating or freeing one, we can derive the state of the
521 * other. That is, if we allocate a small block, and both were
522 * free, the remainder of the region must be split into blocks.
523 * If a block is freed, and its buddy is also free, then this
524 * triggers coalescing into a block of larger size.
525 *
526 * -- wli
527 */
528
529static inline void __free_one_page(struct page *page,
530        struct zone *zone, unsigned int order,
531        int migratetype)
532{
533    unsigned long page_idx;
534    unsigned long combined_idx;
535    unsigned long uninitialized_var(buddy_idx);
536    struct page *buddy;
537
538    if (unlikely(PageCompound(page)))
539        if (unlikely(destroy_compound_page(page, order)))
540            return;
541
542    VM_BUG_ON(migratetype == -1);
543
544    page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
545
546    VM_BUG_ON(page_idx & ((1 << order) - 1));
547    VM_BUG_ON(bad_range(zone, page));
548
549    while (order < MAX_ORDER-1) {
550        buddy_idx = __find_buddy_index(page_idx, order);
551        buddy = page + (buddy_idx - page_idx);
552        if (!page_is_buddy(page, buddy, order))
553            break;
554        /*
555         * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
556         * merge with it and move up one order.
557         */
558        if (page_is_guard(buddy)) {
559            clear_page_guard_flag(buddy);
560            set_page_private(page, 0);
561            __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order);
562        } else {
563            list_del(&buddy->lru);
564            zone->free_area[order].nr_free--;
565            rmv_page_order(buddy);
566        }
567        combined_idx = buddy_idx & page_idx;
568        page = page + (combined_idx - page_idx);
569        page_idx = combined_idx;
570        order++;
571    }
572    set_page_order(page, order);
573
574    /*
575     * If this is not the largest possible page, check if the buddy
576     * of the next-highest order is free. If it is, it's possible
577     * that pages are being freed that will coalesce soon. In case,
578     * that is happening, add the free page to the tail of the list
579     * so it's less likely to be used soon and more likely to be merged
580     * as a higher order page
581     */
582    if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
583        struct page *higher_page, *higher_buddy;
584        combined_idx = buddy_idx & page_idx;
585        higher_page = page + (combined_idx - page_idx);
586        buddy_idx = __find_buddy_index(combined_idx, order + 1);
587        higher_buddy = page + (buddy_idx - combined_idx);
588        if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
589            list_add_tail(&page->lru,
590                &zone->free_area[order].free_list[migratetype]);
591            goto out;
592        }
593    }
594
595    list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
596out:
597    zone->free_area[order].nr_free++;
598}
599
600/*
601 * free_page_mlock() -- clean up attempts to free and mlocked() page.
602 * Page should not be on lru, so no need to fix that up.
603 * free_pages_check() will verify...
604 */
605static inline void free_page_mlock(struct page *page)
606{
607    __dec_zone_page_state(page, NR_MLOCK);
608    __count_vm_event(UNEVICTABLE_MLOCKFREED);
609}
610
611static inline int free_pages_check(struct page *page)
612{
613    if (unlikely(page_mapcount(page) |
614        (page->mapping != NULL) |
615        (atomic_read(&page->_count) != 0) |
616        (page->flags & PAGE_FLAGS_CHECK_AT_FREE) |
617        (mem_cgroup_bad_page_check(page)))) {
618        bad_page(page);
619        return 1;
620    }
621    if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
622        page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
623    return 0;
624}
625
626/*
627 * Frees a number of pages from the PCP lists
628 * Assumes all pages on list are in same zone, and of same order.
629 * count is the number of pages to free.
630 *
631 * If the zone was previously in an "all pages pinned" state then look to
632 * see if this freeing clears that state.
633 *
634 * And clear the zone's pages_scanned counter, to hold off the "all pages are
635 * pinned" detection logic.
636 */
637static void free_pcppages_bulk(struct zone *zone, int count,
638                    struct per_cpu_pages *pcp)
639{
640    int migratetype = 0;
641    int batch_free = 0;
642    int to_free = count;
643
644    spin_lock(&zone->lock);
645    zone->all_unreclaimable = 0;
646    zone->pages_scanned = 0;
647
648    while (to_free) {
649        struct page *page;
650        struct list_head *list;
651
652        /*
653         * Remove pages from lists in a round-robin fashion. A
654         * batch_free count is maintained that is incremented when an
655         * empty list is encountered. This is so more pages are freed
656         * off fuller lists instead of spinning excessively around empty
657         * lists
658         */
659        do {
660            batch_free++;
661            if (++migratetype == MIGRATE_PCPTYPES)
662                migratetype = 0;
663            list = &pcp->lists[migratetype];
664        } while (list_empty(list));
665
666        /* This is the only non-empty list. Free them all. */
667        if (batch_free == MIGRATE_PCPTYPES)
668            batch_free = to_free;
669
670        do {
671            page = list_entry(list->prev, struct page, lru);
672            /* must delete as __free_one_page list manipulates */
673            list_del(&page->lru);
674            /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
675            __free_one_page(page, zone, 0, page_private(page));
676            trace_mm_page_pcpu_drain(page, 0, page_private(page));
677        } while (--to_free && --batch_free && !list_empty(list));
678    }
679    __mod_zone_page_state(zone, NR_FREE_PAGES, count);
680    spin_unlock(&zone->lock);
681}
682
683static void free_one_page(struct zone *zone, struct page *page, int order,
684                int migratetype)
685{
686    spin_lock(&zone->lock);
687    zone->all_unreclaimable = 0;
688    zone->pages_scanned = 0;
689
690    __free_one_page(page, zone, order, migratetype);
691    __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order);
692    spin_unlock(&zone->lock);
693}
694
695static bool free_pages_prepare(struct page *page, unsigned int order)
696{
697    int i;
698    int bad = 0;
699
700    trace_mm_page_free(page, order);
701    kmemcheck_free_shadow(page, order);
702
703    if (PageAnon(page))
704        page->mapping = NULL;
705    for (i = 0; i < (1 << order); i++)
706        bad += free_pages_check(page + i);
707    if (bad)
708        return false;
709
710    if (!PageHighMem(page)) {
711        debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
712        debug_check_no_obj_freed(page_address(page),
713                       PAGE_SIZE << order);
714    }
715    arch_free_page(page, order);
716    kernel_map_pages(page, 1 << order, 0);
717
718    return true;
719}
720
721static void __free_pages_ok(struct page *page, unsigned int order)
722{
723    unsigned long flags;
724    int wasMlocked = __TestClearPageMlocked(page);
725
726    if (!free_pages_prepare(page, order))
727        return;
728
729    local_irq_save(flags);
730    if (unlikely(wasMlocked))
731        free_page_mlock(page);
732    __count_vm_events(PGFREE, 1 << order);
733    free_one_page(page_zone(page), page, order,
734                    get_pageblock_migratetype(page));
735    local_irq_restore(flags);
736}
737
738void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
739{
740    unsigned int nr_pages = 1 << order;
741    unsigned int loop;
742
743    prefetchw(page);
744    for (loop = 0; loop < nr_pages; loop++) {
745        struct page *p = &page[loop];
746
747        if (loop + 1 < nr_pages)
748            prefetchw(p + 1);
749        __ClearPageReserved(p);
750        set_page_count(p, 0);
751    }
752
753    set_page_refcounted(page);
754    __free_pages(page, order);
755}
756
757#ifdef CONFIG_CMA
758/* Free whole pageblock and set it's migration type to MIGRATE_CMA. */
759void __init init_cma_reserved_pageblock(struct page *page)
760{
761    unsigned i = pageblock_nr_pages;
762    struct page *p = page;
763
764    do {
765        __ClearPageReserved(p);
766        set_page_count(p, 0);
767    } while (++p, --i);
768
769    set_page_refcounted(page);
770    set_pageblock_migratetype(page, MIGRATE_CMA);
771    __free_pages(page, pageblock_order);
772    totalram_pages += pageblock_nr_pages;
773}
774#endif
775
776/*
777 * The order of subdivision here is critical for the IO subsystem.
778 * Please do not alter this order without good reasons and regression
779 * testing. Specifically, as large blocks of memory are subdivided,
780 * the order in which smaller blocks are delivered depends on the order
781 * they're subdivided in this function. This is the primary factor
782 * influencing the order in which pages are delivered to the IO
783 * subsystem according to empirical testing, and this is also justified
784 * by considering the behavior of a buddy system containing a single
785 * large block of memory acted on by a series of small allocations.
786 * This behavior is a critical factor in sglist merging's success.
787 *
788 * -- wli
789 */
790static inline void expand(struct zone *zone, struct page *page,
791    int low, int high, struct free_area *area,
792    int migratetype)
793{
794    unsigned long size = 1 << high;
795
796    while (high > low) {
797        area--;
798        high--;
799        size >>= 1;
800        VM_BUG_ON(bad_range(zone, &page[size]));
801
802#ifdef CONFIG_DEBUG_PAGEALLOC
803        if (high < debug_guardpage_minorder()) {
804            /*
805             * Mark as guard pages (or page), that will allow to
806             * merge back to allocator when buddy will be freed.
807             * Corresponding page table entries will not be touched,
808             * pages will stay not present in virtual address space
809             */
810            INIT_LIST_HEAD(&page[size].lru);
811            set_page_guard_flag(&page[size]);
812            set_page_private(&page[size], high);
813            /* Guard pages are not available for any usage */
814            __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << high));
815            continue;
816        }
817#endif
818        list_add(&page[size].lru, &area->free_list[migratetype]);
819        area->nr_free++;
820        set_page_order(&page[size], high);
821    }
822}
823
824/*
825 * This page is about to be returned from the page allocator
826 */
827static inline int check_new_page(struct page *page)
828{
829    if (unlikely(page_mapcount(page) |
830        (page->mapping != NULL) |
831        (atomic_read(&page->_count) != 0) |
832        (page->flags & PAGE_FLAGS_CHECK_AT_PREP) |
833        (mem_cgroup_bad_page_check(page)))) {
834        bad_page(page);
835        return 1;
836    }
837    return 0;
838}
839
840static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
841{
842    int i;
843
844    for (i = 0; i < (1 << order); i++) {
845        struct page *p = page + i;
846        if (unlikely(check_new_page(p)))
847            return 1;
848    }
849
850    set_page_private(page, 0);
851    set_page_refcounted(page);
852
853    arch_alloc_page(page, order);
854    kernel_map_pages(page, 1 << order, 1);
855
856    if (gfp_flags & __GFP_ZERO)
857        prep_zero_page(page, order, gfp_flags);
858
859    if (order && (gfp_flags & __GFP_COMP))
860        prep_compound_page(page, order);
861
862    return 0;
863}
864
865/*
866 * Go through the free lists for the given migratetype and remove
867 * the smallest available page from the freelists
868 */
869static inline
870struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
871                        int migratetype)
872{
873    unsigned int current_order;
874    struct free_area * area;
875    struct page *page;
876
877    /* Find a page of the appropriate size in the preferred list */
878    for (current_order = order; current_order < MAX_ORDER; ++current_order) {
879        area = &(zone->free_area[current_order]);
880        if (list_empty(&area->free_list[migratetype]))
881            continue;
882
883        page = list_entry(area->free_list[migratetype].next,
884                            struct page, lru);
885        list_del(&page->lru);
886        rmv_page_order(page);
887        area->nr_free--;
888        expand(zone, page, order, current_order, area, migratetype);
889        return page;
890    }
891
892    return NULL;
893}
894
895
896/*
897 * This array describes the order lists are fallen back to when
898 * the free lists for the desirable migrate type are depleted
899 */
900static int fallbacks[MIGRATE_TYPES][4] = {
901    [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
902    [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
903#ifdef CONFIG_CMA
904    [MIGRATE_MOVABLE] = { MIGRATE_CMA, MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
905    [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */
906#else
907    [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
908#endif
909    [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */
910    [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */
911};
912
913/*
914 * Move the free pages in a range to the free lists of the requested type.
915 * Note that start_page and end_pages are not aligned on a pageblock
916 * boundary. If alignment is required, use move_freepages_block()
917 */
918static int move_freepages(struct zone *zone,
919              struct page *start_page, struct page *end_page,
920              int migratetype)
921{
922    struct page *page;
923    unsigned long order;
924    int pages_moved = 0;
925
926#ifndef CONFIG_HOLES_IN_ZONE
927    /*
928     * page_zone is not safe to call in this context when
929     * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
930     * anyway as we check zone boundaries in move_freepages_block().
931     * Remove at a later date when no bug reports exist related to
932     * grouping pages by mobility
933     */
934    BUG_ON(page_zone(start_page) != page_zone(end_page));
935#endif
936
937    for (page = start_page; page <= end_page;) {
938        /* Make sure we are not inadvertently changing nodes */
939        VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
940
941        if (!pfn_valid_within(page_to_pfn(page))) {
942            page++;
943            continue;
944        }
945
946        if (!PageBuddy(page)) {
947            page++;
948            continue;
949        }
950
951        order = page_order(page);
952        list_move(&page->lru,
953              &zone->free_area[order].free_list[migratetype]);
954        page += 1 << order;
955        pages_moved += 1 << order;
956    }
957
958    return pages_moved;
959}
960
961int move_freepages_block(struct zone *zone, struct page *page,
962                int migratetype)
963{
964    unsigned long start_pfn, end_pfn;
965    struct page *start_page, *end_page;
966
967    start_pfn = page_to_pfn(page);
968    start_pfn = start_pfn & ~(pageblock_nr_pages-1);
969    start_page = pfn_to_page(start_pfn);
970    end_page = start_page + pageblock_nr_pages - 1;
971    end_pfn = start_pfn + pageblock_nr_pages - 1;
972
973    /* Do not cross zone boundaries */
974    if (start_pfn < zone->zone_start_pfn)
975        start_page = page;
976    if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
977        return 0;
978
979    return move_freepages(zone, start_page, end_page, migratetype);
980}
981
982static void change_pageblock_range(struct page *pageblock_page,
983                    int start_order, int migratetype)
984{
985    int nr_pageblocks = 1 << (start_order - pageblock_order);
986
987    while (nr_pageblocks--) {
988        set_pageblock_migratetype(pageblock_page, migratetype);
989        pageblock_page += pageblock_nr_pages;
990    }
991}
992
993/* Remove an element from the buddy allocator from the fallback list */
994static inline struct page *
995__rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
996{
997    struct free_area * area;
998    int current_order;
999    struct page *page;
1000    int migratetype, i;
1001
1002    /* Find the largest possible block of pages in the other list */
1003    for (current_order = MAX_ORDER-1; current_order >= order;
1004                        --current_order) {
1005        for (i = 0;; i++) {
1006            migratetype = fallbacks[start_migratetype][i];
1007
1008            /* MIGRATE_RESERVE handled later if necessary */
1009            if (migratetype == MIGRATE_RESERVE)
1010                break;
1011
1012            area = &(zone->free_area[current_order]);
1013            if (list_empty(&area->free_list[migratetype]))
1014                continue;
1015
1016            page = list_entry(area->free_list[migratetype].next,
1017                    struct page, lru);
1018            area->nr_free--;
1019
1020            /*
1021             * If breaking a large block of pages, move all free
1022             * pages to the preferred allocation list. If falling
1023             * back for a reclaimable kernel allocation, be more
1024             * aggressive about taking ownership of free pages
1025             *
1026             * On the other hand, never change migration
1027             * type of MIGRATE_CMA pageblocks nor move CMA
1028             * pages on different free lists. We don't
1029             * want unmovable pages to be allocated from
1030             * MIGRATE_CMA areas.
1031             */
1032            if (!is_migrate_cma(migratetype) &&
1033                (unlikely(current_order >= pageblock_order / 2) ||
1034                 start_migratetype == MIGRATE_RECLAIMABLE ||
1035                 page_group_by_mobility_disabled)) {
1036                int pages;
1037                pages = move_freepages_block(zone, page,
1038                                start_migratetype);
1039
1040                /* Claim the whole block if over half of it is free */
1041                if (pages >= (1 << (pageblock_order-1)) ||
1042                        page_group_by_mobility_disabled)
1043                    set_pageblock_migratetype(page,
1044                                start_migratetype);
1045
1046                migratetype = start_migratetype;
1047            }
1048
1049            /* Remove the page from the freelists */
1050            list_del(&page->lru);
1051            rmv_page_order(page);
1052
1053            /* Take ownership for orders >= pageblock_order */
1054            if (current_order >= pageblock_order &&
1055                !is_migrate_cma(migratetype))
1056                change_pageblock_range(page, current_order,
1057                            start_migratetype);
1058
1059            expand(zone, page, order, current_order, area,
1060                   is_migrate_cma(migratetype)
1061                 ? migratetype : start_migratetype);
1062
1063            trace_mm_page_alloc_extfrag(page, order, current_order,
1064                start_migratetype, migratetype);
1065
1066            return page;
1067        }
1068    }
1069
1070    return NULL;
1071}
1072
1073/*
1074 * Do the hard work of removing an element from the buddy allocator.
1075 * Call me with the zone->lock already held.
1076 */
1077static struct page *__rmqueue(struct zone *zone, unsigned int order,
1078                        int migratetype)
1079{
1080    struct page *page;
1081
1082retry_reserve:
1083    page = __rmqueue_smallest(zone, order, migratetype);
1084
1085    if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1086        page = __rmqueue_fallback(zone, order, migratetype);
1087
1088        /*
1089         * Use MIGRATE_RESERVE rather than fail an allocation. goto
1090         * is used because __rmqueue_smallest is an inline function
1091         * and we want just one call site
1092         */
1093        if (!page) {
1094            migratetype = MIGRATE_RESERVE;
1095            goto retry_reserve;
1096        }
1097    }
1098
1099    trace_mm_page_alloc_zone_locked(page, order, migratetype);
1100    return page;
1101}
1102
1103/*
1104 * Obtain a specified number of elements from the buddy allocator, all under
1105 * a single hold of the lock, for efficiency. Add them to the supplied list.
1106 * Returns the number of new pages which were placed at *list.
1107 */
1108static int rmqueue_bulk(struct zone *zone, unsigned int order,
1109            unsigned long count, struct list_head *list,
1110            int migratetype, int cold)
1111{
1112    int mt = migratetype, i;
1113
1114    spin_lock(&zone->lock);
1115    for (i = 0; i < count; ++i) {
1116        struct page *page = __rmqueue(zone, order, migratetype);
1117        if (unlikely(page == NULL))
1118            break;
1119
1120        /*
1121         * Split buddy pages returned by expand() are received here
1122         * in physical page order. The page is added to the callers and
1123         * list and the list head then moves forward. From the callers
1124         * perspective, the linked list is ordered by page number in
1125         * some conditions. This is useful for IO devices that can
1126         * merge IO requests if the physical pages are ordered
1127         * properly.
1128         */
1129        if (likely(cold == 0))
1130            list_add(&page->lru, list);
1131        else
1132            list_add_tail(&page->lru, list);
1133        if (IS_ENABLED(CONFIG_CMA)) {
1134            mt = get_pageblock_migratetype(page);
1135            if (!is_migrate_cma(mt) && mt != MIGRATE_ISOLATE)
1136                mt = migratetype;
1137        }
1138        set_page_private(page, mt);
1139        list = &page->lru;
1140    }
1141    __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1142    spin_unlock(&zone->lock);
1143    return i;
1144}
1145
1146#ifdef CONFIG_NUMA
1147/*
1148 * Called from the vmstat counter updater to drain pagesets of this
1149 * currently executing processor on remote nodes after they have
1150 * expired.
1151 *
1152 * Note that this function must be called with the thread pinned to
1153 * a single processor.
1154 */
1155void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1156{
1157    unsigned long flags;
1158    int to_drain;
1159
1160    local_irq_save(flags);
1161    if (pcp->count >= pcp->batch)
1162        to_drain = pcp->batch;
1163    else
1164        to_drain = pcp->count;
1165    if (to_drain > 0) {
1166        free_pcppages_bulk(zone, to_drain, pcp);
1167        pcp->count -= to_drain;
1168    }
1169    local_irq_restore(flags);
1170}
1171#endif
1172
1173/*
1174 * Drain pages of the indicated processor.
1175 *
1176 * The processor must either be the current processor and the
1177 * thread pinned to the current processor or a processor that
1178 * is not online.
1179 */
1180static void drain_pages(unsigned int cpu)
1181{
1182    unsigned long flags;
1183    struct zone *zone;
1184
1185    for_each_populated_zone(zone) {
1186        struct per_cpu_pageset *pset;
1187        struct per_cpu_pages *pcp;
1188
1189        local_irq_save(flags);
1190        pset = per_cpu_ptr(zone->pageset, cpu);
1191
1192        pcp = &pset->pcp;
1193        if (pcp->count) {
1194            free_pcppages_bulk(zone, pcp->count, pcp);
1195            pcp->count = 0;
1196        }
1197        local_irq_restore(flags);
1198    }
1199}
1200
1201/*
1202 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1203 */
1204void drain_local_pages(void *arg)
1205{
1206    drain_pages(smp_processor_id());
1207}
1208
1209/*
1210 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1211 *
1212 * Note that this code is protected against sending an IPI to an offline
1213 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1214 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1215 * nothing keeps CPUs from showing up after we populated the cpumask and
1216 * before the call to on_each_cpu_mask().
1217 */
1218void drain_all_pages(void)
1219{
1220    int cpu;
1221    struct per_cpu_pageset *pcp;
1222    struct zone *zone;
1223
1224    /*
1225     * Allocate in the BSS so we wont require allocation in
1226     * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1227     */
1228    static cpumask_t cpus_with_pcps;
1229
1230    /*
1231     * We don't care about racing with CPU hotplug event
1232     * as offline notification will cause the notified
1233     * cpu to drain that CPU pcps and on_each_cpu_mask
1234     * disables preemption as part of its processing
1235     */
1236    for_each_online_cpu(cpu) {
1237        bool has_pcps = false;
1238        for_each_populated_zone(zone) {
1239            pcp = per_cpu_ptr(zone->pageset, cpu);
1240            if (pcp->pcp.count) {
1241                has_pcps = true;
1242                break;
1243            }
1244        }
1245        if (has_pcps)
1246            cpumask_set_cpu(cpu, &cpus_with_pcps);
1247        else
1248            cpumask_clear_cpu(cpu, &cpus_with_pcps);
1249    }
1250    on_each_cpu_mask(&cpus_with_pcps, drain_local_pages, NULL, 1);
1251}
1252
1253#ifdef CONFIG_HIBERNATION
1254
1255void mark_free_pages(struct zone *zone)
1256{
1257    unsigned long pfn, max_zone_pfn;
1258    unsigned long flags;
1259    int order, t;
1260    struct list_head *curr;
1261
1262    if (!zone->spanned_pages)
1263        return;
1264
1265    spin_lock_irqsave(&zone->lock, flags);
1266
1267    max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
1268    for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1269        if (pfn_valid(pfn)) {
1270            struct page *page = pfn_to_page(pfn);
1271
1272            if (!swsusp_page_is_forbidden(page))
1273                swsusp_unset_page_free(page);
1274        }
1275
1276    for_each_migratetype_order(order, t) {
1277        list_for_each(curr, &zone->free_area[order].free_list[t]) {
1278            unsigned long i;
1279
1280            pfn = page_to_pfn(list_entry(curr, struct page, lru));
1281            for (i = 0; i < (1UL << order); i++)
1282                swsusp_set_page_free(pfn_to_page(pfn + i));
1283        }
1284    }
1285    spin_unlock_irqrestore(&zone->lock, flags);
1286}
1287#endif /* CONFIG_PM */
1288
1289/*
1290 * Free a 0-order page
1291 * cold == 1 ? free a cold page : free a hot page
1292 */
1293void free_hot_cold_page(struct page *page, int cold)
1294{
1295    struct zone *zone = page_zone(page);
1296    struct per_cpu_pages *pcp;
1297    unsigned long flags;
1298    int migratetype;
1299    int wasMlocked = __TestClearPageMlocked(page);
1300
1301    if (!free_pages_prepare(page, 0))
1302        return;
1303
1304    migratetype = get_pageblock_migratetype(page);
1305    set_page_private(page, migratetype);
1306    local_irq_save(flags);
1307    if (unlikely(wasMlocked))
1308        free_page_mlock(page);
1309    __count_vm_event(PGFREE);
1310
1311    /*
1312     * We only track unmovable, reclaimable and movable on pcp lists.
1313     * Free ISOLATE pages back to the allocator because they are being
1314     * offlined but treat RESERVE as movable pages so we can get those
1315     * areas back if necessary. Otherwise, we may have to free
1316     * excessively into the page allocator
1317     */
1318    if (migratetype >= MIGRATE_PCPTYPES) {
1319        if (unlikely(migratetype == MIGRATE_ISOLATE)) {
1320            free_one_page(zone, page, 0, migratetype);
1321            goto out;
1322        }
1323        migratetype = MIGRATE_MOVABLE;
1324    }
1325
1326    pcp = &this_cpu_ptr(zone->pageset)->pcp;
1327    if (cold)
1328        list_add_tail(&page->lru, &pcp->lists[migratetype]);
1329    else
1330        list_add(&page->lru, &pcp->lists[migratetype]);
1331    pcp->count++;
1332    if (pcp->count >= pcp->high) {
1333        free_pcppages_bulk(zone, pcp->batch, pcp);
1334        pcp->count -= pcp->batch;
1335    }
1336
1337out:
1338    local_irq_restore(flags);
1339}
1340
1341/*
1342 * Free a list of 0-order pages
1343 */
1344void free_hot_cold_page_list(struct list_head *list, int cold)
1345{
1346    struct page *page, *next;
1347
1348    list_for_each_entry_safe(page, next, list, lru) {
1349        trace_mm_page_free_batched(page, cold);
1350        free_hot_cold_page(page, cold);
1351    }
1352}
1353
1354/*
1355 * split_page takes a non-compound higher-order page, and splits it into
1356 * n (1<<order) sub-pages: page[0..n]
1357 * Each sub-page must be freed individually.
1358 *
1359 * Note: this is probably too low level an operation for use in drivers.
1360 * Please consult with lkml before using this in your driver.
1361 */
1362void split_page(struct page *page, unsigned int order)
1363{
1364    int i;
1365
1366    VM_BUG_ON(PageCompound(page));
1367    VM_BUG_ON(!page_count(page));
1368
1369#ifdef CONFIG_KMEMCHECK
1370    /*
1371     * Split shadow pages too, because free(page[0]) would
1372     * otherwise free the whole shadow.
1373     */
1374    if (kmemcheck_page_is_tracked(page))
1375        split_page(virt_to_page(page[0].shadow), order);
1376#endif
1377
1378    for (i = 1; i < (1 << order); i++)
1379        set_page_refcounted(page + i);
1380}
1381
1382/*
1383 * Similar to split_page except the page is already free. As this is only
1384 * being used for migration, the migratetype of the block also changes.
1385 * As this is called with interrupts disabled, the caller is responsible
1386 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1387 * are enabled.
1388 *
1389 * Note: this is probably too low level an operation for use in drivers.
1390 * Please consult with lkml before using this in your driver.
1391 */
1392int split_free_page(struct page *page)
1393{
1394    unsigned int order;
1395    unsigned long watermark;
1396    struct zone *zone;
1397
1398    BUG_ON(!PageBuddy(page));
1399
1400    zone = page_zone(page);
1401    order = page_order(page);
1402
1403    /* Obey watermarks as if the page was being allocated */
1404    watermark = low_wmark_pages(zone) + (1 << order);
1405    if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1406        return 0;
1407
1408    /* Remove page from free list */
1409    list_del(&page->lru);
1410    zone->free_area[order].nr_free--;
1411    rmv_page_order(page);
1412    __mod_zone_page_state(zone, NR_FREE_PAGES, -(1UL << order));
1413
1414    /* Split into individual pages */
1415    set_page_refcounted(page);
1416    split_page(page, order);
1417
1418    if (order >= pageblock_order - 1) {
1419        struct page *endpage = page + (1 << order) - 1;
1420        for (; page < endpage; page += pageblock_nr_pages) {
1421            int mt = get_pageblock_migratetype(page);
1422            if (mt != MIGRATE_ISOLATE && !is_migrate_cma(mt))
1423                set_pageblock_migratetype(page,
1424                              MIGRATE_MOVABLE);
1425        }
1426    }
1427
1428    return 1 << order;
1429}
1430
1431/*
1432 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1433 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1434 * or two.
1435 */
1436static inline
1437struct page *buffered_rmqueue(struct zone *preferred_zone,
1438            struct zone *zone, int order, gfp_t gfp_flags,
1439            int migratetype)
1440{
1441    unsigned long flags;
1442    struct page *page;
1443    int cold = !!(gfp_flags & __GFP_COLD);
1444
1445again:
1446    if (likely(order == 0)) {
1447        struct per_cpu_pages *pcp;
1448        struct list_head *list;
1449
1450        local_irq_save(flags);
1451        pcp = &this_cpu_ptr(zone->pageset)->pcp;
1452        list = &pcp->lists[migratetype];
1453        if (list_empty(list)) {
1454            pcp->count += rmqueue_bulk(zone, 0,
1455                    pcp->batch, list,
1456                    migratetype, cold);
1457            if (unlikely(list_empty(list)))
1458                goto failed;
1459        }
1460
1461        if (cold)
1462            page = list_entry(list->prev, struct page, lru);
1463        else
1464            page = list_entry(list->next, struct page, lru);
1465
1466        list_del(&page->lru);
1467        pcp->count--;
1468    } else {
1469        if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1470            /*
1471             * __GFP_NOFAIL is not to be used in new code.
1472             *
1473             * All __GFP_NOFAIL callers should be fixed so that they
1474             * properly detect and handle allocation failures.
1475             *
1476             * We most definitely don't want callers attempting to
1477             * allocate greater than order-1 page units with
1478             * __GFP_NOFAIL.
1479             */
1480            WARN_ON_ONCE(order > 1);
1481        }
1482        spin_lock_irqsave(&zone->lock, flags);
1483        page = __rmqueue(zone, order, migratetype);
1484        spin_unlock(&zone->lock);
1485        if (!page)
1486            goto failed;
1487        __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << order));
1488    }
1489
1490    __count_zone_vm_events(PGALLOC, zone, 1 << order);
1491    zone_statistics(preferred_zone, zone, gfp_flags);
1492    local_irq_restore(flags);
1493
1494    VM_BUG_ON(bad_range(zone, page));
1495    if (prep_new_page(page, order, gfp_flags))
1496        goto again;
1497    return page;
1498
1499failed:
1500    local_irq_restore(flags);
1501    return NULL;
1502}
1503
1504/* The ALLOC_WMARK bits are used as an index to zone->watermark */
1505#define ALLOC_WMARK_MIN WMARK_MIN
1506#define ALLOC_WMARK_LOW WMARK_LOW
1507#define ALLOC_WMARK_HIGH WMARK_HIGH
1508#define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */
1509
1510/* Mask to get the watermark bits */
1511#define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1)
1512
1513#define ALLOC_HARDER 0x10 /* try to alloc harder */
1514#define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
1515#define ALLOC_CPUSET 0x40 /* check for correct cpuset */
1516
1517#ifdef CONFIG_FAIL_PAGE_ALLOC
1518
1519static struct {
1520    struct fault_attr attr;
1521
1522    u32 ignore_gfp_highmem;
1523    u32 ignore_gfp_wait;
1524    u32 min_order;
1525} fail_page_alloc = {
1526    .attr = FAULT_ATTR_INITIALIZER,
1527    .ignore_gfp_wait = 1,
1528    .ignore_gfp_highmem = 1,
1529    .min_order = 1,
1530};
1531
1532static int __init setup_fail_page_alloc(char *str)
1533{
1534    return setup_fault_attr(&fail_page_alloc.attr, str);
1535}
1536__setup("fail_page_alloc=", setup_fail_page_alloc);
1537
1538static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1539{
1540    if (order < fail_page_alloc.min_order)
1541        return false;
1542    if (gfp_mask & __GFP_NOFAIL)
1543        return false;
1544    if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1545        return false;
1546    if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1547        return false;
1548
1549    return should_fail(&fail_page_alloc.attr, 1 << order);
1550}
1551
1552#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1553
1554static int __init fail_page_alloc_debugfs(void)
1555{
1556    umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1557    struct dentry *dir;
1558
1559    dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1560                    &fail_page_alloc.attr);
1561    if (IS_ERR(dir))
1562        return PTR_ERR(dir);
1563
1564    if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1565                &fail_page_alloc.ignore_gfp_wait))
1566        goto fail;
1567    if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1568                &fail_page_alloc.ignore_gfp_highmem))
1569        goto fail;
1570    if (!debugfs_create_u32("min-order", mode, dir,
1571                &fail_page_alloc.min_order))
1572        goto fail;
1573
1574    return 0;
1575fail:
1576    debugfs_remove_recursive(dir);
1577
1578    return -ENOMEM;
1579}
1580
1581late_initcall(fail_page_alloc_debugfs);
1582
1583#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1584
1585#else /* CONFIG_FAIL_PAGE_ALLOC */
1586
1587static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1588{
1589    return false;
1590}
1591
1592#endif /* CONFIG_FAIL_PAGE_ALLOC */
1593
1594/*
1595 * Return true if free pages are above 'mark'. This takes into account the order
1596 * of the allocation.
1597 */
1598static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1599              int classzone_idx, int alloc_flags, long free_pages)
1600{
1601    /* free_pages my go negative - that's OK */
1602    long min = mark;
1603    long lowmem_reserve = z->lowmem_reserve[classzone_idx];
1604    int o;
1605
1606    free_pages -= (1 << order) - 1;
1607    if (alloc_flags & ALLOC_HIGH)
1608        min -= min / 2;
1609    if (alloc_flags & ALLOC_HARDER)
1610        min -= min / 4;
1611
1612    if (free_pages <= min + lowmem_reserve)
1613        return false;
1614    for (o = 0; o < order; o++) {
1615        /* At the next order, this order's pages become unavailable */
1616        free_pages -= z->free_area[o].nr_free << o;
1617
1618        /* Require fewer higher order pages to be free */
1619        min >>= 1;
1620
1621        if (free_pages <= min)
1622            return false;
1623    }
1624    return true;
1625}
1626
1627#ifdef CONFIG_MEMORY_ISOLATION
1628static inline unsigned long nr_zone_isolate_freepages(struct zone *zone)
1629{
1630    if (unlikely(zone->nr_pageblock_isolate))
1631        return zone->nr_pageblock_isolate * pageblock_nr_pages;
1632    return 0;
1633}
1634#else
1635static inline unsigned long nr_zone_isolate_freepages(struct zone *zone)
1636{
1637    return 0;
1638}
1639#endif
1640
1641bool zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1642              int classzone_idx, int alloc_flags)
1643{
1644    return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1645                    zone_page_state(z, NR_FREE_PAGES));
1646}
1647
1648bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark,
1649              int classzone_idx, int alloc_flags)
1650{
1651    long free_pages = zone_page_state(z, NR_FREE_PAGES);
1652
1653    if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1654        free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1655
1656    /*
1657     * If the zone has MIGRATE_ISOLATE type free pages, we should consider
1658     * it. nr_zone_isolate_freepages is never accurate so kswapd might not
1659     * sleep although it could do so. But this is more desirable for memory
1660     * hotplug than sleeping which can cause a livelock in the direct
1661     * reclaim path.
1662     */
1663    free_pages -= nr_zone_isolate_freepages(z);
1664    return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1665                                free_pages);
1666}
1667
1668#ifdef CONFIG_NUMA
1669/*
1670 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1671 * skip over zones that are not allowed by the cpuset, or that have
1672 * been recently (in last second) found to be nearly full. See further
1673 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1674 * that have to skip over a lot of full or unallowed zones.
1675 *
1676 * If the zonelist cache is present in the passed in zonelist, then
1677 * returns a pointer to the allowed node mask (either the current
1678 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1679 *
1680 * If the zonelist cache is not available for this zonelist, does
1681 * nothing and returns NULL.
1682 *
1683 * If the fullzones BITMAP in the zonelist cache is stale (more than
1684 * a second since last zap'd) then we zap it out (clear its bits.)
1685 *
1686 * We hold off even calling zlc_setup, until after we've checked the
1687 * first zone in the zonelist, on the theory that most allocations will
1688 * be satisfied from that first zone, so best to examine that zone as
1689 * quickly as we can.
1690 */
1691static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1692{
1693    struct zonelist_cache *zlc; /* cached zonelist speedup info */
1694    nodemask_t *allowednodes; /* zonelist_cache approximation */
1695
1696    zlc = zonelist->zlcache_ptr;
1697    if (!zlc)
1698        return NULL;
1699
1700    if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1701        bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1702        zlc->last_full_zap = jiffies;
1703    }
1704
1705    allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1706                    &cpuset_current_mems_allowed :
1707                    &node_states[N_HIGH_MEMORY];
1708    return allowednodes;
1709}
1710
1711/*
1712 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1713 * if it is worth looking at further for free memory:
1714 * 1) Check that the zone isn't thought to be full (doesn't have its
1715 * bit set in the zonelist_cache fullzones BITMAP).
1716 * 2) Check that the zones node (obtained from the zonelist_cache
1717 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1718 * Return true (non-zero) if zone is worth looking at further, or
1719 * else return false (zero) if it is not.
1720 *
1721 * This check -ignores- the distinction between various watermarks,
1722 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1723 * found to be full for any variation of these watermarks, it will
1724 * be considered full for up to one second by all requests, unless
1725 * we are so low on memory on all allowed nodes that we are forced
1726 * into the second scan of the zonelist.
1727 *
1728 * In the second scan we ignore this zonelist cache and exactly
1729 * apply the watermarks to all zones, even it is slower to do so.
1730 * We are low on memory in the second scan, and should leave no stone
1731 * unturned looking for a free page.
1732 */
1733static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1734                        nodemask_t *allowednodes)
1735{
1736    struct zonelist_cache *zlc; /* cached zonelist speedup info */
1737    int i; /* index of *z in zonelist zones */
1738    int n; /* node that zone *z is on */
1739
1740    zlc = zonelist->zlcache_ptr;
1741    if (!zlc)
1742        return 1;
1743
1744    i = z - zonelist->_zonerefs;
1745    n = zlc->z_to_n[i];
1746
1747    /* This zone is worth trying if it is allowed but not full */
1748    return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1749}
1750
1751/*
1752 * Given 'z' scanning a zonelist, set the corresponding bit in
1753 * zlc->fullzones, so that subsequent attempts to allocate a page
1754 * from that zone don't waste time re-examining it.
1755 */
1756static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1757{
1758    struct zonelist_cache *zlc; /* cached zonelist speedup info */
1759    int i; /* index of *z in zonelist zones */
1760
1761    zlc = zonelist->zlcache_ptr;
1762    if (!zlc)
1763        return;
1764
1765    i = z - zonelist->_zonerefs;
1766
1767    set_bit(i, zlc->fullzones);
1768}
1769
1770/*
1771 * clear all zones full, called after direct reclaim makes progress so that
1772 * a zone that was recently full is not skipped over for up to a second
1773 */
1774static void zlc_clear_zones_full(struct zonelist *zonelist)
1775{
1776    struct zonelist_cache *zlc; /* cached zonelist speedup info */
1777
1778    zlc = zonelist->zlcache_ptr;
1779    if (!zlc)
1780        return;
1781
1782    bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1783}
1784
1785#else /* CONFIG_NUMA */
1786
1787static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1788{
1789    return NULL;
1790}
1791
1792static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1793                nodemask_t *allowednodes)
1794{
1795    return 1;
1796}
1797
1798static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1799{
1800}
1801
1802static void zlc_clear_zones_full(struct zonelist *zonelist)
1803{
1804}
1805#endif /* CONFIG_NUMA */
1806
1807/*
1808 * get_page_from_freelist goes through the zonelist trying to allocate
1809 * a page.
1810 */
1811static struct page *
1812get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1813        struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1814        struct zone *preferred_zone, int migratetype)
1815{
1816    struct zoneref *z;
1817    struct page *page = NULL;
1818    int classzone_idx;
1819    struct zone *zone;
1820    nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1821    int zlc_active = 0; /* set if using zonelist_cache */
1822    int did_zlc_setup = 0; /* just call zlc_setup() one time */
1823
1824    classzone_idx = zone_idx(preferred_zone);
1825zonelist_scan:
1826    /*
1827     * Scan zonelist, looking for a zone with enough free.
1828     * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1829     */
1830    for_each_zone_zonelist_nodemask(zone, z, zonelist,
1831                        high_zoneidx, nodemask) {
1832        if (NUMA_BUILD && zlc_active &&
1833            !zlc_zone_worth_trying(zonelist, z, allowednodes))
1834                continue;
1835        if ((alloc_flags & ALLOC_CPUSET) &&
1836            !cpuset_zone_allowed_softwall(zone, gfp_mask))
1837                continue;
1838        /*
1839         * When allocating a page cache page for writing, we
1840         * want to get it from a zone that is within its dirty
1841         * limit, such that no single zone holds more than its
1842         * proportional share of globally allowed dirty pages.
1843         * The dirty limits take into account the zone's
1844         * lowmem reserves and high watermark so that kswapd
1845         * should be able to balance it without having to
1846         * write pages from its LRU list.
1847         *
1848         * This may look like it could increase pressure on
1849         * lower zones by failing allocations in higher zones
1850         * before they are full. But the pages that do spill
1851         * over are limited as the lower zones are protected
1852         * by this very same mechanism. It should not become
1853         * a practical burden to them.
1854         *
1855         * XXX: For now, allow allocations to potentially
1856         * exceed the per-zone dirty limit in the slowpath
1857         * (ALLOC_WMARK_LOW unset) before going into reclaim,
1858         * which is important when on a NUMA setup the allowed
1859         * zones are together not big enough to reach the
1860         * global limit. The proper fix for these situations
1861         * will require awareness of zones in the
1862         * dirty-throttling and the flusher threads.
1863         */
1864        if ((alloc_flags & ALLOC_WMARK_LOW) &&
1865            (gfp_mask & __GFP_WRITE) && !zone_dirty_ok(zone))
1866            goto this_zone_full;
1867
1868        BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1869        if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1870            unsigned long mark;
1871            int ret;
1872
1873            mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1874            if (zone_watermark_ok(zone, order, mark,
1875                    classzone_idx, alloc_flags))
1876                goto try_this_zone;
1877
1878            if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) {
1879                /*
1880                 * we do zlc_setup if there are multiple nodes
1881                 * and before considering the first zone allowed
1882                 * by the cpuset.
1883                 */
1884                allowednodes = zlc_setup(zonelist, alloc_flags);
1885                zlc_active = 1;
1886                did_zlc_setup = 1;
1887            }
1888
1889            if (zone_reclaim_mode == 0)
1890                goto this_zone_full;
1891
1892            /*
1893             * As we may have just activated ZLC, check if the first
1894             * eligible zone has failed zone_reclaim recently.
1895             */
1896            if (NUMA_BUILD && zlc_active &&
1897                !zlc_zone_worth_trying(zonelist, z, allowednodes))
1898                continue;
1899
1900            ret = zone_reclaim(zone, gfp_mask, order);
1901            switch (ret) {
1902            case ZONE_RECLAIM_NOSCAN:
1903                /* did not scan */
1904                continue;
1905            case ZONE_RECLAIM_FULL:
1906                /* scanned but unreclaimable */
1907                continue;
1908            default:
1909                /* did we reclaim enough */
1910                if (!zone_watermark_ok(zone, order, mark,
1911                        classzone_idx, alloc_flags))
1912                    goto this_zone_full;
1913            }
1914        }
1915
1916try_this_zone:
1917        page = buffered_rmqueue(preferred_zone, zone, order,
1918                        gfp_mask, migratetype);
1919        if (page)
1920            break;
1921this_zone_full:
1922        if (NUMA_BUILD)
1923            zlc_mark_zone_full(zonelist, z);
1924    }
1925
1926    if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) {
1927        /* Disable zlc cache for second zonelist scan */
1928        zlc_active = 0;
1929        goto zonelist_scan;
1930    }
1931    return page;
1932}
1933
1934/*
1935 * Large machines with many possible nodes should not always dump per-node
1936 * meminfo in irq context.
1937 */
1938static inline bool should_suppress_show_mem(void)
1939{
1940    bool ret = false;
1941
1942#if NODES_SHIFT > 8
1943    ret = in_interrupt();
1944#endif
1945    return ret;
1946}
1947
1948static DEFINE_RATELIMIT_STATE(nopage_rs,
1949        DEFAULT_RATELIMIT_INTERVAL,
1950        DEFAULT_RATELIMIT_BURST);
1951
1952void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
1953{
1954    unsigned int filter = SHOW_MEM_FILTER_NODES;
1955
1956    if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
1957        debug_guardpage_minorder() > 0)
1958        return;
1959
1960    /*
1961     * This documents exceptions given to allocations in certain
1962     * contexts that are allowed to allocate outside current's set
1963     * of allowed nodes.
1964     */
1965    if (!(gfp_mask & __GFP_NOMEMALLOC))
1966        if (test_thread_flag(TIF_MEMDIE) ||
1967            (current->flags & (PF_MEMALLOC | PF_EXITING)))
1968            filter &= ~SHOW_MEM_FILTER_NODES;
1969    if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
1970        filter &= ~SHOW_MEM_FILTER_NODES;
1971
1972    if (fmt) {
1973        struct va_format vaf;
1974        va_list args;
1975
1976        va_start(args, fmt);
1977
1978        vaf.fmt = fmt;
1979        vaf.va = &args;
1980
1981        pr_warn("%pV", &vaf);
1982
1983        va_end(args);
1984    }
1985
1986    pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
1987        current->comm, order, gfp_mask);
1988
1989    dump_stack();
1990    if (!should_suppress_show_mem())
1991        show_mem(filter);
1992}
1993
1994static inline int
1995should_alloc_retry(gfp_t gfp_mask, unsigned int order,
1996                unsigned long did_some_progress,
1997                unsigned long pages_reclaimed)
1998{
1999    /* Do not loop if specifically requested */
2000    if (gfp_mask & __GFP_NORETRY)
2001        return 0;
2002
2003    /* Always retry if specifically requested */
2004    if (gfp_mask & __GFP_NOFAIL)
2005        return 1;
2006
2007    /*
2008     * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2009     * making forward progress without invoking OOM. Suspend also disables
2010     * storage devices so kswapd will not help. Bail if we are suspending.
2011     */
2012    if (!did_some_progress && pm_suspended_storage())
2013        return 0;
2014
2015    /*
2016     * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2017     * means __GFP_NOFAIL, but that may not be true in other
2018     * implementations.
2019     */
2020    if (order <= PAGE_ALLOC_COSTLY_ORDER)
2021        return 1;
2022
2023    /*
2024     * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2025     * specified, then we retry until we no longer reclaim any pages
2026     * (above), or we've reclaimed an order of pages at least as
2027     * large as the allocation's order. In both cases, if the
2028     * allocation still fails, we stop retrying.
2029     */
2030    if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
2031        return 1;
2032
2033    return 0;
2034}
2035
2036static inline struct page *
2037__alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2038    struct zonelist *zonelist, enum zone_type high_zoneidx,
2039    nodemask_t *nodemask, struct zone *preferred_zone,
2040    int migratetype)
2041{
2042    struct page *page;
2043
2044    /* Acquire the OOM killer lock for the zones in zonelist */
2045    if (!try_set_zonelist_oom(zonelist, gfp_mask)) {
2046        schedule_timeout_uninterruptible(1);
2047        return NULL;
2048    }
2049
2050    /*
2051     * Go through the zonelist yet one more time, keep very high watermark
2052     * here, this is only to catch a parallel oom killing, we must fail if
2053     * we're still under heavy pressure.
2054     */
2055    page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
2056        order, zonelist, high_zoneidx,
2057        ALLOC_WMARK_HIGH|ALLOC_CPUSET,
2058        preferred_zone, migratetype);
2059    if (page)
2060        goto out;
2061
2062    if (!(gfp_mask & __GFP_NOFAIL)) {
2063        /* The OOM killer will not help higher order allocs */
2064        if (order > PAGE_ALLOC_COSTLY_ORDER)
2065            goto out;
2066        /* The OOM killer does not needlessly kill tasks for lowmem */
2067        if (high_zoneidx < ZONE_NORMAL)
2068            goto out;
2069        /*
2070         * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2071         * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2072         * The caller should handle page allocation failure by itself if
2073         * it specifies __GFP_THISNODE.
2074         * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2075         */
2076        if (gfp_mask & __GFP_THISNODE)
2077            goto out;
2078    }
2079    /* Exhausted what can be done so it's blamo time */
2080    out_of_memory(zonelist, gfp_mask, order, nodemask, false);
2081
2082out:
2083    clear_zonelist_oom(zonelist, gfp_mask);
2084    return page;
2085}
2086
2087#ifdef CONFIG_COMPACTION
2088/* Try memory compaction for high-order allocations before reclaim */
2089static struct page *
2090__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2091    struct zonelist *zonelist, enum zone_type high_zoneidx,
2092    nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2093    int migratetype, bool sync_migration,
2094    bool *deferred_compaction,
2095    unsigned long *did_some_progress)
2096{
2097    struct page *page;
2098
2099    if (!order)
2100        return NULL;
2101
2102    if (compaction_deferred(preferred_zone, order)) {
2103        *deferred_compaction = true;
2104        return NULL;
2105    }
2106
2107    current->flags |= PF_MEMALLOC;
2108    *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
2109                        nodemask, sync_migration);
2110    current->flags &= ~PF_MEMALLOC;
2111    if (*did_some_progress != COMPACT_SKIPPED) {
2112
2113        /* Page migration frees to the PCP lists but we want merging */
2114        drain_pages(get_cpu());
2115        put_cpu();
2116
2117        page = get_page_from_freelist(gfp_mask, nodemask,
2118                order, zonelist, high_zoneidx,
2119                alloc_flags & ~ALLOC_NO_WATERMARKS,
2120                preferred_zone, migratetype);
2121        if (page) {
2122            preferred_zone->compact_considered = 0;
2123            preferred_zone->compact_defer_shift = 0;
2124            if (order >= preferred_zone->compact_order_failed)
2125                preferred_zone->compact_order_failed = order + 1;
2126            count_vm_event(COMPACTSUCCESS);
2127            return page;
2128        }
2129
2130        /*
2131         * It's bad if compaction run occurs and fails.
2132         * The most likely reason is that pages exist,
2133         * but not enough to satisfy watermarks.
2134         */
2135        count_vm_event(COMPACTFAIL);
2136
2137        /*
2138         * As async compaction considers a subset of pageblocks, only
2139         * defer if the failure was a sync compaction failure.
2140         */
2141        if (sync_migration)
2142            defer_compaction(preferred_zone, order);
2143
2144        cond_resched();
2145    }
2146
2147    return NULL;
2148}
2149#else
2150static inline struct page *
2151__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2152    struct zonelist *zonelist, enum zone_type high_zoneidx,
2153    nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2154    int migratetype, bool sync_migration,
2155    bool *deferred_compaction,
2156    unsigned long *did_some_progress)
2157{
2158    return NULL;
2159}
2160#endif /* CONFIG_COMPACTION */
2161
2162/* Perform direct synchronous page reclaim */
2163static int
2164__perform_reclaim(gfp_t gfp_mask, unsigned int order, struct zonelist *zonelist,
2165          nodemask_t *nodemask)
2166{
2167    struct reclaim_state reclaim_state;
2168    int progress;
2169
2170    cond_resched();
2171
2172    /* We now go into synchronous reclaim */
2173    cpuset_memory_pressure_bump();
2174    current->flags |= PF_MEMALLOC;
2175    lockdep_set_current_reclaim_state(gfp_mask);
2176    reclaim_state.reclaimed_slab = 0;
2177    current->reclaim_state = &reclaim_state;
2178
2179    progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
2180
2181    current->reclaim_state = NULL;
2182    lockdep_clear_current_reclaim_state();
2183    current->flags &= ~PF_MEMALLOC;
2184
2185    cond_resched();
2186
2187    return progress;
2188}
2189
2190/* The really slow allocator path where we enter direct reclaim */
2191static inline struct page *
2192__alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2193    struct zonelist *zonelist, enum zone_type high_zoneidx,
2194    nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2195    int migratetype, unsigned long *did_some_progress)
2196{
2197    struct page *page = NULL;
2198    bool drained = false;
2199
2200    *did_some_progress = __perform_reclaim(gfp_mask, order, zonelist,
2201                           nodemask);
2202    if (unlikely(!(*did_some_progress)))
2203        return NULL;
2204
2205    /* After successful reclaim, reconsider all zones for allocation */
2206    if (NUMA_BUILD)
2207        zlc_clear_zones_full(zonelist);
2208
2209retry:
2210    page = get_page_from_freelist(gfp_mask, nodemask, order,
2211                    zonelist, high_zoneidx,
2212                    alloc_flags & ~ALLOC_NO_WATERMARKS,
2213                    preferred_zone, migratetype);
2214
2215    /*
2216     * If an allocation failed after direct reclaim, it could be because
2217     * pages are pinned on the per-cpu lists. Drain them and try again
2218     */
2219    if (!page && !drained) {
2220        drain_all_pages();
2221        drained = true;
2222        goto retry;
2223    }
2224
2225    return page;
2226}
2227
2228/*
2229 * This is called in the allocator slow-path if the allocation request is of
2230 * sufficient urgency to ignore watermarks and take other desperate measures
2231 */
2232static inline struct page *
2233__alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2234    struct zonelist *zonelist, enum zone_type high_zoneidx,
2235    nodemask_t *nodemask, struct zone *preferred_zone,
2236    int migratetype)
2237{
2238    struct page *page;
2239
2240    do {
2241        page = get_page_from_freelist(gfp_mask, nodemask, order,
2242            zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
2243            preferred_zone, migratetype);
2244
2245        if (!page && gfp_mask & __GFP_NOFAIL)
2246            wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2247    } while (!page && (gfp_mask & __GFP_NOFAIL));
2248
2249    return page;
2250}
2251
2252static inline
2253void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
2254                        enum zone_type high_zoneidx,
2255                        enum zone_type classzone_idx)
2256{
2257    struct zoneref *z;
2258    struct zone *zone;
2259
2260    for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
2261        wakeup_kswapd(zone, order, classzone_idx);
2262}
2263
2264static inline int
2265gfp_to_alloc_flags(gfp_t gfp_mask)
2266{
2267    int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2268    const gfp_t wait = gfp_mask & __GFP_WAIT;
2269
2270    /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2271    BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2272
2273    /*
2274     * The caller may dip into page reserves a bit more if the caller
2275     * cannot run direct reclaim, or if the caller has realtime scheduling
2276     * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2277     * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2278     */
2279    alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2280
2281    if (!wait) {
2282        /*
2283         * Not worth trying to allocate harder for
2284         * __GFP_NOMEMALLOC even if it can't schedule.
2285         */
2286        if (!(gfp_mask & __GFP_NOMEMALLOC))
2287            alloc_flags |= ALLOC_HARDER;
2288        /*
2289         * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2290         * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
2291         */
2292        alloc_flags &= ~ALLOC_CPUSET;
2293    } else if (unlikely(rt_task(current)) && !in_interrupt())
2294        alloc_flags |= ALLOC_HARDER;
2295
2296    if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2297        if (gfp_mask & __GFP_MEMALLOC)
2298            alloc_flags |= ALLOC_NO_WATERMARKS;
2299        else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2300            alloc_flags |= ALLOC_NO_WATERMARKS;
2301        else if (!in_interrupt() &&
2302                ((current->flags & PF_MEMALLOC) ||
2303                 unlikely(test_thread_flag(TIF_MEMDIE))))
2304            alloc_flags |= ALLOC_NO_WATERMARKS;
2305    }
2306
2307    return alloc_flags;
2308}
2309
2310bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2311{
2312    return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2313}
2314
2315static inline struct page *
2316__alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2317    struct zonelist *zonelist, enum zone_type high_zoneidx,
2318    nodemask_t *nodemask, struct zone *preferred_zone,
2319    int migratetype)
2320{
2321    const gfp_t wait = gfp_mask & __GFP_WAIT;
2322    struct page *page = NULL;
2323    int alloc_flags;
2324    unsigned long pages_reclaimed = 0;
2325    unsigned long did_some_progress;
2326    bool sync_migration = false;
2327    bool deferred_compaction = false;
2328
2329    /*
2330     * In the slowpath, we sanity check order to avoid ever trying to
2331     * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2332     * be using allocators in order of preference for an area that is
2333     * too large.
2334     */
2335    if (order >= MAX_ORDER) {
2336        WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2337        return NULL;
2338    }
2339
2340    /*
2341     * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2342     * __GFP_NOWARN set) should not cause reclaim since the subsystem
2343     * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2344     * using a larger set of nodes after it has established that the
2345     * allowed per node queues are empty and that nodes are
2346     * over allocated.
2347     */
2348    if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2349        goto nopage;
2350
2351restart:
2352    if (!(gfp_mask & __GFP_NO_KSWAPD))
2353        wake_all_kswapd(order, zonelist, high_zoneidx,
2354                        zone_idx(preferred_zone));
2355
2356    /*
2357     * OK, we're below the kswapd watermark and have kicked background
2358     * reclaim. Now things get more complex, so set up alloc_flags according
2359     * to how we want to proceed.
2360     */
2361    alloc_flags = gfp_to_alloc_flags(gfp_mask);
2362
2363    /*
2364     * Find the true preferred zone if the allocation is unconstrained by
2365     * cpusets.
2366     */
2367    if (!(alloc_flags & ALLOC_CPUSET) && !nodemask)
2368        first_zones_zonelist(zonelist, high_zoneidx, NULL,
2369                    &preferred_zone);
2370
2371rebalance:
2372    /* This is the last chance, in general, before the goto nopage. */
2373    page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2374            high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2375            preferred_zone, migratetype);
2376    if (page)
2377        goto got_pg;
2378
2379    /* Allocate without watermarks if the context allows */
2380    if (alloc_flags & ALLOC_NO_WATERMARKS) {
2381        /*
2382         * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2383         * the allocation is high priority and these type of
2384         * allocations are system rather than user orientated
2385         */
2386        zonelist = node_zonelist(numa_node_id(), gfp_mask);
2387
2388        page = __alloc_pages_high_priority(gfp_mask, order,
2389                zonelist, high_zoneidx, nodemask,
2390                preferred_zone, migratetype);
2391        if (page) {
2392            /*
2393             * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was
2394             * necessary to allocate the page. The expectation is
2395             * that the caller is taking steps that will free more
2396             * memory. The caller should avoid the page being used
2397             * for !PFMEMALLOC purposes.
2398             */
2399            page->pfmemalloc = true;
2400            goto got_pg;
2401        }
2402    }
2403
2404    /* Atomic allocations - we can't balance anything */
2405    if (!wait)
2406        goto nopage;
2407
2408    /* Avoid recursion of direct reclaim */
2409    if (current->flags & PF_MEMALLOC)
2410        goto nopage;
2411
2412    /* Avoid allocations with no watermarks from looping endlessly */
2413    if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2414        goto nopage;
2415
2416    /*
2417     * Try direct compaction. The first pass is asynchronous. Subsequent
2418     * attempts after direct reclaim are synchronous
2419     */
2420    page = __alloc_pages_direct_compact(gfp_mask, order,
2421                    zonelist, high_zoneidx,
2422                    nodemask,
2423                    alloc_flags, preferred_zone,
2424                    migratetype, sync_migration,
2425                    &deferred_compaction,
2426                    &did_some_progress);
2427    if (page)
2428        goto got_pg;
2429    sync_migration = true;
2430
2431    /*
2432     * If compaction is deferred for high-order allocations, it is because
2433     * sync compaction recently failed. In this is the case and the caller
2434     * has requested the system not be heavily disrupted, fail the
2435     * allocation now instead of entering direct reclaim
2436     */
2437    if (deferred_compaction && (gfp_mask & __GFP_NO_KSWAPD))
2438        goto nopage;
2439
2440    /* Try direct reclaim and then allocating */
2441    page = __alloc_pages_direct_reclaim(gfp_mask, order,
2442                    zonelist, high_zoneidx,
2443                    nodemask,
2444                    alloc_flags, preferred_zone,
2445                    migratetype, &did_some_progress);
2446    if (page)
2447        goto got_pg;
2448
2449    /*
2450     * If we failed to make any progress reclaiming, then we are
2451     * running out of options and have to consider going OOM
2452     */
2453    if (!did_some_progress) {
2454        if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
2455            if (oom_killer_disabled)
2456                goto nopage;
2457            /* Coredumps can quickly deplete all memory reserves */
2458            if ((current->flags & PF_DUMPCORE) &&
2459                !(gfp_mask & __GFP_NOFAIL))
2460                goto nopage;
2461            page = __alloc_pages_may_oom(gfp_mask, order,
2462                    zonelist, high_zoneidx,
2463                    nodemask, preferred_zone,
2464                    migratetype);
2465            if (page)
2466                goto got_pg;
2467
2468            if (!(gfp_mask & __GFP_NOFAIL)) {
2469                /*
2470                 * The oom killer is not called for high-order
2471                 * allocations that may fail, so if no progress
2472                 * is being made, there are no other options and
2473                 * retrying is unlikely to help.
2474                 */
2475                if (order > PAGE_ALLOC_COSTLY_ORDER)
2476                    goto nopage;
2477                /*
2478                 * The oom killer is not called for lowmem
2479                 * allocations to prevent needlessly killing
2480                 * innocent tasks.
2481                 */
2482                if (high_zoneidx < ZONE_NORMAL)
2483                    goto nopage;
2484            }
2485
2486            goto restart;
2487        }
2488    }
2489
2490    /* Check if we should retry the allocation */
2491    pages_reclaimed += did_some_progress;
2492    if (should_alloc_retry(gfp_mask, order, did_some_progress,
2493                        pages_reclaimed)) {
2494        /* Wait for some write requests to complete then retry */
2495        wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2496        goto rebalance;
2497    } else {
2498        /*
2499         * High-order allocations do not necessarily loop after
2500         * direct reclaim and reclaim/compaction depends on compaction
2501         * being called after reclaim so call directly if necessary
2502         */
2503        page = __alloc_pages_direct_compact(gfp_mask, order,
2504                    zonelist, high_zoneidx,
2505                    nodemask,
2506                    alloc_flags, preferred_zone,
2507                    migratetype, sync_migration,
2508                    &deferred_compaction,
2509                    &did_some_progress);
2510        if (page)
2511            goto got_pg;
2512    }
2513
2514nopage:
2515    warn_alloc_failed(gfp_mask, order, NULL);
2516    return page;
2517got_pg:
2518    if (kmemcheck_enabled)
2519        kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2520
2521    return page;
2522}
2523
2524/*
2525 * This is the 'heart' of the zoned buddy allocator.
2526 */
2527struct page *
2528__alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2529            struct zonelist *zonelist, nodemask_t *nodemask)
2530{
2531    enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2532    struct zone *preferred_zone;
2533    struct page *page = NULL;
2534    int migratetype = allocflags_to_migratetype(gfp_mask);
2535    unsigned int cpuset_mems_cookie;
2536
2537    gfp_mask &= gfp_allowed_mask;
2538
2539    lockdep_trace_alloc(gfp_mask);
2540
2541    might_sleep_if(gfp_mask & __GFP_WAIT);
2542
2543    if (should_fail_alloc_page(gfp_mask, order))
2544        return NULL;
2545
2546    /*
2547     * Check the zones suitable for the gfp_mask contain at least one
2548     * valid zone. It's possible to have an empty zonelist as a result
2549     * of GFP_THISNODE and a memoryless node
2550     */
2551    if (unlikely(!zonelist->_zonerefs->zone))
2552        return NULL;
2553
2554retry_cpuset:
2555    cpuset_mems_cookie = get_mems_allowed();
2556
2557    /* The preferred zone is used for statistics later */
2558    first_zones_zonelist(zonelist, high_zoneidx,
2559                nodemask ? : &cpuset_current_mems_allowed,
2560                &preferred_zone);
2561    if (!preferred_zone)
2562        goto out;
2563
2564    /* First allocation attempt */
2565    page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2566            zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET,
2567            preferred_zone, migratetype);
2568    if (unlikely(!page))
2569        page = __alloc_pages_slowpath(gfp_mask, order,
2570                zonelist, high_zoneidx, nodemask,
2571                preferred_zone, migratetype);
2572    else
2573        page->pfmemalloc = false;
2574
2575    trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2576
2577out:
2578    /*
2579     * When updating a task's mems_allowed, it is possible to race with
2580     * parallel threads in such a way that an allocation can fail while
2581     * the mask is being updated. If a page allocation is about to fail,
2582     * check if the cpuset changed during allocation and if so, retry.
2583     */
2584    if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2585        goto retry_cpuset;
2586
2587    return page;
2588}
2589EXPORT_SYMBOL(__alloc_pages_nodemask);
2590
2591/*
2592 * Common helper functions.
2593 */
2594unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2595{
2596    struct page *page;
2597
2598    /*
2599     * __get_free_pages() returns a 32-bit address, which cannot represent
2600     * a highmem page
2601     */
2602    VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2603
2604    page = alloc_pages(gfp_mask, order);
2605    if (!page)
2606        return 0;
2607    return (unsigned long) page_address(page);
2608}
2609EXPORT_SYMBOL(__get_free_pages);
2610
2611unsigned long get_zeroed_page(gfp_t gfp_mask)
2612{
2613    return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2614}
2615EXPORT_SYMBOL(get_zeroed_page);
2616
2617void __free_pages(struct page *page, unsigned int order)
2618{
2619    if (put_page_testzero(page)) {
2620        if (order == 0)
2621            free_hot_cold_page(page, 0);
2622        else
2623            __free_pages_ok(page, order);
2624    }
2625}
2626
2627EXPORT_SYMBOL(__free_pages);
2628
2629void free_pages(unsigned long addr, unsigned int order)
2630{
2631    if (addr != 0) {
2632        VM_BUG_ON(!virt_addr_valid((void *)addr));
2633        __free_pages(virt_to_page((void *)addr), order);
2634    }
2635}
2636
2637EXPORT_SYMBOL(free_pages);
2638
2639static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
2640{
2641    if (addr) {
2642        unsigned long alloc_end = addr + (PAGE_SIZE << order);
2643        unsigned long used = addr + PAGE_ALIGN(size);
2644
2645        split_page(virt_to_page((void *)addr), order);
2646        while (used < alloc_end) {
2647            free_page(used);
2648            used += PAGE_SIZE;
2649        }
2650    }
2651    return (void *)addr;
2652}
2653
2654/**
2655 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2656 * @size: the number of bytes to allocate
2657 * @gfp_mask: GFP flags for the allocation
2658 *
2659 * This function is similar to alloc_pages(), except that it allocates the
2660 * minimum number of pages to satisfy the request. alloc_pages() can only
2661 * allocate memory in power-of-two pages.
2662 *
2663 * This function is also limited by MAX_ORDER.
2664 *
2665 * Memory allocated by this function must be released by free_pages_exact().
2666 */
2667void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2668{
2669    unsigned int order = get_order(size);
2670    unsigned long addr;
2671
2672    addr = __get_free_pages(gfp_mask, order);
2673    return make_alloc_exact(addr, order, size);
2674}
2675EXPORT_SYMBOL(alloc_pages_exact);
2676
2677/**
2678 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2679 * pages on a node.
2680 * @nid: the preferred node ID where memory should be allocated
2681 * @size: the number of bytes to allocate
2682 * @gfp_mask: GFP flags for the allocation
2683 *
2684 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2685 * back.
2686 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2687 * but is not exact.
2688 */
2689void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
2690{
2691    unsigned order = get_order(size);
2692    struct page *p = alloc_pages_node(nid, gfp_mask, order);
2693    if (!p)
2694        return NULL;
2695    return make_alloc_exact((unsigned long)page_address(p), order, size);
2696}
2697EXPORT_SYMBOL(alloc_pages_exact_nid);
2698
2699/**
2700 * free_pages_exact - release memory allocated via alloc_pages_exact()
2701 * @virt: the value returned by alloc_pages_exact.
2702 * @size: size of allocation, same value as passed to alloc_pages_exact().
2703 *
2704 * Release the memory allocated by a previous call to alloc_pages_exact.
2705 */
2706void free_pages_exact(void *virt, size_t size)
2707{
2708    unsigned long addr = (unsigned long)virt;
2709    unsigned long end = addr + PAGE_ALIGN(size);
2710
2711    while (addr < end) {
2712        free_page(addr);
2713        addr += PAGE_SIZE;
2714    }
2715}
2716EXPORT_SYMBOL(free_pages_exact);
2717
2718static unsigned int nr_free_zone_pages(int offset)
2719{
2720    struct zoneref *z;
2721    struct zone *zone;
2722
2723    /* Just pick one node, since fallback list is circular */
2724    unsigned int sum = 0;
2725
2726    struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2727
2728    for_each_zone_zonelist(zone, z, zonelist, offset) {
2729        unsigned long size = zone->present_pages;
2730        unsigned long high = high_wmark_pages(zone);
2731        if (size > high)
2732            sum += size - high;
2733    }
2734
2735    return sum;
2736}
2737
2738/*
2739 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2740 */
2741unsigned int nr_free_buffer_pages(void)
2742{
2743    return nr_free_zone_pages(gfp_zone(GFP_USER));
2744}
2745EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2746
2747/*
2748 * Amount of free RAM allocatable within all zones
2749 */
2750unsigned int nr_free_pagecache_pages(void)
2751{
2752    return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2753}
2754
2755static inline void show_node(struct zone *zone)
2756{
2757    if (NUMA_BUILD)
2758        printk("Node %d ", zone_to_nid(zone));
2759}
2760
2761void si_meminfo(struct sysinfo *val)
2762{
2763    val->totalram = totalram_pages;
2764    val->sharedram = 0;
2765    val->freeram = global_page_state(NR_FREE_PAGES);
2766    val->bufferram = nr_blockdev_pages();
2767    val->totalhigh = totalhigh_pages;
2768    val->freehigh = nr_free_highpages();
2769    val->mem_unit = PAGE_SIZE;
2770}
2771
2772EXPORT_SYMBOL(si_meminfo);
2773
2774#ifdef CONFIG_NUMA
2775void si_meminfo_node(struct sysinfo *val, int nid)
2776{
2777    pg_data_t *pgdat = NODE_DATA(nid);
2778
2779    val->totalram = pgdat->node_present_pages;
2780    val->freeram = node_page_state(nid, NR_FREE_PAGES);
2781#ifdef CONFIG_HIGHMEM
2782    val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
2783    val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2784            NR_FREE_PAGES);
2785#else
2786    val->totalhigh = 0;
2787    val->freehigh = 0;
2788#endif
2789    val->mem_unit = PAGE_SIZE;
2790}
2791#endif
2792
2793/*
2794 * Determine whether the node should be displayed or not, depending on whether
2795 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
2796 */
2797bool skip_free_areas_node(unsigned int flags, int nid)
2798{
2799    bool ret = false;
2800    unsigned int cpuset_mems_cookie;
2801
2802    if (!(flags & SHOW_MEM_FILTER_NODES))
2803        goto out;
2804
2805    do {
2806        cpuset_mems_cookie = get_mems_allowed();
2807        ret = !node_isset(nid, cpuset_current_mems_allowed);
2808    } while (!put_mems_allowed(cpuset_mems_cookie));
2809out:
2810    return ret;
2811}
2812
2813#define K(x) ((x) << (PAGE_SHIFT-10))
2814
2815/*
2816 * Show free area list (used inside shift_scroll-lock stuff)
2817 * We also calculate the percentage fragmentation. We do this by counting the
2818 * memory on each free list with the exception of the first item on the list.
2819 * Suppresses nodes that are not allowed by current's cpuset if
2820 * SHOW_MEM_FILTER_NODES is passed.
2821 */
2822void show_free_areas(unsigned int filter)
2823{
2824    int cpu;
2825    struct zone *zone;
2826
2827    for_each_populated_zone(zone) {
2828        if (skip_free_areas_node(filter, zone_to_nid(zone)))
2829            continue;
2830        show_node(zone);
2831        printk("%s per-cpu:\n", zone->name);
2832
2833        for_each_online_cpu(cpu) {
2834            struct per_cpu_pageset *pageset;
2835
2836            pageset = per_cpu_ptr(zone->pageset, cpu);
2837
2838            printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2839                   cpu, pageset->pcp.high,
2840                   pageset->pcp.batch, pageset->pcp.count);
2841        }
2842    }
2843
2844    printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2845        " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2846        " unevictable:%lu"
2847        " dirty:%lu writeback:%lu unstable:%lu\n"
2848        " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2849        " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n",
2850        global_page_state(NR_ACTIVE_ANON),
2851        global_page_state(NR_INACTIVE_ANON),
2852        global_page_state(NR_ISOLATED_ANON),
2853        global_page_state(NR_ACTIVE_FILE),
2854        global_page_state(NR_INACTIVE_FILE),
2855        global_page_state(NR_ISOLATED_FILE),
2856        global_page_state(NR_UNEVICTABLE),
2857        global_page_state(NR_FILE_DIRTY),
2858        global_page_state(NR_WRITEBACK),
2859        global_page_state(NR_UNSTABLE_NFS),
2860        global_page_state(NR_FREE_PAGES),
2861        global_page_state(NR_SLAB_RECLAIMABLE),
2862        global_page_state(NR_SLAB_UNRECLAIMABLE),
2863        global_page_state(NR_FILE_MAPPED),
2864        global_page_state(NR_SHMEM),
2865        global_page_state(NR_PAGETABLE),
2866        global_page_state(NR_BOUNCE));
2867
2868    for_each_populated_zone(zone) {
2869        int i;
2870
2871        if (skip_free_areas_node(filter, zone_to_nid(zone)))
2872            continue;
2873        show_node(zone);
2874        printk("%s"
2875            " free:%lukB"
2876            " min:%lukB"
2877            " low:%lukB"
2878            " high:%lukB"
2879            " active_anon:%lukB"
2880            " inactive_anon:%lukB"
2881            " active_file:%lukB"
2882            " inactive_file:%lukB"
2883            " unevictable:%lukB"
2884            " isolated(anon):%lukB"
2885            " isolated(file):%lukB"
2886            " present:%lukB"
2887            " mlocked:%lukB"
2888            " dirty:%lukB"
2889            " writeback:%lukB"
2890            " mapped:%lukB"
2891            " shmem:%lukB"
2892            " slab_reclaimable:%lukB"
2893            " slab_unreclaimable:%lukB"
2894            " kernel_stack:%lukB"
2895            " pagetables:%lukB"
2896            " unstable:%lukB"
2897            " bounce:%lukB"
2898            " writeback_tmp:%lukB"
2899            " pages_scanned:%lu"
2900            " all_unreclaimable? %s"
2901            "\n",
2902            zone->name,
2903            K(zone_page_state(zone, NR_FREE_PAGES)),
2904            K(min_wmark_pages(zone)),
2905            K(low_wmark_pages(zone)),
2906            K(high_wmark_pages(zone)),
2907            K(zone_page_state(zone, NR_ACTIVE_ANON)),
2908            K(zone_page_state(zone, NR_INACTIVE_ANON)),
2909            K(zone_page_state(zone, NR_ACTIVE_FILE)),
2910            K(zone_page_state(zone, NR_INACTIVE_FILE)),
2911            K(zone_page_state(zone, NR_UNEVICTABLE)),
2912            K(zone_page_state(zone, NR_ISOLATED_ANON)),
2913            K(zone_page_state(zone, NR_ISOLATED_FILE)),
2914            K(zone->present_pages),
2915            K(zone_page_state(zone, NR_MLOCK)),
2916            K(zone_page_state(zone, NR_FILE_DIRTY)),
2917            K(zone_page_state(zone, NR_WRITEBACK)),
2918            K(zone_page_state(zone, NR_FILE_MAPPED)),
2919            K(zone_page_state(zone, NR_SHMEM)),
2920            K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
2921            K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
2922            zone_page_state(zone, NR_KERNEL_STACK) *
2923                THREAD_SIZE / 1024,
2924            K(zone_page_state(zone, NR_PAGETABLE)),
2925            K(zone_page_state(zone, NR_UNSTABLE_NFS)),
2926            K(zone_page_state(zone, NR_BOUNCE)),
2927            K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
2928            zone->pages_scanned,
2929            (zone->all_unreclaimable ? "yes" : "no")
2930            );
2931        printk("lowmem_reserve[]:");
2932        for (i = 0; i < MAX_NR_ZONES; i++)
2933            printk(" %lu", zone->lowmem_reserve[i]);
2934        printk("\n");
2935    }
2936
2937    for_each_populated_zone(zone) {
2938         unsigned long nr[MAX_ORDER], flags, order, total = 0;
2939
2940        if (skip_free_areas_node(filter, zone_to_nid(zone)))
2941            continue;
2942        show_node(zone);
2943        printk("%s: ", zone->name);
2944
2945        spin_lock_irqsave(&zone->lock, flags);
2946        for (order = 0; order < MAX_ORDER; order++) {
2947            nr[order] = zone->free_area[order].nr_free;
2948            total += nr[order] << order;
2949        }
2950        spin_unlock_irqrestore(&zone->lock, flags);
2951        for (order = 0; order < MAX_ORDER; order++)
2952            printk("%lu*%lukB ", nr[order], K(1UL) << order);
2953        printk("= %lukB\n", K(total));
2954    }
2955
2956    printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
2957
2958    show_swap_cache_info();
2959}
2960
2961static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
2962{
2963    zoneref->zone = zone;
2964    zoneref->zone_idx = zone_idx(zone);
2965}
2966
2967/*
2968 * Builds allocation fallback zone lists.
2969 *
2970 * Add all populated zones of a node to the zonelist.
2971 */
2972static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
2973                int nr_zones, enum zone_type zone_type)
2974{
2975    struct zone *zone;
2976
2977    BUG_ON(zone_type >= MAX_NR_ZONES);
2978    zone_type++;
2979
2980    do {
2981        zone_type--;
2982        zone = pgdat->node_zones + zone_type;
2983        if (populated_zone(zone)) {
2984            zoneref_set_zone(zone,
2985                &zonelist->_zonerefs[nr_zones++]);
2986            check_highest_zone(zone_type);
2987        }
2988
2989    } while (zone_type);
2990    return nr_zones;
2991}
2992
2993
2994/*
2995 * zonelist_order:
2996 * 0 = automatic detection of better ordering.
2997 * 1 = order by ([node] distance, -zonetype)
2998 * 2 = order by (-zonetype, [node] distance)
2999 *
3000 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3001 * the same zonelist. So only NUMA can configure this param.
3002 */
3003#define ZONELIST_ORDER_DEFAULT 0
3004#define ZONELIST_ORDER_NODE 1
3005#define ZONELIST_ORDER_ZONE 2
3006
3007/* zonelist order in the kernel.
3008 * set_zonelist_order() will set this to NODE or ZONE.
3009 */
3010static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3011static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3012
3013
3014#ifdef CONFIG_NUMA
3015/* The value user specified ....changed by config */
3016static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3017/* string for sysctl */
3018#define NUMA_ZONELIST_ORDER_LEN 16
3019char numa_zonelist_order[16] = "default";
3020
3021/*
3022 * interface for configure zonelist ordering.
3023 * command line option "numa_zonelist_order"
3024 * = "[dD]efault - default, automatic configuration.
3025 * = "[nN]ode - order by node locality, then by zone within node
3026 * = "[zZ]one - order by zone, then by locality within zone
3027 */
3028
3029static int __parse_numa_zonelist_order(char *s)
3030{
3031    if (*s == 'd' || *s == 'D') {
3032        user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3033    } else if (*s == 'n' || *s == 'N') {
3034        user_zonelist_order = ZONELIST_ORDER_NODE;
3035    } else if (*s == 'z' || *s == 'Z') {
3036        user_zonelist_order = ZONELIST_ORDER_ZONE;
3037    } else {
3038        printk(KERN_WARNING
3039            "Ignoring invalid numa_zonelist_order value: "
3040            "%s\n", s);
3041        return -EINVAL;
3042    }
3043    return 0;
3044}
3045
3046static __init int setup_numa_zonelist_order(char *s)
3047{
3048    int ret;
3049
3050    if (!s)
3051        return 0;
3052
3053    ret = __parse_numa_zonelist_order(s);
3054    if (ret == 0)
3055        strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3056
3057    return ret;
3058}
3059early_param("numa_zonelist_order", setup_numa_zonelist_order);
3060
3061/*
3062 * sysctl handler for numa_zonelist_order
3063 */
3064int numa_zonelist_order_handler(ctl_table *table, int write,
3065        void __user *buffer, size_t *length,
3066        loff_t *ppos)
3067{
3068    char saved_string[NUMA_ZONELIST_ORDER_LEN];
3069    int ret;
3070    static DEFINE_MUTEX(zl_order_mutex);
3071
3072    mutex_lock(&zl_order_mutex);
3073    if (write)
3074        strcpy(saved_string, (char*)table->data);
3075    ret = proc_dostring(table, write, buffer, length, ppos);
3076    if (ret)
3077        goto out;
3078    if (write) {
3079        int oldval = user_zonelist_order;
3080        if (__parse_numa_zonelist_order((char*)table->data)) {
3081            /*
3082             * bogus value. restore saved string
3083             */
3084            strncpy((char*)table->data, saved_string,
3085                NUMA_ZONELIST_ORDER_LEN);
3086            user_zonelist_order = oldval;
3087        } else if (oldval != user_zonelist_order) {
3088            mutex_lock(&zonelists_mutex);
3089            build_all_zonelists(NULL, NULL);
3090            mutex_unlock(&zonelists_mutex);
3091        }
3092    }
3093out:
3094    mutex_unlock(&zl_order_mutex);
3095    return ret;
3096}
3097
3098
3099#define MAX_NODE_LOAD (nr_online_nodes)
3100static int node_load[MAX_NUMNODES];
3101
3102/**
3103 * find_next_best_node - find the next node that should appear in a given node's fallback list
3104 * @node: node whose fallback list we're appending
3105 * @used_node_mask: nodemask_t of already used nodes
3106 *
3107 * We use a number of factors to determine which is the next node that should
3108 * appear on a given node's fallback list. The node should not have appeared
3109 * already in @node's fallback list, and it should be the next closest node
3110 * according to the distance array (which contains arbitrary distance values
3111 * from each node to each node in the system), and should also prefer nodes
3112 * with no CPUs, since presumably they'll have very little allocation pressure
3113 * on them otherwise.
3114 * It returns -1 if no node is found.
3115 */
3116static int find_next_best_node(int node, nodemask_t *used_node_mask)
3117{
3118    int n, val;
3119    int min_val = INT_MAX;
3120    int best_node = -1;
3121    const struct cpumask *tmp = cpumask_of_node(0);
3122
3123    /* Use the local node if we haven't already */
3124    if (!node_isset(node, *used_node_mask)) {
3125        node_set(node, *used_node_mask);
3126        return node;
3127    }
3128
3129    for_each_node_state(n, N_HIGH_MEMORY) {
3130
3131        /* Don't want a node to appear more than once */
3132        if (node_isset(n, *used_node_mask))
3133            continue;
3134
3135        /* Use the distance array to find the distance */
3136        val = node_distance(node, n);
3137
3138        /* Penalize nodes under us ("prefer the next node") */
3139        val += (n < node);
3140
3141        /* Give preference to headless and unused nodes */
3142        tmp = cpumask_of_node(n);
3143        if (!cpumask_empty(tmp))
3144            val += PENALTY_FOR_NODE_WITH_CPUS;
3145
3146        /* Slight preference for less loaded node */
3147        val *= (MAX_NODE_LOAD*MAX_NUMNODES);
3148        val += node_load[n];
3149
3150        if (val < min_val) {
3151            min_val = val;
3152            best_node = n;
3153        }
3154    }
3155
3156    if (best_node >= 0)
3157        node_set(best_node, *used_node_mask);
3158
3159    return best_node;
3160}
3161
3162
3163/*
3164 * Build zonelists ordered by node and zones within node.
3165 * This results in maximum locality--normal zone overflows into local
3166 * DMA zone, if any--but risks exhausting DMA zone.
3167 */
3168static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
3169{
3170    int j;
3171    struct zonelist *zonelist;
3172
3173    zonelist = &pgdat->node_zonelists[0];
3174    for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
3175        ;
3176    j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3177                            MAX_NR_ZONES - 1);
3178    zonelist->_zonerefs[j].zone = NULL;
3179    zonelist->_zonerefs[j].zone_idx = 0;
3180}
3181
3182/*
3183 * Build gfp_thisnode zonelists
3184 */
3185static void build_thisnode_zonelists(pg_data_t *pgdat)
3186{
3187    int j;
3188    struct zonelist *zonelist;
3189
3190    zonelist = &pgdat->node_zonelists[1];
3191    j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3192    zonelist->_zonerefs[j].zone = NULL;
3193    zonelist->_zonerefs[j].zone_idx = 0;
3194}
3195
3196/*
3197 * Build zonelists ordered by zone and nodes within zones.
3198 * This results in conserving DMA zone[s] until all Normal memory is
3199 * exhausted, but results in overflowing to remote node while memory
3200 * may still exist in local DMA zone.
3201 */
3202static int node_order[MAX_NUMNODES];
3203
3204static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
3205{
3206    int pos, j, node;
3207    int zone_type; /* needs to be signed */
3208    struct zone *z;
3209    struct zonelist *zonelist;
3210
3211    zonelist = &pgdat->node_zonelists[0];
3212    pos = 0;
3213    for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
3214        for (j = 0; j < nr_nodes; j++) {
3215            node = node_order[j];
3216            z = &NODE_DATA(node)->node_zones[zone_type];
3217            if (populated_zone(z)) {
3218                zoneref_set_zone(z,
3219                    &zonelist->_zonerefs[pos++]);
3220                check_highest_zone(zone_type);
3221            }
3222        }
3223    }
3224    zonelist->_zonerefs[pos].zone = NULL;
3225    zonelist->_zonerefs[pos].zone_idx = 0;
3226}
3227
3228static int default_zonelist_order(void)
3229{
3230    int nid, zone_type;
3231    unsigned long low_kmem_size,total_size;
3232    struct zone *z;
3233    int average_size;
3234    /*
3235         * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
3236     * If they are really small and used heavily, the system can fall
3237     * into OOM very easily.
3238     * This function detect ZONE_DMA/DMA32 size and configures zone order.
3239     */
3240    /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
3241    low_kmem_size = 0;
3242    total_size = 0;
3243    for_each_online_node(nid) {
3244        for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3245            z = &NODE_DATA(nid)->node_zones[zone_type];
3246            if (populated_zone(z)) {
3247                if (zone_type < ZONE_NORMAL)
3248                    low_kmem_size += z->present_pages;
3249                total_size += z->present_pages;
3250            } else if (zone_type == ZONE_NORMAL) {
3251                /*
3252                 * If any node has only lowmem, then node order
3253                 * is preferred to allow kernel allocations
3254                 * locally; otherwise, they can easily infringe
3255                 * on other nodes when there is an abundance of
3256                 * lowmem available to allocate from.
3257                 */
3258                return ZONELIST_ORDER_NODE;
3259            }
3260        }
3261    }
3262    if (!low_kmem_size || /* there are no DMA area. */
3263        low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
3264        return ZONELIST_ORDER_NODE;
3265    /*
3266     * look into each node's config.
3267       * If there is a node whose DMA/DMA32 memory is very big area on
3268      * local memory, NODE_ORDER may be suitable.
3269         */
3270    average_size = total_size /
3271                (nodes_weight(node_states[N_HIGH_MEMORY]) + 1);
3272    for_each_online_node(nid) {
3273        low_kmem_size = 0;
3274        total_size = 0;
3275        for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3276            z = &NODE_DATA(nid)->node_zones[zone_type];
3277            if (populated_zone(z)) {
3278                if (zone_type < ZONE_NORMAL)
3279                    low_kmem_size += z->present_pages;
3280                total_size += z->present_pages;
3281            }
3282        }
3283        if (low_kmem_size &&
3284            total_size > average_size && /* ignore small node */
3285            low_kmem_size > total_size * 70/100)
3286            return ZONELIST_ORDER_NODE;
3287    }
3288    return ZONELIST_ORDER_ZONE;
3289}
3290
3291static void set_zonelist_order(void)
3292{
3293    if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3294        current_zonelist_order = default_zonelist_order();
3295    else
3296        current_zonelist_order = user_zonelist_order;
3297}
3298
3299static void build_zonelists(pg_data_t *pgdat)
3300{
3301    int j, node, load;
3302    enum zone_type i;
3303    nodemask_t used_mask;
3304    int local_node, prev_node;
3305    struct zonelist *zonelist;
3306    int order = current_zonelist_order;
3307
3308    /* initialize zonelists */
3309    for (i = 0; i < MAX_ZONELISTS; i++) {
3310        zonelist = pgdat->node_zonelists + i;
3311        zonelist->_zonerefs[0].zone = NULL;
3312        zonelist->_zonerefs[0].zone_idx = 0;
3313    }
3314
3315    /* NUMA-aware ordering of nodes */
3316    local_node = pgdat->node_id;
3317    load = nr_online_nodes;
3318    prev_node = local_node;
3319    nodes_clear(used_mask);
3320
3321    memset(node_order, 0, sizeof(node_order));
3322    j = 0;
3323
3324    while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3325        int distance = node_distance(local_node, node);
3326
3327        /*
3328         * If another node is sufficiently far away then it is better
3329         * to reclaim pages in a zone before going off node.
3330         */
3331        if (distance > RECLAIM_DISTANCE)
3332            zone_reclaim_mode = 1;
3333
3334        /*
3335         * We don't want to pressure a particular node.
3336         * So adding penalty to the first node in same
3337         * distance group to make it round-robin.
3338         */
3339        if (distance != node_distance(local_node, prev_node))
3340            node_load[node] = load;
3341
3342        prev_node = node;
3343        load--;
3344        if (order == ZONELIST_ORDER_NODE)
3345            build_zonelists_in_node_order(pgdat, node);
3346        else
3347            node_order[j++] = node; /* remember order */
3348    }
3349
3350    if (order == ZONELIST_ORDER_ZONE) {
3351        /* calculate node order -- i.e., DMA last! */
3352        build_zonelists_in_zone_order(pgdat, j);
3353    }
3354
3355    build_thisnode_zonelists(pgdat);
3356}
3357
3358/* Construct the zonelist performance cache - see further mmzone.h */
3359static void build_zonelist_cache(pg_data_t *pgdat)
3360{
3361    struct zonelist *zonelist;
3362    struct zonelist_cache *zlc;
3363    struct zoneref *z;
3364
3365    zonelist = &pgdat->node_zonelists[0];
3366    zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3367    bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3368    for (z = zonelist->_zonerefs; z->zone; z++)
3369        zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3370}
3371
3372#ifdef CONFIG_HAVE_MEMORYLESS_NODES
3373/*
3374 * Return node id of node used for "local" allocations.
3375 * I.e., first node id of first zone in arg node's generic zonelist.
3376 * Used for initializing percpu 'numa_mem', which is used primarily
3377 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3378 */
3379int local_memory_node(int node)
3380{
3381    struct zone *zone;
3382
3383    (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3384                   gfp_zone(GFP_KERNEL),
3385                   NULL,
3386                   &zone);
3387    return zone->node;
3388}
3389#endif
3390
3391#else /* CONFIG_NUMA */
3392
3393static void set_zonelist_order(void)
3394{
3395    current_zonelist_order = ZONELIST_ORDER_ZONE;
3396}
3397
3398static void build_zonelists(pg_data_t *pgdat)
3399{
3400    int node, local_node;
3401    enum zone_type j;
3402    struct zonelist *zonelist;
3403
3404    local_node = pgdat->node_id;
3405
3406    zonelist = &pgdat->node_zonelists[0];
3407    j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3408
3409    /*
3410     * Now we build the zonelist so that it contains the zones
3411     * of all the other nodes.
3412     * We don't want to pressure a particular node, so when
3413     * building the zones for node N, we make sure that the
3414     * zones coming right after the local ones are those from
3415     * node N+1 (modulo N)
3416     */
3417    for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3418        if (!node_online(node))
3419            continue;
3420        j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3421                            MAX_NR_ZONES - 1);
3422    }
3423    for (node = 0; node < local_node; node++) {
3424        if (!node_online(node))
3425            continue;
3426        j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3427                            MAX_NR_ZONES - 1);
3428    }
3429
3430    zonelist->_zonerefs[j].zone = NULL;
3431    zonelist->_zonerefs[j].zone_idx = 0;
3432}
3433
3434/* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3435static void build_zonelist_cache(pg_data_t *pgdat)
3436{
3437    pgdat->node_zonelists[0].zlcache_ptr = NULL;
3438}
3439
3440#endif /* CONFIG_NUMA */
3441
3442/*
3443 * Boot pageset table. One per cpu which is going to be used for all
3444 * zones and all nodes. The parameters will be set in such a way
3445 * that an item put on a list will immediately be handed over to
3446 * the buddy list. This is safe since pageset manipulation is done
3447 * with interrupts disabled.
3448 *
3449 * The boot_pagesets must be kept even after bootup is complete for
3450 * unused processors and/or zones. They do play a role for bootstrapping
3451 * hotplugged processors.
3452 *
3453 * zoneinfo_show() and maybe other functions do
3454 * not check if the processor is online before following the pageset pointer.
3455 * Other parts of the kernel may not check if the zone is available.
3456 */
3457static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3458static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3459static void setup_zone_pageset(struct zone *zone);
3460
3461/*
3462 * Global mutex to protect against size modification of zonelists
3463 * as well as to serialize pageset setup for the new populated zone.
3464 */
3465DEFINE_MUTEX(zonelists_mutex);
3466
3467/* return values int ....just for stop_machine() */
3468static int __build_all_zonelists(void *data)
3469{
3470    int nid;
3471    int cpu;
3472    pg_data_t *self = data;
3473
3474#ifdef CONFIG_NUMA
3475    memset(node_load, 0, sizeof(node_load));
3476#endif
3477
3478    if (self && !node_online(self->node_id)) {
3479        build_zonelists(self);
3480        build_zonelist_cache(self);
3481    }
3482
3483    for_each_online_node(nid) {
3484        pg_data_t *pgdat = NODE_DATA(nid);
3485
3486        build_zonelists(pgdat);
3487        build_zonelist_cache(pgdat);
3488    }
3489
3490    /*
3491     * Initialize the boot_pagesets that are going to be used
3492     * for bootstrapping processors. The real pagesets for
3493     * each zone will be allocated later when the per cpu
3494     * allocator is available.
3495     *
3496     * boot_pagesets are used also for bootstrapping offline
3497     * cpus if the system is already booted because the pagesets
3498     * are needed to initialize allocators on a specific cpu too.
3499     * F.e. the percpu allocator needs the page allocator which
3500     * needs the percpu allocator in order to allocate its pagesets
3501     * (a chicken-egg dilemma).
3502     */
3503    for_each_possible_cpu(cpu) {
3504        setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3505
3506#ifdef CONFIG_HAVE_MEMORYLESS_NODES
3507        /*
3508         * We now know the "local memory node" for each node--
3509         * i.e., the node of the first zone in the generic zonelist.
3510         * Set up numa_mem percpu variable for on-line cpus. During
3511         * boot, only the boot cpu should be on-line; we'll init the
3512         * secondary cpus' numa_mem as they come on-line. During
3513         * node/memory hotplug, we'll fixup all on-line cpus.
3514         */
3515        if (cpu_online(cpu))
3516            set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3517#endif
3518    }
3519
3520    return 0;
3521}
3522
3523/*
3524 * Called with zonelists_mutex held always
3525 * unless system_state == SYSTEM_BOOTING.
3526 */
3527void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
3528{
3529    set_zonelist_order();
3530
3531    if (system_state == SYSTEM_BOOTING) {
3532        __build_all_zonelists(NULL);
3533        mminit_verify_zonelist();
3534        cpuset_init_current_mems_allowed();
3535    } else {
3536        /* we have to stop all cpus to guarantee there is no user
3537           of zonelist */
3538#ifdef CONFIG_MEMORY_HOTPLUG
3539        if (zone)
3540            setup_zone_pageset(zone);
3541#endif
3542        stop_machine(__build_all_zonelists, pgdat, NULL);
3543        /* cpuset refresh routine should be here */
3544    }
3545    vm_total_pages = nr_free_pagecache_pages();
3546    /*
3547     * Disable grouping by mobility if the number of pages in the
3548     * system is too low to allow the mechanism to work. It would be
3549     * more accurate, but expensive to check per-zone. This check is
3550     * made on memory-hotadd so a system can start with mobility
3551     * disabled and enable it later
3552     */
3553    if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3554        page_group_by_mobility_disabled = 1;
3555    else
3556        page_group_by_mobility_disabled = 0;
3557
3558    printk("Built %i zonelists in %s order, mobility grouping %s. "
3559        "Total pages: %ld\n",
3560            nr_online_nodes,
3561            zonelist_order_name[current_zonelist_order],
3562            page_group_by_mobility_disabled ? "off" : "on",
3563            vm_total_pages);
3564#ifdef CONFIG_NUMA
3565    printk("Policy zone: %s\n", zone_names[policy_zone]);
3566#endif
3567}
3568
3569/*
3570 * Helper functions to size the waitqueue hash table.
3571 * Essentially these want to choose hash table sizes sufficiently
3572 * large so that collisions trying to wait on pages are rare.
3573 * But in fact, the number of active page waitqueues on typical
3574 * systems is ridiculously low, less than 200. So this is even
3575 * conservative, even though it seems large.
3576 *
3577 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3578 * waitqueues, i.e. the size of the waitq table given the number of pages.
3579 */
3580#define PAGES_PER_WAITQUEUE 256
3581
3582#ifndef CONFIG_MEMORY_HOTPLUG
3583static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3584{
3585    unsigned long size = 1;
3586
3587    pages /= PAGES_PER_WAITQUEUE;
3588
3589    while (size < pages)
3590        size <<= 1;
3591
3592    /*
3593     * Once we have dozens or even hundreds of threads sleeping
3594     * on IO we've got bigger problems than wait queue collision.
3595     * Limit the size of the wait table to a reasonable size.
3596     */
3597    size = min(size, 4096UL);
3598
3599    return max(size, 4UL);
3600}
3601#else
3602/*
3603 * A zone's size might be changed by hot-add, so it is not possible to determine
3604 * a suitable size for its wait_table. So we use the maximum size now.
3605 *
3606 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3607 *
3608 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3609 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3610 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3611 *
3612 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3613 * or more by the traditional way. (See above). It equals:
3614 *
3615 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3616 * ia64(16K page size) : = ( 8G + 4M)byte.
3617 * powerpc (64K page size) : = (32G +16M)byte.
3618 */
3619static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3620{
3621    return 4096UL;
3622}
3623#endif
3624
3625/*
3626 * This is an integer logarithm so that shifts can be used later
3627 * to extract the more random high bits from the multiplicative
3628 * hash function before the remainder is taken.
3629 */
3630static inline unsigned long wait_table_bits(unsigned long size)
3631{
3632    return ffz(~size);
3633}
3634
3635#define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3636
3637/*
3638 * Check if a pageblock contains reserved pages
3639 */
3640static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
3641{
3642    unsigned long pfn;
3643
3644    for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3645        if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
3646            return 1;
3647    }
3648    return 0;
3649}
3650
3651/*
3652 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3653 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3654 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3655 * higher will lead to a bigger reserve which will get freed as contiguous
3656 * blocks as reclaim kicks in
3657 */
3658static void setup_zone_migrate_reserve(struct zone *zone)
3659{
3660    unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
3661    struct page *page;
3662    unsigned long block_migratetype;
3663    int reserve;
3664
3665    /*
3666     * Get the start pfn, end pfn and the number of blocks to reserve
3667     * We have to be careful to be aligned to pageblock_nr_pages to
3668     * make sure that we always check pfn_valid for the first page in
3669     * the block.
3670     */
3671    start_pfn = zone->zone_start_pfn;
3672    end_pfn = start_pfn + zone->spanned_pages;
3673    start_pfn = roundup(start_pfn, pageblock_nr_pages);
3674    reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3675                            pageblock_order;
3676
3677    /*
3678     * Reserve blocks are generally in place to help high-order atomic
3679     * allocations that are short-lived. A min_free_kbytes value that
3680     * would result in more than 2 reserve blocks for atomic allocations
3681     * is assumed to be in place to help anti-fragmentation for the
3682     * future allocation of hugepages at runtime.
3683     */
3684    reserve = min(2, reserve);
3685
3686    for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
3687        if (!pfn_valid(pfn))
3688            continue;
3689        page = pfn_to_page(pfn);
3690
3691        /* Watch out for overlapping nodes */
3692        if (page_to_nid(page) != zone_to_nid(zone))
3693            continue;
3694
3695        block_migratetype = get_pageblock_migratetype(page);
3696
3697        /* Only test what is necessary when the reserves are not met */
3698        if (reserve > 0) {
3699            /*
3700             * Blocks with reserved pages will never free, skip
3701             * them.
3702             */
3703            block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
3704            if (pageblock_is_reserved(pfn, block_end_pfn))
3705                continue;
3706
3707            /* If this block is reserved, account for it */
3708            if (block_migratetype == MIGRATE_RESERVE) {
3709                reserve--;
3710                continue;
3711            }
3712
3713            /* Suitable for reserving if this block is movable */
3714            if (block_migratetype == MIGRATE_MOVABLE) {
3715                set_pageblock_migratetype(page,
3716                            MIGRATE_RESERVE);
3717                move_freepages_block(zone, page,
3718                            MIGRATE_RESERVE);
3719                reserve--;
3720                continue;
3721            }
3722        }
3723
3724        /*
3725         * If the reserve is met and this is a previous reserved block,
3726         * take it back
3727         */
3728        if (block_migratetype == MIGRATE_RESERVE) {
3729            set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3730            move_freepages_block(zone, page, MIGRATE_MOVABLE);
3731        }
3732    }
3733}
3734
3735/*
3736 * Initially all pages are reserved - free ones are freed
3737 * up by free_all_bootmem() once the early boot process is
3738 * done. Non-atomic initialization, single-pass.
3739 */
3740void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
3741        unsigned long start_pfn, enum memmap_context context)
3742{
3743    struct page *page;
3744    unsigned long end_pfn = start_pfn + size;
3745    unsigned long pfn;
3746    struct zone *z;
3747
3748    if (highest_memmap_pfn < end_pfn - 1)
3749        highest_memmap_pfn = end_pfn - 1;
3750
3751    z = &NODE_DATA(nid)->node_zones[zone];
3752    for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3753        /*
3754         * There can be holes in boot-time mem_map[]s
3755         * handed to this function. They do not
3756         * exist on hotplugged memory.
3757         */
3758        if (context == MEMMAP_EARLY) {
3759            if (!early_pfn_valid(pfn))
3760                continue;
3761            if (!early_pfn_in_nid(pfn, nid))
3762                continue;
3763        }
3764        page = pfn_to_page(pfn);
3765        set_page_links(page, zone, nid, pfn);
3766        mminit_verify_page_links(page, zone, nid, pfn);
3767        init_page_count(page);
3768        reset_page_mapcount(page);
3769        SetPageReserved(page);
3770        /*
3771         * Mark the block movable so that blocks are reserved for
3772         * movable at startup. This will force kernel allocations
3773         * to reserve their blocks rather than leaking throughout
3774         * the address space during boot when many long-lived
3775         * kernel allocations are made. Later some blocks near
3776         * the start are marked MIGRATE_RESERVE by
3777         * setup_zone_migrate_reserve()
3778         *
3779         * bitmap is created for zone's valid pfn range. but memmap
3780         * can be created for invalid pages (for alignment)
3781         * check here not to call set_pageblock_migratetype() against
3782         * pfn out of zone.
3783         */
3784        if ((z->zone_start_pfn <= pfn)
3785            && (pfn < z->zone_start_pfn + z->spanned_pages)
3786            && !(pfn & (pageblock_nr_pages - 1)))
3787            set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3788
3789        INIT_LIST_HEAD(&page->lru);
3790#ifdef WANT_PAGE_VIRTUAL
3791        /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3792        if (!is_highmem_idx(zone))
3793            set_page_address(page, __va(pfn << PAGE_SHIFT));
3794#endif
3795    }
3796}
3797
3798static void __meminit zone_init_free_lists(struct zone *zone)
3799{
3800    int order, t;
3801    for_each_migratetype_order(order, t) {
3802        INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
3803        zone->free_area[order].nr_free = 0;
3804    }
3805}
3806
3807#ifndef __HAVE_ARCH_MEMMAP_INIT
3808#define memmap_init(size, nid, zone, start_pfn) \
3809    memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3810#endif
3811
3812static int __meminit zone_batchsize(struct zone *zone)
3813{
3814#ifdef CONFIG_MMU
3815    int batch;
3816
3817    /*
3818     * The per-cpu-pages pools are set to around 1000th of the
3819     * size of the zone. But no more than 1/2 of a meg.
3820     *
3821     * OK, so we don't know how big the cache is. So guess.
3822     */
3823    batch = zone->present_pages / 1024;
3824    if (batch * PAGE_SIZE > 512 * 1024)
3825        batch = (512 * 1024) / PAGE_SIZE;
3826    batch /= 4; /* We effectively *= 4 below */
3827    if (batch < 1)
3828        batch = 1;
3829
3830    /*
3831     * Clamp the batch to a 2^n - 1 value. Having a power
3832     * of 2 value was found to be more likely to have
3833     * suboptimal cache aliasing properties in some cases.
3834     *
3835     * For example if 2 tasks are alternately allocating
3836     * batches of pages, one task can end up with a lot
3837     * of pages of one half of the possible page colors
3838     * and the other with pages of the other colors.
3839     */
3840    batch = rounddown_pow_of_two(batch + batch/2) - 1;
3841
3842    return batch;
3843
3844#else
3845    /* The deferral and batching of frees should be suppressed under NOMMU
3846     * conditions.
3847     *
3848     * The problem is that NOMMU needs to be able to allocate large chunks
3849     * of contiguous memory as there's no hardware page translation to
3850     * assemble apparent contiguous memory from discontiguous pages.
3851     *
3852     * Queueing large contiguous runs of pages for batching, however,
3853     * causes the pages to actually be freed in smaller chunks. As there
3854     * can be a significant delay between the individual batches being
3855     * recycled, this leads to the once large chunks of space being
3856     * fragmented and becoming unavailable for high-order allocations.
3857     */
3858    return 0;
3859#endif
3860}
3861
3862static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
3863{
3864    struct per_cpu_pages *pcp;
3865    int migratetype;
3866
3867    memset(p, 0, sizeof(*p));
3868
3869    pcp = &p->pcp;
3870    pcp->count = 0;
3871    pcp->high = 6 * batch;
3872    pcp->batch = max(1UL, 1 * batch);
3873    for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
3874        INIT_LIST_HEAD(&pcp->lists[migratetype]);
3875}
3876
3877/*
3878 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3879 * to the value high for the pageset p.
3880 */
3881
3882static void setup_pagelist_highmark(struct per_cpu_pageset *p,
3883                unsigned long high)
3884{
3885    struct per_cpu_pages *pcp;
3886
3887    pcp = &p->pcp;
3888    pcp->high = high;
3889    pcp->batch = max(1UL, high/4);
3890    if ((high/4) > (PAGE_SHIFT * 8))
3891        pcp->batch = PAGE_SHIFT * 8;
3892}
3893
3894static void __meminit setup_zone_pageset(struct zone *zone)
3895{
3896    int cpu;
3897
3898    zone->pageset = alloc_percpu(struct per_cpu_pageset);
3899
3900    for_each_possible_cpu(cpu) {
3901        struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
3902
3903        setup_pageset(pcp, zone_batchsize(zone));
3904
3905        if (percpu_pagelist_fraction)
3906            setup_pagelist_highmark(pcp,
3907                (zone->present_pages /
3908                    percpu_pagelist_fraction));
3909    }
3910}
3911
3912/*
3913 * Allocate per cpu pagesets and initialize them.
3914 * Before this call only boot pagesets were available.
3915 */
3916void __init setup_per_cpu_pageset(void)
3917{
3918    struct zone *zone;
3919
3920    for_each_populated_zone(zone)
3921        setup_zone_pageset(zone);
3922}
3923
3924static noinline __init_refok
3925int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
3926{
3927    int i;
3928    struct pglist_data *pgdat = zone->zone_pgdat;
3929    size_t alloc_size;
3930
3931    /*
3932     * The per-page waitqueue mechanism uses hashed waitqueues
3933     * per zone.
3934     */
3935    zone->wait_table_hash_nr_entries =
3936         wait_table_hash_nr_entries(zone_size_pages);
3937    zone->wait_table_bits =
3938        wait_table_bits(zone->wait_table_hash_nr_entries);
3939    alloc_size = zone->wait_table_hash_nr_entries
3940                    * sizeof(wait_queue_head_t);
3941
3942    if (!slab_is_available()) {
3943        zone->wait_table = (wait_queue_head_t *)
3944            alloc_bootmem_node_nopanic(pgdat, alloc_size);
3945    } else {
3946        /*
3947         * This case means that a zone whose size was 0 gets new memory
3948         * via memory hot-add.
3949         * But it may be the case that a new node was hot-added. In
3950         * this case vmalloc() will not be able to use this new node's
3951         * memory - this wait_table must be initialized to use this new
3952         * node itself as well.
3953         * To use this new node's memory, further consideration will be
3954         * necessary.
3955         */
3956        zone->wait_table = vmalloc(alloc_size);
3957    }
3958    if (!zone->wait_table)
3959        return -ENOMEM;
3960
3961    for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
3962        init_waitqueue_head(zone->wait_table + i);
3963
3964    return 0;
3965}
3966
3967static __meminit void zone_pcp_init(struct zone *zone)
3968{
3969    /*
3970     * per cpu subsystem is not up at this point. The following code
3971     * relies on the ability of the linker to provide the
3972     * offset of a (static) per cpu variable into the per cpu area.
3973     */
3974    zone->pageset = &boot_pageset;
3975
3976    if (zone->present_pages)
3977        printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
3978            zone->name, zone->present_pages,
3979                     zone_batchsize(zone));
3980}
3981
3982int __meminit init_currently_empty_zone(struct zone *zone,
3983                    unsigned long zone_start_pfn,
3984                    unsigned long size,
3985                    enum memmap_context context)
3986{
3987    struct pglist_data *pgdat = zone->zone_pgdat;
3988    int ret;
3989    ret = zone_wait_table_init(zone, size);
3990    if (ret)
3991        return ret;
3992    pgdat->nr_zones = zone_idx(zone) + 1;
3993
3994    zone->zone_start_pfn = zone_start_pfn;
3995
3996    mminit_dprintk(MMINIT_TRACE, "memmap_init",
3997            "Initialising map node %d zone %lu pfns %lu -> %lu\n",
3998            pgdat->node_id,
3999            (unsigned long)zone_idx(zone),
4000            zone_start_pfn, (zone_start_pfn + size));
4001
4002    zone_init_free_lists(zone);
4003
4004    return 0;
4005}
4006
4007#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4008#ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4009/*
4010 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4011 * Architectures may implement their own version but if add_active_range()
4012 * was used and there are no special requirements, this is a convenient
4013 * alternative
4014 */
4015int __meminit __early_pfn_to_nid(unsigned long pfn)
4016{
4017    unsigned long start_pfn, end_pfn;
4018    int i, nid;
4019
4020    for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
4021        if (start_pfn <= pfn && pfn < end_pfn)
4022            return nid;
4023    /* This is a memory hole */
4024    return -1;
4025}
4026#endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4027
4028int __meminit early_pfn_to_nid(unsigned long pfn)
4029{
4030    int nid;
4031
4032    nid = __early_pfn_to_nid(pfn);
4033    if (nid >= 0)
4034        return nid;
4035    /* just returns 0 */
4036    return 0;
4037}
4038
4039#ifdef CONFIG_NODES_SPAN_OTHER_NODES
4040bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
4041{
4042    int nid;
4043
4044    nid = __early_pfn_to_nid(pfn);
4045    if (nid >= 0 && nid != node)
4046        return false;
4047    return true;
4048}
4049#endif
4050
4051/**
4052 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
4053 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4054 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
4055 *
4056 * If an architecture guarantees that all ranges registered with
4057 * add_active_ranges() contain no holes and may be freed, this
4058 * this function may be used instead of calling free_bootmem() manually.
4059 */
4060void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4061{
4062    unsigned long start_pfn, end_pfn;
4063    int i, this_nid;
4064
4065    for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4066        start_pfn = min(start_pfn, max_low_pfn);
4067        end_pfn = min(end_pfn, max_low_pfn);
4068
4069        if (start_pfn < end_pfn)
4070            free_bootmem_node(NODE_DATA(this_nid),
4071                      PFN_PHYS(start_pfn),
4072                      (end_pfn - start_pfn) << PAGE_SHIFT);
4073    }
4074}
4075
4076/**
4077 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4078 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4079 *
4080 * If an architecture guarantees that all ranges registered with
4081 * add_active_ranges() contain no holes and may be freed, this
4082 * function may be used instead of calling memory_present() manually.
4083 */
4084void __init sparse_memory_present_with_active_regions(int nid)
4085{
4086    unsigned long start_pfn, end_pfn;
4087    int i, this_nid;
4088
4089    for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4090        memory_present(this_nid, start_pfn, end_pfn);
4091}
4092
4093/**
4094 * get_pfn_range_for_nid - Return the start and end page frames for a node
4095 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4096 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4097 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4098 *
4099 * It returns the start and end page frame of a node based on information
4100 * provided by an arch calling add_active_range(). If called for a node
4101 * with no available memory, a warning is printed and the start and end
4102 * PFNs will be 0.
4103 */
4104void __meminit get_pfn_range_for_nid(unsigned int nid,
4105            unsigned long *start_pfn, unsigned long *end_pfn)
4106{
4107    unsigned long this_start_pfn, this_end_pfn;
4108    int i;
4109
4110    *start_pfn = -1UL;
4111    *end_pfn = 0;
4112
4113    for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4114        *start_pfn = min(*start_pfn, this_start_pfn);
4115        *end_pfn = max(*end_pfn, this_end_pfn);
4116    }
4117
4118    if (*start_pfn == -1UL)
4119        *start_pfn = 0;
4120}
4121
4122/*
4123 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4124 * assumption is made that zones within a node are ordered in monotonic
4125 * increasing memory addresses so that the "highest" populated zone is used
4126 */
4127static void __init find_usable_zone_for_movable(void)
4128{
4129    int zone_index;
4130    for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4131        if (zone_index == ZONE_MOVABLE)
4132            continue;
4133
4134        if (arch_zone_highest_possible_pfn[zone_index] >
4135                arch_zone_lowest_possible_pfn[zone_index])
4136            break;
4137    }
4138
4139    VM_BUG_ON(zone_index == -1);
4140    movable_zone = zone_index;
4141}
4142
4143/*
4144 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4145 * because it is sized independent of architecture. Unlike the other zones,
4146 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4147 * in each node depending on the size of each node and how evenly kernelcore
4148 * is distributed. This helper function adjusts the zone ranges
4149 * provided by the architecture for a given node by using the end of the
4150 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4151 * zones within a node are in order of monotonic increases memory addresses
4152 */
4153static void __meminit adjust_zone_range_for_zone_movable(int nid,
4154                    unsigned long zone_type,
4155                    unsigned long node_start_pfn,
4156                    unsigned long node_end_pfn,
4157                    unsigned long *zone_start_pfn,
4158                    unsigned long *zone_end_pfn)
4159{
4160    /* Only adjust if ZONE_MOVABLE is on this node */
4161    if (zone_movable_pfn[nid]) {
4162        /* Size ZONE_MOVABLE */
4163        if (zone_type == ZONE_MOVABLE) {
4164            *zone_start_pfn = zone_movable_pfn[nid];
4165            *zone_end_pfn = min(node_end_pfn,
4166                arch_zone_highest_possible_pfn[movable_zone]);
4167
4168        /* Adjust for ZONE_MOVABLE starting within this range */
4169        } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4170                *zone_end_pfn > zone_movable_pfn[nid]) {
4171            *zone_end_pfn = zone_movable_pfn[nid];
4172
4173        /* Check if this whole range is within ZONE_MOVABLE */
4174        } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4175            *zone_start_pfn = *zone_end_pfn;
4176    }
4177}
4178
4179/*
4180 * Return the number of pages a zone spans in a node, including holes
4181 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4182 */
4183static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4184                    unsigned long zone_type,
4185                    unsigned long *ignored)
4186{
4187    unsigned long node_start_pfn, node_end_pfn;
4188    unsigned long zone_start_pfn, zone_end_pfn;
4189
4190    /* Get the start and end of the node and zone */
4191    get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4192    zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4193    zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4194    adjust_zone_range_for_zone_movable(nid, zone_type,
4195                node_start_pfn, node_end_pfn,
4196                &zone_start_pfn, &zone_end_pfn);
4197
4198    /* Check that this node has pages within the zone's required range */
4199    if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4200        return 0;
4201
4202    /* Move the zone boundaries inside the node if necessary */
4203    zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4204    zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4205
4206    /* Return the spanned pages */
4207    return zone_end_pfn - zone_start_pfn;
4208}
4209
4210/*
4211 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4212 * then all holes in the requested range will be accounted for.
4213 */
4214unsigned long __meminit __absent_pages_in_range(int nid,
4215                unsigned long range_start_pfn,
4216                unsigned long range_end_pfn)
4217{
4218    unsigned long nr_absent = range_end_pfn - range_start_pfn;
4219    unsigned long start_pfn, end_pfn;
4220    int i;
4221
4222    for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4223        start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4224        end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4225        nr_absent -= end_pfn - start_pfn;
4226    }
4227    return nr_absent;
4228}
4229
4230/**
4231 * absent_pages_in_range - Return number of page frames in holes within a range
4232 * @start_pfn: The start PFN to start searching for holes
4233 * @end_pfn: The end PFN to stop searching for holes
4234 *
4235 * It returns the number of pages frames in memory holes within a range.
4236 */
4237unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4238                            unsigned long end_pfn)
4239{
4240    return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4241}
4242
4243/* Return the number of page frames in holes in a zone on a node */
4244static unsigned long __meminit zone_absent_pages_in_node(int nid,
4245                    unsigned long zone_type,
4246                    unsigned long *ignored)
4247{
4248    unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
4249    unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
4250    unsigned long node_start_pfn, node_end_pfn;
4251    unsigned long zone_start_pfn, zone_end_pfn;
4252
4253    get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4254    zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
4255    zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
4256
4257    adjust_zone_range_for_zone_movable(nid, zone_type,
4258            node_start_pfn, node_end_pfn,
4259            &zone_start_pfn, &zone_end_pfn);
4260    return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4261}
4262
4263#else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4264static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4265                    unsigned long zone_type,
4266                    unsigned long *zones_size)
4267{
4268    return zones_size[zone_type];
4269}
4270
4271static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4272                        unsigned long zone_type,
4273                        unsigned long *zholes_size)
4274{
4275    if (!zholes_size)
4276        return 0;
4277
4278    return zholes_size[zone_type];
4279}
4280
4281#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4282
4283static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4284        unsigned long *zones_size, unsigned long *zholes_size)
4285{
4286    unsigned long realtotalpages, totalpages = 0;
4287    enum zone_type i;
4288
4289    for (i = 0; i < MAX_NR_ZONES; i++)
4290        totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4291                                zones_size);
4292    pgdat->node_spanned_pages = totalpages;
4293
4294    realtotalpages = totalpages;
4295    for (i = 0; i < MAX_NR_ZONES; i++)
4296        realtotalpages -=
4297            zone_absent_pages_in_node(pgdat->node_id, i,
4298                                zholes_size);
4299    pgdat->node_present_pages = realtotalpages;
4300    printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4301                            realtotalpages);
4302}
4303
4304#ifndef CONFIG_SPARSEMEM
4305/*
4306 * Calculate the size of the zone->blockflags rounded to an unsigned long
4307 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4308 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4309 * round what is now in bits to nearest long in bits, then return it in
4310 * bytes.
4311 */
4312static unsigned long __init usemap_size(unsigned long zonesize)
4313{
4314    unsigned long usemapsize;
4315
4316    usemapsize = roundup(zonesize, pageblock_nr_pages);
4317    usemapsize = usemapsize >> pageblock_order;
4318    usemapsize *= NR_PAGEBLOCK_BITS;
4319    usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4320
4321    return usemapsize / 8;
4322}
4323
4324static void __init setup_usemap(struct pglist_data *pgdat,
4325                struct zone *zone, unsigned long zonesize)
4326{
4327    unsigned long usemapsize = usemap_size(zonesize);
4328    zone->pageblock_flags = NULL;
4329    if (usemapsize)
4330        zone->pageblock_flags = alloc_bootmem_node_nopanic(pgdat,
4331                                   usemapsize);
4332}
4333#else
4334static inline void setup_usemap(struct pglist_data *pgdat,
4335                struct zone *zone, unsigned long zonesize) {}
4336#endif /* CONFIG_SPARSEMEM */
4337
4338#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4339
4340/* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4341void __init set_pageblock_order(void)
4342{
4343    unsigned int order;
4344
4345    /* Check that pageblock_nr_pages has not already been setup */
4346    if (pageblock_order)
4347        return;
4348
4349    if (HPAGE_SHIFT > PAGE_SHIFT)
4350        order = HUGETLB_PAGE_ORDER;
4351    else
4352        order = MAX_ORDER - 1;
4353
4354    /*
4355     * Assume the largest contiguous order of interest is a huge page.
4356     * This value may be variable depending on boot parameters on IA64 and
4357     * powerpc.
4358     */
4359    pageblock_order = order;
4360}
4361#else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4362
4363/*
4364 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4365 * is unused as pageblock_order is set at compile-time. See
4366 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4367 * the kernel config
4368 */
4369void __init set_pageblock_order(void)
4370{
4371}
4372
4373#endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4374
4375/*
4376 * Set up the zone data structures:
4377 * - mark all pages reserved
4378 * - mark all memory queues empty
4379 * - clear the memory bitmaps
4380 *
4381 * NOTE: pgdat should get zeroed by caller.
4382 */
4383static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4384        unsigned long *zones_size, unsigned long *zholes_size)
4385{
4386    enum zone_type j;
4387    int nid = pgdat->node_id;
4388    unsigned long zone_start_pfn = pgdat->node_start_pfn;
4389    int ret;
4390
4391    pgdat_resize_init(pgdat);
4392    init_waitqueue_head(&pgdat->kswapd_wait);
4393    init_waitqueue_head(&pgdat->pfmemalloc_wait);
4394    pgdat_page_cgroup_init(pgdat);
4395
4396    for (j = 0; j < MAX_NR_ZONES; j++) {
4397        struct zone *zone = pgdat->node_zones + j;
4398        unsigned long size, realsize, memmap_pages;
4399
4400        size = zone_spanned_pages_in_node(nid, j, zones_size);
4401        realsize = size - zone_absent_pages_in_node(nid, j,
4402                                zholes_size);
4403
4404        /*
4405         * Adjust realsize so that it accounts for how much memory
4406         * is used by this zone for memmap. This affects the watermark
4407         * and per-cpu initialisations
4408         */
4409        memmap_pages =
4410            PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT;
4411        if (realsize >= memmap_pages) {
4412            realsize -= memmap_pages;
4413            if (memmap_pages)
4414                printk(KERN_DEBUG
4415                       " %s zone: %lu pages used for memmap\n",
4416                       zone_names[j], memmap_pages);
4417        } else
4418            printk(KERN_WARNING
4419                " %s zone: %lu pages exceeds realsize %lu\n",
4420                zone_names[j], memmap_pages, realsize);
4421
4422        /* Account for reserved pages */
4423        if (j == 0 && realsize > dma_reserve) {
4424            realsize -= dma_reserve;
4425            printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4426                    zone_names[0], dma_reserve);
4427        }
4428
4429        if (!is_highmem_idx(j))
4430            nr_kernel_pages += realsize;
4431        nr_all_pages += realsize;
4432
4433        zone->spanned_pages = size;
4434        zone->present_pages = realsize;
4435#if defined CONFIG_COMPACTION || defined CONFIG_CMA
4436        zone->compact_cached_free_pfn = zone->zone_start_pfn +
4437                        zone->spanned_pages;
4438        zone->compact_cached_free_pfn &= ~(pageblock_nr_pages-1);
4439#endif
4440#ifdef CONFIG_NUMA
4441        zone->node = nid;
4442        zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
4443                        / 100;
4444        zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
4445#endif
4446        zone->name = zone_names[j];
4447        spin_lock_init(&zone->lock);
4448        spin_lock_init(&zone->lru_lock);
4449        zone_seqlock_init(zone);
4450        zone->zone_pgdat = pgdat;
4451
4452        zone_pcp_init(zone);
4453        lruvec_init(&zone->lruvec, zone);
4454        if (!size)
4455            continue;
4456
4457        set_pageblock_order();
4458        setup_usemap(pgdat, zone, size);
4459        ret = init_currently_empty_zone(zone, zone_start_pfn,
4460                        size, MEMMAP_EARLY);
4461        BUG_ON(ret);
4462        memmap_init(size, nid, j, zone_start_pfn);
4463        zone_start_pfn += size;
4464    }
4465}
4466
4467static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4468{
4469    /* Skip empty nodes */
4470    if (!pgdat->node_spanned_pages)
4471        return;
4472
4473#ifdef CONFIG_FLAT_NODE_MEM_MAP
4474    /* ia64 gets its own node_mem_map, before this, without bootmem */
4475    if (!pgdat->node_mem_map) {
4476        unsigned long size, start, end;
4477        struct page *map;
4478
4479        /*
4480         * The zone's endpoints aren't required to be MAX_ORDER
4481         * aligned but the node_mem_map endpoints must be in order
4482         * for the buddy allocator to function correctly.
4483         */
4484        start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4485        end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
4486        end = ALIGN(end, MAX_ORDER_NR_PAGES);
4487        size = (end - start) * sizeof(struct page);
4488        map = alloc_remap(pgdat->node_id, size);
4489        if (!map)
4490            map = alloc_bootmem_node_nopanic(pgdat, size);
4491        pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4492    }
4493#ifndef CONFIG_NEED_MULTIPLE_NODES
4494    /*
4495     * With no DISCONTIG, the global mem_map is just set as node 0's
4496     */
4497    if (pgdat == NODE_DATA(0)) {
4498        mem_map = NODE_DATA(0)->node_mem_map;
4499#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4500        if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4501            mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4502#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4503    }
4504#endif
4505#endif /* CONFIG_FLAT_NODE_MEM_MAP */
4506}
4507
4508void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4509        unsigned long node_start_pfn, unsigned long *zholes_size)
4510{
4511    pg_data_t *pgdat = NODE_DATA(nid);
4512
4513    /* pg_data_t should be reset to zero when it's allocated */
4514    WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
4515
4516    pgdat->node_id = nid;
4517    pgdat->node_start_pfn = node_start_pfn;
4518    calculate_node_totalpages(pgdat, zones_size, zholes_size);
4519
4520    alloc_node_mem_map(pgdat);
4521#ifdef CONFIG_FLAT_NODE_MEM_MAP
4522    printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4523        nid, (unsigned long)pgdat,
4524        (unsigned long)pgdat->node_mem_map);
4525#endif
4526
4527    free_area_init_core(pgdat, zones_size, zholes_size);
4528}
4529
4530#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4531
4532#if MAX_NUMNODES > 1
4533/*
4534 * Figure out the number of possible node ids.
4535 */
4536static void __init setup_nr_node_ids(void)
4537{
4538    unsigned int node;
4539    unsigned int highest = 0;
4540
4541    for_each_node_mask(node, node_possible_map)
4542        highest = node;
4543    nr_node_ids = highest + 1;
4544}
4545#else
4546static inline void setup_nr_node_ids(void)
4547{
4548}
4549#endif
4550
4551/**
4552 * node_map_pfn_alignment - determine the maximum internode alignment
4553 *
4554 * This function should be called after node map is populated and sorted.
4555 * It calculates the maximum power of two alignment which can distinguish
4556 * all the nodes.
4557 *
4558 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4559 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
4560 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
4561 * shifted, 1GiB is enough and this function will indicate so.
4562 *
4563 * This is used to test whether pfn -> nid mapping of the chosen memory
4564 * model has fine enough granularity to avoid incorrect mapping for the
4565 * populated node map.
4566 *
4567 * Returns the determined alignment in pfn's. 0 if there is no alignment
4568 * requirement (single node).
4569 */
4570unsigned long __init node_map_pfn_alignment(void)
4571{
4572    unsigned long accl_mask = 0, last_end = 0;
4573    unsigned long start, end, mask;
4574    int last_nid = -1;
4575    int i, nid;
4576
4577    for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
4578        if (!start || last_nid < 0 || last_nid == nid) {
4579            last_nid = nid;
4580            last_end = end;
4581            continue;
4582        }
4583
4584        /*
4585         * Start with a mask granular enough to pin-point to the
4586         * start pfn and tick off bits one-by-one until it becomes
4587         * too coarse to separate the current node from the last.
4588         */
4589        mask = ~((1 << __ffs(start)) - 1);
4590        while (mask && last_end <= (start & (mask << 1)))
4591            mask <<= 1;
4592
4593        /* accumulate all internode masks */
4594        accl_mask |= mask;
4595    }
4596
4597    /* convert mask to number of pages */
4598    return ~accl_mask + 1;
4599}
4600
4601/* Find the lowest pfn for a node */
4602static unsigned long __init find_min_pfn_for_node(int nid)
4603{
4604    unsigned long min_pfn = ULONG_MAX;
4605    unsigned long start_pfn;
4606    int i;
4607
4608    for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
4609        min_pfn = min(min_pfn, start_pfn);
4610
4611    if (min_pfn == ULONG_MAX) {
4612        printk(KERN_WARNING
4613            "Could not find start_pfn for node %d\n", nid);
4614        return 0;
4615    }
4616
4617    return min_pfn;
4618}
4619
4620/**
4621 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4622 *
4623 * It returns the minimum PFN based on information provided via
4624 * add_active_range().
4625 */
4626unsigned long __init find_min_pfn_with_active_regions(void)
4627{
4628    return find_min_pfn_for_node(MAX_NUMNODES);
4629}
4630
4631/*
4632 * early_calculate_totalpages()
4633 * Sum pages in active regions for movable zone.
4634 * Populate N_HIGH_MEMORY for calculating usable_nodes.
4635 */
4636static unsigned long __init early_calculate_totalpages(void)
4637{
4638    unsigned long totalpages = 0;
4639    unsigned long start_pfn, end_pfn;
4640    int i, nid;
4641
4642    for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
4643        unsigned long pages = end_pfn - start_pfn;
4644
4645        totalpages += pages;
4646        if (pages)
4647            node_set_state(nid, N_HIGH_MEMORY);
4648    }
4649      return totalpages;
4650}
4651
4652/*
4653 * Find the PFN the Movable zone begins in each node. Kernel memory
4654 * is spread evenly between nodes as long as the nodes have enough
4655 * memory. When they don't, some nodes will have more kernelcore than
4656 * others
4657 */
4658static void __init find_zone_movable_pfns_for_nodes(void)
4659{
4660    int i, nid;
4661    unsigned long usable_startpfn;
4662    unsigned long kernelcore_node, kernelcore_remaining;
4663    /* save the state before borrow the nodemask */
4664    nodemask_t saved_node_state = node_states[N_HIGH_MEMORY];
4665    unsigned long totalpages = early_calculate_totalpages();
4666    int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
4667
4668    /*
4669     * If movablecore was specified, calculate what size of
4670     * kernelcore that corresponds so that memory usable for
4671     * any allocation type is evenly spread. If both kernelcore
4672     * and movablecore are specified, then the value of kernelcore
4673     * will be used for required_kernelcore if it's greater than
4674     * what movablecore would have allowed.
4675     */
4676    if (required_movablecore) {
4677        unsigned long corepages;
4678
4679        /*
4680         * Round-up so that ZONE_MOVABLE is at least as large as what
4681         * was requested by the user
4682         */
4683        required_movablecore =
4684            roundup(required_movablecore, MAX_ORDER_NR_PAGES);
4685        corepages = totalpages - required_movablecore;
4686
4687        required_kernelcore = max(required_kernelcore, corepages);
4688    }
4689
4690    /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4691    if (!required_kernelcore)
4692        goto out;
4693
4694    /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4695    find_usable_zone_for_movable();
4696    usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
4697
4698restart:
4699    /* Spread kernelcore memory as evenly as possible throughout nodes */
4700    kernelcore_node = required_kernelcore / usable_nodes;
4701    for_each_node_state(nid, N_HIGH_MEMORY) {
4702        unsigned long start_pfn, end_pfn;
4703
4704        /*
4705         * Recalculate kernelcore_node if the division per node
4706         * now exceeds what is necessary to satisfy the requested
4707         * amount of memory for the kernel
4708         */
4709        if (required_kernelcore < kernelcore_node)
4710            kernelcore_node = required_kernelcore / usable_nodes;
4711
4712        /*
4713         * As the map is walked, we track how much memory is usable
4714         * by the kernel using kernelcore_remaining. When it is
4715         * 0, the rest of the node is usable by ZONE_MOVABLE
4716         */
4717        kernelcore_remaining = kernelcore_node;
4718
4719        /* Go through each range of PFNs within this node */
4720        for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4721            unsigned long size_pages;
4722
4723            start_pfn = max(start_pfn, zone_movable_pfn[nid]);
4724            if (start_pfn >= end_pfn)
4725                continue;
4726
4727            /* Account for what is only usable for kernelcore */
4728            if (start_pfn < usable_startpfn) {
4729                unsigned long kernel_pages;
4730                kernel_pages = min(end_pfn, usable_startpfn)
4731                                - start_pfn;
4732
4733                kernelcore_remaining -= min(kernel_pages,
4734                            kernelcore_remaining);
4735                required_kernelcore -= min(kernel_pages,
4736                            required_kernelcore);
4737
4738                /* Continue if range is now fully accounted */
4739                if (end_pfn <= usable_startpfn) {
4740
4741                    /*
4742                     * Push zone_movable_pfn to the end so
4743                     * that if we have to rebalance
4744                     * kernelcore across nodes, we will
4745                     * not double account here
4746                     */
4747                    zone_movable_pfn[nid] = end_pfn;
4748                    continue;
4749                }
4750                start_pfn = usable_startpfn;
4751            }
4752
4753            /*
4754             * The usable PFN range for ZONE_MOVABLE is from
4755             * start_pfn->end_pfn. Calculate size_pages as the
4756             * number of pages used as kernelcore
4757             */
4758            size_pages = end_pfn - start_pfn;
4759            if (size_pages > kernelcore_remaining)
4760                size_pages = kernelcore_remaining;
4761            zone_movable_pfn[nid] = start_pfn + size_pages;
4762
4763            /*
4764             * Some kernelcore has been met, update counts and
4765             * break if the kernelcore for this node has been
4766             * satisified
4767             */
4768            required_kernelcore -= min(required_kernelcore,
4769                                size_pages);
4770            kernelcore_remaining -= size_pages;
4771            if (!kernelcore_remaining)
4772                break;
4773        }
4774    }
4775
4776    /*
4777     * If there is still required_kernelcore, we do another pass with one
4778     * less node in the count. This will push zone_movable_pfn[nid] further
4779     * along on the nodes that still have memory until kernelcore is
4780     * satisified
4781     */
4782    usable_nodes--;
4783    if (usable_nodes && required_kernelcore > usable_nodes)
4784        goto restart;
4785
4786    /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4787    for (nid = 0; nid < MAX_NUMNODES; nid++)
4788        zone_movable_pfn[nid] =
4789            roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
4790
4791out:
4792    /* restore the node_state */
4793    node_states[N_HIGH_MEMORY] = saved_node_state;
4794}
4795
4796/* Any regular memory on that node ? */
4797static void __init check_for_regular_memory(pg_data_t *pgdat)
4798{
4799#ifdef CONFIG_HIGHMEM
4800    enum zone_type zone_type;
4801
4802    for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) {
4803        struct zone *zone = &pgdat->node_zones[zone_type];
4804        if (zone->present_pages) {
4805            node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY);
4806            break;
4807        }
4808    }
4809#endif
4810}
4811
4812/**
4813 * free_area_init_nodes - Initialise all pg_data_t and zone data
4814 * @max_zone_pfn: an array of max PFNs for each zone
4815 *
4816 * This will call free_area_init_node() for each active node in the system.
4817 * Using the page ranges provided by add_active_range(), the size of each
4818 * zone in each node and their holes is calculated. If the maximum PFN
4819 * between two adjacent zones match, it is assumed that the zone is empty.
4820 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4821 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4822 * starts where the previous one ended. For example, ZONE_DMA32 starts
4823 * at arch_max_dma_pfn.
4824 */
4825void __init free_area_init_nodes(unsigned long *max_zone_pfn)
4826{
4827    unsigned long start_pfn, end_pfn;
4828    int i, nid;
4829
4830    /* Record where the zone boundaries are */
4831    memset(arch_zone_lowest_possible_pfn, 0,
4832                sizeof(arch_zone_lowest_possible_pfn));
4833    memset(arch_zone_highest_possible_pfn, 0,
4834                sizeof(arch_zone_highest_possible_pfn));
4835    arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
4836    arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
4837    for (i = 1; i < MAX_NR_ZONES; i++) {
4838        if (i == ZONE_MOVABLE)
4839            continue;
4840        arch_zone_lowest_possible_pfn[i] =
4841            arch_zone_highest_possible_pfn[i-1];
4842        arch_zone_highest_possible_pfn[i] =
4843            max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
4844    }
4845    arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
4846    arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
4847
4848    /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4849    memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
4850    find_zone_movable_pfns_for_nodes();
4851
4852    /* Print out the zone ranges */
4853    printk("Zone ranges:\n");
4854    for (i = 0; i < MAX_NR_ZONES; i++) {
4855        if (i == ZONE_MOVABLE)
4856            continue;
4857        printk(KERN_CONT " %-8s ", zone_names[i]);
4858        if (arch_zone_lowest_possible_pfn[i] ==
4859                arch_zone_highest_possible_pfn[i])
4860            printk(KERN_CONT "empty\n");
4861        else
4862            printk(KERN_CONT "[mem %0#10lx-%0#10lx]\n",
4863                arch_zone_lowest_possible_pfn[i] << PAGE_SHIFT,
4864                (arch_zone_highest_possible_pfn[i]
4865                    << PAGE_SHIFT) - 1);
4866    }
4867
4868    /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4869    printk("Movable zone start for each node\n");
4870    for (i = 0; i < MAX_NUMNODES; i++) {
4871        if (zone_movable_pfn[i])
4872            printk(" Node %d: %#010lx\n", i,
4873                   zone_movable_pfn[i] << PAGE_SHIFT);
4874    }
4875
4876    /* Print out the early_node_map[] */
4877    printk("Early memory node ranges\n");
4878    for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
4879        printk(" node %3d: [mem %#010lx-%#010lx]\n", nid,
4880               start_pfn << PAGE_SHIFT, (end_pfn << PAGE_SHIFT) - 1);
4881
4882    /* Initialise every node */
4883    mminit_verify_pageflags_layout();
4884    setup_nr_node_ids();
4885    for_each_online_node(nid) {
4886        pg_data_t *pgdat = NODE_DATA(nid);
4887        free_area_init_node(nid, NULL,
4888                find_min_pfn_for_node(nid), NULL);
4889
4890        /* Any memory on that node */
4891        if (pgdat->node_present_pages)
4892            node_set_state(nid, N_HIGH_MEMORY);
4893        check_for_regular_memory(pgdat);
4894    }
4895}
4896
4897static int __init cmdline_parse_core(char *p, unsigned long *core)
4898{
4899    unsigned long long coremem;
4900    if (!p)
4901        return -EINVAL;
4902
4903    coremem = memparse(p, &p);
4904    *core = coremem >> PAGE_SHIFT;
4905
4906    /* Paranoid check that UL is enough for the coremem value */
4907    WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
4908
4909    return 0;
4910}
4911
4912/*
4913 * kernelcore=size sets the amount of memory for use for allocations that
4914 * cannot be reclaimed or migrated.
4915 */
4916static int __init cmdline_parse_kernelcore(char *p)
4917{
4918    return cmdline_parse_core(p, &required_kernelcore);
4919}
4920
4921/*
4922 * movablecore=size sets the amount of memory for use for allocations that
4923 * can be reclaimed or migrated.
4924 */
4925static int __init cmdline_parse_movablecore(char *p)
4926{
4927    return cmdline_parse_core(p, &required_movablecore);
4928}
4929
4930early_param("kernelcore", cmdline_parse_kernelcore);
4931early_param("movablecore", cmdline_parse_movablecore);
4932
4933#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4934
4935/**
4936 * set_dma_reserve - set the specified number of pages reserved in the first zone
4937 * @new_dma_reserve: The number of pages to mark reserved
4938 *
4939 * The per-cpu batchsize and zone watermarks are determined by present_pages.
4940 * In the DMA zone, a significant percentage may be consumed by kernel image
4941 * and other unfreeable allocations which can skew the watermarks badly. This
4942 * function may optionally be used to account for unfreeable pages in the
4943 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
4944 * smaller per-cpu batchsize.
4945 */
4946void __init set_dma_reserve(unsigned long new_dma_reserve)
4947{
4948    dma_reserve = new_dma_reserve;
4949}
4950
4951void __init free_area_init(unsigned long *zones_size)
4952{
4953    free_area_init_node(0, zones_size,
4954            __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
4955}
4956
4957static int page_alloc_cpu_notify(struct notifier_block *self,
4958                 unsigned long action, void *hcpu)
4959{
4960    int cpu = (unsigned long)hcpu;
4961
4962    if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
4963        lru_add_drain_cpu(cpu);
4964        drain_pages(cpu);
4965
4966        /*
4967         * Spill the event counters of the dead processor
4968         * into the current processors event counters.
4969         * This artificially elevates the count of the current
4970         * processor.
4971         */
4972        vm_events_fold_cpu(cpu);
4973
4974        /*
4975         * Zero the differential counters of the dead processor
4976         * so that the vm statistics are consistent.
4977         *
4978         * This is only okay since the processor is dead and cannot
4979         * race with what we are doing.
4980         */
4981        refresh_cpu_vm_stats(cpu);
4982    }
4983    return NOTIFY_OK;
4984}
4985
4986void __init page_alloc_init(void)
4987{
4988    hotcpu_notifier(page_alloc_cpu_notify, 0);
4989}
4990
4991/*
4992 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
4993 * or min_free_kbytes changes.
4994 */
4995static void calculate_totalreserve_pages(void)
4996{
4997    struct pglist_data *pgdat;
4998    unsigned long reserve_pages = 0;
4999    enum zone_type i, j;
5000
5001    for_each_online_pgdat(pgdat) {
5002        for (i = 0; i < MAX_NR_ZONES; i++) {
5003            struct zone *zone = pgdat->node_zones + i;
5004            unsigned long max = 0;
5005
5006            /* Find valid and maximum lowmem_reserve in the zone */
5007            for (j = i; j < MAX_NR_ZONES; j++) {
5008                if (zone->lowmem_reserve[j] > max)
5009                    max = zone->lowmem_reserve[j];
5010            }
5011
5012            /* we treat the high watermark as reserved pages. */
5013            max += high_wmark_pages(zone);
5014
5015            if (max > zone->present_pages)
5016                max = zone->present_pages;
5017            reserve_pages += max;
5018            /*
5019             * Lowmem reserves are not available to
5020             * GFP_HIGHUSER page cache allocations and
5021             * kswapd tries to balance zones to their high
5022             * watermark. As a result, neither should be
5023             * regarded as dirtyable memory, to prevent a
5024             * situation where reclaim has to clean pages
5025             * in order to balance the zones.
5026             */
5027            zone->dirty_balance_reserve = max;
5028        }
5029    }
5030    dirty_balance_reserve = reserve_pages;
5031    totalreserve_pages = reserve_pages;
5032}
5033
5034/*
5035 * setup_per_zone_lowmem_reserve - called whenever
5036 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5037 * has a correct pages reserved value, so an adequate number of
5038 * pages are left in the zone after a successful __alloc_pages().
5039 */
5040static void setup_per_zone_lowmem_reserve(void)
5041{
5042    struct pglist_data *pgdat;
5043    enum zone_type j, idx;
5044
5045    for_each_online_pgdat(pgdat) {
5046        for (j = 0; j < MAX_NR_ZONES; j++) {
5047            struct zone *zone = pgdat->node_zones + j;
5048            unsigned long present_pages = zone->present_pages;
5049
5050            zone->lowmem_reserve[j] = 0;
5051
5052            idx = j;
5053            while (idx) {
5054                struct zone *lower_zone;
5055
5056                idx--;
5057
5058                if (sysctl_lowmem_reserve_ratio[idx] < 1)
5059                    sysctl_lowmem_reserve_ratio[idx] = 1;
5060
5061                lower_zone = pgdat->node_zones + idx;
5062                lower_zone->lowmem_reserve[j] = present_pages /
5063                    sysctl_lowmem_reserve_ratio[idx];
5064                present_pages += lower_zone->present_pages;
5065            }
5066        }
5067    }
5068
5069    /* update totalreserve_pages */
5070    calculate_totalreserve_pages();
5071}
5072
5073static void __setup_per_zone_wmarks(void)
5074{
5075    unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5076    unsigned long lowmem_pages = 0;
5077    struct zone *zone;
5078    unsigned long flags;
5079
5080    /* Calculate total number of !ZONE_HIGHMEM pages */
5081    for_each_zone(zone) {
5082        if (!is_highmem(zone))
5083            lowmem_pages += zone->present_pages;
5084    }
5085
5086    for_each_zone(zone) {
5087        u64 tmp;
5088
5089        spin_lock_irqsave(&zone->lock, flags);
5090        tmp = (u64)pages_min * zone->present_pages;
5091        do_div(tmp, lowmem_pages);
5092        if (is_highmem(zone)) {
5093            /*
5094             * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5095             * need highmem pages, so cap pages_min to a small
5096             * value here.
5097             *
5098             * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5099             * deltas controls asynch page reclaim, and so should
5100             * not be capped for highmem.
5101             */
5102            int min_pages;
5103
5104            min_pages = zone->present_pages / 1024;
5105            if (min_pages < SWAP_CLUSTER_MAX)
5106                min_pages = SWAP_CLUSTER_MAX;
5107            if (min_pages > 128)
5108                min_pages = 128;
5109            zone->watermark[WMARK_MIN] = min_pages;
5110        } else {
5111            /*
5112             * If it's a lowmem zone, reserve a number of pages
5113             * proportionate to the zone's size.
5114             */
5115            zone->watermark[WMARK_MIN] = tmp;
5116        }
5117
5118        zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
5119        zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5120
5121        zone->watermark[WMARK_MIN] += cma_wmark_pages(zone);
5122        zone->watermark[WMARK_LOW] += cma_wmark_pages(zone);
5123        zone->watermark[WMARK_HIGH] += cma_wmark_pages(zone);
5124
5125        setup_zone_migrate_reserve(zone);
5126        spin_unlock_irqrestore(&zone->lock, flags);
5127    }
5128
5129    /* update totalreserve_pages */
5130    calculate_totalreserve_pages();
5131}
5132
5133/**
5134 * setup_per_zone_wmarks - called when min_free_kbytes changes
5135 * or when memory is hot-{added|removed}
5136 *
5137 * Ensures that the watermark[min,low,high] values for each zone are set
5138 * correctly with respect to min_free_kbytes.
5139 */
5140void setup_per_zone_wmarks(void)
5141{
5142    mutex_lock(&zonelists_mutex);
5143    __setup_per_zone_wmarks();
5144    mutex_unlock(&zonelists_mutex);
5145}
5146
5147/*
5148 * The inactive anon list should be small enough that the VM never has to
5149 * do too much work, but large enough that each inactive page has a chance
5150 * to be referenced again before it is swapped out.
5151 *
5152 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5153 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5154 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5155 * the anonymous pages are kept on the inactive list.
5156 *
5157 * total target max
5158 * memory ratio inactive anon
5159 * -------------------------------------
5160 * 10MB 1 5MB
5161 * 100MB 1 50MB
5162 * 1GB 3 250MB
5163 * 10GB 10 0.9GB
5164 * 100GB 31 3GB
5165 * 1TB 101 10GB
5166 * 10TB 320 32GB
5167 */
5168static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5169{
5170    unsigned int gb, ratio;
5171
5172    /* Zone size in gigabytes */
5173    gb = zone->present_pages >> (30 - PAGE_SHIFT);
5174    if (gb)
5175        ratio = int_sqrt(10 * gb);
5176    else
5177        ratio = 1;
5178
5179    zone->inactive_ratio = ratio;
5180}
5181
5182static void __meminit setup_per_zone_inactive_ratio(void)
5183{
5184    struct zone *zone;
5185
5186    for_each_zone(zone)
5187        calculate_zone_inactive_ratio(zone);
5188}
5189
5190/*
5191 * Initialise min_free_kbytes.
5192 *
5193 * For small machines we want it small (128k min). For large machines
5194 * we want it large (64MB max). But it is not linear, because network
5195 * bandwidth does not increase linearly with machine size. We use
5196 *
5197 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5198 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5199 *
5200 * which yields
5201 *
5202 * 16MB: 512k
5203 * 32MB: 724k
5204 * 64MB: 1024k
5205 * 128MB: 1448k
5206 * 256MB: 2048k
5207 * 512MB: 2896k
5208 * 1024MB: 4096k
5209 * 2048MB: 5792k
5210 * 4096MB: 8192k
5211 * 8192MB: 11584k
5212 * 16384MB: 16384k
5213 */
5214int __meminit init_per_zone_wmark_min(void)
5215{
5216    unsigned long lowmem_kbytes;
5217
5218    lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5219
5220    min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5221    if (min_free_kbytes < 128)
5222        min_free_kbytes = 128;
5223    if (min_free_kbytes > 65536)
5224        min_free_kbytes = 65536;
5225    setup_per_zone_wmarks();
5226    refresh_zone_stat_thresholds();
5227    setup_per_zone_lowmem_reserve();
5228    setup_per_zone_inactive_ratio();
5229    return 0;
5230}
5231module_init(init_per_zone_wmark_min)
5232
5233/*
5234 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5235 * that we can call two helper functions whenever min_free_kbytes
5236 * changes.
5237 */
5238int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
5239    void __user *buffer, size_t *length, loff_t *ppos)
5240{
5241    proc_dointvec(table, write, buffer, length, ppos);
5242    if (write)
5243        setup_per_zone_wmarks();
5244    return 0;
5245}
5246
5247#ifdef CONFIG_NUMA
5248int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
5249    void __user *buffer, size_t *length, loff_t *ppos)
5250{
5251    struct zone *zone;
5252    int rc;
5253
5254    rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5255    if (rc)
5256        return rc;
5257
5258    for_each_zone(zone)
5259        zone->min_unmapped_pages = (zone->present_pages *
5260                sysctl_min_unmapped_ratio) / 100;
5261    return 0;
5262}
5263
5264int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
5265    void __user *buffer, size_t *length, loff_t *ppos)
5266{
5267    struct zone *zone;
5268    int rc;
5269
5270    rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5271    if (rc)
5272        return rc;
5273
5274    for_each_zone(zone)
5275        zone->min_slab_pages = (zone->present_pages *
5276                sysctl_min_slab_ratio) / 100;
5277    return 0;
5278}
5279#endif
5280
5281/*
5282 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5283 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5284 * whenever sysctl_lowmem_reserve_ratio changes.
5285 *
5286 * The reserve ratio obviously has absolutely no relation with the
5287 * minimum watermarks. The lowmem reserve ratio can only make sense
5288 * if in function of the boot time zone sizes.
5289 */
5290int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
5291    void __user *buffer, size_t *length, loff_t *ppos)
5292{
5293    proc_dointvec_minmax(table, write, buffer, length, ppos);
5294    setup_per_zone_lowmem_reserve();
5295    return 0;
5296}
5297
5298/*
5299 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5300 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
5301 * can have before it gets flushed back to buddy allocator.
5302 */
5303
5304int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
5305    void __user *buffer, size_t *length, loff_t *ppos)
5306{
5307    struct zone *zone;
5308    unsigned int cpu;
5309    int ret;
5310
5311    ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5312    if (!write || (ret < 0))
5313        return ret;
5314    for_each_populated_zone(zone) {
5315        for_each_possible_cpu(cpu) {
5316            unsigned long high;
5317            high = zone->present_pages / percpu_pagelist_fraction;
5318            setup_pagelist_highmark(
5319                per_cpu_ptr(zone->pageset, cpu), high);
5320        }
5321    }
5322    return 0;
5323}
5324
5325int hashdist = HASHDIST_DEFAULT;
5326
5327#ifdef CONFIG_NUMA
5328static int __init set_hashdist(char *str)
5329{
5330    if (!str)
5331        return 0;
5332    hashdist = simple_strtoul(str, &str, 0);
5333    return 1;
5334}
5335__setup("hashdist=", set_hashdist);
5336#endif
5337
5338/*
5339 * allocate a large system hash table from bootmem
5340 * - it is assumed that the hash table must contain an exact power-of-2
5341 * quantity of entries
5342 * - limit is the number of hash buckets, not the total allocation size
5343 */
5344void *__init alloc_large_system_hash(const char *tablename,
5345                     unsigned long bucketsize,
5346                     unsigned long numentries,
5347                     int scale,
5348                     int flags,
5349                     unsigned int *_hash_shift,
5350                     unsigned int *_hash_mask,
5351                     unsigned long low_limit,
5352                     unsigned long high_limit)
5353{
5354    unsigned long long max = high_limit;
5355    unsigned long log2qty, size;
5356    void *table = NULL;
5357
5358    /* allow the kernel cmdline to have a say */
5359    if (!numentries) {
5360        /* round applicable memory size up to nearest megabyte */
5361        numentries = nr_kernel_pages;
5362        numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
5363        numentries >>= 20 - PAGE_SHIFT;
5364        numentries <<= 20 - PAGE_SHIFT;
5365
5366        /* limit to 1 bucket per 2^scale bytes of low memory */
5367        if (scale > PAGE_SHIFT)
5368            numentries >>= (scale - PAGE_SHIFT);
5369        else
5370            numentries <<= (PAGE_SHIFT - scale);
5371
5372        /* Make sure we've got at least a 0-order allocation.. */
5373        if (unlikely(flags & HASH_SMALL)) {
5374            /* Makes no sense without HASH_EARLY */
5375            WARN_ON(!(flags & HASH_EARLY));
5376            if (!(numentries >> *_hash_shift)) {
5377                numentries = 1UL << *_hash_shift;
5378                BUG_ON(!numentries);
5379            }
5380        } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
5381            numentries = PAGE_SIZE / bucketsize;
5382    }
5383    numentries = roundup_pow_of_two(numentries);
5384
5385    /* limit allocation size to 1/16 total memory by default */
5386    if (max == 0) {
5387        max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5388        do_div(max, bucketsize);
5389    }
5390    max = min(max, 0x80000000ULL);
5391
5392    if (numentries < low_limit)
5393        numentries = low_limit;
5394    if (numentries > max)
5395        numentries = max;
5396
5397    log2qty = ilog2(numentries);
5398
5399    do {
5400        size = bucketsize << log2qty;
5401        if (flags & HASH_EARLY)
5402            table = alloc_bootmem_nopanic(size);
5403        else if (hashdist)
5404            table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5405        else {
5406            /*
5407             * If bucketsize is not a power-of-two, we may free
5408             * some pages at the end of hash table which
5409             * alloc_pages_exact() automatically does
5410             */
5411            if (get_order(size) < MAX_ORDER) {
5412                table = alloc_pages_exact(size, GFP_ATOMIC);
5413                kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5414            }
5415        }
5416    } while (!table && size > PAGE_SIZE && --log2qty);
5417
5418    if (!table)
5419        panic("Failed to allocate %s hash table\n", tablename);
5420
5421    printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
5422           tablename,
5423           (1UL << log2qty),
5424           ilog2(size) - PAGE_SHIFT,
5425           size);
5426
5427    if (_hash_shift)
5428        *_hash_shift = log2qty;
5429    if (_hash_mask)
5430        *_hash_mask = (1 << log2qty) - 1;
5431
5432    return table;
5433}
5434
5435/* Return a pointer to the bitmap storing bits affecting a block of pages */
5436static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
5437                            unsigned long pfn)
5438{
5439#ifdef CONFIG_SPARSEMEM
5440    return __pfn_to_section(pfn)->pageblock_flags;
5441#else
5442    return zone->pageblock_flags;
5443#endif /* CONFIG_SPARSEMEM */
5444}
5445
5446static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
5447{
5448#ifdef CONFIG_SPARSEMEM
5449    pfn &= (PAGES_PER_SECTION-1);
5450    return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5451#else
5452    pfn = pfn - zone->zone_start_pfn;
5453    return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5454#endif /* CONFIG_SPARSEMEM */
5455}
5456
5457/**
5458 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5459 * @page: The page within the block of interest
5460 * @start_bitidx: The first bit of interest to retrieve
5461 * @end_bitidx: The last bit of interest
5462 * returns pageblock_bits flags
5463 */
5464unsigned long get_pageblock_flags_group(struct page *page,
5465                    int start_bitidx, int end_bitidx)
5466{
5467    struct zone *zone;
5468    unsigned long *bitmap;
5469    unsigned long pfn, bitidx;
5470    unsigned long flags = 0;
5471    unsigned long value = 1;
5472
5473    zone = page_zone(page);
5474    pfn = page_to_pfn(page);
5475    bitmap = get_pageblock_bitmap(zone, pfn);
5476    bitidx = pfn_to_bitidx(zone, pfn);
5477
5478    for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5479        if (test_bit(bitidx + start_bitidx, bitmap))
5480            flags |= value;
5481
5482    return flags;
5483}
5484
5485/**
5486 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5487 * @page: The page within the block of interest
5488 * @start_bitidx: The first bit of interest
5489 * @end_bitidx: The last bit of interest
5490 * @flags: The flags to set
5491 */
5492void set_pageblock_flags_group(struct page *page, unsigned long flags,
5493                    int start_bitidx, int end_bitidx)
5494{
5495    struct zone *zone;
5496    unsigned long *bitmap;
5497    unsigned long pfn, bitidx;
5498    unsigned long value = 1;
5499
5500    zone = page_zone(page);
5501    pfn = page_to_pfn(page);
5502    bitmap = get_pageblock_bitmap(zone, pfn);
5503    bitidx = pfn_to_bitidx(zone, pfn);
5504    VM_BUG_ON(pfn < zone->zone_start_pfn);
5505    VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages);
5506
5507    for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5508        if (flags & value)
5509            __set_bit(bitidx + start_bitidx, bitmap);
5510        else
5511            __clear_bit(bitidx + start_bitidx, bitmap);
5512}
5513
5514/*
5515 * This function checks whether pageblock includes unmovable pages or not.
5516 * If @count is not zero, it is okay to include less @count unmovable pages
5517 *
5518 * PageLRU check wihtout isolation or lru_lock could race so that
5519 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
5520 * expect this function should be exact.
5521 */
5522bool has_unmovable_pages(struct zone *zone, struct page *page, int count)
5523{
5524    unsigned long pfn, iter, found;
5525    int mt;
5526
5527    /*
5528     * For avoiding noise data, lru_add_drain_all() should be called
5529     * If ZONE_MOVABLE, the zone never contains unmovable pages
5530     */
5531    if (zone_idx(zone) == ZONE_MOVABLE)
5532        return false;
5533    mt = get_pageblock_migratetype(page);
5534    if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
5535        return false;
5536
5537    pfn = page_to_pfn(page);
5538    for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
5539        unsigned long check = pfn + iter;
5540
5541        if (!pfn_valid_within(check))
5542            continue;
5543
5544        page = pfn_to_page(check);
5545        /*
5546         * We can't use page_count without pin a page
5547         * because another CPU can free compound page.
5548         * This check already skips compound tails of THP
5549         * because their page->_count is zero at all time.
5550         */
5551        if (!atomic_read(&page->_count)) {
5552            if (PageBuddy(page))
5553                iter += (1 << page_order(page)) - 1;
5554            continue;
5555        }
5556
5557        if (!PageLRU(page))
5558            found++;
5559        /*
5560         * If there are RECLAIMABLE pages, we need to check it.
5561         * But now, memory offline itself doesn't call shrink_slab()
5562         * and it still to be fixed.
5563         */
5564        /*
5565         * If the page is not RAM, page_count()should be 0.
5566         * we don't need more check. This is an _used_ not-movable page.
5567         *
5568         * The problematic thing here is PG_reserved pages. PG_reserved
5569         * is set to both of a memory hole page and a _used_ kernel
5570         * page at boot.
5571         */
5572        if (found > count)
5573            return true;
5574    }
5575    return false;
5576}
5577
5578bool is_pageblock_removable_nolock(struct page *page)
5579{
5580    struct zone *zone;
5581    unsigned long pfn;
5582
5583    /*
5584     * We have to be careful here because we are iterating over memory
5585     * sections which are not zone aware so we might end up outside of
5586     * the zone but still within the section.
5587     * We have to take care about the node as well. If the node is offline
5588     * its NODE_DATA will be NULL - see page_zone.
5589     */
5590    if (!node_online(page_to_nid(page)))
5591        return false;
5592
5593    zone = page_zone(page);
5594    pfn = page_to_pfn(page);
5595    if (zone->zone_start_pfn > pfn ||
5596            zone->zone_start_pfn + zone->spanned_pages <= pfn)
5597        return false;
5598
5599    return !has_unmovable_pages(zone, page, 0);
5600}
5601
5602#ifdef CONFIG_CMA
5603
5604static unsigned long pfn_max_align_down(unsigned long pfn)
5605{
5606    return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
5607                 pageblock_nr_pages) - 1);
5608}
5609
5610static unsigned long pfn_max_align_up(unsigned long pfn)
5611{
5612    return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
5613                pageblock_nr_pages));
5614}
5615
5616static struct page *
5617__alloc_contig_migrate_alloc(struct page *page, unsigned long private,
5618                 int **resultp)
5619{
5620    gfp_t gfp_mask = GFP_USER | __GFP_MOVABLE;
5621
5622    if (PageHighMem(page))
5623        gfp_mask |= __GFP_HIGHMEM;
5624
5625    return alloc_page(gfp_mask);
5626}
5627
5628/* [start, end) must belong to a single zone. */
5629static int __alloc_contig_migrate_range(unsigned long start, unsigned long end)
5630{
5631    /* This function is based on compact_zone() from compaction.c. */
5632
5633    unsigned long pfn = start;
5634    unsigned int tries = 0;
5635    int ret = 0;
5636
5637    struct compact_control cc = {
5638        .nr_migratepages = 0,
5639        .order = -1,
5640        .zone = page_zone(pfn_to_page(start)),
5641        .sync = true,
5642    };
5643    INIT_LIST_HEAD(&cc.migratepages);
5644
5645    migrate_prep_local();
5646
5647    while (pfn < end || !list_empty(&cc.migratepages)) {
5648        if (fatal_signal_pending(current)) {
5649            ret = -EINTR;
5650            break;
5651        }
5652
5653        if (list_empty(&cc.migratepages)) {
5654            cc.nr_migratepages = 0;
5655            pfn = isolate_migratepages_range(cc.zone, &cc,
5656                             pfn, end);
5657            if (!pfn) {
5658                ret = -EINTR;
5659                break;
5660            }
5661            tries = 0;
5662        } else if (++tries == 5) {
5663            ret = ret < 0 ? ret : -EBUSY;
5664            break;
5665        }
5666
5667        ret = migrate_pages(&cc.migratepages,
5668                    __alloc_contig_migrate_alloc,
5669                    0, false, MIGRATE_SYNC);
5670    }
5671
5672    putback_lru_pages(&cc.migratepages);
5673    return ret > 0 ? 0 : ret;
5674}
5675
5676/*
5677 * Update zone's cma pages counter used for watermark level calculation.
5678 */
5679static inline void __update_cma_watermarks(struct zone *zone, int count)
5680{
5681    unsigned long flags;
5682    spin_lock_irqsave(&zone->lock, flags);
5683    zone->min_cma_pages += count;
5684    spin_unlock_irqrestore(&zone->lock, flags);
5685    setup_per_zone_wmarks();
5686}
5687
5688/*
5689 * Trigger memory pressure bump to reclaim some pages in order to be able to
5690 * allocate 'count' pages in single page units. Does similar work as
5691 *__alloc_pages_slowpath() function.
5692 */
5693static int __reclaim_pages(struct zone *zone, gfp_t gfp_mask, int count)
5694{
5695    enum zone_type high_zoneidx = gfp_zone(gfp_mask);
5696    struct zonelist *zonelist = node_zonelist(0, gfp_mask);
5697    int did_some_progress = 0;
5698    int order = 1;
5699
5700    /*
5701     * Increase level of watermarks to force kswapd do his job
5702     * to stabilise at new watermark level.
5703     */
5704    __update_cma_watermarks(zone, count);
5705
5706    /* Obey watermarks as if the page was being allocated */
5707    while (!zone_watermark_ok(zone, 0, low_wmark_pages(zone), 0, 0)) {
5708        wake_all_kswapd(order, zonelist, high_zoneidx, zone_idx(zone));
5709
5710        did_some_progress = __perform_reclaim(gfp_mask, order, zonelist,
5711                              NULL);
5712        if (!did_some_progress) {
5713            /* Exhausted what can be done so it's blamo time */
5714            out_of_memory(zonelist, gfp_mask, order, NULL, false);
5715        }
5716    }
5717
5718    /* Restore original watermark levels. */
5719    __update_cma_watermarks(zone, -count);
5720
5721    return count;
5722}
5723
5724/**
5725 * alloc_contig_range() -- tries to allocate given range of pages
5726 * @start: start PFN to allocate
5727 * @end: one-past-the-last PFN to allocate
5728 * @migratetype: migratetype of the underlaying pageblocks (either
5729 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
5730 * in range must have the same migratetype and it must
5731 * be either of the two.
5732 *
5733 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
5734 * aligned, however it's the caller's responsibility to guarantee that
5735 * we are the only thread that changes migrate type of pageblocks the
5736 * pages fall in.
5737 *
5738 * The PFN range must belong to a single zone.
5739 *
5740 * Returns zero on success or negative error code. On success all
5741 * pages which PFN is in [start, end) are allocated for the caller and
5742 * need to be freed with free_contig_range().
5743 */
5744int alloc_contig_range(unsigned long start, unsigned long end,
5745               unsigned migratetype)
5746{
5747    struct zone *zone = page_zone(pfn_to_page(start));
5748    unsigned long outer_start, outer_end;
5749    int ret = 0, order;
5750
5751    /*
5752     * What we do here is we mark all pageblocks in range as
5753     * MIGRATE_ISOLATE. Because pageblock and max order pages may
5754     * have different sizes, and due to the way page allocator
5755     * work, we align the range to biggest of the two pages so
5756     * that page allocator won't try to merge buddies from
5757     * different pageblocks and change MIGRATE_ISOLATE to some
5758     * other migration type.
5759     *
5760     * Once the pageblocks are marked as MIGRATE_ISOLATE, we
5761     * migrate the pages from an unaligned range (ie. pages that
5762     * we are interested in). This will put all the pages in
5763     * range back to page allocator as MIGRATE_ISOLATE.
5764     *
5765     * When this is done, we take the pages in range from page
5766     * allocator removing them from the buddy system. This way
5767     * page allocator will never consider using them.
5768     *
5769     * This lets us mark the pageblocks back as
5770     * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
5771     * aligned range but not in the unaligned, original range are
5772     * put back to page allocator so that buddy can use them.
5773     */
5774
5775    ret = start_isolate_page_range(pfn_max_align_down(start),
5776                       pfn_max_align_up(end), migratetype);
5777    if (ret)
5778        goto done;
5779
5780    ret = __alloc_contig_migrate_range(start, end);
5781    if (ret)
5782        goto done;
5783
5784    /*
5785     * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
5786     * aligned blocks that are marked as MIGRATE_ISOLATE. What's
5787     * more, all pages in [start, end) are free in page allocator.
5788     * What we are going to do is to allocate all pages from
5789     * [start, end) (that is remove them from page allocator).
5790     *
5791     * The only problem is that pages at the beginning and at the
5792     * end of interesting range may be not aligned with pages that
5793     * page allocator holds, ie. they can be part of higher order
5794     * pages. Because of this, we reserve the bigger range and
5795     * once this is done free the pages we are not interested in.
5796     *
5797     * We don't have to hold zone->lock here because the pages are
5798     * isolated thus they won't get removed from buddy.
5799     */
5800
5801    lru_add_drain_all();
5802    drain_all_pages();
5803
5804    order = 0;
5805    outer_start = start;
5806    while (!PageBuddy(pfn_to_page(outer_start))) {
5807        if (++order >= MAX_ORDER) {
5808            ret = -EBUSY;
5809            goto done;
5810        }
5811        outer_start &= ~0UL << order;
5812    }
5813
5814    /* Make sure the range is really isolated. */
5815    if (test_pages_isolated(outer_start, end)) {
5816        pr_warn("alloc_contig_range test_pages_isolated(%lx, %lx) failed\n",
5817               outer_start, end);
5818        ret = -EBUSY;
5819        goto done;
5820    }
5821
5822    /*
5823     * Reclaim enough pages to make sure that contiguous allocation
5824     * will not starve the system.
5825     */
5826    __reclaim_pages(zone, GFP_HIGHUSER_MOVABLE, end-start);
5827
5828    /* Grab isolated pages from freelists. */
5829    outer_end = isolate_freepages_range(outer_start, end);
5830    if (!outer_end) {
5831        ret = -EBUSY;
5832        goto done;
5833    }
5834
5835    /* Free head and tail (if any) */
5836    if (start != outer_start)
5837        free_contig_range(outer_start, start - outer_start);
5838    if (end != outer_end)
5