source: src/linux/universal/linux-4.9/mm/page_alloc.c @ 31859

Last change on this file since 31859 was 31859, checked in by brainslayer, 6 weeks ago

kernel update

File size: 202.8 KB
Line 
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/kasan.h>
29#include <linux/module.h>
30#include <linux/suspend.h>
31#include <linux/pagevec.h>
32#include <linux/blkdev.h>
33#include <linux/slab.h>
34#include <linux/ratelimit.h>
35#include <linux/oom.h>
36#include <linux/notifier.h>
37#include <linux/topology.h>
38#include <linux/sysctl.h>
39#include <linux/cpu.h>
40#include <linux/cpuset.h>
41#include <linux/memory_hotplug.h>
42#include <linux/nodemask.h>
43#include <linux/vmalloc.h>
44#include <linux/vmstat.h>
45#include <linux/mempolicy.h>
46#include <linux/memremap.h>
47#include <linux/stop_machine.h>
48#include <linux/sort.h>
49#include <linux/pfn.h>
50#include <linux/backing-dev.h>
51#include <linux/fault-inject.h>
52#include <linux/page-isolation.h>
53#include <linux/page_ext.h>
54#include <linux/debugobjects.h>
55#include <linux/kmemleak.h>
56#include <linux/compaction.h>
57#include <trace/events/kmem.h>
58#include <linux/prefetch.h>
59#include <linux/mm_inline.h>
60#include <linux/migrate.h>
61#include <linux/page_ext.h>
62#include <linux/hugetlb.h>
63#include <linux/sched/rt.h>
64#include <linux/page_owner.h>
65#include <linux/kthread.h>
66#include <linux/memcontrol.h>
67
68#include <asm/sections.h>
69#include <asm/tlbflush.h>
70#include <asm/div64.h>
71#include "internal.h"
72
73/* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
74static DEFINE_MUTEX(pcp_batch_high_lock);
75#define MIN_PERCPU_PAGELIST_FRACTION    (8)
76
77#ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
78DEFINE_PER_CPU(int, numa_node);
79EXPORT_PER_CPU_SYMBOL(numa_node);
80#endif
81
82#ifdef CONFIG_HAVE_MEMORYLESS_NODES
83/*
84 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
85 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
86 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
87 * defined in <linux/topology.h>.
88 */
89DEFINE_PER_CPU(int, _numa_mem_);                /* Kernel "local memory" node */
90EXPORT_PER_CPU_SYMBOL(_numa_mem_);
91int _node_numa_mem_[MAX_NUMNODES];
92#endif
93
94#ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
95volatile unsigned long latent_entropy __latent_entropy;
96EXPORT_SYMBOL(latent_entropy);
97#endif
98
99/*
100 * Array of node states.
101 */
102nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
103        [N_POSSIBLE] = NODE_MASK_ALL,
104        [N_ONLINE] = { { [0] = 1UL } },
105#ifndef CONFIG_NUMA
106        [N_NORMAL_MEMORY] = { { [0] = 1UL } },
107#ifdef CONFIG_HIGHMEM
108        [N_HIGH_MEMORY] = { { [0] = 1UL } },
109#endif
110#ifdef CONFIG_MOVABLE_NODE
111        [N_MEMORY] = { { [0] = 1UL } },
112#endif
113        [N_CPU] = { { [0] = 1UL } },
114#endif  /* NUMA */
115};
116EXPORT_SYMBOL(node_states);
117
118/* Protect totalram_pages and zone->managed_pages */
119static DEFINE_SPINLOCK(managed_page_count_lock);
120
121unsigned long totalram_pages __read_mostly;
122unsigned long totalreserve_pages __read_mostly;
123unsigned long totalcma_pages __read_mostly;
124
125int percpu_pagelist_fraction;
126gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
127
128/*
129 * A cached value of the page's pageblock's migratetype, used when the page is
130 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
131 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
132 * Also the migratetype set in the page does not necessarily match the pcplist
133 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
134 * other index - this ensures that it will be put on the correct CMA freelist.
135 */
136static inline int get_pcppage_migratetype(struct page *page)
137{
138        return page->index;
139}
140
141static inline void set_pcppage_migratetype(struct page *page, int migratetype)
142{
143        page->index = migratetype;
144}
145
146#ifdef CONFIG_PM_SLEEP
147/*
148 * The following functions are used by the suspend/hibernate code to temporarily
149 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
150 * while devices are suspended.  To avoid races with the suspend/hibernate code,
151 * they should always be called with pm_mutex held (gfp_allowed_mask also should
152 * only be modified with pm_mutex held, unless the suspend/hibernate code is
153 * guaranteed not to run in parallel with that modification).
154 */
155
156static gfp_t saved_gfp_mask;
157
158void pm_restore_gfp_mask(void)
159{
160        WARN_ON(!mutex_is_locked(&pm_mutex));
161        if (saved_gfp_mask) {
162                gfp_allowed_mask = saved_gfp_mask;
163                saved_gfp_mask = 0;
164        }
165}
166
167void pm_restrict_gfp_mask(void)
168{
169        WARN_ON(!mutex_is_locked(&pm_mutex));
170        WARN_ON(saved_gfp_mask);
171        saved_gfp_mask = gfp_allowed_mask;
172        gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
173}
174
175bool pm_suspended_storage(void)
176{
177        if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
178                return false;
179        return true;
180}
181#endif /* CONFIG_PM_SLEEP */
182
183#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
184unsigned int pageblock_order __read_mostly;
185#endif
186
187static void __free_pages_ok(struct page *page, unsigned int order);
188
189/*
190 * results with 256, 32 in the lowmem_reserve sysctl:
191 *      1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
192 *      1G machine -> (16M dma, 784M normal, 224M high)
193 *      NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
194 *      HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
195 *      HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
196 *
197 * TBD: should special case ZONE_DMA32 machines here - in those we normally
198 * don't need any ZONE_NORMAL reservation
199 */
200int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
201#ifdef CONFIG_ZONE_DMA
202         256,
203#endif
204#ifdef CONFIG_ZONE_DMA32
205         256,
206#endif
207#ifdef CONFIG_HIGHMEM
208         32,
209#endif
210         32,
211};
212
213EXPORT_SYMBOL(totalram_pages);
214
215static char * const zone_names[MAX_NR_ZONES] = {
216#ifdef CONFIG_ZONE_DMA
217         "DMA",
218#endif
219#ifdef CONFIG_ZONE_DMA32
220         "DMA32",
221#endif
222         "Normal",
223#ifdef CONFIG_HIGHMEM
224         "HighMem",
225#endif
226         "Movable",
227#ifdef CONFIG_ZONE_DEVICE
228         "Device",
229#endif
230};
231
232char * const migratetype_names[MIGRATE_TYPES] = {
233        "Unmovable",
234        "Movable",
235        "Reclaimable",
236        "HighAtomic",
237#ifdef CONFIG_CMA
238        "CMA",
239#endif
240#ifdef CONFIG_MEMORY_ISOLATION
241        "Isolate",
242#endif
243};
244
245compound_page_dtor * const compound_page_dtors[] = {
246        NULL,
247        free_compound_page,
248#ifdef CONFIG_HUGETLB_PAGE
249        free_huge_page,
250#endif
251#ifdef CONFIG_TRANSPARENT_HUGEPAGE
252        free_transhuge_page,
253#endif
254};
255
256int min_free_kbytes = 1024;
257int user_min_free_kbytes = -1;
258int watermark_scale_factor = 10;
259
260static unsigned long __meminitdata nr_kernel_pages;
261static unsigned long __meminitdata nr_all_pages;
262static unsigned long __meminitdata dma_reserve;
263
264#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
265static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
266static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
267static unsigned long __initdata required_kernelcore;
268static unsigned long __initdata required_movablecore;
269static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
270static bool mirrored_kernelcore;
271
272/* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
273int movable_zone;
274EXPORT_SYMBOL(movable_zone);
275#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
276
277#if MAX_NUMNODES > 1
278int nr_node_ids __read_mostly = MAX_NUMNODES;
279int nr_online_nodes __read_mostly = 1;
280EXPORT_SYMBOL(nr_node_ids);
281EXPORT_SYMBOL(nr_online_nodes);
282#endif
283
284int page_group_by_mobility_disabled __read_mostly;
285
286#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
287static inline void reset_deferred_meminit(pg_data_t *pgdat)
288{
289        pgdat->first_deferred_pfn = ULONG_MAX;
290}
291
292/* Returns true if the struct page for the pfn is uninitialised */
293static inline bool __meminit early_page_uninitialised(unsigned long pfn)
294{
295        int nid = early_pfn_to_nid(pfn);
296
297        if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
298                return true;
299
300        return false;
301}
302
303/*
304 * Returns false when the remaining initialisation should be deferred until
305 * later in the boot cycle when it can be parallelised.
306 */
307static inline bool update_defer_init(pg_data_t *pgdat,
308                                unsigned long pfn, unsigned long zone_end,
309                                unsigned long *nr_initialised)
310{
311        unsigned long max_initialise;
312
313        /* Always populate low zones for address-contrained allocations */
314        if (zone_end < pgdat_end_pfn(pgdat))
315                return true;
316        /*
317         * Initialise at least 2G of a node but also take into account that
318         * two large system hashes that can take up 1GB for 0.25TB/node.
319         */
320        max_initialise = max(2UL << (30 - PAGE_SHIFT),
321                (pgdat->node_spanned_pages >> 8));
322
323        (*nr_initialised)++;
324        if ((*nr_initialised > max_initialise) &&
325            (pfn & (PAGES_PER_SECTION - 1)) == 0) {
326                pgdat->first_deferred_pfn = pfn;
327                return false;
328        }
329
330        return true;
331}
332#else
333static inline void reset_deferred_meminit(pg_data_t *pgdat)
334{
335}
336
337static inline bool early_page_uninitialised(unsigned long pfn)
338{
339        return false;
340}
341
342static inline bool update_defer_init(pg_data_t *pgdat,
343                                unsigned long pfn, unsigned long zone_end,
344                                unsigned long *nr_initialised)
345{
346        return true;
347}
348#endif
349
350/* Return a pointer to the bitmap storing bits affecting a block of pages */
351static inline unsigned long *get_pageblock_bitmap(struct page *page,
352                                                        unsigned long pfn)
353{
354#ifdef CONFIG_SPARSEMEM
355        return __pfn_to_section(pfn)->pageblock_flags;
356#else
357        return page_zone(page)->pageblock_flags;
358#endif /* CONFIG_SPARSEMEM */
359}
360
361static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
362{
363#ifdef CONFIG_SPARSEMEM
364        pfn &= (PAGES_PER_SECTION-1);
365        return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
366#else
367        pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
368        return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
369#endif /* CONFIG_SPARSEMEM */
370}
371
372/**
373 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
374 * @page: The page within the block of interest
375 * @pfn: The target page frame number
376 * @end_bitidx: The last bit of interest to retrieve
377 * @mask: mask of bits that the caller is interested in
378 *
379 * Return: pageblock_bits flags
380 */
381static __always_inline unsigned long __get_pfnblock_flags_mask(struct page *page,
382                                        unsigned long pfn,
383                                        unsigned long end_bitidx,
384                                        unsigned long mask)
385{
386        unsigned long *bitmap;
387        unsigned long bitidx, word_bitidx;
388        unsigned long word;
389
390        bitmap = get_pageblock_bitmap(page, pfn);
391        bitidx = pfn_to_bitidx(page, pfn);
392        word_bitidx = bitidx / BITS_PER_LONG;
393        bitidx &= (BITS_PER_LONG-1);
394
395        word = bitmap[word_bitidx];
396        bitidx += end_bitidx;
397        return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
398}
399
400unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
401                                        unsigned long end_bitidx,
402                                        unsigned long mask)
403{
404        return __get_pfnblock_flags_mask(page, pfn, end_bitidx, mask);
405}
406
407static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
408{
409        return __get_pfnblock_flags_mask(page, pfn, PB_migrate_end, MIGRATETYPE_MASK);
410}
411
412/**
413 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
414 * @page: The page within the block of interest
415 * @flags: The flags to set
416 * @pfn: The target page frame number
417 * @end_bitidx: The last bit of interest
418 * @mask: mask of bits that the caller is interested in
419 */
420void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
421                                        unsigned long pfn,
422                                        unsigned long end_bitidx,
423                                        unsigned long mask)
424{
425        unsigned long *bitmap;
426        unsigned long bitidx, word_bitidx;
427        unsigned long old_word, word;
428
429        BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
430
431        bitmap = get_pageblock_bitmap(page, pfn);
432        bitidx = pfn_to_bitidx(page, pfn);
433        word_bitidx = bitidx / BITS_PER_LONG;
434        bitidx &= (BITS_PER_LONG-1);
435
436        VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
437
438        bitidx += end_bitidx;
439        mask <<= (BITS_PER_LONG - bitidx - 1);
440        flags <<= (BITS_PER_LONG - bitidx - 1);
441
442        word = READ_ONCE(bitmap[word_bitidx]);
443        for (;;) {
444                old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
445                if (word == old_word)
446                        break;
447                word = old_word;
448        }
449}
450
451void set_pageblock_migratetype(struct page *page, int migratetype)
452{
453        if (unlikely(page_group_by_mobility_disabled &&
454                     migratetype < MIGRATE_PCPTYPES))
455                migratetype = MIGRATE_UNMOVABLE;
456
457        set_pageblock_flags_group(page, (unsigned long)migratetype,
458                                        PB_migrate, PB_migrate_end);
459}
460
461#ifdef CONFIG_DEBUG_VM
462static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
463{
464        int ret = 0;
465        unsigned seq;
466        unsigned long pfn = page_to_pfn(page);
467        unsigned long sp, start_pfn;
468
469        do {
470                seq = zone_span_seqbegin(zone);
471                start_pfn = zone->zone_start_pfn;
472                sp = zone->spanned_pages;
473                if (!zone_spans_pfn(zone, pfn))
474                        ret = 1;
475        } while (zone_span_seqretry(zone, seq));
476
477        if (ret)
478                pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
479                        pfn, zone_to_nid(zone), zone->name,
480                        start_pfn, start_pfn + sp);
481
482        return ret;
483}
484
485static int page_is_consistent(struct zone *zone, struct page *page)
486{
487        if (!pfn_valid_within(page_to_pfn(page)))
488                return 0;
489        if (zone != page_zone(page))
490                return 0;
491
492        return 1;
493}
494/*
495 * Temporary debugging check for pages not lying within a given zone.
496 */
497static int bad_range(struct zone *zone, struct page *page)
498{
499        if (page_outside_zone_boundaries(zone, page))
500                return 1;
501        if (!page_is_consistent(zone, page))
502                return 1;
503
504        return 0;
505}
506#else
507static inline int bad_range(struct zone *zone, struct page *page)
508{
509        return 0;
510}
511#endif
512
513static void bad_page(struct page *page, const char *reason,
514                unsigned long bad_flags)
515{
516        static unsigned long resume;
517        static unsigned long nr_shown;
518        static unsigned long nr_unshown;
519
520        /*
521         * Allow a burst of 60 reports, then keep quiet for that minute;
522         * or allow a steady drip of one report per second.
523         */
524        if (nr_shown == 60) {
525                if (time_before(jiffies, resume)) {
526                        nr_unshown++;
527                        goto out;
528                }
529                if (nr_unshown) {
530                        pr_alert(
531                              "BUG: Bad page state: %lu messages suppressed\n",
532                                nr_unshown);
533                        nr_unshown = 0;
534                }
535                nr_shown = 0;
536        }
537        if (nr_shown++ == 0)
538                resume = jiffies + 60 * HZ;
539
540        pr_alert("BUG: Bad page state in process %s  pfn:%05lx\n",
541                current->comm, page_to_pfn(page));
542        __dump_page(page, reason);
543        bad_flags &= page->flags;
544        if (bad_flags)
545                pr_alert("bad because of flags: %#lx(%pGp)\n",
546                                                bad_flags, &bad_flags);
547        dump_page_owner(page);
548
549        print_modules();
550        dump_stack();
551out:
552        /* Leave bad fields for debug, except PageBuddy could make trouble */
553        page_mapcount_reset(page); /* remove PageBuddy */
554        add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
555}
556
557/*
558 * Higher-order pages are called "compound pages".  They are structured thusly:
559 *
560 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
561 *
562 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
563 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
564 *
565 * The first tail page's ->compound_dtor holds the offset in array of compound
566 * page destructors. See compound_page_dtors.
567 *
568 * The first tail page's ->compound_order holds the order of allocation.
569 * This usage means that zero-order pages may not be compound.
570 */
571
572void free_compound_page(struct page *page)
573{
574        __free_pages_ok(page, compound_order(page));
575}
576
577void prep_compound_page(struct page *page, unsigned int order)
578{
579        int i;
580        int nr_pages = 1 << order;
581
582        set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
583        set_compound_order(page, order);
584        __SetPageHead(page);
585        for (i = 1; i < nr_pages; i++) {
586                struct page *p = page + i;
587                set_page_count(p, 0);
588                p->mapping = TAIL_MAPPING;
589                set_compound_head(p, page);
590        }
591        atomic_set(compound_mapcount_ptr(page), -1);
592}
593
594#ifdef CONFIG_DEBUG_PAGEALLOC
595unsigned int _debug_guardpage_minorder;
596bool _debug_pagealloc_enabled __read_mostly
597                        = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
598EXPORT_SYMBOL(_debug_pagealloc_enabled);
599bool _debug_guardpage_enabled __read_mostly;
600
601static int __init early_debug_pagealloc(char *buf)
602{
603        if (!buf)
604                return -EINVAL;
605        return kstrtobool(buf, &_debug_pagealloc_enabled);
606}
607early_param("debug_pagealloc", early_debug_pagealloc);
608
609static bool need_debug_guardpage(void)
610{
611        /* If we don't use debug_pagealloc, we don't need guard page */
612        if (!debug_pagealloc_enabled())
613                return false;
614
615        if (!debug_guardpage_minorder())
616                return false;
617
618        return true;
619}
620
621static void init_debug_guardpage(void)
622{
623        if (!debug_pagealloc_enabled())
624                return;
625
626        if (!debug_guardpage_minorder())
627                return;
628
629        _debug_guardpage_enabled = true;
630}
631
632struct page_ext_operations debug_guardpage_ops = {
633        .need = need_debug_guardpage,
634        .init = init_debug_guardpage,
635};
636
637static int __init debug_guardpage_minorder_setup(char *buf)
638{
639        unsigned long res;
640
641        if (kstrtoul(buf, 10, &res) < 0 ||  res > MAX_ORDER / 2) {
642                pr_err("Bad debug_guardpage_minorder value\n");
643                return 0;
644        }
645        _debug_guardpage_minorder = res;
646        pr_info("Setting debug_guardpage_minorder to %lu\n", res);
647        return 0;
648}
649early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
650
651static inline bool set_page_guard(struct zone *zone, struct page *page,
652                                unsigned int order, int migratetype)
653{
654        struct page_ext *page_ext;
655
656        if (!debug_guardpage_enabled())
657                return false;
658
659        if (order >= debug_guardpage_minorder())
660                return false;
661
662        page_ext = lookup_page_ext(page);
663        if (unlikely(!page_ext))
664                return false;
665
666        __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
667
668        INIT_LIST_HEAD(&page->lru);
669        set_page_private(page, order);
670        /* Guard pages are not available for any usage */
671        __mod_zone_freepage_state(zone, -(1 << order), migratetype);
672
673        return true;
674}
675
676static inline void clear_page_guard(struct zone *zone, struct page *page,
677                                unsigned int order, int migratetype)
678{
679        struct page_ext *page_ext;
680
681        if (!debug_guardpage_enabled())
682                return;
683
684        page_ext = lookup_page_ext(page);
685        if (unlikely(!page_ext))
686                return;
687
688        __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
689
690        set_page_private(page, 0);
691        if (!is_migrate_isolate(migratetype))
692                __mod_zone_freepage_state(zone, (1 << order), migratetype);
693}
694#else
695struct page_ext_operations debug_guardpage_ops;
696static inline bool set_page_guard(struct zone *zone, struct page *page,
697                        unsigned int order, int migratetype) { return false; }
698static inline void clear_page_guard(struct zone *zone, struct page *page,
699                                unsigned int order, int migratetype) {}
700#endif
701
702static inline void set_page_order(struct page *page, unsigned int order)
703{
704        set_page_private(page, order);
705        __SetPageBuddy(page);
706}
707
708static inline void rmv_page_order(struct page *page)
709{
710        __ClearPageBuddy(page);
711        set_page_private(page, 0);
712}
713
714/*
715 * This function checks whether a page is free && is the buddy
716 * we can do coalesce a page and its buddy if
717 * (a) the buddy is not in a hole &&
718 * (b) the buddy is in the buddy system &&
719 * (c) a page and its buddy have the same order &&
720 * (d) a page and its buddy are in the same zone.
721 *
722 * For recording whether a page is in the buddy system, we set ->_mapcount
723 * PAGE_BUDDY_MAPCOUNT_VALUE.
724 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
725 * serialized by zone->lock.
726 *
727 * For recording page's order, we use page_private(page).
728 */
729static inline int page_is_buddy(struct page *page, struct page *buddy,
730                                                        unsigned int order)
731{
732        if (!pfn_valid_within(page_to_pfn(buddy)))
733                return 0;
734
735        if (page_is_guard(buddy) && page_order(buddy) == order) {
736                if (page_zone_id(page) != page_zone_id(buddy))
737                        return 0;
738
739                VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
740
741                return 1;
742        }
743
744        if (PageBuddy(buddy) && page_order(buddy) == order) {
745                /*
746                 * zone check is done late to avoid uselessly
747                 * calculating zone/node ids for pages that could
748                 * never merge.
749                 */
750                if (page_zone_id(page) != page_zone_id(buddy))
751                        return 0;
752
753                VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
754
755                return 1;
756        }
757        return 0;
758}
759
760/*
761 * Freeing function for a buddy system allocator.
762 *
763 * The concept of a buddy system is to maintain direct-mapped table
764 * (containing bit values) for memory blocks of various "orders".
765 * The bottom level table contains the map for the smallest allocatable
766 * units of memory (here, pages), and each level above it describes
767 * pairs of units from the levels below, hence, "buddies".
768 * At a high level, all that happens here is marking the table entry
769 * at the bottom level available, and propagating the changes upward
770 * as necessary, plus some accounting needed to play nicely with other
771 * parts of the VM system.
772 * At each level, we keep a list of pages, which are heads of continuous
773 * free pages of length of (1 << order) and marked with _mapcount
774 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
775 * field.
776 * So when we are allocating or freeing one, we can derive the state of the
777 * other.  That is, if we allocate a small block, and both were
778 * free, the remainder of the region must be split into blocks.
779 * If a block is freed, and its buddy is also free, then this
780 * triggers coalescing into a block of larger size.
781 *
782 * -- nyc
783 */
784
785static inline void __free_one_page(struct page *page,
786                unsigned long pfn,
787                struct zone *zone, unsigned int order,
788                int migratetype)
789{
790        unsigned long page_idx;
791        unsigned long combined_idx;
792        unsigned long uninitialized_var(buddy_idx);
793        struct page *buddy;
794        unsigned int max_order;
795
796        max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
797
798        VM_BUG_ON(!zone_is_initialized(zone));
799        VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
800
801        VM_BUG_ON(migratetype == -1);
802        if (likely(!is_migrate_isolate(migratetype)))
803                __mod_zone_freepage_state(zone, 1 << order, migratetype);
804
805        page_idx = pfn & ((1 << MAX_ORDER) - 1);
806
807        VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page);
808        VM_BUG_ON_PAGE(bad_range(zone, page), page);
809
810continue_merging:
811        while (order < max_order - 1) {
812                buddy_idx = __find_buddy_index(page_idx, order);
813                buddy = page + (buddy_idx - page_idx);
814                if (!page_is_buddy(page, buddy, order))
815                        goto done_merging;
816                /*
817                 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
818                 * merge with it and move up one order.
819                 */
820                if (page_is_guard(buddy)) {
821                        clear_page_guard(zone, buddy, order, migratetype);
822                } else {
823                        list_del(&buddy->lru);
824                        zone->free_area[order].nr_free--;
825                        rmv_page_order(buddy);
826                }
827                combined_idx = buddy_idx & page_idx;
828                page = page + (combined_idx - page_idx);
829                page_idx = combined_idx;
830                order++;
831        }
832        if (max_order < MAX_ORDER) {
833                /* If we are here, it means order is >= pageblock_order.
834                 * We want to prevent merge between freepages on isolate
835                 * pageblock and normal pageblock. Without this, pageblock
836                 * isolation could cause incorrect freepage or CMA accounting.
837                 *
838                 * We don't want to hit this code for the more frequent
839                 * low-order merging.
840                 */
841                if (unlikely(has_isolate_pageblock(zone))) {
842                        int buddy_mt;
843
844                        buddy_idx = __find_buddy_index(page_idx, order);
845                        buddy = page + (buddy_idx - page_idx);
846                        buddy_mt = get_pageblock_migratetype(buddy);
847
848                        if (migratetype != buddy_mt
849                                        && (is_migrate_isolate(migratetype) ||
850                                                is_migrate_isolate(buddy_mt)))
851                                goto done_merging;
852                }
853                max_order++;
854                goto continue_merging;
855        }
856
857done_merging:
858        set_page_order(page, order);
859
860        /*
861         * If this is not the largest possible page, check if the buddy
862         * of the next-highest order is free. If it is, it's possible
863         * that pages are being freed that will coalesce soon. In case,
864         * that is happening, add the free page to the tail of the list
865         * so it's less likely to be used soon and more likely to be merged
866         * as a higher order page
867         */
868        if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
869                struct page *higher_page, *higher_buddy;
870                combined_idx = buddy_idx & page_idx;
871                higher_page = page + (combined_idx - page_idx);
872                buddy_idx = __find_buddy_index(combined_idx, order + 1);
873                higher_buddy = higher_page + (buddy_idx - combined_idx);
874                if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
875                        list_add_tail(&page->lru,
876                                &zone->free_area[order].free_list[migratetype]);
877                        goto out;
878                }
879        }
880
881        list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
882out:
883        zone->free_area[order].nr_free++;
884}
885
886/*
887 * A bad page could be due to a number of fields. Instead of multiple branches,
888 * try and check multiple fields with one check. The caller must do a detailed
889 * check if necessary.
890 */
891static inline bool page_expected_state(struct page *page,
892                                        unsigned long check_flags)
893{
894        if (unlikely(atomic_read(&page->_mapcount) != -1))
895                return false;
896
897        if (unlikely((unsigned long)page->mapping |
898                        page_ref_count(page) |
899#ifdef CONFIG_MEMCG
900                        (unsigned long)page->mem_cgroup |
901#endif
902                        (page->flags & check_flags)))
903                return false;
904
905        return true;
906}
907
908static void free_pages_check_bad(struct page *page)
909{
910        const char *bad_reason;
911        unsigned long bad_flags;
912
913        bad_reason = NULL;
914        bad_flags = 0;
915
916        if (unlikely(atomic_read(&page->_mapcount) != -1))
917                bad_reason = "nonzero mapcount";
918        if (unlikely(page->mapping != NULL))
919                bad_reason = "non-NULL mapping";
920        if (unlikely(page_ref_count(page) != 0))
921                bad_reason = "nonzero _refcount";
922        if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
923                bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
924                bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
925        }
926#ifdef CONFIG_MEMCG
927        if (unlikely(page->mem_cgroup))
928                bad_reason = "page still charged to cgroup";
929#endif
930        bad_page(page, bad_reason, bad_flags);
931}
932
933static inline int free_pages_check(struct page *page)
934{
935        if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
936                return 0;
937
938        /* Something has gone sideways, find it */
939        free_pages_check_bad(page);
940        return 1;
941}
942
943static int free_tail_pages_check(struct page *head_page, struct page *page)
944{
945        int ret = 1;
946
947        /*
948         * We rely page->lru.next never has bit 0 set, unless the page
949         * is PageTail(). Let's make sure that's true even for poisoned ->lru.
950         */
951        BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
952
953        if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
954                ret = 0;
955                goto out;
956        }
957        switch (page - head_page) {
958        case 1:
959                /* the first tail page: ->mapping is compound_mapcount() */
960                if (unlikely(compound_mapcount(page))) {
961                        bad_page(page, "nonzero compound_mapcount", 0);
962                        goto out;
963                }
964                break;
965        case 2:
966                /*
967                 * the second tail page: ->mapping is
968                 * page_deferred_list().next -- ignore value.
969                 */
970                break;
971        default:
972                if (page->mapping != TAIL_MAPPING) {
973                        bad_page(page, "corrupted mapping in tail page", 0);
974                        goto out;
975                }
976                break;
977        }
978        if (unlikely(!PageTail(page))) {
979                bad_page(page, "PageTail not set", 0);
980                goto out;
981        }
982        if (unlikely(compound_head(page) != head_page)) {
983                bad_page(page, "compound_head not consistent", 0);
984                goto out;
985        }
986        ret = 0;
987out:
988        page->mapping = NULL;
989        clear_compound_head(page);
990        return ret;
991}
992
993static __always_inline bool free_pages_prepare(struct page *page,
994                                        unsigned int order, bool check_free)
995{
996        int bad = 0;
997
998        VM_BUG_ON_PAGE(PageTail(page), page);
999
1000        trace_mm_page_free(page, order);
1001        kmemcheck_free_shadow(page, order);
1002
1003        /*
1004         * Check tail pages before head page information is cleared to
1005         * avoid checking PageCompound for order-0 pages.
1006         */
1007        if (unlikely(order)) {
1008                bool compound = PageCompound(page);
1009                int i;
1010
1011                VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1012
1013                if (compound)
1014                        ClearPageDoubleMap(page);
1015                for (i = 1; i < (1 << order); i++) {
1016                        if (compound)
1017                                bad += free_tail_pages_check(page, page + i);
1018                        if (unlikely(free_pages_check(page + i))) {
1019                                bad++;
1020                                continue;
1021                        }
1022                        (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1023                }
1024        }
1025        if (PageMappingFlags(page))
1026                page->mapping = NULL;
1027        if (memcg_kmem_enabled() && PageKmemcg(page))
1028                memcg_kmem_uncharge(page, order);
1029        if (check_free)
1030                bad += free_pages_check(page);
1031        if (bad)
1032                return false;
1033
1034        page_cpupid_reset_last(page);
1035        page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1036        reset_page_owner(page, order);
1037
1038        if (!PageHighMem(page)) {
1039                debug_check_no_locks_freed(page_address(page),
1040                                           PAGE_SIZE << order);
1041                debug_check_no_obj_freed(page_address(page),
1042                                           PAGE_SIZE << order);
1043        }
1044        arch_free_page(page, order);
1045        kernel_poison_pages(page, 1 << order, 0);
1046        kernel_map_pages(page, 1 << order, 0);
1047        kasan_free_pages(page, order);
1048
1049        return true;
1050}
1051
1052#ifdef CONFIG_DEBUG_VM
1053static inline bool free_pcp_prepare(struct page *page)
1054{
1055        return free_pages_prepare(page, 0, true);
1056}
1057
1058static inline bool bulkfree_pcp_prepare(struct page *page)
1059{
1060        return false;
1061}
1062#else
1063static bool free_pcp_prepare(struct page *page)
1064{
1065        return free_pages_prepare(page, 0, false);
1066}
1067
1068static bool bulkfree_pcp_prepare(struct page *page)
1069{
1070        return free_pages_check(page);
1071}
1072#endif /* CONFIG_DEBUG_VM */
1073
1074/*
1075 * Frees a number of pages from the PCP lists
1076 * Assumes all pages on list are in same zone, and of same order.
1077 * count is the number of pages to free.
1078 *
1079 * If the zone was previously in an "all pages pinned" state then look to
1080 * see if this freeing clears that state.
1081 *
1082 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1083 * pinned" detection logic.
1084 */
1085static void free_pcppages_bulk(struct zone *zone, int count,
1086                                        struct per_cpu_pages *pcp)
1087{
1088        int migratetype = 0;
1089        int batch_free = 0;
1090        unsigned long nr_scanned;
1091        bool isolated_pageblocks;
1092
1093        spin_lock(&zone->lock);
1094        isolated_pageblocks = has_isolate_pageblock(zone);
1095        nr_scanned = node_page_state(zone->zone_pgdat, NR_PAGES_SCANNED);
1096        if (nr_scanned)
1097                __mod_node_page_state(zone->zone_pgdat, NR_PAGES_SCANNED, -nr_scanned);
1098
1099        while (count) {
1100                struct page *page;
1101                struct list_head *list;
1102
1103                /*
1104                 * Remove pages from lists in a round-robin fashion. A
1105                 * batch_free count is maintained that is incremented when an
1106                 * empty list is encountered.  This is so more pages are freed
1107                 * off fuller lists instead of spinning excessively around empty
1108                 * lists
1109                 */
1110                do {
1111                        batch_free++;
1112                        if (++migratetype == MIGRATE_PCPTYPES)
1113                                migratetype = 0;
1114                        list = &pcp->lists[migratetype];
1115                } while (list_empty(list));
1116
1117                /* This is the only non-empty list. Free them all. */
1118                if (batch_free == MIGRATE_PCPTYPES)
1119                        batch_free = count;
1120
1121                do {
1122                        int mt; /* migratetype of the to-be-freed page */
1123
1124                        page = list_last_entry(list, struct page, lru);
1125                        /* must delete as __free_one_page list manipulates */
1126                        list_del(&page->lru);
1127
1128                        mt = get_pcppage_migratetype(page);
1129                        /* MIGRATE_ISOLATE page should not go to pcplists */
1130                        VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1131                        /* Pageblock could have been isolated meanwhile */
1132                        if (unlikely(isolated_pageblocks))
1133                                mt = get_pageblock_migratetype(page);
1134
1135                        if (bulkfree_pcp_prepare(page))
1136                                continue;
1137
1138                        __free_one_page(page, page_to_pfn(page), zone, 0, mt);
1139                        trace_mm_page_pcpu_drain(page, 0, mt);
1140                } while (--count && --batch_free && !list_empty(list));
1141        }
1142        spin_unlock(&zone->lock);
1143}
1144
1145static void free_one_page(struct zone *zone,
1146                                struct page *page, unsigned long pfn,
1147                                unsigned int order,
1148                                int migratetype)
1149{
1150        unsigned long nr_scanned;
1151        spin_lock(&zone->lock);
1152        nr_scanned = node_page_state(zone->zone_pgdat, NR_PAGES_SCANNED);
1153        if (nr_scanned)
1154                __mod_node_page_state(zone->zone_pgdat, NR_PAGES_SCANNED, -nr_scanned);
1155
1156        if (unlikely(has_isolate_pageblock(zone) ||
1157                is_migrate_isolate(migratetype))) {
1158                migratetype = get_pfnblock_migratetype(page, pfn);
1159        }
1160        __free_one_page(page, pfn, zone, order, migratetype);
1161        spin_unlock(&zone->lock);
1162}
1163
1164static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1165                                unsigned long zone, int nid)
1166{
1167        set_page_links(page, zone, nid, pfn);
1168        init_page_count(page);
1169        page_mapcount_reset(page);
1170        page_cpupid_reset_last(page);
1171
1172        INIT_LIST_HEAD(&page->lru);
1173#ifdef WANT_PAGE_VIRTUAL
1174        /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1175        if (!is_highmem_idx(zone))
1176                set_page_address(page, __va(pfn << PAGE_SHIFT));
1177#endif
1178}
1179
1180static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
1181                                        int nid)
1182{
1183        return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
1184}
1185
1186#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1187static void init_reserved_page(unsigned long pfn)
1188{
1189        pg_data_t *pgdat;
1190        int nid, zid;
1191
1192        if (!early_page_uninitialised(pfn))
1193                return;
1194
1195        nid = early_pfn_to_nid(pfn);
1196        pgdat = NODE_DATA(nid);
1197
1198        for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1199                struct zone *zone = &pgdat->node_zones[zid];
1200
1201                if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1202                        break;
1203        }
1204        __init_single_pfn(pfn, zid, nid);
1205}
1206#else
1207static inline void init_reserved_page(unsigned long pfn)
1208{
1209}
1210#endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1211
1212/*
1213 * Initialised pages do not have PageReserved set. This function is
1214 * called for each range allocated by the bootmem allocator and
1215 * marks the pages PageReserved. The remaining valid pages are later
1216 * sent to the buddy page allocator.
1217 */
1218void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1219{
1220        unsigned long start_pfn = PFN_DOWN(start);
1221        unsigned long end_pfn = PFN_UP(end);
1222
1223        for (; start_pfn < end_pfn; start_pfn++) {
1224                if (pfn_valid(start_pfn)) {
1225                        struct page *page = pfn_to_page(start_pfn);
1226
1227                        init_reserved_page(start_pfn);
1228
1229                        /* Avoid false-positive PageTail() */
1230                        INIT_LIST_HEAD(&page->lru);
1231
1232                        SetPageReserved(page);
1233                }
1234        }
1235}
1236
1237static void __free_pages_ok(struct page *page, unsigned int order)
1238{
1239        unsigned long flags;
1240        int migratetype;
1241        unsigned long pfn = page_to_pfn(page);
1242
1243        if (!free_pages_prepare(page, order, true))
1244                return;
1245
1246        migratetype = get_pfnblock_migratetype(page, pfn);
1247        local_irq_save(flags);
1248        __count_vm_events(PGFREE, 1 << order);
1249        free_one_page(page_zone(page), page, pfn, order, migratetype);
1250        local_irq_restore(flags);
1251}
1252
1253static void __init __free_pages_boot_core(struct page *page, unsigned int order)
1254{
1255        unsigned int nr_pages = 1 << order;
1256        struct page *p = page;
1257        unsigned int loop;
1258
1259        prefetchw(p);
1260        for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1261                prefetchw(p + 1);
1262                __ClearPageReserved(p);
1263                set_page_count(p, 0);
1264        }
1265        __ClearPageReserved(p);
1266        set_page_count(p, 0);
1267
1268        page_zone(page)->managed_pages += nr_pages;
1269        set_page_refcounted(page);
1270        __free_pages(page, order);
1271}
1272
1273#if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1274        defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1275
1276static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1277
1278int __meminit early_pfn_to_nid(unsigned long pfn)
1279{
1280        static DEFINE_SPINLOCK(early_pfn_lock);
1281        int nid;
1282
1283        spin_lock(&early_pfn_lock);
1284        nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1285        if (nid < 0)
1286                nid = first_online_node;
1287        spin_unlock(&early_pfn_lock);
1288
1289        return nid;
1290}
1291#endif
1292
1293#ifdef CONFIG_NODES_SPAN_OTHER_NODES
1294static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1295                                        struct mminit_pfnnid_cache *state)
1296{
1297        int nid;
1298
1299        nid = __early_pfn_to_nid(pfn, state);
1300        if (nid >= 0 && nid != node)
1301                return false;
1302        return true;
1303}
1304
1305/* Only safe to use early in boot when initialisation is single-threaded */
1306static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1307{
1308        return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1309}
1310
1311#else
1312
1313static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1314{
1315        return true;
1316}
1317static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1318                                        struct mminit_pfnnid_cache *state)
1319{
1320        return true;
1321}
1322#endif
1323
1324
1325void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1326                                                        unsigned int order)
1327{
1328        if (early_page_uninitialised(pfn))
1329                return;
1330        return __free_pages_boot_core(page, order);
1331}
1332
1333/*
1334 * Check that the whole (or subset of) a pageblock given by the interval of
1335 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1336 * with the migration of free compaction scanner. The scanners then need to
1337 * use only pfn_valid_within() check for arches that allow holes within
1338 * pageblocks.
1339 *
1340 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1341 *
1342 * It's possible on some configurations to have a setup like node0 node1 node0
1343 * i.e. it's possible that all pages within a zones range of pages do not
1344 * belong to a single zone. We assume that a border between node0 and node1
1345 * can occur within a single pageblock, but not a node0 node1 node0
1346 * interleaving within a single pageblock. It is therefore sufficient to check
1347 * the first and last page of a pageblock and avoid checking each individual
1348 * page in a pageblock.
1349 */
1350struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1351                                     unsigned long end_pfn, struct zone *zone)
1352{
1353        struct page *start_page;
1354        struct page *end_page;
1355
1356        /* end_pfn is one past the range we are checking */
1357        end_pfn--;
1358
1359        if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1360                return NULL;
1361
1362        start_page = pfn_to_page(start_pfn);
1363
1364        if (page_zone(start_page) != zone)
1365                return NULL;
1366
1367        end_page = pfn_to_page(end_pfn);
1368
1369        /* This gives a shorter code than deriving page_zone(end_page) */
1370        if (page_zone_id(start_page) != page_zone_id(end_page))
1371                return NULL;
1372
1373        return start_page;
1374}
1375
1376void set_zone_contiguous(struct zone *zone)
1377{
1378        unsigned long block_start_pfn = zone->zone_start_pfn;
1379        unsigned long block_end_pfn;
1380
1381        block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1382        for (; block_start_pfn < zone_end_pfn(zone);
1383                        block_start_pfn = block_end_pfn,
1384                         block_end_pfn += pageblock_nr_pages) {
1385
1386                block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1387
1388                if (!__pageblock_pfn_to_page(block_start_pfn,
1389                                             block_end_pfn, zone))
1390                        return;
1391        }
1392
1393        /* We confirm that there is no hole */
1394        zone->contiguous = true;
1395}
1396
1397void clear_zone_contiguous(struct zone *zone)
1398{
1399        zone->contiguous = false;
1400}
1401
1402#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1403static void __init deferred_free_range(struct page *page,
1404                                        unsigned long pfn, int nr_pages)
1405{
1406        int i;
1407
1408        if (!page)
1409                return;
1410
1411        /* Free a large naturally-aligned chunk if possible */
1412        if (nr_pages == pageblock_nr_pages &&
1413            (pfn & (pageblock_nr_pages - 1)) == 0) {
1414                set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1415                __free_pages_boot_core(page, pageblock_order);
1416                return;
1417        }
1418
1419        for (i = 0; i < nr_pages; i++, page++, pfn++) {
1420                if ((pfn & (pageblock_nr_pages - 1)) == 0)
1421                        set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1422                __free_pages_boot_core(page, 0);
1423        }
1424}
1425
1426/* Completion tracking for deferred_init_memmap() threads */
1427static atomic_t pgdat_init_n_undone __initdata;
1428static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1429
1430static inline void __init pgdat_init_report_one_done(void)
1431{
1432        if (atomic_dec_and_test(&pgdat_init_n_undone))
1433                complete(&pgdat_init_all_done_comp);
1434}
1435
1436/* Initialise remaining memory on a node */
1437static int __init deferred_init_memmap(void *data)
1438{
1439        pg_data_t *pgdat = data;
1440        int nid = pgdat->node_id;
1441        struct mminit_pfnnid_cache nid_init_state = { };
1442        unsigned long start = jiffies;
1443        unsigned long nr_pages = 0;
1444        unsigned long walk_start, walk_end;
1445        int i, zid;
1446        struct zone *zone;
1447        unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1448        const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1449
1450        if (first_init_pfn == ULONG_MAX) {
1451                pgdat_init_report_one_done();
1452                return 0;
1453        }
1454
1455        /* Bind memory initialisation thread to a local node if possible */
1456        if (!cpumask_empty(cpumask))
1457                set_cpus_allowed_ptr(current, cpumask);
1458
1459        /* Sanity check boundaries */
1460        BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1461        BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1462        pgdat->first_deferred_pfn = ULONG_MAX;
1463
1464        /* Only the highest zone is deferred so find it */
1465        for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1466                zone = pgdat->node_zones + zid;
1467                if (first_init_pfn < zone_end_pfn(zone))
1468                        break;
1469        }
1470
1471        for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) {
1472                unsigned long pfn, end_pfn;
1473                struct page *page = NULL;
1474                struct page *free_base_page = NULL;
1475                unsigned long free_base_pfn = 0;
1476                int nr_to_free = 0;
1477
1478                end_pfn = min(walk_end, zone_end_pfn(zone));
1479                pfn = first_init_pfn;
1480                if (pfn < walk_start)
1481                        pfn = walk_start;
1482                if (pfn < zone->zone_start_pfn)
1483                        pfn = zone->zone_start_pfn;
1484
1485                for (; pfn < end_pfn; pfn++) {
1486                        if (!pfn_valid_within(pfn))
1487                                goto free_range;
1488
1489                        /*
1490                         * Ensure pfn_valid is checked every
1491                         * pageblock_nr_pages for memory holes
1492                         */
1493                        if ((pfn & (pageblock_nr_pages - 1)) == 0) {
1494                                if (!pfn_valid(pfn)) {
1495                                        page = NULL;
1496                                        goto free_range;
1497                                }
1498                        }
1499
1500                        if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1501                                page = NULL;
1502                                goto free_range;
1503                        }
1504
1505                        /* Minimise pfn page lookups and scheduler checks */
1506                        if (page && (pfn & (pageblock_nr_pages - 1)) != 0) {
1507                                page++;
1508                        } else {
1509                                nr_pages += nr_to_free;
1510                                deferred_free_range(free_base_page,
1511                                                free_base_pfn, nr_to_free);
1512                                free_base_page = NULL;
1513                                free_base_pfn = nr_to_free = 0;
1514
1515                                page = pfn_to_page(pfn);
1516                                cond_resched();
1517                        }
1518
1519                        if (page->flags) {
1520                                VM_BUG_ON(page_zone(page) != zone);
1521                                goto free_range;
1522                        }
1523
1524                        __init_single_page(page, pfn, zid, nid);
1525                        if (!free_base_page) {
1526                                free_base_page = page;
1527                                free_base_pfn = pfn;
1528                                nr_to_free = 0;
1529                        }
1530                        nr_to_free++;
1531
1532                        /* Where possible, batch up pages for a single free */
1533                        continue;
1534free_range:
1535                        /* Free the current block of pages to allocator */
1536                        nr_pages += nr_to_free;
1537                        deferred_free_range(free_base_page, free_base_pfn,
1538                                                                nr_to_free);
1539                        free_base_page = NULL;
1540                        free_base_pfn = nr_to_free = 0;
1541                }
1542                /* Free the last block of pages to allocator */
1543                nr_pages += nr_to_free;
1544                deferred_free_range(free_base_page, free_base_pfn, nr_to_free);
1545
1546                first_init_pfn = max(end_pfn, first_init_pfn);
1547        }
1548
1549        /* Sanity check that the next zone really is unpopulated */
1550        WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1551
1552        pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1553                                        jiffies_to_msecs(jiffies - start));
1554
1555        pgdat_init_report_one_done();
1556        return 0;
1557}
1558#endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1559
1560void __init page_alloc_init_late(void)
1561{
1562        struct zone *zone;
1563
1564#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1565        int nid;
1566
1567        /* There will be num_node_state(N_MEMORY) threads */
1568        atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1569        for_each_node_state(nid, N_MEMORY) {
1570                kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1571        }
1572
1573        /* Block until all are initialised */
1574        wait_for_completion(&pgdat_init_all_done_comp);
1575
1576        /* Reinit limits that are based on free pages after the kernel is up */
1577        files_maxfiles_init();
1578#endif
1579
1580        for_each_populated_zone(zone)
1581                set_zone_contiguous(zone);
1582}
1583
1584#ifdef CONFIG_CMA
1585/* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1586void __init init_cma_reserved_pageblock(struct page *page)
1587{
1588        unsigned i = pageblock_nr_pages;
1589        struct page *p = page;
1590
1591        do {
1592                __ClearPageReserved(p);
1593                set_page_count(p, 0);
1594        } while (++p, --i);
1595
1596        set_pageblock_migratetype(page, MIGRATE_CMA);
1597
1598        if (pageblock_order >= MAX_ORDER) {
1599                i = pageblock_nr_pages;
1600                p = page;
1601                do {
1602                        set_page_refcounted(p);
1603                        __free_pages(p, MAX_ORDER - 1);
1604                        p += MAX_ORDER_NR_PAGES;
1605                } while (i -= MAX_ORDER_NR_PAGES);
1606        } else {
1607                set_page_refcounted(page);
1608                __free_pages(page, pageblock_order);
1609        }
1610
1611        adjust_managed_page_count(page, pageblock_nr_pages);
1612}
1613#endif
1614
1615/*
1616 * The order of subdivision here is critical for the IO subsystem.
1617 * Please do not alter this order without good reasons and regression
1618 * testing. Specifically, as large blocks of memory are subdivided,
1619 * the order in which smaller blocks are delivered depends on the order
1620 * they're subdivided in this function. This is the primary factor
1621 * influencing the order in which pages are delivered to the IO
1622 * subsystem according to empirical testing, and this is also justified
1623 * by considering the behavior of a buddy system containing a single
1624 * large block of memory acted on by a series of small allocations.
1625 * This behavior is a critical factor in sglist merging's success.
1626 *
1627 * -- nyc
1628 */
1629static inline void expand(struct zone *zone, struct page *page,
1630        int low, int high, struct free_area *area,
1631        int migratetype)
1632{
1633        unsigned long size = 1 << high;
1634
1635        while (high > low) {
1636                area--;
1637                high--;
1638                size >>= 1;
1639                VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1640
1641                /*
1642                 * Mark as guard pages (or page), that will allow to
1643                 * merge back to allocator when buddy will be freed.
1644                 * Corresponding page table entries will not be touched,
1645                 * pages will stay not present in virtual address space
1646                 */
1647                if (set_page_guard(zone, &page[size], high, migratetype))
1648                        continue;
1649
1650                list_add(&page[size].lru, &area->free_list[migratetype]);
1651                area->nr_free++;
1652                set_page_order(&page[size], high);
1653        }
1654}
1655
1656static void check_new_page_bad(struct page *page)
1657{
1658        const char *bad_reason = NULL;
1659        unsigned long bad_flags = 0;
1660
1661        if (unlikely(atomic_read(&page->_mapcount) != -1))
1662                bad_reason = "nonzero mapcount";
1663        if (unlikely(page->mapping != NULL))
1664                bad_reason = "non-NULL mapping";
1665        if (unlikely(page_ref_count(page) != 0))
1666                bad_reason = "nonzero _count";
1667        if (unlikely(page->flags & __PG_HWPOISON)) {
1668                bad_reason = "HWPoisoned (hardware-corrupted)";
1669                bad_flags = __PG_HWPOISON;
1670                /* Don't complain about hwpoisoned pages */
1671                page_mapcount_reset(page); /* remove PageBuddy */
1672                return;
1673        }
1674        if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1675                bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1676                bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1677        }
1678#ifdef CONFIG_MEMCG
1679        if (unlikely(page->mem_cgroup))
1680                bad_reason = "page still charged to cgroup";
1681#endif
1682        bad_page(page, bad_reason, bad_flags);
1683}
1684
1685/*
1686 * This page is about to be returned from the page allocator
1687 */
1688static inline int check_new_page(struct page *page)
1689{
1690        if (likely(page_expected_state(page,
1691                                PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
1692                return 0;
1693
1694        check_new_page_bad(page);
1695        return 1;
1696}
1697
1698static inline bool free_pages_prezeroed(bool poisoned)
1699{
1700        return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
1701                page_poisoning_enabled() && poisoned;
1702}
1703
1704#ifdef CONFIG_DEBUG_VM
1705static bool check_pcp_refill(struct page *page)
1706{
1707        return false;
1708}
1709
1710static bool check_new_pcp(struct page *page)
1711{
1712        return check_new_page(page);
1713}
1714#else
1715static bool check_pcp_refill(struct page *page)
1716{
1717        return check_new_page(page);
1718}
1719static bool check_new_pcp(struct page *page)
1720{
1721        return false;
1722}
1723#endif /* CONFIG_DEBUG_VM */
1724
1725static bool check_new_pages(struct page *page, unsigned int order)
1726{
1727        int i;
1728        for (i = 0; i < (1 << order); i++) {
1729                struct page *p = page + i;
1730
1731                if (unlikely(check_new_page(p)))
1732                        return true;
1733        }
1734
1735        return false;
1736}
1737
1738inline void post_alloc_hook(struct page *page, unsigned int order,
1739                                gfp_t gfp_flags)
1740{
1741        set_page_private(page, 0);
1742        set_page_refcounted(page);
1743
1744        arch_alloc_page(page, order);
1745        kernel_map_pages(page, 1 << order, 1);
1746        kernel_poison_pages(page, 1 << order, 1);
1747        kasan_alloc_pages(page, order);
1748        set_page_owner(page, order, gfp_flags);
1749}
1750
1751static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1752                                                        unsigned int alloc_flags)
1753{
1754        int i;
1755        bool poisoned = true;
1756
1757        for (i = 0; i < (1 << order); i++) {
1758                struct page *p = page + i;
1759                if (poisoned)
1760                        poisoned &= page_is_poisoned(p);
1761        }
1762
1763        post_alloc_hook(page, order, gfp_flags);
1764
1765        if (!free_pages_prezeroed(poisoned) && (gfp_flags & __GFP_ZERO))
1766                for (i = 0; i < (1 << order); i++)
1767                        clear_highpage(page + i);
1768
1769        if (order && (gfp_flags & __GFP_COMP))
1770                prep_compound_page(page, order);
1771
1772        /*
1773         * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1774         * allocate the page. The expectation is that the caller is taking
1775         * steps that will free more memory. The caller should avoid the page
1776         * being used for !PFMEMALLOC purposes.
1777         */
1778        if (alloc_flags & ALLOC_NO_WATERMARKS)
1779                set_page_pfmemalloc(page);
1780        else
1781                clear_page_pfmemalloc(page);
1782}
1783
1784/*
1785 * Go through the free lists for the given migratetype and remove
1786 * the smallest available page from the freelists
1787 */
1788static inline
1789struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1790                                                int migratetype)
1791{
1792        unsigned int current_order;
1793        struct free_area *area;
1794        struct page *page;
1795
1796        /* Find a page of the appropriate size in the preferred list */
1797        for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1798                area = &(zone->free_area[current_order]);
1799                page = list_first_entry_or_null(&area->free_list[migratetype],
1800                                                        struct page, lru);
1801                if (!page)
1802                        continue;
1803                list_del(&page->lru);
1804                rmv_page_order(page);
1805                area->nr_free--;
1806                expand(zone, page, order, current_order, area, migratetype);
1807                set_pcppage_migratetype(page, migratetype);
1808                return page;
1809        }
1810
1811        return NULL;
1812}
1813
1814
1815/*
1816 * This array describes the order lists are fallen back to when
1817 * the free lists for the desirable migrate type are depleted
1818 */
1819static int fallbacks[MIGRATE_TYPES][4] = {
1820        [MIGRATE_UNMOVABLE]   = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE,   MIGRATE_TYPES },
1821        [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE,   MIGRATE_MOVABLE,   MIGRATE_TYPES },
1822        [MIGRATE_MOVABLE]     = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1823#ifdef CONFIG_CMA
1824        [MIGRATE_CMA]         = { MIGRATE_TYPES }, /* Never used */
1825#endif
1826#ifdef CONFIG_MEMORY_ISOLATION
1827        [MIGRATE_ISOLATE]     = { MIGRATE_TYPES }, /* Never used */
1828#endif
1829};
1830
1831#ifdef CONFIG_CMA
1832static struct page *__rmqueue_cma_fallback(struct zone *zone,
1833                                        unsigned int order)
1834{
1835        return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1836}
1837#else
1838static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1839                                        unsigned int order) { return NULL; }
1840#endif
1841
1842/*
1843 * Move the free pages in a range to the free lists of the requested type.
1844 * Note that start_page and end_pages are not aligned on a pageblock
1845 * boundary. If alignment is required, use move_freepages_block()
1846 */
1847int move_freepages(struct zone *zone,
1848                          struct page *start_page, struct page *end_page,
1849                          int migratetype)
1850{
1851        struct page *page;
1852        unsigned int order;
1853        int pages_moved = 0;
1854
1855#ifndef CONFIG_HOLES_IN_ZONE
1856        /*
1857         * page_zone is not safe to call in this context when
1858         * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1859         * anyway as we check zone boundaries in move_freepages_block().
1860         * Remove at a later date when no bug reports exist related to
1861         * grouping pages by mobility
1862         */
1863        VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1864#endif
1865
1866        for (page = start_page; page <= end_page;) {
1867                /* Make sure we are not inadvertently changing nodes */
1868                VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1869
1870                if (!pfn_valid_within(page_to_pfn(page))) {
1871                        page++;
1872                        continue;
1873                }
1874
1875                if (!PageBuddy(page)) {
1876                        page++;
1877                        continue;
1878                }
1879
1880                order = page_order(page);
1881                list_move(&page->lru,
1882                          &zone->free_area[order].free_list[migratetype]);
1883                page += 1 << order;
1884                pages_moved += 1 << order;
1885        }
1886
1887        return pages_moved;
1888}
1889
1890int move_freepages_block(struct zone *zone, struct page *page,
1891                                int migratetype)
1892{
1893        unsigned long start_pfn, end_pfn;
1894        struct page *start_page, *end_page;
1895
1896        start_pfn = page_to_pfn(page);
1897        start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1898        start_page = pfn_to_page(start_pfn);
1899        end_page = start_page + pageblock_nr_pages - 1;
1900        end_pfn = start_pfn + pageblock_nr_pages - 1;
1901
1902        /* Do not cross zone boundaries */
1903        if (!zone_spans_pfn(zone, start_pfn))
1904                start_page = page;
1905        if (!zone_spans_pfn(zone, end_pfn))
1906                return 0;
1907
1908        return move_freepages(zone, start_page, end_page, migratetype);
1909}
1910
1911static void change_pageblock_range(struct page *pageblock_page,
1912                                        int start_order, int migratetype)
1913{
1914        int nr_pageblocks = 1 << (start_order - pageblock_order);
1915
1916        while (nr_pageblocks--) {
1917                set_pageblock_migratetype(pageblock_page, migratetype);
1918                pageblock_page += pageblock_nr_pages;
1919        }
1920}
1921
1922/*
1923 * When we are falling back to another migratetype during allocation, try to
1924 * steal extra free pages from the same pageblocks to satisfy further
1925 * allocations, instead of polluting multiple pageblocks.
1926 *
1927 * If we are stealing a relatively large buddy page, it is likely there will
1928 * be more free pages in the pageblock, so try to steal them all. For
1929 * reclaimable and unmovable allocations, we steal regardless of page size,
1930 * as fragmentation caused by those allocations polluting movable pageblocks
1931 * is worse than movable allocations stealing from unmovable and reclaimable
1932 * pageblocks.
1933 */
1934static bool can_steal_fallback(unsigned int order, int start_mt)
1935{
1936        /*
1937         * Leaving this order check is intended, although there is
1938         * relaxed order check in next check. The reason is that
1939         * we can actually steal whole pageblock if this condition met,
1940         * but, below check doesn't guarantee it and that is just heuristic
1941         * so could be changed anytime.
1942         */
1943        if (order >= pageblock_order)
1944                return true;
1945
1946        if (order >= pageblock_order / 2 ||
1947                start_mt == MIGRATE_RECLAIMABLE ||
1948                start_mt == MIGRATE_UNMOVABLE ||
1949                page_group_by_mobility_disabled)
1950                return true;
1951
1952        return false;
1953}
1954
1955/*
1956 * This function implements actual steal behaviour. If order is large enough,
1957 * we can steal whole pageblock. If not, we first move freepages in this
1958 * pageblock and check whether half of pages are moved or not. If half of
1959 * pages are moved, we can change migratetype of pageblock and permanently
1960 * use it's pages as requested migratetype in the future.
1961 */
1962static void steal_suitable_fallback(struct zone *zone, struct page *page,
1963                                                          int start_type)
1964{
1965        unsigned int current_order = page_order(page);
1966        int pages;
1967
1968        /* Take ownership for orders >= pageblock_order */
1969        if (current_order >= pageblock_order) {
1970                change_pageblock_range(page, current_order, start_type);
1971                return;
1972        }
1973
1974        pages = move_freepages_block(zone, page, start_type);
1975
1976        /* Claim the whole block if over half of it is free */
1977        if (pages >= (1 << (pageblock_order-1)) ||
1978                        page_group_by_mobility_disabled)
1979                set_pageblock_migratetype(page, start_type);
1980}
1981
1982/*
1983 * Check whether there is a suitable fallback freepage with requested order.
1984 * If only_stealable is true, this function returns fallback_mt only if
1985 * we can steal other freepages all together. This would help to reduce
1986 * fragmentation due to mixed migratetype pages in one pageblock.
1987 */
1988int find_suitable_fallback(struct free_area *area, unsigned int order,
1989                        int migratetype, bool only_stealable, bool *can_steal)
1990{
1991        int i;
1992        int fallback_mt;
1993
1994        if (area->nr_free == 0)
1995                return -1;
1996
1997        *can_steal = false;
1998        for (i = 0;; i++) {
1999                fallback_mt = fallbacks[migratetype][i];
2000                if (fallback_mt == MIGRATE_TYPES)
2001                        break;
2002
2003                if (list_empty(&area->free_list[fallback_mt]))
2004                        continue;
2005
2006                if (can_steal_fallback(order, migratetype))
2007                        *can_steal = true;
2008
2009                if (!only_stealable)
2010                        return fallback_mt;
2011
2012                if (*can_steal)
2013                        return fallback_mt;
2014        }
2015
2016        return -1;
2017}
2018
2019/*
2020 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2021 * there are no empty page blocks that contain a page with a suitable order
2022 */
2023static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2024                                unsigned int alloc_order)
2025{
2026        int mt;
2027        unsigned long max_managed, flags;
2028
2029        /*
2030         * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2031         * Check is race-prone but harmless.
2032         */
2033        max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
2034        if (zone->nr_reserved_highatomic >= max_managed)
2035                return;
2036
2037        spin_lock_irqsave(&zone->lock, flags);
2038
2039        /* Recheck the nr_reserved_highatomic limit under the lock */
2040        if (zone->nr_reserved_highatomic >= max_managed)
2041                goto out_unlock;
2042
2043        /* Yoink! */
2044        mt = get_pageblock_migratetype(page);
2045        if (mt != MIGRATE_HIGHATOMIC &&
2046                        !is_migrate_isolate(mt) && !is_migrate_cma(mt)) {
2047                zone->nr_reserved_highatomic += pageblock_nr_pages;
2048                set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2049                move_freepages_block(zone, page, MIGRATE_HIGHATOMIC);
2050        }
2051
2052out_unlock:
2053        spin_unlock_irqrestore(&zone->lock, flags);
2054}
2055
2056/*
2057 * Used when an allocation is about to fail under memory pressure. This
2058 * potentially hurts the reliability of high-order allocations when under
2059 * intense memory pressure but failed atomic allocations should be easier
2060 * to recover from than an OOM.
2061 */
2062static void unreserve_highatomic_pageblock(const struct alloc_context *ac)
2063{
2064        struct zonelist *zonelist = ac->zonelist;
2065        unsigned long flags;
2066        struct zoneref *z;
2067        struct zone *zone;
2068        struct page *page;
2069        int order;
2070
2071        for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2072                                                                ac->nodemask) {
2073                /* Preserve at least one pageblock */
2074                if (zone->nr_reserved_highatomic <= pageblock_nr_pages)
2075                        continue;
2076
2077                spin_lock_irqsave(&zone->lock, flags);
2078                for (order = 0; order < MAX_ORDER; order++) {
2079                        struct free_area *area = &(zone->free_area[order]);
2080
2081                        page = list_first_entry_or_null(
2082                                        &area->free_list[MIGRATE_HIGHATOMIC],
2083                                        struct page, lru);
2084                        if (!page)
2085                                continue;
2086
2087                        /*
2088                         * It should never happen but changes to locking could
2089                         * inadvertently allow a per-cpu drain to add pages
2090                         * to MIGRATE_HIGHATOMIC while unreserving so be safe
2091                         * and watch for underflows.
2092                         */
2093                        zone->nr_reserved_highatomic -= min(pageblock_nr_pages,
2094                                zone->nr_reserved_highatomic);
2095
2096                        /*
2097                         * Convert to ac->migratetype and avoid the normal
2098                         * pageblock stealing heuristics. Minimally, the caller
2099                         * is doing the work and needs the pages. More
2100                         * importantly, if the block was always converted to
2101                         * MIGRATE_UNMOVABLE or another type then the number
2102                         * of pageblocks that cannot be completely freed
2103                         * may increase.
2104                         */
2105                        set_pageblock_migratetype(page, ac->migratetype);
2106                        move_freepages_block(zone, page, ac->migratetype);
2107                        spin_unlock_irqrestore(&zone->lock, flags);
2108                        return;
2109                }
2110                spin_unlock_irqrestore(&zone->lock, flags);
2111        }
2112}
2113
2114/* Remove an element from the buddy allocator from the fallback list */
2115static inline struct page *
2116__rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
2117{
2118        struct free_area *area;
2119        unsigned int current_order;
2120        struct page *page;
2121        int fallback_mt;
2122        bool can_steal;
2123
2124        /* Find the largest possible block of pages in the other list */
2125        for (current_order = MAX_ORDER-1;
2126                                current_order >= order && current_order <= MAX_ORDER-1;
2127                                --current_order) {
2128                area = &(zone->free_area[current_order]);
2129                fallback_mt = find_suitable_fallback(area, current_order,
2130                                start_migratetype, false, &can_steal);
2131                if (fallback_mt == -1)
2132                        continue;
2133
2134                page = list_first_entry(&area->free_list[fallback_mt],
2135                                                struct page, lru);
2136                if (can_steal)
2137                        steal_suitable_fallback(zone, page, start_migratetype);
2138
2139                /* Remove the page from the freelists */
2140                area->nr_free--;
2141                list_del(&page->lru);
2142                rmv_page_order(page);
2143
2144                expand(zone, page, order, current_order, area,
2145                                        start_migratetype);
2146                /*
2147                 * The pcppage_migratetype may differ from pageblock's
2148                 * migratetype depending on the decisions in
2149                 * find_suitable_fallback(). This is OK as long as it does not
2150                 * differ for MIGRATE_CMA pageblocks. Those can be used as
2151                 * fallback only via special __rmqueue_cma_fallback() function
2152                 */
2153                set_pcppage_migratetype(page, start_migratetype);
2154
2155                trace_mm_page_alloc_extfrag(page, order, current_order,
2156                        start_migratetype, fallback_mt);
2157
2158                return page;
2159        }
2160
2161        return NULL;
2162}
2163
2164/*
2165 * Do the hard work of removing an element from the buddy allocator.
2166 * Call me with the zone->lock already held.
2167 */
2168static struct page *__rmqueue(struct zone *zone, unsigned int order,
2169                                int migratetype)
2170{
2171        struct page *page;
2172
2173        page = __rmqueue_smallest(zone, order, migratetype);
2174        if (unlikely(!page)) {
2175                if (migratetype == MIGRATE_MOVABLE)
2176                        page = __rmqueue_cma_fallback(zone, order);
2177
2178                if (!page)
2179                        page = __rmqueue_fallback(zone, order, migratetype);
2180        }
2181
2182        trace_mm_page_alloc_zone_locked(page, order, migratetype);
2183        return page;
2184}
2185
2186/*
2187 * Obtain a specified number of elements from the buddy allocator, all under
2188 * a single hold of the lock, for efficiency.  Add them to the supplied list.
2189 * Returns the number of new pages which were placed at *list.
2190 */
2191static int rmqueue_bulk(struct zone *zone, unsigned int order,
2192                        unsigned long count, struct list_head *list,
2193                        int migratetype, bool cold)
2194{
2195        int i, alloced = 0;
2196
2197        spin_lock(&zone->lock);
2198        for (i = 0; i < count; ++i) {
2199                struct page *page = __rmqueue(zone, order, migratetype);
2200                if (unlikely(page == NULL))
2201                        break;
2202
2203                if (unlikely(check_pcp_refill(page)))
2204                        continue;
2205
2206                /*
2207                 * Split buddy pages returned by expand() are received here
2208                 * in physical page order. The page is added to the callers and
2209                 * list and the list head then moves forward. From the callers
2210                 * perspective, the linked list is ordered by page number in
2211                 * some conditions. This is useful for IO devices that can
2212                 * merge IO requests if the physical pages are ordered
2213                 * properly.
2214                 */
2215                if (likely(!cold))
2216                        list_add(&page->lru, list);
2217                else
2218                        list_add_tail(&page->lru, list);
2219                list = &page->lru;
2220                alloced++;
2221                if (is_migrate_cma(get_pcppage_migratetype(page)))
2222                        __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2223                                              -(1 << order));
2224        }
2225
2226        /*
2227         * i pages were removed from the buddy list even if some leak due
2228         * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2229         * on i. Do not confuse with 'alloced' which is the number of
2230         * pages added to the pcp list.
2231         */
2232        __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2233        spin_unlock(&zone->lock);
2234        return alloced;
2235}
2236
2237#ifdef CONFIG_NUMA
2238/*
2239 * Called from the vmstat counter updater to drain pagesets of this
2240 * currently executing processor on remote nodes after they have
2241 * expired.
2242 *
2243 * Note that this function must be called with the thread pinned to
2244 * a single processor.
2245 */
2246void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2247{
2248        unsigned long flags;
2249        int to_drain, batch;
2250
2251        local_irq_save(flags);
2252        batch = READ_ONCE(pcp->batch);
2253        to_drain = min(pcp->count, batch);
2254        if (to_drain > 0) {
2255                free_pcppages_bulk(zone, to_drain, pcp);
2256                pcp->count -= to_drain;
2257        }
2258        local_irq_restore(flags);
2259}
2260#endif
2261
2262/*
2263 * Drain pcplists of the indicated processor and zone.
2264 *
2265 * The processor must either be the current processor and the
2266 * thread pinned to the current processor or a processor that
2267 * is not online.
2268 */
2269static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2270{
2271        unsigned long flags;
2272        struct per_cpu_pageset *pset;
2273        struct per_cpu_pages *pcp;
2274
2275        local_irq_save(flags);
2276        pset = per_cpu_ptr(zone->pageset, cpu);
2277
2278        pcp = &pset->pcp;
2279        if (pcp->count) {
2280                free_pcppages_bulk(zone, pcp->count, pcp);
2281                pcp->count = 0;
2282        }
2283        local_irq_restore(flags);
2284}
2285
2286/*
2287 * Drain pcplists of all zones on the indicated processor.
2288 *
2289 * The processor must either be the current processor and the
2290 * thread pinned to the current processor or a processor that
2291 * is not online.
2292 */
2293static void drain_pages(unsigned int cpu)
2294{
2295        struct zone *zone;
2296
2297        for_each_populated_zone(zone) {
2298                drain_pages_zone(cpu, zone);
2299        }
2300}
2301
2302/*
2303 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2304 *
2305 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2306 * the single zone's pages.
2307 */
2308void drain_local_pages(struct zone *zone)
2309{
2310        int cpu = smp_processor_id();
2311
2312        if (zone)
2313                drain_pages_zone(cpu, zone);
2314        else
2315                drain_pages(cpu);
2316}
2317
2318/*
2319 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2320 *
2321 * When zone parameter is non-NULL, spill just the single zone's pages.
2322 *
2323 * Note that this code is protected against sending an IPI to an offline
2324 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
2325 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
2326 * nothing keeps CPUs from showing up after we populated the cpumask and
2327 * before the call to on_each_cpu_mask().
2328 */
2329void drain_all_pages(struct zone *zone)
2330{
2331        int cpu;
2332
2333        /*
2334         * Allocate in the BSS so we wont require allocation in
2335         * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2336         */
2337        static cpumask_t cpus_with_pcps;
2338
2339        /*
2340         * We don't care about racing with CPU hotplug event
2341         * as offline notification will cause the notified
2342         * cpu to drain that CPU pcps and on_each_cpu_mask
2343         * disables preemption as part of its processing
2344         */
2345        for_each_online_cpu(cpu) {
2346                struct per_cpu_pageset *pcp;
2347                struct zone *z;
2348                bool has_pcps = false;
2349
2350                if (zone) {
2351                        pcp = per_cpu_ptr(zone->pageset, cpu);
2352                        if (pcp->pcp.count)
2353                                has_pcps = true;
2354                } else {
2355                        for_each_populated_zone(z) {
2356                                pcp = per_cpu_ptr(z->pageset, cpu);
2357                                if (pcp->pcp.count) {
2358                                        has_pcps = true;
2359                                        break;
2360                                }
2361                        }
2362                }
2363
2364                if (has_pcps)
2365                        cpumask_set_cpu(cpu, &cpus_with_pcps);
2366                else
2367                        cpumask_clear_cpu(cpu, &cpus_with_pcps);
2368        }
2369        on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages,
2370                                                                zone, 1);
2371}
2372
2373#ifdef CONFIG_HIBERNATION
2374
2375void mark_free_pages(struct zone *zone)
2376{
2377        unsigned long pfn, max_zone_pfn;
2378        unsigned long flags;
2379        unsigned int order, t;
2380        struct page *page;
2381
2382        if (zone_is_empty(zone))
2383                return;
2384
2385        spin_lock_irqsave(&zone->lock, flags);
2386
2387        max_zone_pfn = zone_end_pfn(zone);
2388        for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2389                if (pfn_valid(pfn)) {
2390                        page = pfn_to_page(pfn);
2391
2392                        if (page_zone(page) != zone)
2393                                continue;
2394
2395                        if (!swsusp_page_is_forbidden(page))
2396                                swsusp_unset_page_free(page);
2397                }
2398
2399        for_each_migratetype_order(order, t) {
2400                list_for_each_entry(page,
2401                                &zone->free_area[order].free_list[t], lru) {
2402                        unsigned long i;
2403
2404                        pfn = page_to_pfn(page);
2405                        for (i = 0; i < (1UL << order); i++)
2406                                swsusp_set_page_free(pfn_to_page(pfn + i));
2407                }
2408        }
2409        spin_unlock_irqrestore(&zone->lock, flags);
2410}
2411#endif /* CONFIG_PM */
2412
2413/*
2414 * Free a 0-order page
2415 * cold == true ? free a cold page : free a hot page
2416 */
2417void free_hot_cold_page(struct page *page, bool cold)
2418{
2419        struct zone *zone = page_zone(page);
2420        struct per_cpu_pages *pcp;
2421        unsigned long flags;
2422        unsigned long pfn = page_to_pfn(page);
2423        int migratetype;
2424
2425        if (!free_pcp_prepare(page))
2426                return;
2427
2428        migratetype = get_pfnblock_migratetype(page, pfn);
2429        set_pcppage_migratetype(page, migratetype);
2430        local_irq_save(flags);
2431        __count_vm_event(PGFREE);
2432
2433        /*
2434         * We only track unmovable, reclaimable and movable on pcp lists.
2435         * Free ISOLATE pages back to the allocator because they are being
2436         * offlined but treat RESERVE as movable pages so we can get those
2437         * areas back if necessary. Otherwise, we may have to free
2438         * excessively into the page allocator
2439         */
2440        if (migratetype >= MIGRATE_PCPTYPES) {
2441                if (unlikely(is_migrate_isolate(migratetype))) {
2442                        free_one_page(zone, page, pfn, 0, migratetype);
2443                        goto out;
2444                }
2445                migratetype = MIGRATE_MOVABLE;
2446        }
2447
2448        pcp = &this_cpu_ptr(zone->pageset)->pcp;
2449        if (!cold)
2450                list_add(&page->lru, &pcp->lists[migratetype]);
2451        else
2452                list_add_tail(&page->lru, &pcp->lists[migratetype]);
2453        pcp->count++;
2454        if (pcp->count >= pcp->high) {
2455                unsigned long batch = READ_ONCE(pcp->batch);
2456                free_pcppages_bulk(zone, batch, pcp);
2457                pcp->count -= batch;
2458        }
2459
2460out:
2461        local_irq_restore(flags);
2462}
2463
2464/*
2465 * Free a list of 0-order pages
2466 */
2467void free_hot_cold_page_list(struct list_head *list, bool cold)
2468{
2469        struct page *page, *next;
2470
2471        list_for_each_entry_safe(page, next, list, lru) {
2472                trace_mm_page_free_batched(page, cold);
2473                free_hot_cold_page(page, cold);
2474        }
2475}
2476
2477/*
2478 * split_page takes a non-compound higher-order page, and splits it into
2479 * n (1<<order) sub-pages: page[0..n]
2480 * Each sub-page must be freed individually.
2481 *
2482 * Note: this is probably too low level an operation for use in drivers.
2483 * Please consult with lkml before using this in your driver.
2484 */
2485void split_page(struct page *page, unsigned int order)
2486{
2487        int i;
2488
2489        VM_BUG_ON_PAGE(PageCompound(page), page);
2490        VM_BUG_ON_PAGE(!page_count(page), page);
2491
2492#ifdef CONFIG_KMEMCHECK
2493        /*
2494         * Split shadow pages too, because free(page[0]) would
2495         * otherwise free the whole shadow.
2496         */
2497        if (kmemcheck_page_is_tracked(page))
2498                split_page(virt_to_page(page[0].shadow), order);
2499#endif
2500
2501        for (i = 1; i < (1 << order); i++)
2502                set_page_refcounted(page + i);
2503        split_page_owner(page, order);
2504}
2505EXPORT_SYMBOL_GPL(split_page);
2506
2507int __isolate_free_page(struct page *page, unsigned int order)
2508{
2509        unsigned long watermark;
2510        struct zone *zone;
2511        int mt;
2512
2513        BUG_ON(!PageBuddy(page));
2514
2515        zone = page_zone(page);
2516        mt = get_pageblock_migratetype(page);
2517
2518        if (!is_migrate_isolate(mt)) {
2519                /*
2520                 * Obey watermarks as if the page was being allocated. We can
2521                 * emulate a high-order watermark check with a raised order-0
2522                 * watermark, because we already know our high-order page
2523                 * exists.
2524                 */
2525                watermark = min_wmark_pages(zone) + (1UL << order);
2526                if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
2527                        return 0;
2528
2529                __mod_zone_freepage_state(zone, -(1UL << order), mt);
2530        }
2531
2532        /* Remove page from free list */
2533        list_del(&page->lru);
2534        zone->free_area[order].nr_free--;
2535        rmv_page_order(page);
2536
2537        /*
2538         * Set the pageblock if the isolated page is at least half of a
2539         * pageblock
2540         */
2541        if (order >= pageblock_order - 1) {
2542                struct page *endpage = page + (1 << order) - 1;
2543                for (; page < endpage; page += pageblock_nr_pages) {
2544                        int mt = get_pageblock_migratetype(page);
2545                        if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
2546                                set_pageblock_migratetype(page,
2547                                                          MIGRATE_MOVABLE);
2548                }
2549        }
2550
2551
2552        return 1UL << order;
2553}
2554
2555/*
2556 * Update NUMA hit/miss statistics
2557 *
2558 * Must be called with interrupts disabled.
2559 *
2560 * When __GFP_OTHER_NODE is set assume the node of the preferred
2561 * zone is the local node. This is useful for daemons who allocate
2562 * memory on behalf of other processes.
2563 */
2564static inline void zone_statistics(struct zone *preferred_zone, struct zone *z,
2565                                                                gfp_t flags)
2566{
2567#ifdef CONFIG_NUMA
2568        int local_nid = numa_node_id();
2569        enum zone_stat_item local_stat = NUMA_LOCAL;
2570
2571        if (unlikely(flags & __GFP_OTHER_NODE)) {
2572                local_stat = NUMA_OTHER;
2573                local_nid = preferred_zone->node;
2574        }
2575
2576        if (z->node == local_nid) {
2577                __inc_zone_state(z, NUMA_HIT);
2578                __inc_zone_state(z, local_stat);
2579        } else {
2580                __inc_zone_state(z, NUMA_MISS);
2581                __inc_zone_state(preferred_zone, NUMA_FOREIGN);
2582        }
2583#endif
2584}
2585
2586/*
2587 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2588 */
2589static inline
2590struct page *buffered_rmqueue(struct zone *preferred_zone,
2591                        struct zone *zone, unsigned int order,
2592                        gfp_t gfp_flags, unsigned int alloc_flags,
2593                        int migratetype)
2594{
2595        unsigned long flags;
2596        struct page *page;
2597        bool cold = ((gfp_flags & __GFP_COLD) != 0);
2598
2599        if (likely(order == 0)) {
2600                struct per_cpu_pages *pcp;
2601                struct list_head *list;
2602
2603                local_irq_save(flags);
2604                do {
2605                        pcp = &this_cpu_ptr(zone->pageset)->pcp;
2606                        list = &pcp->lists[migratetype];
2607                        if (list_empty(list)) {
2608                                pcp->count += rmqueue_bulk(zone, 0,
2609                                                pcp->batch, list,
2610                                                migratetype, cold);
2611                                if (unlikely(list_empty(list)))
2612                                        goto failed;
2613                        }
2614
2615                        if (cold)
2616                                page = list_last_entry(list, struct page, lru);
2617                        else
2618                                page = list_first_entry(list, struct page, lru);
2619
2620                        list_del(&page->lru);
2621                        pcp->count--;
2622
2623                } while (check_new_pcp(page));
2624        } else {
2625                /*
2626                 * We most definitely don't want callers attempting to
2627                 * allocate greater than order-1 page units with __GFP_NOFAIL.
2628                 */
2629                WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
2630                spin_lock_irqsave(&zone->lock, flags);
2631
2632                do {
2633                        page = NULL;
2634                        if (alloc_flags & ALLOC_HARDER) {
2635                                page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2636                                if (page)
2637                                        trace_mm_page_alloc_zone_locked(page, order, migratetype);
2638                        }
2639                        if (!page)
2640                                page = __rmqueue(zone, order, migratetype);
2641                } while (page && check_new_pages(page, order));
2642                spin_unlock(&zone->lock);
2643                if (!page)
2644                        goto failed;
2645                __mod_zone_freepage_state(zone, -(1 << order),
2646                                          get_pcppage_migratetype(page));
2647        }
2648
2649        __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2650        zone_statistics(preferred_zone, zone, gfp_flags);
2651        local_irq_restore(flags);
2652
2653        VM_BUG_ON_PAGE(bad_range(zone, page), page);
2654        return page;
2655
2656failed:
2657        local_irq_restore(flags);
2658        return NULL;
2659}
2660
2661#ifdef CONFIG_FAIL_PAGE_ALLOC
2662
2663static struct {
2664        struct fault_attr attr;
2665
2666        bool ignore_gfp_highmem;
2667        bool ignore_gfp_reclaim;
2668        u32 min_order;
2669} fail_page_alloc = {
2670        .attr = FAULT_ATTR_INITIALIZER,
2671        .ignore_gfp_reclaim = true,
2672        .ignore_gfp_highmem = true,
2673        .min_order = 1,
2674};
2675
2676static int __init setup_fail_page_alloc(char *str)
2677{
2678        return setup_fault_attr(&fail_page_alloc.attr, str);
2679}
2680__setup("fail_page_alloc=", setup_fail_page_alloc);
2681
2682static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2683{
2684        if (order < fail_page_alloc.min_order)
2685                return false;
2686        if (gfp_mask & __GFP_NOFAIL)
2687                return false;
2688        if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2689                return false;
2690        if (fail_page_alloc.ignore_gfp_reclaim &&
2691                        (gfp_mask & __GFP_DIRECT_RECLAIM))
2692                return false;
2693
2694        return should_fail(&fail_page_alloc.attr, 1 << order);
2695}
2696
2697#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2698
2699static int __init fail_page_alloc_debugfs(void)
2700{
2701        umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2702        struct dentry *dir;
2703
2704        dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2705                                        &fail_page_alloc.attr);
2706        if (IS_ERR(dir))
2707                return PTR_ERR(dir);
2708
2709        if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2710                                &fail_page_alloc.ignore_gfp_reclaim))
2711                goto fail;
2712        if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2713                                &fail_page_alloc.ignore_gfp_highmem))
2714                goto fail;
2715        if (!debugfs_create_u32("min-order", mode, dir,
2716                                &fail_page_alloc.min_order))
2717                goto fail;
2718
2719        return 0;
2720fail:
2721        debugfs_remove_recursive(dir);
2722
2723        return -ENOMEM;
2724}
2725
2726late_initcall(fail_page_alloc_debugfs);
2727
2728#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2729
2730#else /* CONFIG_FAIL_PAGE_ALLOC */
2731
2732static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2733{
2734        return false;
2735}
2736
2737#endif /* CONFIG_FAIL_PAGE_ALLOC */
2738
2739/*
2740 * Return true if free base pages are above 'mark'. For high-order checks it
2741 * will return true of the order-0 watermark is reached and there is at least
2742 * one free page of a suitable size. Checking now avoids taking the zone lock
2743 * to check in the allocation paths if no pages are free.
2744 */
2745bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2746                         int classzone_idx, unsigned int alloc_flags,
2747                         long free_pages)
2748{
2749        long min = mark;
2750        int o;
2751        const bool alloc_harder = (alloc_flags & ALLOC_HARDER);
2752
2753        /* free_pages may go negative - that's OK */
2754        free_pages -= (1 << order) - 1;
2755
2756        if (alloc_flags & ALLOC_HIGH)
2757                min -= min / 2;
2758
2759        /*
2760         * If the caller does not have rights to ALLOC_HARDER then subtract
2761         * the high-atomic reserves. This will over-estimate the size of the
2762         * atomic reserve but it avoids a search.
2763         */
2764        if (likely(!alloc_harder))
2765                free_pages -= z->nr_reserved_highatomic;
2766        else
2767                min -= min / 4;
2768
2769#ifdef CONFIG_CMA
2770        /* If allocation can't use CMA areas don't use free CMA pages */
2771        if (!(alloc_flags & ALLOC_CMA))
2772                free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
2773#endif
2774
2775        /*
2776         * Check watermarks for an order-0 allocation request. If these
2777         * are not met, then a high-order request also cannot go ahead
2778         * even if a suitable page happened to be free.
2779         */
2780        if (free_pages <= min + z->lowmem_reserve[classzone_idx])
2781                return false;
2782
2783        /* If this is an order-0 request then the watermark is fine */
2784        if (!order)
2785                return true;
2786
2787        /* For a high-order request, check at least one suitable page is free */
2788        for (o = order; o < MAX_ORDER; o++) {
2789                struct free_area *area = &z->free_area[o];
2790                int mt;
2791
2792                if (!area->nr_free)
2793                        continue;
2794
2795                if (alloc_harder)
2796                        return true;
2797
2798                for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
2799                        if (!list_empty(&area->free_list[mt]))
2800                                return true;
2801                }
2802
2803#ifdef CONFIG_CMA
2804                if ((alloc_flags & ALLOC_CMA) &&
2805                    !list_empty(&area->free_list[MIGRATE_CMA])) {
2806                        return true;
2807                }
2808#endif
2809        }
2810        return false;
2811}
2812
2813bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2814                      int classzone_idx, unsigned int alloc_flags)
2815{
2816        return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2817                                        zone_page_state(z, NR_FREE_PAGES));
2818}
2819
2820static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
2821                unsigned long mark, int classzone_idx, unsigned int alloc_flags)
2822{
2823        long free_pages = zone_page_state(z, NR_FREE_PAGES);
2824        long cma_pages = 0;
2825
2826#ifdef CONFIG_CMA
2827        /* If allocation can't use CMA areas don't use free CMA pages */
2828        if (!(alloc_flags & ALLOC_CMA))
2829                cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
2830#endif
2831
2832        /*
2833         * Fast check for order-0 only. If this fails then the reserves
2834         * need to be calculated. There is a corner case where the check
2835         * passes but only the high-order atomic reserve are free. If
2836         * the caller is !atomic then it'll uselessly search the free
2837         * list. That corner case is then slower but it is harmless.
2838         */
2839        if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
2840                return true;
2841
2842        return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2843                                        free_pages);
2844}
2845
2846bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
2847                        unsigned long mark, int classzone_idx)
2848{
2849        long free_pages = zone_page_state(z, NR_FREE_PAGES);
2850
2851        if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
2852                free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
2853
2854        return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
2855                                                                free_pages);
2856}
2857
2858#ifdef CONFIG_NUMA
2859static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2860{
2861        return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
2862                                RECLAIM_DISTANCE;
2863}
2864#else   /* CONFIG_NUMA */
2865static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2866{
2867        return true;
2868}
2869#endif  /* CONFIG_NUMA */
2870
2871/*
2872 * get_page_from_freelist goes through the zonelist trying to allocate
2873 * a page.
2874 */
2875static struct page *
2876get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
2877                                                const struct alloc_context *ac)
2878{
2879        struct zoneref *z = ac->preferred_zoneref;
2880        struct zone *zone;
2881        struct pglist_data *last_pgdat_dirty_limit = NULL;
2882
2883        /*
2884         * Scan zonelist, looking for a zone with enough free.
2885         * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2886         */
2887        for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
2888                                                                ac->nodemask) {
2889                struct page *page;
2890                unsigned long mark;
2891
2892                if (cpusets_enabled() &&
2893                        (alloc_flags & ALLOC_CPUSET) &&
2894                        !__cpuset_zone_allowed(zone, gfp_mask))
2895                                continue;
2896                /*
2897                 * When allocating a page cache page for writing, we
2898                 * want to get it from a node that is within its dirty
2899                 * limit, such that no single node holds more than its
2900                 * proportional share of globally allowed dirty pages.
2901                 * The dirty limits take into account the node's
2902                 * lowmem reserves and high watermark so that kswapd
2903                 * should be able to balance it without having to
2904                 * write pages from its LRU list.
2905                 *
2906                 * XXX: For now, allow allocations to potentially
2907                 * exceed the per-node dirty limit in the slowpath
2908                 * (spread_dirty_pages unset) before going into reclaim,
2909                 * which is important when on a NUMA setup the allowed
2910                 * nodes are together not big enough to reach the
2911                 * global limit.  The proper fix for these situations
2912                 * will require awareness of nodes in the
2913                 * dirty-throttling and the flusher threads.
2914                 */
2915                if (ac->spread_dirty_pages) {
2916                        if (last_pgdat_dirty_limit == zone->zone_pgdat)
2917                                continue;
2918
2919                        if (!node_dirty_ok(zone->zone_pgdat)) {
2920                                last_pgdat_dirty_limit = zone->zone_pgdat;
2921                                continue;
2922                        }
2923                }
2924
2925                mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2926                if (!zone_watermark_fast(zone, order, mark,
2927                                       ac_classzone_idx(ac), alloc_flags)) {
2928                        int ret;
2929
2930                        /* Checked here to keep the fast path fast */
2931                        BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2932                        if (alloc_flags & ALLOC_NO_WATERMARKS)
2933                                goto try_this_zone;
2934
2935                        if (node_reclaim_mode == 0 ||
2936                            !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
2937                                continue;
2938
2939                        ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
2940                        switch (ret) {
2941                        case NODE_RECLAIM_NOSCAN:
2942                                /* did not scan */
2943                                continue;
2944                        case NODE_RECLAIM_FULL:
2945                                /* scanned but unreclaimable */
2946                                continue;
2947                        default:
2948                                /* did we reclaim enough */
2949                                if (zone_watermark_ok(zone, order, mark,
2950                                                ac_classzone_idx(ac), alloc_flags))
2951                                        goto try_this_zone;
2952
2953                                continue;
2954                        }
2955                }
2956
2957try_this_zone:
2958                page = buffered_rmqueue(ac->preferred_zoneref->zone, zone, order,
2959                                gfp_mask, alloc_flags, ac->migratetype);
2960                if (page) {
2961                        prep_new_page(page, order, gfp_mask, alloc_flags);
2962
2963                        /*
2964                         * If this is a high-order atomic allocation then check
2965                         * if the pageblock should be reserved for the future
2966                         */
2967                        if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
2968                                reserve_highatomic_pageblock(page, zone, order);
2969
2970                        return page;
2971                }
2972        }
2973
2974        return NULL;
2975}
2976
2977/*
2978 * Large machines with many possible nodes should not always dump per-node
2979 * meminfo in irq context.
2980 */
2981static inline bool should_suppress_show_mem(void)
2982{
2983        bool ret = false;
2984
2985#if NODES_SHIFT > 8
2986        ret = in_interrupt();
2987#endif
2988        return ret;
2989}
2990
2991static DEFINE_RATELIMIT_STATE(nopage_rs,
2992                DEFAULT_RATELIMIT_INTERVAL,
2993                DEFAULT_RATELIMIT_BURST);
2994
2995void warn_alloc(gfp_t gfp_mask, const char *fmt, ...)
2996{
2997        unsigned int filter = SHOW_MEM_FILTER_NODES;
2998        struct va_format vaf;
2999        va_list args;
3000
3001        if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
3002            debug_guardpage_minorder() > 0)
3003                return;
3004
3005        /*
3006         * This documents exceptions given to allocations in certain
3007         * contexts that are allowed to allocate outside current's set
3008         * of allowed nodes.
3009         */
3010        if (!(gfp_mask & __GFP_NOMEMALLOC))
3011                if (test_thread_flag(TIF_MEMDIE) ||
3012                    (current->flags & (PF_MEMALLOC | PF_EXITING)))
3013                        filter &= ~SHOW_MEM_FILTER_NODES;
3014        if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3015                filter &= ~SHOW_MEM_FILTER_NODES;
3016
3017        pr_warn("%s: ", current->comm);
3018
3019        va_start(args, fmt);
3020        vaf.fmt = fmt;
3021        vaf.va = &args;
3022        pr_cont("%pV", &vaf);
3023        va_end(args);
3024
3025        pr_cont(", mode:%#x(%pGg)\n", gfp_mask, &gfp_mask);
3026
3027        dump_stack();
3028        if (!should_suppress_show_mem())
3029                show_mem(filter);
3030}
3031
3032static inline struct page *
3033__alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3034        const struct alloc_context *ac, unsigned long *did_some_progress)
3035{
3036        struct oom_control oc = {
3037                .zonelist = ac->zonelist,
3038                .nodemask = ac->nodemask,
3039                .memcg = NULL,
3040                .gfp_mask = gfp_mask,
3041                .order = order,
3042        };
3043        struct page *page;
3044
3045        *did_some_progress = 0;
3046
3047        /*
3048         * Acquire the oom lock.  If that fails, somebody else is
3049         * making progress for us.
3050         */
3051        if (!mutex_trylock(&oom_lock)) {
3052                *did_some_progress = 1;
3053                schedule_timeout_uninterruptible(1);
3054                return NULL;
3055        }
3056
3057        /*
3058         * Go through the zonelist yet one more time, keep very high watermark
3059         * here, this is only to catch a parallel oom killing, we must fail if
3060         * we're still under heavy pressure.
3061         */
3062        page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order,
3063                                        ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3064        if (page)
3065                goto out;
3066
3067        if (!(gfp_mask & __GFP_NOFAIL)) {
3068                /* Coredumps can quickly deplete all memory reserves */
3069                if (current->flags & PF_DUMPCORE)
3070                        goto out;
3071                /* The OOM killer will not help higher order allocs */
3072                if (order > PAGE_ALLOC_COSTLY_ORDER)
3073                        goto out;
3074                /* The OOM killer does not needlessly kill tasks for lowmem */
3075                if (ac->high_zoneidx < ZONE_NORMAL)
3076                        goto out;
3077                if (pm_suspended_storage())
3078                        goto out;
3079                /*
3080                 * XXX: GFP_NOFS allocations should rather fail than rely on
3081                 * other request to make a forward progress.
3082                 * We are in an unfortunate situation where out_of_memory cannot
3083                 * do much for this context but let's try it to at least get
3084                 * access to memory reserved if the current task is killed (see
3085                 * out_of_memory). Once filesystems are ready to handle allocation
3086                 * failures more gracefully we should just bail out here.
3087                 */
3088
3089                /* The OOM killer may not free memory on a specific node */
3090                if (gfp_mask & __GFP_THISNODE)
3091                        goto out;
3092        }
3093        /* Exhausted what can be done so it's blamo time */
3094        if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3095                *did_some_progress = 1;
3096
3097                if (gfp_mask & __GFP_NOFAIL) {
3098                        page = get_page_from_freelist(gfp_mask, order,
3099                                        ALLOC_NO_WATERMARKS|ALLOC_CPUSET, ac);
3100                        /*
3101                         * fallback to ignore cpuset restriction if our nodes
3102                         * are depleted
3103                         */
3104                        if (!page)
3105                                page = get_page_from_freelist(gfp_mask, order,
3106                                        ALLOC_NO_WATERMARKS, ac);
3107                }
3108        }
3109out:
3110        mutex_unlock(&oom_lock);
3111        return page;
3112}
3113
3114/*
3115 * Maximum number of compaction retries wit a progress before OOM
3116 * killer is consider as the only way to move forward.
3117 */
3118#define MAX_COMPACT_RETRIES 16
3119
3120#ifdef CONFIG_COMPACTION
3121/* Try memory compaction for high-order allocations before reclaim */
3122static struct page *
3123__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3124                unsigned int alloc_flags, const struct alloc_context *ac,
3125                enum compact_priority prio, enum compact_result *compact_result)
3126{
3127        struct page *page;
3128
3129        if (!order)
3130                return NULL;
3131
3132        current->flags |= PF_MEMALLOC;
3133        *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3134                                                                        prio);
3135        current->flags &= ~PF_MEMALLOC;
3136
3137        if (*compact_result <= COMPACT_INACTIVE)
3138                return NULL;
3139
3140        /*
3141         * At least in one zone compaction wasn't deferred or skipped, so let's
3142         * count a compaction stall
3143         */
3144        count_vm_event(COMPACTSTALL);
3145
3146        page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3147
3148        if (page) {
3149                struct zone *zone = page_zone(page);
3150
3151                zone->compact_blockskip_flush = false;
3152                compaction_defer_reset(zone, order, true);
3153                count_vm_event(COMPACTSUCCESS);
3154                return page;
3155        }
3156
3157        /*
3158         * It's bad if compaction run occurs and fails. The most likely reason
3159         * is that pages exist, but not enough to satisfy watermarks.
3160         */
3161        count_vm_event(COMPACTFAIL);
3162
3163        cond_resched();
3164
3165        return NULL;
3166}
3167
3168static inline bool
3169should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3170                     enum compact_result compact_result,
3171                     enum compact_priority *compact_priority,
3172                     int *compaction_retries)
3173{
3174        int max_retries = MAX_COMPACT_RETRIES;
3175        int min_priority;
3176
3177        if (!order)
3178                return false;
3179
3180        if (compaction_made_progress(compact_result))
3181                (*compaction_retries)++;
3182
3183        /*
3184         * compaction considers all the zone as desperately out of memory
3185         * so it doesn't really make much sense to retry except when the
3186         * failure could be caused by insufficient priority
3187         */
3188        if (compaction_failed(compact_result))
3189                goto check_priority;
3190
3191        /*
3192         * make sure the compaction wasn't deferred or didn't bail out early
3193         * due to locks contention before we declare that we should give up.
3194         * But do not retry if the given zonelist is not suitable for
3195         * compaction.
3196         */
3197        if (compaction_withdrawn(compact_result))
3198                return compaction_zonelist_suitable(ac, order, alloc_flags);
3199
3200        /*
3201         * !costly requests are much more important than __GFP_REPEAT
3202         * costly ones because they are de facto nofail and invoke OOM
3203         * killer to move on while costly can fail and users are ready
3204         * to cope with that. 1/4 retries is rather arbitrary but we
3205         * would need much more detailed feedback from compaction to
3206         * make a better decision.
3207         */
3208        if (order > PAGE_ALLOC_COSTLY_ORDER)
3209                max_retries /= 4;
3210        if (*compaction_retries <= max_retries)
3211                return true;
3212
3213        /*
3214         * Make sure there are attempts at the highest priority if we exhausted
3215         * all retries or failed at the lower priorities.
3216         */
3217check_priority:
3218        min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
3219                        MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
3220        if (*compact_priority > min_priority) {
3221                (*compact_priority)--;
3222                *compaction_retries = 0;
3223                return true;
3224        }
3225        return false;
3226}
3227#else
3228static inline struct page *
3229__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3230                unsigned int alloc_flags, const struct alloc_context *ac,
3231                enum compact_priority prio, enum compact_result *compact_result)
3232{
3233        *compact_result = COMPACT_SKIPPED;
3234        return NULL;
3235}
3236
3237static inline bool
3238should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
3239                     enum compact_result compact_result,
3240                     enum compact_priority *compact_priority,
3241                     int *compaction_retries)
3242{
3243        struct zone *zone;
3244        struct zoneref *z;
3245
3246        if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
3247                return false;
3248
3249        /*
3250         * There are setups with compaction disabled which would prefer to loop
3251         * inside the allocator rather than hit the oom killer prematurely.
3252         * Let's give them a good hope and keep retrying while the order-0
3253         * watermarks are OK.
3254         */
3255        for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3256                                        ac->nodemask) {
3257                if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
3258                                        ac_classzone_idx(ac), alloc_flags))
3259                        return true;
3260        }
3261        return false;
3262}
3263#endif /* CONFIG_COMPACTION */
3264
3265/* Perform direct synchronous page reclaim */
3266static int
3267__perform_reclaim(gfp_t gfp_mask, unsigned int order,
3268                                        const struct alloc_context *ac)
3269{
3270        struct reclaim_state reclaim_state;
3271        int progress;
3272
3273        cond_resched();
3274
3275        /* We now go into synchronous reclaim */
3276        cpuset_memory_pressure_bump();
3277        current->flags |= PF_MEMALLOC;
3278        lockdep_set_current_reclaim_state(gfp_mask);
3279        reclaim_state.reclaimed_slab = 0;
3280        current->reclaim_state = &reclaim_state;
3281
3282        progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
3283                                                                ac->nodemask);
3284
3285        current->reclaim_state = NULL;
3286        lockdep_clear_current_reclaim_state();
3287        current->flags &= ~PF_MEMALLOC;
3288
3289        cond_resched();
3290
3291        return progress;
3292}
3293
3294/* The really slow allocator path where we enter direct reclaim */
3295static inline struct page *
3296__alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
3297                unsigned int alloc_flags, const struct alloc_context *ac,
3298                unsigned long *did_some_progress)
3299{
3300        struct page *page = NULL;
3301        bool drained = false;
3302
3303        *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
3304        if (unlikely(!(*did_some_progress)))
3305                return NULL;
3306
3307retry:
3308        page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3309
3310        /*
3311         * If an allocation failed after direct reclaim, it could be because
3312         * pages are pinned on the per-cpu lists or in high alloc reserves.
3313         * Shrink them them and try again
3314         */
3315        if (!page && !drained) {
3316                unreserve_highatomic_pageblock(ac);
3317                drain_all_pages(NULL);
3318                drained = true;
3319                goto retry;
3320        }
3321
3322        return page;
3323}
3324
3325static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
3326{
3327        struct zoneref *z;
3328        struct zone *zone;
3329        pg_data_t *last_pgdat = NULL;
3330
3331        for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
3332                                        ac->high_zoneidx, ac->nodemask) {
3333                if (last_pgdat != zone->zone_pgdat)
3334                        wakeup_kswapd(zone, order, ac->high_zoneidx);
3335                last_pgdat = zone->zone_pgdat;
3336        }
3337}
3338
3339static inline unsigned int
3340gfp_to_alloc_flags(gfp_t gfp_mask)
3341{
3342        unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
3343
3344        /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3345        BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
3346
3347        /*
3348         * The caller may dip into page reserves a bit more if the caller
3349         * cannot run direct reclaim, or if the caller has realtime scheduling
3350         * policy or is asking for __GFP_HIGH memory.  GFP_ATOMIC requests will
3351         * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3352         */
3353        alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
3354
3355        if (gfp_mask & __GFP_ATOMIC) {
3356                /*
3357                 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3358                 * if it can't schedule.
3359                 */
3360                if (!(gfp_mask & __GFP_NOMEMALLOC))
3361                        alloc_flags |= ALLOC_HARDER;
3362                /*
3363                 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3364                 * comment for __cpuset_node_allowed().
3365                 */
3366                alloc_flags &= ~ALLOC_CPUSET;
3367        } else if (unlikely(rt_task(current)) && !in_interrupt())
3368                alloc_flags |= ALLOC_HARDER;
3369
3370#ifdef CONFIG_CMA
3371        if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3372                alloc_flags |= ALLOC_CMA;
3373#endif
3374        return alloc_flags;
3375}
3376
3377bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
3378{
3379        if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
3380                return false;
3381
3382        if (gfp_mask & __GFP_MEMALLOC)
3383                return true;
3384        if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
3385                return true;
3386        if (!in_interrupt() &&
3387                        ((current->flags & PF_MEMALLOC) ||
3388                         unlikely(test_thread_flag(TIF_MEMDIE))))
3389                return true;
3390
3391        return false;
3392}
3393
3394/*
3395 * Maximum number of reclaim retries without any progress before OOM killer
3396 * is consider as the only way to move forward.
3397 */
3398#define MAX_RECLAIM_RETRIES 16
3399
3400/*
3401 * Checks whether it makes sense to retry the reclaim to make a forward progress
3402 * for the given allocation request.
3403 * The reclaim feedback represented by did_some_progress (any progress during
3404 * the last reclaim round) and no_progress_loops (number of reclaim rounds without
3405 * any progress in a row) is considered as well as the reclaimable pages on the
3406 * applicable zone list (with a backoff mechanism which is a function of
3407 * no_progress_loops).
3408 *
3409 * Returns true if a retry is viable or false to enter the oom path.
3410 */
3411static inline bool
3412should_reclaim_retry(gfp_t gfp_mask, unsigned order,
3413                     struct alloc_context *ac, int alloc_flags,
3414                     bool did_some_progress, int *no_progress_loops)
3415{
3416        struct zone *zone;
3417        struct zoneref *z;
3418
3419        /*
3420         * Costly allocations might have made a progress but this doesn't mean
3421         * their order will become available due to high fragmentation so
3422         * always increment the no progress counter for them
3423         */
3424        if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
3425                *no_progress_loops = 0;
3426        else
3427                (*no_progress_loops)++;
3428
3429        /*
3430         * Make sure we converge to OOM if we cannot make any progress
3431         * several times in the row.
3432         */
3433        if (*no_progress_loops > MAX_RECLAIM_RETRIES)
3434                return false;
3435
3436        /*
3437         * Keep reclaiming pages while there is a chance this will lead
3438         * somewhere.  If none of the target zones can satisfy our allocation
3439         * request even if all reclaimable pages are considered then we are
3440         * screwed and have to go OOM.
3441         */
3442        for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3443                                        ac->nodemask) {
3444                unsigned long available;
3445                unsigned long reclaimable;
3446
3447                available = reclaimable = zone_reclaimable_pages(zone);
3448                available -= DIV_ROUND_UP((*no_progress_loops) * available,
3449                                          MAX_RECLAIM_RETRIES);
3450                available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
3451
3452                /*
3453                 * Would the allocation succeed if we reclaimed the whole
3454                 * available?
3455                 */
3456                if (__zone_watermark_ok(zone, order, min_wmark_pages(zone),
3457                                ac_classzone_idx(ac), alloc_flags, available)) {
3458                        /*
3459                         * If we didn't make any progress and have a lot of
3460                         * dirty + writeback pages then we should wait for
3461                         * an IO to complete to slow down the reclaim and
3462                         * prevent from pre mature OOM
3463                         */
3464                        if (!did_some_progress) {
3465                                unsigned long write_pending;
3466
3467                                write_pending = zone_page_state_snapshot(zone,
3468                                                        NR_ZONE_WRITE_PENDING);
3469
3470                                if (2 * write_pending > reclaimable) {
3471                                        congestion_wait(BLK_RW_ASYNC, HZ/10);
3472                                        return true;
3473                                }
3474                        }
3475
3476                        /*
3477                         * Memory allocation/reclaim might be called from a WQ
3478                         * context and the current implementation of the WQ
3479                         * concurrency control doesn't recognize that
3480                         * a particular WQ is congested if the worker thread is
3481                         * looping without ever sleeping. Therefore we have to
3482                         * do a short sleep here rather than calling
3483                         * cond_resched().
3484                         */
3485                        if (current->flags & PF_WQ_WORKER)
3486                                schedule_timeout_uninterruptible(1);
3487                        else
3488                                cond_resched();
3489
3490                        return true;
3491                }
3492        }
3493
3494        return false;
3495}
3496
3497static inline struct page *
3498__alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
3499                                                struct alloc_context *ac)
3500{
3501        bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
3502        struct page *page = NULL;
3503        unsigned int alloc_flags;
3504        unsigned long did_some_progress;
3505        enum compact_priority compact_priority;
3506        enum compact_result compact_result;
3507        int compaction_retries;
3508        int no_progress_loops;
3509        unsigned long alloc_start = jiffies;
3510        unsigned int stall_timeout = 10 * HZ;
3511        unsigned int cpuset_mems_cookie;
3512
3513        /*
3514         * In the slowpath, we sanity check order to avoid ever trying to
3515         * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3516         * be using allocators in order of preference for an area that is
3517         * too large.
3518         */
3519        if (order >= MAX_ORDER) {
3520                WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
3521                return NULL;
3522        }
3523
3524        /*
3525         * We also sanity check to catch abuse of atomic reserves being used by
3526         * callers that are not in atomic context.
3527         */
3528        if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
3529                                (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
3530                gfp_mask &= ~__GFP_ATOMIC;
3531
3532retry_cpuset:
3533        compaction_retries = 0;
3534        no_progress_loops = 0;
3535        compact_priority = DEF_COMPACT_PRIORITY;
3536        cpuset_mems_cookie = read_mems_allowed_begin();
3537        /*
3538         * We need to recalculate the starting point for the zonelist iterator
3539         * because we might have used different nodemask in the fast path, or
3540         * there was a cpuset modification and we are retrying - otherwise we
3541         * could end up iterating over non-eligible zones endlessly.
3542         */
3543        ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
3544                                        ac->high_zoneidx, ac->nodemask);
3545        if (!ac->preferred_zoneref->zone)
3546                goto nopage;
3547
3548
3549        /*
3550         * The fast path uses conservative alloc_flags to succeed only until
3551         * kswapd needs to be woken up, and to avoid the cost of setting up
3552         * alloc_flags precisely. So we do that now.
3553         */
3554        alloc_flags = gfp_to_alloc_flags(gfp_mask);
3555
3556        if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3557                wake_all_kswapds(order, ac);
3558
3559        /*
3560         * The adjusted alloc_flags might result in immediate success, so try
3561         * that first
3562         */
3563        page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3564        if (page)
3565                goto got_pg;
3566
3567        /*
3568         * For costly allocations, try direct compaction first, as it's likely
3569         * that we have enough base pages and don't need to reclaim. Don't try
3570         * that for allocations that are allowed to ignore watermarks, as the
3571         * ALLOC_NO_WATERMARKS attempt didn't yet happen.
3572         */
3573        if (can_direct_reclaim && order > PAGE_ALLOC_COSTLY_ORDER &&
3574                !gfp_pfmemalloc_allowed(gfp_mask)) {
3575                page = __alloc_pages_direct_compact(gfp_mask, order,
3576                                                alloc_flags, ac,
3577                                                INIT_COMPACT_PRIORITY,
3578                                                &compact_result);
3579                if (page)
3580                        goto got_pg;
3581
3582                /*
3583                 * Checks for costly allocations with __GFP_NORETRY, which
3584                 * includes THP page fault allocations
3585                 */
3586                if (gfp_mask & __GFP_NORETRY) {
3587                        /*
3588                         * If compaction is deferred for high-order allocations,
3589                         * it is because sync compaction recently failed. If
3590                         * this is the case and the caller requested a THP
3591                         * allocation, we do not want to heavily disrupt the
3592                         * system, so we fail the allocation instead of entering
3593                         * direct reclaim.
3594                         */
3595                        if (compact_result == COMPACT_DEFERRED)
3596                                goto nopage;
3597
3598                        /*
3599                         * Looks like reclaim/compaction is worth trying, but
3600                         * sync compaction could be very expensive, so keep
3601                         * using async compaction.
3602                         */
3603                        compact_priority = INIT_COMPACT_PRIORITY;
3604                }
3605        }
3606
3607retry:
3608        /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
3609        if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3610                wake_all_kswapds(order, ac);
3611
3612        if (gfp_pfmemalloc_allowed(gfp_mask))
3613                alloc_flags = ALLOC_NO_WATERMARKS;
3614
3615        /*
3616         * Reset the zonelist iterators if memory policies can be ignored.
3617         * These allocations are high priority and system rather than user
3618         * orientated.
3619         */
3620        if (!(alloc_flags & ALLOC_CPUSET) || (alloc_flags & ALLOC_NO_WATERMARKS)) {
3621                ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
3622                ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
3623                                        ac->high_zoneidx, ac->nodemask);
3624        }
3625
3626        /* Attempt with potentially adjusted zonelist and alloc_flags */
3627        page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3628        if (page)
3629                goto got_pg;
3630
3631        /* Caller is not willing to reclaim, we can't balance anything */
3632        if (!can_direct_reclaim) {
3633                /*
3634                 * All existing users of the __GFP_NOFAIL are blockable, so warn
3635                 * of any new users that actually allow this type of allocation
3636                 * to fail.
3637                 */
3638                WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
3639                goto nopage;
3640        }
3641
3642        /* Avoid recursion of direct reclaim */
3643        if (current->flags & PF_MEMALLOC) {
3644                /*
3645                 * __GFP_NOFAIL request from this context is rather bizarre
3646                 * because we cannot reclaim anything and only can loop waiting
3647                 * for somebody to do a work for us.
3648                 */
3649                if (WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3650                        cond_resched();
3651                        goto retry;
3652                }
3653                goto nopage;
3654        }
3655
3656        /* Avoid allocations with no watermarks from looping endlessly */
3657        if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
3658                goto nopage;
3659
3660
3661        /* Try direct reclaim and then allocating */
3662        page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
3663                                                        &did_some_progress);
3664        if (page)
3665                goto got_pg;
3666
3667        /* Try direct compaction and then allocating */
3668        page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
3669                                        compact_priority, &compact_result);
3670        if (page)
3671                goto got_pg;
3672
3673        /* Do not loop if specifically requested */
3674        if (gfp_mask & __GFP_NORETRY)
3675                goto nopage;
3676
3677        /*
3678         * Do not retry costly high order allocations unless they are
3679         * __GFP_REPEAT
3680         */
3681        if (order > PAGE_ALLOC_COSTLY_ORDER && !(gfp_mask & __GFP_REPEAT))
3682                goto nopage;
3683
3684        /* Make sure we know about allocations which stall for too long */
3685        if (time_after(jiffies, alloc_start + stall_timeout)) {
3686                warn_alloc(gfp_mask,
3687                        "page allocation stalls for %ums, order:%u",
3688                        jiffies_to_msecs(jiffies-alloc_start), order);
3689                stall_timeout += 10 * HZ;
3690        }
3691
3692        if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
3693                                 did_some_progress > 0, &no_progress_loops))
3694                goto retry;
3695
3696        /*
3697         * It doesn't make any sense to retry for the compaction if the order-0
3698         * reclaim is not able to make any progress because the current
3699         * implementation of the compaction depends on the sufficient amount
3700         * of free memory (see __compaction_suitable)
3701         */
3702        if (did_some_progress > 0 &&
3703                        should_compact_retry(ac, order, alloc_flags,
3704                                compact_result, &compact_priority,
3705                                &compaction_retries))
3706                goto retry;
3707
3708        /*
3709         * It's possible we raced with cpuset update so the OOM would be
3710         * premature (see below the nopage: label for full explanation).
3711         */
3712        if (read_mems_allowed_retry(cpuset_mems_cookie))
3713                goto retry_cpuset;
3714
3715        /* Reclaim has failed us, start killing things */
3716        page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
3717        if (page)
3718                goto got_pg;
3719
3720        /* Retry as long as the OOM killer is making progress */
3721        if (did_some_progress) {
3722                no_progress_loops = 0;
3723                goto retry;
3724        }
3725
3726nopage:
3727        /*
3728         * When updating a task's mems_allowed or mempolicy nodemask, it is
3729         * possible to race with parallel threads in such a way that our
3730         * allocation can fail while the mask is being updated. If we are about
3731         * to fail, check if the cpuset changed during allocation and if so,
3732         * retry.
3733         */
3734        if (read_mems_allowed_retry(cpuset_mems_cookie))
3735                goto retry_cpuset;
3736
3737        warn_alloc(gfp_mask,
3738                        "page allocation failure: order:%u", order);
3739got_pg:
3740        return page;
3741}
3742
3743/*
3744 * This is the 'heart' of the zoned buddy allocator.
3745 */
3746struct page *
3747__alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
3748                        struct zonelist *zonelist, nodemask_t *nodemask)
3749{
3750        struct page *page;
3751        unsigned int alloc_flags = ALLOC_WMARK_LOW;
3752        gfp_t alloc_mask = gfp_mask; /* The gfp_t that was actually used for allocation */
3753        struct alloc_context ac = {
3754                .high_zoneidx = gfp_zone(gfp_mask),
3755                .zonelist = zonelist,
3756                .nodemask = nodemask,
3757                .migratetype = gfpflags_to_migratetype(gfp_mask),
3758        };
3759
3760        if (cpusets_enabled()) {
3761                alloc_mask |= __GFP_HARDWALL;
3762                alloc_flags |= ALLOC_CPUSET;
3763                if (!ac.nodemask)
3764                        ac.nodemask = &cpuset_current_mems_allowed;
3765        }
3766
3767        gfp_mask &= gfp_allowed_mask;
3768
3769        lockdep_trace_alloc(gfp_mask);
3770
3771        might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
3772
3773        if (should_fail_alloc_page(gfp_mask, order))
3774                return NULL;
3775
3776        /*
3777         * Check the zones suitable for the gfp_mask contain at least one
3778         * valid zone. It's possible to have an empty zonelist as a result
3779         * of __GFP_THISNODE and a memoryless node
3780         */
3781        if (unlikely(!zonelist->_zonerefs->zone))
3782                return NULL;
3783
3784        if (IS_ENABLED(CONFIG_CMA) && ac.migratetype == MIGRATE_MOVABLE)
3785                alloc_flags |= ALLOC_CMA;
3786
3787        /* Dirty zone balancing only done in the fast path */
3788        ac.spread_dirty_pages = (gfp_mask & __GFP_WRITE);
3789
3790        /*
3791         * The preferred zone is used for statistics but crucially it is
3792         * also used as the starting point for the zonelist iterator. It
3793         * may get reset for allocations that ignore memory policies.
3794         */
3795        ac.preferred_zoneref = first_zones_zonelist(ac.zonelist,
3796                                        ac.high_zoneidx, ac.nodemask);
3797        if (!ac.preferred_zoneref->zone) {
3798                page = NULL;
3799                /*
3800                 * This might be due to race with cpuset_current_mems_allowed
3801                 * update, so make sure we retry with original nodemask in the
3802                 * slow path.
3803                 */
3804                goto no_zone;
3805        }
3806
3807        /* First allocation attempt */
3808        page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
3809        if (likely(page))
3810                goto out;
3811
3812no_zone:
3813        /*
3814         * Runtime PM, block IO and its error handling path can deadlock
3815         * because I/O on the device might not complete.
3816         */
3817        alloc_mask = memalloc_noio_flags(gfp_mask);
3818        ac.spread_dirty_pages = false;
3819
3820        /*
3821         * Restore the original nodemask if it was potentially replaced with
3822         * &cpuset_current_mems_allowed to optimize the fast-path attempt.
3823         */
3824        if (unlikely(ac.nodemask != nodemask))
3825                ac.nodemask = nodemask;
3826
3827        page = __alloc_pages_slowpath(alloc_mask, order, &ac);
3828
3829out:
3830        if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
3831            unlikely(memcg_kmem_charge(page, gfp_mask, order) != 0)) {
3832                __free_pages(page, order);
3833                page = NULL;
3834        }
3835
3836        if (kmemcheck_enabled && page)
3837                kmemcheck_pagealloc_alloc(page, order, gfp_mask);
3838
3839        trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
3840
3841        return page;
3842}
3843EXPORT_SYMBOL(__alloc_pages_nodemask);
3844
3845/*
3846 * Common helper functions.
3847 */
3848unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
3849{
3850        struct page *page;
3851
3852        /*
3853         * __get_free_pages() returns a 32-bit address, which cannot represent
3854         * a highmem page
3855         */
3856        VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
3857
3858        page = alloc_pages(gfp_mask, order);
3859        if (!page)
3860                return 0;
3861        return (unsigned long) page_address(page);
3862}
3863EXPORT_SYMBOL(__get_free_pages);
3864
3865unsigned long get_zeroed_page(gfp_t gfp_mask)
3866{
3867        return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
3868}
3869EXPORT_SYMBOL(get_zeroed_page);
3870
3871void __free_pages(struct page *page, unsigned int order)
3872{
3873        if (put_page_testzero(page)) {
3874                if (order == 0)
3875                        free_hot_cold_page(page, false);
3876                else
3877                        __free_pages_ok(page, order);
3878        }
3879}
3880
3881EXPORT_SYMBOL(__free_pages);
3882
3883void free_pages(unsigned long addr, unsigned int order)
3884{
3885        if (addr != 0) {
3886                VM_BUG_ON(!virt_addr_valid((void *)addr));
3887                __free_pages(virt_to_page((void *)addr), order);
3888        }
3889}
3890
3891EXPORT_SYMBOL(free_pages);
3892
3893/*
3894 * Page Fragment:
3895 *  An arbitrary-length arbitrary-offset area of memory which resides
3896 *  within a 0 or higher order page.  Multiple fragments within that page
3897 *  are individually refcounted, in the page's reference counter.
3898 *
3899 * The page_frag functions below provide a simple allocation framework for
3900 * page fragments.  This is used by the network stack and network device
3901 * drivers to provide a backing region of memory for use as either an
3902 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3903 */
3904static struct page *__page_frag_refill(struct page_frag_cache *nc,
3905                                       gfp_t gfp_mask)
3906{
3907        struct page *page = NULL;
3908        gfp_t gfp = gfp_mask;
3909
3910#if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3911        gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
3912                    __GFP_NOMEMALLOC;
3913        page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
3914                                PAGE_FRAG_CACHE_MAX_ORDER);
3915        nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
3916#endif
3917        if (unlikely(!page))
3918                page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
3919
3920        nc->va = page ? page_address(page) : NULL;
3921
3922        return page;
3923}
3924
3925void *__alloc_page_frag(struct page_frag_cache *nc,
3926                        unsigned int fragsz, gfp_t gfp_mask)
3927{
3928        unsigned int size = PAGE_SIZE;
3929        struct page *page;
3930        int offset;
3931
3932        if (unlikely(!nc->va)) {
3933refill:
3934                page = __page_frag_refill(nc, gfp_mask);
3935                if (!page)
3936                        return NULL;
3937
3938#if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3939                /* if size can vary use size else just use PAGE_SIZE */
3940                size = nc->size;
3941#endif
3942                /* Even if we own the page, we do not use atomic_set().
3943                 * This would break get_page_unless_zero() users.
3944                 */
3945                page_ref_add(page, size - 1);
3946
3947                /* reset page count bias and offset to start of new frag */
3948                nc->pfmemalloc = page_is_pfmemalloc(page);
3949                nc->pagecnt_bias = size;
3950                nc->offset = size;
3951        }
3952
3953        offset = nc->offset - fragsz;
3954        if (unlikely(offset < 0)) {
3955                page = virt_to_page(nc->va);
3956
3957                if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
3958                        goto refill;
3959
3960#if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3961                /* if size can vary use size else just use PAGE_SIZE */
3962                size = nc->size;
3963#endif
3964                /* OK, page count is 0, we can safely set it */
3965                set_page_count(page, size);
3966
3967                /* reset page count bias and offset to start of new frag */
3968                nc->pagecnt_bias = size;
3969                offset = size - fragsz;
3970        }
3971
3972        nc->pagecnt_bias--;
3973        nc->offset = offset;
3974
3975        return nc->va + offset;
3976}
3977EXPORT_SYMBOL(__alloc_page_frag);
3978
3979/*
3980 * Frees a page fragment allocated out of either a compound or order 0 page.
3981 */
3982void __free_page_frag(void *addr)
3983{
3984        struct page *page = virt_to_head_page(addr);
3985
3986        if (unlikely(put_page_testzero(page)))
3987                __free_pages_ok(page, compound_order(page));
3988}
3989EXPORT_SYMBOL(__free_page_frag);
3990
3991static void *make_alloc_exact(unsigned long addr, unsigned int order,
3992                size_t size)
3993{
3994        if (addr) {
3995                unsigned long alloc_end = addr + (PAGE_SIZE << order);
3996                unsigned long used = addr + PAGE_ALIGN(size);
3997
3998                split_page(virt_to_page((void *)addr), order);
3999                while (used < alloc_end) {
4000                        free_page(used);
4001                        used += PAGE_SIZE;
4002                }
4003        }
4004        return (void *)addr;
4005}
4006
4007/**
4008 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4009 * @size: the number of bytes to allocate
4010 * @gfp_mask: GFP flags for the allocation
4011 *
4012 * This function is similar to alloc_pages(), except that it allocates the
4013 * minimum number of pages to satisfy the request.  alloc_pages() can only
4014 * allocate memory in power-of-two pages.
4015 *
4016 * This function is also limited by MAX_ORDER.
4017 *
4018 * Memory allocated by this function must be released by free_pages_exact().
4019 */
4020void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
4021{
4022        unsigned int order = get_order(size);
4023        unsigned long addr;
4024
4025        addr = __get_free_pages(gfp_mask, order);
4026        return make_alloc_exact(addr, order, size);
4027}
4028EXPORT_SYMBOL(alloc_pages_exact);
4029
4030/**
4031 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4032 *                         pages on a node.
4033 * @nid: the preferred node ID where memory should be allocated
4034 * @size: the number of bytes to allocate
4035 * @gfp_mask: GFP flags for the allocation
4036 *
4037 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4038 * back.
4039 */
4040void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
4041{
4042        unsigned int order = get_order(size);
4043        struct page *p = alloc_pages_node(nid, gfp_mask, order);
4044        if (!p)
4045                return NULL;
4046        return make_alloc_exact((unsigned long)page_address(p), order, size);
4047}
4048
4049/**
4050 * free_pages_exact - release memory allocated via alloc_pages_exact()
4051 * @virt: the value returned by alloc_pages_exact.
4052 * @size: size of allocation, same value as passed to alloc_pages_exact().
4053 *
4054 * Release the memory allocated by a previous call to alloc_pages_exact.
4055 */
4056void free_pages_exact(void *virt, size_t size)
4057{
4058        unsigned long addr = (unsigned long)virt;
4059        unsigned long end = addr + PAGE_ALIGN(size);
4060
4061        while (addr < end) {
4062                free_page(addr);
4063                addr += PAGE_SIZE;
4064        }
4065}
4066EXPORT_SYMBOL(free_pages_exact);
4067
4068/**
4069 * nr_free_zone_pages - count number of pages beyond high watermark
4070 * @offset: The zone index of the highest zone
4071 *
4072 * nr_free_zone_pages() counts the number of counts pages which are beyond the
4073 * high watermark within all zones at or below a given zone index.  For each
4074 * zone, the number of pages is calculated as:
4075 *     managed_pages - high_pages
4076 */
4077static unsigned long nr_free_zone_pages(int offset)
4078{
4079        struct zoneref *z;
4080        struct zone *zone;
4081
4082        /* Just pick one node, since fallback list is circular */
4083        unsigned long sum = 0;
4084
4085        struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
4086
4087        for_each_zone_zonelist(zone, z, zonelist, offset) {
4088                unsigned long size = zone->managed_pages;
4089                unsigned long high = high_wmark_pages(zone);
4090                if (size > high)
4091                        sum += size - high;
4092        }
4093
4094        return sum;
4095}
4096
4097/**
4098 * nr_free_buffer_pages - count number of pages beyond high watermark
4099 *
4100 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4101 * watermark within ZONE_DMA and ZONE_NORMAL.
4102 */
4103unsigned long nr_free_buffer_pages(void)
4104{
4105        return nr_free_zone_pages(gfp_zone(GFP_USER));
4106}
4107EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
4108
4109/**
4110 * nr_free_pagecache_pages - count number of pages beyond high watermark
4111 *
4112 * nr_free_pagecache_pages() counts the number of pages which are beyond the
4113 * high watermark within all zones.
4114 */
4115unsigned long nr_free_pagecache_pages(void)
4116{
4117        return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
4118}
4119
4120static inline void show_node(struct zone *zone)
4121{
4122        if (IS_ENABLED(CONFIG_NUMA))
4123                printk("Node %d ", zone_to_nid(zone));
4124}
4125
4126long si_mem_available(void)
4127{
4128        long available;
4129        unsigned long pagecache;
4130        unsigned long wmark_low = 0;
4131        unsigned long pages[NR_LRU_LISTS];
4132        struct zone *zone;
4133        int lru;
4134
4135        for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
4136                pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
4137
4138        for_each_zone(zone)
4139                wmark_low += zone->watermark[WMARK_LOW];
4140
4141        /*
4142         * Estimate the amount of memory available for userspace allocations,
4143         * without causing swapping.
4144         */
4145        available = global_page_state(NR_FREE_PAGES) - totalreserve_pages;
4146
4147        /*
4148         * Not all the page cache can be freed, otherwise the system will
4149         * start swapping. Assume at least half of the page cache, or the
4150         * low watermark worth of cache, needs to stay.
4151         */
4152        pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
4153        pagecache -= min(pagecache / 2, wmark_low);
4154        available += pagecache;
4155
4156        /*
4157         * Part of the reclaimable slab consists of items that are in use,
4158         * and cannot be freed. Cap this estimate at the low watermark.
4159         */
4160        available += global_page_state(NR_SLAB_RECLAIMABLE) -
4161                     min(global_page_state(NR_SLAB_RECLAIMABLE) / 2, wmark_low);
4162
4163        if (available < 0)
4164                available = 0;
4165        return available;
4166}
4167EXPORT_SYMBOL_GPL(si_mem_available);
4168
4169void si_meminfo(struct sysinfo *val)
4170{
4171        val->totalram = totalram_pages;
4172        val->sharedram = global_node_page_state(NR_SHMEM);
4173        val->freeram = global_page_state(NR_FREE_PAGES);
4174        val->bufferram = nr_blockdev_pages();
4175        val->totalhigh = totalhigh_pages;
4176        val->freehigh = nr_free_highpages();
4177        val->mem_unit = PAGE_SIZE;
4178}
4179
4180EXPORT_SYMBOL(si_meminfo);
4181
4182#ifdef CONFIG_NUMA
4183void si_meminfo_node(struct sysinfo *val, int nid)
4184{
4185        int zone_type;          /* needs to be signed */
4186        unsigned long managed_pages = 0;
4187        unsigned long managed_highpages = 0;
4188        unsigned long free_highpages = 0;
4189        pg_data_t *pgdat = NODE_DATA(nid);
4190
4191        for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
4192                managed_pages += pgdat->node_zones[zone_type].managed_pages;
4193        val->totalram = managed_pages;
4194        val->sharedram = node_page_state(pgdat, NR_SHMEM);
4195        val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
4196#ifdef CONFIG_HIGHMEM
4197        for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
4198                struct zone *zone = &pgdat->node_zones[zone_type];
4199
4200                if (is_highmem(zone)) {
4201                        managed_highpages += zone->managed_pages;
4202                        free_highpages += zone_page_state(zone, NR_FREE_PAGES);
4203                }
4204        }
4205        val->totalhigh = managed_highpages;
4206        val->freehigh = free_highpages;
4207#else
4208        val->totalhigh = managed_highpages;
4209        val->freehigh = free_highpages;
4210#endif
4211        val->mem_unit = PAGE_SIZE;
4212}
4213#endif
4214
4215/*
4216 * Determine whether the node should be displayed or not, depending on whether
4217 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
4218 */
4219bool skip_free_areas_node(unsigned int flags, int nid)
4220{
4221        bool ret = false;
4222        unsigned int cpuset_mems_cookie;
4223
4224        if (!(flags & SHOW_MEM_FILTER_NODES))
4225                goto out;
4226
4227        do {
4228                cpuset_mems_cookie = read_mems_allowed_begin();
4229                ret = !node_isset(nid, cpuset_current_mems_allowed);
4230        } while (read_mems_allowed_retry(cpuset_mems_cookie));
4231out:
4232        return ret;
4233}
4234
4235#define K(x) ((x) << (PAGE_SHIFT-10))
4236
4237static void show_migration_types(unsigned char type)
4238{
4239        static const char types[MIGRATE_TYPES] = {
4240                [MIGRATE_UNMOVABLE]     = 'U',
4241                [MIGRATE_MOVABLE]       = 'M',
4242                [MIGRATE_RECLAIMABLE]   = 'E',
4243                [MIGRATE_HIGHATOMIC]    = 'H',
4244#ifdef CONFIG_CMA
4245                [MIGRATE_CMA]           = 'C',
4246#endif
4247#ifdef CONFIG_MEMORY_ISOLATION
4248                [MIGRATE_ISOLATE]       = 'I',
4249#endif
4250        };
4251        char tmp[MIGRATE_TYPES + 1];
4252        char *p = tmp;
4253        int i;
4254
4255        for (i = 0; i < MIGRATE_TYPES; i++) {
4256                if (type & (1 << i))
4257                        *p++ = types[i];
4258        }
4259
4260        *p = '\0';
4261        printk(KERN_CONT "(%s) ", tmp);
4262}
4263
4264/*
4265 * Show free area list (used inside shift_scroll-lock stuff)
4266 * We also calculate the percentage fragmentation. We do this by counting the
4267 * memory on each free list with the exception of the first item on the list.
4268 *
4269 * Bits in @filter:
4270 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
4271 *   cpuset.
4272 */
4273void show_free_areas(unsigned int filter)
4274{
4275        unsigned long free_pcp = 0;
4276        int cpu;
4277        struct zone *zone;
4278        pg_data_t *pgdat;
4279
4280        for_each_populated_zone(zone) {
4281                if (skip_free_areas_node(filter, zone_to_nid(zone)))
4282                        continue;
4283
4284                for_each_online_cpu(cpu)
4285                        free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4286        }
4287
4288        printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
4289                " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
4290                " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
4291                " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
4292                " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
4293                " free:%lu free_pcp:%lu free_cma:%lu\n",
4294                global_node_page_state(NR_ACTIVE_ANON),
4295                global_node_page_state(NR_INACTIVE_ANON),
4296                global_node_page_state(NR_ISOLATED_ANON),
4297                global_node_page_state(NR_ACTIVE_FILE),
4298                global_node_page_state(NR_INACTIVE_FILE),
4299                global_node_page_state(NR_ISOLATED_FILE),
4300                global_node_page_state(NR_UNEVICTABLE),
4301                global_node_page_state(NR_FILE_DIRTY),
4302                global_node_page_state(NR_WRITEBACK),
4303                global_node_page_state(NR_UNSTABLE_NFS),
4304                global_page_state(NR_SLAB_RECLAIMABLE),
4305                global_page_state(NR_SLAB_UNRECLAIMABLE),
4306                global_node_page_state(NR_FILE_MAPPED),
4307                global_node_page_state(NR_SHMEM),
4308                global_page_state(NR_PAGETABLE),
4309                global_page_state(NR_BOUNCE),
4310                global_page_state(NR_FREE_PAGES),
4311                free_pcp,
4312                global_page_state(NR_FREE_CMA_PAGES));
4313
4314        for_each_online_pgdat(pgdat) {
4315                printk("Node %d"
4316                        " active_anon:%lukB"
4317                        " inactive_anon:%lukB"
4318                        " active_file:%lukB"
4319                        " inactive_file:%lukB"
4320                        " unevictable:%lukB"
4321                        " isolated(anon):%lukB"
4322                        " isolated(file):%lukB"
4323                        " mapped:%lukB"
4324                        " dirty:%lukB"
4325                        " writeback:%lukB"
4326                        " shmem:%lukB"
4327#ifdef CONFIG_TRANSPARENT_HUGEPAGE
4328                        " shmem_thp: %lukB"
4329                        " shmem_pmdmapped: %lukB"
4330                        " anon_thp: %lukB"
4331#endif
4332                        " writeback_tmp:%lukB"
4333                        " unstable:%lukB"
4334                        " pages_scanned:%lu"
4335                        " all_unreclaimable? %s"
4336                        "\n",
4337                        pgdat->node_id,
4338                        K(node_page_state(pgdat, NR_ACTIVE_ANON)),
4339                        K(node_page_state(pgdat, NR_INACTIVE_ANON)),
4340                        K(node_page_state(pgdat, NR_ACTIVE_FILE)),
4341                        K(node_page_state(pgdat, NR_INACTIVE_FILE)),
4342                        K(node_page_state(pgdat, NR_UNEVICTABLE)),
4343                        K(node_page_state(pgdat, NR_ISOLATED_ANON)),
4344                        K(node_page_state(pgdat, NR_ISOLATED_FILE)),
4345                        K(node_page_state(pgdat, NR_FILE_MAPPED)),
4346                        K(node_page_state(pgdat, NR_FILE_DIRTY)),
4347                        K(node_page_state(pgdat, NR_WRITEBACK)),
4348                        K(node_page_state(pgdat, NR_SHMEM)),
4349#ifdef CONFIG_TRANSPARENT_HUGEPAGE
4350                        K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
4351                        K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
4352                                        * HPAGE_PMD_NR),
4353                        K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
4354#endif
4355                        K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
4356                        K(node_page_state(pgdat, NR_UNSTABLE_NFS)),
4357                        node_page_state(pgdat, NR_PAGES_SCANNED),
4358                        !pgdat_reclaimable(pgdat) ? "yes" : "no");
4359        }
4360
4361        for_each_populated_zone(zone) {
4362                int i;
4363
4364                if (skip_free_areas_node(filter, zone_to_nid(zone)))
4365                        continue;
4366
4367                free_pcp = 0;
4368                for_each_online_cpu(cpu)
4369                        free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4370
4371                show_node(zone);
4372                printk(KERN_CONT
4373                        "%s"
4374                        " free:%lukB"
4375                        " min:%lukB"
4376                        " low:%lukB"
4377                        " high:%lukB"
4378                        " active_anon:%lukB"
4379                        " inactive_anon:%lukB"
4380                        " active_file:%lukB"
4381                        " inactive_file:%lukB"
4382                        " unevictable:%lukB"
4383                        " writepending:%lukB"
4384                        " present:%lukB"
4385                        " managed:%lukB"
4386                        " mlocked:%lukB"
4387                        " slab_reclaimable:%lukB"
4388                        " slab_unreclaimable:%lukB"
4389                        " kernel_stack:%lukB"
4390                        " pagetables:%lukB"
4391                        " bounce:%lukB"
4392                        " free_pcp:%lukB"
4393                        " local_pcp:%ukB"
4394                        " free_cma:%lukB"
4395                        "\n",
4396                        zone->name,
4397                        K(zone_page_state(zone, NR_FREE_PAGES)),
4398                        K(min_wmark_pages(zone)),
4399                        K(low_wmark_pages(zone)),
4400                        K(high_wmark_pages(zone)),
4401                        K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
4402                        K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
4403                        K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
4404                        K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
4405                        K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
4406                        K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
4407                        K(zone->present_pages),
4408                        K(zone->managed_pages),
4409                        K(zone_page_state(zone, NR_MLOCK)),
4410                        K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
4411                        K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
4412                        zone_page_state(zone, NR_KERNEL_STACK_KB),
4413                        K(zone_page_state(zone, NR_PAGETABLE)),
4414                        K(zone_page_state(zone, NR_BOUNCE)),
4415                        K(free_pcp),
4416                        K(this_cpu_read(zone->pageset->pcp.count)),
4417                        K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
4418                printk("lowmem_reserve[]:");
4419                for (i = 0; i < MAX_NR_ZONES; i++)
4420                        printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
4421                printk(KERN_CONT "\n");
4422        }
4423
4424        for_each_populated_zone(zone) {
4425                unsigned int order;
4426                unsigned long nr[MAX_ORDER], flags, total = 0;
4427                unsigned char types[MAX_ORDER];
4428
4429                if (skip_free_areas_node(filter, zone_to_nid(zone)))
4430                        continue;
4431                show_node(zone);
4432                printk(KERN_CONT "%s: ", zone->name);
4433
4434                spin_lock_irqsave(&zone->lock, flags);
4435                for (order = 0; order < MAX_ORDER; order++) {
4436                        struct free_area *area = &zone->free_area[order];
4437                        int type;
4438
4439                        nr[order] = area->nr_free;
4440                        total += nr[order] << order;
4441
4442                        types[order] = 0;
4443                        for (type = 0; type < MIGRATE_TYPES; type++) {
4444                                if (!list_empty(&area->free_list[type]))
4445                                        types[order] |= 1 << type;
4446                        }
4447                }
4448                spin_unlock_irqrestore(&zone->lock, flags);
4449                for (order = 0; order < MAX_ORDER; order++) {
4450                        printk(KERN_CONT "%lu*%lukB ",
4451                               nr[order], K(1UL) << order);
4452                        if (nr[order])
4453                                show_migration_types(types[order]);
4454                }
4455                printk(KERN_CONT "= %lukB\n", K(total));
4456        }
4457
4458        hugetlb_show_meminfo();
4459
4460        printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
4461
4462        show_swap_cache_info();
4463}
4464
4465static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
4466{
4467        zoneref->zone = zone;
4468        zoneref->zone_idx = zone_idx(zone);
4469}
4470
4471/*
4472 * Builds allocation fallback zone lists.
4473 *
4474 * Add all populated zones of a node to the zonelist.
4475 */
4476static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
4477                                int nr_zones)
4478{
4479        struct zone *zone;
4480        enum zone_type zone_type = MAX_NR_ZONES;
4481
4482        do {
4483                zone_type--;
4484                zone = pgdat->node_zones + zone_type;
4485                if (managed_zone(zone)) {
4486                        zoneref_set_zone(zone,
4487                                &zonelist->_zonerefs[nr_zones++]);
4488                        check_highest_zone(zone_type);
4489                }
4490        } while (zone_type);
4491
4492        return nr_zones;
4493}
4494
4495
4496/*
4497 *  zonelist_order:
4498 *  0 = automatic detection of better ordering.
4499 *  1 = order by ([node] distance, -zonetype)
4500 *  2 = order by (-zonetype, [node] distance)
4501 *
4502 *  If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
4503 *  the same zonelist. So only NUMA can configure this param.
4504 */
4505#define ZONELIST_ORDER_DEFAULT  0
4506#define ZONELIST_ORDER_NODE     1
4507#define ZONELIST_ORDER_ZONE     2
4508
4509/* zonelist order in the kernel.
4510 * set_zonelist_order() will set this to NODE or ZONE.
4511 */
4512static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
4513static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
4514
4515
4516#ifdef CONFIG_NUMA
4517/* The value user specified ....changed by config */
4518static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
4519/* string for sysctl */
4520#define NUMA_ZONELIST_ORDER_LEN 16
4521char numa_zonelist_order[16] = "default";
4522
4523/*
4524 * interface for configure zonelist ordering.
4525 * command line option "numa_zonelist_order"
4526 *      = "[dD]efault   - default, automatic configuration.
4527 *      = "[nN]ode      - order by node locality, then by zone within node
4528 *      = "[zZ]one      - order by zone, then by locality within zone
4529 */
4530
4531static int __parse_numa_zonelist_order(char *s)
4532{
4533        if (*s == 'd' || *s == 'D') {
4534                user_zonelist_order = ZONELIST_ORDER_DEFAULT;
4535        } else if (*s == 'n' || *s == 'N') {
4536                user_zonelist_order = ZONELIST_ORDER_NODE;
4537        } else if (*s == 'z' || *s == 'Z') {
4538                user_zonelist_order = ZONELIST_ORDER_ZONE;
4539        } else {
4540                pr_warn("Ignoring invalid numa_zonelist_order value:  %s\n", s);
4541                return -EINVAL;
4542        }
4543        return 0;
4544}
4545
4546static __init int setup_numa_zonelist_order(char *s)
4547{
4548        int ret;
4549
4550        if (!s)
4551                return 0;
4552
4553        ret = __parse_numa_zonelist_order(s);
4554        if (ret == 0)
4555                strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
4556
4557        return ret;
4558}
4559early_param("numa_zonelist_order", setup_numa_zonelist_order);
4560
4561/*
4562 * sysctl handler for numa_zonelist_order
4563 */
4564int numa_zonelist_order_handler(struct ctl_table *table, int write,
4565                void __user *buffer, size_t *length,
4566                loff_t *ppos)
4567{
4568        char saved_string[NUMA_ZONELIST_ORDER_LEN];
4569        int ret;
4570        static DEFINE_MUTEX(zl_order_mutex);
4571
4572        mutex_lock(&zl_order_mutex);
4573        if (write) {
4574                if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
4575                        ret = -EINVAL;
4576                        goto out;
4577                }
4578                strcpy(saved_string, (char *)table->data);
4579        }
4580        ret = proc_dostring(table, write, buffer, length, ppos);
4581        if (ret)
4582                goto out;
4583        if (write) {
4584                int oldval = user_zonelist_order;
4585
4586                ret = __parse_numa_zonelist_order((char *)table->data);
4587                if (ret) {
4588                        /*
4589                         * bogus value.  restore saved string
4590                         */
4591                        strncpy((char *)table->data, saved_string,
4592                                NUMA_ZONELIST_ORDER_LEN);
4593                        user_zonelist_order = oldval;
4594                } else if (oldval != user_zonelist_order) {
4595                        mutex_lock(&zonelists_mutex);
4596                        build_all_zonelists(NULL, NULL);
4597                        mutex_unlock(&zonelists_mutex);
4598                }
4599        }
4600out:
4601        mutex_unlock(&zl_order_mutex);
4602        return ret;
4603}
4604
4605
4606#define MAX_NODE_LOAD (nr_online_nodes)
4607static int node_load[MAX_NUMNODES];
4608
4609/**
4610 * find_next_best_node - find the next node that should appear in a given node's fallback list
4611 * @node: node whose fallback list we're appending
4612 * @used_node_mask: nodemask_t of already used nodes
4613 *
4614 * We use a number of factors to determine which is the next node that should
4615 * appear on a given node's fallback list.  The node should not have appeared
4616 * already in @node's fallback list, and it should be the next closest node
4617 * according to the distance array (which contains arbitrary distance values
4618 * from each node to each node in the system), and should also prefer nodes
4619 * with no CPUs, since presumably they'll have very little allocation pressure
4620 * on them otherwise.
4621 * It returns -1 if no node is found.
4622 */
4623static int find_next_best_node(int node, nodemask_t *used_node_mask)
4624{
4625        int n, val;
4626        int min_val = INT_MAX;
4627        int best_node = NUMA_NO_NODE;
4628        const struct cpumask *tmp = cpumask_of_node(0);
4629
4630        /* Use the local node if we haven't already */
4631        if (!node_isset(node, *used_node_mask)) {
4632                node_set(node, *used_node_mask);
4633                return node;
4634        }
4635
4636        for_each_node_state(n, N_MEMORY) {
4637
4638                /* Don't want a node to appear more than once */
4639                if (node_isset(n, *used_node_mask))
4640                        continue;
4641
4642                /* Use the distance array to find the distance */
4643                val = node_distance(node, n);
4644
4645                /* Penalize nodes under us ("prefer the next node") */
4646                val += (n < node);
4647
4648                /* Give preference to headless and unused nodes */
4649                tmp = cpumask_of_node(n);
4650                if (!cpumask_empty(tmp))
4651                        val += PENALTY_FOR_NODE_WITH_CPUS;
4652
4653                /* Slight preference for less loaded node */
4654                val *= (MAX_NODE_LOAD*MAX_NUMNODES);
4655                val += node_load[n];
4656
4657                if (val < min_val) {
4658                        min_val = val;
4659                        best_node = n;
4660                }
4661        }
4662
4663        if (best_node >= 0)
4664                node_set(best_node, *used_node_mask);
4665
4666        return best_node;
4667}
4668
4669
4670/*
4671 * Build zonelists ordered by node and zones within node.
4672 * This results in maximum locality--normal zone overflows into local
4673 * DMA zone, if any--but risks exhausting DMA zone.
4674 */
4675static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
4676{
4677        int j;
4678        struct zonelist *zonelist;
4679
4680        zonelist = &pgdat->node_zonelists[ZONELIST_FALLBACK];
4681        for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
4682                ;
4683        j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4684        zonelist->_zonerefs[j].zone = NULL;
4685        zonelist->_zonerefs[j].zone_idx = 0;
4686}
4687
4688/*
4689 * Build gfp_thisnode zonelists
4690 */
4691static void build_thisnode_zonelists(pg_data_t *pgdat)
4692{
4693        int j;
4694        struct zonelist *zonelist;
4695
4696        zonelist = &pgdat->node_zonelists[ZONELIST_NOFALLBACK];
4697        j = build_zonelists_node(pgdat, zonelist, 0);
4698        zonelist->_zonerefs[j].zone = NULL;
4699        zonelist->_zonerefs[j].zone_idx = 0;
4700}
4701
4702/*
4703 * Build zonelists ordered by zone and nodes within zones.
4704 * This results in conserving DMA zone[s] until all Normal memory is
4705 * exhausted, but results in overflowing to remote node while memory
4706 * may still exist in local DMA zone.
4707 */
4708static int node_order[MAX_NUMNODES];
4709
4710static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
4711{
4712        int pos, j, node;
4713        int zone_type;          /* needs to be signed */
4714        struct zone *z;
4715        struct zonelist *zonelist;
4716
4717        zonelist = &pgdat->node_zonelists[ZONELIST_FALLBACK];
4718        pos = 0;
4719        for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
4720                for (j = 0; j < nr_nodes; j++) {
4721                        node = node_order[j];
4722                        z = &NODE_DATA(node)->node_zones[zone_type];
4723                        if (managed_zone(z)) {
4724                                zoneref_set_zone(z,
4725                                        &zonelist->_zonerefs[pos++]);
4726                                check_highest_zone(zone_type);
4727                        }
4728                }
4729        }
4730        zonelist->_zonerefs[pos].zone = NULL;
4731        zonelist->_zonerefs[pos].zone_idx = 0;
4732}
4733
4734#if defined(CONFIG_64BIT)
4735/*
4736 * Devices that require DMA32/DMA are relatively rare and do not justify a
4737 * penalty to every machine in case the specialised case applies. Default
4738 * to Node-ordering on 64-bit NUMA machines
4739 */
4740static int default_zonelist_order(void)
4741{
4742        return ZONELIST_ORDER_NODE;
4743}
4744#else
4745/*
4746 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4747 * by the kernel. If processes running on node 0 deplete the low memory zone
4748 * then reclaim will occur more frequency increasing stalls and potentially
4749 * be easier to OOM if a large percentage of the zone is under writeback or
4750 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4751 * Hence, default to zone ordering on 32-bit.
4752 */
4753static int default_zonelist_order(void)
4754{
4755        return ZONELIST_ORDER_ZONE;
4756}
4757#endif /* CONFIG_64BIT */
4758
4759static void set_zonelist_order(void)
4760{
4761        if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
4762                current_zonelist_order = default_zonelist_order();
4763        else
4764                current_zonelist_order = user_zonelist_order;
4765}
4766
4767static void build_zonelists(pg_data_t *pgdat)
4768{
4769        int i, node, load;
4770        nodemask_t used_mask;
4771        int local_node, prev_node;
4772        struct zonelist *zonelist;
4773        unsigned int order = current_zonelist_order;
4774
4775        /* initialize zonelists */
4776        for (i = 0; i < MAX_ZONELISTS; i++) {
4777                zonelist = pgdat->node_zonelists + i;
4778                zonelist->_zonerefs[0].zone = NULL;
4779                zonelist->_zonerefs[0].zone_idx = 0;
4780        }
4781
4782        /* NUMA-aware ordering of nodes */
4783        local_node = pgdat->node_id;
4784        load = nr_online_nodes;
4785        prev_node = local_node;
4786        nodes_clear(used_mask);
4787
4788        memset(node_order, 0, sizeof(node_order));
4789        i = 0;
4790
4791        while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
4792                /*
4793                 * We don't want to pressure a particular node.
4794                 * So adding penalty to the first node in same
4795                 * distance group to make it round-robin.
4796                 */
4797                if (node_distance(local_node, node) !=
4798                    node_distance(local_node, prev_node))
4799                        node_load[node] = load;
4800
4801                prev_node = node;
4802                load--;
4803                if (order == ZONELIST_ORDER_NODE)
4804                        build_zonelists_in_node_order(pgdat, node);
4805                else
4806                        node_order[i++] = node; /* remember order */
4807        }
4808
4809        if (order == ZONELIST_ORDER_ZONE) {
4810                /* calculate node order -- i.e., DMA last! */
4811                build_zonelists_in_zone_order(pgdat, i);
4812        }
4813
4814        build_thisnode_zonelists(pgdat);
4815}
4816
4817#ifdef CONFIG_HAVE_MEMORYLESS_NODES
4818/*
4819 * Return node id of node used for "local" allocations.
4820 * I.e., first node id of first zone in arg node's generic zonelist.
4821 * Used for initializing percpu 'numa_mem', which is used primarily
4822 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4823 */
4824int local_memory_node(int node)
4825{
4826        struct zoneref *z;
4827
4828        z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
4829                                   gfp_zone(GFP_KERNEL),
4830                                   NULL);
4831        return z->zone->node;
4832}
4833#endif
4834
4835static void setup_min_unmapped_ratio(void);
4836static void setup_min_slab_ratio(void);
4837#else   /* CONFIG_NUMA */
4838
4839static void set_zonelist_order(void)
4840{
4841        current_zonelist_order = ZONELIST_ORDER_ZONE;
4842}
4843
4844static void build_zonelists(pg_data_t *pgdat)
4845{
4846        int node, local_node;
4847        enum zone_type j;
4848        struct zonelist *zonelist;
4849
4850        local_node = pgdat->node_id;
4851
4852        zonelist = &pgdat->node_zonelists[ZONELIST_FALLBACK];
4853        j = build_zonelists_node(pgdat, zonelist, 0);
4854
4855        /*
4856         * Now we build the zonelist so that it contains the zones
4857         * of all the other nodes.
4858         * We don't want to pressure a particular node, so when
4859         * building the zones for node N, we make sure that the
4860         * zones coming right after the local ones are those from
4861         * node N+1 (modulo N)
4862         */
4863        for (node = local_node + 1; node < MAX_NUMNODES; node++) {
4864                if (!node_online(node))
4865                        continue;
4866                j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4867        }
4868        for (node = 0; node < local_node; node++) {
4869                if (!node_online(node))
4870                        continue;
4871                j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4872        }
4873
4874        zonelist->_zonerefs[j].zone = NULL;
4875        zonelist->_zonerefs[j].zone_idx = 0;
4876}
4877
4878#endif  /* CONFIG_NUMA */
4879
4880/*
4881 * Boot pageset table. One per cpu which is going to be used for all
4882 * zones and all nodes. The parameters will be set in such a way
4883 * that an item put on a list will immediately be handed over to
4884 * the buddy list. This is safe since pageset manipulation is done
4885 * with interrupts disabled.
4886 *
4887 * The boot_pagesets must be kept even after bootup is complete for
4888 * unused processors and/or zones. They do play a role for bootstrapping
4889 * hotplugged processors.
4890 *
4891 * zoneinfo_show() and maybe other functions do
4892 * not check if the processor is online before following the pageset pointer.
4893 * Other parts of the kernel may not check if the zone is available.
4894 */
4895static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
4896static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
4897static void setup_zone_pageset(struct zone *zone);
4898
4899/*
4900 * Global mutex to protect against size modification of zonelists
4901 * as well as to serialize pageset setup for the new populated zone.
4902 */
4903DEFINE_MUTEX(zonelists_mutex);
4904
4905/* return values int ....just for stop_machine() */
4906static int __build_all_zonelists(void *data)
4907{
4908        int nid;
4909        int cpu;
4910        pg_data_t *self = data;
4911
4912#ifdef CONFIG_NUMA
4913        memset(node_load, 0, sizeof(node_load));
4914#endif
4915
4916        if (self && !node_online(self->node_id)) {
4917                build_zonelists(self);
4918        }
4919
4920        for_each_online_node(nid) {
4921                pg_data_t *pgdat = NODE_DATA(nid);
4922
4923                build_zonelists(pgdat);
4924        }
4925
4926        /*
4927         * Initialize the boot_pagesets that are going to be used
4928         * for bootstrapping processors. The real pagesets for
4929         * each zone will be allocated later when the per cpu
4930         * allocator is available.
4931         *
4932         * boot_pagesets are used also for bootstrapping offline
4933         * cpus if the system is already booted because the pagesets
4934         * are needed to initialize allocators on a specific cpu too.
4935         * F.e. the percpu allocator needs the page allocator which
4936         * needs the percpu allocator in order to allocate its pagesets
4937         * (a chicken-egg dilemma).
4938         */
4939        for_each_possible_cpu(cpu) {
4940                setup_pageset(&per_cpu(boot_pageset, cpu), 0);
4941
4942#ifdef CONFIG_HAVE_MEMORYLESS_NODES
4943                /*
4944                 * We now know the "local memory node" for each node--
4945                 * i.e., the node of the first zone in the generic zonelist.
4946                 * Set up numa_mem percpu variable for on-line cpus.  During
4947                 * boot, only the boot cpu should be on-line;  we'll init the
4948                 * secondary cpus' numa_mem as they come on-line.  During
4949                 * node/memory hotplug, we'll fixup all on-line cpus.
4950                 */
4951                if (cpu_online(cpu))
4952                        set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
4953#endif
4954        }
4955
4956        return 0;
4957}
4958
4959static noinline void __init
4960build_all_zonelists_init(void)
4961{
4962        __build_all_zonelists(NULL);
4963        mminit_verify_zonelist();
4964        cpuset_init_current_mems_allowed();
4965}
4966
4967/*
4968 * Called with zonelists_mutex held always
4969 * unless system_state == SYSTEM_BOOTING.
4970 *
4971 * __ref due to (1) call of __meminit annotated setup_zone_pageset
4972 * [we're only called with non-NULL zone through __meminit paths] and
4973 * (2) call of __init annotated helper build_all_zonelists_init
4974 * [protected by SYSTEM_BOOTING].
4975 */
4976void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
4977{
4978        set_zonelist_order();
4979
4980        if (system_state == SYSTEM_BOOTING) {
4981                build_all_zonelists_init();
4982        } else {
4983#ifdef CONFIG_MEMORY_HOTPLUG
4984                if (zone)
4985                        setup_zone_pageset(zone);
4986#endif
4987                /* we have to stop all cpus to guarantee there is no user
4988                   of zonelist */
4989                stop_machine(__build_all_zonelists, pgdat, NULL);
4990                /* cpuset refresh routine should be here */
4991        }
4992        vm_total_pages = nr_free_pagecache_pages();
4993        /*
4994         * Disable grouping by mobility if the number of pages in the
4995         * system is too low to allow the mechanism to work. It would be
4996         * more accurate, but expensive to check per-zone. This check is
4997         * made on memory-hotadd so a system can start with mobility
4998         * disabled and enable it later
4999         */
5000        if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5001                page_group_by_mobility_disabled = 1;
5002        else
5003                page_group_by_mobility_disabled = 0;
5004
5005        pr_info("Built %i zonelists in %s order, mobility grouping %s.  Total pages: %ld\n",
5006                nr_online_nodes,
5007                zonelist_order_name[current_zonelist_order],
5008                page_group_by_mobility_disabled ? "off" : "on",
5009                vm_total_pages);
5010#ifdef CONFIG_NUMA
5011        pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5012#endif
5013}
5014
5015/*
5016 * Initially all pages are reserved - free ones are freed
5017 * up by free_all_bootmem() once the early boot process is
5018 * done. Non-atomic initialization, single-pass.
5019 */
5020void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5021                unsigned long start_pfn, enum memmap_context context)
5022{
5023        struct vmem_altmap *altmap = to_vmem_altmap(__pfn_to_phys(start_pfn));
5024        unsigned long end_pfn = start_pfn + size;
5025        pg_data_t *pgdat = NODE_DATA(nid);
5026        unsigned long pfn;
5027        unsigned long nr_initialised = 0;
5028#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5029        struct memblock_region *r = NULL, *tmp;
5030#endif
5031
5032        if (highest_memmap_pfn < end_pfn - 1)
5033                highest_memmap_pfn = end_pfn - 1;
5034
5035        /*
5036         * Honor reservation requested by the driver for this ZONE_DEVICE
5037         * memory
5038         */
5039        if (altmap && start_pfn == altmap->base_pfn)
5040                start_pfn += altmap->reserve;
5041
5042        for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5043                /*
5044                 * There can be holes in boot-time mem_map[]s handed to this
5045                 * function.  They do not exist on hotplugged memory.
5046                 */
5047                if (context != MEMMAP_EARLY)
5048                        goto not_early;
5049
5050                if (!early_pfn_valid(pfn))
5051                        continue;
5052                if (!early_pfn_in_nid(pfn, nid))
5053                        continue;
5054                if (!update_defer_init(pgdat, pfn, end_pfn, &nr_initialised))
5055                        break;
5056
5057#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5058                /*
5059                 * Check given memblock attribute by firmware which can affect
5060                 * kernel memory layout.  If zone==ZONE_MOVABLE but memory is
5061                 * mirrored, it's an overlapped memmap init. skip it.
5062                 */
5063                if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5064                        if (!r || pfn >= memblock_region_memory_end_pfn(r)) {
5065                                for_each_memblock(memory, tmp)
5066                                        if (pfn < memblock_region_memory_end_pfn(tmp))
5067                                                break;
5068                                r = tmp;
5069                        }
5070                        if (pfn >= memblock_region_memory_base_pfn(r) &&
5071                            memblock_is_mirror(r)) {
5072                                /* already initialized as NORMAL */
5073                                pfn = memblock_region_memory_end_pfn(r);
5074                                continue;
5075                        }
5076                }
5077#endif
5078
5079not_early:
5080                /*
5081                 * Mark the block movable so that blocks are reserved for
5082                 * movable at startup. This will force kernel allocations
5083                 * to reserve their blocks rather than leaking throughout
5084                 * the address space during boot when many long-lived
5085                 * kernel allocations are made.
5086                 *
5087                 * bitmap is created for zone's valid pfn range. but memmap
5088                 * can be created for invalid pages (for alignment)
5089                 * check here not to call set_pageblock_migratetype() against
5090                 * pfn out of zone.
5091                 */
5092                if (!(pfn & (pageblock_nr_pages - 1))) {
5093                        struct page *page = pfn_to_page(pfn);
5094
5095                        __init_single_page(page, pfn, zone, nid);
5096                        set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5097                } else {
5098                        __init_single_pfn(pfn, zone, nid);
5099                }
5100        }
5101}
5102
5103static void __meminit zone_init_free_lists(struct zone *zone)
5104{
5105        unsigned int order, t;
5106        for_each_migratetype_order(order, t) {
5107                INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
5108                zone->free_area[order].nr_free = 0;
5109        }
5110}
5111
5112#ifndef __HAVE_ARCH_MEMMAP_INIT
5113#define memmap_init(size, nid, zone, start_pfn) \
5114        memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
5115#endif
5116
5117static int zone_batchsize(struct zone *zone)
5118{
5119#ifdef CONFIG_MMU
5120        int batch;
5121
5122        /*
5123         * The per-cpu-pages pools are set to around 1000th of the
5124         * size of the zone.  But no more than 1/2 of a meg.
5125         *
5126         * OK, so we don't know how big the cache is.  So guess.
5127         */
5128        batch = zone->managed_pages / 1024;
5129        if (batch * PAGE_SIZE > 512 * 1024)
5130                batch = (512 * 1024) / PAGE_SIZE;
5131        batch /= 4;             /* We effectively *= 4 below */
5132        if (batch < 1)
5133                batch = 1;
5134
5135        /*
5136         * Clamp the batch to a 2^n - 1 value. Having a power
5137         * of 2 value was found to be more likely to have
5138         * suboptimal cache aliasing properties in some cases.
5139         *
5140         * For example if 2 tasks are alternately allocating
5141         * batches of pages, one task can end up with a lot
5142         * of pages of one half of the possible page colors
5143         * and the other with pages of the other colors.
5144         */
5145        batch = rounddown_pow_of_two(batch + batch/2) - 1;
5146
5147        return batch;
5148
5149#else
5150        /* The deferral and batching of frees should be suppressed under NOMMU
5151         * conditions.
5152         *
5153         * The problem is that NOMMU needs to be able to allocate large chunks
5154         * of contiguous memory as there's no hardware page translation to
5155         * assemble apparent contiguous memory from discontiguous pages.
5156         *
5157         * Queueing large contiguous runs of pages for batching, however,
5158         * causes the pages to actually be freed in smaller chunks.  As there
5159         * can be a significant delay between the individual batches being
5160         * recycled, this leads to the once large chunks of space being
5161         * fragmented and becoming unavailable for high-order allocations.
5162         */
5163        return 0;
5164#endif
5165}
5166
5167/*
5168 * pcp->high and pcp->batch values are related and dependent on one another:
5169 * ->batch must never be higher then ->high.
5170 * The following function updates them in a safe manner without read side
5171 * locking.
5172 *
5173 * Any new users of pcp->batch and pcp->high should ensure they can cope with
5174 * those fields changing asynchronously (acording the the above rule).
5175 *
5176 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5177 * outside of boot time (or some other assurance that no concurrent updaters
5178 * exist).
5179 */
5180static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
5181                unsigned long batch)
5182{
5183       /* start with a fail safe value for batch */
5184        pcp->batch = 1;
5185        smp_wmb();
5186
5187       /* Update high, then batch, in order */
5188        pcp->high = high;
5189        smp_wmb();
5190
5191        pcp->batch = batch;
5192}
5193
5194/* a companion to pageset_set_high() */
5195static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
5196{
5197        pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
5198}
5199
5200static void pageset_init(struct per_cpu_pageset *p)
5201{
5202        struct per_cpu_pages *pcp;
5203        int migratetype;
5204
5205        memset(p, 0, sizeof(*p));
5206
5207        pcp = &p->pcp;
5208        pcp->count = 0;
5209        for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
5210                INIT_LIST_HEAD(&pcp->lists[migratetype]);
5211}
5212
5213static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
5214{
5215        pageset_init(p);
5216        pageset_set_batch(p, batch);
5217}
5218
5219/*
5220 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5221 * to the value high for the pageset p.
5222 */
5223static void pageset_set_high(struct per_cpu_pageset *p,
5224                                unsigned long high)
5225{
5226        unsigned long batch = max(1UL, high / 4);
5227        if ((high / 4) > (PAGE_SHIFT * 8))
5228                batch = PAGE_SHIFT * 8;
5229
5230        pageset_update(&p->pcp, high, batch);
5231}
5232
5233static void pageset_set_high_and_batch(struct zone *zone,
5234                                       struct per_cpu_pageset *pcp)
5235{
5236        if (percpu_pagelist_fraction)
5237                pageset_set_high(pcp,
5238                        (zone->managed_pages /
5239                                percpu_pagelist_fraction));
5240        else
5241                pageset_set_batch(pcp, zone_batchsize(zone));
5242}
5243
5244static void __meminit zone_pageset_init(struct zone *zone, int cpu)
5245{
5246        struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
5247
5248        pageset_init(pcp);
5249        pageset_set_high_and_batch(zone, pcp);
5250}
5251
5252static void __meminit setup_zone_pageset(struct zone *zone)
5253{
5254        int cpu;
5255        zone->pageset = alloc_percpu(struct per_cpu_pageset);
5256        for_each_possible_cpu(cpu)
5257                zone_pageset_init(zone, cpu);
5258}
5259
5260/*
5261 * Allocate per cpu pagesets and initialize them.
5262 * Before this call only boot pagesets were available.
5263 */
5264void __init setup_per_cpu_pageset(void)
5265{
5266        struct pglist_data *pgdat;
5267        struct zone *zone;
5268
5269        for_each_populated_zone(zone)
5270                setup_zone_pageset(zone);
5271
5272        for_each_online_pgdat(pgdat)
5273                pgdat->per_cpu_nodestats =
5274                        alloc_percpu(struct per_cpu_nodestat);
5275}
5276
5277static __meminit void zone_pcp_init(struct zone *zone)
5278{
5279        /*
5280         * per cpu subsystem is not up at this point. The following code
5281         * relies on the ability of the linker to provide the
5282         * offset of a (static) per cpu variable into the per cpu area.
5283         */
5284        zone->pageset = &boot_pageset;
5285
5286        if (populated_zone(zone))
5287                printk(KERN_DEBUG "  %s zone: %lu pages, LIFO batch:%u\n",
5288                        zone->name, zone->present_pages,
5289                                         zone_batchsize(zone));
5290}
5291
5292int __meminit init_currently_empty_zone(struct zone *zone,
5293                                        unsigned long zone_start_pfn,
5294                                        unsigned long size)
5295{
5296        struct pglist_data *pgdat = zone->zone_pgdat;
5297
5298        pgdat->nr_zones = zone_idx(zone) + 1;
5299
5300        zone->zone_start_pfn = zone_start_pfn;
5301
5302        mminit_dprintk(MMINIT_TRACE, "memmap_init",
5303                        "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5304                        pgdat->node_id,
5305                        (unsigned long)zone_idx(zone),
5306                        zone_start_pfn, (zone_start_pfn + size));
5307
5308        zone_init_free_lists(zone);
5309        zone->initialized = 1;
5310
5311        return 0;
5312}
5313
5314#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5315#ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5316
5317/*
5318 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5319 */
5320int __meminit __early_pfn_to_nid(unsigned long pfn,
5321                                        struct mminit_pfnnid_cache *state)
5322{
5323        unsigned long start_pfn, end_pfn;
5324        int nid;
5325
5326        if (state->last_start <= pfn && pfn < state->last_end)
5327                return state->last_nid;
5328
5329        nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
5330        if (nid != -1) {
5331                state->last_start = start_pfn;
5332                state->last_end = end_pfn;
5333                state->last_nid = nid;
5334        }
5335
5336        return nid;
5337}
5338#endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5339
5340/**
5341 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5342 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5343 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5344 *
5345 * If an architecture guarantees that all ranges registered contain no holes
5346 * and may be freed, this this function may be used instead of calling
5347 * memblock_free_early_nid() manually.
5348 */
5349void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
5350{
5351        unsigned long start_pfn, end_pfn;
5352        int i, this_nid;
5353
5354        for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
5355                start_pfn = min(start_pfn, max_low_pfn);
5356                end_pfn = min(end_pfn, max_low_pfn);
5357
5358                if (start_pfn < end_pfn)
5359                        memblock_free_early_nid(PFN_PHYS(start_pfn),
5360                                        (end_pfn - start_pfn) << PAGE_SHIFT,
5361                                        this_nid);
5362        }
5363}
5364
5365/**
5366 * sparse_memory_present_with_active_regions - Call memory_present for each active range
5367 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
5368 *
5369 * If an architecture guarantees that all ranges registered contain no holes and may
5370 * be freed, this function may be used instead of calling memory_present() manually.
5371 */
5372void __init sparse_memory_present_with_active_regions(int nid)
5373{
5374        unsigned long start_pfn, end_pfn;
5375        int i, this_nid;
5376
5377        for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
5378                memory_present(this_nid, start_pfn, end_pfn);
5379}
5380
5381/**
5382 * get_pfn_range_for_nid - Return the start and end page frames for a node
5383 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5384 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5385 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5386 *
5387 * It returns the start and end page frame of a node based on information
5388 * provided by memblock_set_node(). If called for a node
5389 * with no available memory, a warning is printed and the start and end
5390 * PFNs will be 0.
5391 */
5392void __meminit get_pfn_range_for_nid(unsigned int nid,
5393                        unsigned long *start_pfn, unsigned long *end_pfn)
5394{
5395        unsigned long this_start_pfn, this_end_pfn;
5396        int i;
5397
5398        *start_pfn = -1UL;
5399        *end_pfn = 0;
5400
5401        for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
5402                *start_pfn = min(*start_pfn, this_start_pfn);
5403                *end_pfn = max(*end_pfn, this_end_pfn);
5404        }
5405
5406        if (*start_pfn == -1UL)
5407                *start_pfn = 0;
5408}
5409
5410/*
5411 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5412 * assumption is made that zones within a node are ordered in monotonic
5413 * increasing memory addresses so that the "highest" populated zone is used
5414 */
5415static void __init find_usable_zone_for_movable(void)
5416{
5417        int zone_index;
5418        for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
5419                if (zone_index == ZONE_MOVABLE)
5420                        continue;
5421
5422                if (arch_zone_highest_possible_pfn[zone_index] >
5423                                arch_zone_lowest_possible_pfn[zone_index])
5424                        break;
5425        }
5426
5427        VM_BUG_ON(zone_index == -1);
5428        movable_zone = zone_index;
5429}
5430
5431/*
5432 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5433 * because it is sized independent of architecture. Unlike the other zones,
5434 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5435 * in each node depending on the size of each node and how evenly kernelcore
5436 * is distributed. This helper function adjusts the zone ranges
5437 * provided by the architecture for a given node by using the end of the
5438 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5439 * zones within a node are in order of monotonic increases memory addresses
5440 */
5441static void __meminit adjust_zone_range_for_zone_movable(int nid,
5442                                        unsigned long zone_type,
5443                                        unsigned long node_start_pfn,
5444                                        unsigned long node_end_pfn,
5445                                        unsigned long *zone_start_pfn,
5446                                        unsigned long *zone_end_pfn)
5447{
5448        /* Only adjust if ZONE_MOVABLE is on this node */
5449        if (zone_movable_pfn[nid]) {
5450                /* Size ZONE_MOVABLE */
5451                if (zone_type == ZONE_MOVABLE) {
5452                        *zone_start_pfn = zone_movable_pfn[nid];
5453                        *zone_end_pfn = min(node_end_pfn,
5454                                arch_zone_highest_possible_pfn[movable_zone]);
5455
5456                /* Adjust for ZONE_MOVABLE starting within this range */
5457                } else if (!mirrored_kernelcore &&
5458                        *zone_start_pfn < zone_movable_pfn[nid] &&
5459                        *zone_end_pfn > zone_movable_pfn[nid]) {
5460                        *zone_end_pfn = zone_movable_pfn[nid];
5461
5462                /* Check if this whole range is within ZONE_MOVABLE */
5463                } else if (*zone_start_pfn >= zone_movable_pfn[nid])
5464                        *zone_start_pfn = *zone_end_pfn;
5465        }
5466}
5467
5468/*
5469 * Return the number of pages a zone spans in a node, including holes
5470 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5471 */
5472static unsigned long __meminit zone_spanned_pages_in_node(int nid,
5473                                        unsigned long zone_type,
5474                                        unsigned long node_start_pfn,
5475                                        unsigned long node_end_pfn,
5476                                        unsigned long *zone_start_pfn,
5477                                        unsigned long *zone_end_pfn,
5478                                        unsigned long *ignored)
5479{
5480        /* When hotadd a new node from cpu_up(), the node should be empty */
5481        if (!node_start_pfn && !node_end_pfn)
5482                return 0;
5483
5484        /* Get the start and end of the zone */
5485        *zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
5486        *zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
5487        adjust_zone_range_for_zone_movable(nid, zone_type,
5488                                node_start_pfn, node_end_pfn,
5489                                zone_start_pfn, zone_end_pfn);
5490
5491        /* Check that this node has pages within the zone's required range */
5492        if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
5493                return 0;
5494
5495        /* Move the zone boundaries inside the node if necessary */
5496        *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
5497        *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
5498
5499        /* Return the spanned pages */
5500        return *zone_end_pfn - *zone_start_pfn;
5501}
5502
5503/*
5504 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5505 * then all holes in the requested range will be accounted for.
5506 */
5507unsigned long __meminit __absent_pages_in_range(int nid,
5508                                unsigned long range_start_pfn,
5509                                unsigned long range_end_pfn)
5510{
5511        unsigned long nr_absent = range_end_pfn - range_start_pfn;
5512        unsigned long start_pfn, end_pfn;
5513        int i;
5514
5515        for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5516                start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
5517                end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
5518                nr_absent -= end_pfn - start_pfn;
5519        }
5520        return nr_absent;
5521}
5522
5523/**
5524 * absent_pages_in_range - Return number of page frames in holes within a range
5525 * @start_pfn: The start PFN to start searching for holes
5526 * @end_pfn: The end PFN to stop searching for holes
5527 *
5528 * It returns the number of pages frames in memory holes within a range.
5529 */
5530unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5531                                                        unsigned long end_pfn)
5532{
5533        return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5534}
5535
5536/* Return the number of page frames in holes in a zone on a node */
5537static unsigned long __meminit zone_absent_pages_in_node(int nid,
5538                                        unsigned long zone_type,
5539                                        unsigned long node_start_pfn,
5540                                        unsigned long node_end_pfn,
5541                                        unsigned long *ignored)
5542{
5543        unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5544        unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5545        unsigned long zone_start_pfn, zone_end_pfn;
5546        unsigned long nr_absent;
5547
5548        /* When hotadd a new node from cpu_up(), the node should be empty */
5549        if (!node_start_pfn && !node_end_pfn)
5550                return 0;
5551
5552        zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5553        zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5554
5555        adjust_zone_range_for_zone_movable(nid, zone_type,
5556                        node_start_pfn, node_end_pfn,
5557                        &zone_start_pfn, &zone_end_pfn);
5558        nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5559
5560        /*
5561         * ZONE_MOVABLE handling.
5562         * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
5563         * and vice versa.
5564         */
5565        if (mirrored_kernelcore && zone_movable_pfn[nid]) {
5566                unsigned long start_pfn, end_pfn;
5567                struct memblock_region *r;
5568
5569                for_each_memblock(memory, r) {
5570                        start_pfn = clamp(memblock_region_memory_base_pfn(r),
5571                                          zone_start_pfn, zone_end_pfn);
5572                        end_pfn = clamp(memblock_region_memory_end_pfn(r),
5573                                        zone_start_pfn, zone_end_pfn);
5574
5575                        if (zone_type == ZONE_MOVABLE &&
5576                            memblock_is_mirror(r))
5577                                nr_absent += end_pfn - start_pfn;
5578
5579                        if (zone_type == ZONE_NORMAL &&
5580                            !memblock_is_mirror(r))
5581                                nr_absent += end_pfn - start_pfn;
5582                }
5583        }
5584
5585        return nr_absent;
5586}
5587
5588#else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5589static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5590                                        unsigned long zone_type,
5591                                        unsigned long node_start_pfn,
5592                                        unsigned long node_end_pfn,
5593                                        unsigned long *zone_start_pfn,
5594                                        unsigned long *zone_end_pfn,
5595                                        unsigned long *zones_size)
5596{
5597        unsigned int zone;
5598
5599        *zone_start_pfn = node_start_pfn;
5600        for (zone = 0; zone < zone_type; zone++)
5601                *zone_start_pfn += zones_size[zone];
5602
5603        *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
5604
5605        return zones_size[zone_type];
5606}
5607
5608static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
5609                                                unsigned long zone_type,
5610                                                unsigned long node_start_pfn,
5611                                                unsigned long node_end_pfn,
5612                                                unsigned long *zholes_size)
5613{
5614        if (!zholes_size)
5615                return 0;
5616
5617        return zholes_size[zone_type];
5618}
5619
5620#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5621
5622static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5623                                                unsigned long node_start_pfn,
5624                                                unsigned long node_end_pfn,
5625                                                unsigned long *zones_size,
5626                                                unsigned long *zholes_size)
5627{
5628        unsigned long realtotalpages = 0, totalpages = 0;
5629        enum zone_type i;
5630
5631        for (i = 0; i < MAX_NR_ZONES; i++) {
5632                struct zone *zone = pgdat->node_zones + i;
5633                unsigned long zone_start_pfn, zone_end_pfn;
5634                unsigned long size, real_size;
5635
5636                size = zone_spanned_pages_in_node(pgdat->node_id, i,
5637                                                  node_start_pfn,
5638                                                  node_end_pfn,
5639                                                  &zone_start_pfn,
5640                                                  &zone_end_pfn,
5641                                                  zones_size);
5642                real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5643                                                  node_start_pfn, node_end_pfn,
5644                                                  zholes_size);
5645                if (size)
5646                        zone->zone_start_pfn = zone_start_pfn;
5647                else
5648                        zone->zone_start_pfn = 0;
5649                zone->spanned_pages = size;
5650                zone->present_pages = real_size;
5651
5652                totalpages += size;
5653                realtotalpages += real_size;
5654        }
5655
5656        pgdat->node_spanned_pages = totalpages;
5657        pgdat->node_present_pages = realtotalpages;
5658        printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5659                                                        realtotalpages);
5660}
5661
5662#ifndef CONFIG_SPARSEMEM
5663/*
5664 * Calculate the size of the zone->blockflags rounded to an unsigned long
5665 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5666 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5667 * round what is now in bits to nearest long in bits, then return it in
5668 * bytes.
5669 */
5670static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
5671{
5672        unsigned long usemapsize;
5673
5674        zonesize += zone_start_pfn & (pageblock_nr_pages-1);
5675        usemapsize = roundup(zonesize, pageblock_nr_pages);
5676        usemapsize = usemapsize >> pageblock_order;
5677        usemapsize *= NR_PAGEBLOCK_BITS;
5678        usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
5679
5680        return usemapsize / 8;
5681}
5682
5683static void __init setup_usemap(struct pglist_data *pgdat,
5684                                struct zone *zone,
5685                                unsigned long zone_start_pfn,
5686                                unsigned long zonesize)
5687{
5688        unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
5689        zone->pageblock_flags = NULL;
5690        if (usemapsize)
5691                zone->pageblock_flags =
5692                        memblock_virt_alloc_node_nopanic(usemapsize,
5693                                                         pgdat->node_id);
5694}
5695#else
5696static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
5697                                unsigned long zone_start_pfn, unsigned long zonesize) {}
5698#endif /* CONFIG_SPARSEMEM */
5699
5700#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5701
5702/* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5703void __paginginit set_pageblock_order(void)
5704{
5705        unsigned int order;
5706
5707        /* Check that pageblock_nr_pages has not already been setup */
5708        if (pageblock_order)
5709                return;
5710
5711        if (HPAGE_SHIFT > PAGE_SHIFT)
5712                order = HUGETLB_PAGE_ORDER;
5713        else
5714                order = MAX_ORDER - 1;
5715
5716        /*
5717         * Assume the largest contiguous order of interest is a huge page.
5718         * This value may be variable depending on boot parameters on IA64 and
5719         * powerpc.
5720         */
5721        pageblock_order = order;
5722}
5723#else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5724
5725/*
5726 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5727 * is unused as pageblock_order is set at compile-time. See
5728 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5729 * the kernel config
5730 */
5731void __paginginit set_pageblock_order(void)
5732{
5733}
5734
5735#endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5736
5737static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
5738                                                   unsigned long present_pages)
5739{
5740        unsigned long pages = spanned_pages;
5741
5742        /*
5743         * Provide a more accurate estimation if there are holes within
5744         * the zone and SPARSEMEM is in use. If there are holes within the
5745         * zone, each populated memory region may cost us one or two extra
5746         * memmap pages due to alignment because memmap pages for each
5747         * populated regions may not naturally algined on page boundary.
5748         * So the (present_pages >> 4) heuristic is a tradeoff for that.
5749         */
5750        if (spanned_pages > present_pages + (present_pages >> 4) &&
5751            IS_ENABLED(CONFIG_SPARSEMEM))
5752                pages = present_pages;
5753
5754        return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
5755}
5756
5757/*
5758 * Set up the zone data structures:
5759 *   - mark all pages reserved
5760 *   - mark all memory queues empty
5761 *   - clear the memory bitmaps
5762 *
5763 * NOTE: pgdat should get zeroed by caller.
5764 */
5765static void __paginginit free_area_init_core(struct pglist_data *pgdat)
5766{
5767        enum zone_type j;
5768        int nid = pgdat->node_id;
5769        int ret;
5770
5771        pgdat_resize_init(pgdat);
5772#ifdef CONFIG_NUMA_BALANCING
5773        spin_lock_init(&pgdat->numabalancing_migrate_lock);
5774        pgdat->numabalancing_migrate_nr_pages = 0;
5775        pgdat->numabalancing_migrate_next_window = jiffies;
5776#endif
5777#ifdef CONFIG_TRANSPARENT_HUGEPAGE
5778        spin_lock_init(&pgdat->split_queue_lock);
5779        INIT_LIST_HEAD(&pgdat->split_queue);
5780        pgdat->split_queue_len = 0;
5781#endif
5782        init_waitqueue_head(&pgdat->kswapd_wait);
5783        init_waitqueue_head(&pgdat->pfmemalloc_wait);
5784#ifdef CONFIG_COMPACTION
5785        init_waitqueue_head(&pgdat->kcompactd_wait);
5786#endif
5787        pgdat_page_ext_init(pgdat);
5788        spin_lock_init(&pgdat->lru_lock);
5789        lruvec_init(node_lruvec(pgdat));
5790
5791        for (j = 0; j < MAX_NR_ZONES; j++) {
5792                struct zone *zone = pgdat->node_zones + j;
5793                unsigned long size, realsize, freesize, memmap_pages;
5794                unsigned long zone_start_pfn = zone->zone_start_pfn;
5795
5796                size = zone->spanned_pages;
5797                realsize = freesize = zone->present_pages;
5798
5799                /*
5800                 * Adjust freesize so that it accounts for how much memory
5801                 * is used by this zone for memmap. This affects the watermark
5802                 * and per-cpu initialisations
5803                 */
5804                memmap_pages = calc_memmap_size(size, realsize);
5805                if (!is_highmem_idx(j)) {
5806                        if (freesize >= memmap_pages) {
5807                                freesize -= memmap_pages;
5808                                if (memmap_pages)
5809                                        printk(KERN_DEBUG
5810                                               "  %s zone: %lu pages used for memmap\n",
5811                                               zone_names[j], memmap_pages);
5812                        } else
5813                                pr_warn("  %s zone: %lu pages exceeds freesize %lu\n",
5814                                        zone_names[j], memmap_pages, freesize);
5815                }
5816
5817                /* Account for reserved pages */
5818                if (j == 0 && freesize > dma_reserve) {
5819                        freesize -= dma_reserve;
5820                        printk(KERN_DEBUG "  %s zone: %lu pages reserved\n",
5821                                        zone_names[0], dma_reserve);
5822                }
5823
5824                if (!is_highmem_idx(j))
5825                        nr_kernel_pages += freesize;
5826                /* Charge for highmem memmap if there are enough kernel pages */
5827                else if (nr_kernel_pages > memmap_pages * 2)
5828                        nr_kernel_pages -= memmap_pages;
5829                nr_all_pages += freesize;
5830
5831                /*
5832                 * Set an approximate value for lowmem here, it will be adjusted
5833                 * when the bootmem allocator frees pages into the buddy system.
5834                 * And all highmem pages will be managed by the buddy system.
5835                 */
5836                zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
5837#ifdef CONFIG_NUMA
5838                zone->node = nid;
5839#endif
5840                zone->name = zone_names[j];
5841                zone->zone_pgdat = pgdat;
5842                spin_lock_init(&zone->lock);
5843                zone_seqlock_init(zone);
5844                zone_pcp_init(zone);
5845
5846                if (!size)
5847                        continue;
5848
5849                set_pageblock_order();
5850                setup_usemap(pgdat, zone, zone_start_pfn, size);
5851                ret = init_currently_empty_zone(zone, zone_start_pfn, size);
5852                BUG_ON(ret);
5853                memmap_init(size, nid, j, zone_start_pfn);
5854        }
5855}
5856
5857static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
5858{
5859        unsigned long __maybe_unused start = 0;
5860        unsigned long __maybe_unused offset = 0;
5861
5862        /* Skip empty nodes */
5863        if (!pgdat->node_spanned_pages)
5864                return;
5865
5866#ifdef CONFIG_FLAT_NODE_MEM_MAP
5867        start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
5868        offset = pgdat->node_start_pfn - start;
5869        /* ia64 gets its own node_mem_map, before this, without bootmem */
5870        if (!pgdat->node_mem_map) {
5871                unsigned long size, end;
5872                struct page *map;
5873
5874                /*
5875                 * The zone's endpoints aren't required to be MAX_ORDER
5876                 * aligned but the node_mem_map endpoints must be in order
5877                 * for the buddy allocator to function correctly.
5878                 */
5879                end = pgdat_end_pfn(pgdat);
5880                end = ALIGN(end, MAX_ORDER_NR_PAGES);
5881                size =  (end - start) * sizeof(struct page);
5882                map = alloc_remap(pgdat->node_id, size);
5883                if (!map)
5884                        map = memblock_virt_alloc_node_nopanic(size,
5885                                                               pgdat->node_id);
5886                pgdat->node_mem_map = map + offset;
5887        }
5888#ifndef CONFIG_NEED_MULTIPLE_NODES
5889        /*
5890         * With no DISCONTIG, the global mem_map is just set as node 0's
5891         */
5892        if (pgdat == NODE_DATA(0)) {
5893                mem_map = NODE_DATA(0)->node_mem_map;
5894#if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
5895                if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
5896                        mem_map -= offset;
5897#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5898        }
5899#endif
5900#endif /* CONFIG_FLAT_NODE_MEM_MAP */
5901}
5902
5903void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
5904                unsigned long node_start_pfn, unsigned long *zholes_size)
5905{
5906        pg_data_t *pgdat = NODE_DATA(nid);
5907        unsigned long start_pfn = 0;
5908        unsigned long end_pfn = 0;
5909
5910        /* pg_data_t should be reset to zero when it's allocated */
5911        WARN_ON(pgdat->nr_zones || pgdat->kswapd_classzone_idx);
5912
5913        reset_deferred_meminit(pgdat);
5914        pgdat->node_id = nid;
5915        pgdat->node_start_pfn = node_start_pfn;
5916        pgdat->per_cpu_nodestats = NULL;
5917#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5918        get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
5919        pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
5920                (u64)start_pfn << PAGE_SHIFT,
5921                end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
5922#else
5923        start_pfn = node_start_pfn;
5924#endif
5925        calculate_node_totalpages(pgdat, start_pfn, end_pfn,
5926                                  zones_size, zholes_size);
5927
5928        alloc_node_mem_map(pgdat);
5929#ifdef CONFIG_FLAT_NODE_MEM_MAP
5930        printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5931                nid, (unsigned long)pgdat,
5932                (unsigned long)pgdat->node_mem_map);
5933#endif
5934
5935        free_area_init_core(pgdat);
5936}
5937
5938#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5939
5940#if MAX_NUMNODES > 1
5941/*
5942 * Figure out the number of possible node ids.
5943 */
5944void __init setup_nr_node_ids(void)
5945{
5946        unsigned int highest;
5947
5948        highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
5949        nr_node_ids = highest + 1;
5950}
5951#endif
5952
5953/**
5954 * node_map_pfn_alignment - determine the maximum internode alignment
5955 *
5956 * This function should be called after node map is populated and sorted.
5957 * It calculates the maximum power of two alignment which can distinguish
5958 * all the nodes.
5959 *
5960 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5961 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)).  If the
5962 * nodes are shifted by 256MiB, 256MiB.  Note that if only the last node is
5963 * shifted, 1GiB is enough and this function will indicate so.
5964 *
5965 * This is used to test whether pfn -> nid mapping of the chosen memory
5966 * model has fine enough granularity to avoid incorrect mapping for the
5967 * populated node map.
5968 *
5969 * Returns the determined alignment in pfn's.  0 if there is no alignment
5970 * requirement (single node).
5971 */
5972unsigned long __init node_map_pfn_alignment(void)
5973{
5974        unsigned long accl_mask = 0, last_end = 0;
5975        unsigned long start, end, mask;
5976        int last_nid = -1;
5977        int i, nid;
5978
5979        for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
5980                if (!start || last_nid < 0 || last_nid == nid) {
5981                        last_nid = nid;
5982                        last_end = end;
5983                        continue;
5984                }
5985
5986                /*
5987                 * Start with a mask granular enough to pin-point to the
5988                 * start pfn and tick off bits one-by-one until it becomes
5989                 * too coarse to separate the current node from the last.
5990                 */
5991                mask = ~((1 << __ffs(start)) - 1);
5992                while (mask && last_end <= (start & (mask << 1)))
5993                        mask <<= 1;
5994
5995                /* accumulate all internode masks */
5996                accl_mask |= mask;
5997        }
5998
5999        /* convert mask to number of pages */
6000        return ~accl_mask + 1;
6001}
6002
6003/* Find the lowest pfn for a node */
6004static unsigned long __init find_min_pfn_for_node(int nid)
6005{
6006        unsigned long min_pfn = ULONG_MAX;
6007        unsigned long start_pfn;
6008        int i;
6009
6010        for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
6011                min_pfn = min(min_pfn, start_pfn);
6012
6013        if (min_pfn == ULONG_MAX) {
6014                pr_warn("Could not find start_pfn for node %d\n", nid);
6015                return 0;
6016        }
6017
6018        return min_pfn;
6019}
6020
6021/**
6022 * find_min_pfn_with_active_regions - Find the minimum PFN registered
6023 *
6024 * It returns the minimum PFN based on information provided via
6025 * memblock_set_node().
6026 */
6027unsigned long __init find_min_pfn_with_active_regions(void)
6028{
6029        return find_min_pfn_for_node(MAX_NUMNODES);
6030}
6031
6032/*
6033 * early_calculate_totalpages()
6034 * Sum pages in active regions for movable zone.
6035 * Populate N_MEMORY for calculating usable_nodes.
6036 */
6037static unsigned long __init early_calculate_totalpages(void)
6038{
6039        unsigned long totalpages = 0;
6040        unsigned long start_pfn, end_pfn;
6041        int i, nid;
6042
6043        for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6044                unsigned long pages = end_pfn - start_pfn;
6045
6046                totalpages += pages;
6047                if (pages)
6048                        node_set_state(nid, N_MEMORY);
6049        }
6050        return totalpages;
6051}
6052
6053/*
6054 * Find the PFN the Movable zone begins in each node. Kernel memory
6055 * is spread evenly between nodes as long as the nodes have enough
6056 * memory. When they don't, some nodes will have more kernelcore than
6057 * others
6058 */
6059static void __init find_zone_movable_pfns_for_nodes(void)
6060{
6061        int i, nid;
6062        unsigned long usable_startpfn;
6063        unsigned long kernelcore_node, kernelcore_remaining;
6064        /* save the state before borrow the nodemask */
6065        nodemask_t saved_node_state = node_states[N_MEMORY];
6066        unsigned long totalpages = early_calculate_totalpages();
6067        int usable_nodes = nodes_weight(node_states[N_MEMORY]);
6068        struct memblock_region *r;
6069
6070        /* Need to find movable_zone earlier when movable_node is specified. */
6071        find_usable_zone_for_movable();
6072
6073        /*
6074         * If movable_node is specified, ignore kernelcore and movablecore
6075         * options.
6076         */
6077        if (movable_node_is_enabled()) {
6078                for_each_memblock(memory, r) {
6079                        if (!memblock_is_hotpluggable(r))
6080                                continue;
6081
6082                        nid = r->nid;
6083
6084                        usable_startpfn = PFN_DOWN(r->base);
6085                        zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6086                                min(usable_startpfn, zone_movable_pfn[nid]) :
6087                                usable_startpfn;
6088                }
6089
6090                goto out2;
6091        }
6092
6093        /*
6094         * If kernelcore=mirror is specified, ignore movablecore option
6095         */
6096        if (mirrored_kernelcore) {
6097                bool mem_below_4gb_not_mirrored = false;
6098
6099                for_each_memblock(memory, r) {
6100                        if (memblock_is_mirror(r))
6101                                continue;
6102
6103                        nid = r->nid;
6104
6105                        usable_startpfn = memblock_region_memory_base_pfn(r);
6106
6107                        if (usable_startpfn < 0x100000) {
6108                                mem_below_4gb_not_mirrored = true;
6109                                continue;
6110                        }
6111
6112                        zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6113                                min(usable_startpfn, zone_movable_pfn[nid]) :
6114                                usable_startpfn;
6115                }
6116
6117                if (mem_below_4gb_not_mirrored)
6118                        pr_warn("This configuration results in unmirrored kernel memory.");
6119
6120                goto out2;
6121        }
6122
6123        /*
6124         * If movablecore=nn[KMG] was specified, calculate what size of
6125         * kernelcore that corresponds so that memory usable for
6126         * any allocation type is evenly spread. If both kernelcore
6127         * and movablecore are specified, then the value of kernelcore
6128         * will be used for required_kernelcore if it's greater than
6129         * what movablecore would have allowed.
6130         */
6131        if (required_movablecore) {
6132                unsigned long corepages;
6133
6134                /*
6135                 * Round-up so that ZONE_MOVABLE is at least as large as what
6136                 * was requested by the user
6137                 */
6138                required_movablecore =
6139                        roundup(required_movablecore, MAX_ORDER_NR_PAGES);
6140                required_movablecore = min(totalpages, required_movablecore);
6141                corepages = totalpages - required_movablecore;
6142
6143                required_kernelcore = max(required_kernelcore, corepages);
6144        }
6145
6146        /*
6147         * If kernelcore was not specified or kernelcore size is larger
6148         * than totalpages, there is no ZONE_MOVABLE.
6149         */
6150        if (!required_kernelcore || required_kernelcore >= totalpages)
6151                goto out;
6152
6153        /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
6154        usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
6155
6156restart:
6157        /* Spread kernelcore memory as evenly as possible throughout nodes */
6158        kernelcore_node = required_kernelcore / usable_nodes;
6159        for_each_node_state(nid, N_MEMORY) {
6160                unsigned long start_pfn, end_pfn;
6161
6162                /*
6163                 * Recalculate kernelcore_node if the division per node
6164                 * now exceeds what is necessary to satisfy the requested
6165                 * amount of memory for the kernel
6166                 */
6167                if (required_kernelcore < kernelcore_node)
6168                        kernelcore_node = required_kernelcore / usable_nodes;
6169
6170                /*
6171                 * As the map is walked, we track how much memory is usable
6172                 * by the kernel using kernelcore_remaining. When it is
6173                 * 0, the rest of the node is usable by ZONE_MOVABLE
6174                 */
6175                kernelcore_remaining = kernelcore_node;
6176
6177                /* Go through each range of PFNs within this node */
6178                for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6179                        unsigned long size_pages;
6180
6181                        start_pfn = max(start_pfn, zone_movable_pfn[nid]);
6182                        if (start_pfn >= end_pfn)
6183                                continue;
6184
6185                        /* Account for what is only usable for kernelcore */
6186                        if (start_pfn < usable_startpfn) {
6187                                unsigned long kernel_pages;
6188                                kernel_pages = min(end_pfn, usable_startpfn)
6189                                                                - start_pfn;
6190
6191                                kernelcore_remaining -= min(kernel_pages,
6192                                                        kernelcore_remaining);
6193                                required_kernelcore -= min(kernel_pages,
6194                                                        required_kernelcore);
6195
6196                                /* Continue if range is now fully accounted */
6197                                if (end_pfn <= usable_startpfn) {
6198
6199                                        /*
6200                                         * Push zone_movable_pfn to the end so
6201                                         * that if we have to rebalance
6202                                         * kernelcore across nodes, we will
6203                                         * not double account here
6204                                         */
6205                                        zone_movable_pfn[nid] = end_pfn;
6206                                        continue;
6207                                }
6208                                start_pfn = usable_startpfn;
6209                        }
6210
6211                        /*
6212                         * The usable PFN range for ZONE_MOVABLE is from
6213                         * start_pfn->end_pfn. Calculate size_pages as the
6214                         * number of pages used as kernelcore
6215                         */
6216                        size_pages = end_pfn - start_pfn;
6217                        if (size_pages > kernelcore_remaining)
6218                                size_pages = kernelcore_remaining;
6219                        zone_movable_pfn[nid] = start_pfn + size_pages;
6220
6221                        /*
6222                         * Some kernelcore has been met, update counts and
6223                         * break if the kernelcore for this node has been
6224                         * satisfied
6225                         */
6226                        required_kernelcore -= min(required_kernelcore,
6227                                                                size_pages);
6228                        kernelcore_remaining -= size_pages;
6229                        if (!kernelcore_remaining)
6230                                break;
6231                }
6232        }
6233
6234        /*
6235         * If there is still required_kernelcore, we do another pass with one
6236         * less node in the count. This will push zone_movable_pfn[nid] further
6237         * along on the nodes that still have memory until kernelcore is
6238         * satisfied
6239         */
6240        usable_nodes--;
6241        if (usable_nodes && required_kernelcore > usable_nodes)
6242                goto restart;
6243
6244out2:
6245        /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
6246        for (nid = 0; nid < MAX_NUMNODES; nid++)
6247                zone_movable_pfn[nid] =
6248                        roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
6249
6250out:
6251        /* restore the node_state */
6252        node_states[N_MEMORY] = saved_node_state;
6253}
6254
6255/* Any regular or high memory on that node ? */
6256static void check_for_memory(pg_data_t *pgdat, int nid)
6257{
6258        enum zone_type zone_type;
6259
6260        if (N_MEMORY == N_NORMAL_MEMORY)
6261                return;
6262
6263        for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
6264                struct zone *zone = &pgdat->node_zones[zone_type];
6265                if (populated_zone(zone)) {
6266                        node_set_state(nid, N_HIGH_MEMORY);
6267                        if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
6268                            zone_type <= ZONE_NORMAL)
6269                                node_set_state(nid, N_NORMAL_MEMORY);
6270                        break;
6271                }
6272        }
6273}
6274
6275/**
6276 * free_area_init_nodes - Initialise all pg_data_t and zone data
6277 * @max_zone_pfn: an array of max PFNs for each zone
6278 *
6279 * This will call free_area_init_node() for each active node in the system.
6280 * Using the page ranges provided by memblock_set_node(), the size of each
6281 * zone in each node and their holes is calculated. If the maximum PFN
6282 * between two adjacent zones match, it is assumed that the zone is empty.
6283 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
6284 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
6285 * starts where the previous one ended. For example, ZONE_DMA32 starts
6286 * at arch_max_dma_pfn.
6287 */
6288void __init free_area_init_nodes(unsigned long *max_zone_pfn)
6289{
6290        unsigned long start_pfn, end_pfn;
6291        int i, nid;
6292
6293        /* Record where the zone boundaries are */
6294        memset(arch_zone_lowest_possible_pfn, 0,
6295                                sizeof(arch_zone_lowest_possible_pfn));
6296        memset(arch_zone_highest_possible_pfn, 0,
6297                                sizeof(arch_zone_highest_possible_pfn));
6298
6299        start_pfn = find_min_pfn_with_active_regions();
6300
6301        for (i = 0; i < MAX_NR_ZONES; i++) {
6302                if (i == ZONE_MOVABLE)
6303                        continue;
6304
6305                end_pfn = max(max_zone_pfn[i], start_pfn);
6306                arch_zone_lowest_possible_pfn[i] = start_pfn;
6307                arch_zone_highest_possible_pfn[i] = end_pfn;
6308
6309                start_pfn = end_pfn;
6310        }
6311        arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
6312        arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
6313
6314        /* Find the PFNs that ZONE_MOVABLE begins at in each node */
6315        memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
6316        find_zone_movable_pfns_for_nodes();
6317
6318        /* Print out the zone ranges */
6319        pr_info("Zone ranges:\n");
6320        for (i = 0; i < MAX_NR_ZONES; i++) {
6321                if (i == ZONE_MOVABLE)
6322                        continue;
6323                pr_info("  %-8s ", zone_names[i]);
6324                if (arch_zone_lowest_possible_pfn[i] ==
6325                                arch_zone_highest_possible_pfn[i])
6326                        pr_cont("empty\n");
6327                else
6328                        pr_cont("[mem %#018Lx-%#018Lx]\n",
6329                                (u64)arch_zone_lowest_possible_pfn[i]
6330                                        << PAGE_SHIFT,
6331                                ((u64)arch_zone_highest_possible_pfn[i]
6332                                        << PAGE_SHIFT) - 1);
6333        }
6334
6335        /* Print out the PFNs ZONE_MOVABLE begins at in each node */
6336        pr_info("Movable zone start for each node\n");
6337        for (i = 0; i < MAX_NUMNODES; i++) {
6338                if (zone_movable_pfn[i])
6339                        pr_info("  Node %d: %#018Lx\n", i,
6340                               (u64)zone_movable_pfn[i] << PAGE_SHIFT);
6341        }
6342
6343        /* Print out the early node map */
6344        pr_info("Early memory node ranges\n");
6345        for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
6346                pr_info("  node %3d: [mem %#018Lx-%#018Lx]\n", nid,
6347                        (u64)start_pfn << PAGE_SHIFT,
6348                        ((u64)end_pfn << PAGE_SHIFT) - 1);
6349
6350        /* Initialise every node */
6351        mminit_verify_pageflags_layout();
6352        setup_nr_node_ids();
6353        for_each_online_node(nid) {
6354                pg_data_t *pgdat = NODE_DATA(nid);
6355                free_area_init_node(nid, NULL,
6356                                find_min_pfn_for_node(nid), NULL);
6357
6358                /* Any memory on that node */
6359                if (pgdat->node_present_pages)
6360                        node_set_state(nid, N_MEMORY);
6361                check_for_memory(pgdat, nid);
6362        }
6363}
6364
6365static int __init cmdline_parse_core(char *p, unsigned long *core)
6366{
6367        unsigned long long coremem;
6368        if (!p)
6369                return -EINVAL;
6370
6371        coremem = memparse(p, &p);
6372        *core = coremem >> PAGE_SHIFT;
6373
6374        /* Paranoid check that UL is enough for the coremem value */
6375        WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
6376
6377        return 0;
6378}
6379
6380/*
6381 * kernelcore=size sets the amount of memory for use for allocations that
6382 * cannot be reclaimed or migrated.
6383 */
6384static int __init cmdline_parse_kernelcore(char *p)
6385{
6386        /* parse kernelcore=mirror */
6387        if (parse_option_str(p, "mirror")) {
6388                mirrored_kernelcore = true;
6389                return 0;
6390        }
6391
6392        return cmdline_parse_core(p, &required_kernelcore);
6393}
6394
6395/*
6396 * movablecore=size sets the amount of memory for use for allocations that
6397 * can be reclaimed or migrated.
6398 */
6399static int __init cmdline_parse_movablecore(char *p)
6400{
6401        return cmdline_parse_core(p, &required_movablecore);
6402}
6403
6404early_param("kernelcore", cmdline_parse_kernelcore);
6405early_param("movablecore", cmdline_parse_movablecore);
6406
6407#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6408
6409void adjust_managed_page_count(struct page *page, long count)
6410{
6411        spin_lock(&managed_page_count_lock);
6412        page_zone(page)->managed_pages += count;
6413        totalram_pages += count;
6414#ifdef CONFIG_HIGHMEM
6415        if (PageHighMem(page))
6416                totalhigh_pages += count;
6417#endif
6418        spin_unlock(&managed_page_count_lock);
6419}
6420EXPORT_SYMBOL(adjust_managed_page_count);
6421
6422unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
6423{
6424        void *pos;
6425        unsigned long pages = 0;
6426
6427        start = (void *)PAGE_ALIGN((unsigned long)start);
6428        end = (void *)((unsigned long)end & PAGE_MASK);
6429        for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
6430                if ((unsigned int)poison <= 0xFF)
6431                        memset(pos, poison, PAGE_SIZE);
6432                free_reserved_page(virt_to_page(pos));
6433        }
6434
6435        if (pages && s)
6436                pr_info("Freeing %s memory: %ldK (%p - %p)\n",
6437                        s, pages << (PAGE_SHIFT - 10), start, end);
6438
6439        return pages;
6440}
6441EXPORT_SYMBOL(free_reserved_area);
6442
6443#ifdef  CONFIG_HIGHMEM
6444void free_highmem_page(struct page *page)
6445{
6446        __free_reserved_page(page);
6447        totalram_pages++;
6448        page_zone(page)->managed_pages++;
6449        totalhigh_pages++;
6450}
6451#endif
6452
6453
6454void __init mem_init_print_info(const char *str)
6455{
6456        unsigned long physpages, codesize, datasize, rosize, bss_size;
6457        unsigned long init_code_size, init_data_size;
6458
6459        physpages = get_num_physpages();
6460        codesize = _etext - _stext;
6461        datasize = _edata - _sdata;
6462        rosize = __end_rodata - __start_rodata;
6463        bss_size = __bss_stop - __bss_start;
6464        init_data_size = __init_end - __init_begin;
6465        init_code_size = _einittext - _sinittext;
6466
6467        /*
6468         * Detect special cases and adjust section sizes accordingly:
6469         * 1) .init.* may be embedded into .data sections
6470         * 2) .init.text.* may be out of [__init_begin, __init_end],
6471         *    please refer to arch/tile/kernel/vmlinux.lds.S.
6472         * 3) .rodata.* may be embedded into .text or .data sections.
6473         */
6474#define adj_init_size(start, end, size, pos, adj) \
6475        do { \
6476                if (start <= pos && pos < end && size > adj) \
6477                        size -= adj; \
6478        } while (0)
6479
6480        adj_init_size(__init_begin, __init_end, init_data_size,
6481                     _sinittext, init_code_size);
6482        adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
6483        adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
6484        adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
6485        adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
6486
6487#undef  adj_init_size
6488
6489        pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
6490#ifdef  CONFIG_HIGHMEM
6491                ", %luK highmem"
6492#endif
6493                "%s%s)\n",
6494                nr_free_pages() << (PAGE_SHIFT - 10),
6495                physpages << (PAGE_SHIFT - 10),
6496                codesize >> 10, datasize >> 10, rosize >> 10,
6497                (init_data_size + init_code_size) >> 10, bss_size >> 10,
6498                (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT - 10),
6499                totalcma_pages << (PAGE_SHIFT - 10),
6500#ifdef  CONFIG_HIGHMEM
6501                totalhigh_pages << (PAGE_SHIFT - 10),
6502#endif
6503                str ? ", " : "", str ? str : "");
6504}
6505
6506/**
6507 * set_dma_reserve - set the specified number of pages reserved in the first zone
6508 * @new_dma_reserve: The number of pages to mark reserved
6509 *
6510 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
6511 * In the DMA zone, a significant percentage may be consumed by kernel image
6512 * and other unfreeable allocations which can skew the watermarks badly. This
6513 * function may optionally be used to account for unfreeable pages in the
6514 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
6515 * smaller per-cpu batchsize.
6516 */
6517void __init set_dma_reserve(unsigned long new_dma_reserve)
6518{
6519        dma_reserve = new_dma_reserve;
6520}
6521
6522void __init free_area_init(unsigned long *zones_size)
6523{
6524        free_area_init_node(0, zones_size,
6525                        __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
6526}
6527
6528static int page_alloc_cpu_notify(struct notifier_block *self,
6529                                 unsigned long action, void *hcpu)
6530{
6531        int cpu = (unsigned long)hcpu;
6532
6533        if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
6534                lru_add_drain_cpu(cpu);
6535                drain_pages(cpu);
6536
6537                /*
6538                 * Spill the event counters of the dead processor
6539                 * into the current processors event counters.
6540                 * This artificially elevates the count of the current
6541                 * processor.
6542                 */
6543                vm_events_fold_cpu(cpu);
6544
6545                /*
6546                 * Zero the differential counters of the dead processor
6547                 * so that the vm statistics are consistent.
6548                 *
6549                 * This is only okay since the processor is dead and cannot
6550                 * race with what we are doing.
6551                 */
6552                cpu_vm_stats_fold(cpu);
6553        }
6554        return NOTIFY_OK;
6555}
6556
6557void __init page_alloc_init(void)
6558{
6559        hotcpu_notifier(page_alloc_cpu_notify, 0);
6560}
6561
6562/*
6563 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6564 *      or min_free_kbytes changes.
6565 */
6566static void calculate_totalreserve_pages(void)
6567{
6568        struct pglist_data *pgdat;
6569        unsigned long reserve_pages = 0;
6570        enum zone_type i, j;
6571
6572        for_each_online_pgdat(pgdat) {
6573
6574                pgdat->totalreserve_pages = 0;
6575
6576                for (i = 0; i < MAX_NR_ZONES; i++) {
6577                        struct zone *zone = pgdat->node_zones + i;
6578                        long max = 0;
6579
6580                        /* Find valid and maximum lowmem_reserve in the zone */
6581                        for (j = i; j < MAX_NR_ZONES; j++) {
6582                                if (zone->lowmem_reserve[j] > max)
6583                                        max = zone->lowmem_reserve[j];
6584                        }
6585
6586                        /* we treat the high watermark as reserved pages. */
6587                        max += high_wmark_pages(zone);
6588
6589                        if (max > zone->managed_pages)
6590                                max = zone->managed_pages;
6591
6592                        pgdat->totalreserve_pages += max;
6593
6594                        reserve_pages += max;
6595                }
6596        }
6597        totalreserve_pages = reserve_pages;
6598}
6599
6600/*
6601 * setup_per_zone_lowmem_reserve - called whenever
6602 *      sysctl_lowmem_reserve_ratio changes.  Ensures that each zone
6603 *      has a correct pages reserved value, so an adequate number of
6604 *      pages are left in the zone after a successful __alloc_pages().
6605 */
6606static void setup_per_zone_lowmem_reserve(void)
6607{
6608        struct pglist_data *pgdat;
6609        enum zone_type j, idx;
6610
6611        for_each_online_pgdat(pgdat) {
6612                for (j = 0; j < MAX_NR_ZONES; j++) {
6613                        struct zone *zone = pgdat->node_zones + j;
6614                        unsigned long managed_pages = zone->managed_pages;
6615
6616                        zone->lowmem_reserve[j] = 0;
6617
6618                        idx = j;
6619                        while (idx) {
6620                                struct zone *lower_zone;
6621
6622                                idx--;
6623
6624                                if (sysctl_lowmem_reserve_ratio[idx] < 1)
6625                                        sysctl_lowmem_reserve_ratio[idx] = 1;
6626
6627                                lower_zone = pgdat->node_zones + idx;
6628                                lower_zone->lowmem_reserve[j] = managed_pages /
6629                                        sysctl_lowmem_reserve_ratio[idx];
6630                                managed_pages += lower_zone->managed_pages;
6631                        }
6632                }
6633        }
6634
6635        /* update totalreserve_pages */
6636        calculate_totalreserve_pages();
6637}
6638
6639static void __setup_per_zone_wmarks(void)
6640{
6641        unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6642        unsigned long lowmem_pages = 0;
6643        struct zone *zone;
6644        unsigned long flags;
6645
6646        /* Calculate total number of !ZONE_HIGHMEM pages */
6647        for_each_zone(zone) {
6648                if (!is_highmem(zone))
6649                        lowmem_pages += zone->managed_pages;
6650        }
6651
6652        for_each_zone(zone) {
6653                u64 tmp;
6654
6655                spin_lock_irqsave(&zone->lock, flags);
6656                tmp = (u64)pages_min * zone->managed_pages;
6657                do_div(tmp, lowmem_pages);
6658                if (is_highmem(zone)) {
6659                        /*
6660                         * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6661                         * need highmem pages, so cap pages_min to a small
6662                         * value here.
6663                         *
6664                         * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6665                         * deltas control asynch page reclaim, and so should
6666                         * not be capped for highmem.
6667                         */
6668                        unsigned long min_pages;
6669
6670                        min_pages = zone->managed_pages / 1024;
6671                        min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
6672                        zone->watermark[WMARK_MIN] = min_pages;
6673                } else {
6674                        /*
6675                         * If it's a lowmem zone, reserve a number of pages
6676                         * proportionate to the zone's size.
6677                         */
6678                        zone->watermark[WMARK_MIN] = tmp;
6679                }
6680
6681                /*
6682                 * Set the kswapd watermarks distance according to the
6683                 * scale factor in proportion to available memory, but
6684                 * ensure a minimum size on small systems.
6685                 */
6686                tmp = max_t(u64, tmp >> 2,
6687                            mult_frac(zone->managed_pages,
6688                                      watermark_scale_factor, 10000));
6689
6690                zone->watermark[WMARK_LOW]  = min_wmark_pages(zone) + tmp;
6691                zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
6692
6693                spin_unlock_irqrestore(&zone->lock, flags);
6694        }
6695
6696        /* update totalreserve_pages */
6697        calculate_totalreserve_pages();
6698}
6699
6700/**
6701 * setup_per_zone_wmarks - called when min_free_kbytes changes
6702 * or when memory is hot-{added|removed}
6703 *
6704 * Ensures that the watermark[min,low,high] values for each zone are set
6705 * correctly with respect to min_free_kbytes.
6706 */
6707void setup_per_zone_wmarks(void)
6708{
6709        mutex_lock(&zonelists_mutex);
6710        __setup_per_zone_wmarks();
6711        mutex_unlock(&zonelists_mutex);
6712}
6713
6714/*
6715 * Initialise min_free_kbytes.
6716 *
6717 * For small machines we want it small (128k min).  For large machines
6718 * we want it large (64MB max).  But it is not linear, because network
6719 * bandwidth does not increase linearly with machine size.  We use
6720 *
6721 *      min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6722 *      min_free_kbytes = sqrt(lowmem_kbytes * 16)
6723 *
6724 * which yields
6725 *
6726 * 16MB:        512k
6727 * 32MB:        724k
6728 * 64MB:        1024k
6729 * 128MB:       1448k
6730 * 256MB:       2048k
6731 * 512MB:       2896k
6732 * 1024MB:      4096k
6733 * 2048MB:      5792k
6734 * 4096MB:      8192k
6735 * 8192MB:      11584k
6736 * 16384MB:     16384k
6737 */
6738int __meminit init_per_zone_wmark_min(void)
6739{
6740        unsigned long lowmem_kbytes;
6741        int new_min_free_kbytes;
6742
6743        lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
6744        new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
6745
6746        if (new_min_free_kbytes > user_min_free_kbytes) {
6747                min_free_kbytes = new_min_free_kbytes;
6748                if (min_free_kbytes < 128)
6749                        min_free_kbytes = 128;
6750                if (min_free_kbytes > 65536)
6751                        min_free_kbytes = 65536;
6752        } else {
6753                pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6754                                new_min_free_kbytes, user_min_free_kbytes);
6755        }
6756        setup_per_zone_wmarks();
6757        refresh_zone_stat_thresholds();
6758        setup_per_zone_lowmem_reserve();
6759
6760#ifdef CONFIG_NUMA
6761        setup_min_unmapped_ratio();
6762        setup_min_slab_ratio();
6763#endif
6764
6765        return 0;
6766}
6767core_initcall(init_per_zone_wmark_min)
6768
6769/*
6770 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6771 *      that we can call two helper functions whenever min_free_kbytes
6772 *      changes.
6773 */
6774int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
6775        void __user *buffer, size_t *length, loff_t *ppos)
6776{
6777        int rc;
6778
6779        rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6780        if (rc)
6781                return rc;
6782
6783        if (write) {
6784                user_min_free_kbytes = min_free_kbytes;
6785                setup_per_zone_wmarks();
6786        }
6787        return 0;
6788}
6789
6790int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
6791        void __user *buffer, size_t *length, loff_t *ppos)
6792{
6793        int rc;
6794
6795        rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6796        if (rc)
6797                return rc;
6798
6799        if (write)
6800                setup_per_zone_wmarks();
6801
6802        return 0;
6803}
6804
6805#ifdef CONFIG_NUMA
6806static void setup_min_unmapped_ratio(void)
6807{
6808        pg_data_t *pgdat;
6809        struct zone *zone;
6810
6811        for_each_online_pgdat(pgdat)
6812                pgdat->min_unmapped_pages = 0;
6813
6814        for_each_zone(zone)
6815                zone->zone_pgdat->min_unmapped_pages += (zone->managed_pages *
6816                                sysctl_min_unmapped_ratio) / 100;
6817}
6818
6819
6820int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
6821        void __user *buffer, size_t *length, loff_t *ppos)
6822{
6823        int rc;
6824
6825        rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6826        if (rc)
6827                return rc;
6828
6829        setup_min_unmapped_ratio();
6830
6831        return 0;
6832}
6833
6834static void setup_min_slab_ratio(void)
6835{
6836        pg_data_t *pgdat;
6837        struct zone *zone;
6838
6839        for_each_online_pgdat(pgdat)
6840                pgdat->min_slab_pages = 0;
6841
6842        for_each_zone(zone)
6843                zone->zone_pgdat->min_slab_pages += (zone->managed_pages *
6844                                sysctl_min_slab_ratio) / 100;
6845}
6846
6847int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
6848        void __user *buffer, size_t *length, loff_t *ppos)
6849{
6850        int rc;
6851
6852        rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6853        if (rc)
6854                return rc;
6855
6856        setup_min_slab_ratio();
6857
6858        return 0;
6859}
6860#endif
6861
6862/*
6863 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6864 *      proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6865 *      whenever sysctl_lowmem_reserve_ratio changes.
6866 *
6867 * The reserve ratio obviously has absolutely no relation with the
6868 * minimum watermarks. The lowmem reserve ratio can only make sense
6869 * if in function of the boot time zone sizes.
6870 */
6871int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
6872        void __user *buffer, size_t *length, loff_t *ppos)
6873{
6874        proc_dointvec_minmax(table, write, buffer, length, ppos);
6875        setup_per_zone_lowmem_reserve();
6876        return 0;
6877}
6878
6879/*
6880 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6881 * cpu.  It is the fraction of total pages in each zone that a hot per cpu
6882 * pagelist can have before it gets flushed back to buddy allocator.
6883 */
6884int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
6885        void __user *buffer, size_t *length, loff_t *ppos)
6886{
6887        struct zone *zone;
6888        int old_percpu_pagelist_fraction;
6889        int ret;
6890
6891        mutex_lock(&pcp_batch_high_lock);
6892        old_percpu_pagelist_fraction = percpu_pagelist_fraction;
6893
6894        ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
6895        if (!write || ret < 0)
6896                goto out;
6897
6898        /* Sanity checking to avoid pcp imbalance */
6899        if (percpu_pagelist_fraction &&
6900            percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
6901                percpu_pagelist_fraction = old_percpu_pagelist_fraction;
6902                ret = -EINVAL;
6903                goto out;
6904        }
6905
6906        /* No change? */
6907        if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
6908                goto out;
6909
6910        for_each_populated_zone(zone) {
6911                unsigned int cpu;
6912
6913                for_each_possible_cpu(cpu)
6914                        pageset_set_high_and_batch(zone,
6915                                        per_cpu_ptr(zone->pageset, cpu));
6916        }
6917out:
6918        mutex_unlock(&pcp_batch_high_lock);
6919        return ret;
6920}
6921
6922#ifdef CONFIG_NUMA
6923int hashdist = HASHDIST_DEFAULT;
6924
6925static int __init set_hashdist(char *str)
6926{
6927        if (!str)
6928                return 0;
6929        hashdist = simple_strtoul(str, &str, 0);
6930        return 1;
6931}
6932__setup("hashdist=", set_hashdist);
6933#endif
6934
6935#ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
6936/*
6937 * Returns the number of pages that arch has reserved but
6938 * is not known to alloc_large_system_hash().
6939 */
6940static unsigned long __init arch_reserved_kernel_pages(void)
6941{
6942        return 0;
6943}
6944#endif
6945
6946/*
6947 * allocate a large system hash table from bootmem
6948 * - it is assumed that the hash table must contain an exact power-of-2
6949 *   quantity of entries
6950 * - limit is the number of hash buckets, not the total allocation size
6951 */
6952void *__init alloc_large_system_hash(const char *tablename,
6953                                     unsigned long bucketsize,
6954                                     unsigned long numentries,
6955                                     int scale,
6956                                     int flags,
6957                                     unsigned int *_hash_shift,
6958                                     unsigned int *_hash_mask,
6959                                     unsigned long low_limit,
6960                                     unsigned long high_limit)
6961{
6962        unsigned long long max = high_limit;
6963        unsigned long log2qty, size;
6964        void *table = NULL;
6965
6966        /* allow the kernel cmdline to have a say */
6967        if (!numentries) {
6968                /* round applicable memory size up to nearest megabyte */
6969                numentries = nr_kernel_pages;
6970                numentries -= arch_reserved_kernel_pages();
6971
6972                /* It isn't necessary when PAGE_SIZE >= 1MB */
6973                if (PAGE_SHIFT < 20)
6974                        numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
6975
6976                /* limit to 1 bucket per 2^scale bytes of low memory */
6977                if (scale > PAGE_SHIFT)
6978                        numentries >>= (scale - PAGE_SHIFT);
6979                else
6980                        numentries <<= (PAGE_SHIFT - scale);
6981
6982                /* Make sure we've got at least a 0-order allocation.. */
6983                if (unlikely(flags & HASH_SMALL)) {
6984                        /* Makes no sense without HASH_EARLY */
6985                        WARN_ON(!(flags & HASH_EARLY));
6986                        if (!(numentries >> *_hash_shift)) {
6987                                numentries = 1UL << *_hash_shift;
6988                                BUG_ON(!numentries);
6989                        }
6990                } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
6991                        numentries = PAGE_SIZE / bucketsize;
6992        }
6993        numentries = roundup_pow_of_two(numentries);
6994
6995        /* limit allocation size to 1/16 total memory by default */
6996        if (max == 0) {
6997                max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
6998                do_div(max, bucketsize);
6999        }
7000        max = min(max, 0x80000000ULL);
7001
7002        if (numentries < low_limit)
7003                numentries = low_limit;
7004        if (numentries > max)
7005                numentries = max;
7006
7007        log2qty = ilog2(numentries);
7008
7009        do {
7010                size = bucketsize << log2qty;
7011                if (flags & HASH_EARLY)
7012                        table = memblock_virt_alloc_nopanic(size, 0);
7013                else if (hashdist)
7014                        table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
7015                else {
7016                        /*
7017                         * If bucketsize is not a power-of-two, we may free
7018                         * some pages at the end of hash table which
7019                         * alloc_pages_exact() automatically does
7020                         */
7021                        if (get_order(size) < MAX_ORDER) {
7022                                table = alloc_pages_exact(size, GFP_ATOMIC);
7023                                kmemleak_alloc(table, size, 1, GFP_ATOMIC);
7024                        }
7025                }
7026        } while (!table && size > PAGE_SIZE && --log2qty);
7027
7028        if (!table)
7029                panic("Failed to allocate %s hash table\n", tablename);
7030
7031        pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
7032                tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size);
7033
7034        if (_hash_shift)
7035                *_hash_shift = log2qty;
7036        if (_hash_mask)
7037                *_hash_mask = (1 << log2qty) - 1;
7038
7039        return table;
7040}
7041
7042/*
7043 * This function checks whether pageblock includes unmovable pages or not.
7044 * If @count is not zero, it is okay to include less @count unmovable pages
7045 *
7046 * PageLRU check without isolation or lru_lock could race so that
7047 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
7048 * expect this function should be exact.
7049 */
7050bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
7051                         bool skip_hwpoisoned_pages)
7052{
7053        unsigned long pfn, iter, found;
7054        int mt;
7055
7056        /*
7057         * For avoiding noise data, lru_add_drain_all() should be called
7058         * If ZONE_MOVABLE, the zone never contains unmovable pages
7059         */
7060        if (zone_idx(zone) == ZONE_MOVABLE)
7061                return false;
7062        mt = get_pageblock_migratetype(page);
7063        if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
7064                return false;
7065
7066        pfn = page_to_pfn(page);
7067        for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
7068                unsigned long check = pfn + iter;
7069
7070                if (!pfn_valid_within(check))
7071                        continue;
7072
7073                page = pfn_to_page(check);
7074
7075                /*
7076                 * Hugepages are not in LRU lists, but they're movable.
7077                 * We need not scan over tail pages bacause we don't
7078                 * handle each tail page individually in migration.
7079                 */
7080                if (PageHuge(page)) {
7081                        iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
7082                        continue;
7083                }
7084
7085                /*
7086                 * We can't use page_count without pin a page
7087                 * because another CPU can free compound page.
7088                 * This check already skips compound tails of THP
7089                 * because their page->_refcount is zero at all time.
7090                 */
7091                if (!page_ref_count(page)) {
7092                        if (PageBuddy(page))
7093                                iter += (1 << page_order(page)) - 1;
7094                        continue;
7095                }
7096
7097                /*
7098                 * The HWPoisoned page may be not in buddy system, and
7099                 * page_count() is not 0.
7100                 */
7101                if (skip_hwpoisoned_pages && PageHWPoison(page))
7102                        continue;
7103
7104                if (!PageLRU(page))
7105                        found++;
7106                /*
7107                 * If there are RECLAIMABLE pages, we need to check
7108                 * it.  But now, memory offline itself doesn't call
7109                 * shrink_node_slabs() and it still to be fixed.
7110                 */
7111                /*
7112                 * If the page is not RAM, page_count()should be 0.
7113                 * we don't need more check. This is an _used_ not-movable page.
7114                 *
7115                 * The problematic thing here is PG_reserved pages. PG_reserved
7116                 * is set to both of a memory hole page and a _used_ kernel
7117                 * page at boot.
7118                 */
7119                if (found > count)
7120                        return true;
7121        }
7122        return false;
7123}
7124
7125bool is_pageblock_removable_nolock(struct page *page)
7126{
7127        struct zone *zone;
7128        unsigned long pfn;
7129
7130        /*
7131         * We have to be careful here because we are iterating over memory
7132         * sections which are not zone aware so we might end up outside of
7133         * the zone but still within the section.
7134         * We have to take care about the node as well. If the node is offline
7135         * its NODE_DATA will be NULL - see page_zone.
7136         */
7137        if (!node_online(page_to_nid(page)))
7138                return false;
7139
7140        zone = page_zone(page);
7141        pfn = page_to_pfn(page);
7142        if (!zone_spans_pfn(zone, pfn))
7143                return false;
7144
7145        return !has_unmovable_pages(zone, page, 0, true);
7146}
7147
7148#if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
7149
7150static unsigned long pfn_max_align_down(unsigned long pfn)
7151{
7152        return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
7153                             pageblock_nr_pages) - 1);
7154}
7155
7156static unsigned long pfn_max_align_up(unsigned long pfn)
7157{
7158        return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
7159                                pageblock_nr_pages));
7160}
7161
7162/* [start, end) must belong to a single zone. */
7163static int __alloc_contig_migrate_range(struct compact_control *cc,
7164                                        unsigned long start, unsigned long end)
7165{
7166        /* This function is based on compact_zone() from compaction.c. */
7167        unsigned long nr_reclaimed;
7168        unsigned long pfn = start;
7169        unsigned int tries = 0;
7170        int ret = 0;
7171
7172        migrate_prep();
7173
7174        while (pfn < end || !list_empty(&cc->migratepages)) {
7175                if (fatal_signal_pending(current)) {
7176                        ret = -EINTR;
7177                        break;
7178                }
7179
7180                if (list_empty(&cc->migratepages)) {
7181                        cc->nr_migratepages = 0;
7182                        pfn = isolate_migratepages_range(cc, pfn, end);
7183                        if (!pfn) {
7184                                ret = -EINTR;
7185                                break;
7186                        }
7187                        tries = 0;
7188                } else if (++tries == 5) {
7189                        ret = ret < 0 ? ret : -EBUSY;
7190                        break;
7191                }
7192
7193                nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
7194                                                        &cc->migratepages);
7195                cc->nr_migratepages -= nr_reclaimed;
7196
7197                ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
7198                                    NULL, 0, cc->mode, MR_CMA);
7199        }
7200        if (ret < 0) {
7201                putback_movable_pages(&cc->migratepages);
7202                return ret;
7203        }
7204        return 0;
7205}
7206
7207/**
7208 * alloc_contig_range() -- tries to allocate given range of pages
7209 * @start:      start PFN to allocate
7210 * @end:        one-past-the-last PFN to allocate
7211 * @migratetype:        migratetype of the underlaying pageblocks (either
7212 *                      #MIGRATE_MOVABLE or #MIGRATE_CMA).  All pageblocks
7213 *                      in range must have the same migratetype and it must
7214 *                      be either of the two.
7215 *
7216 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
7217 * aligned, however it's the caller's responsibility to guarantee that
7218 * we are the only thread that changes migrate type of pageblocks the
7219 * pages fall in.
7220 *
7221 * The PFN range must belong to a single zone.
7222 *
7223 * Returns zero on success or negative error code.  On success all
7224 * pages which PFN is in [start, end) are allocated for the caller and
7225 * need to be freed with free_contig_range().
7226 */
7227int alloc_contig_range(unsigned long start, unsigned long end,
7228                       unsigned migratetype)
7229{
7230        unsigned long outer_start, outer_end;
7231        unsigned int order;
7232        int ret = 0;
7233
7234        struct compact_control cc = {
7235                .nr_migratepages = 0,
7236                .order = -1,
7237                .zone = page_zone(pfn_to_page(start)),
7238                .mode = MIGRATE_SYNC,
7239                .ignore_skip_hint = true,
7240        };
7241        INIT_LIST_HEAD(&cc.migratepages);
7242
7243        /*
7244         * What we do here is we mark all pageblocks in range as
7245         * MIGRATE_ISOLATE.  Because pageblock and max order pages may
7246         * have different sizes, and due to the way page allocator
7247         * work, we align the range to biggest of the two pages so
7248         * that page allocator won't try to merge buddies from
7249         * different pageblocks and change MIGRATE_ISOLATE to some
7250         * other migration type.
7251         *
7252         * Once the pageblocks are marked as MIGRATE_ISOLATE, we
7253         * migrate the pages from an unaligned range (ie. pages that
7254         * we are interested in).  This will put all the pages in
7255         * range back to page allocator as MIGRATE_ISOLATE.
7256         *
7257         * When this is done, we take the pages in range from page
7258         * allocator removing them from the buddy system.  This way
7259         * page allocator will never consider using them.
7260         *
7261         * This lets us mark the pageblocks back as
7262         * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
7263         * aligned range but not in the unaligned, original range are
7264         * put back to page allocator so that buddy can use them.
7265         */
7266
7267        ret = start_isolate_page_range(pfn_max_align_down(start),
7268                                       pfn_max_align_up(end), migratetype,
7269                                       false);
7270        if (ret)
7271                return ret;
7272
7273        /*
7274         * In case of -EBUSY, we'd like to know which page causes problem.
7275         * So, just fall through. We will check it in test_pages_isolated().
7276         */
7277        ret = __alloc_contig_migrate_range(&cc, start, end);
7278        if (ret && ret != -EBUSY)
7279                goto done;
7280
7281        /*
7282         * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
7283         * aligned blocks that are marked as MIGRATE_ISOLATE.  What's
7284         * more, all pages in [start, end) are free in page allocator.
7285         * What we are going to do is to allocate all pages from
7286         * [start, end) (that is remove them from page allocator).
7287         *
7288         * The only problem is that pages at the beginning and at the
7289         * end of interesting range may be not aligned with pages that
7290         * page allocator holds, ie. they can be part of higher order
7291         * pages.  Because of this, we reserve the bigger range and
7292         * once this is done free the pages we are not interested in.
7293         *
7294         * We don't have to hold zone->lock here because the pages are
7295         * isolated thus they won't get removed from buddy.
7296         */
7297
7298        lru_add_drain_all();
7299        drain_all_pages(cc.zone);
7300
7301        order = 0;
7302        outer_start = start;
7303        while (!PageBuddy(pfn_to_page(outer_start))) {
7304                if (++order >= MAX_ORDER) {
7305                        outer_start = start;
7306                        break;
7307                }
7308                outer_start &= ~0UL << order;
7309        }
7310
7311        if (outer_start != start) {
7312                order = page_order(pfn_to_page(outer_start));
7313
7314                /*
7315                 * outer_start page could be small order buddy page and
7316                 * it doesn't include start page. Adjust outer_start
7317                 * in this case to report failed page properly
7318                 * on tracepoint in test_pages_isolated()
7319                 */
7320                if (outer_start + (1UL << order) <= start)
7321                        outer_start = start;
7322        }
7323
7324        /* Make sure the range is really isolated. */
7325        if (test_pages_isolated(outer_start, end, false)) {
7326                pr_info("%s: [%lx, %lx) PFNs busy\n",
7327                        __func__, outer_start, end);
7328                ret = -EBUSY;
7329                goto done;
7330        }
7331
7332        /* Grab isolated pages from freelists. */
7333        outer_end = isolate_freepages_range(&cc, outer_start, end);
7334        if (!outer_end) {
7335                ret = -EBUSY;
7336                goto done;
7337        }
7338
7339        /* Free head and tail (if any) */
7340        if (start != outer_start)
7341                free_contig_range(outer_start, start - outer_start);
7342        if (end != outer_end)
7343                free_contig_range(end, outer_end - end);
7344
7345done:
7346        undo_isolate_page_range(pfn_max_align_down(start),
7347                                pfn_max_align_up(end), migratetype);
7348        return ret;
7349}
7350
7351void free_contig_range(unsigned long pfn, unsigned nr_pages)
7352{
7353        unsigned int count = 0;
7354
7355        for (; nr_pages--; pfn++) {
7356                struct page *page = pfn_to_page(pfn);
7357
7358                count += page_count(page) != 1;
7359                __free_page(page);
7360        }
7361        WARN(count != 0, "%d pages are still in use!\n", count);
7362}
7363#endif
7364
7365#ifdef CONFIG_MEMORY_HOTPLUG
7366/*
7367 * The zone indicated has a new number of managed_pages; batch sizes and percpu
7368 * page high values need to be recalulated.
7369 */
7370void __meminit zone_pcp_update(struct zone *zone)
7371{
7372        unsigned cpu;
7373        mutex_lock(&pcp_batch_high_lock);
7374        for_each_possible_cpu(cpu)
7375                pageset_set_high_and_batch(zone,
7376                                per_cpu_ptr(zone->pageset, cpu));
7377        mutex_unlock(&pcp_batch_high_lock);
7378}
7379#endif
7380
7381void zone_pcp_reset(struct zone *zone)
7382{
7383        unsigned long flags;
7384        int cpu;
7385        struct per_cpu_pageset *pset;
7386
7387        /* avoid races with drain_pages()  */
7388        local_irq_save(flags);
7389        if (zone->pageset != &boot_pageset) {
7390                for_each_online_cpu(cpu) {
7391                        pset = per_cpu_ptr(zone->pageset, cpu);
7392                        drain_zonestat(zone, pset);
7393                }
7394                free_percpu(zone->pageset);
7395                zone->pageset = &boot_pageset;
7396        }
7397        local_irq_restore(flags);
7398}
7399
7400#ifdef CONFIG_MEMORY_HOTREMOVE
7401/*
7402 * All pages in the range must be in a single zone and isolated
7403 * before calling this.
7404 */
7405void
7406__offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
7407{
7408        struct page *page;
7409        struct zone *zone;
7410        unsigned int order, i;
7411        unsigned long pfn;
7412        unsigned long flags;
7413        /* find the first valid pfn */
7414        for (pfn = start_pfn; pfn < end_pfn; pfn++)
7415                if (pfn_valid(pfn))
7416                        break;
7417        if (pfn == end_pfn)
7418                return;
7419        zone = page_zone(pfn_to_page(pfn));
7420        spin_lock_irqsave(&zone->lock, flags);
7421        pfn = start_pfn;
7422        while (pfn < end_pfn) {
7423                if (!pfn_valid(pfn)) {
7424                        pfn++;
7425                        continue;
7426                }
7427                page = pfn_to_page(pfn);
7428                /*
7429                 * The HWPoisoned page may be not in buddy system, and
7430                 * page_count() is not 0.
7431                 */
7432                if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
7433                        pfn++;
7434                        SetPageReserved(page);
7435                        continue;
7436                }
7437
7438                BUG_ON(page_count(page));
7439                BUG_ON(!PageBuddy(page));
7440                order = page_order(page);
7441#ifdef CONFIG_DEBUG_VM
7442                pr_info("remove from free list %lx %d %lx\n",
7443                        pfn, 1 << order, end_pfn);
7444#endif
7445                list_del(&page->lru);
7446                rmv_page_order(page);
7447                zone->free_area[order].nr_free--;
7448                for (i = 0; i < (1 << order); i++)
7449                        SetPageReserved((page+i));
7450                pfn += (1 << order);
7451        }
7452        spin_unlock_irqrestore(&zone->lock, flags);
7453}
7454#endif
7455
7456bool is_free_buddy_page(struct page *page)
7457{
7458        struct zone *zone = page_zone(page);
7459        unsigned long pfn = page_to_pfn(page);
7460        unsigned long flags;
7461        unsigned int order;
7462
7463        spin_lock_irqsave(&zone->lock, flags);
7464        for (order = 0; order < MAX_ORDER; order++) {
7465                struct page *page_head = page - (pfn & ((1 << order) - 1));
7466
7467                if (PageBuddy(page_head) && page_order(page_head) >= order)
7468                        break;
7469        }
7470        spin_unlock_irqrestore(&zone->lock, flags);
7471
7472        return order < MAX_ORDER;
7473}
Note: See TracBrowser for help on using the repository browser.