Root/mm/page_alloc.c

1/*
2 * linux/mm/page_alloc.c
3 *
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
6 *
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
15 */
16
17#include <linux/stddef.h>
18#include <linux/mm.h>
19#include <linux/swap.h>
20#include <linux/interrupt.h>
21#include <linux/pagemap.h>
22#include <linux/jiffies.h>
23#include <linux/bootmem.h>
24#include <linux/memblock.h>
25#include <linux/compiler.h>
26#include <linux/kernel.h>
27#include <linux/kmemcheck.h>
28#include <linux/module.h>
29#include <linux/suspend.h>
30#include <linux/pagevec.h>
31#include <linux/blkdev.h>
32#include <linux/slab.h>
33#include <linux/ratelimit.h>
34#include <linux/oom.h>
35#include <linux/notifier.h>
36#include <linux/topology.h>
37#include <linux/sysctl.h>
38#include <linux/cpu.h>
39#include <linux/cpuset.h>
40#include <linux/memory_hotplug.h>
41#include <linux/nodemask.h>
42#include <linux/vmalloc.h>
43#include <linux/vmstat.h>
44#include <linux/mempolicy.h>
45#include <linux/stop_machine.h>
46#include <linux/sort.h>
47#include <linux/pfn.h>
48#include <linux/backing-dev.h>
49#include <linux/fault-inject.h>
50#include <linux/page-isolation.h>
51#include <linux/page_cgroup.h>
52#include <linux/debugobjects.h>
53#include <linux/kmemleak.h>
54#include <linux/compaction.h>
55#include <trace/events/kmem.h>
56#include <linux/ftrace_event.h>
57#include <linux/memcontrol.h>
58#include <linux/prefetch.h>
59#include <linux/mm_inline.h>
60#include <linux/migrate.h>
61#include <linux/page-debug-flags.h>
62#include <linux/hugetlb.h>
63#include <linux/sched/rt.h>
64
65#include <asm/sections.h>
66#include <asm/tlbflush.h>
67#include <asm/div64.h>
68#include "internal.h"
69
70/* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
71static DEFINE_MUTEX(pcp_batch_high_lock);
72
73#ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
74DEFINE_PER_CPU(int, numa_node);
75EXPORT_PER_CPU_SYMBOL(numa_node);
76#endif
77
78#ifdef CONFIG_HAVE_MEMORYLESS_NODES
79/*
80 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
81 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
82 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
83 * defined in <linux/topology.h>.
84 */
85DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
86EXPORT_PER_CPU_SYMBOL(_numa_mem_);
87#endif
88
89/*
90 * Array of node states.
91 */
92nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
93    [N_POSSIBLE] = NODE_MASK_ALL,
94    [N_ONLINE] = { { [0] = 1UL } },
95#ifndef CONFIG_NUMA
96    [N_NORMAL_MEMORY] = { { [0] = 1UL } },
97#ifdef CONFIG_HIGHMEM
98    [N_HIGH_MEMORY] = { { [0] = 1UL } },
99#endif
100#ifdef CONFIG_MOVABLE_NODE
101    [N_MEMORY] = { { [0] = 1UL } },
102#endif
103    [N_CPU] = { { [0] = 1UL } },
104#endif /* NUMA */
105};
106EXPORT_SYMBOL(node_states);
107
108/* Protect totalram_pages and zone->managed_pages */
109static DEFINE_SPINLOCK(managed_page_count_lock);
110
111unsigned long totalram_pages __read_mostly;
112unsigned long totalreserve_pages __read_mostly;
113/*
114 * When calculating the number of globally allowed dirty pages, there
115 * is a certain number of per-zone reserves that should not be
116 * considered dirtyable memory. This is the sum of those reserves
117 * over all existing zones that contribute dirtyable memory.
118 */
119unsigned long dirty_balance_reserve __read_mostly;
120
121int percpu_pagelist_fraction;
122gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
123
124#ifdef CONFIG_PM_SLEEP
125/*
126 * The following functions are used by the suspend/hibernate code to temporarily
127 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
128 * while devices are suspended. To avoid races with the suspend/hibernate code,
129 * they should always be called with pm_mutex held (gfp_allowed_mask also should
130 * only be modified with pm_mutex held, unless the suspend/hibernate code is
131 * guaranteed not to run in parallel with that modification).
132 */
133
134static gfp_t saved_gfp_mask;
135
136void pm_restore_gfp_mask(void)
137{
138    WARN_ON(!mutex_is_locked(&pm_mutex));
139    if (saved_gfp_mask) {
140        gfp_allowed_mask = saved_gfp_mask;
141        saved_gfp_mask = 0;
142    }
143}
144
145void pm_restrict_gfp_mask(void)
146{
147    WARN_ON(!mutex_is_locked(&pm_mutex));
148    WARN_ON(saved_gfp_mask);
149    saved_gfp_mask = gfp_allowed_mask;
150    gfp_allowed_mask &= ~GFP_IOFS;
151}
152
153bool pm_suspended_storage(void)
154{
155    if ((gfp_allowed_mask & GFP_IOFS) == GFP_IOFS)
156        return false;
157    return true;
158}
159#endif /* CONFIG_PM_SLEEP */
160
161#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
162int pageblock_order __read_mostly;
163#endif
164
165static void __free_pages_ok(struct page *page, unsigned int order);
166
167/*
168 * results with 256, 32 in the lowmem_reserve sysctl:
169 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
170 * 1G machine -> (16M dma, 784M normal, 224M high)
171 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
172 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
173 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
174 *
175 * TBD: should special case ZONE_DMA32 machines here - in those we normally
176 * don't need any ZONE_NORMAL reservation
177 */
178int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
179#ifdef CONFIG_ZONE_DMA
180     256,
181#endif
182#ifdef CONFIG_ZONE_DMA32
183     256,
184#endif
185#ifdef CONFIG_HIGHMEM
186     32,
187#endif
188     32,
189};
190
191EXPORT_SYMBOL(totalram_pages);
192
193static char * const zone_names[MAX_NR_ZONES] = {
194#ifdef CONFIG_ZONE_DMA
195     "DMA",
196#endif
197#ifdef CONFIG_ZONE_DMA32
198     "DMA32",
199#endif
200     "Normal",
201#ifdef CONFIG_HIGHMEM
202     "HighMem",
203#endif
204     "Movable",
205};
206
207int min_free_kbytes = 1024;
208int user_min_free_kbytes;
209
210static unsigned long __meminitdata nr_kernel_pages;
211static unsigned long __meminitdata nr_all_pages;
212static unsigned long __meminitdata dma_reserve;
213
214#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
215static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
216static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
217static unsigned long __initdata required_kernelcore;
218static unsigned long __initdata required_movablecore;
219static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
220
221/* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
222int movable_zone;
223EXPORT_SYMBOL(movable_zone);
224#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
225
226#if MAX_NUMNODES > 1
227int nr_node_ids __read_mostly = MAX_NUMNODES;
228int nr_online_nodes __read_mostly = 1;
229EXPORT_SYMBOL(nr_node_ids);
230EXPORT_SYMBOL(nr_online_nodes);
231#endif
232
233int page_group_by_mobility_disabled __read_mostly;
234
235void set_pageblock_migratetype(struct page *page, int migratetype)
236{
237
238    if (unlikely(page_group_by_mobility_disabled))
239        migratetype = MIGRATE_UNMOVABLE;
240
241    set_pageblock_flags_group(page, (unsigned long)migratetype,
242                    PB_migrate, PB_migrate_end);
243}
244
245bool oom_killer_disabled __read_mostly;
246
247#ifdef CONFIG_DEBUG_VM
248static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
249{
250    int ret = 0;
251    unsigned seq;
252    unsigned long pfn = page_to_pfn(page);
253    unsigned long sp, start_pfn;
254
255    do {
256        seq = zone_span_seqbegin(zone);
257        start_pfn = zone->zone_start_pfn;
258        sp = zone->spanned_pages;
259        if (!zone_spans_pfn(zone, pfn))
260            ret = 1;
261    } while (zone_span_seqretry(zone, seq));
262
263    if (ret)
264        pr_err("page %lu outside zone [ %lu - %lu ]\n",
265            pfn, start_pfn, start_pfn + sp);
266
267    return ret;
268}
269
270static int page_is_consistent(struct zone *zone, struct page *page)
271{
272    if (!pfn_valid_within(page_to_pfn(page)))
273        return 0;
274    if (zone != page_zone(page))
275        return 0;
276
277    return 1;
278}
279/*
280 * Temporary debugging check for pages not lying within a given zone.
281 */
282static int bad_range(struct zone *zone, struct page *page)
283{
284    if (page_outside_zone_boundaries(zone, page))
285        return 1;
286    if (!page_is_consistent(zone, page))
287        return 1;
288
289    return 0;
290}
291#else
292static inline int bad_range(struct zone *zone, struct page *page)
293{
294    return 0;
295}
296#endif
297
298static void bad_page(struct page *page)
299{
300    static unsigned long resume;
301    static unsigned long nr_shown;
302    static unsigned long nr_unshown;
303
304    /* Don't complain about poisoned pages */
305    if (PageHWPoison(page)) {
306        page_mapcount_reset(page); /* remove PageBuddy */
307        return;
308    }
309
310    /*
311     * Allow a burst of 60 reports, then keep quiet for that minute;
312     * or allow a steady drip of one report per second.
313     */
314    if (nr_shown == 60) {
315        if (time_before(jiffies, resume)) {
316            nr_unshown++;
317            goto out;
318        }
319        if (nr_unshown) {
320            printk(KERN_ALERT
321                  "BUG: Bad page state: %lu messages suppressed\n",
322                nr_unshown);
323            nr_unshown = 0;
324        }
325        nr_shown = 0;
326    }
327    if (nr_shown++ == 0)
328        resume = jiffies + 60 * HZ;
329
330    printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
331        current->comm, page_to_pfn(page));
332    dump_page(page);
333
334    print_modules();
335    dump_stack();
336out:
337    /* Leave bad fields for debug, except PageBuddy could make trouble */
338    page_mapcount_reset(page); /* remove PageBuddy */
339    add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
340}
341
342/*
343 * Higher-order pages are called "compound pages". They are structured thusly:
344 *
345 * The first PAGE_SIZE page is called the "head page".
346 *
347 * The remaining PAGE_SIZE pages are called "tail pages".
348 *
349 * All pages have PG_compound set. All tail pages have their ->first_page
350 * pointing at the head page.
351 *
352 * The first tail page's ->lru.next holds the address of the compound page's
353 * put_page() function. Its ->lru.prev holds the order of allocation.
354 * This usage means that zero-order pages may not be compound.
355 */
356
357static void free_compound_page(struct page *page)
358{
359    __free_pages_ok(page, compound_order(page));
360}
361
362void prep_compound_page(struct page *page, unsigned long order)
363{
364    int i;
365    int nr_pages = 1 << order;
366
367    set_compound_page_dtor(page, free_compound_page);
368    set_compound_order(page, order);
369    __SetPageHead(page);
370    for (i = 1; i < nr_pages; i++) {
371        struct page *p = page + i;
372        __SetPageTail(p);
373        set_page_count(p, 0);
374        p->first_page = page;
375    }
376}
377
378/* update __split_huge_page_refcount if you change this function */
379static int destroy_compound_page(struct page *page, unsigned long order)
380{
381    int i;
382    int nr_pages = 1 << order;
383    int bad = 0;
384
385    if (unlikely(compound_order(page) != order)) {
386        bad_page(page);
387        bad++;
388    }
389
390    __ClearPageHead(page);
391
392    for (i = 1; i < nr_pages; i++) {
393        struct page *p = page + i;
394
395        if (unlikely(!PageTail(p) || (p->first_page != page))) {
396            bad_page(page);
397            bad++;
398        }
399        __ClearPageTail(p);
400    }
401
402    return bad;
403}
404
405static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
406{
407    int i;
408
409    /*
410     * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
411     * and __GFP_HIGHMEM from hard or soft interrupt context.
412     */
413    VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
414    for (i = 0; i < (1 << order); i++)
415        clear_highpage(page + i);
416}
417
418#ifdef CONFIG_DEBUG_PAGEALLOC
419unsigned int _debug_guardpage_minorder;
420
421static int __init debug_guardpage_minorder_setup(char *buf)
422{
423    unsigned long res;
424
425    if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
426        printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
427        return 0;
428    }
429    _debug_guardpage_minorder = res;
430    printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
431    return 0;
432}
433__setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
434
435static inline void set_page_guard_flag(struct page *page)
436{
437    __set_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
438}
439
440static inline void clear_page_guard_flag(struct page *page)
441{
442    __clear_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
443}
444#else
445static inline void set_page_guard_flag(struct page *page) { }
446static inline void clear_page_guard_flag(struct page *page) { }
447#endif
448
449static inline void set_page_order(struct page *page, int order)
450{
451    set_page_private(page, order);
452    __SetPageBuddy(page);
453}
454
455static inline void rmv_page_order(struct page *page)
456{
457    __ClearPageBuddy(page);
458    set_page_private(page, 0);
459}
460
461/*
462 * Locate the struct page for both the matching buddy in our
463 * pair (buddy1) and the combined O(n+1) page they form (page).
464 *
465 * 1) Any buddy B1 will have an order O twin B2 which satisfies
466 * the following equation:
467 * B2 = B1 ^ (1 << O)
468 * For example, if the starting buddy (buddy2) is #8 its order
469 * 1 buddy is #10:
470 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
471 *
472 * 2) Any buddy B will have an order O+1 parent P which
473 * satisfies the following equation:
474 * P = B & ~(1 << O)
475 *
476 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
477 */
478static inline unsigned long
479__find_buddy_index(unsigned long page_idx, unsigned int order)
480{
481    return page_idx ^ (1 << order);
482}
483
484/*
485 * This function checks whether a page is free && is the buddy
486 * we can do coalesce a page and its buddy if
487 * (a) the buddy is not in a hole &&
488 * (b) the buddy is in the buddy system &&
489 * (c) a page and its buddy have the same order &&
490 * (d) a page and its buddy are in the same zone.
491 *
492 * For recording whether a page is in the buddy system, we set ->_mapcount
493 * PAGE_BUDDY_MAPCOUNT_VALUE.
494 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
495 * serialized by zone->lock.
496 *
497 * For recording page's order, we use page_private(page).
498 */
499static inline int page_is_buddy(struct page *page, struct page *buddy,
500                                int order)
501{
502    if (!pfn_valid_within(page_to_pfn(buddy)))
503        return 0;
504
505    if (page_zone_id(page) != page_zone_id(buddy))
506        return 0;
507
508    if (page_is_guard(buddy) && page_order(buddy) == order) {
509        VM_BUG_ON(page_count(buddy) != 0);
510        return 1;
511    }
512
513    if (PageBuddy(buddy) && page_order(buddy) == order) {
514        VM_BUG_ON(page_count(buddy) != 0);
515        return 1;
516    }
517    return 0;
518}
519
520/*
521 * Freeing function for a buddy system allocator.
522 *
523 * The concept of a buddy system is to maintain direct-mapped table
524 * (containing bit values) for memory blocks of various "orders".
525 * The bottom level table contains the map for the smallest allocatable
526 * units of memory (here, pages), and each level above it describes
527 * pairs of units from the levels below, hence, "buddies".
528 * At a high level, all that happens here is marking the table entry
529 * at the bottom level available, and propagating the changes upward
530 * as necessary, plus some accounting needed to play nicely with other
531 * parts of the VM system.
532 * At each level, we keep a list of pages, which are heads of continuous
533 * free pages of length of (1 << order) and marked with _mapcount
534 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
535 * field.
536 * So when we are allocating or freeing one, we can derive the state of the
537 * other. That is, if we allocate a small block, and both were
538 * free, the remainder of the region must be split into blocks.
539 * If a block is freed, and its buddy is also free, then this
540 * triggers coalescing into a block of larger size.
541 *
542 * -- nyc
543 */
544
545static inline void __free_one_page(struct page *page,
546        struct zone *zone, unsigned int order,
547        int migratetype)
548{
549    unsigned long page_idx;
550    unsigned long combined_idx;
551    unsigned long uninitialized_var(buddy_idx);
552    struct page *buddy;
553
554    VM_BUG_ON(!zone_is_initialized(zone));
555
556    if (unlikely(PageCompound(page)))
557        if (unlikely(destroy_compound_page(page, order)))
558            return;
559
560    VM_BUG_ON(migratetype == -1);
561
562    page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
563
564    VM_BUG_ON(page_idx & ((1 << order) - 1));
565    VM_BUG_ON(bad_range(zone, page));
566
567    while (order < MAX_ORDER-1) {
568        buddy_idx = __find_buddy_index(page_idx, order);
569        buddy = page + (buddy_idx - page_idx);
570        if (!page_is_buddy(page, buddy, order))
571            break;
572        /*
573         * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
574         * merge with it and move up one order.
575         */
576        if (page_is_guard(buddy)) {
577            clear_page_guard_flag(buddy);
578            set_page_private(page, 0);
579            __mod_zone_freepage_state(zone, 1 << order,
580                          migratetype);
581        } else {
582            list_del(&buddy->lru);
583            zone->free_area[order].nr_free--;
584            rmv_page_order(buddy);
585        }
586        combined_idx = buddy_idx & page_idx;
587        page = page + (combined_idx - page_idx);
588        page_idx = combined_idx;
589        order++;
590    }
591    set_page_order(page, order);
592
593    /*
594     * If this is not the largest possible page, check if the buddy
595     * of the next-highest order is free. If it is, it's possible
596     * that pages are being freed that will coalesce soon. In case,
597     * that is happening, add the free page to the tail of the list
598     * so it's less likely to be used soon and more likely to be merged
599     * as a higher order page
600     */
601    if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
602        struct page *higher_page, *higher_buddy;
603        combined_idx = buddy_idx & page_idx;
604        higher_page = page + (combined_idx - page_idx);
605        buddy_idx = __find_buddy_index(combined_idx, order + 1);
606        higher_buddy = higher_page + (buddy_idx - combined_idx);
607        if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
608            list_add_tail(&page->lru,
609                &zone->free_area[order].free_list[migratetype]);
610            goto out;
611        }
612    }
613
614    list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
615out:
616    zone->free_area[order].nr_free++;
617}
618
619static inline int free_pages_check(struct page *page)
620{
621    if (unlikely(page_mapcount(page) |
622        (page->mapping != NULL) |
623        (atomic_read(&page->_count) != 0) |
624        (page->flags & PAGE_FLAGS_CHECK_AT_FREE) |
625        (mem_cgroup_bad_page_check(page)))) {
626        bad_page(page);
627        return 1;
628    }
629    page_nid_reset_last(page);
630    if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
631        page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
632    return 0;
633}
634
635/*
636 * Frees a number of pages from the PCP lists
637 * Assumes all pages on list are in same zone, and of same order.
638 * count is the number of pages to free.
639 *
640 * If the zone was previously in an "all pages pinned" state then look to
641 * see if this freeing clears that state.
642 *
643 * And clear the zone's pages_scanned counter, to hold off the "all pages are
644 * pinned" detection logic.
645 */
646static void free_pcppages_bulk(struct zone *zone, int count,
647                    struct per_cpu_pages *pcp)
648{
649    int migratetype = 0;
650    int batch_free = 0;
651    int to_free = count;
652
653    spin_lock(&zone->lock);
654    zone->pages_scanned = 0;
655
656    while (to_free) {
657        struct page *page;
658        struct list_head *list;
659
660        /*
661         * Remove pages from lists in a round-robin fashion. A
662         * batch_free count is maintained that is incremented when an
663         * empty list is encountered. This is so more pages are freed
664         * off fuller lists instead of spinning excessively around empty
665         * lists
666         */
667        do {
668            batch_free++;
669            if (++migratetype == MIGRATE_PCPTYPES)
670                migratetype = 0;
671            list = &pcp->lists[migratetype];
672        } while (list_empty(list));
673
674        /* This is the only non-empty list. Free them all. */
675        if (batch_free == MIGRATE_PCPTYPES)
676            batch_free = to_free;
677
678        do {
679            int mt; /* migratetype of the to-be-freed page */
680
681            page = list_entry(list->prev, struct page, lru);
682            /* must delete as __free_one_page list manipulates */
683            list_del(&page->lru);
684            mt = get_freepage_migratetype(page);
685            /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
686            __free_one_page(page, zone, 0, mt);
687            trace_mm_page_pcpu_drain(page, 0, mt);
688            if (likely(!is_migrate_isolate_page(page))) {
689                __mod_zone_page_state(zone, NR_FREE_PAGES, 1);
690                if (is_migrate_cma(mt))
691                    __mod_zone_page_state(zone, NR_FREE_CMA_PAGES, 1);
692            }
693        } while (--to_free && --batch_free && !list_empty(list));
694    }
695    spin_unlock(&zone->lock);
696}
697
698static void free_one_page(struct zone *zone, struct page *page, int order,
699                int migratetype)
700{
701    spin_lock(&zone->lock);
702    zone->pages_scanned = 0;
703
704    __free_one_page(page, zone, order, migratetype);
705    if (unlikely(!is_migrate_isolate(migratetype)))
706        __mod_zone_freepage_state(zone, 1 << order, migratetype);
707    spin_unlock(&zone->lock);
708}
709
710static bool free_pages_prepare(struct page *page, unsigned int order)
711{
712    int i;
713    int bad = 0;
714
715    trace_mm_page_free(page, order);
716    kmemcheck_free_shadow(page, order);
717
718    if (PageAnon(page))
719        page->mapping = NULL;
720    for (i = 0; i < (1 << order); i++)
721        bad += free_pages_check(page + i);
722    if (bad)
723        return false;
724
725    if (!PageHighMem(page)) {
726        debug_check_no_locks_freed(page_address(page),
727                       PAGE_SIZE << order);
728        debug_check_no_obj_freed(page_address(page),
729                       PAGE_SIZE << order);
730    }
731    arch_free_page(page, order);
732    kernel_map_pages(page, 1 << order, 0);
733
734    return true;
735}
736
737static void __free_pages_ok(struct page *page, unsigned int order)
738{
739    unsigned long flags;
740    int migratetype;
741
742    if (!free_pages_prepare(page, order))
743        return;
744
745    local_irq_save(flags);
746    __count_vm_events(PGFREE, 1 << order);
747    migratetype = get_pageblock_migratetype(page);
748    set_freepage_migratetype(page, migratetype);
749    free_one_page(page_zone(page), page, order, migratetype);
750    local_irq_restore(flags);
751}
752
753void __init __free_pages_bootmem(struct page *page, unsigned int order)
754{
755    unsigned int nr_pages = 1 << order;
756    struct page *p = page;
757    unsigned int loop;
758
759    prefetchw(p);
760    for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
761        prefetchw(p + 1);
762        __ClearPageReserved(p);
763        set_page_count(p, 0);
764    }
765    __ClearPageReserved(p);
766    set_page_count(p, 0);
767
768    page_zone(page)->managed_pages += nr_pages;
769    set_page_refcounted(page);
770    __free_pages(page, order);
771}
772
773#ifdef CONFIG_CMA
774/* Free whole pageblock and set its migration type to MIGRATE_CMA. */
775void __init init_cma_reserved_pageblock(struct page *page)
776{
777    unsigned i = pageblock_nr_pages;
778    struct page *p = page;
779
780    do {
781        __ClearPageReserved(p);
782        set_page_count(p, 0);
783    } while (++p, --i);
784
785    set_page_refcounted(page);
786    set_pageblock_migratetype(page, MIGRATE_CMA);
787    __free_pages(page, pageblock_order);
788    adjust_managed_page_count(page, pageblock_nr_pages);
789}
790#endif
791
792/*
793 * The order of subdivision here is critical for the IO subsystem.
794 * Please do not alter this order without good reasons and regression
795 * testing. Specifically, as large blocks of memory are subdivided,
796 * the order in which smaller blocks are delivered depends on the order
797 * they're subdivided in this function. This is the primary factor
798 * influencing the order in which pages are delivered to the IO
799 * subsystem according to empirical testing, and this is also justified
800 * by considering the behavior of a buddy system containing a single
801 * large block of memory acted on by a series of small allocations.
802 * This behavior is a critical factor in sglist merging's success.
803 *
804 * -- nyc
805 */
806static inline void expand(struct zone *zone, struct page *page,
807    int low, int high, struct free_area *area,
808    int migratetype)
809{
810    unsigned long size = 1 << high;
811
812    while (high > low) {
813        area--;
814        high--;
815        size >>= 1;
816        VM_BUG_ON(bad_range(zone, &page[size]));
817
818#ifdef CONFIG_DEBUG_PAGEALLOC
819        if (high < debug_guardpage_minorder()) {
820            /*
821             * Mark as guard pages (or page), that will allow to
822             * merge back to allocator when buddy will be freed.
823             * Corresponding page table entries will not be touched,
824             * pages will stay not present in virtual address space
825             */
826            INIT_LIST_HEAD(&page[size].lru);
827            set_page_guard_flag(&page[size]);
828            set_page_private(&page[size], high);
829            /* Guard pages are not available for any usage */
830            __mod_zone_freepage_state(zone, -(1 << high),
831                          migratetype);
832            continue;
833        }
834#endif
835        list_add(&page[size].lru, &area->free_list[migratetype]);
836        area->nr_free++;
837        set_page_order(&page[size], high);
838    }
839}
840
841/*
842 * This page is about to be returned from the page allocator
843 */
844static inline int check_new_page(struct page *page)
845{
846    if (unlikely(page_mapcount(page) |
847        (page->mapping != NULL) |
848        (atomic_read(&page->_count) != 0) |
849        (page->flags & PAGE_FLAGS_CHECK_AT_PREP) |
850        (mem_cgroup_bad_page_check(page)))) {
851        bad_page(page);
852        return 1;
853    }
854    return 0;
855}
856
857static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
858{
859    int i;
860
861    for (i = 0; i < (1 << order); i++) {
862        struct page *p = page + i;
863        if (unlikely(check_new_page(p)))
864            return 1;
865    }
866
867    set_page_private(page, 0);
868    set_page_refcounted(page);
869
870    arch_alloc_page(page, order);
871    kernel_map_pages(page, 1 << order, 1);
872
873    if (gfp_flags & __GFP_ZERO)
874        prep_zero_page(page, order, gfp_flags);
875
876    if (order && (gfp_flags & __GFP_COMP))
877        prep_compound_page(page, order);
878
879    return 0;
880}
881
882/*
883 * Go through the free lists for the given migratetype and remove
884 * the smallest available page from the freelists
885 */
886static inline
887struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
888                        int migratetype)
889{
890    unsigned int current_order;
891    struct free_area *area;
892    struct page *page;
893
894    /* Find a page of the appropriate size in the preferred list */
895    for (current_order = order; current_order < MAX_ORDER; ++current_order) {
896        area = &(zone->free_area[current_order]);
897        if (list_empty(&area->free_list[migratetype]))
898            continue;
899
900        page = list_entry(area->free_list[migratetype].next,
901                            struct page, lru);
902        list_del(&page->lru);
903        rmv_page_order(page);
904        area->nr_free--;
905        expand(zone, page, order, current_order, area, migratetype);
906        return page;
907    }
908
909    return NULL;
910}
911
912
913/*
914 * This array describes the order lists are fallen back to when
915 * the free lists for the desirable migrate type are depleted
916 */
917static int fallbacks[MIGRATE_TYPES][4] = {
918    [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
919    [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
920#ifdef CONFIG_CMA
921    [MIGRATE_MOVABLE] = { MIGRATE_CMA, MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
922    [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */
923#else
924    [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
925#endif
926    [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */
927#ifdef CONFIG_MEMORY_ISOLATION
928    [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */
929#endif
930};
931
932/*
933 * Move the free pages in a range to the free lists of the requested type.
934 * Note that start_page and end_pages are not aligned on a pageblock
935 * boundary. If alignment is required, use move_freepages_block()
936 */
937int move_freepages(struct zone *zone,
938              struct page *start_page, struct page *end_page,
939              int migratetype)
940{
941    struct page *page;
942    unsigned long order;
943    int pages_moved = 0;
944
945#ifndef CONFIG_HOLES_IN_ZONE
946    /*
947     * page_zone is not safe to call in this context when
948     * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
949     * anyway as we check zone boundaries in move_freepages_block().
950     * Remove at a later date when no bug reports exist related to
951     * grouping pages by mobility
952     */
953    BUG_ON(page_zone(start_page) != page_zone(end_page));
954#endif
955
956    for (page = start_page; page <= end_page;) {
957        /* Make sure we are not inadvertently changing nodes */
958        VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
959
960        if (!pfn_valid_within(page_to_pfn(page))) {
961            page++;
962            continue;
963        }
964
965        if (!PageBuddy(page)) {
966            page++;
967            continue;
968        }
969
970        order = page_order(page);
971        list_move(&page->lru,
972              &zone->free_area[order].free_list[migratetype]);
973        set_freepage_migratetype(page, migratetype);
974        page += 1 << order;
975        pages_moved += 1 << order;
976    }
977
978    return pages_moved;
979}
980
981int move_freepages_block(struct zone *zone, struct page *page,
982                int migratetype)
983{
984    unsigned long start_pfn, end_pfn;
985    struct page *start_page, *end_page;
986
987    start_pfn = page_to_pfn(page);
988    start_pfn = start_pfn & ~(pageblock_nr_pages-1);
989    start_page = pfn_to_page(start_pfn);
990    end_page = start_page + pageblock_nr_pages - 1;
991    end_pfn = start_pfn + pageblock_nr_pages - 1;
992
993    /* Do not cross zone boundaries */
994    if (!zone_spans_pfn(zone, start_pfn))
995        start_page = page;
996    if (!zone_spans_pfn(zone, end_pfn))
997        return 0;
998
999    return move_freepages(zone, start_page, end_page, migratetype);
1000}
1001
1002static void change_pageblock_range(struct page *pageblock_page,
1003                    int start_order, int migratetype)
1004{
1005    int nr_pageblocks = 1 << (start_order - pageblock_order);
1006
1007    while (nr_pageblocks--) {
1008        set_pageblock_migratetype(pageblock_page, migratetype);
1009        pageblock_page += pageblock_nr_pages;
1010    }
1011}
1012
1013/*
1014 * If breaking a large block of pages, move all free pages to the preferred
1015 * allocation list. If falling back for a reclaimable kernel allocation, be
1016 * more aggressive about taking ownership of free pages.
1017 *
1018 * On the other hand, never change migration type of MIGRATE_CMA pageblocks
1019 * nor move CMA pages to different free lists. We don't want unmovable pages
1020 * to be allocated from MIGRATE_CMA areas.
1021 *
1022 * Returns the new migratetype of the pageblock (or the same old migratetype
1023 * if it was unchanged).
1024 */
1025static int try_to_steal_freepages(struct zone *zone, struct page *page,
1026                  int start_type, int fallback_type)
1027{
1028    int current_order = page_order(page);
1029
1030    if (is_migrate_cma(fallback_type))
1031        return fallback_type;
1032
1033    /* Take ownership for orders >= pageblock_order */
1034    if (current_order >= pageblock_order) {
1035        change_pageblock_range(page, current_order, start_type);
1036        return start_type;
1037    }
1038
1039    if (current_order >= pageblock_order / 2 ||
1040        start_type == MIGRATE_RECLAIMABLE ||
1041        page_group_by_mobility_disabled) {
1042        int pages;
1043
1044        pages = move_freepages_block(zone, page, start_type);
1045
1046        /* Claim the whole block if over half of it is free */
1047        if (pages >= (1 << (pageblock_order-1)) ||
1048                page_group_by_mobility_disabled) {
1049
1050            set_pageblock_migratetype(page, start_type);
1051            return start_type;
1052        }
1053
1054    }
1055
1056    return fallback_type;
1057}
1058
1059/* Remove an element from the buddy allocator from the fallback list */
1060static inline struct page *
1061__rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
1062{
1063    struct free_area *area;
1064    int current_order;
1065    struct page *page;
1066    int migratetype, new_type, i;
1067
1068    /* Find the largest possible block of pages in the other list */
1069    for (current_order = MAX_ORDER-1; current_order >= order;
1070                        --current_order) {
1071        for (i = 0;; i++) {
1072            migratetype = fallbacks[start_migratetype][i];
1073
1074            /* MIGRATE_RESERVE handled later if necessary */
1075            if (migratetype == MIGRATE_RESERVE)
1076                break;
1077
1078            area = &(zone->free_area[current_order]);
1079            if (list_empty(&area->free_list[migratetype]))
1080                continue;
1081
1082            page = list_entry(area->free_list[migratetype].next,
1083                    struct page, lru);
1084            area->nr_free--;
1085
1086            new_type = try_to_steal_freepages(zone, page,
1087                              start_migratetype,
1088                              migratetype);
1089
1090            /* Remove the page from the freelists */
1091            list_del(&page->lru);
1092            rmv_page_order(page);
1093
1094            /*
1095             * Borrow the excess buddy pages as well, irrespective
1096             * of whether we stole freepages, or took ownership of
1097             * the pageblock or not.
1098             *
1099             * Exception: When borrowing from MIGRATE_CMA, release
1100             * the excess buddy pages to CMA itself.
1101             */
1102            expand(zone, page, order, current_order, area,
1103                   is_migrate_cma(migratetype)
1104                 ? migratetype : start_migratetype);
1105
1106            trace_mm_page_alloc_extfrag(page, order,
1107                current_order, start_migratetype, migratetype,
1108                new_type == start_migratetype);
1109
1110            return page;
1111        }
1112    }
1113
1114    return NULL;
1115}
1116
1117/*
1118 * Do the hard work of removing an element from the buddy allocator.
1119 * Call me with the zone->lock already held.
1120 */
1121static struct page *__rmqueue(struct zone *zone, unsigned int order,
1122                        int migratetype)
1123{
1124    struct page *page;
1125
1126retry_reserve:
1127    page = __rmqueue_smallest(zone, order, migratetype);
1128
1129    if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1130        page = __rmqueue_fallback(zone, order, migratetype);
1131
1132        /*
1133         * Use MIGRATE_RESERVE rather than fail an allocation. goto
1134         * is used because __rmqueue_smallest is an inline function
1135         * and we want just one call site
1136         */
1137        if (!page) {
1138            migratetype = MIGRATE_RESERVE;
1139            goto retry_reserve;
1140        }
1141    }
1142
1143    trace_mm_page_alloc_zone_locked(page, order, migratetype);
1144    return page;
1145}
1146
1147/*
1148 * Obtain a specified number of elements from the buddy allocator, all under
1149 * a single hold of the lock, for efficiency. Add them to the supplied list.
1150 * Returns the number of new pages which were placed at *list.
1151 */
1152static int rmqueue_bulk(struct zone *zone, unsigned int order,
1153            unsigned long count, struct list_head *list,
1154            int migratetype, int cold)
1155{
1156    int mt = migratetype, i;
1157
1158    spin_lock(&zone->lock);
1159    for (i = 0; i < count; ++i) {
1160        struct page *page = __rmqueue(zone, order, migratetype);
1161        if (unlikely(page == NULL))
1162            break;
1163
1164        /*
1165         * Split buddy pages returned by expand() are received here
1166         * in physical page order. The page is added to the callers and
1167         * list and the list head then moves forward. From the callers
1168         * perspective, the linked list is ordered by page number in
1169         * some conditions. This is useful for IO devices that can
1170         * merge IO requests if the physical pages are ordered
1171         * properly.
1172         */
1173        if (likely(cold == 0))
1174            list_add(&page->lru, list);
1175        else
1176            list_add_tail(&page->lru, list);
1177        if (IS_ENABLED(CONFIG_CMA)) {
1178            mt = get_pageblock_migratetype(page);
1179            if (!is_migrate_cma(mt) && !is_migrate_isolate(mt))
1180                mt = migratetype;
1181        }
1182        set_freepage_migratetype(page, mt);
1183        list = &page->lru;
1184        if (is_migrate_cma(mt))
1185            __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1186                          -(1 << order));
1187    }
1188    __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1189    spin_unlock(&zone->lock);
1190    return i;
1191}
1192
1193#ifdef CONFIG_NUMA
1194/*
1195 * Called from the vmstat counter updater to drain pagesets of this
1196 * currently executing processor on remote nodes after they have
1197 * expired.
1198 *
1199 * Note that this function must be called with the thread pinned to
1200 * a single processor.
1201 */
1202void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1203{
1204    unsigned long flags;
1205    int to_drain;
1206    unsigned long batch;
1207
1208    local_irq_save(flags);
1209    batch = ACCESS_ONCE(pcp->batch);
1210    if (pcp->count >= batch)
1211        to_drain = batch;
1212    else
1213        to_drain = pcp->count;
1214    if (to_drain > 0) {
1215        free_pcppages_bulk(zone, to_drain, pcp);
1216        pcp->count -= to_drain;
1217    }
1218    local_irq_restore(flags);
1219}
1220#endif
1221
1222/*
1223 * Drain pages of the indicated processor.
1224 *
1225 * The processor must either be the current processor and the
1226 * thread pinned to the current processor or a processor that
1227 * is not online.
1228 */
1229static void drain_pages(unsigned int cpu)
1230{
1231    unsigned long flags;
1232    struct zone *zone;
1233
1234    for_each_populated_zone(zone) {
1235        struct per_cpu_pageset *pset;
1236        struct per_cpu_pages *pcp;
1237
1238        local_irq_save(flags);
1239        pset = per_cpu_ptr(zone->pageset, cpu);
1240
1241        pcp = &pset->pcp;
1242        if (pcp->count) {
1243            free_pcppages_bulk(zone, pcp->count, pcp);
1244            pcp->count = 0;
1245        }
1246        local_irq_restore(flags);
1247    }
1248}
1249
1250/*
1251 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1252 */
1253void drain_local_pages(void *arg)
1254{
1255    drain_pages(smp_processor_id());
1256}
1257
1258/*
1259 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1260 *
1261 * Note that this code is protected against sending an IPI to an offline
1262 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1263 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1264 * nothing keeps CPUs from showing up after we populated the cpumask and
1265 * before the call to on_each_cpu_mask().
1266 */
1267void drain_all_pages(void)
1268{
1269    int cpu;
1270    struct per_cpu_pageset *pcp;
1271    struct zone *zone;
1272
1273    /*
1274     * Allocate in the BSS so we wont require allocation in
1275     * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1276     */
1277    static cpumask_t cpus_with_pcps;
1278
1279    /*
1280     * We don't care about racing with CPU hotplug event
1281     * as offline notification will cause the notified
1282     * cpu to drain that CPU pcps and on_each_cpu_mask
1283     * disables preemption as part of its processing
1284     */
1285    for_each_online_cpu(cpu) {
1286        bool has_pcps = false;
1287        for_each_populated_zone(zone) {
1288            pcp = per_cpu_ptr(zone->pageset, cpu);
1289            if (pcp->pcp.count) {
1290                has_pcps = true;
1291                break;
1292            }
1293        }
1294        if (has_pcps)
1295            cpumask_set_cpu(cpu, &cpus_with_pcps);
1296        else
1297            cpumask_clear_cpu(cpu, &cpus_with_pcps);
1298    }
1299    on_each_cpu_mask(&cpus_with_pcps, drain_local_pages, NULL, 1);
1300}
1301
1302#ifdef CONFIG_HIBERNATION
1303
1304void mark_free_pages(struct zone *zone)
1305{
1306    unsigned long pfn, max_zone_pfn;
1307    unsigned long flags;
1308    int order, t;
1309    struct list_head *curr;
1310
1311    if (zone_is_empty(zone))
1312        return;
1313
1314    spin_lock_irqsave(&zone->lock, flags);
1315
1316    max_zone_pfn = zone_end_pfn(zone);
1317    for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1318        if (pfn_valid(pfn)) {
1319            struct page *page = pfn_to_page(pfn);
1320
1321            if (!swsusp_page_is_forbidden(page))
1322                swsusp_unset_page_free(page);
1323        }
1324
1325    for_each_migratetype_order(order, t) {
1326        list_for_each(curr, &zone->free_area[order].free_list[t]) {
1327            unsigned long i;
1328
1329            pfn = page_to_pfn(list_entry(curr, struct page, lru));
1330            for (i = 0; i < (1UL << order); i++)
1331                swsusp_set_page_free(pfn_to_page(pfn + i));
1332        }
1333    }
1334    spin_unlock_irqrestore(&zone->lock, flags);
1335}
1336#endif /* CONFIG_PM */
1337
1338/*
1339 * Free a 0-order page
1340 * cold == 1 ? free a cold page : free a hot page
1341 */
1342void free_hot_cold_page(struct page *page, int cold)
1343{
1344    struct zone *zone = page_zone(page);
1345    struct per_cpu_pages *pcp;
1346    unsigned long flags;
1347    int migratetype;
1348
1349    if (!free_pages_prepare(page, 0))
1350        return;
1351
1352    migratetype = get_pageblock_migratetype(page);
1353    set_freepage_migratetype(page, migratetype);
1354    local_irq_save(flags);
1355    __count_vm_event(PGFREE);
1356
1357    /*
1358     * We only track unmovable, reclaimable and movable on pcp lists.
1359     * Free ISOLATE pages back to the allocator because they are being
1360     * offlined but treat RESERVE as movable pages so we can get those
1361     * areas back if necessary. Otherwise, we may have to free
1362     * excessively into the page allocator
1363     */
1364    if (migratetype >= MIGRATE_PCPTYPES) {
1365        if (unlikely(is_migrate_isolate(migratetype))) {
1366            free_one_page(zone, page, 0, migratetype);
1367            goto out;
1368        }
1369        migratetype = MIGRATE_MOVABLE;
1370    }
1371
1372    pcp = &this_cpu_ptr(zone->pageset)->pcp;
1373    if (cold)
1374        list_add_tail(&page->lru, &pcp->lists[migratetype]);
1375    else
1376        list_add(&page->lru, &pcp->lists[migratetype]);
1377    pcp->count++;
1378    if (pcp->count >= pcp->high) {
1379        unsigned long batch = ACCESS_ONCE(pcp->batch);
1380        free_pcppages_bulk(zone, batch, pcp);
1381        pcp->count -= batch;
1382    }
1383
1384out:
1385    local_irq_restore(flags);
1386}
1387
1388/*
1389 * Free a list of 0-order pages
1390 */
1391void free_hot_cold_page_list(struct list_head *list, int cold)
1392{
1393    struct page *page, *next;
1394
1395    list_for_each_entry_safe(page, next, list, lru) {
1396        trace_mm_page_free_batched(page, cold);
1397        free_hot_cold_page(page, cold);
1398    }
1399}
1400
1401/*
1402 * split_page takes a non-compound higher-order page, and splits it into
1403 * n (1<<order) sub-pages: page[0..n]
1404 * Each sub-page must be freed individually.
1405 *
1406 * Note: this is probably too low level an operation for use in drivers.
1407 * Please consult with lkml before using this in your driver.
1408 */
1409void split_page(struct page *page, unsigned int order)
1410{
1411    int i;
1412
1413    VM_BUG_ON(PageCompound(page));
1414    VM_BUG_ON(!page_count(page));
1415
1416#ifdef CONFIG_KMEMCHECK
1417    /*
1418     * Split shadow pages too, because free(page[0]) would
1419     * otherwise free the whole shadow.
1420     */
1421    if (kmemcheck_page_is_tracked(page))
1422        split_page(virt_to_page(page[0].shadow), order);
1423#endif
1424
1425    for (i = 1; i < (1 << order); i++)
1426        set_page_refcounted(page + i);
1427}
1428EXPORT_SYMBOL_GPL(split_page);
1429
1430static int __isolate_free_page(struct page *page, unsigned int order)
1431{
1432    unsigned long watermark;
1433    struct zone *zone;
1434    int mt;
1435
1436    BUG_ON(!PageBuddy(page));
1437
1438    zone = page_zone(page);
1439    mt = get_pageblock_migratetype(page);
1440
1441    if (!is_migrate_isolate(mt)) {
1442        /* Obey watermarks as if the page was being allocated */
1443        watermark = low_wmark_pages(zone) + (1 << order);
1444        if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1445            return 0;
1446
1447        __mod_zone_freepage_state(zone, -(1UL << order), mt);
1448    }
1449
1450    /* Remove page from free list */
1451    list_del(&page->lru);
1452    zone->free_area[order].nr_free--;
1453    rmv_page_order(page);
1454
1455    /* Set the pageblock if the isolated page is at least a pageblock */
1456    if (order >= pageblock_order - 1) {
1457        struct page *endpage = page + (1 << order) - 1;
1458        for (; page < endpage; page += pageblock_nr_pages) {
1459            int mt = get_pageblock_migratetype(page);
1460            if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
1461                set_pageblock_migratetype(page,
1462                              MIGRATE_MOVABLE);
1463        }
1464    }
1465
1466    return 1UL << order;
1467}
1468
1469/*
1470 * Similar to split_page except the page is already free. As this is only
1471 * being used for migration, the migratetype of the block also changes.
1472 * As this is called with interrupts disabled, the caller is responsible
1473 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1474 * are enabled.
1475 *
1476 * Note: this is probably too low level an operation for use in drivers.
1477 * Please consult with lkml before using this in your driver.
1478 */
1479int split_free_page(struct page *page)
1480{
1481    unsigned int order;
1482    int nr_pages;
1483
1484    order = page_order(page);
1485
1486    nr_pages = __isolate_free_page(page, order);
1487    if (!nr_pages)
1488        return 0;
1489
1490    /* Split into individual pages */
1491    set_page_refcounted(page);
1492    split_page(page, order);
1493    return nr_pages;
1494}
1495
1496/*
1497 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1498 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1499 * or two.
1500 */
1501static inline
1502struct page *buffered_rmqueue(struct zone *preferred_zone,
1503            struct zone *zone, int order, gfp_t gfp_flags,
1504            int migratetype)
1505{
1506    unsigned long flags;
1507    struct page *page;
1508    int cold = !!(gfp_flags & __GFP_COLD);
1509
1510again:
1511    if (likely(order == 0)) {
1512        struct per_cpu_pages *pcp;
1513        struct list_head *list;
1514
1515        local_irq_save(flags);
1516        pcp = &this_cpu_ptr(zone->pageset)->pcp;
1517        list = &pcp->lists[migratetype];
1518        if (list_empty(list)) {
1519            pcp->count += rmqueue_bulk(zone, 0,
1520                    pcp->batch, list,
1521                    migratetype, cold);
1522            if (unlikely(list_empty(list)))
1523                goto failed;
1524        }
1525
1526        if (cold)
1527            page = list_entry(list->prev, struct page, lru);
1528        else
1529            page = list_entry(list->next, struct page, lru);
1530
1531        list_del(&page->lru);
1532        pcp->count--;
1533    } else {
1534        if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1535            /*
1536             * __GFP_NOFAIL is not to be used in new code.
1537             *
1538             * All __GFP_NOFAIL callers should be fixed so that they
1539             * properly detect and handle allocation failures.
1540             *
1541             * We most definitely don't want callers attempting to
1542             * allocate greater than order-1 page units with
1543             * __GFP_NOFAIL.
1544             */
1545            WARN_ON_ONCE(order > 1);
1546        }
1547        spin_lock_irqsave(&zone->lock, flags);
1548        page = __rmqueue(zone, order, migratetype);
1549        spin_unlock(&zone->lock);
1550        if (!page)
1551            goto failed;
1552        __mod_zone_freepage_state(zone, -(1 << order),
1553                      get_pageblock_migratetype(page));
1554    }
1555
1556    __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
1557    __count_zone_vm_events(PGALLOC, zone, 1 << order);
1558    zone_statistics(preferred_zone, zone, gfp_flags);
1559    local_irq_restore(flags);
1560
1561    VM_BUG_ON(bad_range(zone, page));
1562    if (prep_new_page(page, order, gfp_flags))
1563        goto again;
1564    return page;
1565
1566failed:
1567    local_irq_restore(flags);
1568    return NULL;
1569}
1570
1571#ifdef CONFIG_FAIL_PAGE_ALLOC
1572
1573static struct {
1574    struct fault_attr attr;
1575
1576    u32 ignore_gfp_highmem;
1577    u32 ignore_gfp_wait;
1578    u32 min_order;
1579} fail_page_alloc = {
1580    .attr = FAULT_ATTR_INITIALIZER,
1581    .ignore_gfp_wait = 1,
1582    .ignore_gfp_highmem = 1,
1583    .min_order = 1,
1584};
1585
1586static int __init setup_fail_page_alloc(char *str)
1587{
1588    return setup_fault_attr(&fail_page_alloc.attr, str);
1589}
1590__setup("fail_page_alloc=", setup_fail_page_alloc);
1591
1592static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1593{
1594    if (order < fail_page_alloc.min_order)
1595        return false;
1596    if (gfp_mask & __GFP_NOFAIL)
1597        return false;
1598    if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1599        return false;
1600    if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1601        return false;
1602
1603    return should_fail(&fail_page_alloc.attr, 1 << order);
1604}
1605
1606#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1607
1608static int __init fail_page_alloc_debugfs(void)
1609{
1610    umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1611    struct dentry *dir;
1612
1613    dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1614                    &fail_page_alloc.attr);
1615    if (IS_ERR(dir))
1616        return PTR_ERR(dir);
1617
1618    if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1619                &fail_page_alloc.ignore_gfp_wait))
1620        goto fail;
1621    if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1622                &fail_page_alloc.ignore_gfp_highmem))
1623        goto fail;
1624    if (!debugfs_create_u32("min-order", mode, dir,
1625                &fail_page_alloc.min_order))
1626        goto fail;
1627
1628    return 0;
1629fail:
1630    debugfs_remove_recursive(dir);
1631
1632    return -ENOMEM;
1633}
1634
1635late_initcall(fail_page_alloc_debugfs);
1636
1637#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1638
1639#else /* CONFIG_FAIL_PAGE_ALLOC */
1640
1641static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1642{
1643    return false;
1644}
1645
1646#endif /* CONFIG_FAIL_PAGE_ALLOC */
1647
1648/*
1649 * Return true if free pages are above 'mark'. This takes into account the order
1650 * of the allocation.
1651 */
1652static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1653              int classzone_idx, int alloc_flags, long free_pages)
1654{
1655    /* free_pages my go negative - that's OK */
1656    long min = mark;
1657    long lowmem_reserve = z->lowmem_reserve[classzone_idx];
1658    int o;
1659    long free_cma = 0;
1660
1661    free_pages -= (1 << order) - 1;
1662    if (alloc_flags & ALLOC_HIGH)
1663        min -= min / 2;
1664    if (alloc_flags & ALLOC_HARDER)
1665        min -= min / 4;
1666#ifdef CONFIG_CMA
1667    /* If allocation can't use CMA areas don't use free CMA pages */
1668    if (!(alloc_flags & ALLOC_CMA))
1669        free_cma = zone_page_state(z, NR_FREE_CMA_PAGES);
1670#endif
1671
1672    if (free_pages - free_cma <= min + lowmem_reserve)
1673        return false;
1674    for (o = 0; o < order; o++) {
1675        /* At the next order, this order's pages become unavailable */
1676        free_pages -= z->free_area[o].nr_free << o;
1677
1678        /* Require fewer higher order pages to be free */
1679        min >>= 1;
1680
1681        if (free_pages <= min)
1682            return false;
1683    }
1684    return true;
1685}
1686
1687bool zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1688              int classzone_idx, int alloc_flags)
1689{
1690    return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1691                    zone_page_state(z, NR_FREE_PAGES));
1692}
1693
1694bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark,
1695              int classzone_idx, int alloc_flags)
1696{
1697    long free_pages = zone_page_state(z, NR_FREE_PAGES);
1698
1699    if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1700        free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1701
1702    return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1703                                free_pages);
1704}
1705
1706#ifdef CONFIG_NUMA
1707/*
1708 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1709 * skip over zones that are not allowed by the cpuset, or that have
1710 * been recently (in last second) found to be nearly full. See further
1711 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1712 * that have to skip over a lot of full or unallowed zones.
1713 *
1714 * If the zonelist cache is present in the passed in zonelist, then
1715 * returns a pointer to the allowed node mask (either the current
1716 * tasks mems_allowed, or node_states[N_MEMORY].)
1717 *
1718 * If the zonelist cache is not available for this zonelist, does
1719 * nothing and returns NULL.
1720 *
1721 * If the fullzones BITMAP in the zonelist cache is stale (more than
1722 * a second since last zap'd) then we zap it out (clear its bits.)
1723 *
1724 * We hold off even calling zlc_setup, until after we've checked the
1725 * first zone in the zonelist, on the theory that most allocations will
1726 * be satisfied from that first zone, so best to examine that zone as
1727 * quickly as we can.
1728 */
1729static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1730{
1731    struct zonelist_cache *zlc; /* cached zonelist speedup info */
1732    nodemask_t *allowednodes; /* zonelist_cache approximation */
1733
1734    zlc = zonelist->zlcache_ptr;
1735    if (!zlc)
1736        return NULL;
1737
1738    if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1739        bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1740        zlc->last_full_zap = jiffies;
1741    }
1742
1743    allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1744                    &cpuset_current_mems_allowed :
1745                    &node_states[N_MEMORY];
1746    return allowednodes;
1747}
1748
1749/*
1750 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1751 * if it is worth looking at further for free memory:
1752 * 1) Check that the zone isn't thought to be full (doesn't have its
1753 * bit set in the zonelist_cache fullzones BITMAP).
1754 * 2) Check that the zones node (obtained from the zonelist_cache
1755 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1756 * Return true (non-zero) if zone is worth looking at further, or
1757 * else return false (zero) if it is not.
1758 *
1759 * This check -ignores- the distinction between various watermarks,
1760 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1761 * found to be full for any variation of these watermarks, it will
1762 * be considered full for up to one second by all requests, unless
1763 * we are so low on memory on all allowed nodes that we are forced
1764 * into the second scan of the zonelist.
1765 *
1766 * In the second scan we ignore this zonelist cache and exactly
1767 * apply the watermarks to all zones, even it is slower to do so.
1768 * We are low on memory in the second scan, and should leave no stone
1769 * unturned looking for a free page.
1770 */
1771static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1772                        nodemask_t *allowednodes)
1773{
1774    struct zonelist_cache *zlc; /* cached zonelist speedup info */
1775    int i; /* index of *z in zonelist zones */
1776    int n; /* node that zone *z is on */
1777
1778    zlc = zonelist->zlcache_ptr;
1779    if (!zlc)
1780        return 1;
1781
1782    i = z - zonelist->_zonerefs;
1783    n = zlc->z_to_n[i];
1784
1785    /* This zone is worth trying if it is allowed but not full */
1786    return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1787}
1788
1789/*
1790 * Given 'z' scanning a zonelist, set the corresponding bit in
1791 * zlc->fullzones, so that subsequent attempts to allocate a page
1792 * from that zone don't waste time re-examining it.
1793 */
1794static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1795{
1796    struct zonelist_cache *zlc; /* cached zonelist speedup info */
1797    int i; /* index of *z in zonelist zones */
1798
1799    zlc = zonelist->zlcache_ptr;
1800    if (!zlc)
1801        return;
1802
1803    i = z - zonelist->_zonerefs;
1804
1805    set_bit(i, zlc->fullzones);
1806}
1807
1808/*
1809 * clear all zones full, called after direct reclaim makes progress so that
1810 * a zone that was recently full is not skipped over for up to a second
1811 */
1812static void zlc_clear_zones_full(struct zonelist *zonelist)
1813{
1814    struct zonelist_cache *zlc; /* cached zonelist speedup info */
1815
1816    zlc = zonelist->zlcache_ptr;
1817    if (!zlc)
1818        return;
1819
1820    bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1821}
1822
1823static bool zone_local(struct zone *local_zone, struct zone *zone)
1824{
1825    return node_distance(local_zone->node, zone->node) == LOCAL_DISTANCE;
1826}
1827
1828static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1829{
1830    return node_isset(local_zone->node, zone->zone_pgdat->reclaim_nodes);
1831}
1832
1833static void __paginginit init_zone_allows_reclaim(int nid)
1834{
1835    int i;
1836
1837    for_each_online_node(i)
1838        if (node_distance(nid, i) <= RECLAIM_DISTANCE)
1839            node_set(i, NODE_DATA(nid)->reclaim_nodes);
1840        else
1841            zone_reclaim_mode = 1;
1842}
1843
1844#else /* CONFIG_NUMA */
1845
1846static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1847{
1848    return NULL;
1849}
1850
1851static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1852                nodemask_t *allowednodes)
1853{
1854    return 1;
1855}
1856
1857static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1858{
1859}
1860
1861static void zlc_clear_zones_full(struct zonelist *zonelist)
1862{
1863}
1864
1865static bool zone_local(struct zone *local_zone, struct zone *zone)
1866{
1867    return true;
1868}
1869
1870static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1871{
1872    return true;
1873}
1874
1875static inline void init_zone_allows_reclaim(int nid)
1876{
1877}
1878#endif /* CONFIG_NUMA */
1879
1880/*
1881 * get_page_from_freelist goes through the zonelist trying to allocate
1882 * a page.
1883 */
1884static struct page *
1885get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1886        struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1887        struct zone *preferred_zone, int migratetype)
1888{
1889    struct zoneref *z;
1890    struct page *page = NULL;
1891    int classzone_idx;
1892    struct zone *zone;
1893    nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1894    int zlc_active = 0; /* set if using zonelist_cache */
1895    int did_zlc_setup = 0; /* just call zlc_setup() one time */
1896
1897    classzone_idx = zone_idx(preferred_zone);
1898zonelist_scan:
1899    /*
1900     * Scan zonelist, looking for a zone with enough free.
1901     * See also __cpuset_node_allowed_softwall() comment in kernel/cpuset.c.
1902     */
1903    for_each_zone_zonelist_nodemask(zone, z, zonelist,
1904                        high_zoneidx, nodemask) {
1905        unsigned long mark;
1906
1907        if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
1908            !zlc_zone_worth_trying(zonelist, z, allowednodes))
1909                continue;
1910        if ((alloc_flags & ALLOC_CPUSET) &&
1911            !cpuset_zone_allowed_softwall(zone, gfp_mask))
1912                continue;
1913        BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1914        if (unlikely(alloc_flags & ALLOC_NO_WATERMARKS))
1915            goto try_this_zone;
1916        /*
1917         * Distribute pages in proportion to the individual
1918         * zone size to ensure fair page aging. The zone a
1919         * page was allocated in should have no effect on the
1920         * time the page has in memory before being reclaimed.
1921         *
1922         * When zone_reclaim_mode is enabled, try to stay in
1923         * local zones in the fastpath. If that fails, the
1924         * slowpath is entered, which will do another pass
1925         * starting with the local zones, but ultimately fall
1926         * back to remote zones that do not partake in the
1927         * fairness round-robin cycle of this zonelist.
1928         */
1929        if (alloc_flags & ALLOC_WMARK_LOW) {
1930            if (zone_page_state(zone, NR_ALLOC_BATCH) <= 0)
1931                continue;
1932            if (zone_reclaim_mode &&
1933                !zone_local(preferred_zone, zone))
1934                continue;
1935        }
1936        /*
1937         * When allocating a page cache page for writing, we
1938         * want to get it from a zone that is within its dirty
1939         * limit, such that no single zone holds more than its
1940         * proportional share of globally allowed dirty pages.
1941         * The dirty limits take into account the zone's
1942         * lowmem reserves and high watermark so that kswapd
1943         * should be able to balance it without having to
1944         * write pages from its LRU list.
1945         *
1946         * This may look like it could increase pressure on
1947         * lower zones by failing allocations in higher zones
1948         * before they are full. But the pages that do spill
1949         * over are limited as the lower zones are protected
1950         * by this very same mechanism. It should not become
1951         * a practical burden to them.
1952         *
1953         * XXX: For now, allow allocations to potentially
1954         * exceed the per-zone dirty limit in the slowpath
1955         * (ALLOC_WMARK_LOW unset) before going into reclaim,
1956         * which is important when on a NUMA setup the allowed
1957         * zones are together not big enough to reach the
1958         * global limit. The proper fix for these situations
1959         * will require awareness of zones in the
1960         * dirty-throttling and the flusher threads.
1961         */
1962        if ((alloc_flags & ALLOC_WMARK_LOW) &&
1963            (gfp_mask & __GFP_WRITE) && !zone_dirty_ok(zone))
1964            goto this_zone_full;
1965
1966        mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1967        if (!zone_watermark_ok(zone, order, mark,
1968                       classzone_idx, alloc_flags)) {
1969            int ret;
1970
1971            if (IS_ENABLED(CONFIG_NUMA) &&
1972                    !did_zlc_setup && nr_online_nodes > 1) {
1973                /*
1974                 * we do zlc_setup if there are multiple nodes
1975                 * and before considering the first zone allowed
1976                 * by the cpuset.
1977                 */
1978                allowednodes = zlc_setup(zonelist, alloc_flags);
1979                zlc_active = 1;
1980                did_zlc_setup = 1;
1981            }
1982
1983            if (zone_reclaim_mode == 0 ||
1984                !zone_allows_reclaim(preferred_zone, zone))
1985                goto this_zone_full;
1986
1987            /*
1988             * As we may have just activated ZLC, check if the first
1989             * eligible zone has failed zone_reclaim recently.
1990             */
1991            if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
1992                !zlc_zone_worth_trying(zonelist, z, allowednodes))
1993                continue;
1994
1995            ret = zone_reclaim(zone, gfp_mask, order);
1996            switch (ret) {
1997            case ZONE_RECLAIM_NOSCAN:
1998                /* did not scan */
1999                continue;
2000            case ZONE_RECLAIM_FULL:
2001                /* scanned but unreclaimable */
2002                continue;
2003            default:
2004                /* did we reclaim enough */
2005                if (zone_watermark_ok(zone, order, mark,
2006                        classzone_idx, alloc_flags))
2007                    goto try_this_zone;
2008
2009                /*
2010                 * Failed to reclaim enough to meet watermark.
2011                 * Only mark the zone full if checking the min
2012                 * watermark or if we failed to reclaim just
2013                 * 1<<order pages or else the page allocator
2014                 * fastpath will prematurely mark zones full
2015                 * when the watermark is between the low and
2016                 * min watermarks.
2017                 */
2018                if (((alloc_flags & ALLOC_WMARK_MASK) == ALLOC_WMARK_MIN) ||
2019                    ret == ZONE_RECLAIM_SOME)
2020                    goto this_zone_full;
2021
2022                continue;
2023            }
2024        }
2025
2026try_this_zone:
2027        page = buffered_rmqueue(preferred_zone, zone, order,
2028                        gfp_mask, migratetype);
2029        if (page)
2030            break;
2031this_zone_full:
2032        if (IS_ENABLED(CONFIG_NUMA))
2033            zlc_mark_zone_full(zonelist, z);
2034    }
2035
2036    if (unlikely(IS_ENABLED(CONFIG_NUMA) && page == NULL && zlc_active)) {
2037        /* Disable zlc cache for second zonelist scan */
2038        zlc_active = 0;
2039        goto zonelist_scan;
2040    }
2041
2042    if (page)
2043        /*
2044         * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was
2045         * necessary to allocate the page. The expectation is
2046         * that the caller is taking steps that will free more
2047         * memory. The caller should avoid the page being used
2048         * for !PFMEMALLOC purposes.
2049         */
2050        page->pfmemalloc = !!(alloc_flags & ALLOC_NO_WATERMARKS);
2051
2052    return page;
2053}
2054
2055/*
2056 * Large machines with many possible nodes should not always dump per-node
2057 * meminfo in irq context.
2058 */
2059static inline bool should_suppress_show_mem(void)
2060{
2061    bool ret = false;
2062
2063#if NODES_SHIFT > 8
2064    ret = in_interrupt();
2065#endif
2066    return ret;
2067}
2068
2069static DEFINE_RATELIMIT_STATE(nopage_rs,
2070        DEFAULT_RATELIMIT_INTERVAL,
2071        DEFAULT_RATELIMIT_BURST);
2072
2073void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
2074{
2075    unsigned int filter = SHOW_MEM_FILTER_NODES;
2076
2077    if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2078        debug_guardpage_minorder() > 0)
2079        return;
2080
2081    /*
2082     * Walking all memory to count page types is very expensive and should
2083     * be inhibited in non-blockable contexts.
2084     */
2085    if (!(gfp_mask & __GFP_WAIT))
2086        filter |= SHOW_MEM_FILTER_PAGE_COUNT;
2087
2088    /*
2089     * This documents exceptions given to allocations in certain
2090     * contexts that are allowed to allocate outside current's set
2091     * of allowed nodes.
2092     */
2093    if (!(gfp_mask & __GFP_NOMEMALLOC))
2094        if (test_thread_flag(TIF_MEMDIE) ||
2095            (current->flags & (PF_MEMALLOC | PF_EXITING)))
2096            filter &= ~SHOW_MEM_FILTER_NODES;
2097    if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
2098        filter &= ~SHOW_MEM_FILTER_NODES;
2099
2100    if (fmt) {
2101        struct va_format vaf;
2102        va_list args;
2103
2104        va_start(args, fmt);
2105
2106        vaf.fmt = fmt;
2107        vaf.va = &args;
2108
2109        pr_warn("%pV", &vaf);
2110
2111        va_end(args);
2112    }
2113
2114    pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2115        current->comm, order, gfp_mask);
2116
2117    dump_stack();
2118    if (!should_suppress_show_mem())
2119        show_mem(filter);
2120}
2121
2122static inline int
2123should_alloc_retry(gfp_t gfp_mask, unsigned int order,
2124                unsigned long did_some_progress,
2125                unsigned long pages_reclaimed)
2126{
2127    /* Do not loop if specifically requested */
2128    if (gfp_mask & __GFP_NORETRY)
2129        return 0;
2130
2131    /* Always retry if specifically requested */
2132    if (gfp_mask & __GFP_NOFAIL)
2133        return 1;
2134
2135    /*
2136     * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2137     * making forward progress without invoking OOM. Suspend also disables
2138     * storage devices so kswapd will not help. Bail if we are suspending.
2139     */
2140    if (!did_some_progress && pm_suspended_storage())
2141        return 0;
2142
2143    /*
2144     * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2145     * means __GFP_NOFAIL, but that may not be true in other
2146     * implementations.
2147     */
2148    if (order <= PAGE_ALLOC_COSTLY_ORDER)
2149        return 1;
2150
2151    /*
2152     * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2153     * specified, then we retry until we no longer reclaim any pages
2154     * (above), or we've reclaimed an order of pages at least as
2155     * large as the allocation's order. In both cases, if the
2156     * allocation still fails, we stop retrying.
2157     */
2158    if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
2159        return 1;
2160
2161    return 0;
2162}
2163
2164static inline struct page *
2165__alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2166    struct zonelist *zonelist, enum zone_type high_zoneidx,
2167    nodemask_t *nodemask, struct zone *preferred_zone,
2168    int migratetype)
2169{
2170    struct page *page;
2171
2172    /* Acquire the OOM killer lock for the zones in zonelist */
2173    if (!try_set_zonelist_oom(zonelist, gfp_mask)) {
2174        schedule_timeout_uninterruptible(1);
2175        return NULL;
2176    }
2177
2178    /*
2179     * Go through the zonelist yet one more time, keep very high watermark
2180     * here, this is only to catch a parallel oom killing, we must fail if
2181     * we're still under heavy pressure.
2182     */
2183    page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
2184        order, zonelist, high_zoneidx,
2185        ALLOC_WMARK_HIGH|ALLOC_CPUSET,
2186        preferred_zone, migratetype);
2187    if (page)
2188        goto out;
2189
2190    if (!(gfp_mask & __GFP_NOFAIL)) {
2191        /* The OOM killer will not help higher order allocs */
2192        if (order > PAGE_ALLOC_COSTLY_ORDER)
2193            goto out;
2194        /* The OOM killer does not needlessly kill tasks for lowmem */
2195        if (high_zoneidx < ZONE_NORMAL)
2196            goto out;
2197        /*
2198         * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2199         * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2200         * The caller should handle page allocation failure by itself if
2201         * it specifies __GFP_THISNODE.
2202         * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2203         */
2204        if (gfp_mask & __GFP_THISNODE)
2205            goto out;
2206    }
2207    /* Exhausted what can be done so it's blamo time */
2208    out_of_memory(zonelist, gfp_mask, order, nodemask, false);
2209
2210out:
2211    clear_zonelist_oom(zonelist, gfp_mask);
2212    return page;
2213}
2214
2215#ifdef CONFIG_COMPACTION
2216/* Try memory compaction for high-order allocations before reclaim */
2217static struct page *
2218__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2219    struct zonelist *zonelist, enum zone_type high_zoneidx,
2220    nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2221    int migratetype, bool sync_migration,
2222    bool *contended_compaction, bool *deferred_compaction,
2223    unsigned long *did_some_progress)
2224{
2225    if (!order)
2226        return NULL;
2227
2228    if (compaction_deferred(preferred_zone, order)) {
2229        *deferred_compaction = true;
2230        return NULL;
2231    }
2232
2233    current->flags |= PF_MEMALLOC;
2234    *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
2235                        nodemask, sync_migration,
2236                        contended_compaction);
2237    current->flags &= ~PF_MEMALLOC;
2238
2239    if (*did_some_progress != COMPACT_SKIPPED) {
2240        struct page *page;
2241
2242        /* Page migration frees to the PCP lists but we want merging */
2243        drain_pages(get_cpu());
2244        put_cpu();
2245
2246        page = get_page_from_freelist(gfp_mask, nodemask,
2247                order, zonelist, high_zoneidx,
2248                alloc_flags & ~ALLOC_NO_WATERMARKS,
2249                preferred_zone, migratetype);
2250        if (page) {
2251            preferred_zone->compact_blockskip_flush = false;
2252            preferred_zone->compact_considered = 0;
2253            preferred_zone->compact_defer_shift = 0;
2254            if (order >= preferred_zone->compact_order_failed)
2255                preferred_zone->compact_order_failed = order + 1;
2256            count_vm_event(COMPACTSUCCESS);
2257            return page;
2258        }
2259
2260        /*
2261         * It's bad if compaction run occurs and fails.
2262         * The most likely reason is that pages exist,
2263         * but not enough to satisfy watermarks.
2264         */
2265        count_vm_event(COMPACTFAIL);
2266
2267        /*
2268         * As async compaction considers a subset of pageblocks, only
2269         * defer if the failure was a sync compaction failure.
2270         */
2271        if (sync_migration)
2272            defer_compaction(preferred_zone, order);
2273
2274        cond_resched();
2275    }
2276
2277    return NULL;
2278}
2279#else
2280static inline struct page *
2281__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2282    struct zonelist *zonelist, enum zone_type high_zoneidx,
2283    nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2284    int migratetype, bool sync_migration,
2285    bool *contended_compaction, bool *deferred_compaction,
2286    unsigned long *did_some_progress)
2287{
2288    return NULL;
2289}
2290#endif /* CONFIG_COMPACTION */
2291
2292/* Perform direct synchronous page reclaim */
2293static int
2294__perform_reclaim(gfp_t gfp_mask, unsigned int order, struct zonelist *zonelist,
2295          nodemask_t *nodemask)
2296{
2297    struct reclaim_state reclaim_state;
2298    int progress;
2299
2300    cond_resched();
2301
2302    /* We now go into synchronous reclaim */
2303    cpuset_memory_pressure_bump();
2304    current->flags |= PF_MEMALLOC;
2305    lockdep_set_current_reclaim_state(gfp_mask);
2306    reclaim_state.reclaimed_slab = 0;
2307    current->reclaim_state = &reclaim_state;
2308
2309    progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
2310
2311    current->reclaim_state = NULL;
2312    lockdep_clear_current_reclaim_state();
2313    current->flags &= ~PF_MEMALLOC;
2314
2315    cond_resched();
2316
2317    return progress;
2318}
2319
2320/* The really slow allocator path where we enter direct reclaim */
2321static inline struct page *
2322__alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2323    struct zonelist *zonelist, enum zone_type high_zoneidx,
2324    nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2325    int migratetype, unsigned long *did_some_progress)
2326{
2327    struct page *page = NULL;
2328    bool drained = false;
2329
2330    *did_some_progress = __perform_reclaim(gfp_mask, order, zonelist,
2331                           nodemask);
2332    if (unlikely(!(*did_some_progress)))
2333        return NULL;
2334
2335    /* After successful reclaim, reconsider all zones for allocation */
2336    if (IS_ENABLED(CONFIG_NUMA))
2337        zlc_clear_zones_full(zonelist);
2338
2339retry:
2340    page = get_page_from_freelist(gfp_mask, nodemask, order,
2341                    zonelist, high_zoneidx,
2342                    alloc_flags & ~ALLOC_NO_WATERMARKS,
2343                    preferred_zone, migratetype);
2344
2345    /*
2346     * If an allocation failed after direct reclaim, it could be because
2347     * pages are pinned on the per-cpu lists. Drain them and try again
2348     */
2349    if (!page && !drained) {
2350        drain_all_pages();
2351        drained = true;
2352        goto retry;
2353    }
2354
2355    return page;
2356}
2357
2358/*
2359 * This is called in the allocator slow-path if the allocation request is of
2360 * sufficient urgency to ignore watermarks and take other desperate measures
2361 */
2362static inline struct page *
2363__alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2364    struct zonelist *zonelist, enum zone_type high_zoneidx,
2365    nodemask_t *nodemask, struct zone *preferred_zone,
2366    int migratetype)
2367{
2368    struct page *page;
2369
2370    do {
2371        page = get_page_from_freelist(gfp_mask, nodemask, order,
2372            zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
2373            preferred_zone, migratetype);
2374
2375        if (!page && gfp_mask & __GFP_NOFAIL)
2376            wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2377    } while (!page && (gfp_mask & __GFP_NOFAIL));
2378
2379    return page;
2380}
2381
2382static void prepare_slowpath(gfp_t gfp_mask, unsigned int order,
2383                 struct zonelist *zonelist,
2384                 enum zone_type high_zoneidx,
2385                 struct zone *preferred_zone)
2386{
2387    struct zoneref *z;
2388    struct zone *zone;
2389
2390    for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
2391        if (!(gfp_mask & __GFP_NO_KSWAPD))
2392            wakeup_kswapd(zone, order, zone_idx(preferred_zone));
2393        /*
2394         * Only reset the batches of zones that were actually
2395         * considered in the fast path, we don't want to
2396         * thrash fairness information for zones that are not
2397         * actually part of this zonelist's round-robin cycle.
2398         */
2399        if (zone_reclaim_mode && !zone_local(preferred_zone, zone))
2400            continue;
2401        mod_zone_page_state(zone, NR_ALLOC_BATCH,
2402                    high_wmark_pages(zone) -
2403                    low_wmark_pages(zone) -
2404                    zone_page_state(zone, NR_ALLOC_BATCH));
2405    }
2406}
2407
2408static inline int
2409gfp_to_alloc_flags(gfp_t gfp_mask)
2410{
2411    int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2412    const gfp_t wait = gfp_mask & __GFP_WAIT;
2413
2414    /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2415    BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2416
2417    /*
2418     * The caller may dip into page reserves a bit more if the caller
2419     * cannot run direct reclaim, or if the caller has realtime scheduling
2420     * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2421     * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2422     */
2423    alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2424
2425    if (!wait) {
2426        /*
2427         * Not worth trying to allocate harder for
2428         * __GFP_NOMEMALLOC even if it can't schedule.
2429         */
2430        if (!(gfp_mask & __GFP_NOMEMALLOC))
2431            alloc_flags |= ALLOC_HARDER;
2432        /*
2433         * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2434         * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
2435         */
2436        alloc_flags &= ~ALLOC_CPUSET;
2437    } else if (unlikely(rt_task(current)) && !in_interrupt())
2438        alloc_flags |= ALLOC_HARDER;
2439
2440    if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2441        if (gfp_mask & __GFP_MEMALLOC)
2442            alloc_flags |= ALLOC_NO_WATERMARKS;
2443        else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2444            alloc_flags |= ALLOC_NO_WATERMARKS;
2445        else if (!in_interrupt() &&
2446                ((current->flags & PF_MEMALLOC) ||
2447                 unlikely(test_thread_flag(TIF_MEMDIE))))
2448            alloc_flags |= ALLOC_NO_WATERMARKS;
2449    }
2450#ifdef CONFIG_CMA
2451    if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2452        alloc_flags |= ALLOC_CMA;
2453#endif
2454    return alloc_flags;
2455}
2456
2457bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2458{
2459    return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2460}
2461
2462static inline struct page *
2463__alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2464    struct zonelist *zonelist, enum zone_type high_zoneidx,
2465    nodemask_t *nodemask, struct zone *preferred_zone,
2466    int migratetype)
2467{
2468    const gfp_t wait = gfp_mask & __GFP_WAIT;
2469    struct page *page = NULL;
2470    int alloc_flags;
2471    unsigned long pages_reclaimed = 0;
2472    unsigned long did_some_progress;
2473    bool sync_migration = false;
2474    bool deferred_compaction = false;
2475    bool contended_compaction = false;
2476
2477    /*
2478     * In the slowpath, we sanity check order to avoid ever trying to
2479     * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2480     * be using allocators in order of preference for an area that is
2481     * too large.
2482     */
2483    if (order >= MAX_ORDER) {
2484        WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2485        return NULL;
2486    }
2487
2488    /*
2489     * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2490     * __GFP_NOWARN set) should not cause reclaim since the subsystem
2491     * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2492     * using a larger set of nodes after it has established that the
2493     * allowed per node queues are empty and that nodes are
2494     * over allocated.
2495     */
2496    if (IS_ENABLED(CONFIG_NUMA) &&
2497            (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2498        goto nopage;
2499
2500restart:
2501    prepare_slowpath(gfp_mask, order, zonelist,
2502             high_zoneidx, preferred_zone);
2503
2504    /*
2505     * OK, we're below the kswapd watermark and have kicked background
2506     * reclaim. Now things get more complex, so set up alloc_flags according
2507     * to how we want to proceed.
2508     */
2509    alloc_flags = gfp_to_alloc_flags(gfp_mask);
2510
2511    /*
2512     * Find the true preferred zone if the allocation is unconstrained by
2513     * cpusets.
2514     */
2515    if (!(alloc_flags & ALLOC_CPUSET) && !nodemask)
2516        first_zones_zonelist(zonelist, high_zoneidx, NULL,
2517                    &preferred_zone);
2518
2519rebalance:
2520    /* This is the last chance, in general, before the goto nopage. */
2521    page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2522            high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2523            preferred_zone, migratetype);
2524    if (page)
2525        goto got_pg;
2526
2527    /* Allocate without watermarks if the context allows */
2528    if (alloc_flags & ALLOC_NO_WATERMARKS) {
2529        /*
2530         * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2531         * the allocation is high priority and these type of
2532         * allocations are system rather than user orientated
2533         */
2534        zonelist = node_zonelist(numa_node_id(), gfp_mask);
2535
2536        page = __alloc_pages_high_priority(gfp_mask, order,
2537                zonelist, high_zoneidx, nodemask,
2538                preferred_zone, migratetype);
2539        if (page) {
2540            goto got_pg;
2541        }
2542    }
2543
2544    /* Atomic allocations - we can't balance anything */
2545    if (!wait)
2546        goto nopage;
2547
2548    /* Avoid recursion of direct reclaim */
2549    if (current->flags & PF_MEMALLOC)
2550        goto nopage;
2551
2552    /* Avoid allocations with no watermarks from looping endlessly */
2553    if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2554        goto nopage;
2555
2556    /*
2557     * Try direct compaction. The first pass is asynchronous. Subsequent
2558     * attempts after direct reclaim are synchronous
2559     */
2560    page = __alloc_pages_direct_compact(gfp_mask, order,
2561                    zonelist, high_zoneidx,
2562                    nodemask,
2563                    alloc_flags, preferred_zone,
2564                    migratetype, sync_migration,
2565                    &contended_compaction,
2566                    &deferred_compaction,
2567                    &did_some_progress);
2568    if (page)
2569        goto got_pg;
2570    sync_migration = true;
2571
2572    /*
2573     * If compaction is deferred for high-order allocations, it is because
2574     * sync compaction recently failed. In this is the case and the caller
2575     * requested a movable allocation that does not heavily disrupt the
2576     * system then fail the allocation instead of entering direct reclaim.
2577     */
2578    if ((deferred_compaction || contended_compaction) &&
2579                        (gfp_mask & __GFP_NO_KSWAPD))
2580        goto nopage;
2581
2582    /* Try direct reclaim and then allocating */
2583    page = __alloc_pages_direct_reclaim(gfp_mask, order,
2584                    zonelist, high_zoneidx,
2585                    nodemask,
2586                    alloc_flags, preferred_zone,
2587                    migratetype, &did_some_progress);
2588    if (page)
2589        goto got_pg;
2590
2591    /*
2592     * If we failed to make any progress reclaiming, then we are
2593     * running out of options and have to consider going OOM
2594     */
2595    if (!did_some_progress) {
2596        if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
2597            if (oom_killer_disabled)
2598                goto nopage;
2599            /* Coredumps can quickly deplete all memory reserves */
2600            if ((current->flags & PF_DUMPCORE) &&
2601                !(gfp_mask & __GFP_NOFAIL))
2602                goto nopage;
2603            page = __alloc_pages_may_oom(gfp_mask, order,
2604                    zonelist, high_zoneidx,
2605                    nodemask, preferred_zone,
2606                    migratetype);
2607            if (page)
2608                goto got_pg;
2609
2610            if (!(gfp_mask & __GFP_NOFAIL)) {
2611                /*
2612                 * The oom killer is not called for high-order
2613                 * allocations that may fail, so if no progress
2614                 * is being made, there are no other options and
2615                 * retrying is unlikely to help.
2616                 */
2617                if (order > PAGE_ALLOC_COSTLY_ORDER)
2618                    goto nopage;
2619                /*
2620                 * The oom killer is not called for lowmem
2621                 * allocations to prevent needlessly killing
2622                 * innocent tasks.
2623                 */
2624                if (high_zoneidx < ZONE_NORMAL)
2625                    goto nopage;
2626            }
2627
2628            goto restart;
2629        }
2630    }
2631
2632    /* Check if we should retry the allocation */
2633    pages_reclaimed += did_some_progress;
2634    if (should_alloc_retry(gfp_mask, order, did_some_progress,
2635                        pages_reclaimed)) {
2636        /* Wait for some write requests to complete then retry */
2637        wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2638        goto rebalance;
2639    } else {
2640        /*
2641         * High-order allocations do not necessarily loop after
2642         * direct reclaim and reclaim/compaction depends on compaction
2643         * being called after reclaim so call directly if necessary
2644         */
2645        page = __alloc_pages_direct_compact(gfp_mask, order,
2646                    zonelist, high_zoneidx,
2647                    nodemask,
2648                    alloc_flags, preferred_zone,
2649                    migratetype, sync_migration,
2650                    &contended_compaction,
2651                    &deferred_compaction,
2652                    &did_some_progress);
2653        if (page)
2654            goto got_pg;
2655    }
2656
2657nopage:
2658    warn_alloc_failed(gfp_mask, order, NULL);
2659    return page;
2660got_pg:
2661    if (kmemcheck_enabled)
2662        kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2663
2664    return page;
2665}
2666
2667/*
2668 * This is the 'heart' of the zoned buddy allocator.
2669 */
2670struct page *
2671__alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2672            struct zonelist *zonelist, nodemask_t *nodemask)
2673{
2674    enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2675    struct zone *preferred_zone;
2676    struct page *page = NULL;
2677    int migratetype = allocflags_to_migratetype(gfp_mask);
2678    unsigned int cpuset_mems_cookie;
2679    int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET;
2680    struct mem_cgroup *memcg = NULL;
2681
2682    gfp_mask &= gfp_allowed_mask;
2683
2684    lockdep_trace_alloc(gfp_mask);
2685
2686    might_sleep_if(gfp_mask & __GFP_WAIT);
2687
2688    if (should_fail_alloc_page(gfp_mask, order))
2689        return NULL;
2690
2691    /*
2692     * Check the zones suitable for the gfp_mask contain at least one
2693     * valid zone. It's possible to have an empty zonelist as a result
2694     * of GFP_THISNODE and a memoryless node
2695     */
2696    if (unlikely(!zonelist->_zonerefs->zone))
2697        return NULL;
2698
2699    /*
2700     * Will only have any effect when __GFP_KMEMCG is set. This is
2701     * verified in the (always inline) callee
2702     */
2703    if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
2704        return NULL;
2705
2706retry_cpuset:
2707    cpuset_mems_cookie = get_mems_allowed();
2708
2709    /* The preferred zone is used for statistics later */
2710    first_zones_zonelist(zonelist, high_zoneidx,
2711                nodemask ? : &cpuset_current_mems_allowed,
2712                &preferred_zone);
2713    if (!preferred_zone)
2714        goto out;
2715
2716#ifdef CONFIG_CMA
2717    if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2718        alloc_flags |= ALLOC_CMA;
2719#endif
2720    /* First allocation attempt */
2721    page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2722            zonelist, high_zoneidx, alloc_flags,
2723            preferred_zone, migratetype);
2724    if (unlikely(!page)) {
2725        /*
2726         * Runtime PM, block IO and its error handling path
2727         * can deadlock because I/O on the device might not
2728         * complete.
2729         */
2730        gfp_mask = memalloc_noio_flags(gfp_mask);
2731        page = __alloc_pages_slowpath(gfp_mask, order,
2732                zonelist, high_zoneidx, nodemask,
2733                preferred_zone, migratetype);
2734    }
2735
2736    trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2737
2738out:
2739    /*
2740     * When updating a task's mems_allowed, it is possible to race with
2741     * parallel threads in such a way that an allocation can fail while
2742     * the mask is being updated. If a page allocation is about to fail,
2743     * check if the cpuset changed during allocation and if so, retry.
2744     */
2745    if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2746        goto retry_cpuset;
2747
2748    memcg_kmem_commit_charge(page, memcg, order);
2749
2750    return page;
2751}
2752EXPORT_SYMBOL(__alloc_pages_nodemask);
2753
2754/*
2755 * Common helper functions.
2756 */
2757unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2758{
2759    struct page *page;
2760
2761    /*
2762     * __get_free_pages() returns a 32-bit address, which cannot represent
2763     * a highmem page
2764     */
2765    VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2766
2767    page = alloc_pages(gfp_mask, order);
2768    if (!page)
2769        return 0;
2770    return (unsigned long) page_address(page);
2771}
2772EXPORT_SYMBOL(__get_free_pages);
2773
2774unsigned long get_zeroed_page(gfp_t gfp_mask)
2775{
2776    return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2777}
2778EXPORT_SYMBOL(get_zeroed_page);
2779
2780void __free_pages(struct page *page, unsigned int order)
2781{
2782    if (put_page_testzero(page)) {
2783        if (order == 0)
2784            free_hot_cold_page(page, 0);
2785        else
2786            __free_pages_ok(page, order);
2787    }
2788}
2789
2790EXPORT_SYMBOL(__free_pages);
2791
2792void free_pages(unsigned long addr, unsigned int order)
2793{
2794    if (addr != 0) {
2795        VM_BUG_ON(!virt_addr_valid((void *)addr));
2796        __free_pages(virt_to_page((void *)addr), order);
2797    }
2798}
2799
2800EXPORT_SYMBOL(free_pages);
2801
2802/*
2803 * __free_memcg_kmem_pages and free_memcg_kmem_pages will free
2804 * pages allocated with __GFP_KMEMCG.
2805 *
2806 * Those pages are accounted to a particular memcg, embedded in the
2807 * corresponding page_cgroup. To avoid adding a hit in the allocator to search
2808 * for that information only to find out that it is NULL for users who have no
2809 * interest in that whatsoever, we provide these functions.
2810 *
2811 * The caller knows better which flags it relies on.
2812 */
2813void __free_memcg_kmem_pages(struct page *page, unsigned int order)
2814{
2815    memcg_kmem_uncharge_pages(page, order);
2816    __free_pages(page, order);
2817}
2818
2819void free_memcg_kmem_pages(unsigned long addr, unsigned int order)
2820{
2821    if (addr != 0) {
2822        VM_BUG_ON(!virt_addr_valid((void *)addr));
2823        __free_memcg_kmem_pages(virt_to_page((void *)addr), order);
2824    }
2825}
2826
2827static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
2828{
2829    if (addr) {
2830        unsigned long alloc_end = addr + (PAGE_SIZE << order);
2831        unsigned long used = addr + PAGE_ALIGN(size);
2832
2833        split_page(virt_to_page((void *)addr), order);
2834        while (used < alloc_end) {
2835            free_page(used);
2836            used += PAGE_SIZE;
2837        }
2838    }
2839    return (void *)addr;
2840}
2841
2842/**
2843 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2844 * @size: the number of bytes to allocate
2845 * @gfp_mask: GFP flags for the allocation
2846 *
2847 * This function is similar to alloc_pages(), except that it allocates the
2848 * minimum number of pages to satisfy the request. alloc_pages() can only
2849 * allocate memory in power-of-two pages.
2850 *
2851 * This function is also limited by MAX_ORDER.
2852 *
2853 * Memory allocated by this function must be released by free_pages_exact().
2854 */
2855void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2856{
2857    unsigned int order = get_order(size);
2858    unsigned long addr;
2859
2860    addr = __get_free_pages(gfp_mask, order);
2861    return make_alloc_exact(addr, order, size);
2862}
2863EXPORT_SYMBOL(alloc_pages_exact);
2864
2865/**
2866 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2867 * pages on a node.
2868 * @nid: the preferred node ID where memory should be allocated
2869 * @size: the number of bytes to allocate
2870 * @gfp_mask: GFP flags for the allocation
2871 *
2872 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2873 * back.
2874 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2875 * but is not exact.
2876 */
2877void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
2878{
2879    unsigned order = get_order(size);
2880    struct page *p = alloc_pages_node(nid, gfp_mask, order);
2881    if (!p)
2882        return NULL;
2883    return make_alloc_exact((unsigned long)page_address(p), order, size);
2884}
2885EXPORT_SYMBOL(alloc_pages_exact_nid);
2886
2887/**
2888 * free_pages_exact - release memory allocated via alloc_pages_exact()
2889 * @virt: the value returned by alloc_pages_exact.
2890 * @size: size of allocation, same value as passed to alloc_pages_exact().
2891 *
2892 * Release the memory allocated by a previous call to alloc_pages_exact.
2893 */
2894void free_pages_exact(void *virt, size_t size)
2895{
2896    unsigned long addr = (unsigned long)virt;
2897    unsigned long end = addr + PAGE_ALIGN(size);
2898
2899    while (addr < end) {
2900        free_page(addr);
2901        addr += PAGE_SIZE;
2902    }
2903}
2904EXPORT_SYMBOL(free_pages_exact);
2905
2906/**
2907 * nr_free_zone_pages - count number of pages beyond high watermark
2908 * @offset: The zone index of the highest zone
2909 *
2910 * nr_free_zone_pages() counts the number of counts pages which are beyond the
2911 * high watermark within all zones at or below a given zone index. For each
2912 * zone, the number of pages is calculated as:
2913 * managed_pages - high_pages
2914 */
2915static unsigned long nr_free_zone_pages(int offset)
2916{
2917    struct zoneref *z;
2918    struct zone *zone;
2919
2920    /* Just pick one node, since fallback list is circular */
2921    unsigned long sum = 0;
2922
2923    struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2924
2925    for_each_zone_zonelist(zone, z, zonelist, offset) {
2926        unsigned long size = zone->managed_pages;
2927        unsigned long high = high_wmark_pages(zone);
2928        if (size > high)
2929            sum += size - high;
2930    }
2931
2932    return sum;
2933}
2934
2935/**
2936 * nr_free_buffer_pages - count number of pages beyond high watermark
2937 *
2938 * nr_free_buffer_pages() counts the number of pages which are beyond the high
2939 * watermark within ZONE_DMA and ZONE_NORMAL.
2940 */
2941unsigned long nr_free_buffer_pages(void)
2942{
2943    return nr_free_zone_pages(gfp_zone(GFP_USER));
2944}
2945EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2946
2947/**
2948 * nr_free_pagecache_pages - count number of pages beyond high watermark
2949 *
2950 * nr_free_pagecache_pages() counts the number of pages which are beyond the
2951 * high watermark within all zones.
2952 */
2953unsigned long nr_free_pagecache_pages(void)
2954{
2955    return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2956}
2957
2958static inline void show_node(struct zone *zone)
2959{
2960    if (IS_ENABLED(CONFIG_NUMA))
2961        printk("Node %d ", zone_to_nid(zone));
2962}
2963
2964void si_meminfo(struct sysinfo *val)
2965{
2966    val->totalram = totalram_pages;
2967    val->sharedram = 0;
2968    val->freeram = global_page_state(NR_FREE_PAGES);
2969    val->bufferram = nr_blockdev_pages();
2970    val->totalhigh = totalhigh_pages;
2971    val->freehigh = nr_free_highpages();
2972    val->mem_unit = PAGE_SIZE;
2973}
2974
2975EXPORT_SYMBOL(si_meminfo);
2976
2977#ifdef CONFIG_NUMA
2978void si_meminfo_node(struct sysinfo *val, int nid)
2979{
2980    int zone_type; /* needs to be signed */
2981    unsigned long managed_pages = 0;
2982    pg_data_t *pgdat = NODE_DATA(nid);
2983
2984    for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
2985        managed_pages += pgdat->node_zones[zone_type].managed_pages;
2986    val->totalram = managed_pages;
2987    val->freeram = node_page_state(nid, NR_FREE_PAGES);
2988#ifdef CONFIG_HIGHMEM
2989    val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
2990    val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2991            NR_FREE_PAGES);
2992#else
2993    val->totalhigh = 0;
2994    val->freehigh = 0;
2995#endif
2996    val->mem_unit = PAGE_SIZE;
2997}
2998#endif
2999
3000/*
3001 * Determine whether the node should be displayed or not, depending on whether
3002 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3003 */
3004bool skip_free_areas_node(unsigned int flags, int nid)
3005{
3006    bool ret = false;
3007    unsigned int cpuset_mems_cookie;
3008
3009    if (!(flags & SHOW_MEM_FILTER_NODES))
3010        goto out;
3011
3012    do {
3013        cpuset_mems_cookie = get_mems_allowed();
3014        ret = !node_isset(nid, cpuset_current_mems_allowed);
3015    } while (!put_mems_allowed(cpuset_mems_cookie));
3016out:
3017    return ret;
3018}
3019
3020#define K(x) ((x) << (PAGE_SHIFT-10))
3021
3022static void show_migration_types(unsigned char type)
3023{
3024    static const char types[MIGRATE_TYPES] = {
3025        [MIGRATE_UNMOVABLE] = 'U',
3026        [MIGRATE_RECLAIMABLE] = 'E',
3027        [MIGRATE_MOVABLE] = 'M',
3028        [MIGRATE_RESERVE] = 'R',
3029#ifdef CONFIG_CMA
3030        [MIGRATE_CMA] = 'C',
3031#endif
3032#ifdef CONFIG_MEMORY_ISOLATION
3033        [MIGRATE_ISOLATE] = 'I',
3034#endif
3035    };
3036    char tmp[MIGRATE_TYPES + 1];
3037    char *p = tmp;
3038    int i;
3039
3040    for (i = 0; i < MIGRATE_TYPES; i++) {
3041        if (type & (1 << i))
3042            *p++ = types[i];
3043    }
3044
3045    *p = '\0';
3046    printk("(%s) ", tmp);
3047}
3048
3049/*
3050 * Show free area list (used inside shift_scroll-lock stuff)
3051 * We also calculate the percentage fragmentation. We do this by counting the
3052 * memory on each free list with the exception of the first item on the list.
3053 * Suppresses nodes that are not allowed by current's cpuset if
3054 * SHOW_MEM_FILTER_NODES is passed.
3055 */
3056void show_free_areas(unsigned int filter)
3057{
3058    int cpu;
3059    struct zone *zone;
3060
3061    for_each_populated_zone(zone) {
3062        if (skip_free_areas_node(filter, zone_to_nid(zone)))
3063            continue;
3064        show_node(zone);
3065        printk("%s per-cpu:\n", zone->name);
3066
3067        for_each_online_cpu(cpu) {
3068            struct per_cpu_pageset *pageset;
3069
3070            pageset = per_cpu_ptr(zone->pageset, cpu);
3071
3072            printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
3073                   cpu, pageset->pcp.high,
3074                   pageset->pcp.batch, pageset->pcp.count);
3075        }
3076    }
3077
3078    printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3079        " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3080        " unevictable:%lu"
3081        " dirty:%lu writeback:%lu unstable:%lu\n"
3082        " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3083        " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3084        " free_cma:%lu\n",
3085        global_page_state(NR_ACTIVE_ANON),
3086        global_page_state(NR_INACTIVE_ANON),
3087        global_page_state(NR_ISOLATED_ANON),
3088        global_page_state(NR_ACTIVE_FILE),
3089        global_page_state(NR_INACTIVE_FILE),
3090        global_page_state(NR_ISOLATED_FILE),
3091        global_page_state(NR_UNEVICTABLE),
3092        global_page_state(NR_FILE_DIRTY),
3093        global_page_state(NR_WRITEBACK),
3094        global_page_state(NR_UNSTABLE_NFS),
3095        global_page_state(NR_FREE_PAGES),
3096        global_page_state(NR_SLAB_RECLAIMABLE),
3097        global_page_state(NR_SLAB_UNRECLAIMABLE),
3098        global_page_state(NR_FILE_MAPPED),
3099        global_page_state(NR_SHMEM),
3100        global_page_state(NR_PAGETABLE),
3101        global_page_state(NR_BOUNCE),
3102        global_page_state(NR_FREE_CMA_PAGES));
3103
3104    for_each_populated_zone(zone) {
3105        int i;
3106
3107        if (skip_free_areas_node(filter, zone_to_nid(zone)))
3108            continue;
3109        show_node(zone);
3110        printk("%s"
3111            " free:%lukB"
3112            " min:%lukB"
3113            " low:%lukB"
3114            " high:%lukB"
3115            " active_anon:%lukB"
3116            " inactive_anon:%lukB"
3117            " active_file:%lukB"
3118            " inactive_file:%lukB"
3119            " unevictable:%lukB"
3120            " isolated(anon):%lukB"
3121            " isolated(file):%lukB"
3122            " present:%lukB"
3123            " managed:%lukB"
3124            " mlocked:%lukB"
3125            " dirty:%lukB"
3126            " writeback:%lukB"
3127            " mapped:%lukB"
3128            " shmem:%lukB"
3129            " slab_reclaimable:%lukB"
3130            " slab_unreclaimable:%lukB"
3131            " kernel_stack:%lukB"
3132            " pagetables:%lukB"
3133            " unstable:%lukB"
3134            " bounce:%lukB"
3135            " free_cma:%lukB"
3136            " writeback_tmp:%lukB"
3137            " pages_scanned:%lu"
3138            " all_unreclaimable? %s"
3139            "\n",
3140            zone->name,
3141            K(zone_page_state(zone, NR_FREE_PAGES)),
3142            K(min_wmark_pages(zone)),
3143            K(low_wmark_pages(zone)),
3144            K(high_wmark_pages(zone)),
3145            K(zone_page_state(zone, NR_ACTIVE_ANON)),
3146            K(zone_page_state(zone, NR_INACTIVE_ANON)),
3147            K(zone_page_state(zone, NR_ACTIVE_FILE)),
3148            K(zone_page_state(zone, NR_INACTIVE_FILE)),
3149            K(zone_page_state(zone, NR_UNEVICTABLE)),
3150            K(zone_page_state(zone, NR_ISOLATED_ANON)),
3151            K(zone_page_state(zone, NR_ISOLATED_FILE)),
3152            K(zone->present_pages),
3153            K(zone->managed_pages),
3154            K(zone_page_state(zone, NR_MLOCK)),
3155            K(zone_page_state(zone, NR_FILE_DIRTY)),
3156            K(zone_page_state(zone, NR_WRITEBACK)),
3157            K(zone_page_state(zone, NR_FILE_MAPPED)),
3158            K(zone_page_state(zone, NR_SHMEM)),
3159            K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3160            K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3161            zone_page_state(zone, NR_KERNEL_STACK) *
3162                THREAD_SIZE / 1024,
3163            K(zone_page_state(zone, NR_PAGETABLE)),
3164            K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3165            K(zone_page_state(zone, NR_BOUNCE)),
3166            K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3167            K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3168            zone->pages_scanned,
3169            (!zone_reclaimable(zone) ? "yes" : "no")
3170            );
3171        printk("lowmem_reserve[]:");
3172        for (i = 0; i < MAX_NR_ZONES; i++)
3173            printk(" %lu", zone->lowmem_reserve[i]);
3174        printk("\n");
3175    }
3176
3177    for_each_populated_zone(zone) {
3178        unsigned long nr[MAX_ORDER], flags, order, total = 0;
3179        unsigned char types[MAX_ORDER];
3180
3181        if (skip_free_areas_node(filter, zone_to_nid(zone)))
3182            continue;
3183        show_node(zone);
3184        printk("%s: ", zone->name);
3185
3186        spin_lock_irqsave(&zone->lock, flags);
3187        for (order = 0; order < MAX_ORDER; order++) {
3188            struct free_area *area = &zone->free_area[order];
3189            int type;
3190
3191            nr[order] = area->nr_free;
3192            total += nr[order] << order;
3193
3194            types[order] = 0;
3195            for (type = 0; type < MIGRATE_TYPES; type++) {
3196                if (!list_empty(&area->free_list[type]))
3197                    types[order] |= 1 << type;
3198            }
3199        }
3200        spin_unlock_irqrestore(&zone->lock, flags);
3201        for (order = 0; order < MAX_ORDER; order++) {
3202            printk("%lu*%lukB ", nr[order], K(1UL) << order);
3203            if (nr[order])
3204                show_migration_types(types[order]);
3205        }
3206        printk("= %lukB\n", K(total));
3207    }
3208
3209    hugetlb_show_meminfo();
3210
3211    printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3212
3213    show_swap_cache_info();
3214}
3215
3216static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3217{
3218    zoneref->zone = zone;
3219    zoneref->zone_idx = zone_idx(zone);
3220}
3221
3222/*
3223 * Builds allocation fallback zone lists.
3224 *
3225 * Add all populated zones of a node to the zonelist.
3226 */
3227static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3228                int nr_zones)
3229{
3230    struct zone *zone;
3231    enum zone_type zone_type = MAX_NR_ZONES;
3232
3233    do {
3234        zone_type--;
3235        zone = pgdat->node_zones + zone_type;
3236        if (populated_zone(zone)) {
3237            zoneref_set_zone(zone,
3238                &zonelist->_zonerefs[nr_zones++]);
3239            check_highest_zone(zone_type);
3240        }
3241    } while (zone_type);
3242
3243    return nr_zones;
3244}
3245
3246
3247/*
3248 * zonelist_order:
3249 * 0 = automatic detection of better ordering.
3250 * 1 = order by ([node] distance, -zonetype)
3251 * 2 = order by (-zonetype, [node] distance)
3252 *
3253 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3254 * the same zonelist. So only NUMA can configure this param.
3255 */
3256#define ZONELIST_ORDER_DEFAULT 0
3257#define ZONELIST_ORDER_NODE 1
3258#define ZONELIST_ORDER_ZONE 2
3259
3260/* zonelist order in the kernel.
3261 * set_zonelist_order() will set this to NODE or ZONE.
3262 */
3263static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3264static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3265
3266
3267#ifdef CONFIG_NUMA
3268/* The value user specified ....changed by config */
3269static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3270/* string for sysctl */
3271#define NUMA_ZONELIST_ORDER_LEN 16
3272char numa_zonelist_order[16] = "default";
3273
3274/*
3275 * interface for configure zonelist ordering.
3276 * command line option "numa_zonelist_order"
3277 * = "[dD]efault - default, automatic configuration.
3278 * = "[nN]ode - order by node locality, then by zone within node
3279 * = "[zZ]one - order by zone, then by locality within zone
3280 */
3281
3282static int __parse_numa_zonelist_order(char *s)
3283{
3284    if (*s == 'd' || *s == 'D') {
3285        user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3286    } else if (*s == 'n' || *s == 'N') {
3287        user_zonelist_order = ZONELIST_ORDER_NODE;
3288    } else if (*s == 'z' || *s == 'Z') {
3289        user_zonelist_order = ZONELIST_ORDER_ZONE;
3290    } else {
3291        printk(KERN_WARNING
3292            "Ignoring invalid numa_zonelist_order value: "
3293            "%s\n", s);
3294        return -EINVAL;
3295    }
3296    return 0;
3297}
3298
3299static __init int setup_numa_zonelist_order(char *s)
3300{
3301    int ret;
3302
3303    if (!s)
3304        return 0;
3305
3306    ret = __parse_numa_zonelist_order(s);
3307    if (ret == 0)
3308        strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3309
3310    return ret;
3311}
3312early_param("numa_zonelist_order", setup_numa_zonelist_order);
3313
3314/*
3315 * sysctl handler for numa_zonelist_order
3316 */
3317int numa_zonelist_order_handler(ctl_table *table, int write,
3318        void __user *buffer, size_t *length,
3319        loff_t *ppos)
3320{
3321    char saved_string[NUMA_ZONELIST_ORDER_LEN];
3322    int ret;
3323    static DEFINE_MUTEX(zl_order_mutex);
3324
3325    mutex_lock(&zl_order_mutex);
3326    if (write) {
3327        if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
3328            ret = -EINVAL;
3329            goto out;
3330        }
3331        strcpy(saved_string, (char *)table->data);
3332    }
3333    ret = proc_dostring(table, write, buffer, length, ppos);
3334    if (ret)
3335        goto out;
3336    if (write) {
3337        int oldval = user_zonelist_order;
3338
3339        ret = __parse_numa_zonelist_order((char *)table->data);
3340        if (ret) {
3341            /*
3342             * bogus value. restore saved string
3343             */
3344            strncpy((char *)table->data, saved_string,
3345                NUMA_ZONELIST_ORDER_LEN);
3346            user_zonelist_order = oldval;
3347        } else if (oldval != user_zonelist_order) {
3348            mutex_lock(&zonelists_mutex);
3349            build_all_zonelists(NULL, NULL);
3350            mutex_unlock(&zonelists_mutex);
3351        }
3352    }
3353out:
3354    mutex_unlock(&zl_order_mutex);
3355    return ret;
3356}
3357
3358
3359#define MAX_NODE_LOAD (nr_online_nodes)
3360static int node_load[MAX_NUMNODES];
3361
3362/**
3363 * find_next_best_node - find the next node that should appear in a given node's fallback list
3364 * @node: node whose fallback list we're appending
3365 * @used_node_mask: nodemask_t of already used nodes
3366 *
3367 * We use a number of factors to determine which is the next node that should
3368 * appear on a given node's fallback list. The node should not have appeared
3369 * already in @node's fallback list, and it should be the next closest node
3370 * according to the distance array (which contains arbitrary distance values
3371 * from each node to each node in the system), and should also prefer nodes
3372 * with no CPUs, since presumably they'll have very little allocation pressure
3373 * on them otherwise.
3374 * It returns -1 if no node is found.
3375 */
3376static int find_next_best_node(int node, nodemask_t *used_node_mask)
3377{
3378    int n, val;
3379    int min_val = INT_MAX;
3380    int best_node = NUMA_NO_NODE;
3381    const struct cpumask *tmp = cpumask_of_node(0);
3382
3383    /* Use the local node if we haven't already */
3384    if (!node_isset(node, *used_node_mask)) {
3385        node_set(node, *used_node_mask);
3386        return node;
3387    }
3388
3389    for_each_node_state(n, N_MEMORY) {
3390
3391        /* Don't want a node to appear more than once */
3392        if (node_isset(n, *used_node_mask))
3393            continue;
3394
3395        /* Use the distance array to find the distance */
3396        val = node_distance(node, n);
3397
3398        /* Penalize nodes under us ("prefer the next node") */
3399        val += (n < node);
3400
3401        /* Give preference to headless and unused nodes */
3402        tmp = cpumask_of_node(n);
3403        if (!cpumask_empty(tmp))
3404            val += PENALTY_FOR_NODE_WITH_CPUS;
3405
3406        /* Slight preference for less loaded node */
3407        val *= (MAX_NODE_LOAD*MAX_NUMNODES);
3408        val += node_load[n];
3409
3410        if (val < min_val) {
3411            min_val = val;
3412            best_node = n;
3413        }
3414    }
3415
3416    if (best_node >= 0)
3417        node_set(best_node, *used_node_mask);
3418
3419    return best_node;
3420}
3421
3422
3423/*
3424 * Build zonelists ordered by node and zones within node.
3425 * This results in maximum locality--normal zone overflows into local
3426 * DMA zone, if any--but risks exhausting DMA zone.
3427 */
3428static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
3429{
3430    int j;
3431    struct zonelist *zonelist;
3432
3433    zonelist = &pgdat->node_zonelists[0];
3434    for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
3435        ;
3436    j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3437    zonelist->_zonerefs[j].zone = NULL;
3438    zonelist->_zonerefs[j].zone_idx = 0;
3439}
3440
3441/*
3442 * Build gfp_thisnode zonelists
3443 */
3444static void build_thisnode_zonelists(pg_data_t *pgdat)
3445{
3446    int j;
3447    struct zonelist *zonelist;
3448
3449    zonelist = &pgdat->node_zonelists[1];
3450    j = build_zonelists_node(pgdat, zonelist, 0);
3451    zonelist->_zonerefs[j].zone = NULL;
3452    zonelist->_zonerefs[j].zone_idx = 0;
3453}
3454
3455/*
3456 * Build zonelists ordered by zone and nodes within zones.
3457 * This results in conserving DMA zone[s] until all Normal memory is
3458 * exhausted, but results in overflowing to remote node while memory
3459 * may still exist in local DMA zone.
3460 */
3461static int node_order[MAX_NUMNODES];
3462
3463static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
3464{
3465    int pos, j, node;
3466    int zone_type; /* needs to be signed */
3467    struct zone *z;
3468    struct zonelist *zonelist;
3469
3470    zonelist = &pgdat->node_zonelists[0];
3471    pos = 0;
3472    for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
3473        for (j = 0; j < nr_nodes; j++) {
3474            node = node_order[j];
3475            z = &NODE_DATA(node)->node_zones[zone_type];
3476            if (populated_zone(z)) {
3477                zoneref_set_zone(z,
3478                    &zonelist->_zonerefs[pos++]);
3479                check_highest_zone(zone_type);
3480            }
3481        }
3482    }
3483    zonelist->_zonerefs[pos].zone = NULL;
3484    zonelist->_zonerefs[pos].zone_idx = 0;
3485}
3486
3487static int default_zonelist_order(void)
3488{
3489    int nid, zone_type;
3490    unsigned long low_kmem_size, total_size;
3491    struct zone *z;
3492    int average_size;
3493    /*
3494     * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
3495     * If they are really small and used heavily, the system can fall
3496     * into OOM very easily.
3497     * This function detect ZONE_DMA/DMA32 size and configures zone order.
3498     */
3499    /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
3500    low_kmem_size = 0;
3501    total_size = 0;
3502    for_each_online_node(nid) {
3503        for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3504            z = &NODE_DATA(nid)->node_zones[zone_type];
3505            if (populated_zone(z)) {
3506                if (zone_type < ZONE_NORMAL)
3507                    low_kmem_size += z->managed_pages;
3508                total_size += z->managed_pages;
3509            } else if (zone_type == ZONE_NORMAL) {
3510                /*
3511                 * If any node has only lowmem, then node order
3512                 * is preferred to allow kernel allocations
3513                 * locally; otherwise, they can easily infringe
3514                 * on other nodes when there is an abundance of
3515                 * lowmem available to allocate from.
3516                 */
3517                return ZONELIST_ORDER_NODE;
3518            }
3519        }
3520    }
3521    if (!low_kmem_size || /* there are no DMA area. */
3522        low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
3523        return ZONELIST_ORDER_NODE;
3524    /*
3525     * look into each node's config.
3526     * If there is a node whose DMA/DMA32 memory is very big area on
3527     * local memory, NODE_ORDER may be suitable.
3528     */
3529    average_size = total_size /
3530                (nodes_weight(node_states[N_MEMORY]) + 1);
3531    for_each_online_node(nid) {
3532        low_kmem_size = 0;
3533        total_size = 0;
3534        for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3535            z = &NODE_DATA(nid)->node_zones[zone_type];
3536            if (populated_zone(z)) {
3537                if (zone_type < ZONE_NORMAL)
3538                    low_kmem_size += z->present_pages;
3539                total_size += z->present_pages;
3540            }
3541        }
3542        if (low_kmem_size &&
3543            total_size > average_size && /* ignore small node */
3544            low_kmem_size > total_size * 70/100)
3545            return ZONELIST_ORDER_NODE;
3546    }
3547    return ZONELIST_ORDER_ZONE;
3548}
3549
3550static void set_zonelist_order(void)
3551{
3552    if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3553        current_zonelist_order = default_zonelist_order();
3554    else
3555        current_zonelist_order = user_zonelist_order;
3556}
3557
3558static void build_zonelists(pg_data_t *pgdat)
3559{
3560    int j, node, load;
3561    enum zone_type i;
3562    nodemask_t used_mask;
3563    int local_node, prev_node;
3564    struct zonelist *zonelist;
3565    int order = current_zonelist_order;
3566
3567    /* initialize zonelists */
3568    for (i = 0; i < MAX_ZONELISTS; i++) {
3569        zonelist = pgdat->node_zonelists + i;
3570        zonelist->_zonerefs[0].zone = NULL;
3571        zonelist->_zonerefs[0].zone_idx = 0;
3572    }
3573
3574    /* NUMA-aware ordering of nodes */
3575    local_node = pgdat->node_id;
3576    load = nr_online_nodes;
3577    prev_node = local_node;
3578    nodes_clear(used_mask);
3579
3580    memset(node_order, 0, sizeof(node_order));
3581    j = 0;
3582
3583    while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3584        /*
3585         * We don't want to pressure a particular node.
3586         * So adding penalty to the first node in same
3587         * distance group to make it round-robin.
3588         */
3589        if (node_distance(local_node, node) !=
3590            node_distance(local_node, prev_node))
3591            node_load[node] = load;
3592
3593        prev_node = node;
3594        load--;
3595        if (order == ZONELIST_ORDER_NODE)
3596            build_zonelists_in_node_order(pgdat, node);
3597        else
3598            node_order[j++] = node; /* remember order */
3599    }
3600
3601    if (order == ZONELIST_ORDER_ZONE) {
3602        /* calculate node order -- i.e., DMA last! */
3603        build_zonelists_in_zone_order(pgdat, j);
3604    }
3605
3606    build_thisnode_zonelists(pgdat);
3607}
3608
3609/* Construct the zonelist performance cache - see further mmzone.h */
3610static void build_zonelist_cache(pg_data_t *pgdat)
3611{
3612    struct zonelist *zonelist;
3613    struct zonelist_cache *zlc;
3614    struct zoneref *z;
3615
3616    zonelist = &pgdat->node_zonelists[0];
3617    zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3618    bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3619    for (z = zonelist->_zonerefs; z->zone; z++)
3620        zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3621}
3622
3623#ifdef CONFIG_HAVE_MEMORYLESS_NODES
3624/*
3625 * Return node id of node used for "local" allocations.
3626 * I.e., first node id of first zone in arg node's generic zonelist.
3627 * Used for initializing percpu 'numa_mem', which is used primarily
3628 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3629 */
3630int local_memory_node(int node)
3631{
3632    struct zone *zone;
3633
3634    (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3635                   gfp_zone(GFP_KERNEL),
3636                   NULL,
3637                   &zone);
3638    return zone->node;
3639}
3640#endif
3641
3642#else /* CONFIG_NUMA */
3643
3644static void set_zonelist_order(void)
3645{
3646    current_zonelist_order = ZONELIST_ORDER_ZONE;
3647}
3648
3649static void build_zonelists(pg_data_t *pgdat)
3650{
3651    int node, local_node;
3652    enum zone_type j;
3653    struct zonelist *zonelist;
3654
3655    local_node = pgdat->node_id;
3656
3657    zonelist = &pgdat->node_zonelists[0];
3658    j = build_zonelists_node(pgdat, zonelist, 0);
3659
3660    /*
3661     * Now we build the zonelist so that it contains the zones
3662     * of all the other nodes.
3663     * We don't want to pressure a particular node, so when
3664     * building the zones for node N, we make sure that the
3665     * zones coming right after the local ones are those from
3666     * node N+1 (modulo N)
3667     */
3668    for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3669        if (!node_online(node))
3670            continue;
3671        j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3672    }
3673    for (node = 0; node < local_node; node++) {
3674        if (!node_online(node))
3675            continue;
3676        j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3677    }
3678
3679    zonelist->_zonerefs[j].zone = NULL;
3680    zonelist->_zonerefs[j].zone_idx = 0;
3681}
3682
3683/* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3684static void build_zonelist_cache(pg_data_t *pgdat)
3685{
3686    pgdat->node_zonelists[0].zlcache_ptr = NULL;
3687}
3688
3689#endif /* CONFIG_NUMA */
3690
3691/*
3692 * Boot pageset table. One per cpu which is going to be used for all
3693 * zones and all nodes. The parameters will be set in such a way
3694 * that an item put on a list will immediately be handed over to
3695 * the buddy list. This is safe since pageset manipulation is done
3696 * with interrupts disabled.
3697 *
3698 * The boot_pagesets must be kept even after bootup is complete for
3699 * unused processors and/or zones. They do play a role for bootstrapping
3700 * hotplugged processors.
3701 *
3702 * zoneinfo_show() and maybe other functions do
3703 * not check if the processor is online before following the pageset pointer.
3704 * Other parts of the kernel may not check if the zone is available.
3705 */
3706static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3707static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3708static void setup_zone_pageset(struct zone *zone);
3709
3710/*
3711 * Global mutex to protect against size modification of zonelists
3712 * as well as to serialize pageset setup for the new populated zone.
3713 */
3714DEFINE_MUTEX(zonelists_mutex);
3715
3716/* return values int ....just for stop_machine() */
3717static int __build_all_zonelists(void *data)
3718{
3719    int nid;
3720    int cpu;
3721    pg_data_t *self = data;
3722
3723#ifdef CONFIG_NUMA
3724    memset(node_load, 0, sizeof(node_load));
3725#endif
3726
3727    if (self && !node_online(self->node_id)) {
3728        build_zonelists(self);
3729        build_zonelist_cache(self);
3730    }
3731
3732    for_each_online_node(nid) {
3733        pg_data_t *pgdat = NODE_DATA(nid);
3734
3735        build_zonelists(pgdat);
3736        build_zonelist_cache(pgdat);
3737    }
3738
3739    /*
3740     * Initialize the boot_pagesets that are going to be used
3741     * for bootstrapping processors. The real pagesets for
3742     * each zone will be allocated later when the per cpu
3743     * allocator is available.
3744     *
3745     * boot_pagesets are used also for bootstrapping offline
3746     * cpus if the system is already booted because the pagesets
3747     * are needed to initialize allocators on a specific cpu too.
3748     * F.e. the percpu allocator needs the page allocator which
3749     * needs the percpu allocator in order to allocate its pagesets
3750     * (a chicken-egg dilemma).
3751     */
3752    for_each_possible_cpu(cpu) {
3753        setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3754
3755#ifdef CONFIG_HAVE_MEMORYLESS_NODES
3756        /*
3757         * We now know the "local memory node" for each node--
3758         * i.e., the node of the first zone in the generic zonelist.
3759         * Set up numa_mem percpu variable for on-line cpus. During
3760         * boot, only the boot cpu should be on-line; we'll init the
3761         * secondary cpus' numa_mem as they come on-line. During
3762         * node/memory hotplug, we'll fixup all on-line cpus.
3763         */
3764        if (cpu_online(cpu))
3765            set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3766#endif
3767    }
3768
3769    return 0;
3770}
3771
3772/*
3773 * Called with zonelists_mutex held always
3774 * unless system_state == SYSTEM_BOOTING.
3775 */
3776void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
3777{
3778    set_zonelist_order();
3779
3780    if (system_state == SYSTEM_BOOTING) {
3781        __build_all_zonelists(NULL);
3782        mminit_verify_zonelist();
3783        cpuset_init_current_mems_allowed();
3784    } else {
3785#ifdef CONFIG_MEMORY_HOTPLUG
3786        if (zone)
3787            setup_zone_pageset(zone);
3788#endif
3789        /* we have to stop all cpus to guarantee there is no user
3790           of zonelist */
3791        stop_machine(__build_all_zonelists, pgdat, NULL);
3792        /* cpuset refresh routine should be here */
3793    }
3794    vm_total_pages = nr_free_pagecache_pages();
3795    /*
3796     * Disable grouping by mobility if the number of pages in the
3797     * system is too low to allow the mechanism to work. It would be
3798     * more accurate, but expensive to check per-zone. This check is
3799     * made on memory-hotadd so a system can start with mobility
3800     * disabled and enable it later
3801     */
3802    if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3803        page_group_by_mobility_disabled = 1;
3804    else
3805        page_group_by_mobility_disabled = 0;
3806
3807    printk("Built %i zonelists in %s order, mobility grouping %s. "
3808        "Total pages: %ld\n",
3809            nr_online_nodes,
3810            zonelist_order_name[current_zonelist_order],
3811            page_group_by_mobility_disabled ? "off" : "on",
3812            vm_total_pages);
3813#ifdef CONFIG_NUMA
3814    printk("Policy zone: %s\n", zone_names[policy_zone]);
3815#endif
3816}
3817
3818/*
3819 * Helper functions to size the waitqueue hash table.
3820 * Essentially these want to choose hash table sizes sufficiently
3821 * large so that collisions trying to wait on pages are rare.
3822 * But in fact, the number of active page waitqueues on typical
3823 * systems is ridiculously low, less than 200. So this is even
3824 * conservative, even though it seems large.
3825 *
3826 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3827 * waitqueues, i.e. the size of the waitq table given the number of pages.
3828 */
3829#define PAGES_PER_WAITQUEUE 256
3830
3831#ifndef CONFIG_MEMORY_HOTPLUG
3832static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3833{
3834    unsigned long size = 1;
3835
3836    pages /= PAGES_PER_WAITQUEUE;
3837
3838    while (size < pages)
3839        size <<= 1;
3840
3841    /*
3842     * Once we have dozens or even hundreds of threads sleeping
3843     * on IO we've got bigger problems than wait queue collision.
3844     * Limit the size of the wait table to a reasonable size.
3845     */
3846    size = min(size, 4096UL);
3847
3848    return max(size, 4UL);
3849}
3850#else
3851/*
3852 * A zone's size might be changed by hot-add, so it is not possible to determine
3853 * a suitable size for its wait_table. So we use the maximum size now.
3854 *
3855 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3856 *
3857 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3858 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3859 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3860 *
3861 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3862 * or more by the traditional way. (See above). It equals:
3863 *
3864 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3865 * ia64(16K page size) : = ( 8G + 4M)byte.
3866 * powerpc (64K page size) : = (32G +16M)byte.
3867 */
3868static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3869{
3870    return 4096UL;
3871}
3872#endif
3873
3874/*
3875 * This is an integer logarithm so that shifts can be used later
3876 * to extract the more random high bits from the multiplicative
3877 * hash function before the remainder is taken.
3878 */
3879static inline unsigned long wait_table_bits(unsigned long size)
3880{
3881    return ffz(~size);
3882}
3883
3884#define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3885
3886/*
3887 * Check if a pageblock contains reserved pages
3888 */
3889static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
3890{
3891    unsigned long pfn;
3892
3893    for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3894        if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
3895            return 1;
3896    }
3897    return 0;
3898}
3899
3900/*
3901 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3902 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3903 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3904 * higher will lead to a bigger reserve which will get freed as contiguous
3905 * blocks as reclaim kicks in
3906 */
3907static void setup_zone_migrate_reserve(struct zone *zone)
3908{
3909    unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
3910    struct page *page;
3911    unsigned long block_migratetype;
3912    int reserve;
3913
3914    /*
3915     * Get the start pfn, end pfn and the number of blocks to reserve
3916     * We have to be careful to be aligned to pageblock_nr_pages to
3917     * make sure that we always check pfn_valid for the first page in
3918     * the block.
3919     */
3920    start_pfn = zone->zone_start_pfn;
3921    end_pfn = zone_end_pfn(zone);
3922    start_pfn = roundup(start_pfn, pageblock_nr_pages);
3923    reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3924                            pageblock_order;
3925
3926    /*
3927     * Reserve blocks are generally in place to help high-order atomic
3928     * allocations that are short-lived. A min_free_kbytes value that
3929     * would result in more than 2 reserve blocks for atomic allocations
3930     * is assumed to be in place to help anti-fragmentation for the
3931     * future allocation of hugepages at runtime.
3932     */
3933    reserve = min(2, reserve);
3934
3935    for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
3936        if (!pfn_valid(pfn))
3937            continue;
3938        page = pfn_to_page(pfn);
3939
3940        /* Watch out for overlapping nodes */
3941        if (page_to_nid(page) != zone_to_nid(zone))
3942            continue;
3943
3944        block_migratetype = get_pageblock_migratetype(page);
3945
3946        /* Only test what is necessary when the reserves are not met */
3947        if (reserve > 0) {
3948            /*
3949             * Blocks with reserved pages will never free, skip
3950             * them.
3951             */
3952            block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
3953            if (pageblock_is_reserved(pfn, block_end_pfn))
3954                continue;
3955
3956            /* If this block is reserved, account for it */
3957            if (block_migratetype == MIGRATE_RESERVE) {
3958                reserve--;
3959                continue;
3960            }
3961
3962            /* Suitable for reserving if this block is movable */
3963            if (block_migratetype == MIGRATE_MOVABLE) {
3964                set_pageblock_migratetype(page,
3965                            MIGRATE_RESERVE);
3966                move_freepages_block(zone, page,
3967                            MIGRATE_RESERVE);
3968                reserve--;
3969                continue;
3970            }
3971        }
3972
3973        /*
3974         * If the reserve is met and this is a previous reserved block,
3975         * take it back
3976         */
3977        if (block_migratetype == MIGRATE_RESERVE) {
3978            set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3979            move_freepages_block(zone, page, MIGRATE_MOVABLE);
3980        }
3981    }
3982}
3983
3984/*
3985 * Initially all pages are reserved - free ones are freed
3986 * up by free_all_bootmem() once the early boot process is
3987 * done. Non-atomic initialization, single-pass.
3988 */
3989void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
3990        unsigned long start_pfn, enum memmap_context context)
3991{
3992    struct page *page;
3993    unsigned long end_pfn = start_pfn + size;
3994    unsigned long pfn;
3995    struct zone *z;
3996
3997    if (highest_memmap_pfn < end_pfn - 1)
3998        highest_memmap_pfn = end_pfn - 1;
3999
4000    z = &NODE_DATA(nid)->node_zones[zone];
4001    for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4002        /*
4003         * There can be holes in boot-time mem_map[]s
4004         * handed to this function. They do not
4005         * exist on hotplugged memory.
4006         */
4007        if (context == MEMMAP_EARLY) {
4008            if (!early_pfn_valid(pfn))
4009                continue;
4010            if (!early_pfn_in_nid(pfn, nid))
4011                continue;
4012        }
4013        page = pfn_to_page(pfn);
4014        set_page_links(page, zone, nid, pfn);
4015        mminit_verify_page_links(page, zone, nid, pfn);
4016        init_page_count(page);
4017        page_mapcount_reset(page);
4018        page_nid_reset_last(page);
4019        SetPageReserved(page);
4020        /*
4021         * Mark the block movable so that blocks are reserved for
4022         * movable at startup. This will force kernel allocations
4023         * to reserve their blocks rather than leaking throughout
4024         * the address space during boot when many long-lived
4025         * kernel allocations are made. Later some blocks near
4026         * the start are marked MIGRATE_RESERVE by
4027         * setup_zone_migrate_reserve()
4028         *
4029         * bitmap is created for zone's valid pfn range. but memmap
4030         * can be created for invalid pages (for alignment)
4031         * check here not to call set_pageblock_migratetype() against
4032         * pfn out of zone.
4033         */
4034        if ((z->zone_start_pfn <= pfn)
4035            && (pfn < zone_end_pfn(z))
4036            && !(pfn & (pageblock_nr_pages - 1)))
4037            set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4038
4039        INIT_LIST_HEAD(&page->lru);
4040#ifdef WANT_PAGE_VIRTUAL
4041        /* The shift won't overflow because ZONE_NORMAL is below 4G. */
4042        if (!is_highmem_idx(zone))
4043            set_page_address(page, __va(pfn << PAGE_SHIFT));
4044#endif
4045    }
4046}
4047
4048static void __meminit zone_init_free_lists(struct zone *zone)
4049{
4050    int order, t;
4051    for_each_migratetype_order(order, t) {
4052        INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
4053        zone->free_area[order].nr_free = 0;
4054    }
4055}
4056
4057#ifndef __HAVE_ARCH_MEMMAP_INIT
4058#define memmap_init(size, nid, zone, start_pfn) \
4059    memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4060#endif
4061
4062static int __meminit zone_batchsize(struct zone *zone)
4063{
4064#ifdef CONFIG_MMU
4065    int batch;
4066
4067    /*
4068     * The per-cpu-pages pools are set to around 1000th of the
4069     * size of the zone. But no more than 1/2 of a meg.
4070     *
4071     * OK, so we don't know how big the cache is. So guess.
4072     */
4073    batch = zone->managed_pages / 1024;
4074    if (batch * PAGE_SIZE > 512 * 1024)
4075        batch = (512 * 1024) / PAGE_SIZE;
4076    batch /= 4; /* We effectively *= 4 below */
4077    if (batch < 1)
4078        batch = 1;
4079
4080    /*
4081     * Clamp the batch to a 2^n - 1 value. Having a power
4082     * of 2 value was found to be more likely to have
4083     * suboptimal cache aliasing properties in some cases.
4084     *
4085     * For example if 2 tasks are alternately allocating
4086     * batches of pages, one task can end up with a lot
4087     * of pages of one half of the possible page colors
4088     * and the other with pages of the other colors.
4089     */
4090    batch = rounddown_pow_of_two(batch + batch/2) - 1;
4091
4092    return batch;
4093
4094#else
4095    /* The deferral and batching of frees should be suppressed under NOMMU
4096     * conditions.
4097     *
4098     * The problem is that NOMMU needs to be able to allocate large chunks
4099     * of contiguous memory as there's no hardware page translation to
4100     * assemble apparent contiguous memory from discontiguous pages.
4101     *
4102     * Queueing large contiguous runs of pages for batching, however,
4103     * causes the pages to actually be freed in smaller chunks. As there
4104     * can be a significant delay between the individual batches being
4105     * recycled, this leads to the once large chunks of space being
4106     * fragmented and becoming unavailable for high-order allocations.
4107     */
4108    return 0;
4109#endif
4110}
4111
4112/*
4113 * pcp->high and pcp->batch values are related and dependent on one another:
4114 * ->batch must never be higher then ->high.
4115 * The following function updates them in a safe manner without read side
4116 * locking.
4117 *
4118 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4119 * those fields changing asynchronously (acording the the above rule).
4120 *
4121 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4122 * outside of boot time (or some other assurance that no concurrent updaters
4123 * exist).
4124 */
4125static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4126        unsigned long batch)
4127{
4128       /* start with a fail safe value for batch */
4129    pcp->batch = 1;
4130    smp_wmb();
4131
4132       /* Update high, then batch, in order */
4133    pcp->high = high;
4134    smp_wmb();
4135
4136    pcp->batch = batch;
4137}
4138
4139/* a companion to pageset_set_high() */
4140static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4141{
4142    pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4143}
4144
4145static void pageset_init(struct per_cpu_pageset *p)
4146{
4147    struct per_cpu_pages *pcp;
4148    int migratetype;
4149
4150    memset(p, 0, sizeof(*p));
4151
4152    pcp = &p->pcp;
4153    pcp->count = 0;
4154    for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4155        INIT_LIST_HEAD(&pcp->lists[migratetype]);
4156}
4157
4158static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4159{
4160    pageset_init(p);
4161    pageset_set_batch(p, batch);
4162}
4163
4164/*
4165 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4166 * to the value high for the pageset p.
4167 */
4168static void pageset_set_high(struct per_cpu_pageset *p,
4169                unsigned long high)
4170{
4171    unsigned long batch = max(1UL, high / 4);
4172    if ((high / 4) > (PAGE_SHIFT * 8))
4173        batch = PAGE_SHIFT * 8;
4174
4175    pageset_update(&p->pcp, high, batch);
4176}
4177
4178static void __meminit pageset_set_high_and_batch(struct zone *zone,
4179        struct per_cpu_pageset *pcp)
4180{
4181    if (percpu_pagelist_fraction)
4182        pageset_set_high(pcp,
4183            (zone->managed_pages /
4184                percpu_pagelist_fraction));
4185    else
4186        pageset_set_batch(pcp, zone_batchsize(zone));
4187}
4188
4189static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4190{
4191    struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4192
4193    pageset_init(pcp);
4194    pageset_set_high_and_batch(zone, pcp);
4195}
4196
4197static void __meminit setup_zone_pageset(struct zone *zone)
4198{
4199    int cpu;
4200    zone->pageset = alloc_percpu(struct per_cpu_pageset);
4201    for_each_possible_cpu(cpu)
4202        zone_pageset_init(zone, cpu);
4203}
4204
4205/*
4206 * Allocate per cpu pagesets and initialize them.
4207 * Before this call only boot pagesets were available.
4208 */
4209void __init setup_per_cpu_pageset(void)
4210{
4211    struct zone *zone;
4212
4213    for_each_populated_zone(zone)
4214        setup_zone_pageset(zone);
4215}
4216
4217static noinline __init_refok
4218int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4219{
4220    int i;
4221    struct pglist_data *pgdat = zone->zone_pgdat;
4222    size_t alloc_size;
4223
4224    /*
4225     * The per-page waitqueue mechanism uses hashed waitqueues
4226     * per zone.
4227     */
4228    zone->wait_table_hash_nr_entries =
4229         wait_table_hash_nr_entries(zone_size_pages);
4230    zone->wait_table_bits =
4231        wait_table_bits(zone->wait_table_hash_nr_entries);
4232    alloc_size = zone->wait_table_hash_nr_entries
4233                    * sizeof(wait_queue_head_t);
4234
4235    if (!slab_is_available()) {
4236        zone->wait_table = (wait_queue_head_t *)
4237            alloc_bootmem_node_nopanic(pgdat, alloc_size);
4238    } else {
4239        /*
4240         * This case means that a zone whose size was 0 gets new memory
4241         * via memory hot-add.
4242         * But it may be the case that a new node was hot-added. In
4243         * this case vmalloc() will not be able to use this new node's
4244         * memory - this wait_table must be initialized to use this new
4245         * node itself as well.
4246         * To use this new node's memory, further consideration will be
4247         * necessary.
4248         */
4249        zone->wait_table = vmalloc(alloc_size);
4250    }
4251    if (!zone->wait_table)
4252        return -ENOMEM;
4253
4254    for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4255        init_waitqueue_head(zone->wait_table + i);
4256
4257    return 0;
4258}
4259
4260static __meminit void zone_pcp_init(struct zone *zone)
4261{
4262    /*
4263     * per cpu subsystem is not up at this point. The following code
4264     * relies on the ability of the linker to provide the
4265     * offset of a (static) per cpu variable into the per cpu area.
4266     */
4267    zone->pageset = &boot_pageset;
4268
4269    if (zone->present_pages)
4270        printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4271            zone->name, zone->present_pages,
4272                     zone_batchsize(zone));
4273}
4274
4275int __meminit init_currently_empty_zone(struct zone *zone,
4276                    unsigned long zone_start_pfn,
4277                    unsigned long size,
4278                    enum memmap_context context)
4279{
4280    struct pglist_data *pgdat = zone->zone_pgdat;
4281    int ret;
4282    ret = zone_wait_table_init(zone, size);
4283    if (ret)
4284        return ret;
4285    pgdat->nr_zones = zone_idx(zone) + 1;
4286
4287    zone->zone_start_pfn = zone_start_pfn;
4288
4289    mminit_dprintk(MMINIT_TRACE, "memmap_init",
4290            "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4291            pgdat->node_id,
4292            (unsigned long)zone_idx(zone),
4293            zone_start_pfn, (zone_start_pfn + size));
4294
4295    zone_init_free_lists(zone);
4296
4297    return 0;
4298}
4299
4300#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4301#ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4302/*
4303 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4304 * Architectures may implement their own version but if add_active_range()
4305 * was used and there are no special requirements, this is a convenient
4306 * alternative
4307 */
4308int __meminit __early_pfn_to_nid(unsigned long pfn)
4309{
4310    unsigned long start_pfn, end_pfn;
4311    int nid;
4312    /*
4313     * NOTE: The following SMP-unsafe globals are only used early in boot
4314     * when the kernel is running single-threaded.
4315     */
4316    static unsigned long __meminitdata last_start_pfn, last_end_pfn;
4317    static int __meminitdata last_nid;
4318
4319    if (last_start_pfn <= pfn && pfn < last_end_pfn)
4320        return last_nid;
4321
4322    nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
4323    if (nid != -1) {
4324        last_start_pfn = start_pfn;
4325        last_end_pfn = end_pfn;
4326        last_nid = nid;
4327    }
4328
4329    return nid;
4330}
4331#endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4332
4333int __meminit early_pfn_to_nid(unsigned long pfn)
4334{
4335    int nid;
4336
4337    nid = __early_pfn_to_nid(pfn);
4338    if (nid >= 0)
4339        return nid;
4340    /* just returns 0 */
4341    return 0;
4342}
4343
4344#ifdef CONFIG_NODES_SPAN_OTHER_NODES
4345bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
4346{
4347    int nid;
4348
4349    nid = __early_pfn_to_nid(pfn);
4350    if (nid >= 0 && nid != node)
4351        return false;
4352    return true;
4353}
4354#endif
4355
4356/**
4357 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
4358 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4359 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
4360 *
4361 * If an architecture guarantees that all ranges registered with
4362 * add_active_ranges() contain no holes and may be freed, this
4363 * this function may be used instead of calling free_bootmem() manually.
4364 */
4365void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4366{
4367    unsigned long start_pfn, end_pfn;
4368    int i, this_nid;
4369
4370    for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4371        start_pfn = min(start_pfn, max_low_pfn);
4372        end_pfn = min(end_pfn, max_low_pfn);
4373
4374        if (start_pfn < end_pfn)
4375            free_bootmem_node(NODE_DATA(this_nid),
4376                      PFN_PHYS(start_pfn),
4377                      (end_pfn - start_pfn) << PAGE_SHIFT);
4378    }
4379}
4380
4381/**
4382 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4383 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4384 *
4385 * If an architecture guarantees that all ranges registered with
4386 * add_active_ranges() contain no holes and may be freed, this
4387 * function may be used instead of calling memory_present() manually.
4388 */
4389void __init sparse_memory_present_with_active_regions(int nid)
4390{
4391    unsigned long start_pfn, end_pfn;
4392    int i, this_nid;
4393
4394    for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4395        memory_present(this_nid, start_pfn, end_pfn);
4396}
4397
4398/**
4399 * get_pfn_range_for_nid - Return the start and end page frames for a node
4400 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4401 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4402 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4403 *
4404 * It returns the start and end page frame of a node based on information
4405 * provided by an arch calling add_active_range(). If called for a node
4406 * with no available memory, a warning is printed and the start and end
4407 * PFNs will be 0.
4408 */
4409void __meminit get_pfn_range_for_nid(unsigned int nid,
4410            unsigned long *start_pfn, unsigned long *end_pfn)
4411{
4412    unsigned long this_start_pfn, this_end_pfn;
4413    int i;
4414
4415    *start_pfn = -1UL;
4416    *end_pfn = 0;
4417
4418    for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4419        *start_pfn = min(*start_pfn, this_start_pfn);
4420        *end_pfn = max(*end_pfn, this_end_pfn);
4421    }
4422
4423    if (*start_pfn == -1UL)
4424        *start_pfn = 0;
4425}
4426
4427/*
4428 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4429 * assumption is made that zones within a node are ordered in monotonic
4430 * increasing memory addresses so that the "highest" populated zone is used
4431 */
4432static void __init find_usable_zone_for_movable(void)
4433{
4434    int zone_index;
4435    for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4436        if (zone_index == ZONE_MOVABLE)
4437            continue;
4438
4439        if (arch_zone_highest_possible_pfn[zone_index] >
4440                arch_zone_lowest_possible_pfn[zone_index])
4441            break;
4442    }
4443
4444    VM_BUG_ON(zone_index == -1);
4445    movable_zone = zone_index;
4446}
4447
4448/*
4449 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4450 * because it is sized independent of architecture. Unlike the other zones,
4451 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4452 * in each node depending on the size of each node and how evenly kernelcore
4453 * is distributed. This helper function adjusts the zone ranges
4454 * provided by the architecture for a given node by using the end of the
4455 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4456 * zones within a node are in order of monotonic increases memory addresses
4457 */
4458static void __meminit adjust_zone_range_for_zone_movable(int nid,
4459                    unsigned long zone_type,
4460                    unsigned long node_start_pfn,
4461                    unsigned long node_end_pfn,
4462                    unsigned long *zone_start_pfn,
4463                    unsigned long *zone_end_pfn)
4464{
4465    /* Only adjust if ZONE_MOVABLE is on this node */
4466    if (zone_movable_pfn[nid]) {
4467        /* Size ZONE_MOVABLE */
4468        if (zone_type == ZONE_MOVABLE) {
4469            *zone_start_pfn = zone_movable_pfn[nid];
4470            *zone_end_pfn = min(node_end_pfn,
4471                arch_zone_highest_possible_pfn[movable_zone]);
4472
4473        /* Adjust for ZONE_MOVABLE starting within this range */
4474        } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4475                *zone_end_pfn > zone_movable_pfn[nid]) {
4476            *zone_end_pfn = zone_movable_pfn[nid];
4477
4478        /* Check if this whole range is within ZONE_MOVABLE */
4479        } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4480            *zone_start_pfn = *zone_end_pfn;
4481    }
4482}
4483
4484/*
4485 * Return the number of pages a zone spans in a node, including holes
4486 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4487 */
4488static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4489                    unsigned long zone_type,
4490                    unsigned long node_start_pfn,
4491                    unsigned long node_end_pfn,
4492                    unsigned long *ignored)
4493{
4494    unsigned long zone_start_pfn, zone_end_pfn;
4495
4496    /* Get the start and end of the zone */
4497    zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4498    zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4499    adjust_zone_range_for_zone_movable(nid, zone_type,
4500                node_start_pfn, node_end_pfn,
4501                &zone_start_pfn, &zone_end_pfn);
4502
4503    /* Check that this node has pages within the zone's required range */
4504    if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4505        return 0;
4506
4507    /* Move the zone boundaries inside the node if necessary */
4508    zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4509    zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4510
4511    /* Return the spanned pages */
4512    return zone_end_pfn - zone_start_pfn;
4513}
4514
4515/*
4516 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4517 * then all holes in the requested range will be accounted for.
4518 */
4519unsigned long __meminit __absent_pages_in_range(int nid,
4520                unsigned long range_start_pfn,
4521                unsigned long range_end_pfn)
4522{
4523    unsigned long nr_absent = range_end_pfn - range_start_pfn;
4524    unsigned long start_pfn, end_pfn;
4525    int i;
4526
4527    for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4528        start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4529        end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4530        nr_absent -= end_pfn - start_pfn;
4531    }
4532    return nr_absent;
4533}
4534
4535/**
4536 * absent_pages_in_range - Return number of page frames in holes within a range
4537 * @start_pfn: The start PFN to start searching for holes
4538 * @end_pfn: The end PFN to stop searching for holes
4539 *
4540 * It returns the number of pages frames in memory holes within a range.
4541 */
4542unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4543                            unsigned long end_pfn)
4544{
4545    return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4546}
4547
4548/* Return the number of page frames in holes in a zone on a node */
4549static unsigned long __meminit zone_absent_pages_in_node(int nid,
4550                    unsigned long zone_type,
4551                    unsigned long node_start_pfn,
4552                    unsigned long node_end_pfn,
4553                    unsigned long *ignored)
4554{
4555    unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
4556    unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
4557    unsigned long zone_start_pfn, zone_end_pfn;
4558
4559    zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
4560    zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
4561
4562    adjust_zone_range_for_zone_movable(nid, zone_type,
4563            node_start_pfn, node_end_pfn,
4564            &zone_start_pfn, &zone_end_pfn);
4565    return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4566}
4567
4568#else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4569static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4570                    unsigned long zone_type,
4571                    unsigned long node_start_pfn,
4572                    unsigned long node_end_pfn,
4573                    unsigned long *zones_size)
4574{
4575    return zones_size[zone_type];
4576}
4577
4578static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4579                        unsigned long zone_type,
4580                        unsigned long node_start_pfn,
4581                        unsigned long node_end_pfn,
4582                        unsigned long *zholes_size)
4583{
4584    if (!zholes_size)
4585        return 0;
4586
4587    return zholes_size[zone_type];
4588}
4589
4590#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4591
4592static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4593                        unsigned long node_start_pfn,
4594                        unsigned long node_end_pfn,
4595                        unsigned long *zones_size,
4596                        unsigned long *zholes_size)
4597{
4598    unsigned long realtotalpages, totalpages = 0;
4599    enum zone_type i;
4600
4601    for (i = 0; i < MAX_NR_ZONES; i++)
4602        totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4603                             node_start_pfn,
4604                             node_end_pfn,
4605                             zones_size);
4606    pgdat->node_spanned_pages = totalpages;
4607
4608    realtotalpages = totalpages;
4609    for (i = 0; i < MAX_NR_ZONES; i++)
4610        realtotalpages -=
4611            zone_absent_pages_in_node(pgdat->node_id, i,
4612                          node_start_pfn, node_end_pfn,
4613                          zholes_size);
4614    pgdat->node_present_pages = realtotalpages;
4615    printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4616                            realtotalpages);
4617}
4618
4619#ifndef CONFIG_SPARSEMEM
4620/*
4621 * Calculate the size of the zone->blockflags rounded to an unsigned long
4622 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4623 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4624 * round what is now in bits to nearest long in bits, then return it in
4625 * bytes.
4626 */
4627static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
4628{
4629    unsigned long usemapsize;
4630
4631    zonesize += zone_start_pfn & (pageblock_nr_pages-1);
4632    usemapsize = roundup(zonesize, pageblock_nr_pages);
4633    usemapsize = usemapsize >> pageblock_order;
4634    usemapsize *= NR_PAGEBLOCK_BITS;
4635    usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4636
4637    return usemapsize / 8;
4638}
4639
4640static void __init setup_usemap(struct pglist_data *pgdat,
4641                struct zone *zone,
4642                unsigned long zone_start_pfn,
4643                unsigned long zonesize)
4644{
4645    unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
4646    zone->pageblock_flags = NULL;
4647    if (usemapsize)
4648        zone->pageblock_flags = alloc_bootmem_node_nopanic(pgdat,
4649                                   usemapsize);
4650}
4651#else
4652static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
4653                unsigned long zone_start_pfn, unsigned long zonesize) {}
4654#endif /* CONFIG_SPARSEMEM */
4655
4656#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4657
4658/* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4659void __paginginit set_pageblock_order(void)
4660{
4661    unsigned int order;
4662
4663    /* Check that pageblock_nr_pages has not already been setup */
4664    if (pageblock_order)
4665        return;
4666
4667    if (HPAGE_SHIFT > PAGE_SHIFT)
4668        order = HUGETLB_PAGE_ORDER;
4669    else
4670        order = MAX_ORDER - 1;
4671
4672    /*
4673     * Assume the largest contiguous order of interest is a huge page.
4674     * This value may be variable depending on boot parameters on IA64 and
4675     * powerpc.
4676     */
4677    pageblock_order = order;
4678}
4679#else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4680
4681/*
4682 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4683 * is unused as pageblock_order is set at compile-time. See
4684 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4685 * the kernel config
4686 */
4687void __paginginit set_pageblock_order(void)
4688{
4689}
4690
4691#endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4692
4693static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
4694                           unsigned long present_pages)
4695{
4696    unsigned long pages = spanned_pages;
4697
4698    /*
4699     * Provide a more accurate estimation if there are holes within
4700     * the zone and SPARSEMEM is in use. If there are holes within the
4701     * zone, each populated memory region may cost us one or two extra
4702     * memmap pages due to alignment because memmap pages for each
4703     * populated regions may not naturally algined on page boundary.
4704     * So the (present_pages >> 4) heuristic is a tradeoff for that.
4705     */
4706    if (spanned_pages > present_pages + (present_pages >> 4) &&
4707        IS_ENABLED(CONFIG_SPARSEMEM))
4708        pages = present_pages;
4709
4710    return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
4711}
4712
4713/*
4714 * Set up the zone data structures:
4715 * - mark all pages reserved
4716 * - mark all memory queues empty
4717 * - clear the memory bitmaps
4718 *
4719 * NOTE: pgdat should get zeroed by caller.
4720 */
4721static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4722        unsigned long node_start_pfn, unsigned long node_end_pfn,
4723        unsigned long *zones_size, unsigned long *zholes_size)
4724{
4725    enum zone_type j;
4726    int nid = pgdat->node_id;
4727    unsigned long zone_start_pfn = pgdat->node_start_pfn;
4728    int ret;
4729
4730    pgdat_resize_init(pgdat);
4731#ifdef CONFIG_NUMA_BALANCING
4732    spin_lock_init(&pgdat->numabalancing_migrate_lock);
4733    pgdat->numabalancing_migrate_nr_pages = 0;
4734    pgdat->numabalancing_migrate_next_window = jiffies;
4735#endif
4736    init_waitqueue_head(&pgdat->kswapd_wait);
4737    init_waitqueue_head(&pgdat->pfmemalloc_wait);
4738    pgdat_page_cgroup_init(pgdat);
4739
4740    for (j = 0; j < MAX_NR_ZONES; j++) {
4741        struct zone *zone = pgdat->node_zones + j;
4742        unsigned long size, realsize, freesize, memmap_pages;
4743
4744        size = zone_spanned_pages_in_node(nid, j, node_start_pfn,
4745                          node_end_pfn, zones_size);
4746        realsize = freesize = size - zone_absent_pages_in_node(nid, j,
4747                                node_start_pfn,
4748                                node_end_pfn,
4749                                zholes_size);
4750
4751        /*
4752         * Adjust freesize so that it accounts for how much memory
4753         * is used by this zone for memmap. This affects the watermark
4754         * and per-cpu initialisations
4755         */
4756        memmap_pages = calc_memmap_size(size, realsize);
4757        if (freesize >= memmap_pages) {
4758            freesize -= memmap_pages;
4759            if (memmap_pages)
4760                printk(KERN_DEBUG
4761                       " %s zone: %lu pages used for memmap\n",
4762                       zone_names[j], memmap_pages);
4763        } else
4764            printk(KERN_WARNING
4765                " %s zone: %lu pages exceeds freesize %lu\n",
4766                zone_names[j], memmap_pages, freesize);
4767
4768        /* Account for reserved pages */
4769        if (j == 0 && freesize > dma_reserve) {
4770            freesize -= dma_reserve;
4771            printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4772                    zone_names[0], dma_reserve);
4773        }
4774
4775        if (!is_highmem_idx(j))
4776            nr_kernel_pages += freesize;
4777        /* Charge for highmem memmap if there are enough kernel pages */
4778        else if (nr_kernel_pages > memmap_pages * 2)
4779            nr_kernel_pages -= memmap_pages;
4780        nr_all_pages += freesize;
4781
4782        zone->spanned_pages = size;
4783        zone->present_pages = realsize;
4784        /*
4785         * Set an approximate value for lowmem here, it will be adjusted
4786         * when the bootmem allocator frees pages into the buddy system.
4787         * And all highmem pages will be managed by the buddy system.
4788         */
4789        zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
4790#ifdef CONFIG_NUMA
4791        zone->node = nid;
4792        zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
4793                        / 100;
4794        zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
4795#endif
4796        zone->name = zone_names[j];
4797        spin_lock_init(&zone->lock);
4798        spin_lock_init(&zone->lru_lock);
4799        zone_seqlock_init(zone);
4800        zone->zone_pgdat = pgdat;
4801        zone_pcp_init(zone);
4802
4803        /* For bootup, initialized properly in watermark setup */
4804        mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
4805
4806        lruvec_init(&zone->lruvec);
4807        if (!size)
4808            continue;
4809
4810        set_pageblock_order();
4811        setup_usemap(pgdat, zone, zone_start_pfn, size);
4812        ret = init_currently_empty_zone(zone, zone_start_pfn,
4813                        size, MEMMAP_EARLY);
4814        BUG_ON(ret);
4815        memmap_init(size, nid, j, zone_start_pfn);
4816        zone_start_pfn += size;
4817    }
4818}
4819
4820static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4821{
4822    /* Skip empty nodes */
4823    if (!pgdat->node_spanned_pages)
4824        return;
4825
4826#ifdef CONFIG_FLAT_NODE_MEM_MAP
4827    /* ia64 gets its own node_mem_map, before this, without bootmem */
4828    if (!pgdat->node_mem_map) {
4829        unsigned long size, start, end;
4830        struct page *map;
4831
4832        /*
4833         * The zone's endpoints aren't required to be MAX_ORDER
4834         * aligned but the node_mem_map endpoints must be in order
4835         * for the buddy allocator to function correctly.
4836         */
4837        start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4838        end = pgdat_end_pfn(pgdat);
4839        end = ALIGN(end, MAX_ORDER_NR_PAGES);
4840        size = (end - start) * sizeof(struct page);
4841        map = alloc_remap(pgdat->node_id, size);
4842        if (!map)
4843            map = alloc_bootmem_node_nopanic(pgdat, size);
4844        pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4845    }
4846#ifndef CONFIG_NEED_MULTIPLE_NODES
4847    /*
4848     * With no DISCONTIG, the global mem_map is just set as node 0's
4849     */
4850    if (pgdat == NODE_DATA(0)) {
4851        mem_map = NODE_DATA(0)->node_mem_map;
4852#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4853        if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4854            mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4855#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4856    }
4857#endif
4858#endif /* CONFIG_FLAT_NODE_MEM_MAP */
4859}
4860
4861void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4862        unsigned long node_start_pfn, unsigned long *zholes_size)
4863{
4864    pg_data_t *pgdat = NODE_DATA(nid);
4865    unsigned long start_pfn = 0;
4866    unsigned long end_pfn = 0;
4867
4868    /* pg_data_t should be reset to zero when it's allocated */
4869    WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
4870
4871    pgdat->node_id = nid;
4872    pgdat->node_start_pfn = node_start_pfn;
4873    init_zone_allows_reclaim(nid);
4874#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4875    get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
4876#endif
4877    calculate_node_totalpages(pgdat, start_pfn, end_pfn,
4878                  zones_size, zholes_size);
4879
4880    alloc_node_mem_map(pgdat);
4881#ifdef CONFIG_FLAT_NODE_MEM_MAP
4882    printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4883        nid, (unsigned long)pgdat,
4884        (unsigned long)pgdat->node_mem_map);
4885#endif
4886
4887    free_area_init_core(pgdat, start_pfn, end_pfn,
4888                zones_size, zholes_size);
4889}
4890
4891#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4892
4893#if MAX_NUMNODES > 1
4894/*
4895 * Figure out the number of possible node ids.
4896 */
4897void __init setup_nr_node_ids(void)
4898{
4899    unsigned int node;
4900    unsigned int highest = 0;
4901
4902    for_each_node_mask(node, node_possible_map)
4903        highest = node;
4904    nr_node_ids = highest + 1;
4905}
4906#endif
4907
4908/**
4909 * node_map_pfn_alignment - determine the maximum internode alignment
4910 *
4911 * This function should be called after node map is populated and sorted.
4912 * It calculates the maximum power of two alignment which can distinguish
4913 * all the nodes.
4914 *
4915 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4916 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
4917 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
4918 * shifted, 1GiB is enough and this function will indicate so.
4919 *
4920 * This is used to test whether pfn -> nid mapping of the chosen memory
4921 * model has fine enough granularity to avoid incorrect mapping for the
4922 * populated node map.
4923 *
4924 * Returns the determined alignment in pfn's. 0 if there is no alignment
4925 * requirement (single node).
4926 */
4927unsigned long __init node_map_pfn_alignment(void)
4928{
4929    unsigned long accl_mask = 0, last_end = 0;
4930    unsigned long start, end, mask;
4931    int last_nid = -1;
4932    int i, nid;
4933
4934    for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
4935        if (!start || last_nid < 0 || last_nid == nid) {
4936            last_nid = nid;
4937            last_end = end;
4938            continue;
4939        }
4940
4941        /*
4942         * Start with a mask granular enough to pin-point to the
4943         * start pfn and tick off bits one-by-one until it becomes
4944         * too coarse to separate the current node from the last.
4945         */
4946        mask = ~((1 << __ffs(start)) - 1);
4947        while (mask && last_end <= (start & (mask << 1)))
4948            mask <<= 1;
4949
4950        /* accumulate all internode masks */
4951        accl_mask |= mask;
4952    }
4953
4954    /* convert mask to number of pages */
4955    return ~accl_mask + 1;
4956}
4957
4958/* Find the lowest pfn for a node */
4959static unsigned long __init find_min_pfn_for_node(int nid)
4960{
4961    unsigned long min_pfn = ULONG_MAX;
4962    unsigned long start_pfn;
4963    int i;
4964
4965    for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
4966        min_pfn = min(min_pfn, start_pfn);
4967
4968    if (min_pfn == ULONG_MAX) {
4969        printk(KERN_WARNING
4970            "Could not find start_pfn for node %d\n", nid);
4971        return 0;
4972    }
4973
4974    return min_pfn;
4975}
4976
4977/**
4978 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4979 *
4980 * It returns the minimum PFN based on information provided via
4981 * add_active_range().
4982 */
4983unsigned long __init find_min_pfn_with_active_regions(void)
4984{
4985    return find_min_pfn_for_node(MAX_NUMNODES);
4986}
4987
4988/*
4989 * early_calculate_totalpages()
4990 * Sum pages in active regions for movable zone.
4991 * Populate N_MEMORY for calculating usable_nodes.
4992 */
4993static unsigned long __init early_calculate_totalpages(void)
4994{
4995    unsigned long totalpages = 0;
4996    unsigned long start_pfn, end_pfn;
4997    int i, nid;
4998
4999    for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
5000        unsigned long pages = end_pfn - start_pfn;
5001
5002        totalpages += pages;
5003        if (pages)
5004            node_set_state(nid, N_MEMORY);
5005    }
5006    return totalpages;
5007}
5008
5009/*
5010 * Find the PFN the Movable zone begins in each node. Kernel memory
5011 * is spread evenly between nodes as long as the nodes have enough
5012 * memory. When they don't, some nodes will have more kernelcore than
5013 * others
5014 */
5015static void __init find_zone_movable_pfns_for_nodes(void)
5016{
5017    int i, nid;
5018    unsigned long usable_startpfn;
5019    unsigned long kernelcore_node, kernelcore_remaining;
5020    /* save the state before borrow the nodemask */
5021    nodemask_t saved_node_state = node_states[N_MEMORY];
5022    unsigned long totalpages = early_calculate_totalpages();
5023    int usable_nodes = nodes_weight(node_states[N_MEMORY]);
5024
5025    /*
5026     * If movablecore was specified, calculate what size of
5027     * kernelcore that corresponds so that memory usable for
5028     * any allocation type is evenly spread. If both kernelcore
5029     * and movablecore are specified, then the value of kernelcore
5030     * will be used for required_kernelcore if it's greater than
5031     * what movablecore would have allowed.
5032     */
5033    if (required_movablecore) {
5034        unsigned long corepages;
5035
5036        /*
5037         * Round-up so that ZONE_MOVABLE is at least as large as what
5038         * was requested by the user
5039         */
5040        required_movablecore =
5041            roundup(required_movablecore, MAX_ORDER_NR_PAGES);
5042        corepages = totalpages - required_movablecore;
5043
5044        required_kernelcore = max(required_kernelcore, corepages);
5045    }
5046
5047    /* If kernelcore was not specified, there is no ZONE_MOVABLE */
5048    if (!required_kernelcore)
5049        goto out;
5050
5051    /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5052    find_usable_zone_for_movable();
5053    usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
5054
5055restart:
5056    /* Spread kernelcore memory as evenly as possible throughout nodes */
5057    kernelcore_node = required_kernelcore / usable_nodes;
5058    for_each_node_state(nid, N_MEMORY) {
5059        unsigned long start_pfn, end_pfn;
5060
5061        /*
5062         * Recalculate kernelcore_node if the division per node
5063         * now exceeds what is necessary to satisfy the requested
5064         * amount of memory for the kernel
5065         */
5066        if (required_kernelcore < kernelcore_node)
5067            kernelcore_node = required_kernelcore / usable_nodes;
5068
5069        /*
5070         * As the map is walked, we track how much memory is usable
5071         * by the kernel using kernelcore_remaining. When it is
5072         * 0, the rest of the node is usable by ZONE_MOVABLE
5073         */
5074        kernelcore_remaining = kernelcore_node;
5075
5076        /* Go through each range of PFNs within this node */
5077        for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5078            unsigned long size_pages;
5079
5080            start_pfn = max(start_pfn, zone_movable_pfn[nid]);
5081            if (start_pfn >= end_pfn)
5082                continue;
5083
5084            /* Account for what is only usable for kernelcore */
5085            if (start_pfn < usable_startpfn) {
5086                unsigned long kernel_pages;
5087                kernel_pages = min(end_pfn, usable_startpfn)
5088                                - start_pfn;
5089
5090                kernelcore_remaining -= min(kernel_pages,
5091                            kernelcore_remaining);
5092                required_kernelcore -= min(kernel_pages,
5093                            required_kernelcore);
5094
5095                /* Continue if range is now fully accounted */
5096                if (end_pfn <= usable_startpfn) {
5097
5098                    /*
5099                     * Push zone_movable_pfn to the end so
5100                     * that if we have to rebalance
5101                     * kernelcore across nodes, we will
5102                     * not double account here
5103                     */
5104                    zone_movable_pfn[nid] = end_pfn;
5105                    continue;
5106                }
5107                start_pfn = usable_startpfn;
5108            }
5109
5110            /*
5111             * The usable PFN range for ZONE_MOVABLE is from
5112             * start_pfn->end_pfn. Calculate size_pages as the
5113             * number of pages used as kernelcore
5114             */
5115            size_pages = end_pfn - start_pfn;
5116            if (size_pages > kernelcore_remaining)
5117                size_pages = kernelcore_remaining;
5118            zone_movable_pfn[nid] = start_pfn + size_pages;
5119
5120            /*
5121             * Some kernelcore has been met, update counts and
5122             * break if the kernelcore for this node has been
5123             * satisfied
5124             */
5125            required_kernelcore -= min(required_kernelcore,
5126                                size_pages);
5127            kernelcore_remaining -= size_pages;
5128            if (!kernelcore_remaining)
5129                break;
5130        }
5131    }
5132
5133    /*
5134     * If there is still required_kernelcore, we do another pass with one
5135     * less node in the count. This will push zone_movable_pfn[nid] further
5136     * along on the nodes that still have memory until kernelcore is
5137     * satisfied
5138     */
5139    usable_nodes--;
5140    if (usable_nodes && required_kernelcore > usable_nodes)
5141        goto restart;
5142
5143    /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5144    for (nid = 0; nid < MAX_NUMNODES; nid++)
5145        zone_movable_pfn[nid] =
5146            roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5147
5148out:
5149    /* restore the node_state */
5150    node_states[N_MEMORY] = saved_node_state;
5151}
5152
5153/* Any regular or high memory on that node ? */
5154static void check_for_memory(pg_data_t *pgdat, int nid)
5155{
5156    enum zone_type zone_type;
5157
5158    if (N_MEMORY == N_NORMAL_MEMORY)
5159        return;
5160
5161    for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5162        struct zone *zone = &pgdat->node_zones[zone_type];
5163        if (zone->present_pages) {
5164            node_set_state(nid, N_HIGH_MEMORY);
5165            if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5166                zone_type <= ZONE_NORMAL)
5167                node_set_state(nid, N_NORMAL_MEMORY);
5168            break;
5169        }
5170    }
5171}
5172
5173/**
5174 * free_area_init_nodes - Initialise all pg_data_t and zone data
5175 * @max_zone_pfn: an array of max PFNs for each zone
5176 *
5177 * This will call free_area_init_node() for each active node in the system.
5178 * Using the page ranges provided by add_active_range(), the size of each
5179 * zone in each node and their holes is calculated. If the maximum PFN
5180 * between two adjacent zones match, it is assumed that the zone is empty.
5181 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5182 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5183 * starts where the previous one ended. For example, ZONE_DMA32 starts
5184 * at arch_max_dma_pfn.
5185 */
5186void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5187{
5188    unsigned long start_pfn, end_pfn;
5189    int i, nid;
5190
5191    /* Record where the zone boundaries are */
5192    memset(arch_zone_lowest_possible_pfn, 0,
5193                sizeof(arch_zone_lowest_possible_pfn));
5194    memset(arch_zone_highest_possible_pfn, 0,
5195                sizeof(arch_zone_highest_possible_pfn));
5196    arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5197    arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5198    for (i = 1; i < MAX_NR_ZONES; i++) {
5199        if (i == ZONE_MOVABLE)
5200            continue;
5201        arch_zone_lowest_possible_pfn[i] =
5202            arch_zone_highest_possible_pfn[i-1];
5203        arch_zone_highest_possible_pfn[i] =
5204            max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5205    }
5206    arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5207    arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5208
5209    /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5210    memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5211    find_zone_movable_pfns_for_nodes();
5212
5213    /* Print out the zone ranges */
5214    printk("Zone ranges:\n");
5215    for (i = 0; i < MAX_NR_ZONES; i++) {
5216        if (i == ZONE_MOVABLE)
5217            continue;
5218        printk(KERN_CONT " %-8s ", zone_names[i]);
5219        if (arch_zone_lowest_possible_pfn[i] ==
5220                arch_zone_highest_possible_pfn[i])
5221            printk(KERN_CONT "empty\n");
5222        else
5223            printk(KERN_CONT "[mem %0#10lx-%0#10lx]\n",
5224                arch_zone_lowest_possible_pfn[i] << PAGE_SHIFT,
5225                (arch_zone_highest_possible_pfn[i]
5226                    << PAGE_SHIFT) - 1);
5227    }
5228
5229    /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5230    printk("Movable zone start for each node\n");
5231    for (i = 0; i < MAX_NUMNODES; i++) {
5232        if (zone_movable_pfn[i])
5233            printk(" Node %d: %#010lx\n", i,
5234                   zone_movable_pfn[i] << PAGE_SHIFT);
5235    }
5236
5237    /* Print out the early node map */
5238    printk("Early memory node ranges\n");
5239    for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5240        printk(" node %3d: [mem %#010lx-%#010lx]\n", nid,
5241               start_pfn << PAGE_SHIFT, (end_pfn << PAGE_SHIFT) - 1);
5242
5243    /* Initialise every node */
5244    mminit_verify_pageflags_layout();
5245    setup_nr_node_ids();
5246    for_each_online_node(nid) {
5247        pg_data_t *pgdat = NODE_DATA(nid);
5248        free_area_init_node(nid, NULL,
5249                find_min_pfn_for_node(nid), NULL);
5250
5251        /* Any memory on that node */
5252        if (pgdat->node_present_pages)
5253            node_set_state(nid, N_MEMORY);
5254        check_for_memory(pgdat, nid);
5255    }
5256}
5257
5258static int __init cmdline_parse_core(char *p, unsigned long *core)
5259{
5260    unsigned long long coremem;
5261    if (!p)
5262        return -EINVAL;
5263
5264    coremem = memparse(p, &p);
5265    *core = coremem >> PAGE_SHIFT;
5266
5267    /* Paranoid check that UL is enough for the coremem value */
5268    WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5269
5270    return 0;
5271}
5272
5273/*
5274 * kernelcore=size sets the amount of memory for use for allocations that
5275 * cannot be reclaimed or migrated.
5276 */
5277static int __init cmdline_parse_kernelcore(char *p)
5278{
5279    return cmdline_parse_core(p, &required_kernelcore);
5280}
5281
5282/*
5283 * movablecore=size sets the amount of memory for use for allocations that
5284 * can be reclaimed or migrated.
5285 */
5286static int __init cmdline_parse_movablecore(char *p)
5287{
5288    return cmdline_parse_core(p, &required_movablecore);
5289}
5290
5291early_param("kernelcore", cmdline_parse_kernelcore);
5292early_param("movablecore", cmdline_parse_movablecore);
5293
5294#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5295
5296void adjust_managed_page_count(struct page *page, long count)
5297{
5298    spin_lock(&managed_page_count_lock);
5299    page_zone(page)->managed_pages += count;
5300    totalram_pages += count;
5301#ifdef CONFIG_HIGHMEM
5302    if (PageHighMem(page))
5303        totalhigh_pages += count;
5304#endif
5305    spin_unlock(&managed_page_count_lock);
5306}
5307EXPORT_SYMBOL(adjust_managed_page_count);
5308
5309unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
5310{
5311    void *pos;
5312    unsigned long pages = 0;
5313
5314    start = (void *)PAGE_ALIGN((unsigned long)start);
5315    end = (void *)((unsigned long)end & PAGE_MASK);
5316    for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5317        if ((unsigned int)poison <= 0xFF)
5318            memset(pos, poison, PAGE_SIZE);
5319        free_reserved_page(virt_to_page(pos));
5320    }
5321
5322    if (pages && s)
5323        pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5324            s, pages << (PAGE_SHIFT - 10), start, end);
5325
5326    return pages;
5327}
5328EXPORT_SYMBOL(free_reserved_area);
5329
5330#ifdef CONFIG_HIGHMEM
5331void free_highmem_page(struct page *page)
5332{
5333    __free_reserved_page(page);
5334    totalram_pages++;
5335    page_zone(page)->managed_pages++;
5336    totalhigh_pages++;
5337}
5338#endif
5339
5340
5341void __init mem_init_print_info(const char *str)
5342{
5343    unsigned long physpages, codesize, datasize, rosize, bss_size;
5344    unsigned long init_code_size, init_data_size;
5345
5346    physpages = get_num_physpages();
5347    codesize = _etext - _stext;
5348    datasize = _edata - _sdata;
5349    rosize = __end_rodata - __start_rodata;
5350    bss_size = __bss_stop - __bss_start;
5351    init_data_size = __init_end - __init_begin;
5352    init_code_size = _einittext - _sinittext;
5353
5354    /*
5355     * Detect special cases and adjust section sizes accordingly:
5356     * 1) .init.* may be embedded into .data sections
5357     * 2) .init.text.* may be out of [__init_begin, __init_end],
5358     * please refer to arch/tile/kernel/vmlinux.lds.S.
5359     * 3) .rodata.* may be embedded into .text or .data sections.
5360     */
5361#define adj_init_size(start, end, size, pos, adj) \
5362    do { \
5363        if (start <= pos && pos < end && size > adj) \
5364            size -= adj; \
5365    } while (0)
5366
5367    adj_init_size(__init_begin, __init_end, init_data_size,
5368             _sinittext, init_code_size);
5369    adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
5370    adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
5371    adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
5372    adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
5373
5374#undef adj_init_size
5375
5376    printk("Memory: %luK/%luK available "
5377           "(%luK kernel code, %luK rwdata, %luK rodata, "
5378           "%luK init, %luK bss, %luK reserved"
5379#ifdef CONFIG_HIGHMEM
5380           ", %luK highmem"
5381#endif
5382           "%s%s)\n",
5383           nr_free_pages() << (PAGE_SHIFT-10), physpages << (PAGE_SHIFT-10),
5384           codesize >> 10, datasize >> 10, rosize >> 10,
5385           (init_data_size + init_code_size) >> 10, bss_size >> 10,
5386           (physpages - totalram_pages) << (PAGE_SHIFT-10),
5387#ifdef CONFIG_HIGHMEM
5388           totalhigh_pages << (PAGE_SHIFT-10),
5389#endif
5390           str ? ", " : "", str ? str : "");
5391}
5392
5393/**
5394 * set_dma_reserve - set the specified number of pages reserved in the first zone
5395 * @new_dma_reserve: The number of pages to mark reserved
5396 *
5397 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5398 * In the DMA zone, a significant percentage may be consumed by kernel image
5399 * and other unfreeable allocations which can skew the watermarks badly. This
5400 * function may optionally be used to account for unfreeable pages in the
5401 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5402 * smaller per-cpu batchsize.
5403 */
5404void __init set_dma_reserve(unsigned long new_dma_reserve)
5405{
5406    dma_reserve = new_dma_reserve;
5407}
5408
5409void __init free_area_init(unsigned long *zones_size)
5410{
5411    free_area_init_node(0, zones_size,
5412            __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5413}
5414
5415static int page_alloc_cpu_notify(struct notifier_block *self,
5416                 unsigned long action, void *hcpu)
5417{
5418    int cpu = (unsigned long)hcpu;
5419
5420    if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5421        lru_add_drain_cpu(cpu);
5422        drain_pages(cpu);
5423
5424        /*
5425         * Spill the event counters of the dead processor
5426         * into the current processors event counters.
5427         * This artificially elevates the count of the current
5428         * processor.
5429         */
5430        vm_events_fold_cpu(cpu);
5431
5432        /*
5433         * Zero the differential counters of the dead processor
5434         * so that the vm statistics are consistent.
5435         *
5436         * This is only okay since the processor is dead and cannot
5437         * race with what we are doing.
5438         */
5439        cpu_vm_stats_fold(cpu);
5440    }
5441    return NOTIFY_OK;
5442}
5443
5444void __init page_alloc_init(void)
5445{
5446    hotcpu_notifier(page_alloc_cpu_notify, 0);
5447}
5448
5449/*
5450 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5451 * or min_free_kbytes changes.
5452 */
5453static void calculate_totalreserve_pages(void)
5454{
5455    struct pglist_data *pgdat;
5456    unsigned long reserve_pages = 0;
5457    enum zone_type i, j;
5458
5459    for_each_online_pgdat(pgdat) {
5460        for (i = 0; i < MAX_NR_ZONES; i++) {
5461            struct zone *zone = pgdat->node_zones + i;
5462            unsigned long max = 0;
5463
5464            /* Find valid and maximum lowmem_reserve in the zone */
5465            for (j = i; j < MAX_NR_ZONES; j++) {
5466                if (zone->lowmem_reserve[j] > max)
5467                    max = zone->lowmem_reserve[j];
5468            }
5469
5470            /* we treat the high watermark as reserved pages. */
5471            max += high_wmark_pages(zone);
5472
5473            if (max > zone->managed_pages)
5474                max = zone->managed_pages;
5475            reserve_pages += max;
5476            /*
5477             * Lowmem reserves are not available to
5478             * GFP_HIGHUSER page cache allocations and
5479             * kswapd tries to balance zones to their high
5480             * watermark. As a result, neither should be
5481             * regarded as dirtyable memory, to prevent a
5482             * situation where reclaim has to clean pages
5483             * in order to balance the zones.
5484             */
5485            zone->dirty_balance_reserve = max;
5486        }
5487    }
5488    dirty_balance_reserve = reserve_pages;
5489    totalreserve_pages = reserve_pages;
5490}
5491
5492/*
5493 * setup_per_zone_lowmem_reserve - called whenever
5494 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5495 * has a correct pages reserved value, so an adequate number of
5496 * pages are left in the zone after a successful __alloc_pages().
5497 */
5498static void setup_per_zone_lowmem_reserve(void)
5499{
5500    struct pglist_data *pgdat;
5501    enum zone_type j, idx;
5502
5503    for_each_online_pgdat(pgdat) {
5504        for (j = 0; j < MAX_NR_ZONES; j++) {
5505            struct zone *zone = pgdat->node_zones + j;
5506            unsigned long managed_pages = zone->managed_pages;
5507
5508            zone->lowmem_reserve[j] = 0;
5509
5510            idx = j;
5511            while (idx) {
5512                struct zone *lower_zone;
5513
5514                idx--;
5515
5516                if (sysctl_lowmem_reserve_ratio[idx] < 1)
5517                    sysctl_lowmem_reserve_ratio[idx] = 1;
5518
5519                lower_zone = pgdat->node_zones + idx;
5520                lower_zone->lowmem_reserve[j] = managed_pages /
5521                    sysctl_lowmem_reserve_ratio[idx];
5522                managed_pages += lower_zone->managed_pages;
5523            }
5524        }
5525    }
5526
5527    /* update totalreserve_pages */
5528    calculate_totalreserve_pages();
5529}
5530
5531static void __setup_per_zone_wmarks(void)
5532{
5533    unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5534    unsigned long lowmem_pages = 0;
5535    struct zone *zone;
5536    unsigned long flags;
5537
5538    /* Calculate total number of !ZONE_HIGHMEM pages */
5539    for_each_zone(zone) {
5540        if (!is_highmem(zone))
5541            lowmem_pages += zone->managed_pages;
5542    }
5543
5544    for_each_zone(zone) {
5545        u64 tmp;
5546
5547        spin_lock_irqsave(&zone->lock, flags);
5548        tmp = (u64)pages_min * zone->managed_pages;
5549        do_div(tmp, lowmem_pages);
5550        if (is_highmem(zone)) {
5551            /*
5552             * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5553             * need highmem pages, so cap pages_min to a small
5554             * value here.
5555             *
5556             * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5557             * deltas controls asynch page reclaim, and so should
5558             * not be capped for highmem.
5559             */
5560            unsigned long min_pages;
5561
5562            min_pages = zone->managed_pages / 1024;
5563            min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
5564            zone->watermark[WMARK_MIN] = min_pages;
5565        } else {
5566            /*
5567             * If it's a lowmem zone, reserve a number of pages
5568             * proportionate to the zone's size.
5569             */
5570            zone->watermark[WMARK_MIN] = tmp;
5571        }
5572
5573        zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
5574        zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5575
5576        __mod_zone_page_state(zone, NR_ALLOC_BATCH,
5577                      high_wmark_pages(zone) -
5578                      low_wmark_pages(zone) -
5579                      zone_page_state(zone, NR_ALLOC_BATCH));
5580
5581        setup_zone_migrate_reserve(zone);
5582        spin_unlock_irqrestore(&zone->lock, flags);
5583    }
5584
5585    /* update totalreserve_pages */
5586    calculate_totalreserve_pages();
5587}
5588
5589/**
5590 * setup_per_zone_wmarks - called when min_free_kbytes changes
5591 * or when memory is hot-{added|removed}
5592 *
5593 * Ensures that the watermark[min,low,high] values for each zone are set
5594 * correctly with respect to min_free_kbytes.
5595 */
5596void setup_per_zone_wmarks(void)
5597{
5598    mutex_lock(&zonelists_mutex);
5599    __setup_per_zone_wmarks();
5600    mutex_unlock(&zonelists_mutex);
5601}
5602
5603/*
5604 * The inactive anon list should be small enough that the VM never has to
5605 * do too much work, but large enough that each inactive page has a chance
5606 * to be referenced again before it is swapped out.
5607 *
5608 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5609 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5610 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5611 * the anonymous pages are kept on the inactive list.
5612 *
5613 * total target max
5614 * memory ratio inactive anon
5615 * -------------------------------------
5616 * 10MB 1 5MB
5617 * 100MB 1 50MB
5618 * 1GB 3 250MB
5619 * 10GB 10 0.9GB
5620 * 100GB 31 3GB
5621 * 1TB 101 10GB
5622 * 10TB 320 32GB
5623 */
5624static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5625{
5626    unsigned int gb, ratio;
5627
5628    /* Zone size in gigabytes */
5629    gb = zone->managed_pages >> (30 - PAGE_SHIFT);
5630    if (gb)
5631        ratio = int_sqrt(10 * gb);
5632    else
5633        ratio = 1;
5634
5635    zone->inactive_ratio = ratio;
5636}
5637
5638static void __meminit setup_per_zone_inactive_ratio(void)
5639{
5640    struct zone *zone;
5641
5642    for_each_zone(zone)
5643        calculate_zone_inactive_ratio(zone);
5644}
5645
5646/*
5647 * Initialise min_free_kbytes.
5648 *
5649 * For small machines we want it small (128k min). For large machines
5650 * we want it large (64MB max). But it is not linear, because network
5651 * bandwidth does not increase linearly with machine size. We use
5652 *
5653 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5654 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5655 *
5656 * which yields
5657 *
5658 * 16MB: 512k
5659 * 32MB: 724k
5660 * 64MB: 1024k
5661 * 128MB: 1448k
5662 * 256MB: 2048k
5663 * 512MB: 2896k
5664 * 1024MB: 4096k
5665 * 2048MB: 5792k
5666 * 4096MB: 8192k
5667 * 8192MB: 11584k
5668 * 16384MB: 16384k
5669 */
5670int __meminit init_per_zone_wmark_min(void)
5671{
5672    unsigned long lowmem_kbytes;
5673    int new_min_free_kbytes;
5674
5675    lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5676    new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5677
5678    if (new_min_free_kbytes > user_min_free_kbytes) {
5679        min_free_kbytes = new_min_free_kbytes;
5680        if (min_free_kbytes < 128)
5681            min_free_kbytes = 128;
5682        if (min_free_kbytes > 65536)
5683            min_free_kbytes = 65536;
5684    } else {
5685        pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
5686                new_min_free_kbytes, user_min_free_kbytes);
5687    }
5688    setup_per_zone_wmarks();
5689    refresh_zone_stat_thresholds();
5690    setup_per_zone_lowmem_reserve();
5691    setup_per_zone_inactive_ratio();
5692    return 0;
5693}
5694module_init(init_per_zone_wmark_min)
5695
5696/*
5697 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5698 * that we can call two helper functions whenever min_free_kbytes
5699 * changes.
5700 */
5701int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
5702    void __user *buffer, size_t *length, loff_t *ppos)
5703{
5704    proc_dointvec(table, write, buffer, length, ppos);
5705    if (write) {
5706        user_min_free_kbytes = min_free_kbytes;
5707        setup_per_zone_wmarks();
5708    }
5709    return 0;
5710}
5711
5712#ifdef CONFIG_NUMA
5713int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
5714    void __user *buffer, size_t *length, loff_t *ppos)
5715{
5716    struct zone *zone;
5717    int rc;
5718
5719    rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5720    if (rc)
5721        return rc;
5722
5723    for_each_zone(zone)
5724        zone->min_unmapped_pages = (zone->managed_pages *
5725                sysctl_min_unmapped_ratio) / 100;
5726    return 0;
5727}
5728
5729int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
5730    void __user *buffer, size_t *length, loff_t *ppos)
5731{
5732    struct zone *zone;
5733    int rc;
5734
5735    rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5736    if (rc)
5737        return rc;
5738
5739    for_each_zone(zone)
5740        zone->min_slab_pages = (zone->managed_pages *
5741                sysctl_min_slab_ratio) / 100;
5742    return 0;
5743}
5744#endif
5745
5746/*
5747 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5748 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5749 * whenever sysctl_lowmem_reserve_ratio changes.
5750 *
5751 * The reserve ratio obviously has absolutely no relation with the
5752 * minimum watermarks. The lowmem reserve ratio can only make sense
5753 * if in function of the boot time zone sizes.
5754 */
5755int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
5756    void __user *buffer, size_t *length, loff_t *ppos)
5757{
5758    proc_dointvec_minmax(table, write, buffer, length, ppos);
5759    setup_per_zone_lowmem_reserve();
5760    return 0;
5761}
5762
5763/*
5764 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5765 * cpu. It is the fraction of total pages in each zone that a hot per cpu
5766 * pagelist can have before it gets flushed back to buddy allocator.
5767 */
5768int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
5769    void __user *buffer, size_t *length, loff_t *ppos)
5770{
5771    struct zone *zone;
5772    unsigned int cpu;
5773    int ret;
5774
5775    ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5776    if (!write || (ret < 0))
5777        return ret;
5778
5779    mutex_lock(&pcp_batch_high_lock);
5780    for_each_populated_zone(zone) {
5781        unsigned long high;
5782        high = zone->managed_pages / percpu_pagelist_fraction;
5783        for_each_possible_cpu(cpu)
5784            pageset_set_high(per_cpu_ptr(zone->pageset, cpu),
5785                     high);
5786    }
5787    mutex_unlock(&pcp_batch_high_lock);
5788    return 0;
5789}
5790
5791int hashdist = HASHDIST_DEFAULT;
5792
5793#ifdef CONFIG_NUMA
5794static int __init set_hashdist(char *str)
5795{
5796    if (!str)
5797        return 0;
5798    hashdist = simple_strtoul(str, &str, 0);
5799    return 1;
5800}
5801__setup("hashdist=", set_hashdist);
5802#endif
5803
5804/*
5805 * allocate a large system hash table from bootmem
5806 * - it is assumed that the hash table must contain an exact power-of-2
5807 * quantity of entries
5808 * - limit is the number of hash buckets, not the total allocation size
5809 */
5810void *__init alloc_large_system_hash(const char *tablename,
5811                     unsigned long bucketsize,
5812                     unsigned long numentries,
5813                     int scale,
5814                     int flags,
5815                     unsigned int *_hash_shift,
5816                     unsigned int *_hash_mask,
5817                     unsigned long low_limit,
5818                     unsigned long high_limit)
5819{
5820    unsigned long long max = high_limit;
5821    unsigned long log2qty, size;
5822    void *table = NULL;
5823
5824    /* allow the kernel cmdline to have a say */
5825    if (!numentries) {
5826        /* round applicable memory size up to nearest megabyte */
5827        numentries = nr_kernel_pages;