Loading...
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/kasan.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/memremap.h>
46#include <linux/stop_machine.h>
47#include <linux/sort.h>
48#include <linux/pfn.h>
49#include <linux/backing-dev.h>
50#include <linux/fault-inject.h>
51#include <linux/page-isolation.h>
52#include <linux/page_ext.h>
53#include <linux/debugobjects.h>
54#include <linux/kmemleak.h>
55#include <linux/compaction.h>
56#include <trace/events/kmem.h>
57#include <trace/events/oom.h>
58#include <linux/prefetch.h>
59#include <linux/mm_inline.h>
60#include <linux/migrate.h>
61#include <linux/hugetlb.h>
62#include <linux/sched/rt.h>
63#include <linux/sched/mm.h>
64#include <linux/page_owner.h>
65#include <linux/kthread.h>
66#include <linux/memcontrol.h>
67#include <linux/ftrace.h>
68#include <linux/lockdep.h>
69#include <linux/nmi.h>
70
71#include <asm/sections.h>
72#include <asm/tlbflush.h>
73#include <asm/div64.h>
74#include "internal.h"
75
76/* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
77static DEFINE_MUTEX(pcp_batch_high_lock);
78#define MIN_PERCPU_PAGELIST_FRACTION (8)
79
80#ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
81DEFINE_PER_CPU(int, numa_node);
82EXPORT_PER_CPU_SYMBOL(numa_node);
83#endif
84
85DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
86
87#ifdef CONFIG_HAVE_MEMORYLESS_NODES
88/*
89 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
90 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
91 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
92 * defined in <linux/topology.h>.
93 */
94DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
95EXPORT_PER_CPU_SYMBOL(_numa_mem_);
96int _node_numa_mem_[MAX_NUMNODES];
97#endif
98
99/* work_structs for global per-cpu drains */
100DEFINE_MUTEX(pcpu_drain_mutex);
101DEFINE_PER_CPU(struct work_struct, pcpu_drain);
102
103#ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
104volatile unsigned long latent_entropy __latent_entropy;
105EXPORT_SYMBOL(latent_entropy);
106#endif
107
108/*
109 * Array of node states.
110 */
111nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
112 [N_POSSIBLE] = NODE_MASK_ALL,
113 [N_ONLINE] = { { [0] = 1UL } },
114#ifndef CONFIG_NUMA
115 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
116#ifdef CONFIG_HIGHMEM
117 [N_HIGH_MEMORY] = { { [0] = 1UL } },
118#endif
119 [N_MEMORY] = { { [0] = 1UL } },
120 [N_CPU] = { { [0] = 1UL } },
121#endif /* NUMA */
122};
123EXPORT_SYMBOL(node_states);
124
125/* Protect totalram_pages and zone->managed_pages */
126static DEFINE_SPINLOCK(managed_page_count_lock);
127
128unsigned long totalram_pages __read_mostly;
129unsigned long totalreserve_pages __read_mostly;
130unsigned long totalcma_pages __read_mostly;
131
132int percpu_pagelist_fraction;
133gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
134
135/*
136 * A cached value of the page's pageblock's migratetype, used when the page is
137 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
138 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
139 * Also the migratetype set in the page does not necessarily match the pcplist
140 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
141 * other index - this ensures that it will be put on the correct CMA freelist.
142 */
143static inline int get_pcppage_migratetype(struct page *page)
144{
145 return page->index;
146}
147
148static inline void set_pcppage_migratetype(struct page *page, int migratetype)
149{
150 page->index = migratetype;
151}
152
153#ifdef CONFIG_PM_SLEEP
154/*
155 * The following functions are used by the suspend/hibernate code to temporarily
156 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
157 * while devices are suspended. To avoid races with the suspend/hibernate code,
158 * they should always be called with pm_mutex held (gfp_allowed_mask also should
159 * only be modified with pm_mutex held, unless the suspend/hibernate code is
160 * guaranteed not to run in parallel with that modification).
161 */
162
163static gfp_t saved_gfp_mask;
164
165void pm_restore_gfp_mask(void)
166{
167 WARN_ON(!mutex_is_locked(&pm_mutex));
168 if (saved_gfp_mask) {
169 gfp_allowed_mask = saved_gfp_mask;
170 saved_gfp_mask = 0;
171 }
172}
173
174void pm_restrict_gfp_mask(void)
175{
176 WARN_ON(!mutex_is_locked(&pm_mutex));
177 WARN_ON(saved_gfp_mask);
178 saved_gfp_mask = gfp_allowed_mask;
179 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
180}
181
182bool pm_suspended_storage(void)
183{
184 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
185 return false;
186 return true;
187}
188#endif /* CONFIG_PM_SLEEP */
189
190#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
191unsigned int pageblock_order __read_mostly;
192#endif
193
194static void __free_pages_ok(struct page *page, unsigned int order);
195
196/*
197 * results with 256, 32 in the lowmem_reserve sysctl:
198 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
199 * 1G machine -> (16M dma, 784M normal, 224M high)
200 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
201 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
202 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
203 *
204 * TBD: should special case ZONE_DMA32 machines here - in those we normally
205 * don't need any ZONE_NORMAL reservation
206 */
207int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
208#ifdef CONFIG_ZONE_DMA
209 [ZONE_DMA] = 256,
210#endif
211#ifdef CONFIG_ZONE_DMA32
212 [ZONE_DMA32] = 256,
213#endif
214 [ZONE_NORMAL] = 32,
215#ifdef CONFIG_HIGHMEM
216 [ZONE_HIGHMEM] = 0,
217#endif
218 [ZONE_MOVABLE] = 0,
219};
220
221EXPORT_SYMBOL(totalram_pages);
222
223static char * const zone_names[MAX_NR_ZONES] = {
224#ifdef CONFIG_ZONE_DMA
225 "DMA",
226#endif
227#ifdef CONFIG_ZONE_DMA32
228 "DMA32",
229#endif
230 "Normal",
231#ifdef CONFIG_HIGHMEM
232 "HighMem",
233#endif
234 "Movable",
235#ifdef CONFIG_ZONE_DEVICE
236 "Device",
237#endif
238};
239
240char * const migratetype_names[MIGRATE_TYPES] = {
241 "Unmovable",
242 "Movable",
243 "Reclaimable",
244 "HighAtomic",
245#ifdef CONFIG_CMA
246 "CMA",
247#endif
248#ifdef CONFIG_MEMORY_ISOLATION
249 "Isolate",
250#endif
251};
252
253compound_page_dtor * const compound_page_dtors[] = {
254 NULL,
255 free_compound_page,
256#ifdef CONFIG_HUGETLB_PAGE
257 free_huge_page,
258#endif
259#ifdef CONFIG_TRANSPARENT_HUGEPAGE
260 free_transhuge_page,
261#endif
262};
263
264int min_free_kbytes = 1024;
265int user_min_free_kbytes = -1;
266int watermark_scale_factor = 10;
267
268static unsigned long nr_kernel_pages __meminitdata;
269static unsigned long nr_all_pages __meminitdata;
270static unsigned long dma_reserve __meminitdata;
271
272#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
273static unsigned long arch_zone_lowest_possible_pfn[MAX_NR_ZONES] __meminitdata;
274static unsigned long arch_zone_highest_possible_pfn[MAX_NR_ZONES] __meminitdata;
275static unsigned long required_kernelcore __initdata;
276static unsigned long required_kernelcore_percent __initdata;
277static unsigned long required_movablecore __initdata;
278static unsigned long required_movablecore_percent __initdata;
279static unsigned long zone_movable_pfn[MAX_NUMNODES] __meminitdata;
280static bool mirrored_kernelcore __meminitdata;
281
282/* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
283int movable_zone;
284EXPORT_SYMBOL(movable_zone);
285#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
286
287#if MAX_NUMNODES > 1
288int nr_node_ids __read_mostly = MAX_NUMNODES;
289int nr_online_nodes __read_mostly = 1;
290EXPORT_SYMBOL(nr_node_ids);
291EXPORT_SYMBOL(nr_online_nodes);
292#endif
293
294int page_group_by_mobility_disabled __read_mostly;
295
296#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
297/* Returns true if the struct page for the pfn is uninitialised */
298static inline bool __meminit early_page_uninitialised(unsigned long pfn)
299{
300 int nid = early_pfn_to_nid(pfn);
301
302 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
303 return true;
304
305 return false;
306}
307
308/*
309 * Returns false when the remaining initialisation should be deferred until
310 * later in the boot cycle when it can be parallelised.
311 */
312static inline bool update_defer_init(pg_data_t *pgdat,
313 unsigned long pfn, unsigned long zone_end,
314 unsigned long *nr_initialised)
315{
316 /* Always populate low zones for address-constrained allocations */
317 if (zone_end < pgdat_end_pfn(pgdat))
318 return true;
319 (*nr_initialised)++;
320 if ((*nr_initialised > pgdat->static_init_pgcnt) &&
321 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
322 pgdat->first_deferred_pfn = pfn;
323 return false;
324 }
325
326 return true;
327}
328#else
329static inline bool early_page_uninitialised(unsigned long pfn)
330{
331 return false;
332}
333
334static inline bool update_defer_init(pg_data_t *pgdat,
335 unsigned long pfn, unsigned long zone_end,
336 unsigned long *nr_initialised)
337{
338 return true;
339}
340#endif
341
342/* Return a pointer to the bitmap storing bits affecting a block of pages */
343static inline unsigned long *get_pageblock_bitmap(struct page *page,
344 unsigned long pfn)
345{
346#ifdef CONFIG_SPARSEMEM
347 return __pfn_to_section(pfn)->pageblock_flags;
348#else
349 return page_zone(page)->pageblock_flags;
350#endif /* CONFIG_SPARSEMEM */
351}
352
353static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
354{
355#ifdef CONFIG_SPARSEMEM
356 pfn &= (PAGES_PER_SECTION-1);
357 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
358#else
359 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
360 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
361#endif /* CONFIG_SPARSEMEM */
362}
363
364/**
365 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
366 * @page: The page within the block of interest
367 * @pfn: The target page frame number
368 * @end_bitidx: The last bit of interest to retrieve
369 * @mask: mask of bits that the caller is interested in
370 *
371 * Return: pageblock_bits flags
372 */
373static __always_inline unsigned long __get_pfnblock_flags_mask(struct page *page,
374 unsigned long pfn,
375 unsigned long end_bitidx,
376 unsigned long mask)
377{
378 unsigned long *bitmap;
379 unsigned long bitidx, word_bitidx;
380 unsigned long word;
381
382 bitmap = get_pageblock_bitmap(page, pfn);
383 bitidx = pfn_to_bitidx(page, pfn);
384 word_bitidx = bitidx / BITS_PER_LONG;
385 bitidx &= (BITS_PER_LONG-1);
386
387 word = bitmap[word_bitidx];
388 bitidx += end_bitidx;
389 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
390}
391
392unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
393 unsigned long end_bitidx,
394 unsigned long mask)
395{
396 return __get_pfnblock_flags_mask(page, pfn, end_bitidx, mask);
397}
398
399static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
400{
401 return __get_pfnblock_flags_mask(page, pfn, PB_migrate_end, MIGRATETYPE_MASK);
402}
403
404/**
405 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
406 * @page: The page within the block of interest
407 * @flags: The flags to set
408 * @pfn: The target page frame number
409 * @end_bitidx: The last bit of interest
410 * @mask: mask of bits that the caller is interested in
411 */
412void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
413 unsigned long pfn,
414 unsigned long end_bitidx,
415 unsigned long mask)
416{
417 unsigned long *bitmap;
418 unsigned long bitidx, word_bitidx;
419 unsigned long old_word, word;
420
421 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
422
423 bitmap = get_pageblock_bitmap(page, pfn);
424 bitidx = pfn_to_bitidx(page, pfn);
425 word_bitidx = bitidx / BITS_PER_LONG;
426 bitidx &= (BITS_PER_LONG-1);
427
428 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
429
430 bitidx += end_bitidx;
431 mask <<= (BITS_PER_LONG - bitidx - 1);
432 flags <<= (BITS_PER_LONG - bitidx - 1);
433
434 word = READ_ONCE(bitmap[word_bitidx]);
435 for (;;) {
436 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
437 if (word == old_word)
438 break;
439 word = old_word;
440 }
441}
442
443void set_pageblock_migratetype(struct page *page, int migratetype)
444{
445 if (unlikely(page_group_by_mobility_disabled &&
446 migratetype < MIGRATE_PCPTYPES))
447 migratetype = MIGRATE_UNMOVABLE;
448
449 set_pageblock_flags_group(page, (unsigned long)migratetype,
450 PB_migrate, PB_migrate_end);
451}
452
453#ifdef CONFIG_DEBUG_VM
454static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
455{
456 int ret = 0;
457 unsigned seq;
458 unsigned long pfn = page_to_pfn(page);
459 unsigned long sp, start_pfn;
460
461 do {
462 seq = zone_span_seqbegin(zone);
463 start_pfn = zone->zone_start_pfn;
464 sp = zone->spanned_pages;
465 if (!zone_spans_pfn(zone, pfn))
466 ret = 1;
467 } while (zone_span_seqretry(zone, seq));
468
469 if (ret)
470 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
471 pfn, zone_to_nid(zone), zone->name,
472 start_pfn, start_pfn + sp);
473
474 return ret;
475}
476
477static int page_is_consistent(struct zone *zone, struct page *page)
478{
479 if (!pfn_valid_within(page_to_pfn(page)))
480 return 0;
481 if (zone != page_zone(page))
482 return 0;
483
484 return 1;
485}
486/*
487 * Temporary debugging check for pages not lying within a given zone.
488 */
489static int __maybe_unused bad_range(struct zone *zone, struct page *page)
490{
491 if (page_outside_zone_boundaries(zone, page))
492 return 1;
493 if (!page_is_consistent(zone, page))
494 return 1;
495
496 return 0;
497}
498#else
499static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
500{
501 return 0;
502}
503#endif
504
505static void bad_page(struct page *page, const char *reason,
506 unsigned long bad_flags)
507{
508 static unsigned long resume;
509 static unsigned long nr_shown;
510 static unsigned long nr_unshown;
511
512 /*
513 * Allow a burst of 60 reports, then keep quiet for that minute;
514 * or allow a steady drip of one report per second.
515 */
516 if (nr_shown == 60) {
517 if (time_before(jiffies, resume)) {
518 nr_unshown++;
519 goto out;
520 }
521 if (nr_unshown) {
522 pr_alert(
523 "BUG: Bad page state: %lu messages suppressed\n",
524 nr_unshown);
525 nr_unshown = 0;
526 }
527 nr_shown = 0;
528 }
529 if (nr_shown++ == 0)
530 resume = jiffies + 60 * HZ;
531
532 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
533 current->comm, page_to_pfn(page));
534 __dump_page(page, reason);
535 bad_flags &= page->flags;
536 if (bad_flags)
537 pr_alert("bad because of flags: %#lx(%pGp)\n",
538 bad_flags, &bad_flags);
539 dump_page_owner(page);
540
541 print_modules();
542 dump_stack();
543out:
544 /* Leave bad fields for debug, except PageBuddy could make trouble */
545 page_mapcount_reset(page); /* remove PageBuddy */
546 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
547}
548
549/*
550 * Higher-order pages are called "compound pages". They are structured thusly:
551 *
552 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
553 *
554 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
555 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
556 *
557 * The first tail page's ->compound_dtor holds the offset in array of compound
558 * page destructors. See compound_page_dtors.
559 *
560 * The first tail page's ->compound_order holds the order of allocation.
561 * This usage means that zero-order pages may not be compound.
562 */
563
564void free_compound_page(struct page *page)
565{
566 __free_pages_ok(page, compound_order(page));
567}
568
569void prep_compound_page(struct page *page, unsigned int order)
570{
571 int i;
572 int nr_pages = 1 << order;
573
574 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
575 set_compound_order(page, order);
576 __SetPageHead(page);
577 for (i = 1; i < nr_pages; i++) {
578 struct page *p = page + i;
579 set_page_count(p, 0);
580 p->mapping = TAIL_MAPPING;
581 set_compound_head(p, page);
582 }
583 atomic_set(compound_mapcount_ptr(page), -1);
584}
585
586#ifdef CONFIG_DEBUG_PAGEALLOC
587unsigned int _debug_guardpage_minorder;
588bool _debug_pagealloc_enabled __read_mostly
589 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
590EXPORT_SYMBOL(_debug_pagealloc_enabled);
591bool _debug_guardpage_enabled __read_mostly;
592
593static int __init early_debug_pagealloc(char *buf)
594{
595 if (!buf)
596 return -EINVAL;
597 return kstrtobool(buf, &_debug_pagealloc_enabled);
598}
599early_param("debug_pagealloc", early_debug_pagealloc);
600
601static bool need_debug_guardpage(void)
602{
603 /* If we don't use debug_pagealloc, we don't need guard page */
604 if (!debug_pagealloc_enabled())
605 return false;
606
607 if (!debug_guardpage_minorder())
608 return false;
609
610 return true;
611}
612
613static void init_debug_guardpage(void)
614{
615 if (!debug_pagealloc_enabled())
616 return;
617
618 if (!debug_guardpage_minorder())
619 return;
620
621 _debug_guardpage_enabled = true;
622}
623
624struct page_ext_operations debug_guardpage_ops = {
625 .need = need_debug_guardpage,
626 .init = init_debug_guardpage,
627};
628
629static int __init debug_guardpage_minorder_setup(char *buf)
630{
631 unsigned long res;
632
633 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
634 pr_err("Bad debug_guardpage_minorder value\n");
635 return 0;
636 }
637 _debug_guardpage_minorder = res;
638 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
639 return 0;
640}
641early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
642
643static inline bool set_page_guard(struct zone *zone, struct page *page,
644 unsigned int order, int migratetype)
645{
646 struct page_ext *page_ext;
647
648 if (!debug_guardpage_enabled())
649 return false;
650
651 if (order >= debug_guardpage_minorder())
652 return false;
653
654 page_ext = lookup_page_ext(page);
655 if (unlikely(!page_ext))
656 return false;
657
658 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
659
660 INIT_LIST_HEAD(&page->lru);
661 set_page_private(page, order);
662 /* Guard pages are not available for any usage */
663 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
664
665 return true;
666}
667
668static inline void clear_page_guard(struct zone *zone, struct page *page,
669 unsigned int order, int migratetype)
670{
671 struct page_ext *page_ext;
672
673 if (!debug_guardpage_enabled())
674 return;
675
676 page_ext = lookup_page_ext(page);
677 if (unlikely(!page_ext))
678 return;
679
680 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
681
682 set_page_private(page, 0);
683 if (!is_migrate_isolate(migratetype))
684 __mod_zone_freepage_state(zone, (1 << order), migratetype);
685}
686#else
687struct page_ext_operations debug_guardpage_ops;
688static inline bool set_page_guard(struct zone *zone, struct page *page,
689 unsigned int order, int migratetype) { return false; }
690static inline void clear_page_guard(struct zone *zone, struct page *page,
691 unsigned int order, int migratetype) {}
692#endif
693
694static inline void set_page_order(struct page *page, unsigned int order)
695{
696 set_page_private(page, order);
697 __SetPageBuddy(page);
698}
699
700static inline void rmv_page_order(struct page *page)
701{
702 __ClearPageBuddy(page);
703 set_page_private(page, 0);
704}
705
706/*
707 * This function checks whether a page is free && is the buddy
708 * we can do coalesce a page and its buddy if
709 * (a) the buddy is not in a hole (check before calling!) &&
710 * (b) the buddy is in the buddy system &&
711 * (c) a page and its buddy have the same order &&
712 * (d) a page and its buddy are in the same zone.
713 *
714 * For recording whether a page is in the buddy system, we set ->_mapcount
715 * PAGE_BUDDY_MAPCOUNT_VALUE.
716 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
717 * serialized by zone->lock.
718 *
719 * For recording page's order, we use page_private(page).
720 */
721static inline int page_is_buddy(struct page *page, struct page *buddy,
722 unsigned int order)
723{
724 if (page_is_guard(buddy) && page_order(buddy) == order) {
725 if (page_zone_id(page) != page_zone_id(buddy))
726 return 0;
727
728 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
729
730 return 1;
731 }
732
733 if (PageBuddy(buddy) && page_order(buddy) == order) {
734 /*
735 * zone check is done late to avoid uselessly
736 * calculating zone/node ids for pages that could
737 * never merge.
738 */
739 if (page_zone_id(page) != page_zone_id(buddy))
740 return 0;
741
742 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
743
744 return 1;
745 }
746 return 0;
747}
748
749/*
750 * Freeing function for a buddy system allocator.
751 *
752 * The concept of a buddy system is to maintain direct-mapped table
753 * (containing bit values) for memory blocks of various "orders".
754 * The bottom level table contains the map for the smallest allocatable
755 * units of memory (here, pages), and each level above it describes
756 * pairs of units from the levels below, hence, "buddies".
757 * At a high level, all that happens here is marking the table entry
758 * at the bottom level available, and propagating the changes upward
759 * as necessary, plus some accounting needed to play nicely with other
760 * parts of the VM system.
761 * At each level, we keep a list of pages, which are heads of continuous
762 * free pages of length of (1 << order) and marked with _mapcount
763 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
764 * field.
765 * So when we are allocating or freeing one, we can derive the state of the
766 * other. That is, if we allocate a small block, and both were
767 * free, the remainder of the region must be split into blocks.
768 * If a block is freed, and its buddy is also free, then this
769 * triggers coalescing into a block of larger size.
770 *
771 * -- nyc
772 */
773
774static inline void __free_one_page(struct page *page,
775 unsigned long pfn,
776 struct zone *zone, unsigned int order,
777 int migratetype)
778{
779 unsigned long combined_pfn;
780 unsigned long uninitialized_var(buddy_pfn);
781 struct page *buddy;
782 unsigned int max_order;
783
784 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
785
786 VM_BUG_ON(!zone_is_initialized(zone));
787 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
788
789 VM_BUG_ON(migratetype == -1);
790 if (likely(!is_migrate_isolate(migratetype)))
791 __mod_zone_freepage_state(zone, 1 << order, migratetype);
792
793 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
794 VM_BUG_ON_PAGE(bad_range(zone, page), page);
795
796continue_merging:
797 while (order < max_order - 1) {
798 buddy_pfn = __find_buddy_pfn(pfn, order);
799 buddy = page + (buddy_pfn - pfn);
800
801 if (!pfn_valid_within(buddy_pfn))
802 goto done_merging;
803 if (!page_is_buddy(page, buddy, order))
804 goto done_merging;
805 /*
806 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
807 * merge with it and move up one order.
808 */
809 if (page_is_guard(buddy)) {
810 clear_page_guard(zone, buddy, order, migratetype);
811 } else {
812 list_del(&buddy->lru);
813 zone->free_area[order].nr_free--;
814 rmv_page_order(buddy);
815 }
816 combined_pfn = buddy_pfn & pfn;
817 page = page + (combined_pfn - pfn);
818 pfn = combined_pfn;
819 order++;
820 }
821 if (max_order < MAX_ORDER) {
822 /* If we are here, it means order is >= pageblock_order.
823 * We want to prevent merge between freepages on isolate
824 * pageblock and normal pageblock. Without this, pageblock
825 * isolation could cause incorrect freepage or CMA accounting.
826 *
827 * We don't want to hit this code for the more frequent
828 * low-order merging.
829 */
830 if (unlikely(has_isolate_pageblock(zone))) {
831 int buddy_mt;
832
833 buddy_pfn = __find_buddy_pfn(pfn, order);
834 buddy = page + (buddy_pfn - pfn);
835 buddy_mt = get_pageblock_migratetype(buddy);
836
837 if (migratetype != buddy_mt
838 && (is_migrate_isolate(migratetype) ||
839 is_migrate_isolate(buddy_mt)))
840 goto done_merging;
841 }
842 max_order++;
843 goto continue_merging;
844 }
845
846done_merging:
847 set_page_order(page, order);
848
849 /*
850 * If this is not the largest possible page, check if the buddy
851 * of the next-highest order is free. If it is, it's possible
852 * that pages are being freed that will coalesce soon. In case,
853 * that is happening, add the free page to the tail of the list
854 * so it's less likely to be used soon and more likely to be merged
855 * as a higher order page
856 */
857 if ((order < MAX_ORDER-2) && pfn_valid_within(buddy_pfn)) {
858 struct page *higher_page, *higher_buddy;
859 combined_pfn = buddy_pfn & pfn;
860 higher_page = page + (combined_pfn - pfn);
861 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
862 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
863 if (pfn_valid_within(buddy_pfn) &&
864 page_is_buddy(higher_page, higher_buddy, order + 1)) {
865 list_add_tail(&page->lru,
866 &zone->free_area[order].free_list[migratetype]);
867 goto out;
868 }
869 }
870
871 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
872out:
873 zone->free_area[order].nr_free++;
874}
875
876/*
877 * A bad page could be due to a number of fields. Instead of multiple branches,
878 * try and check multiple fields with one check. The caller must do a detailed
879 * check if necessary.
880 */
881static inline bool page_expected_state(struct page *page,
882 unsigned long check_flags)
883{
884 if (unlikely(atomic_read(&page->_mapcount) != -1))
885 return false;
886
887 if (unlikely((unsigned long)page->mapping |
888 page_ref_count(page) |
889#ifdef CONFIG_MEMCG
890 (unsigned long)page->mem_cgroup |
891#endif
892 (page->flags & check_flags)))
893 return false;
894
895 return true;
896}
897
898static void free_pages_check_bad(struct page *page)
899{
900 const char *bad_reason;
901 unsigned long bad_flags;
902
903 bad_reason = NULL;
904 bad_flags = 0;
905
906 if (unlikely(atomic_read(&page->_mapcount) != -1))
907 bad_reason = "nonzero mapcount";
908 if (unlikely(page->mapping != NULL))
909 bad_reason = "non-NULL mapping";
910 if (unlikely(page_ref_count(page) != 0))
911 bad_reason = "nonzero _refcount";
912 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
913 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
914 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
915 }
916#ifdef CONFIG_MEMCG
917 if (unlikely(page->mem_cgroup))
918 bad_reason = "page still charged to cgroup";
919#endif
920 bad_page(page, bad_reason, bad_flags);
921}
922
923static inline int free_pages_check(struct page *page)
924{
925 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
926 return 0;
927
928 /* Something has gone sideways, find it */
929 free_pages_check_bad(page);
930 return 1;
931}
932
933static int free_tail_pages_check(struct page *head_page, struct page *page)
934{
935 int ret = 1;
936
937 /*
938 * We rely page->lru.next never has bit 0 set, unless the page
939 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
940 */
941 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
942
943 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
944 ret = 0;
945 goto out;
946 }
947 switch (page - head_page) {
948 case 1:
949 /* the first tail page: ->mapping is compound_mapcount() */
950 if (unlikely(compound_mapcount(page))) {
951 bad_page(page, "nonzero compound_mapcount", 0);
952 goto out;
953 }
954 break;
955 case 2:
956 /*
957 * the second tail page: ->mapping is
958 * page_deferred_list().next -- ignore value.
959 */
960 break;
961 default:
962 if (page->mapping != TAIL_MAPPING) {
963 bad_page(page, "corrupted mapping in tail page", 0);
964 goto out;
965 }
966 break;
967 }
968 if (unlikely(!PageTail(page))) {
969 bad_page(page, "PageTail not set", 0);
970 goto out;
971 }
972 if (unlikely(compound_head(page) != head_page)) {
973 bad_page(page, "compound_head not consistent", 0);
974 goto out;
975 }
976 ret = 0;
977out:
978 page->mapping = NULL;
979 clear_compound_head(page);
980 return ret;
981}
982
983static __always_inline bool free_pages_prepare(struct page *page,
984 unsigned int order, bool check_free)
985{
986 int bad = 0;
987
988 VM_BUG_ON_PAGE(PageTail(page), page);
989
990 trace_mm_page_free(page, order);
991
992 /*
993 * Check tail pages before head page information is cleared to
994 * avoid checking PageCompound for order-0 pages.
995 */
996 if (unlikely(order)) {
997 bool compound = PageCompound(page);
998 int i;
999
1000 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1001
1002 if (compound)
1003 ClearPageDoubleMap(page);
1004 for (i = 1; i < (1 << order); i++) {
1005 if (compound)
1006 bad += free_tail_pages_check(page, page + i);
1007 if (unlikely(free_pages_check(page + i))) {
1008 bad++;
1009 continue;
1010 }
1011 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1012 }
1013 }
1014 if (PageMappingFlags(page))
1015 page->mapping = NULL;
1016 if (memcg_kmem_enabled() && PageKmemcg(page))
1017 memcg_kmem_uncharge(page, order);
1018 if (check_free)
1019 bad += free_pages_check(page);
1020 if (bad)
1021 return false;
1022
1023 page_cpupid_reset_last(page);
1024 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1025 reset_page_owner(page, order);
1026
1027 if (!PageHighMem(page)) {
1028 debug_check_no_locks_freed(page_address(page),
1029 PAGE_SIZE << order);
1030 debug_check_no_obj_freed(page_address(page),
1031 PAGE_SIZE << order);
1032 }
1033 arch_free_page(page, order);
1034 kernel_poison_pages(page, 1 << order, 0);
1035 kernel_map_pages(page, 1 << order, 0);
1036 kasan_free_pages(page, order);
1037
1038 return true;
1039}
1040
1041#ifdef CONFIG_DEBUG_VM
1042static inline bool free_pcp_prepare(struct page *page)
1043{
1044 return free_pages_prepare(page, 0, true);
1045}
1046
1047static inline bool bulkfree_pcp_prepare(struct page *page)
1048{
1049 return false;
1050}
1051#else
1052static bool free_pcp_prepare(struct page *page)
1053{
1054 return free_pages_prepare(page, 0, false);
1055}
1056
1057static bool bulkfree_pcp_prepare(struct page *page)
1058{
1059 return free_pages_check(page);
1060}
1061#endif /* CONFIG_DEBUG_VM */
1062
1063static inline void prefetch_buddy(struct page *page)
1064{
1065 unsigned long pfn = page_to_pfn(page);
1066 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1067 struct page *buddy = page + (buddy_pfn - pfn);
1068
1069 prefetch(buddy);
1070}
1071
1072/*
1073 * Frees a number of pages from the PCP lists
1074 * Assumes all pages on list are in same zone, and of same order.
1075 * count is the number of pages to free.
1076 *
1077 * If the zone was previously in an "all pages pinned" state then look to
1078 * see if this freeing clears that state.
1079 *
1080 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1081 * pinned" detection logic.
1082 */
1083static void free_pcppages_bulk(struct zone *zone, int count,
1084 struct per_cpu_pages *pcp)
1085{
1086 int migratetype = 0;
1087 int batch_free = 0;
1088 int prefetch_nr = 0;
1089 bool isolated_pageblocks;
1090 struct page *page, *tmp;
1091 LIST_HEAD(head);
1092
1093 while (count) {
1094 struct list_head *list;
1095
1096 /*
1097 * Remove pages from lists in a round-robin fashion. A
1098 * batch_free count is maintained that is incremented when an
1099 * empty list is encountered. This is so more pages are freed
1100 * off fuller lists instead of spinning excessively around empty
1101 * lists
1102 */
1103 do {
1104 batch_free++;
1105 if (++migratetype == MIGRATE_PCPTYPES)
1106 migratetype = 0;
1107 list = &pcp->lists[migratetype];
1108 } while (list_empty(list));
1109
1110 /* This is the only non-empty list. Free them all. */
1111 if (batch_free == MIGRATE_PCPTYPES)
1112 batch_free = count;
1113
1114 do {
1115 page = list_last_entry(list, struct page, lru);
1116 /* must delete to avoid corrupting pcp list */
1117 list_del(&page->lru);
1118 pcp->count--;
1119
1120 if (bulkfree_pcp_prepare(page))
1121 continue;
1122
1123 list_add_tail(&page->lru, &head);
1124
1125 /*
1126 * We are going to put the page back to the global
1127 * pool, prefetch its buddy to speed up later access
1128 * under zone->lock. It is believed the overhead of
1129 * an additional test and calculating buddy_pfn here
1130 * can be offset by reduced memory latency later. To
1131 * avoid excessive prefetching due to large count, only
1132 * prefetch buddy for the first pcp->batch nr of pages.
1133 */
1134 if (prefetch_nr++ < pcp->batch)
1135 prefetch_buddy(page);
1136 } while (--count && --batch_free && !list_empty(list));
1137 }
1138
1139 spin_lock(&zone->lock);
1140 isolated_pageblocks = has_isolate_pageblock(zone);
1141
1142 /*
1143 * Use safe version since after __free_one_page(),
1144 * page->lru.next will not point to original list.
1145 */
1146 list_for_each_entry_safe(page, tmp, &head, lru) {
1147 int mt = get_pcppage_migratetype(page);
1148 /* MIGRATE_ISOLATE page should not go to pcplists */
1149 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1150 /* Pageblock could have been isolated meanwhile */
1151 if (unlikely(isolated_pageblocks))
1152 mt = get_pageblock_migratetype(page);
1153
1154 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
1155 trace_mm_page_pcpu_drain(page, 0, mt);
1156 }
1157 spin_unlock(&zone->lock);
1158}
1159
1160static void free_one_page(struct zone *zone,
1161 struct page *page, unsigned long pfn,
1162 unsigned int order,
1163 int migratetype)
1164{
1165 spin_lock(&zone->lock);
1166 if (unlikely(has_isolate_pageblock(zone) ||
1167 is_migrate_isolate(migratetype))) {
1168 migratetype = get_pfnblock_migratetype(page, pfn);
1169 }
1170 __free_one_page(page, pfn, zone, order, migratetype);
1171 spin_unlock(&zone->lock);
1172}
1173
1174static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1175 unsigned long zone, int nid)
1176{
1177 mm_zero_struct_page(page);
1178 set_page_links(page, zone, nid, pfn);
1179 init_page_count(page);
1180 page_mapcount_reset(page);
1181 page_cpupid_reset_last(page);
1182
1183 INIT_LIST_HEAD(&page->lru);
1184#ifdef WANT_PAGE_VIRTUAL
1185 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1186 if (!is_highmem_idx(zone))
1187 set_page_address(page, __va(pfn << PAGE_SHIFT));
1188#endif
1189}
1190
1191#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1192static void __meminit init_reserved_page(unsigned long pfn)
1193{
1194 pg_data_t *pgdat;
1195 int nid, zid;
1196
1197 if (!early_page_uninitialised(pfn))
1198 return;
1199
1200 nid = early_pfn_to_nid(pfn);
1201 pgdat = NODE_DATA(nid);
1202
1203 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1204 struct zone *zone = &pgdat->node_zones[zid];
1205
1206 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1207 break;
1208 }
1209 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1210}
1211#else
1212static inline void init_reserved_page(unsigned long pfn)
1213{
1214}
1215#endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1216
1217/*
1218 * Initialised pages do not have PageReserved set. This function is
1219 * called for each range allocated by the bootmem allocator and
1220 * marks the pages PageReserved. The remaining valid pages are later
1221 * sent to the buddy page allocator.
1222 */
1223void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1224{
1225 unsigned long start_pfn = PFN_DOWN(start);
1226 unsigned long end_pfn = PFN_UP(end);
1227
1228 for (; start_pfn < end_pfn; start_pfn++) {
1229 if (pfn_valid(start_pfn)) {
1230 struct page *page = pfn_to_page(start_pfn);
1231
1232 init_reserved_page(start_pfn);
1233
1234 /* Avoid false-positive PageTail() */
1235 INIT_LIST_HEAD(&page->lru);
1236
1237 SetPageReserved(page);
1238 }
1239 }
1240}
1241
1242static void __free_pages_ok(struct page *page, unsigned int order)
1243{
1244 unsigned long flags;
1245 int migratetype;
1246 unsigned long pfn = page_to_pfn(page);
1247
1248 if (!free_pages_prepare(page, order, true))
1249 return;
1250
1251 migratetype = get_pfnblock_migratetype(page, pfn);
1252 local_irq_save(flags);
1253 __count_vm_events(PGFREE, 1 << order);
1254 free_one_page(page_zone(page), page, pfn, order, migratetype);
1255 local_irq_restore(flags);
1256}
1257
1258static void __init __free_pages_boot_core(struct page *page, unsigned int order)
1259{
1260 unsigned int nr_pages = 1 << order;
1261 struct page *p = page;
1262 unsigned int loop;
1263
1264 prefetchw(p);
1265 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1266 prefetchw(p + 1);
1267 __ClearPageReserved(p);
1268 set_page_count(p, 0);
1269 }
1270 __ClearPageReserved(p);
1271 set_page_count(p, 0);
1272
1273 page_zone(page)->managed_pages += nr_pages;
1274 set_page_refcounted(page);
1275 __free_pages(page, order);
1276}
1277
1278#if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1279 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1280
1281static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1282
1283int __meminit early_pfn_to_nid(unsigned long pfn)
1284{
1285 static DEFINE_SPINLOCK(early_pfn_lock);
1286 int nid;
1287
1288 spin_lock(&early_pfn_lock);
1289 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1290 if (nid < 0)
1291 nid = first_online_node;
1292 spin_unlock(&early_pfn_lock);
1293
1294 return nid;
1295}
1296#endif
1297
1298#ifdef CONFIG_NODES_SPAN_OTHER_NODES
1299static inline bool __meminit __maybe_unused
1300meminit_pfn_in_nid(unsigned long pfn, int node,
1301 struct mminit_pfnnid_cache *state)
1302{
1303 int nid;
1304
1305 nid = __early_pfn_to_nid(pfn, state);
1306 if (nid >= 0 && nid != node)
1307 return false;
1308 return true;
1309}
1310
1311/* Only safe to use early in boot when initialisation is single-threaded */
1312static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1313{
1314 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1315}
1316
1317#else
1318
1319static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1320{
1321 return true;
1322}
1323static inline bool __meminit __maybe_unused
1324meminit_pfn_in_nid(unsigned long pfn, int node,
1325 struct mminit_pfnnid_cache *state)
1326{
1327 return true;
1328}
1329#endif
1330
1331
1332void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1333 unsigned int order)
1334{
1335 if (early_page_uninitialised(pfn))
1336 return;
1337 return __free_pages_boot_core(page, order);
1338}
1339
1340/*
1341 * Check that the whole (or subset of) a pageblock given by the interval of
1342 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1343 * with the migration of free compaction scanner. The scanners then need to
1344 * use only pfn_valid_within() check for arches that allow holes within
1345 * pageblocks.
1346 *
1347 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1348 *
1349 * It's possible on some configurations to have a setup like node0 node1 node0
1350 * i.e. it's possible that all pages within a zones range of pages do not
1351 * belong to a single zone. We assume that a border between node0 and node1
1352 * can occur within a single pageblock, but not a node0 node1 node0
1353 * interleaving within a single pageblock. It is therefore sufficient to check
1354 * the first and last page of a pageblock and avoid checking each individual
1355 * page in a pageblock.
1356 */
1357struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1358 unsigned long end_pfn, struct zone *zone)
1359{
1360 struct page *start_page;
1361 struct page *end_page;
1362
1363 /* end_pfn is one past the range we are checking */
1364 end_pfn--;
1365
1366 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1367 return NULL;
1368
1369 start_page = pfn_to_online_page(start_pfn);
1370 if (!start_page)
1371 return NULL;
1372
1373 if (page_zone(start_page) != zone)
1374 return NULL;
1375
1376 end_page = pfn_to_page(end_pfn);
1377
1378 /* This gives a shorter code than deriving page_zone(end_page) */
1379 if (page_zone_id(start_page) != page_zone_id(end_page))
1380 return NULL;
1381
1382 return start_page;
1383}
1384
1385void set_zone_contiguous(struct zone *zone)
1386{
1387 unsigned long block_start_pfn = zone->zone_start_pfn;
1388 unsigned long block_end_pfn;
1389
1390 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1391 for (; block_start_pfn < zone_end_pfn(zone);
1392 block_start_pfn = block_end_pfn,
1393 block_end_pfn += pageblock_nr_pages) {
1394
1395 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1396
1397 if (!__pageblock_pfn_to_page(block_start_pfn,
1398 block_end_pfn, zone))
1399 return;
1400 }
1401
1402 /* We confirm that there is no hole */
1403 zone->contiguous = true;
1404}
1405
1406void clear_zone_contiguous(struct zone *zone)
1407{
1408 zone->contiguous = false;
1409}
1410
1411#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1412static void __init deferred_free_range(unsigned long pfn,
1413 unsigned long nr_pages)
1414{
1415 struct page *page;
1416 unsigned long i;
1417
1418 if (!nr_pages)
1419 return;
1420
1421 page = pfn_to_page(pfn);
1422
1423 /* Free a large naturally-aligned chunk if possible */
1424 if (nr_pages == pageblock_nr_pages &&
1425 (pfn & (pageblock_nr_pages - 1)) == 0) {
1426 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1427 __free_pages_boot_core(page, pageblock_order);
1428 return;
1429 }
1430
1431 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1432 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1433 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1434 __free_pages_boot_core(page, 0);
1435 }
1436}
1437
1438/* Completion tracking for deferred_init_memmap() threads */
1439static atomic_t pgdat_init_n_undone __initdata;
1440static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1441
1442static inline void __init pgdat_init_report_one_done(void)
1443{
1444 if (atomic_dec_and_test(&pgdat_init_n_undone))
1445 complete(&pgdat_init_all_done_comp);
1446}
1447
1448/*
1449 * Returns true if page needs to be initialized or freed to buddy allocator.
1450 *
1451 * First we check if pfn is valid on architectures where it is possible to have
1452 * holes within pageblock_nr_pages. On systems where it is not possible, this
1453 * function is optimized out.
1454 *
1455 * Then, we check if a current large page is valid by only checking the validity
1456 * of the head pfn.
1457 *
1458 * Finally, meminit_pfn_in_nid is checked on systems where pfns can interleave
1459 * within a node: a pfn is between start and end of a node, but does not belong
1460 * to this memory node.
1461 */
1462static inline bool __init
1463deferred_pfn_valid(int nid, unsigned long pfn,
1464 struct mminit_pfnnid_cache *nid_init_state)
1465{
1466 if (!pfn_valid_within(pfn))
1467 return false;
1468 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1469 return false;
1470 if (!meminit_pfn_in_nid(pfn, nid, nid_init_state))
1471 return false;
1472 return true;
1473}
1474
1475/*
1476 * Free pages to buddy allocator. Try to free aligned pages in
1477 * pageblock_nr_pages sizes.
1478 */
1479static void __init deferred_free_pages(int nid, int zid, unsigned long pfn,
1480 unsigned long end_pfn)
1481{
1482 struct mminit_pfnnid_cache nid_init_state = { };
1483 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1484 unsigned long nr_free = 0;
1485
1486 for (; pfn < end_pfn; pfn++) {
1487 if (!deferred_pfn_valid(nid, pfn, &nid_init_state)) {
1488 deferred_free_range(pfn - nr_free, nr_free);
1489 nr_free = 0;
1490 } else if (!(pfn & nr_pgmask)) {
1491 deferred_free_range(pfn - nr_free, nr_free);
1492 nr_free = 1;
1493 touch_nmi_watchdog();
1494 } else {
1495 nr_free++;
1496 }
1497 }
1498 /* Free the last block of pages to allocator */
1499 deferred_free_range(pfn - nr_free, nr_free);
1500}
1501
1502/*
1503 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1504 * by performing it only once every pageblock_nr_pages.
1505 * Return number of pages initialized.
1506 */
1507static unsigned long __init deferred_init_pages(int nid, int zid,
1508 unsigned long pfn,
1509 unsigned long end_pfn)
1510{
1511 struct mminit_pfnnid_cache nid_init_state = { };
1512 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1513 unsigned long nr_pages = 0;
1514 struct page *page = NULL;
1515
1516 for (; pfn < end_pfn; pfn++) {
1517 if (!deferred_pfn_valid(nid, pfn, &nid_init_state)) {
1518 page = NULL;
1519 continue;
1520 } else if (!page || !(pfn & nr_pgmask)) {
1521 page = pfn_to_page(pfn);
1522 touch_nmi_watchdog();
1523 } else {
1524 page++;
1525 }
1526 __init_single_page(page, pfn, zid, nid);
1527 nr_pages++;
1528 }
1529 return (nr_pages);
1530}
1531
1532/* Initialise remaining memory on a node */
1533static int __init deferred_init_memmap(void *data)
1534{
1535 pg_data_t *pgdat = data;
1536 int nid = pgdat->node_id;
1537 unsigned long start = jiffies;
1538 unsigned long nr_pages = 0;
1539 unsigned long spfn, epfn, first_init_pfn, flags;
1540 phys_addr_t spa, epa;
1541 int zid;
1542 struct zone *zone;
1543 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1544 u64 i;
1545
1546 /* Bind memory initialisation thread to a local node if possible */
1547 if (!cpumask_empty(cpumask))
1548 set_cpus_allowed_ptr(current, cpumask);
1549
1550 pgdat_resize_lock(pgdat, &flags);
1551 first_init_pfn = pgdat->first_deferred_pfn;
1552 if (first_init_pfn == ULONG_MAX) {
1553 pgdat_resize_unlock(pgdat, &flags);
1554 pgdat_init_report_one_done();
1555 return 0;
1556 }
1557
1558 /* Sanity check boundaries */
1559 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1560 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1561 pgdat->first_deferred_pfn = ULONG_MAX;
1562
1563 /* Only the highest zone is deferred so find it */
1564 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1565 zone = pgdat->node_zones + zid;
1566 if (first_init_pfn < zone_end_pfn(zone))
1567 break;
1568 }
1569 first_init_pfn = max(zone->zone_start_pfn, first_init_pfn);
1570
1571 /*
1572 * Initialize and free pages. We do it in two loops: first we initialize
1573 * struct page, than free to buddy allocator, because while we are
1574 * freeing pages we can access pages that are ahead (computing buddy
1575 * page in __free_one_page()).
1576 */
1577 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1578 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1579 epfn = min_t(unsigned long, zone_end_pfn(zone), PFN_DOWN(epa));
1580 nr_pages += deferred_init_pages(nid, zid, spfn, epfn);
1581 }
1582 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1583 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1584 epfn = min_t(unsigned long, zone_end_pfn(zone), PFN_DOWN(epa));
1585 deferred_free_pages(nid, zid, spfn, epfn);
1586 }
1587 pgdat_resize_unlock(pgdat, &flags);
1588
1589 /* Sanity check that the next zone really is unpopulated */
1590 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1591
1592 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1593 jiffies_to_msecs(jiffies - start));
1594
1595 pgdat_init_report_one_done();
1596 return 0;
1597}
1598
1599/*
1600 * During boot we initialize deferred pages on-demand, as needed, but once
1601 * page_alloc_init_late() has finished, the deferred pages are all initialized,
1602 * and we can permanently disable that path.
1603 */
1604static DEFINE_STATIC_KEY_TRUE(deferred_pages);
1605
1606/*
1607 * If this zone has deferred pages, try to grow it by initializing enough
1608 * deferred pages to satisfy the allocation specified by order, rounded up to
1609 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
1610 * of SECTION_SIZE bytes by initializing struct pages in increments of
1611 * PAGES_PER_SECTION * sizeof(struct page) bytes.
1612 *
1613 * Return true when zone was grown, otherwise return false. We return true even
1614 * when we grow less than requested, to let the caller decide if there are
1615 * enough pages to satisfy the allocation.
1616 *
1617 * Note: We use noinline because this function is needed only during boot, and
1618 * it is called from a __ref function _deferred_grow_zone. This way we are
1619 * making sure that it is not inlined into permanent text section.
1620 */
1621static noinline bool __init
1622deferred_grow_zone(struct zone *zone, unsigned int order)
1623{
1624 int zid = zone_idx(zone);
1625 int nid = zone_to_nid(zone);
1626 pg_data_t *pgdat = NODE_DATA(nid);
1627 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
1628 unsigned long nr_pages = 0;
1629 unsigned long first_init_pfn, spfn, epfn, t, flags;
1630 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
1631 phys_addr_t spa, epa;
1632 u64 i;
1633
1634 /* Only the last zone may have deferred pages */
1635 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
1636 return false;
1637
1638 pgdat_resize_lock(pgdat, &flags);
1639
1640 /*
1641 * If deferred pages have been initialized while we were waiting for
1642 * the lock, return true, as the zone was grown. The caller will retry
1643 * this zone. We won't return to this function since the caller also
1644 * has this static branch.
1645 */
1646 if (!static_branch_unlikely(&deferred_pages)) {
1647 pgdat_resize_unlock(pgdat, &flags);
1648 return true;
1649 }
1650
1651 /*
1652 * If someone grew this zone while we were waiting for spinlock, return
1653 * true, as there might be enough pages already.
1654 */
1655 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
1656 pgdat_resize_unlock(pgdat, &flags);
1657 return true;
1658 }
1659
1660 first_init_pfn = max(zone->zone_start_pfn, first_deferred_pfn);
1661
1662 if (first_init_pfn >= pgdat_end_pfn(pgdat)) {
1663 pgdat_resize_unlock(pgdat, &flags);
1664 return false;
1665 }
1666
1667 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1668 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1669 epfn = min_t(unsigned long, zone_end_pfn(zone), PFN_DOWN(epa));
1670
1671 while (spfn < epfn && nr_pages < nr_pages_needed) {
1672 t = ALIGN(spfn + PAGES_PER_SECTION, PAGES_PER_SECTION);
1673 first_deferred_pfn = min(t, epfn);
1674 nr_pages += deferred_init_pages(nid, zid, spfn,
1675 first_deferred_pfn);
1676 spfn = first_deferred_pfn;
1677 }
1678
1679 if (nr_pages >= nr_pages_needed)
1680 break;
1681 }
1682
1683 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1684 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1685 epfn = min_t(unsigned long, first_deferred_pfn, PFN_DOWN(epa));
1686 deferred_free_pages(nid, zid, spfn, epfn);
1687
1688 if (first_deferred_pfn == epfn)
1689 break;
1690 }
1691 pgdat->first_deferred_pfn = first_deferred_pfn;
1692 pgdat_resize_unlock(pgdat, &flags);
1693
1694 return nr_pages > 0;
1695}
1696
1697/*
1698 * deferred_grow_zone() is __init, but it is called from
1699 * get_page_from_freelist() during early boot until deferred_pages permanently
1700 * disables this call. This is why we have refdata wrapper to avoid warning,
1701 * and to ensure that the function body gets unloaded.
1702 */
1703static bool __ref
1704_deferred_grow_zone(struct zone *zone, unsigned int order)
1705{
1706 return deferred_grow_zone(zone, order);
1707}
1708
1709#endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1710
1711void __init page_alloc_init_late(void)
1712{
1713 struct zone *zone;
1714
1715#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1716 int nid;
1717
1718 /* There will be num_node_state(N_MEMORY) threads */
1719 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1720 for_each_node_state(nid, N_MEMORY) {
1721 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1722 }
1723
1724 /* Block until all are initialised */
1725 wait_for_completion(&pgdat_init_all_done_comp);
1726
1727 /*
1728 * We initialized the rest of the deferred pages. Permanently disable
1729 * on-demand struct page initialization.
1730 */
1731 static_branch_disable(&deferred_pages);
1732
1733 /* Reinit limits that are based on free pages after the kernel is up */
1734 files_maxfiles_init();
1735#endif
1736#ifdef CONFIG_ARCH_DISCARD_MEMBLOCK
1737 /* Discard memblock private memory */
1738 memblock_discard();
1739#endif
1740
1741 for_each_populated_zone(zone)
1742 set_zone_contiguous(zone);
1743}
1744
1745#ifdef CONFIG_CMA
1746/* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1747void __init init_cma_reserved_pageblock(struct page *page)
1748{
1749 unsigned i = pageblock_nr_pages;
1750 struct page *p = page;
1751
1752 do {
1753 __ClearPageReserved(p);
1754 set_page_count(p, 0);
1755 } while (++p, --i);
1756
1757 set_pageblock_migratetype(page, MIGRATE_CMA);
1758
1759 if (pageblock_order >= MAX_ORDER) {
1760 i = pageblock_nr_pages;
1761 p = page;
1762 do {
1763 set_page_refcounted(p);
1764 __free_pages(p, MAX_ORDER - 1);
1765 p += MAX_ORDER_NR_PAGES;
1766 } while (i -= MAX_ORDER_NR_PAGES);
1767 } else {
1768 set_page_refcounted(page);
1769 __free_pages(page, pageblock_order);
1770 }
1771
1772 adjust_managed_page_count(page, pageblock_nr_pages);
1773}
1774#endif
1775
1776/*
1777 * The order of subdivision here is critical for the IO subsystem.
1778 * Please do not alter this order without good reasons and regression
1779 * testing. Specifically, as large blocks of memory are subdivided,
1780 * the order in which smaller blocks are delivered depends on the order
1781 * they're subdivided in this function. This is the primary factor
1782 * influencing the order in which pages are delivered to the IO
1783 * subsystem according to empirical testing, and this is also justified
1784 * by considering the behavior of a buddy system containing a single
1785 * large block of memory acted on by a series of small allocations.
1786 * This behavior is a critical factor in sglist merging's success.
1787 *
1788 * -- nyc
1789 */
1790static inline void expand(struct zone *zone, struct page *page,
1791 int low, int high, struct free_area *area,
1792 int migratetype)
1793{
1794 unsigned long size = 1 << high;
1795
1796 while (high > low) {
1797 area--;
1798 high--;
1799 size >>= 1;
1800 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1801
1802 /*
1803 * Mark as guard pages (or page), that will allow to
1804 * merge back to allocator when buddy will be freed.
1805 * Corresponding page table entries will not be touched,
1806 * pages will stay not present in virtual address space
1807 */
1808 if (set_page_guard(zone, &page[size], high, migratetype))
1809 continue;
1810
1811 list_add(&page[size].lru, &area->free_list[migratetype]);
1812 area->nr_free++;
1813 set_page_order(&page[size], high);
1814 }
1815}
1816
1817static void check_new_page_bad(struct page *page)
1818{
1819 const char *bad_reason = NULL;
1820 unsigned long bad_flags = 0;
1821
1822 if (unlikely(atomic_read(&page->_mapcount) != -1))
1823 bad_reason = "nonzero mapcount";
1824 if (unlikely(page->mapping != NULL))
1825 bad_reason = "non-NULL mapping";
1826 if (unlikely(page_ref_count(page) != 0))
1827 bad_reason = "nonzero _count";
1828 if (unlikely(page->flags & __PG_HWPOISON)) {
1829 bad_reason = "HWPoisoned (hardware-corrupted)";
1830 bad_flags = __PG_HWPOISON;
1831 /* Don't complain about hwpoisoned pages */
1832 page_mapcount_reset(page); /* remove PageBuddy */
1833 return;
1834 }
1835 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1836 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1837 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1838 }
1839#ifdef CONFIG_MEMCG
1840 if (unlikely(page->mem_cgroup))
1841 bad_reason = "page still charged to cgroup";
1842#endif
1843 bad_page(page, bad_reason, bad_flags);
1844}
1845
1846/*
1847 * This page is about to be returned from the page allocator
1848 */
1849static inline int check_new_page(struct page *page)
1850{
1851 if (likely(page_expected_state(page,
1852 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
1853 return 0;
1854
1855 check_new_page_bad(page);
1856 return 1;
1857}
1858
1859static inline bool free_pages_prezeroed(void)
1860{
1861 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
1862 page_poisoning_enabled();
1863}
1864
1865#ifdef CONFIG_DEBUG_VM
1866static bool check_pcp_refill(struct page *page)
1867{
1868 return false;
1869}
1870
1871static bool check_new_pcp(struct page *page)
1872{
1873 return check_new_page(page);
1874}
1875#else
1876static bool check_pcp_refill(struct page *page)
1877{
1878 return check_new_page(page);
1879}
1880static bool check_new_pcp(struct page *page)
1881{
1882 return false;
1883}
1884#endif /* CONFIG_DEBUG_VM */
1885
1886static bool check_new_pages(struct page *page, unsigned int order)
1887{
1888 int i;
1889 for (i = 0; i < (1 << order); i++) {
1890 struct page *p = page + i;
1891
1892 if (unlikely(check_new_page(p)))
1893 return true;
1894 }
1895
1896 return false;
1897}
1898
1899inline void post_alloc_hook(struct page *page, unsigned int order,
1900 gfp_t gfp_flags)
1901{
1902 set_page_private(page, 0);
1903 set_page_refcounted(page);
1904
1905 arch_alloc_page(page, order);
1906 kernel_map_pages(page, 1 << order, 1);
1907 kernel_poison_pages(page, 1 << order, 1);
1908 kasan_alloc_pages(page, order);
1909 set_page_owner(page, order, gfp_flags);
1910}
1911
1912static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1913 unsigned int alloc_flags)
1914{
1915 int i;
1916
1917 post_alloc_hook(page, order, gfp_flags);
1918
1919 if (!free_pages_prezeroed() && (gfp_flags & __GFP_ZERO))
1920 for (i = 0; i < (1 << order); i++)
1921 clear_highpage(page + i);
1922
1923 if (order && (gfp_flags & __GFP_COMP))
1924 prep_compound_page(page, order);
1925
1926 /*
1927 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1928 * allocate the page. The expectation is that the caller is taking
1929 * steps that will free more memory. The caller should avoid the page
1930 * being used for !PFMEMALLOC purposes.
1931 */
1932 if (alloc_flags & ALLOC_NO_WATERMARKS)
1933 set_page_pfmemalloc(page);
1934 else
1935 clear_page_pfmemalloc(page);
1936}
1937
1938/*
1939 * Go through the free lists for the given migratetype and remove
1940 * the smallest available page from the freelists
1941 */
1942static __always_inline
1943struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1944 int migratetype)
1945{
1946 unsigned int current_order;
1947 struct free_area *area;
1948 struct page *page;
1949
1950 /* Find a page of the appropriate size in the preferred list */
1951 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1952 area = &(zone->free_area[current_order]);
1953 page = list_first_entry_or_null(&area->free_list[migratetype],
1954 struct page, lru);
1955 if (!page)
1956 continue;
1957 list_del(&page->lru);
1958 rmv_page_order(page);
1959 area->nr_free--;
1960 expand(zone, page, order, current_order, area, migratetype);
1961 set_pcppage_migratetype(page, migratetype);
1962 return page;
1963 }
1964
1965 return NULL;
1966}
1967
1968
1969/*
1970 * This array describes the order lists are fallen back to when
1971 * the free lists for the desirable migrate type are depleted
1972 */
1973static int fallbacks[MIGRATE_TYPES][4] = {
1974 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1975 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1976 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1977#ifdef CONFIG_CMA
1978 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1979#endif
1980#ifdef CONFIG_MEMORY_ISOLATION
1981 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1982#endif
1983};
1984
1985#ifdef CONFIG_CMA
1986static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1987 unsigned int order)
1988{
1989 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1990}
1991#else
1992static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1993 unsigned int order) { return NULL; }
1994#endif
1995
1996/*
1997 * Move the free pages in a range to the free lists of the requested type.
1998 * Note that start_page and end_pages are not aligned on a pageblock
1999 * boundary. If alignment is required, use move_freepages_block()
2000 */
2001static int move_freepages(struct zone *zone,
2002 struct page *start_page, struct page *end_page,
2003 int migratetype, int *num_movable)
2004{
2005 struct page *page;
2006 unsigned int order;
2007 int pages_moved = 0;
2008
2009#ifndef CONFIG_HOLES_IN_ZONE
2010 /*
2011 * page_zone is not safe to call in this context when
2012 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
2013 * anyway as we check zone boundaries in move_freepages_block().
2014 * Remove at a later date when no bug reports exist related to
2015 * grouping pages by mobility
2016 */
2017 VM_BUG_ON(pfn_valid(page_to_pfn(start_page)) &&
2018 pfn_valid(page_to_pfn(end_page)) &&
2019 page_zone(start_page) != page_zone(end_page));
2020#endif
2021
2022 if (num_movable)
2023 *num_movable = 0;
2024
2025 for (page = start_page; page <= end_page;) {
2026 if (!pfn_valid_within(page_to_pfn(page))) {
2027 page++;
2028 continue;
2029 }
2030
2031 /* Make sure we are not inadvertently changing nodes */
2032 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2033
2034 if (!PageBuddy(page)) {
2035 /*
2036 * We assume that pages that could be isolated for
2037 * migration are movable. But we don't actually try
2038 * isolating, as that would be expensive.
2039 */
2040 if (num_movable &&
2041 (PageLRU(page) || __PageMovable(page)))
2042 (*num_movable)++;
2043
2044 page++;
2045 continue;
2046 }
2047
2048 order = page_order(page);
2049 list_move(&page->lru,
2050 &zone->free_area[order].free_list[migratetype]);
2051 page += 1 << order;
2052 pages_moved += 1 << order;
2053 }
2054
2055 return pages_moved;
2056}
2057
2058int move_freepages_block(struct zone *zone, struct page *page,
2059 int migratetype, int *num_movable)
2060{
2061 unsigned long start_pfn, end_pfn;
2062 struct page *start_page, *end_page;
2063
2064 start_pfn = page_to_pfn(page);
2065 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
2066 start_page = pfn_to_page(start_pfn);
2067 end_page = start_page + pageblock_nr_pages - 1;
2068 end_pfn = start_pfn + pageblock_nr_pages - 1;
2069
2070 /* Do not cross zone boundaries */
2071 if (!zone_spans_pfn(zone, start_pfn))
2072 start_page = page;
2073 if (!zone_spans_pfn(zone, end_pfn))
2074 return 0;
2075
2076 return move_freepages(zone, start_page, end_page, migratetype,
2077 num_movable);
2078}
2079
2080static void change_pageblock_range(struct page *pageblock_page,
2081 int start_order, int migratetype)
2082{
2083 int nr_pageblocks = 1 << (start_order - pageblock_order);
2084
2085 while (nr_pageblocks--) {
2086 set_pageblock_migratetype(pageblock_page, migratetype);
2087 pageblock_page += pageblock_nr_pages;
2088 }
2089}
2090
2091/*
2092 * When we are falling back to another migratetype during allocation, try to
2093 * steal extra free pages from the same pageblocks to satisfy further
2094 * allocations, instead of polluting multiple pageblocks.
2095 *
2096 * If we are stealing a relatively large buddy page, it is likely there will
2097 * be more free pages in the pageblock, so try to steal them all. For
2098 * reclaimable and unmovable allocations, we steal regardless of page size,
2099 * as fragmentation caused by those allocations polluting movable pageblocks
2100 * is worse than movable allocations stealing from unmovable and reclaimable
2101 * pageblocks.
2102 */
2103static bool can_steal_fallback(unsigned int order, int start_mt)
2104{
2105 /*
2106 * Leaving this order check is intended, although there is
2107 * relaxed order check in next check. The reason is that
2108 * we can actually steal whole pageblock if this condition met,
2109 * but, below check doesn't guarantee it and that is just heuristic
2110 * so could be changed anytime.
2111 */
2112 if (order >= pageblock_order)
2113 return true;
2114
2115 if (order >= pageblock_order / 2 ||
2116 start_mt == MIGRATE_RECLAIMABLE ||
2117 start_mt == MIGRATE_UNMOVABLE ||
2118 page_group_by_mobility_disabled)
2119 return true;
2120
2121 return false;
2122}
2123
2124/*
2125 * This function implements actual steal behaviour. If order is large enough,
2126 * we can steal whole pageblock. If not, we first move freepages in this
2127 * pageblock to our migratetype and determine how many already-allocated pages
2128 * are there in the pageblock with a compatible migratetype. If at least half
2129 * of pages are free or compatible, we can change migratetype of the pageblock
2130 * itself, so pages freed in the future will be put on the correct free list.
2131 */
2132static void steal_suitable_fallback(struct zone *zone, struct page *page,
2133 int start_type, bool whole_block)
2134{
2135 unsigned int current_order = page_order(page);
2136 struct free_area *area;
2137 int free_pages, movable_pages, alike_pages;
2138 int old_block_type;
2139
2140 old_block_type = get_pageblock_migratetype(page);
2141
2142 /*
2143 * This can happen due to races and we want to prevent broken
2144 * highatomic accounting.
2145 */
2146 if (is_migrate_highatomic(old_block_type))
2147 goto single_page;
2148
2149 /* Take ownership for orders >= pageblock_order */
2150 if (current_order >= pageblock_order) {
2151 change_pageblock_range(page, current_order, start_type);
2152 goto single_page;
2153 }
2154
2155 /* We are not allowed to try stealing from the whole block */
2156 if (!whole_block)
2157 goto single_page;
2158
2159 free_pages = move_freepages_block(zone, page, start_type,
2160 &movable_pages);
2161 /*
2162 * Determine how many pages are compatible with our allocation.
2163 * For movable allocation, it's the number of movable pages which
2164 * we just obtained. For other types it's a bit more tricky.
2165 */
2166 if (start_type == MIGRATE_MOVABLE) {
2167 alike_pages = movable_pages;
2168 } else {
2169 /*
2170 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2171 * to MOVABLE pageblock, consider all non-movable pages as
2172 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2173 * vice versa, be conservative since we can't distinguish the
2174 * exact migratetype of non-movable pages.
2175 */
2176 if (old_block_type == MIGRATE_MOVABLE)
2177 alike_pages = pageblock_nr_pages
2178 - (free_pages + movable_pages);
2179 else
2180 alike_pages = 0;
2181 }
2182
2183 /* moving whole block can fail due to zone boundary conditions */
2184 if (!free_pages)
2185 goto single_page;
2186
2187 /*
2188 * If a sufficient number of pages in the block are either free or of
2189 * comparable migratability as our allocation, claim the whole block.
2190 */
2191 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2192 page_group_by_mobility_disabled)
2193 set_pageblock_migratetype(page, start_type);
2194
2195 return;
2196
2197single_page:
2198 area = &zone->free_area[current_order];
2199 list_move(&page->lru, &area->free_list[start_type]);
2200}
2201
2202/*
2203 * Check whether there is a suitable fallback freepage with requested order.
2204 * If only_stealable is true, this function returns fallback_mt only if
2205 * we can steal other freepages all together. This would help to reduce
2206 * fragmentation due to mixed migratetype pages in one pageblock.
2207 */
2208int find_suitable_fallback(struct free_area *area, unsigned int order,
2209 int migratetype, bool only_stealable, bool *can_steal)
2210{
2211 int i;
2212 int fallback_mt;
2213
2214 if (area->nr_free == 0)
2215 return -1;
2216
2217 *can_steal = false;
2218 for (i = 0;; i++) {
2219 fallback_mt = fallbacks[migratetype][i];
2220 if (fallback_mt == MIGRATE_TYPES)
2221 break;
2222
2223 if (list_empty(&area->free_list[fallback_mt]))
2224 continue;
2225
2226 if (can_steal_fallback(order, migratetype))
2227 *can_steal = true;
2228
2229 if (!only_stealable)
2230 return fallback_mt;
2231
2232 if (*can_steal)
2233 return fallback_mt;
2234 }
2235
2236 return -1;
2237}
2238
2239/*
2240 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2241 * there are no empty page blocks that contain a page with a suitable order
2242 */
2243static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2244 unsigned int alloc_order)
2245{
2246 int mt;
2247 unsigned long max_managed, flags;
2248
2249 /*
2250 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2251 * Check is race-prone but harmless.
2252 */
2253 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
2254 if (zone->nr_reserved_highatomic >= max_managed)
2255 return;
2256
2257 spin_lock_irqsave(&zone->lock, flags);
2258
2259 /* Recheck the nr_reserved_highatomic limit under the lock */
2260 if (zone->nr_reserved_highatomic >= max_managed)
2261 goto out_unlock;
2262
2263 /* Yoink! */
2264 mt = get_pageblock_migratetype(page);
2265 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2266 && !is_migrate_cma(mt)) {
2267 zone->nr_reserved_highatomic += pageblock_nr_pages;
2268 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2269 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2270 }
2271
2272out_unlock:
2273 spin_unlock_irqrestore(&zone->lock, flags);
2274}
2275
2276/*
2277 * Used when an allocation is about to fail under memory pressure. This
2278 * potentially hurts the reliability of high-order allocations when under
2279 * intense memory pressure but failed atomic allocations should be easier
2280 * to recover from than an OOM.
2281 *
2282 * If @force is true, try to unreserve a pageblock even though highatomic
2283 * pageblock is exhausted.
2284 */
2285static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2286 bool force)
2287{
2288 struct zonelist *zonelist = ac->zonelist;
2289 unsigned long flags;
2290 struct zoneref *z;
2291 struct zone *zone;
2292 struct page *page;
2293 int order;
2294 bool ret;
2295
2296 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2297 ac->nodemask) {
2298 /*
2299 * Preserve at least one pageblock unless memory pressure
2300 * is really high.
2301 */
2302 if (!force && zone->nr_reserved_highatomic <=
2303 pageblock_nr_pages)
2304 continue;
2305
2306 spin_lock_irqsave(&zone->lock, flags);
2307 for (order = 0; order < MAX_ORDER; order++) {
2308 struct free_area *area = &(zone->free_area[order]);
2309
2310 page = list_first_entry_or_null(
2311 &area->free_list[MIGRATE_HIGHATOMIC],
2312 struct page, lru);
2313 if (!page)
2314 continue;
2315
2316 /*
2317 * In page freeing path, migratetype change is racy so
2318 * we can counter several free pages in a pageblock
2319 * in this loop althoug we changed the pageblock type
2320 * from highatomic to ac->migratetype. So we should
2321 * adjust the count once.
2322 */
2323 if (is_migrate_highatomic_page(page)) {
2324 /*
2325 * It should never happen but changes to
2326 * locking could inadvertently allow a per-cpu
2327 * drain to add pages to MIGRATE_HIGHATOMIC
2328 * while unreserving so be safe and watch for
2329 * underflows.
2330 */
2331 zone->nr_reserved_highatomic -= min(
2332 pageblock_nr_pages,
2333 zone->nr_reserved_highatomic);
2334 }
2335
2336 /*
2337 * Convert to ac->migratetype and avoid the normal
2338 * pageblock stealing heuristics. Minimally, the caller
2339 * is doing the work and needs the pages. More
2340 * importantly, if the block was always converted to
2341 * MIGRATE_UNMOVABLE or another type then the number
2342 * of pageblocks that cannot be completely freed
2343 * may increase.
2344 */
2345 set_pageblock_migratetype(page, ac->migratetype);
2346 ret = move_freepages_block(zone, page, ac->migratetype,
2347 NULL);
2348 if (ret) {
2349 spin_unlock_irqrestore(&zone->lock, flags);
2350 return ret;
2351 }
2352 }
2353 spin_unlock_irqrestore(&zone->lock, flags);
2354 }
2355
2356 return false;
2357}
2358
2359/*
2360 * Try finding a free buddy page on the fallback list and put it on the free
2361 * list of requested migratetype, possibly along with other pages from the same
2362 * block, depending on fragmentation avoidance heuristics. Returns true if
2363 * fallback was found so that __rmqueue_smallest() can grab it.
2364 *
2365 * The use of signed ints for order and current_order is a deliberate
2366 * deviation from the rest of this file, to make the for loop
2367 * condition simpler.
2368 */
2369static __always_inline bool
2370__rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
2371{
2372 struct free_area *area;
2373 int current_order;
2374 struct page *page;
2375 int fallback_mt;
2376 bool can_steal;
2377
2378 /*
2379 * Find the largest available free page in the other list. This roughly
2380 * approximates finding the pageblock with the most free pages, which
2381 * would be too costly to do exactly.
2382 */
2383 for (current_order = MAX_ORDER - 1; current_order >= order;
2384 --current_order) {
2385 area = &(zone->free_area[current_order]);
2386 fallback_mt = find_suitable_fallback(area, current_order,
2387 start_migratetype, false, &can_steal);
2388 if (fallback_mt == -1)
2389 continue;
2390
2391 /*
2392 * We cannot steal all free pages from the pageblock and the
2393 * requested migratetype is movable. In that case it's better to
2394 * steal and split the smallest available page instead of the
2395 * largest available page, because even if the next movable
2396 * allocation falls back into a different pageblock than this
2397 * one, it won't cause permanent fragmentation.
2398 */
2399 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2400 && current_order > order)
2401 goto find_smallest;
2402
2403 goto do_steal;
2404 }
2405
2406 return false;
2407
2408find_smallest:
2409 for (current_order = order; current_order < MAX_ORDER;
2410 current_order++) {
2411 area = &(zone->free_area[current_order]);
2412 fallback_mt = find_suitable_fallback(area, current_order,
2413 start_migratetype, false, &can_steal);
2414 if (fallback_mt != -1)
2415 break;
2416 }
2417
2418 /*
2419 * This should not happen - we already found a suitable fallback
2420 * when looking for the largest page.
2421 */
2422 VM_BUG_ON(current_order == MAX_ORDER);
2423
2424do_steal:
2425 page = list_first_entry(&area->free_list[fallback_mt],
2426 struct page, lru);
2427
2428 steal_suitable_fallback(zone, page, start_migratetype, can_steal);
2429
2430 trace_mm_page_alloc_extfrag(page, order, current_order,
2431 start_migratetype, fallback_mt);
2432
2433 return true;
2434
2435}
2436
2437/*
2438 * Do the hard work of removing an element from the buddy allocator.
2439 * Call me with the zone->lock already held.
2440 */
2441static __always_inline struct page *
2442__rmqueue(struct zone *zone, unsigned int order, int migratetype)
2443{
2444 struct page *page;
2445
2446retry:
2447 page = __rmqueue_smallest(zone, order, migratetype);
2448 if (unlikely(!page)) {
2449 if (migratetype == MIGRATE_MOVABLE)
2450 page = __rmqueue_cma_fallback(zone, order);
2451
2452 if (!page && __rmqueue_fallback(zone, order, migratetype))
2453 goto retry;
2454 }
2455
2456 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2457 return page;
2458}
2459
2460/*
2461 * Obtain a specified number of elements from the buddy allocator, all under
2462 * a single hold of the lock, for efficiency. Add them to the supplied list.
2463 * Returns the number of new pages which were placed at *list.
2464 */
2465static int rmqueue_bulk(struct zone *zone, unsigned int order,
2466 unsigned long count, struct list_head *list,
2467 int migratetype)
2468{
2469 int i, alloced = 0;
2470
2471 spin_lock(&zone->lock);
2472 for (i = 0; i < count; ++i) {
2473 struct page *page = __rmqueue(zone, order, migratetype);
2474 if (unlikely(page == NULL))
2475 break;
2476
2477 if (unlikely(check_pcp_refill(page)))
2478 continue;
2479
2480 /*
2481 * Split buddy pages returned by expand() are received here in
2482 * physical page order. The page is added to the tail of
2483 * caller's list. From the callers perspective, the linked list
2484 * is ordered by page number under some conditions. This is
2485 * useful for IO devices that can forward direction from the
2486 * head, thus also in the physical page order. This is useful
2487 * for IO devices that can merge IO requests if the physical
2488 * pages are ordered properly.
2489 */
2490 list_add_tail(&page->lru, list);
2491 alloced++;
2492 if (is_migrate_cma(get_pcppage_migratetype(page)))
2493 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2494 -(1 << order));
2495 }
2496
2497 /*
2498 * i pages were removed from the buddy list even if some leak due
2499 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2500 * on i. Do not confuse with 'alloced' which is the number of
2501 * pages added to the pcp list.
2502 */
2503 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2504 spin_unlock(&zone->lock);
2505 return alloced;
2506}
2507
2508#ifdef CONFIG_NUMA
2509/*
2510 * Called from the vmstat counter updater to drain pagesets of this
2511 * currently executing processor on remote nodes after they have
2512 * expired.
2513 *
2514 * Note that this function must be called with the thread pinned to
2515 * a single processor.
2516 */
2517void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2518{
2519 unsigned long flags;
2520 int to_drain, batch;
2521
2522 local_irq_save(flags);
2523 batch = READ_ONCE(pcp->batch);
2524 to_drain = min(pcp->count, batch);
2525 if (to_drain > 0)
2526 free_pcppages_bulk(zone, to_drain, pcp);
2527 local_irq_restore(flags);
2528}
2529#endif
2530
2531/*
2532 * Drain pcplists of the indicated processor and zone.
2533 *
2534 * The processor must either be the current processor and the
2535 * thread pinned to the current processor or a processor that
2536 * is not online.
2537 */
2538static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2539{
2540 unsigned long flags;
2541 struct per_cpu_pageset *pset;
2542 struct per_cpu_pages *pcp;
2543
2544 local_irq_save(flags);
2545 pset = per_cpu_ptr(zone->pageset, cpu);
2546
2547 pcp = &pset->pcp;
2548 if (pcp->count)
2549 free_pcppages_bulk(zone, pcp->count, pcp);
2550 local_irq_restore(flags);
2551}
2552
2553/*
2554 * Drain pcplists of all zones on the indicated processor.
2555 *
2556 * The processor must either be the current processor and the
2557 * thread pinned to the current processor or a processor that
2558 * is not online.
2559 */
2560static void drain_pages(unsigned int cpu)
2561{
2562 struct zone *zone;
2563
2564 for_each_populated_zone(zone) {
2565 drain_pages_zone(cpu, zone);
2566 }
2567}
2568
2569/*
2570 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2571 *
2572 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2573 * the single zone's pages.
2574 */
2575void drain_local_pages(struct zone *zone)
2576{
2577 int cpu = smp_processor_id();
2578
2579 if (zone)
2580 drain_pages_zone(cpu, zone);
2581 else
2582 drain_pages(cpu);
2583}
2584
2585static void drain_local_pages_wq(struct work_struct *work)
2586{
2587 /*
2588 * drain_all_pages doesn't use proper cpu hotplug protection so
2589 * we can race with cpu offline when the WQ can move this from
2590 * a cpu pinned worker to an unbound one. We can operate on a different
2591 * cpu which is allright but we also have to make sure to not move to
2592 * a different one.
2593 */
2594 preempt_disable();
2595 drain_local_pages(NULL);
2596 preempt_enable();
2597}
2598
2599/*
2600 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2601 *
2602 * When zone parameter is non-NULL, spill just the single zone's pages.
2603 *
2604 * Note that this can be extremely slow as the draining happens in a workqueue.
2605 */
2606void drain_all_pages(struct zone *zone)
2607{
2608 int cpu;
2609
2610 /*
2611 * Allocate in the BSS so we wont require allocation in
2612 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2613 */
2614 static cpumask_t cpus_with_pcps;
2615
2616 /*
2617 * Make sure nobody triggers this path before mm_percpu_wq is fully
2618 * initialized.
2619 */
2620 if (WARN_ON_ONCE(!mm_percpu_wq))
2621 return;
2622
2623 /*
2624 * Do not drain if one is already in progress unless it's specific to
2625 * a zone. Such callers are primarily CMA and memory hotplug and need
2626 * the drain to be complete when the call returns.
2627 */
2628 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2629 if (!zone)
2630 return;
2631 mutex_lock(&pcpu_drain_mutex);
2632 }
2633
2634 /*
2635 * We don't care about racing with CPU hotplug event
2636 * as offline notification will cause the notified
2637 * cpu to drain that CPU pcps and on_each_cpu_mask
2638 * disables preemption as part of its processing
2639 */
2640 for_each_online_cpu(cpu) {
2641 struct per_cpu_pageset *pcp;
2642 struct zone *z;
2643 bool has_pcps = false;
2644
2645 if (zone) {
2646 pcp = per_cpu_ptr(zone->pageset, cpu);
2647 if (pcp->pcp.count)
2648 has_pcps = true;
2649 } else {
2650 for_each_populated_zone(z) {
2651 pcp = per_cpu_ptr(z->pageset, cpu);
2652 if (pcp->pcp.count) {
2653 has_pcps = true;
2654 break;
2655 }
2656 }
2657 }
2658
2659 if (has_pcps)
2660 cpumask_set_cpu(cpu, &cpus_with_pcps);
2661 else
2662 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2663 }
2664
2665 for_each_cpu(cpu, &cpus_with_pcps) {
2666 struct work_struct *work = per_cpu_ptr(&pcpu_drain, cpu);
2667 INIT_WORK(work, drain_local_pages_wq);
2668 queue_work_on(cpu, mm_percpu_wq, work);
2669 }
2670 for_each_cpu(cpu, &cpus_with_pcps)
2671 flush_work(per_cpu_ptr(&pcpu_drain, cpu));
2672
2673 mutex_unlock(&pcpu_drain_mutex);
2674}
2675
2676#ifdef CONFIG_HIBERNATION
2677
2678/*
2679 * Touch the watchdog for every WD_PAGE_COUNT pages.
2680 */
2681#define WD_PAGE_COUNT (128*1024)
2682
2683void mark_free_pages(struct zone *zone)
2684{
2685 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
2686 unsigned long flags;
2687 unsigned int order, t;
2688 struct page *page;
2689
2690 if (zone_is_empty(zone))
2691 return;
2692
2693 spin_lock_irqsave(&zone->lock, flags);
2694
2695 max_zone_pfn = zone_end_pfn(zone);
2696 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2697 if (pfn_valid(pfn)) {
2698 page = pfn_to_page(pfn);
2699
2700 if (!--page_count) {
2701 touch_nmi_watchdog();
2702 page_count = WD_PAGE_COUNT;
2703 }
2704
2705 if (page_zone(page) != zone)
2706 continue;
2707
2708 if (!swsusp_page_is_forbidden(page))
2709 swsusp_unset_page_free(page);
2710 }
2711
2712 for_each_migratetype_order(order, t) {
2713 list_for_each_entry(page,
2714 &zone->free_area[order].free_list[t], lru) {
2715 unsigned long i;
2716
2717 pfn = page_to_pfn(page);
2718 for (i = 0; i < (1UL << order); i++) {
2719 if (!--page_count) {
2720 touch_nmi_watchdog();
2721 page_count = WD_PAGE_COUNT;
2722 }
2723 swsusp_set_page_free(pfn_to_page(pfn + i));
2724 }
2725 }
2726 }
2727 spin_unlock_irqrestore(&zone->lock, flags);
2728}
2729#endif /* CONFIG_PM */
2730
2731static bool free_unref_page_prepare(struct page *page, unsigned long pfn)
2732{
2733 int migratetype;
2734
2735 if (!free_pcp_prepare(page))
2736 return false;
2737
2738 migratetype = get_pfnblock_migratetype(page, pfn);
2739 set_pcppage_migratetype(page, migratetype);
2740 return true;
2741}
2742
2743static void free_unref_page_commit(struct page *page, unsigned long pfn)
2744{
2745 struct zone *zone = page_zone(page);
2746 struct per_cpu_pages *pcp;
2747 int migratetype;
2748
2749 migratetype = get_pcppage_migratetype(page);
2750 __count_vm_event(PGFREE);
2751
2752 /*
2753 * We only track unmovable, reclaimable and movable on pcp lists.
2754 * Free ISOLATE pages back to the allocator because they are being
2755 * offlined but treat HIGHATOMIC as movable pages so we can get those
2756 * areas back if necessary. Otherwise, we may have to free
2757 * excessively into the page allocator
2758 */
2759 if (migratetype >= MIGRATE_PCPTYPES) {
2760 if (unlikely(is_migrate_isolate(migratetype))) {
2761 free_one_page(zone, page, pfn, 0, migratetype);
2762 return;
2763 }
2764 migratetype = MIGRATE_MOVABLE;
2765 }
2766
2767 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2768 list_add(&page->lru, &pcp->lists[migratetype]);
2769 pcp->count++;
2770 if (pcp->count >= pcp->high) {
2771 unsigned long batch = READ_ONCE(pcp->batch);
2772 free_pcppages_bulk(zone, batch, pcp);
2773 }
2774}
2775
2776/*
2777 * Free a 0-order page
2778 */
2779void free_unref_page(struct page *page)
2780{
2781 unsigned long flags;
2782 unsigned long pfn = page_to_pfn(page);
2783
2784 if (!free_unref_page_prepare(page, pfn))
2785 return;
2786
2787 local_irq_save(flags);
2788 free_unref_page_commit(page, pfn);
2789 local_irq_restore(flags);
2790}
2791
2792/*
2793 * Free a list of 0-order pages
2794 */
2795void free_unref_page_list(struct list_head *list)
2796{
2797 struct page *page, *next;
2798 unsigned long flags, pfn;
2799 int batch_count = 0;
2800
2801 /* Prepare pages for freeing */
2802 list_for_each_entry_safe(page, next, list, lru) {
2803 pfn = page_to_pfn(page);
2804 if (!free_unref_page_prepare(page, pfn))
2805 list_del(&page->lru);
2806 set_page_private(page, pfn);
2807 }
2808
2809 local_irq_save(flags);
2810 list_for_each_entry_safe(page, next, list, lru) {
2811 unsigned long pfn = page_private(page);
2812
2813 set_page_private(page, 0);
2814 trace_mm_page_free_batched(page);
2815 free_unref_page_commit(page, pfn);
2816
2817 /*
2818 * Guard against excessive IRQ disabled times when we get
2819 * a large list of pages to free.
2820 */
2821 if (++batch_count == SWAP_CLUSTER_MAX) {
2822 local_irq_restore(flags);
2823 batch_count = 0;
2824 local_irq_save(flags);
2825 }
2826 }
2827 local_irq_restore(flags);
2828}
2829
2830/*
2831 * split_page takes a non-compound higher-order page, and splits it into
2832 * n (1<<order) sub-pages: page[0..n]
2833 * Each sub-page must be freed individually.
2834 *
2835 * Note: this is probably too low level an operation for use in drivers.
2836 * Please consult with lkml before using this in your driver.
2837 */
2838void split_page(struct page *page, unsigned int order)
2839{
2840 int i;
2841
2842 VM_BUG_ON_PAGE(PageCompound(page), page);
2843 VM_BUG_ON_PAGE(!page_count(page), page);
2844
2845 for (i = 1; i < (1 << order); i++)
2846 set_page_refcounted(page + i);
2847 split_page_owner(page, order);
2848}
2849EXPORT_SYMBOL_GPL(split_page);
2850
2851int __isolate_free_page(struct page *page, unsigned int order)
2852{
2853 unsigned long watermark;
2854 struct zone *zone;
2855 int mt;
2856
2857 BUG_ON(!PageBuddy(page));
2858
2859 zone = page_zone(page);
2860 mt = get_pageblock_migratetype(page);
2861
2862 if (!is_migrate_isolate(mt)) {
2863 /*
2864 * Obey watermarks as if the page was being allocated. We can
2865 * emulate a high-order watermark check with a raised order-0
2866 * watermark, because we already know our high-order page
2867 * exists.
2868 */
2869 watermark = min_wmark_pages(zone) + (1UL << order);
2870 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
2871 return 0;
2872
2873 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2874 }
2875
2876 /* Remove page from free list */
2877 list_del(&page->lru);
2878 zone->free_area[order].nr_free--;
2879 rmv_page_order(page);
2880
2881 /*
2882 * Set the pageblock if the isolated page is at least half of a
2883 * pageblock
2884 */
2885 if (order >= pageblock_order - 1) {
2886 struct page *endpage = page + (1 << order) - 1;
2887 for (; page < endpage; page += pageblock_nr_pages) {
2888 int mt = get_pageblock_migratetype(page);
2889 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
2890 && !is_migrate_highatomic(mt))
2891 set_pageblock_migratetype(page,
2892 MIGRATE_MOVABLE);
2893 }
2894 }
2895
2896
2897 return 1UL << order;
2898}
2899
2900/*
2901 * Update NUMA hit/miss statistics
2902 *
2903 * Must be called with interrupts disabled.
2904 */
2905static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
2906{
2907#ifdef CONFIG_NUMA
2908 enum numa_stat_item local_stat = NUMA_LOCAL;
2909
2910 /* skip numa counters update if numa stats is disabled */
2911 if (!static_branch_likely(&vm_numa_stat_key))
2912 return;
2913
2914 if (z->node != numa_node_id())
2915 local_stat = NUMA_OTHER;
2916
2917 if (z->node == preferred_zone->node)
2918 __inc_numa_state(z, NUMA_HIT);
2919 else {
2920 __inc_numa_state(z, NUMA_MISS);
2921 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
2922 }
2923 __inc_numa_state(z, local_stat);
2924#endif
2925}
2926
2927/* Remove page from the per-cpu list, caller must protect the list */
2928static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
2929 struct per_cpu_pages *pcp,
2930 struct list_head *list)
2931{
2932 struct page *page;
2933
2934 do {
2935 if (list_empty(list)) {
2936 pcp->count += rmqueue_bulk(zone, 0,
2937 pcp->batch, list,
2938 migratetype);
2939 if (unlikely(list_empty(list)))
2940 return NULL;
2941 }
2942
2943 page = list_first_entry(list, struct page, lru);
2944 list_del(&page->lru);
2945 pcp->count--;
2946 } while (check_new_pcp(page));
2947
2948 return page;
2949}
2950
2951/* Lock and remove page from the per-cpu list */
2952static struct page *rmqueue_pcplist(struct zone *preferred_zone,
2953 struct zone *zone, unsigned int order,
2954 gfp_t gfp_flags, int migratetype)
2955{
2956 struct per_cpu_pages *pcp;
2957 struct list_head *list;
2958 struct page *page;
2959 unsigned long flags;
2960
2961 local_irq_save(flags);
2962 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2963 list = &pcp->lists[migratetype];
2964 page = __rmqueue_pcplist(zone, migratetype, pcp, list);
2965 if (page) {
2966 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2967 zone_statistics(preferred_zone, zone);
2968 }
2969 local_irq_restore(flags);
2970 return page;
2971}
2972
2973/*
2974 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2975 */
2976static inline
2977struct page *rmqueue(struct zone *preferred_zone,
2978 struct zone *zone, unsigned int order,
2979 gfp_t gfp_flags, unsigned int alloc_flags,
2980 int migratetype)
2981{
2982 unsigned long flags;
2983 struct page *page;
2984
2985 if (likely(order == 0)) {
2986 page = rmqueue_pcplist(preferred_zone, zone, order,
2987 gfp_flags, migratetype);
2988 goto out;
2989 }
2990
2991 /*
2992 * We most definitely don't want callers attempting to
2993 * allocate greater than order-1 page units with __GFP_NOFAIL.
2994 */
2995 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
2996 spin_lock_irqsave(&zone->lock, flags);
2997
2998 do {
2999 page = NULL;
3000 if (alloc_flags & ALLOC_HARDER) {
3001 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3002 if (page)
3003 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3004 }
3005 if (!page)
3006 page = __rmqueue(zone, order, migratetype);
3007 } while (page && check_new_pages(page, order));
3008 spin_unlock(&zone->lock);
3009 if (!page)
3010 goto failed;
3011 __mod_zone_freepage_state(zone, -(1 << order),
3012 get_pcppage_migratetype(page));
3013
3014 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3015 zone_statistics(preferred_zone, zone);
3016 local_irq_restore(flags);
3017
3018out:
3019 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3020 return page;
3021
3022failed:
3023 local_irq_restore(flags);
3024 return NULL;
3025}
3026
3027#ifdef CONFIG_FAIL_PAGE_ALLOC
3028
3029static struct {
3030 struct fault_attr attr;
3031
3032 bool ignore_gfp_highmem;
3033 bool ignore_gfp_reclaim;
3034 u32 min_order;
3035} fail_page_alloc = {
3036 .attr = FAULT_ATTR_INITIALIZER,
3037 .ignore_gfp_reclaim = true,
3038 .ignore_gfp_highmem = true,
3039 .min_order = 1,
3040};
3041
3042static int __init setup_fail_page_alloc(char *str)
3043{
3044 return setup_fault_attr(&fail_page_alloc.attr, str);
3045}
3046__setup("fail_page_alloc=", setup_fail_page_alloc);
3047
3048static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3049{
3050 if (order < fail_page_alloc.min_order)
3051 return false;
3052 if (gfp_mask & __GFP_NOFAIL)
3053 return false;
3054 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3055 return false;
3056 if (fail_page_alloc.ignore_gfp_reclaim &&
3057 (gfp_mask & __GFP_DIRECT_RECLAIM))
3058 return false;
3059
3060 return should_fail(&fail_page_alloc.attr, 1 << order);
3061}
3062
3063#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3064
3065static int __init fail_page_alloc_debugfs(void)
3066{
3067 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
3068 struct dentry *dir;
3069
3070 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3071 &fail_page_alloc.attr);
3072 if (IS_ERR(dir))
3073 return PTR_ERR(dir);
3074
3075 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
3076 &fail_page_alloc.ignore_gfp_reclaim))
3077 goto fail;
3078 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3079 &fail_page_alloc.ignore_gfp_highmem))
3080 goto fail;
3081 if (!debugfs_create_u32("min-order", mode, dir,
3082 &fail_page_alloc.min_order))
3083 goto fail;
3084
3085 return 0;
3086fail:
3087 debugfs_remove_recursive(dir);
3088
3089 return -ENOMEM;
3090}
3091
3092late_initcall(fail_page_alloc_debugfs);
3093
3094#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3095
3096#else /* CONFIG_FAIL_PAGE_ALLOC */
3097
3098static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3099{
3100 return false;
3101}
3102
3103#endif /* CONFIG_FAIL_PAGE_ALLOC */
3104
3105/*
3106 * Return true if free base pages are above 'mark'. For high-order checks it
3107 * will return true of the order-0 watermark is reached and there is at least
3108 * one free page of a suitable size. Checking now avoids taking the zone lock
3109 * to check in the allocation paths if no pages are free.
3110 */
3111bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3112 int classzone_idx, unsigned int alloc_flags,
3113 long free_pages)
3114{
3115 long min = mark;
3116 int o;
3117 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3118
3119 /* free_pages may go negative - that's OK */
3120 free_pages -= (1 << order) - 1;
3121
3122 if (alloc_flags & ALLOC_HIGH)
3123 min -= min / 2;
3124
3125 /*
3126 * If the caller does not have rights to ALLOC_HARDER then subtract
3127 * the high-atomic reserves. This will over-estimate the size of the
3128 * atomic reserve but it avoids a search.
3129 */
3130 if (likely(!alloc_harder)) {
3131 free_pages -= z->nr_reserved_highatomic;
3132 } else {
3133 /*
3134 * OOM victims can try even harder than normal ALLOC_HARDER
3135 * users on the grounds that it's definitely going to be in
3136 * the exit path shortly and free memory. Any allocation it
3137 * makes during the free path will be small and short-lived.
3138 */
3139 if (alloc_flags & ALLOC_OOM)
3140 min -= min / 2;
3141 else
3142 min -= min / 4;
3143 }
3144
3145
3146#ifdef CONFIG_CMA
3147 /* If allocation can't use CMA areas don't use free CMA pages */
3148 if (!(alloc_flags & ALLOC_CMA))
3149 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
3150#endif
3151
3152 /*
3153 * Check watermarks for an order-0 allocation request. If these
3154 * are not met, then a high-order request also cannot go ahead
3155 * even if a suitable page happened to be free.
3156 */
3157 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
3158 return false;
3159
3160 /* If this is an order-0 request then the watermark is fine */
3161 if (!order)
3162 return true;
3163
3164 /* For a high-order request, check at least one suitable page is free */
3165 for (o = order; o < MAX_ORDER; o++) {
3166 struct free_area *area = &z->free_area[o];
3167 int mt;
3168
3169 if (!area->nr_free)
3170 continue;
3171
3172 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3173 if (!list_empty(&area->free_list[mt]))
3174 return true;
3175 }
3176
3177#ifdef CONFIG_CMA
3178 if ((alloc_flags & ALLOC_CMA) &&
3179 !list_empty(&area->free_list[MIGRATE_CMA])) {
3180 return true;
3181 }
3182#endif
3183 if (alloc_harder &&
3184 !list_empty(&area->free_list[MIGRATE_HIGHATOMIC]))
3185 return true;
3186 }
3187 return false;
3188}
3189
3190bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3191 int classzone_idx, unsigned int alloc_flags)
3192{
3193 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3194 zone_page_state(z, NR_FREE_PAGES));
3195}
3196
3197static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3198 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
3199{
3200 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3201 long cma_pages = 0;
3202
3203#ifdef CONFIG_CMA
3204 /* If allocation can't use CMA areas don't use free CMA pages */
3205 if (!(alloc_flags & ALLOC_CMA))
3206 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
3207#endif
3208
3209 /*
3210 * Fast check for order-0 only. If this fails then the reserves
3211 * need to be calculated. There is a corner case where the check
3212 * passes but only the high-order atomic reserve are free. If
3213 * the caller is !atomic then it'll uselessly search the free
3214 * list. That corner case is then slower but it is harmless.
3215 */
3216 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
3217 return true;
3218
3219 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3220 free_pages);
3221}
3222
3223bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3224 unsigned long mark, int classzone_idx)
3225{
3226 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3227
3228 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3229 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3230
3231 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
3232 free_pages);
3233}
3234
3235#ifdef CONFIG_NUMA
3236static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3237{
3238 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3239 RECLAIM_DISTANCE;
3240}
3241#else /* CONFIG_NUMA */
3242static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3243{
3244 return true;
3245}
3246#endif /* CONFIG_NUMA */
3247
3248/*
3249 * get_page_from_freelist goes through the zonelist trying to allocate
3250 * a page.
3251 */
3252static struct page *
3253get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3254 const struct alloc_context *ac)
3255{
3256 struct zoneref *z = ac->preferred_zoneref;
3257 struct zone *zone;
3258 struct pglist_data *last_pgdat_dirty_limit = NULL;
3259
3260 /*
3261 * Scan zonelist, looking for a zone with enough free.
3262 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3263 */
3264 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3265 ac->nodemask) {
3266 struct page *page;
3267 unsigned long mark;
3268
3269 if (cpusets_enabled() &&
3270 (alloc_flags & ALLOC_CPUSET) &&
3271 !__cpuset_zone_allowed(zone, gfp_mask))
3272 continue;
3273 /*
3274 * When allocating a page cache page for writing, we
3275 * want to get it from a node that is within its dirty
3276 * limit, such that no single node holds more than its
3277 * proportional share of globally allowed dirty pages.
3278 * The dirty limits take into account the node's
3279 * lowmem reserves and high watermark so that kswapd
3280 * should be able to balance it without having to
3281 * write pages from its LRU list.
3282 *
3283 * XXX: For now, allow allocations to potentially
3284 * exceed the per-node dirty limit in the slowpath
3285 * (spread_dirty_pages unset) before going into reclaim,
3286 * which is important when on a NUMA setup the allowed
3287 * nodes are together not big enough to reach the
3288 * global limit. The proper fix for these situations
3289 * will require awareness of nodes in the
3290 * dirty-throttling and the flusher threads.
3291 */
3292 if (ac->spread_dirty_pages) {
3293 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3294 continue;
3295
3296 if (!node_dirty_ok(zone->zone_pgdat)) {
3297 last_pgdat_dirty_limit = zone->zone_pgdat;
3298 continue;
3299 }
3300 }
3301
3302 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
3303 if (!zone_watermark_fast(zone, order, mark,
3304 ac_classzone_idx(ac), alloc_flags)) {
3305 int ret;
3306
3307#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3308 /*
3309 * Watermark failed for this zone, but see if we can
3310 * grow this zone if it contains deferred pages.
3311 */
3312 if (static_branch_unlikely(&deferred_pages)) {
3313 if (_deferred_grow_zone(zone, order))
3314 goto try_this_zone;
3315 }
3316#endif
3317 /* Checked here to keep the fast path fast */
3318 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3319 if (alloc_flags & ALLOC_NO_WATERMARKS)
3320 goto try_this_zone;
3321
3322 if (node_reclaim_mode == 0 ||
3323 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3324 continue;
3325
3326 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3327 switch (ret) {
3328 case NODE_RECLAIM_NOSCAN:
3329 /* did not scan */
3330 continue;
3331 case NODE_RECLAIM_FULL:
3332 /* scanned but unreclaimable */
3333 continue;
3334 default:
3335 /* did we reclaim enough */
3336 if (zone_watermark_ok(zone, order, mark,
3337 ac_classzone_idx(ac), alloc_flags))
3338 goto try_this_zone;
3339
3340 continue;
3341 }
3342 }
3343
3344try_this_zone:
3345 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3346 gfp_mask, alloc_flags, ac->migratetype);
3347 if (page) {
3348 prep_new_page(page, order, gfp_mask, alloc_flags);
3349
3350 /*
3351 * If this is a high-order atomic allocation then check
3352 * if the pageblock should be reserved for the future
3353 */
3354 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3355 reserve_highatomic_pageblock(page, zone, order);
3356
3357 return page;
3358 } else {
3359#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3360 /* Try again if zone has deferred pages */
3361 if (static_branch_unlikely(&deferred_pages)) {
3362 if (_deferred_grow_zone(zone, order))
3363 goto try_this_zone;
3364 }
3365#endif
3366 }
3367 }
3368
3369 return NULL;
3370}
3371
3372/*
3373 * Large machines with many possible nodes should not always dump per-node
3374 * meminfo in irq context.
3375 */
3376static inline bool should_suppress_show_mem(void)
3377{
3378 bool ret = false;
3379
3380#if NODES_SHIFT > 8
3381 ret = in_interrupt();
3382#endif
3383 return ret;
3384}
3385
3386static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3387{
3388 unsigned int filter = SHOW_MEM_FILTER_NODES;
3389 static DEFINE_RATELIMIT_STATE(show_mem_rs, HZ, 1);
3390
3391 if (should_suppress_show_mem() || !__ratelimit(&show_mem_rs))
3392 return;
3393
3394 /*
3395 * This documents exceptions given to allocations in certain
3396 * contexts that are allowed to allocate outside current's set
3397 * of allowed nodes.
3398 */
3399 if (!(gfp_mask & __GFP_NOMEMALLOC))
3400 if (tsk_is_oom_victim(current) ||
3401 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3402 filter &= ~SHOW_MEM_FILTER_NODES;
3403 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3404 filter &= ~SHOW_MEM_FILTER_NODES;
3405
3406 show_mem(filter, nodemask);
3407}
3408
3409void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3410{
3411 struct va_format vaf;
3412 va_list args;
3413 static DEFINE_RATELIMIT_STATE(nopage_rs, DEFAULT_RATELIMIT_INTERVAL,
3414 DEFAULT_RATELIMIT_BURST);
3415
3416 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
3417 return;
3418
3419 va_start(args, fmt);
3420 vaf.fmt = fmt;
3421 vaf.va = &args;
3422 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl\n",
3423 current->comm, &vaf, gfp_mask, &gfp_mask,
3424 nodemask_pr_args(nodemask));
3425 va_end(args);
3426
3427 cpuset_print_current_mems_allowed();
3428
3429 dump_stack();
3430 warn_alloc_show_mem(gfp_mask, nodemask);
3431}
3432
3433static inline struct page *
3434__alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3435 unsigned int alloc_flags,
3436 const struct alloc_context *ac)
3437{
3438 struct page *page;
3439
3440 page = get_page_from_freelist(gfp_mask, order,
3441 alloc_flags|ALLOC_CPUSET, ac);
3442 /*
3443 * fallback to ignore cpuset restriction if our nodes
3444 * are depleted
3445 */
3446 if (!page)
3447 page = get_page_from_freelist(gfp_mask, order,
3448 alloc_flags, ac);
3449
3450 return page;
3451}
3452
3453static inline struct page *
3454__alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3455 const struct alloc_context *ac, unsigned long *did_some_progress)
3456{
3457 struct oom_control oc = {
3458 .zonelist = ac->zonelist,
3459 .nodemask = ac->nodemask,
3460 .memcg = NULL,
3461 .gfp_mask = gfp_mask,
3462 .order = order,
3463 };
3464 struct page *page;
3465
3466 *did_some_progress = 0;
3467
3468 /*
3469 * Acquire the oom lock. If that fails, somebody else is
3470 * making progress for us.
3471 */
3472 if (!mutex_trylock(&oom_lock)) {
3473 *did_some_progress = 1;
3474 schedule_timeout_uninterruptible(1);
3475 return NULL;
3476 }
3477
3478 /*
3479 * Go through the zonelist yet one more time, keep very high watermark
3480 * here, this is only to catch a parallel oom killing, we must fail if
3481 * we're still under heavy pressure. But make sure that this reclaim
3482 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3483 * allocation which will never fail due to oom_lock already held.
3484 */
3485 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3486 ~__GFP_DIRECT_RECLAIM, order,
3487 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3488 if (page)
3489 goto out;
3490
3491 /* Coredumps can quickly deplete all memory reserves */
3492 if (current->flags & PF_DUMPCORE)
3493 goto out;
3494 /* The OOM killer will not help higher order allocs */
3495 if (order > PAGE_ALLOC_COSTLY_ORDER)
3496 goto out;
3497 /*
3498 * We have already exhausted all our reclaim opportunities without any
3499 * success so it is time to admit defeat. We will skip the OOM killer
3500 * because it is very likely that the caller has a more reasonable
3501 * fallback than shooting a random task.
3502 */
3503 if (gfp_mask & __GFP_RETRY_MAYFAIL)
3504 goto out;
3505 /* The OOM killer does not needlessly kill tasks for lowmem */
3506 if (ac->high_zoneidx < ZONE_NORMAL)
3507 goto out;
3508 if (pm_suspended_storage())
3509 goto out;
3510 /*
3511 * XXX: GFP_NOFS allocations should rather fail than rely on
3512 * other request to make a forward progress.
3513 * We are in an unfortunate situation where out_of_memory cannot
3514 * do much for this context but let's try it to at least get
3515 * access to memory reserved if the current task is killed (see
3516 * out_of_memory). Once filesystems are ready to handle allocation
3517 * failures more gracefully we should just bail out here.
3518 */
3519
3520 /* The OOM killer may not free memory on a specific node */
3521 if (gfp_mask & __GFP_THISNODE)
3522 goto out;
3523
3524 /* Exhausted what can be done so it's blame time */
3525 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3526 *did_some_progress = 1;
3527
3528 /*
3529 * Help non-failing allocations by giving them access to memory
3530 * reserves
3531 */
3532 if (gfp_mask & __GFP_NOFAIL)
3533 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3534 ALLOC_NO_WATERMARKS, ac);
3535 }
3536out:
3537 mutex_unlock(&oom_lock);
3538 return page;
3539}
3540
3541/*
3542 * Maximum number of compaction retries wit a progress before OOM
3543 * killer is consider as the only way to move forward.
3544 */
3545#define MAX_COMPACT_RETRIES 16
3546
3547#ifdef CONFIG_COMPACTION
3548/* Try memory compaction for high-order allocations before reclaim */
3549static struct page *
3550__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3551 unsigned int alloc_flags, const struct alloc_context *ac,
3552 enum compact_priority prio, enum compact_result *compact_result)
3553{
3554 struct page *page;
3555 unsigned int noreclaim_flag;
3556
3557 if (!order)
3558 return NULL;
3559
3560 noreclaim_flag = memalloc_noreclaim_save();
3561 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3562 prio);
3563 memalloc_noreclaim_restore(noreclaim_flag);
3564
3565 if (*compact_result <= COMPACT_INACTIVE)
3566 return NULL;
3567
3568 /*
3569 * At least in one zone compaction wasn't deferred or skipped, so let's
3570 * count a compaction stall
3571 */
3572 count_vm_event(COMPACTSTALL);
3573
3574 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3575
3576 if (page) {
3577 struct zone *zone = page_zone(page);
3578
3579 zone->compact_blockskip_flush = false;
3580 compaction_defer_reset(zone, order, true);
3581 count_vm_event(COMPACTSUCCESS);
3582 return page;
3583 }
3584
3585 /*
3586 * It's bad if compaction run occurs and fails. The most likely reason
3587 * is that pages exist, but not enough to satisfy watermarks.
3588 */
3589 count_vm_event(COMPACTFAIL);
3590
3591 cond_resched();
3592
3593 return NULL;
3594}
3595
3596static inline bool
3597should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3598 enum compact_result compact_result,
3599 enum compact_priority *compact_priority,
3600 int *compaction_retries)
3601{
3602 int max_retries = MAX_COMPACT_RETRIES;
3603 int min_priority;
3604 bool ret = false;
3605 int retries = *compaction_retries;
3606 enum compact_priority priority = *compact_priority;
3607
3608 if (!order)
3609 return false;
3610
3611 if (compaction_made_progress(compact_result))
3612 (*compaction_retries)++;
3613
3614 /*
3615 * compaction considers all the zone as desperately out of memory
3616 * so it doesn't really make much sense to retry except when the
3617 * failure could be caused by insufficient priority
3618 */
3619 if (compaction_failed(compact_result))
3620 goto check_priority;
3621
3622 /*
3623 * make sure the compaction wasn't deferred or didn't bail out early
3624 * due to locks contention before we declare that we should give up.
3625 * But do not retry if the given zonelist is not suitable for
3626 * compaction.
3627 */
3628 if (compaction_withdrawn(compact_result)) {
3629 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
3630 goto out;
3631 }
3632
3633 /*
3634 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
3635 * costly ones because they are de facto nofail and invoke OOM
3636 * killer to move on while costly can fail and users are ready
3637 * to cope with that. 1/4 retries is rather arbitrary but we
3638 * would need much more detailed feedback from compaction to
3639 * make a better decision.
3640 */
3641 if (order > PAGE_ALLOC_COSTLY_ORDER)
3642 max_retries /= 4;
3643 if (*compaction_retries <= max_retries) {
3644 ret = true;
3645 goto out;
3646 }
3647
3648 /*
3649 * Make sure there are attempts at the highest priority if we exhausted
3650 * all retries or failed at the lower priorities.
3651 */
3652check_priority:
3653 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
3654 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
3655
3656 if (*compact_priority > min_priority) {
3657 (*compact_priority)--;
3658 *compaction_retries = 0;
3659 ret = true;
3660 }
3661out:
3662 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
3663 return ret;
3664}
3665#else
3666static inline struct page *
3667__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3668 unsigned int alloc_flags, const struct alloc_context *ac,
3669 enum compact_priority prio, enum compact_result *compact_result)
3670{
3671 *compact_result = COMPACT_SKIPPED;
3672 return NULL;
3673}
3674
3675static inline bool
3676should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
3677 enum compact_result compact_result,
3678 enum compact_priority *compact_priority,
3679 int *compaction_retries)
3680{
3681 struct zone *zone;
3682 struct zoneref *z;
3683
3684 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
3685 return false;
3686
3687 /*
3688 * There are setups with compaction disabled which would prefer to loop
3689 * inside the allocator rather than hit the oom killer prematurely.
3690 * Let's give them a good hope and keep retrying while the order-0
3691 * watermarks are OK.
3692 */
3693 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3694 ac->nodemask) {
3695 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
3696 ac_classzone_idx(ac), alloc_flags))
3697 return true;
3698 }
3699 return false;
3700}
3701#endif /* CONFIG_COMPACTION */
3702
3703#ifdef CONFIG_LOCKDEP
3704struct lockdep_map __fs_reclaim_map =
3705 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
3706
3707static bool __need_fs_reclaim(gfp_t gfp_mask)
3708{
3709 gfp_mask = current_gfp_context(gfp_mask);
3710
3711 /* no reclaim without waiting on it */
3712 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
3713 return false;
3714
3715 /* this guy won't enter reclaim */
3716 if (current->flags & PF_MEMALLOC)
3717 return false;
3718
3719 /* We're only interested __GFP_FS allocations for now */
3720 if (!(gfp_mask & __GFP_FS))
3721 return false;
3722
3723 if (gfp_mask & __GFP_NOLOCKDEP)
3724 return false;
3725
3726 return true;
3727}
3728
3729void fs_reclaim_acquire(gfp_t gfp_mask)
3730{
3731 if (__need_fs_reclaim(gfp_mask))
3732 lock_map_acquire(&__fs_reclaim_map);
3733}
3734EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
3735
3736void fs_reclaim_release(gfp_t gfp_mask)
3737{
3738 if (__need_fs_reclaim(gfp_mask))
3739 lock_map_release(&__fs_reclaim_map);
3740}
3741EXPORT_SYMBOL_GPL(fs_reclaim_release);
3742#endif
3743
3744/* Perform direct synchronous page reclaim */
3745static int
3746__perform_reclaim(gfp_t gfp_mask, unsigned int order,
3747 const struct alloc_context *ac)
3748{
3749 struct reclaim_state reclaim_state;
3750 int progress;
3751 unsigned int noreclaim_flag;
3752
3753 cond_resched();
3754
3755 /* We now go into synchronous reclaim */
3756 cpuset_memory_pressure_bump();
3757 noreclaim_flag = memalloc_noreclaim_save();
3758 fs_reclaim_acquire(gfp_mask);
3759 reclaim_state.reclaimed_slab = 0;
3760 current->reclaim_state = &reclaim_state;
3761
3762 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
3763 ac->nodemask);
3764
3765 current->reclaim_state = NULL;
3766 fs_reclaim_release(gfp_mask);
3767 memalloc_noreclaim_restore(noreclaim_flag);
3768
3769 cond_resched();
3770
3771 return progress;
3772}
3773
3774/* The really slow allocator path where we enter direct reclaim */
3775static inline struct page *
3776__alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
3777 unsigned int alloc_flags, const struct alloc_context *ac,
3778 unsigned long *did_some_progress)
3779{
3780 struct page *page = NULL;
3781 bool drained = false;
3782
3783 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
3784 if (unlikely(!(*did_some_progress)))
3785 return NULL;
3786
3787retry:
3788 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3789
3790 /*
3791 * If an allocation failed after direct reclaim, it could be because
3792 * pages are pinned on the per-cpu lists or in high alloc reserves.
3793 * Shrink them them and try again
3794 */
3795 if (!page && !drained) {
3796 unreserve_highatomic_pageblock(ac, false);
3797 drain_all_pages(NULL);
3798 drained = true;
3799 goto retry;
3800 }
3801
3802 return page;
3803}
3804
3805static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
3806 const struct alloc_context *ac)
3807{
3808 struct zoneref *z;
3809 struct zone *zone;
3810 pg_data_t *last_pgdat = NULL;
3811 enum zone_type high_zoneidx = ac->high_zoneidx;
3812
3813 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, high_zoneidx,
3814 ac->nodemask) {
3815 if (last_pgdat != zone->zone_pgdat)
3816 wakeup_kswapd(zone, gfp_mask, order, high_zoneidx);
3817 last_pgdat = zone->zone_pgdat;
3818 }
3819}
3820
3821static inline unsigned int
3822gfp_to_alloc_flags(gfp_t gfp_mask)
3823{
3824 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
3825
3826 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3827 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
3828
3829 /*
3830 * The caller may dip into page reserves a bit more if the caller
3831 * cannot run direct reclaim, or if the caller has realtime scheduling
3832 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3833 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3834 */
3835 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
3836
3837 if (gfp_mask & __GFP_ATOMIC) {
3838 /*
3839 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3840 * if it can't schedule.
3841 */
3842 if (!(gfp_mask & __GFP_NOMEMALLOC))
3843 alloc_flags |= ALLOC_HARDER;
3844 /*
3845 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3846 * comment for __cpuset_node_allowed().
3847 */
3848 alloc_flags &= ~ALLOC_CPUSET;
3849 } else if (unlikely(rt_task(current)) && !in_interrupt())
3850 alloc_flags |= ALLOC_HARDER;
3851
3852#ifdef CONFIG_CMA
3853 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3854 alloc_flags |= ALLOC_CMA;
3855#endif
3856 return alloc_flags;
3857}
3858
3859static bool oom_reserves_allowed(struct task_struct *tsk)
3860{
3861 if (!tsk_is_oom_victim(tsk))
3862 return false;
3863
3864 /*
3865 * !MMU doesn't have oom reaper so give access to memory reserves
3866 * only to the thread with TIF_MEMDIE set
3867 */
3868 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
3869 return false;
3870
3871 return true;
3872}
3873
3874/*
3875 * Distinguish requests which really need access to full memory
3876 * reserves from oom victims which can live with a portion of it
3877 */
3878static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
3879{
3880 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
3881 return 0;
3882 if (gfp_mask & __GFP_MEMALLOC)
3883 return ALLOC_NO_WATERMARKS;
3884 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
3885 return ALLOC_NO_WATERMARKS;
3886 if (!in_interrupt()) {
3887 if (current->flags & PF_MEMALLOC)
3888 return ALLOC_NO_WATERMARKS;
3889 else if (oom_reserves_allowed(current))
3890 return ALLOC_OOM;
3891 }
3892
3893 return 0;
3894}
3895
3896bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
3897{
3898 return !!__gfp_pfmemalloc_flags(gfp_mask);
3899}
3900
3901/*
3902 * Checks whether it makes sense to retry the reclaim to make a forward progress
3903 * for the given allocation request.
3904 *
3905 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
3906 * without success, or when we couldn't even meet the watermark if we
3907 * reclaimed all remaining pages on the LRU lists.
3908 *
3909 * Returns true if a retry is viable or false to enter the oom path.
3910 */
3911static inline bool
3912should_reclaim_retry(gfp_t gfp_mask, unsigned order,
3913 struct alloc_context *ac, int alloc_flags,
3914 bool did_some_progress, int *no_progress_loops)
3915{
3916 struct zone *zone;
3917 struct zoneref *z;
3918
3919 /*
3920 * Costly allocations might have made a progress but this doesn't mean
3921 * their order will become available due to high fragmentation so
3922 * always increment the no progress counter for them
3923 */
3924 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
3925 *no_progress_loops = 0;
3926 else
3927 (*no_progress_loops)++;
3928
3929 /*
3930 * Make sure we converge to OOM if we cannot make any progress
3931 * several times in the row.
3932 */
3933 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
3934 /* Before OOM, exhaust highatomic_reserve */
3935 return unreserve_highatomic_pageblock(ac, true);
3936 }
3937
3938 /*
3939 * Keep reclaiming pages while there is a chance this will lead
3940 * somewhere. If none of the target zones can satisfy our allocation
3941 * request even if all reclaimable pages are considered then we are
3942 * screwed and have to go OOM.
3943 */
3944 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3945 ac->nodemask) {
3946 unsigned long available;
3947 unsigned long reclaimable;
3948 unsigned long min_wmark = min_wmark_pages(zone);
3949 bool wmark;
3950
3951 available = reclaimable = zone_reclaimable_pages(zone);
3952 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
3953
3954 /*
3955 * Would the allocation succeed if we reclaimed all
3956 * reclaimable pages?
3957 */
3958 wmark = __zone_watermark_ok(zone, order, min_wmark,
3959 ac_classzone_idx(ac), alloc_flags, available);
3960 trace_reclaim_retry_zone(z, order, reclaimable,
3961 available, min_wmark, *no_progress_loops, wmark);
3962 if (wmark) {
3963 /*
3964 * If we didn't make any progress and have a lot of
3965 * dirty + writeback pages then we should wait for
3966 * an IO to complete to slow down the reclaim and
3967 * prevent from pre mature OOM
3968 */
3969 if (!did_some_progress) {
3970 unsigned long write_pending;
3971
3972 write_pending = zone_page_state_snapshot(zone,
3973 NR_ZONE_WRITE_PENDING);
3974
3975 if (2 * write_pending > reclaimable) {
3976 congestion_wait(BLK_RW_ASYNC, HZ/10);
3977 return true;
3978 }
3979 }
3980
3981 /*
3982 * Memory allocation/reclaim might be called from a WQ
3983 * context and the current implementation of the WQ
3984 * concurrency control doesn't recognize that
3985 * a particular WQ is congested if the worker thread is
3986 * looping without ever sleeping. Therefore we have to
3987 * do a short sleep here rather than calling
3988 * cond_resched().
3989 */
3990 if (current->flags & PF_WQ_WORKER)
3991 schedule_timeout_uninterruptible(1);
3992 else
3993 cond_resched();
3994
3995 return true;
3996 }
3997 }
3998
3999 return false;
4000}
4001
4002static inline bool
4003check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4004{
4005 /*
4006 * It's possible that cpuset's mems_allowed and the nodemask from
4007 * mempolicy don't intersect. This should be normally dealt with by
4008 * policy_nodemask(), but it's possible to race with cpuset update in
4009 * such a way the check therein was true, and then it became false
4010 * before we got our cpuset_mems_cookie here.
4011 * This assumes that for all allocations, ac->nodemask can come only
4012 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4013 * when it does not intersect with the cpuset restrictions) or the
4014 * caller can deal with a violated nodemask.
4015 */
4016 if (cpusets_enabled() && ac->nodemask &&
4017 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4018 ac->nodemask = NULL;
4019 return true;
4020 }
4021
4022 /*
4023 * When updating a task's mems_allowed or mempolicy nodemask, it is
4024 * possible to race with parallel threads in such a way that our
4025 * allocation can fail while the mask is being updated. If we are about
4026 * to fail, check if the cpuset changed during allocation and if so,
4027 * retry.
4028 */
4029 if (read_mems_allowed_retry(cpuset_mems_cookie))
4030 return true;
4031
4032 return false;
4033}
4034
4035static inline struct page *
4036__alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4037 struct alloc_context *ac)
4038{
4039 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4040 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4041 struct page *page = NULL;
4042 unsigned int alloc_flags;
4043 unsigned long did_some_progress;
4044 enum compact_priority compact_priority;
4045 enum compact_result compact_result;
4046 int compaction_retries;
4047 int no_progress_loops;
4048 unsigned int cpuset_mems_cookie;
4049 int reserve_flags;
4050
4051 /*
4052 * In the slowpath, we sanity check order to avoid ever trying to
4053 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
4054 * be using allocators in order of preference for an area that is
4055 * too large.
4056 */
4057 if (order >= MAX_ORDER) {
4058 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
4059 return NULL;
4060 }
4061
4062 /*
4063 * We also sanity check to catch abuse of atomic reserves being used by
4064 * callers that are not in atomic context.
4065 */
4066 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4067 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4068 gfp_mask &= ~__GFP_ATOMIC;
4069
4070retry_cpuset:
4071 compaction_retries = 0;
4072 no_progress_loops = 0;
4073 compact_priority = DEF_COMPACT_PRIORITY;
4074 cpuset_mems_cookie = read_mems_allowed_begin();
4075
4076 /*
4077 * The fast path uses conservative alloc_flags to succeed only until
4078 * kswapd needs to be woken up, and to avoid the cost of setting up
4079 * alloc_flags precisely. So we do that now.
4080 */
4081 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4082
4083 /*
4084 * We need to recalculate the starting point for the zonelist iterator
4085 * because we might have used different nodemask in the fast path, or
4086 * there was a cpuset modification and we are retrying - otherwise we
4087 * could end up iterating over non-eligible zones endlessly.
4088 */
4089 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4090 ac->high_zoneidx, ac->nodemask);
4091 if (!ac->preferred_zoneref->zone)
4092 goto nopage;
4093
4094 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
4095 wake_all_kswapds(order, gfp_mask, ac);
4096
4097 /*
4098 * The adjusted alloc_flags might result in immediate success, so try
4099 * that first
4100 */
4101 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4102 if (page)
4103 goto got_pg;
4104
4105 /*
4106 * For costly allocations, try direct compaction first, as it's likely
4107 * that we have enough base pages and don't need to reclaim. For non-
4108 * movable high-order allocations, do that as well, as compaction will
4109 * try prevent permanent fragmentation by migrating from blocks of the
4110 * same migratetype.
4111 * Don't try this for allocations that are allowed to ignore
4112 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4113 */
4114 if (can_direct_reclaim &&
4115 (costly_order ||
4116 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4117 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4118 page = __alloc_pages_direct_compact(gfp_mask, order,
4119 alloc_flags, ac,
4120 INIT_COMPACT_PRIORITY,
4121 &compact_result);
4122 if (page)
4123 goto got_pg;
4124
4125 /*
4126 * Checks for costly allocations with __GFP_NORETRY, which
4127 * includes THP page fault allocations
4128 */
4129 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4130 /*
4131 * If compaction is deferred for high-order allocations,
4132 * it is because sync compaction recently failed. If
4133 * this is the case and the caller requested a THP
4134 * allocation, we do not want to heavily disrupt the
4135 * system, so we fail the allocation instead of entering
4136 * direct reclaim.
4137 */
4138 if (compact_result == COMPACT_DEFERRED)
4139 goto nopage;
4140
4141 /*
4142 * Looks like reclaim/compaction is worth trying, but
4143 * sync compaction could be very expensive, so keep
4144 * using async compaction.
4145 */
4146 compact_priority = INIT_COMPACT_PRIORITY;
4147 }
4148 }
4149
4150retry:
4151 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4152 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
4153 wake_all_kswapds(order, gfp_mask, ac);
4154
4155 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4156 if (reserve_flags)
4157 alloc_flags = reserve_flags;
4158
4159 /*
4160 * Reset the zonelist iterators if memory policies can be ignored.
4161 * These allocations are high priority and system rather than user
4162 * orientated.
4163 */
4164 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4165 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
4166 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4167 ac->high_zoneidx, ac->nodemask);
4168 }
4169
4170 /* Attempt with potentially adjusted zonelist and alloc_flags */
4171 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4172 if (page)
4173 goto got_pg;
4174
4175 /* Caller is not willing to reclaim, we can't balance anything */
4176 if (!can_direct_reclaim)
4177 goto nopage;
4178
4179 /* Avoid recursion of direct reclaim */
4180 if (current->flags & PF_MEMALLOC)
4181 goto nopage;
4182
4183 /* Try direct reclaim and then allocating */
4184 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4185 &did_some_progress);
4186 if (page)
4187 goto got_pg;
4188
4189 /* Try direct compaction and then allocating */
4190 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4191 compact_priority, &compact_result);
4192 if (page)
4193 goto got_pg;
4194
4195 /* Do not loop if specifically requested */
4196 if (gfp_mask & __GFP_NORETRY)
4197 goto nopage;
4198
4199 /*
4200 * Do not retry costly high order allocations unless they are
4201 * __GFP_RETRY_MAYFAIL
4202 */
4203 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4204 goto nopage;
4205
4206 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4207 did_some_progress > 0, &no_progress_loops))
4208 goto retry;
4209
4210 /*
4211 * It doesn't make any sense to retry for the compaction if the order-0
4212 * reclaim is not able to make any progress because the current
4213 * implementation of the compaction depends on the sufficient amount
4214 * of free memory (see __compaction_suitable)
4215 */
4216 if (did_some_progress > 0 &&
4217 should_compact_retry(ac, order, alloc_flags,
4218 compact_result, &compact_priority,
4219 &compaction_retries))
4220 goto retry;
4221
4222
4223 /* Deal with possible cpuset update races before we start OOM killing */
4224 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4225 goto retry_cpuset;
4226
4227 /* Reclaim has failed us, start killing things */
4228 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4229 if (page)
4230 goto got_pg;
4231
4232 /* Avoid allocations with no watermarks from looping endlessly */
4233 if (tsk_is_oom_victim(current) &&
4234 (alloc_flags == ALLOC_OOM ||
4235 (gfp_mask & __GFP_NOMEMALLOC)))
4236 goto nopage;
4237
4238 /* Retry as long as the OOM killer is making progress */
4239 if (did_some_progress) {
4240 no_progress_loops = 0;
4241 goto retry;
4242 }
4243
4244nopage:
4245 /* Deal with possible cpuset update races before we fail */
4246 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4247 goto retry_cpuset;
4248
4249 /*
4250 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4251 * we always retry
4252 */
4253 if (gfp_mask & __GFP_NOFAIL) {
4254 /*
4255 * All existing users of the __GFP_NOFAIL are blockable, so warn
4256 * of any new users that actually require GFP_NOWAIT
4257 */
4258 if (WARN_ON_ONCE(!can_direct_reclaim))
4259 goto fail;
4260
4261 /*
4262 * PF_MEMALLOC request from this context is rather bizarre
4263 * because we cannot reclaim anything and only can loop waiting
4264 * for somebody to do a work for us
4265 */
4266 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4267
4268 /*
4269 * non failing costly orders are a hard requirement which we
4270 * are not prepared for much so let's warn about these users
4271 * so that we can identify them and convert them to something
4272 * else.
4273 */
4274 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4275
4276 /*
4277 * Help non-failing allocations by giving them access to memory
4278 * reserves but do not use ALLOC_NO_WATERMARKS because this
4279 * could deplete whole memory reserves which would just make
4280 * the situation worse
4281 */
4282 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4283 if (page)
4284 goto got_pg;
4285
4286 cond_resched();
4287 goto retry;
4288 }
4289fail:
4290 warn_alloc(gfp_mask, ac->nodemask,
4291 "page allocation failure: order:%u", order);
4292got_pg:
4293 return page;
4294}
4295
4296static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4297 int preferred_nid, nodemask_t *nodemask,
4298 struct alloc_context *ac, gfp_t *alloc_mask,
4299 unsigned int *alloc_flags)
4300{
4301 ac->high_zoneidx = gfp_zone(gfp_mask);
4302 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4303 ac->nodemask = nodemask;
4304 ac->migratetype = gfpflags_to_migratetype(gfp_mask);
4305
4306 if (cpusets_enabled()) {
4307 *alloc_mask |= __GFP_HARDWALL;
4308 if (!ac->nodemask)
4309 ac->nodemask = &cpuset_current_mems_allowed;
4310 else
4311 *alloc_flags |= ALLOC_CPUSET;
4312 }
4313
4314 fs_reclaim_acquire(gfp_mask);
4315 fs_reclaim_release(gfp_mask);
4316
4317 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4318
4319 if (should_fail_alloc_page(gfp_mask, order))
4320 return false;
4321
4322 if (IS_ENABLED(CONFIG_CMA) && ac->migratetype == MIGRATE_MOVABLE)
4323 *alloc_flags |= ALLOC_CMA;
4324
4325 return true;
4326}
4327
4328/* Determine whether to spread dirty pages and what the first usable zone */
4329static inline void finalise_ac(gfp_t gfp_mask,
4330 unsigned int order, struct alloc_context *ac)
4331{
4332 /* Dirty zone balancing only done in the fast path */
4333 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4334
4335 /*
4336 * The preferred zone is used for statistics but crucially it is
4337 * also used as the starting point for the zonelist iterator. It
4338 * may get reset for allocations that ignore memory policies.
4339 */
4340 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4341 ac->high_zoneidx, ac->nodemask);
4342}
4343
4344/*
4345 * This is the 'heart' of the zoned buddy allocator.
4346 */
4347struct page *
4348__alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4349 nodemask_t *nodemask)
4350{
4351 struct page *page;
4352 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4353 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
4354 struct alloc_context ac = { };
4355
4356 gfp_mask &= gfp_allowed_mask;
4357 alloc_mask = gfp_mask;
4358 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
4359 return NULL;
4360
4361 finalise_ac(gfp_mask, order, &ac);
4362
4363 /* First allocation attempt */
4364 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
4365 if (likely(page))
4366 goto out;
4367
4368 /*
4369 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4370 * resp. GFP_NOIO which has to be inherited for all allocation requests
4371 * from a particular context which has been marked by
4372 * memalloc_no{fs,io}_{save,restore}.
4373 */
4374 alloc_mask = current_gfp_context(gfp_mask);
4375 ac.spread_dirty_pages = false;
4376
4377 /*
4378 * Restore the original nodemask if it was potentially replaced with
4379 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4380 */
4381 if (unlikely(ac.nodemask != nodemask))
4382 ac.nodemask = nodemask;
4383
4384 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
4385
4386out:
4387 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
4388 unlikely(memcg_kmem_charge(page, gfp_mask, order) != 0)) {
4389 __free_pages(page, order);
4390 page = NULL;
4391 }
4392
4393 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
4394
4395 return page;
4396}
4397EXPORT_SYMBOL(__alloc_pages_nodemask);
4398
4399/*
4400 * Common helper functions.
4401 */
4402unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4403{
4404 struct page *page;
4405
4406 /*
4407 * __get_free_pages() returns a virtual address, which cannot represent
4408 * a highmem page
4409 */
4410 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
4411
4412 page = alloc_pages(gfp_mask, order);
4413 if (!page)
4414 return 0;
4415 return (unsigned long) page_address(page);
4416}
4417EXPORT_SYMBOL(__get_free_pages);
4418
4419unsigned long get_zeroed_page(gfp_t gfp_mask)
4420{
4421 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
4422}
4423EXPORT_SYMBOL(get_zeroed_page);
4424
4425void __free_pages(struct page *page, unsigned int order)
4426{
4427 if (put_page_testzero(page)) {
4428 if (order == 0)
4429 free_unref_page(page);
4430 else
4431 __free_pages_ok(page, order);
4432 }
4433}
4434
4435EXPORT_SYMBOL(__free_pages);
4436
4437void free_pages(unsigned long addr, unsigned int order)
4438{
4439 if (addr != 0) {
4440 VM_BUG_ON(!virt_addr_valid((void *)addr));
4441 __free_pages(virt_to_page((void *)addr), order);
4442 }
4443}
4444
4445EXPORT_SYMBOL(free_pages);
4446
4447/*
4448 * Page Fragment:
4449 * An arbitrary-length arbitrary-offset area of memory which resides
4450 * within a 0 or higher order page. Multiple fragments within that page
4451 * are individually refcounted, in the page's reference counter.
4452 *
4453 * The page_frag functions below provide a simple allocation framework for
4454 * page fragments. This is used by the network stack and network device
4455 * drivers to provide a backing region of memory for use as either an
4456 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4457 */
4458static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4459 gfp_t gfp_mask)
4460{
4461 struct page *page = NULL;
4462 gfp_t gfp = gfp_mask;
4463
4464#if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4465 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
4466 __GFP_NOMEMALLOC;
4467 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4468 PAGE_FRAG_CACHE_MAX_ORDER);
4469 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4470#endif
4471 if (unlikely(!page))
4472 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4473
4474 nc->va = page ? page_address(page) : NULL;
4475
4476 return page;
4477}
4478
4479void __page_frag_cache_drain(struct page *page, unsigned int count)
4480{
4481 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4482
4483 if (page_ref_sub_and_test(page, count)) {
4484 unsigned int order = compound_order(page);
4485
4486 if (order == 0)
4487 free_unref_page(page);
4488 else
4489 __free_pages_ok(page, order);
4490 }
4491}
4492EXPORT_SYMBOL(__page_frag_cache_drain);
4493
4494void *page_frag_alloc(struct page_frag_cache *nc,
4495 unsigned int fragsz, gfp_t gfp_mask)
4496{
4497 unsigned int size = PAGE_SIZE;
4498 struct page *page;
4499 int offset;
4500
4501 if (unlikely(!nc->va)) {
4502refill:
4503 page = __page_frag_cache_refill(nc, gfp_mask);
4504 if (!page)
4505 return NULL;
4506
4507#if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4508 /* if size can vary use size else just use PAGE_SIZE */
4509 size = nc->size;
4510#endif
4511 /* Even if we own the page, we do not use atomic_set().
4512 * This would break get_page_unless_zero() users.
4513 */
4514 page_ref_add(page, size - 1);
4515
4516 /* reset page count bias and offset to start of new frag */
4517 nc->pfmemalloc = page_is_pfmemalloc(page);
4518 nc->pagecnt_bias = size;
4519 nc->offset = size;
4520 }
4521
4522 offset = nc->offset - fragsz;
4523 if (unlikely(offset < 0)) {
4524 page = virt_to_page(nc->va);
4525
4526 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
4527 goto refill;
4528
4529#if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4530 /* if size can vary use size else just use PAGE_SIZE */
4531 size = nc->size;
4532#endif
4533 /* OK, page count is 0, we can safely set it */
4534 set_page_count(page, size);
4535
4536 /* reset page count bias and offset to start of new frag */
4537 nc->pagecnt_bias = size;
4538 offset = size - fragsz;
4539 }
4540
4541 nc->pagecnt_bias--;
4542 nc->offset = offset;
4543
4544 return nc->va + offset;
4545}
4546EXPORT_SYMBOL(page_frag_alloc);
4547
4548/*
4549 * Frees a page fragment allocated out of either a compound or order 0 page.
4550 */
4551void page_frag_free(void *addr)
4552{
4553 struct page *page = virt_to_head_page(addr);
4554
4555 if (unlikely(put_page_testzero(page)))
4556 __free_pages_ok(page, compound_order(page));
4557}
4558EXPORT_SYMBOL(page_frag_free);
4559
4560static void *make_alloc_exact(unsigned long addr, unsigned int order,
4561 size_t size)
4562{
4563 if (addr) {
4564 unsigned long alloc_end = addr + (PAGE_SIZE << order);
4565 unsigned long used = addr + PAGE_ALIGN(size);
4566
4567 split_page(virt_to_page((void *)addr), order);
4568 while (used < alloc_end) {
4569 free_page(used);
4570 used += PAGE_SIZE;
4571 }
4572 }
4573 return (void *)addr;
4574}
4575
4576/**
4577 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4578 * @size: the number of bytes to allocate
4579 * @gfp_mask: GFP flags for the allocation
4580 *
4581 * This function is similar to alloc_pages(), except that it allocates the
4582 * minimum number of pages to satisfy the request. alloc_pages() can only
4583 * allocate memory in power-of-two pages.
4584 *
4585 * This function is also limited by MAX_ORDER.
4586 *
4587 * Memory allocated by this function must be released by free_pages_exact().
4588 */
4589void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
4590{
4591 unsigned int order = get_order(size);
4592 unsigned long addr;
4593
4594 addr = __get_free_pages(gfp_mask, order);
4595 return make_alloc_exact(addr, order, size);
4596}
4597EXPORT_SYMBOL(alloc_pages_exact);
4598
4599/**
4600 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4601 * pages on a node.
4602 * @nid: the preferred node ID where memory should be allocated
4603 * @size: the number of bytes to allocate
4604 * @gfp_mask: GFP flags for the allocation
4605 *
4606 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4607 * back.
4608 */
4609void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
4610{
4611 unsigned int order = get_order(size);
4612 struct page *p = alloc_pages_node(nid, gfp_mask, order);
4613 if (!p)
4614 return NULL;
4615 return make_alloc_exact((unsigned long)page_address(p), order, size);
4616}
4617
4618/**
4619 * free_pages_exact - release memory allocated via alloc_pages_exact()
4620 * @virt: the value returned by alloc_pages_exact.
4621 * @size: size of allocation, same value as passed to alloc_pages_exact().
4622 *
4623 * Release the memory allocated by a previous call to alloc_pages_exact.
4624 */
4625void free_pages_exact(void *virt, size_t size)
4626{
4627 unsigned long addr = (unsigned long)virt;
4628 unsigned long end = addr + PAGE_ALIGN(size);
4629
4630 while (addr < end) {
4631 free_page(addr);
4632 addr += PAGE_SIZE;
4633 }
4634}
4635EXPORT_SYMBOL(free_pages_exact);
4636
4637/**
4638 * nr_free_zone_pages - count number of pages beyond high watermark
4639 * @offset: The zone index of the highest zone
4640 *
4641 * nr_free_zone_pages() counts the number of counts pages which are beyond the
4642 * high watermark within all zones at or below a given zone index. For each
4643 * zone, the number of pages is calculated as:
4644 *
4645 * nr_free_zone_pages = managed_pages - high_pages
4646 */
4647static unsigned long nr_free_zone_pages(int offset)
4648{
4649 struct zoneref *z;
4650 struct zone *zone;
4651
4652 /* Just pick one node, since fallback list is circular */
4653 unsigned long sum = 0;
4654
4655 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
4656
4657 for_each_zone_zonelist(zone, z, zonelist, offset) {
4658 unsigned long size = zone->managed_pages;
4659 unsigned long high = high_wmark_pages(zone);
4660 if (size > high)
4661 sum += size - high;
4662 }
4663
4664 return sum;
4665}
4666
4667/**
4668 * nr_free_buffer_pages - count number of pages beyond high watermark
4669 *
4670 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4671 * watermark within ZONE_DMA and ZONE_NORMAL.
4672 */
4673unsigned long nr_free_buffer_pages(void)
4674{
4675 return nr_free_zone_pages(gfp_zone(GFP_USER));
4676}
4677EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
4678
4679/**
4680 * nr_free_pagecache_pages - count number of pages beyond high watermark
4681 *
4682 * nr_free_pagecache_pages() counts the number of pages which are beyond the
4683 * high watermark within all zones.
4684 */
4685unsigned long nr_free_pagecache_pages(void)
4686{
4687 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
4688}
4689
4690static inline void show_node(struct zone *zone)
4691{
4692 if (IS_ENABLED(CONFIG_NUMA))
4693 printk("Node %d ", zone_to_nid(zone));
4694}
4695
4696long si_mem_available(void)
4697{
4698 long available;
4699 unsigned long pagecache;
4700 unsigned long wmark_low = 0;
4701 unsigned long pages[NR_LRU_LISTS];
4702 struct zone *zone;
4703 int lru;
4704
4705 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
4706 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
4707
4708 for_each_zone(zone)
4709 wmark_low += zone->watermark[WMARK_LOW];
4710
4711 /*
4712 * Estimate the amount of memory available for userspace allocations,
4713 * without causing swapping.
4714 */
4715 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
4716
4717 /*
4718 * Not all the page cache can be freed, otherwise the system will
4719 * start swapping. Assume at least half of the page cache, or the
4720 * low watermark worth of cache, needs to stay.
4721 */
4722 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
4723 pagecache -= min(pagecache / 2, wmark_low);
4724 available += pagecache;
4725
4726 /*
4727 * Part of the reclaimable slab consists of items that are in use,
4728 * and cannot be freed. Cap this estimate at the low watermark.
4729 */
4730 available += global_node_page_state(NR_SLAB_RECLAIMABLE) -
4731 min(global_node_page_state(NR_SLAB_RECLAIMABLE) / 2,
4732 wmark_low);
4733
4734 /*
4735 * Part of the kernel memory, which can be released under memory
4736 * pressure.
4737 */
4738 available += global_node_page_state(NR_INDIRECTLY_RECLAIMABLE_BYTES) >>
4739 PAGE_SHIFT;
4740
4741 if (available < 0)
4742 available = 0;
4743 return available;
4744}
4745EXPORT_SYMBOL_GPL(si_mem_available);
4746
4747void si_meminfo(struct sysinfo *val)
4748{
4749 val->totalram = totalram_pages;
4750 val->sharedram = global_node_page_state(NR_SHMEM);
4751 val->freeram = global_zone_page_state(NR_FREE_PAGES);
4752 val->bufferram = nr_blockdev_pages();
4753 val->totalhigh = totalhigh_pages;
4754 val->freehigh = nr_free_highpages();
4755 val->mem_unit = PAGE_SIZE;
4756}
4757
4758EXPORT_SYMBOL(si_meminfo);
4759
4760#ifdef CONFIG_NUMA
4761void si_meminfo_node(struct sysinfo *val, int nid)
4762{
4763 int zone_type; /* needs to be signed */
4764 unsigned long managed_pages = 0;
4765 unsigned long managed_highpages = 0;
4766 unsigned long free_highpages = 0;
4767 pg_data_t *pgdat = NODE_DATA(nid);
4768
4769 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
4770 managed_pages += pgdat->node_zones[zone_type].managed_pages;
4771 val->totalram = managed_pages;
4772 val->sharedram = node_page_state(pgdat, NR_SHMEM);
4773 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
4774#ifdef CONFIG_HIGHMEM
4775 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
4776 struct zone *zone = &pgdat->node_zones[zone_type];
4777
4778 if (is_highmem(zone)) {
4779 managed_highpages += zone->managed_pages;
4780 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
4781 }
4782 }
4783 val->totalhigh = managed_highpages;
4784 val->freehigh = free_highpages;
4785#else
4786 val->totalhigh = managed_highpages;
4787 val->freehigh = free_highpages;
4788#endif
4789 val->mem_unit = PAGE_SIZE;
4790}
4791#endif
4792
4793/*
4794 * Determine whether the node should be displayed or not, depending on whether
4795 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
4796 */
4797static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
4798{
4799 if (!(flags & SHOW_MEM_FILTER_NODES))
4800 return false;
4801
4802 /*
4803 * no node mask - aka implicit memory numa policy. Do not bother with
4804 * the synchronization - read_mems_allowed_begin - because we do not
4805 * have to be precise here.
4806 */
4807 if (!nodemask)
4808 nodemask = &cpuset_current_mems_allowed;
4809
4810 return !node_isset(nid, *nodemask);
4811}
4812
4813#define K(x) ((x) << (PAGE_SHIFT-10))
4814
4815static void show_migration_types(unsigned char type)
4816{
4817 static const char types[MIGRATE_TYPES] = {
4818 [MIGRATE_UNMOVABLE] = 'U',
4819 [MIGRATE_MOVABLE] = 'M',
4820 [MIGRATE_RECLAIMABLE] = 'E',
4821 [MIGRATE_HIGHATOMIC] = 'H',
4822#ifdef CONFIG_CMA
4823 [MIGRATE_CMA] = 'C',
4824#endif
4825#ifdef CONFIG_MEMORY_ISOLATION
4826 [MIGRATE_ISOLATE] = 'I',
4827#endif
4828 };
4829 char tmp[MIGRATE_TYPES + 1];
4830 char *p = tmp;
4831 int i;
4832
4833 for (i = 0; i < MIGRATE_TYPES; i++) {
4834 if (type & (1 << i))
4835 *p++ = types[i];
4836 }
4837
4838 *p = '\0';
4839 printk(KERN_CONT "(%s) ", tmp);
4840}
4841
4842/*
4843 * Show free area list (used inside shift_scroll-lock stuff)
4844 * We also calculate the percentage fragmentation. We do this by counting the
4845 * memory on each free list with the exception of the first item on the list.
4846 *
4847 * Bits in @filter:
4848 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
4849 * cpuset.
4850 */
4851void show_free_areas(unsigned int filter, nodemask_t *nodemask)
4852{
4853 unsigned long free_pcp = 0;
4854 int cpu;
4855 struct zone *zone;
4856 pg_data_t *pgdat;
4857
4858 for_each_populated_zone(zone) {
4859 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4860 continue;
4861
4862 for_each_online_cpu(cpu)
4863 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4864 }
4865
4866 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
4867 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
4868 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
4869 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
4870 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
4871 " free:%lu free_pcp:%lu free_cma:%lu\n",
4872 global_node_page_state(NR_ACTIVE_ANON),
4873 global_node_page_state(NR_INACTIVE_ANON),
4874 global_node_page_state(NR_ISOLATED_ANON),
4875 global_node_page_state(NR_ACTIVE_FILE),
4876 global_node_page_state(NR_INACTIVE_FILE),
4877 global_node_page_state(NR_ISOLATED_FILE),
4878 global_node_page_state(NR_UNEVICTABLE),
4879 global_node_page_state(NR_FILE_DIRTY),
4880 global_node_page_state(NR_WRITEBACK),
4881 global_node_page_state(NR_UNSTABLE_NFS),
4882 global_node_page_state(NR_SLAB_RECLAIMABLE),
4883 global_node_page_state(NR_SLAB_UNRECLAIMABLE),
4884 global_node_page_state(NR_FILE_MAPPED),
4885 global_node_page_state(NR_SHMEM),
4886 global_zone_page_state(NR_PAGETABLE),
4887 global_zone_page_state(NR_BOUNCE),
4888 global_zone_page_state(NR_FREE_PAGES),
4889 free_pcp,
4890 global_zone_page_state(NR_FREE_CMA_PAGES));
4891
4892 for_each_online_pgdat(pgdat) {
4893 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
4894 continue;
4895
4896 printk("Node %d"
4897 " active_anon:%lukB"
4898 " inactive_anon:%lukB"
4899 " active_file:%lukB"
4900 " inactive_file:%lukB"
4901 " unevictable:%lukB"
4902 " isolated(anon):%lukB"
4903 " isolated(file):%lukB"
4904 " mapped:%lukB"
4905 " dirty:%lukB"
4906 " writeback:%lukB"
4907 " shmem:%lukB"
4908#ifdef CONFIG_TRANSPARENT_HUGEPAGE
4909 " shmem_thp: %lukB"
4910 " shmem_pmdmapped: %lukB"
4911 " anon_thp: %lukB"
4912#endif
4913 " writeback_tmp:%lukB"
4914 " unstable:%lukB"
4915 " all_unreclaimable? %s"
4916 "\n",
4917 pgdat->node_id,
4918 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
4919 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
4920 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
4921 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
4922 K(node_page_state(pgdat, NR_UNEVICTABLE)),
4923 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
4924 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
4925 K(node_page_state(pgdat, NR_FILE_MAPPED)),
4926 K(node_page_state(pgdat, NR_FILE_DIRTY)),
4927 K(node_page_state(pgdat, NR_WRITEBACK)),
4928 K(node_page_state(pgdat, NR_SHMEM)),
4929#ifdef CONFIG_TRANSPARENT_HUGEPAGE
4930 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
4931 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
4932 * HPAGE_PMD_NR),
4933 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
4934#endif
4935 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
4936 K(node_page_state(pgdat, NR_UNSTABLE_NFS)),
4937 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
4938 "yes" : "no");
4939 }
4940
4941 for_each_populated_zone(zone) {
4942 int i;
4943
4944 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4945 continue;
4946
4947 free_pcp = 0;
4948 for_each_online_cpu(cpu)
4949 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4950
4951 show_node(zone);
4952 printk(KERN_CONT
4953 "%s"
4954 " free:%lukB"
4955 " min:%lukB"
4956 " low:%lukB"
4957 " high:%lukB"
4958 " active_anon:%lukB"
4959 " inactive_anon:%lukB"
4960 " active_file:%lukB"
4961 " inactive_file:%lukB"
4962 " unevictable:%lukB"
4963 " writepending:%lukB"
4964 " present:%lukB"
4965 " managed:%lukB"
4966 " mlocked:%lukB"
4967 " kernel_stack:%lukB"
4968 " pagetables:%lukB"
4969 " bounce:%lukB"
4970 " free_pcp:%lukB"
4971 " local_pcp:%ukB"
4972 " free_cma:%lukB"
4973 "\n",
4974 zone->name,
4975 K(zone_page_state(zone, NR_FREE_PAGES)),
4976 K(min_wmark_pages(zone)),
4977 K(low_wmark_pages(zone)),
4978 K(high_wmark_pages(zone)),
4979 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
4980 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
4981 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
4982 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
4983 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
4984 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
4985 K(zone->present_pages),
4986 K(zone->managed_pages),
4987 K(zone_page_state(zone, NR_MLOCK)),
4988 zone_page_state(zone, NR_KERNEL_STACK_KB),
4989 K(zone_page_state(zone, NR_PAGETABLE)),
4990 K(zone_page_state(zone, NR_BOUNCE)),
4991 K(free_pcp),
4992 K(this_cpu_read(zone->pageset->pcp.count)),
4993 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
4994 printk("lowmem_reserve[]:");
4995 for (i = 0; i < MAX_NR_ZONES; i++)
4996 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
4997 printk(KERN_CONT "\n");
4998 }
4999
5000 for_each_populated_zone(zone) {
5001 unsigned int order;
5002 unsigned long nr[MAX_ORDER], flags, total = 0;
5003 unsigned char types[MAX_ORDER];
5004
5005 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5006 continue;
5007 show_node(zone);
5008 printk(KERN_CONT "%s: ", zone->name);
5009
5010 spin_lock_irqsave(&zone->lock, flags);
5011 for (order = 0; order < MAX_ORDER; order++) {
5012 struct free_area *area = &zone->free_area[order];
5013 int type;
5014
5015 nr[order] = area->nr_free;
5016 total += nr[order] << order;
5017
5018 types[order] = 0;
5019 for (type = 0; type < MIGRATE_TYPES; type++) {
5020 if (!list_empty(&area->free_list[type]))
5021 types[order] |= 1 << type;
5022 }
5023 }
5024 spin_unlock_irqrestore(&zone->lock, flags);
5025 for (order = 0; order < MAX_ORDER; order++) {
5026 printk(KERN_CONT "%lu*%lukB ",
5027 nr[order], K(1UL) << order);
5028 if (nr[order])
5029 show_migration_types(types[order]);
5030 }
5031 printk(KERN_CONT "= %lukB\n", K(total));
5032 }
5033
5034 hugetlb_show_meminfo();
5035
5036 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
5037
5038 show_swap_cache_info();
5039}
5040
5041static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5042{
5043 zoneref->zone = zone;
5044 zoneref->zone_idx = zone_idx(zone);
5045}
5046
5047/*
5048 * Builds allocation fallback zone lists.
5049 *
5050 * Add all populated zones of a node to the zonelist.
5051 */
5052static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5053{
5054 struct zone *zone;
5055 enum zone_type zone_type = MAX_NR_ZONES;
5056 int nr_zones = 0;
5057
5058 do {
5059 zone_type--;
5060 zone = pgdat->node_zones + zone_type;
5061 if (managed_zone(zone)) {
5062 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5063 check_highest_zone(zone_type);
5064 }
5065 } while (zone_type);
5066
5067 return nr_zones;
5068}
5069
5070#ifdef CONFIG_NUMA
5071
5072static int __parse_numa_zonelist_order(char *s)
5073{
5074 /*
5075 * We used to support different zonlists modes but they turned
5076 * out to be just not useful. Let's keep the warning in place
5077 * if somebody still use the cmd line parameter so that we do
5078 * not fail it silently
5079 */
5080 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5081 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5082 return -EINVAL;
5083 }
5084 return 0;
5085}
5086
5087static __init int setup_numa_zonelist_order(char *s)
5088{
5089 if (!s)
5090 return 0;
5091
5092 return __parse_numa_zonelist_order(s);
5093}
5094early_param("numa_zonelist_order", setup_numa_zonelist_order);
5095
5096char numa_zonelist_order[] = "Node";
5097
5098/*
5099 * sysctl handler for numa_zonelist_order
5100 */
5101int numa_zonelist_order_handler(struct ctl_table *table, int write,
5102 void __user *buffer, size_t *length,
5103 loff_t *ppos)
5104{
5105 char *str;
5106 int ret;
5107
5108 if (!write)
5109 return proc_dostring(table, write, buffer, length, ppos);
5110 str = memdup_user_nul(buffer, 16);
5111 if (IS_ERR(str))
5112 return PTR_ERR(str);
5113
5114 ret = __parse_numa_zonelist_order(str);
5115 kfree(str);
5116 return ret;
5117}
5118
5119
5120#define MAX_NODE_LOAD (nr_online_nodes)
5121static int node_load[MAX_NUMNODES];
5122
5123/**
5124 * find_next_best_node - find the next node that should appear in a given node's fallback list
5125 * @node: node whose fallback list we're appending
5126 * @used_node_mask: nodemask_t of already used nodes
5127 *
5128 * We use a number of factors to determine which is the next node that should
5129 * appear on a given node's fallback list. The node should not have appeared
5130 * already in @node's fallback list, and it should be the next closest node
5131 * according to the distance array (which contains arbitrary distance values
5132 * from each node to each node in the system), and should also prefer nodes
5133 * with no CPUs, since presumably they'll have very little allocation pressure
5134 * on them otherwise.
5135 * It returns -1 if no node is found.
5136 */
5137static int find_next_best_node(int node, nodemask_t *used_node_mask)
5138{
5139 int n, val;
5140 int min_val = INT_MAX;
5141 int best_node = NUMA_NO_NODE;
5142 const struct cpumask *tmp = cpumask_of_node(0);
5143
5144 /* Use the local node if we haven't already */
5145 if (!node_isset(node, *used_node_mask)) {
5146 node_set(node, *used_node_mask);
5147 return node;
5148 }
5149
5150 for_each_node_state(n, N_MEMORY) {
5151
5152 /* Don't want a node to appear more than once */
5153 if (node_isset(n, *used_node_mask))
5154 continue;
5155
5156 /* Use the distance array to find the distance */
5157 val = node_distance(node, n);
5158
5159 /* Penalize nodes under us ("prefer the next node") */
5160 val += (n < node);
5161
5162 /* Give preference to headless and unused nodes */
5163 tmp = cpumask_of_node(n);
5164 if (!cpumask_empty(tmp))
5165 val += PENALTY_FOR_NODE_WITH_CPUS;
5166
5167 /* Slight preference for less loaded node */
5168 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
5169 val += node_load[n];
5170
5171 if (val < min_val) {
5172 min_val = val;
5173 best_node = n;
5174 }
5175 }
5176
5177 if (best_node >= 0)
5178 node_set(best_node, *used_node_mask);
5179
5180 return best_node;
5181}
5182
5183
5184/*
5185 * Build zonelists ordered by node and zones within node.
5186 * This results in maximum locality--normal zone overflows into local
5187 * DMA zone, if any--but risks exhausting DMA zone.
5188 */
5189static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5190 unsigned nr_nodes)
5191{
5192 struct zoneref *zonerefs;
5193 int i;
5194
5195 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5196
5197 for (i = 0; i < nr_nodes; i++) {
5198 int nr_zones;
5199
5200 pg_data_t *node = NODE_DATA(node_order[i]);
5201
5202 nr_zones = build_zonerefs_node(node, zonerefs);
5203 zonerefs += nr_zones;
5204 }
5205 zonerefs->zone = NULL;
5206 zonerefs->zone_idx = 0;
5207}
5208
5209/*
5210 * Build gfp_thisnode zonelists
5211 */
5212static void build_thisnode_zonelists(pg_data_t *pgdat)
5213{
5214 struct zoneref *zonerefs;
5215 int nr_zones;
5216
5217 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5218 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5219 zonerefs += nr_zones;
5220 zonerefs->zone = NULL;
5221 zonerefs->zone_idx = 0;
5222}
5223
5224/*
5225 * Build zonelists ordered by zone and nodes within zones.
5226 * This results in conserving DMA zone[s] until all Normal memory is
5227 * exhausted, but results in overflowing to remote node while memory
5228 * may still exist in local DMA zone.
5229 */
5230
5231static void build_zonelists(pg_data_t *pgdat)
5232{
5233 static int node_order[MAX_NUMNODES];
5234 int node, load, nr_nodes = 0;
5235 nodemask_t used_mask;
5236 int local_node, prev_node;
5237
5238 /* NUMA-aware ordering of nodes */
5239 local_node = pgdat->node_id;
5240 load = nr_online_nodes;
5241 prev_node = local_node;
5242 nodes_clear(used_mask);
5243
5244 memset(node_order, 0, sizeof(node_order));
5245 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5246 /*
5247 * We don't want to pressure a particular node.
5248 * So adding penalty to the first node in same
5249 * distance group to make it round-robin.
5250 */
5251 if (node_distance(local_node, node) !=
5252 node_distance(local_node, prev_node))
5253 node_load[node] = load;
5254
5255 node_order[nr_nodes++] = node;
5256 prev_node = node;
5257 load--;
5258 }
5259
5260 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5261 build_thisnode_zonelists(pgdat);
5262}
5263
5264#ifdef CONFIG_HAVE_MEMORYLESS_NODES
5265/*
5266 * Return node id of node used for "local" allocations.
5267 * I.e., first node id of first zone in arg node's generic zonelist.
5268 * Used for initializing percpu 'numa_mem', which is used primarily
5269 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5270 */
5271int local_memory_node(int node)
5272{
5273 struct zoneref *z;
5274
5275 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5276 gfp_zone(GFP_KERNEL),
5277 NULL);
5278 return z->zone->node;
5279}
5280#endif
5281
5282static void setup_min_unmapped_ratio(void);
5283static void setup_min_slab_ratio(void);
5284#else /* CONFIG_NUMA */
5285
5286static void build_zonelists(pg_data_t *pgdat)
5287{
5288 int node, local_node;
5289 struct zoneref *zonerefs;
5290 int nr_zones;
5291
5292 local_node = pgdat->node_id;
5293
5294 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5295 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5296 zonerefs += nr_zones;
5297
5298 /*
5299 * Now we build the zonelist so that it contains the zones
5300 * of all the other nodes.
5301 * We don't want to pressure a particular node, so when
5302 * building the zones for node N, we make sure that the
5303 * zones coming right after the local ones are those from
5304 * node N+1 (modulo N)
5305 */
5306 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5307 if (!node_online(node))
5308 continue;
5309 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5310 zonerefs += nr_zones;
5311 }
5312 for (node = 0; node < local_node; node++) {
5313 if (!node_online(node))
5314 continue;
5315 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5316 zonerefs += nr_zones;
5317 }
5318
5319 zonerefs->zone = NULL;
5320 zonerefs->zone_idx = 0;
5321}
5322
5323#endif /* CONFIG_NUMA */
5324
5325/*
5326 * Boot pageset table. One per cpu which is going to be used for all
5327 * zones and all nodes. The parameters will be set in such a way
5328 * that an item put on a list will immediately be handed over to
5329 * the buddy list. This is safe since pageset manipulation is done
5330 * with interrupts disabled.
5331 *
5332 * The boot_pagesets must be kept even after bootup is complete for
5333 * unused processors and/or zones. They do play a role for bootstrapping
5334 * hotplugged processors.
5335 *
5336 * zoneinfo_show() and maybe other functions do
5337 * not check if the processor is online before following the pageset pointer.
5338 * Other parts of the kernel may not check if the zone is available.
5339 */
5340static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
5341static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5342static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5343
5344static void __build_all_zonelists(void *data)
5345{
5346 int nid;
5347 int __maybe_unused cpu;
5348 pg_data_t *self = data;
5349 static DEFINE_SPINLOCK(lock);
5350
5351 spin_lock(&lock);
5352
5353#ifdef CONFIG_NUMA
5354 memset(node_load, 0, sizeof(node_load));
5355#endif
5356
5357 /*
5358 * This node is hotadded and no memory is yet present. So just
5359 * building zonelists is fine - no need to touch other nodes.
5360 */
5361 if (self && !node_online(self->node_id)) {
5362 build_zonelists(self);
5363 } else {
5364 for_each_online_node(nid) {
5365 pg_data_t *pgdat = NODE_DATA(nid);
5366
5367 build_zonelists(pgdat);
5368 }
5369
5370#ifdef CONFIG_HAVE_MEMORYLESS_NODES
5371 /*
5372 * We now know the "local memory node" for each node--
5373 * i.e., the node of the first zone in the generic zonelist.
5374 * Set up numa_mem percpu variable for on-line cpus. During
5375 * boot, only the boot cpu should be on-line; we'll init the
5376 * secondary cpus' numa_mem as they come on-line. During
5377 * node/memory hotplug, we'll fixup all on-line cpus.
5378 */
5379 for_each_online_cpu(cpu)
5380 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5381#endif
5382 }
5383
5384 spin_unlock(&lock);
5385}
5386
5387static noinline void __init
5388build_all_zonelists_init(void)
5389{
5390 int cpu;
5391
5392 __build_all_zonelists(NULL);
5393
5394 /*
5395 * Initialize the boot_pagesets that are going to be used
5396 * for bootstrapping processors. The real pagesets for
5397 * each zone will be allocated later when the per cpu
5398 * allocator is available.
5399 *
5400 * boot_pagesets are used also for bootstrapping offline
5401 * cpus if the system is already booted because the pagesets
5402 * are needed to initialize allocators on a specific cpu too.
5403 * F.e. the percpu allocator needs the page allocator which
5404 * needs the percpu allocator in order to allocate its pagesets
5405 * (a chicken-egg dilemma).
5406 */
5407 for_each_possible_cpu(cpu)
5408 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
5409
5410 mminit_verify_zonelist();
5411 cpuset_init_current_mems_allowed();
5412}
5413
5414/*
5415 * unless system_state == SYSTEM_BOOTING.
5416 *
5417 * __ref due to call of __init annotated helper build_all_zonelists_init
5418 * [protected by SYSTEM_BOOTING].
5419 */
5420void __ref build_all_zonelists(pg_data_t *pgdat)
5421{
5422 if (system_state == SYSTEM_BOOTING) {
5423 build_all_zonelists_init();
5424 } else {
5425 __build_all_zonelists(pgdat);
5426 /* cpuset refresh routine should be here */
5427 }
5428 vm_total_pages = nr_free_pagecache_pages();
5429 /*
5430 * Disable grouping by mobility if the number of pages in the
5431 * system is too low to allow the mechanism to work. It would be
5432 * more accurate, but expensive to check per-zone. This check is
5433 * made on memory-hotadd so a system can start with mobility
5434 * disabled and enable it later
5435 */
5436 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5437 page_group_by_mobility_disabled = 1;
5438 else
5439 page_group_by_mobility_disabled = 0;
5440
5441 pr_info("Built %i zonelists, mobility grouping %s. Total pages: %ld\n",
5442 nr_online_nodes,
5443 page_group_by_mobility_disabled ? "off" : "on",
5444 vm_total_pages);
5445#ifdef CONFIG_NUMA
5446 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5447#endif
5448}
5449
5450/*
5451 * Initially all pages are reserved - free ones are freed
5452 * up by free_all_bootmem() once the early boot process is
5453 * done. Non-atomic initialization, single-pass.
5454 */
5455void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5456 unsigned long start_pfn, enum memmap_context context,
5457 struct vmem_altmap *altmap)
5458{
5459 unsigned long end_pfn = start_pfn + size;
5460 pg_data_t *pgdat = NODE_DATA(nid);
5461 unsigned long pfn;
5462 unsigned long nr_initialised = 0;
5463 struct page *page;
5464#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5465 struct memblock_region *r = NULL, *tmp;
5466#endif
5467
5468 if (highest_memmap_pfn < end_pfn - 1)
5469 highest_memmap_pfn = end_pfn - 1;
5470
5471 /*
5472 * Honor reservation requested by the driver for this ZONE_DEVICE
5473 * memory
5474 */
5475 if (altmap && start_pfn == altmap->base_pfn)
5476 start_pfn += altmap->reserve;
5477
5478 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5479 /*
5480 * There can be holes in boot-time mem_map[]s handed to this
5481 * function. They do not exist on hotplugged memory.
5482 */
5483 if (context != MEMMAP_EARLY)
5484 goto not_early;
5485
5486 if (!early_pfn_valid(pfn))
5487 continue;
5488 if (!early_pfn_in_nid(pfn, nid))
5489 continue;
5490 if (!update_defer_init(pgdat, pfn, end_pfn, &nr_initialised))
5491 break;
5492
5493#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5494 /*
5495 * Check given memblock attribute by firmware which can affect
5496 * kernel memory layout. If zone==ZONE_MOVABLE but memory is
5497 * mirrored, it's an overlapped memmap init. skip it.
5498 */
5499 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5500 if (!r || pfn >= memblock_region_memory_end_pfn(r)) {
5501 for_each_memblock(memory, tmp)
5502 if (pfn < memblock_region_memory_end_pfn(tmp))
5503 break;
5504 r = tmp;
5505 }
5506 if (pfn >= memblock_region_memory_base_pfn(r) &&
5507 memblock_is_mirror(r)) {
5508 /* already initialized as NORMAL */
5509 pfn = memblock_region_memory_end_pfn(r);
5510 continue;
5511 }
5512 }
5513#endif
5514
5515not_early:
5516 page = pfn_to_page(pfn);
5517 __init_single_page(page, pfn, zone, nid);
5518 if (context == MEMMAP_HOTPLUG)
5519 SetPageReserved(page);
5520
5521 /*
5522 * Mark the block movable so that blocks are reserved for
5523 * movable at startup. This will force kernel allocations
5524 * to reserve their blocks rather than leaking throughout
5525 * the address space during boot when many long-lived
5526 * kernel allocations are made.
5527 *
5528 * bitmap is created for zone's valid pfn range. but memmap
5529 * can be created for invalid pages (for alignment)
5530 * check here not to call set_pageblock_migratetype() against
5531 * pfn out of zone.
5532 *
5533 * Please note that MEMMAP_HOTPLUG path doesn't clear memmap
5534 * because this is done early in sparse_add_one_section
5535 */
5536 if (!(pfn & (pageblock_nr_pages - 1))) {
5537 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5538 cond_resched();
5539 }
5540 }
5541}
5542
5543static void __meminit zone_init_free_lists(struct zone *zone)
5544{
5545 unsigned int order, t;
5546 for_each_migratetype_order(order, t) {
5547 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
5548 zone->free_area[order].nr_free = 0;
5549 }
5550}
5551
5552#ifndef __HAVE_ARCH_MEMMAP_INIT
5553#define memmap_init(size, nid, zone, start_pfn) \
5554 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY, NULL)
5555#endif
5556
5557static int zone_batchsize(struct zone *zone)
5558{
5559#ifdef CONFIG_MMU
5560 int batch;
5561
5562 /*
5563 * The per-cpu-pages pools are set to around 1000th of the
5564 * size of the zone. But no more than 1/2 of a meg.
5565 *
5566 * OK, so we don't know how big the cache is. So guess.
5567 */
5568 batch = zone->managed_pages / 1024;
5569 if (batch * PAGE_SIZE > 512 * 1024)
5570 batch = (512 * 1024) / PAGE_SIZE;
5571 batch /= 4; /* We effectively *= 4 below */
5572 if (batch < 1)
5573 batch = 1;
5574
5575 /*
5576 * Clamp the batch to a 2^n - 1 value. Having a power
5577 * of 2 value was found to be more likely to have
5578 * suboptimal cache aliasing properties in some cases.
5579 *
5580 * For example if 2 tasks are alternately allocating
5581 * batches of pages, one task can end up with a lot
5582 * of pages of one half of the possible page colors
5583 * and the other with pages of the other colors.
5584 */
5585 batch = rounddown_pow_of_two(batch + batch/2) - 1;
5586
5587 return batch;
5588
5589#else
5590 /* The deferral and batching of frees should be suppressed under NOMMU
5591 * conditions.
5592 *
5593 * The problem is that NOMMU needs to be able to allocate large chunks
5594 * of contiguous memory as there's no hardware page translation to
5595 * assemble apparent contiguous memory from discontiguous pages.
5596 *
5597 * Queueing large contiguous runs of pages for batching, however,
5598 * causes the pages to actually be freed in smaller chunks. As there
5599 * can be a significant delay between the individual batches being
5600 * recycled, this leads to the once large chunks of space being
5601 * fragmented and becoming unavailable for high-order allocations.
5602 */
5603 return 0;
5604#endif
5605}
5606
5607/*
5608 * pcp->high and pcp->batch values are related and dependent on one another:
5609 * ->batch must never be higher then ->high.
5610 * The following function updates them in a safe manner without read side
5611 * locking.
5612 *
5613 * Any new users of pcp->batch and pcp->high should ensure they can cope with
5614 * those fields changing asynchronously (acording the the above rule).
5615 *
5616 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5617 * outside of boot time (or some other assurance that no concurrent updaters
5618 * exist).
5619 */
5620static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
5621 unsigned long batch)
5622{
5623 /* start with a fail safe value for batch */
5624 pcp->batch = 1;
5625 smp_wmb();
5626
5627 /* Update high, then batch, in order */
5628 pcp->high = high;
5629 smp_wmb();
5630
5631 pcp->batch = batch;
5632}
5633
5634/* a companion to pageset_set_high() */
5635static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
5636{
5637 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
5638}
5639
5640static void pageset_init(struct per_cpu_pageset *p)
5641{
5642 struct per_cpu_pages *pcp;
5643 int migratetype;
5644
5645 memset(p, 0, sizeof(*p));
5646
5647 pcp = &p->pcp;
5648 pcp->count = 0;
5649 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
5650 INIT_LIST_HEAD(&pcp->lists[migratetype]);
5651}
5652
5653static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
5654{
5655 pageset_init(p);
5656 pageset_set_batch(p, batch);
5657}
5658
5659/*
5660 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5661 * to the value high for the pageset p.
5662 */
5663static void pageset_set_high(struct per_cpu_pageset *p,
5664 unsigned long high)
5665{
5666 unsigned long batch = max(1UL, high / 4);
5667 if ((high / 4) > (PAGE_SHIFT * 8))
5668 batch = PAGE_SHIFT * 8;
5669
5670 pageset_update(&p->pcp, high, batch);
5671}
5672
5673static void pageset_set_high_and_batch(struct zone *zone,
5674 struct per_cpu_pageset *pcp)
5675{
5676 if (percpu_pagelist_fraction)
5677 pageset_set_high(pcp,
5678 (zone->managed_pages /
5679 percpu_pagelist_fraction));
5680 else
5681 pageset_set_batch(pcp, zone_batchsize(zone));
5682}
5683
5684static void __meminit zone_pageset_init(struct zone *zone, int cpu)
5685{
5686 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
5687
5688 pageset_init(pcp);
5689 pageset_set_high_and_batch(zone, pcp);
5690}
5691
5692void __meminit setup_zone_pageset(struct zone *zone)
5693{
5694 int cpu;
5695 zone->pageset = alloc_percpu(struct per_cpu_pageset);
5696 for_each_possible_cpu(cpu)
5697 zone_pageset_init(zone, cpu);
5698}
5699
5700/*
5701 * Allocate per cpu pagesets and initialize them.
5702 * Before this call only boot pagesets were available.
5703 */
5704void __init setup_per_cpu_pageset(void)
5705{
5706 struct pglist_data *pgdat;
5707 struct zone *zone;
5708
5709 for_each_populated_zone(zone)
5710 setup_zone_pageset(zone);
5711
5712 for_each_online_pgdat(pgdat)
5713 pgdat->per_cpu_nodestats =
5714 alloc_percpu(struct per_cpu_nodestat);
5715}
5716
5717static __meminit void zone_pcp_init(struct zone *zone)
5718{
5719 /*
5720 * per cpu subsystem is not up at this point. The following code
5721 * relies on the ability of the linker to provide the
5722 * offset of a (static) per cpu variable into the per cpu area.
5723 */
5724 zone->pageset = &boot_pageset;
5725
5726 if (populated_zone(zone))
5727 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
5728 zone->name, zone->present_pages,
5729 zone_batchsize(zone));
5730}
5731
5732void __meminit init_currently_empty_zone(struct zone *zone,
5733 unsigned long zone_start_pfn,
5734 unsigned long size)
5735{
5736 struct pglist_data *pgdat = zone->zone_pgdat;
5737
5738 pgdat->nr_zones = zone_idx(zone) + 1;
5739
5740 zone->zone_start_pfn = zone_start_pfn;
5741
5742 mminit_dprintk(MMINIT_TRACE, "memmap_init",
5743 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5744 pgdat->node_id,
5745 (unsigned long)zone_idx(zone),
5746 zone_start_pfn, (zone_start_pfn + size));
5747
5748 zone_init_free_lists(zone);
5749 zone->initialized = 1;
5750}
5751
5752#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5753#ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5754
5755/*
5756 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5757 */
5758int __meminit __early_pfn_to_nid(unsigned long pfn,
5759 struct mminit_pfnnid_cache *state)
5760{
5761 unsigned long start_pfn, end_pfn;
5762 int nid;
5763
5764 if (state->last_start <= pfn && pfn < state->last_end)
5765 return state->last_nid;
5766
5767 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
5768 if (nid != -1) {
5769 state->last_start = start_pfn;
5770 state->last_end = end_pfn;
5771 state->last_nid = nid;
5772 }
5773
5774 return nid;
5775}
5776#endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5777
5778/**
5779 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5780 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5781 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5782 *
5783 * If an architecture guarantees that all ranges registered contain no holes
5784 * and may be freed, this this function may be used instead of calling
5785 * memblock_free_early_nid() manually.
5786 */
5787void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
5788{
5789 unsigned long start_pfn, end_pfn;
5790 int i, this_nid;
5791
5792 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
5793 start_pfn = min(start_pfn, max_low_pfn);
5794 end_pfn = min(end_pfn, max_low_pfn);
5795
5796 if (start_pfn < end_pfn)
5797 memblock_free_early_nid(PFN_PHYS(start_pfn),
5798 (end_pfn - start_pfn) << PAGE_SHIFT,
5799 this_nid);
5800 }
5801}
5802
5803/**
5804 * sparse_memory_present_with_active_regions - Call memory_present for each active range
5805 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
5806 *
5807 * If an architecture guarantees that all ranges registered contain no holes and may
5808 * be freed, this function may be used instead of calling memory_present() manually.
5809 */
5810void __init sparse_memory_present_with_active_regions(int nid)
5811{
5812 unsigned long start_pfn, end_pfn;
5813 int i, this_nid;
5814
5815 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
5816 memory_present(this_nid, start_pfn, end_pfn);
5817}
5818
5819/**
5820 * get_pfn_range_for_nid - Return the start and end page frames for a node
5821 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5822 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5823 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5824 *
5825 * It returns the start and end page frame of a node based on information
5826 * provided by memblock_set_node(). If called for a node
5827 * with no available memory, a warning is printed and the start and end
5828 * PFNs will be 0.
5829 */
5830void __meminit get_pfn_range_for_nid(unsigned int nid,
5831 unsigned long *start_pfn, unsigned long *end_pfn)
5832{
5833 unsigned long this_start_pfn, this_end_pfn;
5834 int i;
5835
5836 *start_pfn = -1UL;
5837 *end_pfn = 0;
5838
5839 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
5840 *start_pfn = min(*start_pfn, this_start_pfn);
5841 *end_pfn = max(*end_pfn, this_end_pfn);
5842 }
5843
5844 if (*start_pfn == -1UL)
5845 *start_pfn = 0;
5846}
5847
5848/*
5849 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5850 * assumption is made that zones within a node are ordered in monotonic
5851 * increasing memory addresses so that the "highest" populated zone is used
5852 */
5853static void __init find_usable_zone_for_movable(void)
5854{
5855 int zone_index;
5856 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
5857 if (zone_index == ZONE_MOVABLE)
5858 continue;
5859
5860 if (arch_zone_highest_possible_pfn[zone_index] >
5861 arch_zone_lowest_possible_pfn[zone_index])
5862 break;
5863 }
5864
5865 VM_BUG_ON(zone_index == -1);
5866 movable_zone = zone_index;
5867}
5868
5869/*
5870 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5871 * because it is sized independent of architecture. Unlike the other zones,
5872 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5873 * in each node depending on the size of each node and how evenly kernelcore
5874 * is distributed. This helper function adjusts the zone ranges
5875 * provided by the architecture for a given node by using the end of the
5876 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5877 * zones within a node are in order of monotonic increases memory addresses
5878 */
5879static void __meminit adjust_zone_range_for_zone_movable(int nid,
5880 unsigned long zone_type,
5881 unsigned long node_start_pfn,
5882 unsigned long node_end_pfn,
5883 unsigned long *zone_start_pfn,
5884 unsigned long *zone_end_pfn)
5885{
5886 /* Only adjust if ZONE_MOVABLE is on this node */
5887 if (zone_movable_pfn[nid]) {
5888 /* Size ZONE_MOVABLE */
5889 if (zone_type == ZONE_MOVABLE) {
5890 *zone_start_pfn = zone_movable_pfn[nid];
5891 *zone_end_pfn = min(node_end_pfn,
5892 arch_zone_highest_possible_pfn[movable_zone]);
5893
5894 /* Adjust for ZONE_MOVABLE starting within this range */
5895 } else if (!mirrored_kernelcore &&
5896 *zone_start_pfn < zone_movable_pfn[nid] &&
5897 *zone_end_pfn > zone_movable_pfn[nid]) {
5898 *zone_end_pfn = zone_movable_pfn[nid];
5899
5900 /* Check if this whole range is within ZONE_MOVABLE */
5901 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
5902 *zone_start_pfn = *zone_end_pfn;
5903 }
5904}
5905
5906/*
5907 * Return the number of pages a zone spans in a node, including holes
5908 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5909 */
5910static unsigned long __meminit zone_spanned_pages_in_node(int nid,
5911 unsigned long zone_type,
5912 unsigned long node_start_pfn,
5913 unsigned long node_end_pfn,
5914 unsigned long *zone_start_pfn,
5915 unsigned long *zone_end_pfn,
5916 unsigned long *ignored)
5917{
5918 /* When hotadd a new node from cpu_up(), the node should be empty */
5919 if (!node_start_pfn && !node_end_pfn)
5920 return 0;
5921
5922 /* Get the start and end of the zone */
5923 *zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
5924 *zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
5925 adjust_zone_range_for_zone_movable(nid, zone_type,
5926 node_start_pfn, node_end_pfn,
5927 zone_start_pfn, zone_end_pfn);
5928
5929 /* Check that this node has pages within the zone's required range */
5930 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
5931 return 0;
5932
5933 /* Move the zone boundaries inside the node if necessary */
5934 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
5935 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
5936
5937 /* Return the spanned pages */
5938 return *zone_end_pfn - *zone_start_pfn;
5939}
5940
5941/*
5942 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5943 * then all holes in the requested range will be accounted for.
5944 */
5945unsigned long __meminit __absent_pages_in_range(int nid,
5946 unsigned long range_start_pfn,
5947 unsigned long range_end_pfn)
5948{
5949 unsigned long nr_absent = range_end_pfn - range_start_pfn;
5950 unsigned long start_pfn, end_pfn;
5951 int i;
5952
5953 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5954 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
5955 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
5956 nr_absent -= end_pfn - start_pfn;
5957 }
5958 return nr_absent;
5959}
5960
5961/**
5962 * absent_pages_in_range - Return number of page frames in holes within a range
5963 * @start_pfn: The start PFN to start searching for holes
5964 * @end_pfn: The end PFN to stop searching for holes
5965 *
5966 * It returns the number of pages frames in memory holes within a range.
5967 */
5968unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5969 unsigned long end_pfn)
5970{
5971 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5972}
5973
5974/* Return the number of page frames in holes in a zone on a node */
5975static unsigned long __meminit zone_absent_pages_in_node(int nid,
5976 unsigned long zone_type,
5977 unsigned long node_start_pfn,
5978 unsigned long node_end_pfn,
5979 unsigned long *ignored)
5980{
5981 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5982 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5983 unsigned long zone_start_pfn, zone_end_pfn;
5984 unsigned long nr_absent;
5985
5986 /* When hotadd a new node from cpu_up(), the node should be empty */
5987 if (!node_start_pfn && !node_end_pfn)
5988 return 0;
5989
5990 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5991 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5992
5993 adjust_zone_range_for_zone_movable(nid, zone_type,
5994 node_start_pfn, node_end_pfn,
5995 &zone_start_pfn, &zone_end_pfn);
5996 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5997
5998 /*
5999 * ZONE_MOVABLE handling.
6000 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6001 * and vice versa.
6002 */
6003 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
6004 unsigned long start_pfn, end_pfn;
6005 struct memblock_region *r;
6006
6007 for_each_memblock(memory, r) {
6008 start_pfn = clamp(memblock_region_memory_base_pfn(r),
6009 zone_start_pfn, zone_end_pfn);
6010 end_pfn = clamp(memblock_region_memory_end_pfn(r),
6011 zone_start_pfn, zone_end_pfn);
6012
6013 if (zone_type == ZONE_MOVABLE &&
6014 memblock_is_mirror(r))
6015 nr_absent += end_pfn - start_pfn;
6016
6017 if (zone_type == ZONE_NORMAL &&
6018 !memblock_is_mirror(r))
6019 nr_absent += end_pfn - start_pfn;
6020 }
6021 }
6022
6023 return nr_absent;
6024}
6025
6026#else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6027static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
6028 unsigned long zone_type,
6029 unsigned long node_start_pfn,
6030 unsigned long node_end_pfn,
6031 unsigned long *zone_start_pfn,
6032 unsigned long *zone_end_pfn,
6033 unsigned long *zones_size)
6034{
6035 unsigned int zone;
6036
6037 *zone_start_pfn = node_start_pfn;
6038 for (zone = 0; zone < zone_type; zone++)
6039 *zone_start_pfn += zones_size[zone];
6040
6041 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
6042
6043 return zones_size[zone_type];
6044}
6045
6046static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
6047 unsigned long zone_type,
6048 unsigned long node_start_pfn,
6049 unsigned long node_end_pfn,
6050 unsigned long *zholes_size)
6051{
6052 if (!zholes_size)
6053 return 0;
6054
6055 return zholes_size[zone_type];
6056}
6057
6058#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6059
6060static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
6061 unsigned long node_start_pfn,
6062 unsigned long node_end_pfn,
6063 unsigned long *zones_size,
6064 unsigned long *zholes_size)
6065{
6066 unsigned long realtotalpages = 0, totalpages = 0;
6067 enum zone_type i;
6068
6069 for (i = 0; i < MAX_NR_ZONES; i++) {
6070 struct zone *zone = pgdat->node_zones + i;
6071 unsigned long zone_start_pfn, zone_end_pfn;
6072 unsigned long size, real_size;
6073
6074 size = zone_spanned_pages_in_node(pgdat->node_id, i,
6075 node_start_pfn,
6076 node_end_pfn,
6077 &zone_start_pfn,
6078 &zone_end_pfn,
6079 zones_size);
6080 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
6081 node_start_pfn, node_end_pfn,
6082 zholes_size);
6083 if (size)
6084 zone->zone_start_pfn = zone_start_pfn;
6085 else
6086 zone->zone_start_pfn = 0;
6087 zone->spanned_pages = size;
6088 zone->present_pages = real_size;
6089
6090 totalpages += size;
6091 realtotalpages += real_size;
6092 }
6093
6094 pgdat->node_spanned_pages = totalpages;
6095 pgdat->node_present_pages = realtotalpages;
6096 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
6097 realtotalpages);
6098}
6099
6100#ifndef CONFIG_SPARSEMEM
6101/*
6102 * Calculate the size of the zone->blockflags rounded to an unsigned long
6103 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6104 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6105 * round what is now in bits to nearest long in bits, then return it in
6106 * bytes.
6107 */
6108static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
6109{
6110 unsigned long usemapsize;
6111
6112 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
6113 usemapsize = roundup(zonesize, pageblock_nr_pages);
6114 usemapsize = usemapsize >> pageblock_order;
6115 usemapsize *= NR_PAGEBLOCK_BITS;
6116 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
6117
6118 return usemapsize / 8;
6119}
6120
6121static void __init setup_usemap(struct pglist_data *pgdat,
6122 struct zone *zone,
6123 unsigned long zone_start_pfn,
6124 unsigned long zonesize)
6125{
6126 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
6127 zone->pageblock_flags = NULL;
6128 if (usemapsize)
6129 zone->pageblock_flags =
6130 memblock_virt_alloc_node_nopanic(usemapsize,
6131 pgdat->node_id);
6132}
6133#else
6134static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
6135 unsigned long zone_start_pfn, unsigned long zonesize) {}
6136#endif /* CONFIG_SPARSEMEM */
6137
6138#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6139
6140/* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6141void __paginginit set_pageblock_order(void)
6142{
6143 unsigned int order;
6144
6145 /* Check that pageblock_nr_pages has not already been setup */
6146 if (pageblock_order)
6147 return;
6148
6149 if (HPAGE_SHIFT > PAGE_SHIFT)
6150 order = HUGETLB_PAGE_ORDER;
6151 else
6152 order = MAX_ORDER - 1;
6153
6154 /*
6155 * Assume the largest contiguous order of interest is a huge page.
6156 * This value may be variable depending on boot parameters on IA64 and
6157 * powerpc.
6158 */
6159 pageblock_order = order;
6160}
6161#else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6162
6163/*
6164 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6165 * is unused as pageblock_order is set at compile-time. See
6166 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6167 * the kernel config
6168 */
6169void __paginginit set_pageblock_order(void)
6170{
6171}
6172
6173#endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6174
6175static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
6176 unsigned long present_pages)
6177{
6178 unsigned long pages = spanned_pages;
6179
6180 /*
6181 * Provide a more accurate estimation if there are holes within
6182 * the zone and SPARSEMEM is in use. If there are holes within the
6183 * zone, each populated memory region may cost us one or two extra
6184 * memmap pages due to alignment because memmap pages for each
6185 * populated regions may not be naturally aligned on page boundary.
6186 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6187 */
6188 if (spanned_pages > present_pages + (present_pages >> 4) &&
6189 IS_ENABLED(CONFIG_SPARSEMEM))
6190 pages = present_pages;
6191
6192 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6193}
6194
6195/*
6196 * Set up the zone data structures:
6197 * - mark all pages reserved
6198 * - mark all memory queues empty
6199 * - clear the memory bitmaps
6200 *
6201 * NOTE: pgdat should get zeroed by caller.
6202 */
6203static void __paginginit free_area_init_core(struct pglist_data *pgdat)
6204{
6205 enum zone_type j;
6206 int nid = pgdat->node_id;
6207
6208 pgdat_resize_init(pgdat);
6209#ifdef CONFIG_NUMA_BALANCING
6210 spin_lock_init(&pgdat->numabalancing_migrate_lock);
6211 pgdat->numabalancing_migrate_nr_pages = 0;
6212 pgdat->numabalancing_migrate_next_window = jiffies;
6213#endif
6214#ifdef CONFIG_TRANSPARENT_HUGEPAGE
6215 spin_lock_init(&pgdat->split_queue_lock);
6216 INIT_LIST_HEAD(&pgdat->split_queue);
6217 pgdat->split_queue_len = 0;
6218#endif
6219 init_waitqueue_head(&pgdat->kswapd_wait);
6220 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6221#ifdef CONFIG_COMPACTION
6222 init_waitqueue_head(&pgdat->kcompactd_wait);
6223#endif
6224 pgdat_page_ext_init(pgdat);
6225 spin_lock_init(&pgdat->lru_lock);
6226 lruvec_init(node_lruvec(pgdat));
6227
6228 pgdat->per_cpu_nodestats = &boot_nodestats;
6229
6230 for (j = 0; j < MAX_NR_ZONES; j++) {
6231 struct zone *zone = pgdat->node_zones + j;
6232 unsigned long size, realsize, freesize, memmap_pages;
6233 unsigned long zone_start_pfn = zone->zone_start_pfn;
6234
6235 size = zone->spanned_pages;
6236 realsize = freesize = zone->present_pages;
6237
6238 /*
6239 * Adjust freesize so that it accounts for how much memory
6240 * is used by this zone for memmap. This affects the watermark
6241 * and per-cpu initialisations
6242 */
6243 memmap_pages = calc_memmap_size(size, realsize);
6244 if (!is_highmem_idx(j)) {
6245 if (freesize >= memmap_pages) {
6246 freesize -= memmap_pages;
6247 if (memmap_pages)
6248 printk(KERN_DEBUG
6249 " %s zone: %lu pages used for memmap\n",
6250 zone_names[j], memmap_pages);
6251 } else
6252 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6253 zone_names[j], memmap_pages, freesize);
6254 }
6255
6256 /* Account for reserved pages */
6257 if (j == 0 && freesize > dma_reserve) {
6258 freesize -= dma_reserve;
6259 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6260 zone_names[0], dma_reserve);
6261 }
6262
6263 if (!is_highmem_idx(j))
6264 nr_kernel_pages += freesize;
6265 /* Charge for highmem memmap if there are enough kernel pages */
6266 else if (nr_kernel_pages > memmap_pages * 2)
6267 nr_kernel_pages -= memmap_pages;
6268 nr_all_pages += freesize;
6269
6270 /*
6271 * Set an approximate value for lowmem here, it will be adjusted
6272 * when the bootmem allocator frees pages into the buddy system.
6273 * And all highmem pages will be managed by the buddy system.
6274 */
6275 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
6276#ifdef CONFIG_NUMA
6277 zone->node = nid;
6278#endif
6279 zone->name = zone_names[j];
6280 zone->zone_pgdat = pgdat;
6281 spin_lock_init(&zone->lock);
6282 zone_seqlock_init(zone);
6283 zone_pcp_init(zone);
6284
6285 if (!size)
6286 continue;
6287
6288 set_pageblock_order();
6289 setup_usemap(pgdat, zone, zone_start_pfn, size);
6290 init_currently_empty_zone(zone, zone_start_pfn, size);
6291 memmap_init(size, nid, j, zone_start_pfn);
6292 }
6293}
6294
6295#ifdef CONFIG_FLAT_NODE_MEM_MAP
6296static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
6297{
6298 unsigned long __maybe_unused start = 0;
6299 unsigned long __maybe_unused offset = 0;
6300
6301 /* Skip empty nodes */
6302 if (!pgdat->node_spanned_pages)
6303 return;
6304
6305 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6306 offset = pgdat->node_start_pfn - start;
6307 /* ia64 gets its own node_mem_map, before this, without bootmem */
6308 if (!pgdat->node_mem_map) {
6309 unsigned long size, end;
6310 struct page *map;
6311
6312 /*
6313 * The zone's endpoints aren't required to be MAX_ORDER
6314 * aligned but the node_mem_map endpoints must be in order
6315 * for the buddy allocator to function correctly.
6316 */
6317 end = pgdat_end_pfn(pgdat);
6318 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6319 size = (end - start) * sizeof(struct page);
6320 map = memblock_virt_alloc_node_nopanic(size, pgdat->node_id);
6321 pgdat->node_mem_map = map + offset;
6322 }
6323 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
6324 __func__, pgdat->node_id, (unsigned long)pgdat,
6325 (unsigned long)pgdat->node_mem_map);
6326#ifndef CONFIG_NEED_MULTIPLE_NODES
6327 /*
6328 * With no DISCONTIG, the global mem_map is just set as node 0's
6329 */
6330 if (pgdat == NODE_DATA(0)) {
6331 mem_map = NODE_DATA(0)->node_mem_map;
6332#if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6333 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6334 mem_map -= offset;
6335#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6336 }
6337#endif
6338}
6339#else
6340static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
6341#endif /* CONFIG_FLAT_NODE_MEM_MAP */
6342
6343void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
6344 unsigned long node_start_pfn, unsigned long *zholes_size)
6345{
6346 pg_data_t *pgdat = NODE_DATA(nid);
6347 unsigned long start_pfn = 0;
6348 unsigned long end_pfn = 0;
6349
6350 /* pg_data_t should be reset to zero when it's allocated */
6351 WARN_ON(pgdat->nr_zones || pgdat->kswapd_classzone_idx);
6352
6353 pgdat->node_id = nid;
6354 pgdat->node_start_pfn = node_start_pfn;
6355 pgdat->per_cpu_nodestats = NULL;
6356#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6357 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6358 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6359 (u64)start_pfn << PAGE_SHIFT,
6360 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6361#else
6362 start_pfn = node_start_pfn;
6363#endif
6364 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
6365 zones_size, zholes_size);
6366
6367 alloc_node_mem_map(pgdat);
6368
6369#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
6370 /*
6371 * We start only with one section of pages, more pages are added as
6372 * needed until the rest of deferred pages are initialized.
6373 */
6374 pgdat->static_init_pgcnt = min_t(unsigned long, PAGES_PER_SECTION,
6375 pgdat->node_spanned_pages);
6376 pgdat->first_deferred_pfn = ULONG_MAX;
6377#endif
6378 free_area_init_core(pgdat);
6379}
6380
6381#ifdef CONFIG_HAVE_MEMBLOCK
6382/*
6383 * Only struct pages that are backed by physical memory are zeroed and
6384 * initialized by going through __init_single_page(). But, there are some
6385 * struct pages which are reserved in memblock allocator and their fields
6386 * may be accessed (for example page_to_pfn() on some configuration accesses
6387 * flags). We must explicitly zero those struct pages.
6388 */
6389void __paginginit zero_resv_unavail(void)
6390{
6391 phys_addr_t start, end;
6392 unsigned long pfn;
6393 u64 i, pgcnt;
6394
6395 /*
6396 * Loop through ranges that are reserved, but do not have reported
6397 * physical memory backing.
6398 */
6399 pgcnt = 0;
6400 for_each_resv_unavail_range(i, &start, &end) {
6401 for (pfn = PFN_DOWN(start); pfn < PFN_UP(end); pfn++) {
6402 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages)))
6403 continue;
6404 mm_zero_struct_page(pfn_to_page(pfn));
6405 pgcnt++;
6406 }
6407 }
6408
6409 /*
6410 * Struct pages that do not have backing memory. This could be because
6411 * firmware is using some of this memory, or for some other reasons.
6412 * Once memblock is changed so such behaviour is not allowed: i.e.
6413 * list of "reserved" memory must be a subset of list of "memory", then
6414 * this code can be removed.
6415 */
6416 if (pgcnt)
6417 pr_info("Reserved but unavailable: %lld pages", pgcnt);
6418}
6419#endif /* CONFIG_HAVE_MEMBLOCK */
6420
6421#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6422
6423#if MAX_NUMNODES > 1
6424/*
6425 * Figure out the number of possible node ids.
6426 */
6427void __init setup_nr_node_ids(void)
6428{
6429 unsigned int highest;
6430
6431 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
6432 nr_node_ids = highest + 1;
6433}
6434#endif
6435
6436/**
6437 * node_map_pfn_alignment - determine the maximum internode alignment
6438 *
6439 * This function should be called after node map is populated and sorted.
6440 * It calculates the maximum power of two alignment which can distinguish
6441 * all the nodes.
6442 *
6443 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
6444 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
6445 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
6446 * shifted, 1GiB is enough and this function will indicate so.
6447 *
6448 * This is used to test whether pfn -> nid mapping of the chosen memory
6449 * model has fine enough granularity to avoid incorrect mapping for the
6450 * populated node map.
6451 *
6452 * Returns the determined alignment in pfn's. 0 if there is no alignment
6453 * requirement (single node).
6454 */
6455unsigned long __init node_map_pfn_alignment(void)
6456{
6457 unsigned long accl_mask = 0, last_end = 0;
6458 unsigned long start, end, mask;
6459 int last_nid = -1;
6460 int i, nid;
6461
6462 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
6463 if (!start || last_nid < 0 || last_nid == nid) {
6464 last_nid = nid;
6465 last_end = end;
6466 continue;
6467 }
6468
6469 /*
6470 * Start with a mask granular enough to pin-point to the
6471 * start pfn and tick off bits one-by-one until it becomes
6472 * too coarse to separate the current node from the last.
6473 */
6474 mask = ~((1 << __ffs(start)) - 1);
6475 while (mask && last_end <= (start & (mask << 1)))
6476 mask <<= 1;
6477
6478 /* accumulate all internode masks */
6479 accl_mask |= mask;
6480 }
6481
6482 /* convert mask to number of pages */
6483 return ~accl_mask + 1;
6484}
6485
6486/* Find the lowest pfn for a node */
6487static unsigned long __init find_min_pfn_for_node(int nid)
6488{
6489 unsigned long min_pfn = ULONG_MAX;
6490 unsigned long start_pfn;
6491 int i;
6492
6493 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
6494 min_pfn = min(min_pfn, start_pfn);
6495
6496 if (min_pfn == ULONG_MAX) {
6497 pr_warn("Could not find start_pfn for node %d\n", nid);
6498 return 0;
6499 }
6500
6501 return min_pfn;
6502}
6503
6504/**
6505 * find_min_pfn_with_active_regions - Find the minimum PFN registered
6506 *
6507 * It returns the minimum PFN based on information provided via
6508 * memblock_set_node().
6509 */
6510unsigned long __init find_min_pfn_with_active_regions(void)
6511{
6512 return find_min_pfn_for_node(MAX_NUMNODES);
6513}
6514
6515/*
6516 * early_calculate_totalpages()
6517 * Sum pages in active regions for movable zone.
6518 * Populate N_MEMORY for calculating usable_nodes.
6519 */
6520static unsigned long __init early_calculate_totalpages(void)
6521{
6522 unsigned long totalpages = 0;
6523 unsigned long start_pfn, end_pfn;
6524 int i, nid;
6525
6526 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6527 unsigned long pages = end_pfn - start_pfn;
6528
6529 totalpages += pages;
6530 if (pages)
6531 node_set_state(nid, N_MEMORY);
6532 }
6533 return totalpages;
6534}
6535
6536/*
6537 * Find the PFN the Movable zone begins in each node. Kernel memory
6538 * is spread evenly between nodes as long as the nodes have enough
6539 * memory. When they don't, some nodes will have more kernelcore than
6540 * others
6541 */
6542static void __init find_zone_movable_pfns_for_nodes(void)
6543{
6544 int i, nid;
6545 unsigned long usable_startpfn;
6546 unsigned long kernelcore_node, kernelcore_remaining;
6547 /* save the state before borrow the nodemask */
6548 nodemask_t saved_node_state = node_states[N_MEMORY];
6549 unsigned long totalpages = early_calculate_totalpages();
6550 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
6551 struct memblock_region *r;
6552
6553 /* Need to find movable_zone earlier when movable_node is specified. */
6554 find_usable_zone_for_movable();
6555
6556 /*
6557 * If movable_node is specified, ignore kernelcore and movablecore
6558 * options.
6559 */
6560 if (movable_node_is_enabled()) {
6561 for_each_memblock(memory, r) {
6562 if (!memblock_is_hotpluggable(r))
6563 continue;
6564
6565 nid = r->nid;
6566
6567 usable_startpfn = PFN_DOWN(r->base);
6568 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6569 min(usable_startpfn, zone_movable_pfn[nid]) :
6570 usable_startpfn;
6571 }
6572
6573 goto out2;
6574 }
6575
6576 /*
6577 * If kernelcore=mirror is specified, ignore movablecore option
6578 */
6579 if (mirrored_kernelcore) {
6580 bool mem_below_4gb_not_mirrored = false;
6581
6582 for_each_memblock(memory, r) {
6583 if (memblock_is_mirror(r))
6584 continue;
6585
6586 nid = r->nid;
6587
6588 usable_startpfn = memblock_region_memory_base_pfn(r);
6589
6590 if (usable_startpfn < 0x100000) {
6591 mem_below_4gb_not_mirrored = true;
6592 continue;
6593 }
6594
6595 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6596 min(usable_startpfn, zone_movable_pfn[nid]) :
6597 usable_startpfn;
6598 }
6599
6600 if (mem_below_4gb_not_mirrored)
6601 pr_warn("This configuration results in unmirrored kernel memory.");
6602
6603 goto out2;
6604 }
6605
6606 /*
6607 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
6608 * amount of necessary memory.
6609 */
6610 if (required_kernelcore_percent)
6611 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
6612 10000UL;
6613 if (required_movablecore_percent)
6614 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
6615 10000UL;
6616
6617 /*
6618 * If movablecore= was specified, calculate what size of
6619 * kernelcore that corresponds so that memory usable for
6620 * any allocation type is evenly spread. If both kernelcore
6621 * and movablecore are specified, then the value of kernelcore
6622 * will be used for required_kernelcore if it's greater than
6623 * what movablecore would have allowed.
6624 */
6625 if (required_movablecore) {
6626 unsigned long corepages;
6627
6628 /*
6629 * Round-up so that ZONE_MOVABLE is at least as large as what
6630 * was requested by the user
6631 */
6632 required_movablecore =
6633 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
6634 required_movablecore = min(totalpages, required_movablecore);
6635 corepages = totalpages - required_movablecore;
6636
6637 required_kernelcore = max(required_kernelcore, corepages);
6638 }
6639
6640 /*
6641 * If kernelcore was not specified or kernelcore size is larger
6642 * than totalpages, there is no ZONE_MOVABLE.
6643 */
6644 if (!required_kernelcore || required_kernelcore >= totalpages)
6645 goto out;
6646
6647 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
6648 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
6649
6650restart:
6651 /* Spread kernelcore memory as evenly as possible throughout nodes */
6652 kernelcore_node = required_kernelcore / usable_nodes;
6653 for_each_node_state(nid, N_MEMORY) {
6654 unsigned long start_pfn, end_pfn;
6655
6656 /*
6657 * Recalculate kernelcore_node if the division per node
6658 * now exceeds what is necessary to satisfy the requested
6659 * amount of memory for the kernel
6660 */
6661 if (required_kernelcore < kernelcore_node)
6662 kernelcore_node = required_kernelcore / usable_nodes;
6663
6664 /*
6665 * As the map is walked, we track how much memory is usable
6666 * by the kernel using kernelcore_remaining. When it is
6667 * 0, the rest of the node is usable by ZONE_MOVABLE
6668 */
6669 kernelcore_remaining = kernelcore_node;
6670
6671 /* Go through each range of PFNs within this node */
6672 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6673 unsigned long size_pages;
6674
6675 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
6676 if (start_pfn >= end_pfn)
6677 continue;
6678
6679 /* Account for what is only usable for kernelcore */
6680 if (start_pfn < usable_startpfn) {
6681 unsigned long kernel_pages;
6682 kernel_pages = min(end_pfn, usable_startpfn)
6683 - start_pfn;
6684
6685 kernelcore_remaining -= min(kernel_pages,
6686 kernelcore_remaining);
6687 required_kernelcore -= min(kernel_pages,
6688 required_kernelcore);
6689
6690 /* Continue if range is now fully accounted */
6691 if (end_pfn <= usable_startpfn) {
6692
6693 /*
6694 * Push zone_movable_pfn to the end so
6695 * that if we have to rebalance
6696 * kernelcore across nodes, we will
6697 * not double account here
6698 */
6699 zone_movable_pfn[nid] = end_pfn;
6700 continue;
6701 }
6702 start_pfn = usable_startpfn;
6703 }
6704
6705 /*
6706 * The usable PFN range for ZONE_MOVABLE is from
6707 * start_pfn->end_pfn. Calculate size_pages as the
6708 * number of pages used as kernelcore
6709 */
6710 size_pages = end_pfn - start_pfn;
6711 if (size_pages > kernelcore_remaining)
6712 size_pages = kernelcore_remaining;
6713 zone_movable_pfn[nid] = start_pfn + size_pages;
6714
6715 /*
6716 * Some kernelcore has been met, update counts and
6717 * break if the kernelcore for this node has been
6718 * satisfied
6719 */
6720 required_kernelcore -= min(required_kernelcore,
6721 size_pages);
6722 kernelcore_remaining -= size_pages;
6723 if (!kernelcore_remaining)
6724 break;
6725 }
6726 }
6727
6728 /*
6729 * If there is still required_kernelcore, we do another pass with one
6730 * less node in the count. This will push zone_movable_pfn[nid] further
6731 * along on the nodes that still have memory until kernelcore is
6732 * satisfied
6733 */
6734 usable_nodes--;
6735 if (usable_nodes && required_kernelcore > usable_nodes)
6736 goto restart;
6737
6738out2:
6739 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
6740 for (nid = 0; nid < MAX_NUMNODES; nid++)
6741 zone_movable_pfn[nid] =
6742 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
6743
6744out:
6745 /* restore the node_state */
6746 node_states[N_MEMORY] = saved_node_state;
6747}
6748
6749/* Any regular or high memory on that node ? */
6750static void check_for_memory(pg_data_t *pgdat, int nid)
6751{
6752 enum zone_type zone_type;
6753
6754 if (N_MEMORY == N_NORMAL_MEMORY)
6755 return;
6756
6757 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
6758 struct zone *zone = &pgdat->node_zones[zone_type];
6759 if (populated_zone(zone)) {
6760 node_set_state(nid, N_HIGH_MEMORY);
6761 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
6762 zone_type <= ZONE_NORMAL)
6763 node_set_state(nid, N_NORMAL_MEMORY);
6764 break;
6765 }
6766 }
6767}
6768
6769/**
6770 * free_area_init_nodes - Initialise all pg_data_t and zone data
6771 * @max_zone_pfn: an array of max PFNs for each zone
6772 *
6773 * This will call free_area_init_node() for each active node in the system.
6774 * Using the page ranges provided by memblock_set_node(), the size of each
6775 * zone in each node and their holes is calculated. If the maximum PFN
6776 * between two adjacent zones match, it is assumed that the zone is empty.
6777 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
6778 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
6779 * starts where the previous one ended. For example, ZONE_DMA32 starts
6780 * at arch_max_dma_pfn.
6781 */
6782void __init free_area_init_nodes(unsigned long *max_zone_pfn)
6783{
6784 unsigned long start_pfn, end_pfn;
6785 int i, nid;
6786
6787 /* Record where the zone boundaries are */
6788 memset(arch_zone_lowest_possible_pfn, 0,
6789 sizeof(arch_zone_lowest_possible_pfn));
6790 memset(arch_zone_highest_possible_pfn, 0,
6791 sizeof(arch_zone_highest_possible_pfn));
6792
6793 start_pfn = find_min_pfn_with_active_regions();
6794
6795 for (i = 0; i < MAX_NR_ZONES; i++) {
6796 if (i == ZONE_MOVABLE)
6797 continue;
6798
6799 end_pfn = max(max_zone_pfn[i], start_pfn);
6800 arch_zone_lowest_possible_pfn[i] = start_pfn;
6801 arch_zone_highest_possible_pfn[i] = end_pfn;
6802
6803 start_pfn = end_pfn;
6804 }
6805
6806 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
6807 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
6808 find_zone_movable_pfns_for_nodes();
6809
6810 /* Print out the zone ranges */
6811 pr_info("Zone ranges:\n");
6812 for (i = 0; i < MAX_NR_ZONES; i++) {
6813 if (i == ZONE_MOVABLE)
6814 continue;
6815 pr_info(" %-8s ", zone_names[i]);
6816 if (arch_zone_lowest_possible_pfn[i] ==
6817 arch_zone_highest_possible_pfn[i])
6818 pr_cont("empty\n");
6819 else
6820 pr_cont("[mem %#018Lx-%#018Lx]\n",
6821 (u64)arch_zone_lowest_possible_pfn[i]
6822 << PAGE_SHIFT,
6823 ((u64)arch_zone_highest_possible_pfn[i]
6824 << PAGE_SHIFT) - 1);
6825 }
6826
6827 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
6828 pr_info("Movable zone start for each node\n");
6829 for (i = 0; i < MAX_NUMNODES; i++) {
6830 if (zone_movable_pfn[i])
6831 pr_info(" Node %d: %#018Lx\n", i,
6832 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
6833 }
6834
6835 /* Print out the early node map */
6836 pr_info("Early memory node ranges\n");
6837 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
6838 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
6839 (u64)start_pfn << PAGE_SHIFT,
6840 ((u64)end_pfn << PAGE_SHIFT) - 1);
6841
6842 /* Initialise every node */
6843 mminit_verify_pageflags_layout();
6844 setup_nr_node_ids();
6845 for_each_online_node(nid) {
6846 pg_data_t *pgdat = NODE_DATA(nid);
6847 free_area_init_node(nid, NULL,
6848 find_min_pfn_for_node(nid), NULL);
6849
6850 /* Any memory on that node */
6851 if (pgdat->node_present_pages)
6852 node_set_state(nid, N_MEMORY);
6853 check_for_memory(pgdat, nid);
6854 }
6855 zero_resv_unavail();
6856}
6857
6858static int __init cmdline_parse_core(char *p, unsigned long *core,
6859 unsigned long *percent)
6860{
6861 unsigned long long coremem;
6862 char *endptr;
6863
6864 if (!p)
6865 return -EINVAL;
6866
6867 /* Value may be a percentage of total memory, otherwise bytes */
6868 coremem = simple_strtoull(p, &endptr, 0);
6869 if (*endptr == '%') {
6870 /* Paranoid check for percent values greater than 100 */
6871 WARN_ON(coremem > 100);
6872
6873 *percent = coremem;
6874 } else {
6875 coremem = memparse(p, &p);
6876 /* Paranoid check that UL is enough for the coremem value */
6877 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
6878
6879 *core = coremem >> PAGE_SHIFT;
6880 *percent = 0UL;
6881 }
6882 return 0;
6883}
6884
6885/*
6886 * kernelcore=size sets the amount of memory for use for allocations that
6887 * cannot be reclaimed or migrated.
6888 */
6889static int __init cmdline_parse_kernelcore(char *p)
6890{
6891 /* parse kernelcore=mirror */
6892 if (parse_option_str(p, "mirror")) {
6893 mirrored_kernelcore = true;
6894 return 0;
6895 }
6896
6897 return cmdline_parse_core(p, &required_kernelcore,
6898 &required_kernelcore_percent);
6899}
6900
6901/*
6902 * movablecore=size sets the amount of memory for use for allocations that
6903 * can be reclaimed or migrated.
6904 */
6905static int __init cmdline_parse_movablecore(char *p)
6906{
6907 return cmdline_parse_core(p, &required_movablecore,
6908 &required_movablecore_percent);
6909}
6910
6911early_param("kernelcore", cmdline_parse_kernelcore);
6912early_param("movablecore", cmdline_parse_movablecore);
6913
6914#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6915
6916void adjust_managed_page_count(struct page *page, long count)
6917{
6918 spin_lock(&managed_page_count_lock);
6919 page_zone(page)->managed_pages += count;
6920 totalram_pages += count;
6921#ifdef CONFIG_HIGHMEM
6922 if (PageHighMem(page))
6923 totalhigh_pages += count;
6924#endif
6925 spin_unlock(&managed_page_count_lock);
6926}
6927EXPORT_SYMBOL(adjust_managed_page_count);
6928
6929unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
6930{
6931 void *pos;
6932 unsigned long pages = 0;
6933
6934 start = (void *)PAGE_ALIGN((unsigned long)start);
6935 end = (void *)((unsigned long)end & PAGE_MASK);
6936 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
6937 if ((unsigned int)poison <= 0xFF)
6938 memset(pos, poison, PAGE_SIZE);
6939 free_reserved_page(virt_to_page(pos));
6940 }
6941
6942 if (pages && s)
6943 pr_info("Freeing %s memory: %ldK\n",
6944 s, pages << (PAGE_SHIFT - 10));
6945
6946 return pages;
6947}
6948EXPORT_SYMBOL(free_reserved_area);
6949
6950#ifdef CONFIG_HIGHMEM
6951void free_highmem_page(struct page *page)
6952{
6953 __free_reserved_page(page);
6954 totalram_pages++;
6955 page_zone(page)->managed_pages++;
6956 totalhigh_pages++;
6957}
6958#endif
6959
6960
6961void __init mem_init_print_info(const char *str)
6962{
6963 unsigned long physpages, codesize, datasize, rosize, bss_size;
6964 unsigned long init_code_size, init_data_size;
6965
6966 physpages = get_num_physpages();
6967 codesize = _etext - _stext;
6968 datasize = _edata - _sdata;
6969 rosize = __end_rodata - __start_rodata;
6970 bss_size = __bss_stop - __bss_start;
6971 init_data_size = __init_end - __init_begin;
6972 init_code_size = _einittext - _sinittext;
6973
6974 /*
6975 * Detect special cases and adjust section sizes accordingly:
6976 * 1) .init.* may be embedded into .data sections
6977 * 2) .init.text.* may be out of [__init_begin, __init_end],
6978 * please refer to arch/tile/kernel/vmlinux.lds.S.
6979 * 3) .rodata.* may be embedded into .text or .data sections.
6980 */
6981#define adj_init_size(start, end, size, pos, adj) \
6982 do { \
6983 if (start <= pos && pos < end && size > adj) \
6984 size -= adj; \
6985 } while (0)
6986
6987 adj_init_size(__init_begin, __init_end, init_data_size,
6988 _sinittext, init_code_size);
6989 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
6990 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
6991 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
6992 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
6993
6994#undef adj_init_size
6995
6996 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
6997#ifdef CONFIG_HIGHMEM
6998 ", %luK highmem"
6999#endif
7000 "%s%s)\n",
7001 nr_free_pages() << (PAGE_SHIFT - 10),
7002 physpages << (PAGE_SHIFT - 10),
7003 codesize >> 10, datasize >> 10, rosize >> 10,
7004 (init_data_size + init_code_size) >> 10, bss_size >> 10,
7005 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT - 10),
7006 totalcma_pages << (PAGE_SHIFT - 10),
7007#ifdef CONFIG_HIGHMEM
7008 totalhigh_pages << (PAGE_SHIFT - 10),
7009#endif
7010 str ? ", " : "", str ? str : "");
7011}
7012
7013/**
7014 * set_dma_reserve - set the specified number of pages reserved in the first zone
7015 * @new_dma_reserve: The number of pages to mark reserved
7016 *
7017 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7018 * In the DMA zone, a significant percentage may be consumed by kernel image
7019 * and other unfreeable allocations which can skew the watermarks badly. This
7020 * function may optionally be used to account for unfreeable pages in the
7021 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7022 * smaller per-cpu batchsize.
7023 */
7024void __init set_dma_reserve(unsigned long new_dma_reserve)
7025{
7026 dma_reserve = new_dma_reserve;
7027}
7028
7029void __init free_area_init(unsigned long *zones_size)
7030{
7031 free_area_init_node(0, zones_size,
7032 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
7033 zero_resv_unavail();
7034}
7035
7036static int page_alloc_cpu_dead(unsigned int cpu)
7037{
7038
7039 lru_add_drain_cpu(cpu);
7040 drain_pages(cpu);
7041
7042 /*
7043 * Spill the event counters of the dead processor
7044 * into the current processors event counters.
7045 * This artificially elevates the count of the current
7046 * processor.
7047 */
7048 vm_events_fold_cpu(cpu);
7049
7050 /*
7051 * Zero the differential counters of the dead processor
7052 * so that the vm statistics are consistent.
7053 *
7054 * This is only okay since the processor is dead and cannot
7055 * race with what we are doing.
7056 */
7057 cpu_vm_stats_fold(cpu);
7058 return 0;
7059}
7060
7061void __init page_alloc_init(void)
7062{
7063 int ret;
7064
7065 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
7066 "mm/page_alloc:dead", NULL,
7067 page_alloc_cpu_dead);
7068 WARN_ON(ret < 0);
7069}
7070
7071/*
7072 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7073 * or min_free_kbytes changes.
7074 */
7075static void calculate_totalreserve_pages(void)
7076{
7077 struct pglist_data *pgdat;
7078 unsigned long reserve_pages = 0;
7079 enum zone_type i, j;
7080
7081 for_each_online_pgdat(pgdat) {
7082
7083 pgdat->totalreserve_pages = 0;
7084
7085 for (i = 0; i < MAX_NR_ZONES; i++) {
7086 struct zone *zone = pgdat->node_zones + i;
7087 long max = 0;
7088
7089 /* Find valid and maximum lowmem_reserve in the zone */
7090 for (j = i; j < MAX_NR_ZONES; j++) {
7091 if (zone->lowmem_reserve[j] > max)
7092 max = zone->lowmem_reserve[j];
7093 }
7094
7095 /* we treat the high watermark as reserved pages. */
7096 max += high_wmark_pages(zone);
7097
7098 if (max > zone->managed_pages)
7099 max = zone->managed_pages;
7100
7101 pgdat->totalreserve_pages += max;
7102
7103 reserve_pages += max;
7104 }
7105 }
7106 totalreserve_pages = reserve_pages;
7107}
7108
7109/*
7110 * setup_per_zone_lowmem_reserve - called whenever
7111 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7112 * has a correct pages reserved value, so an adequate number of
7113 * pages are left in the zone after a successful __alloc_pages().
7114 */
7115static void setup_per_zone_lowmem_reserve(void)
7116{
7117 struct pglist_data *pgdat;
7118 enum zone_type j, idx;
7119
7120 for_each_online_pgdat(pgdat) {
7121 for (j = 0; j < MAX_NR_ZONES; j++) {
7122 struct zone *zone = pgdat->node_zones + j;
7123 unsigned long managed_pages = zone->managed_pages;
7124
7125 zone->lowmem_reserve[j] = 0;
7126
7127 idx = j;
7128 while (idx) {
7129 struct zone *lower_zone;
7130
7131 idx--;
7132 lower_zone = pgdat->node_zones + idx;
7133
7134 if (sysctl_lowmem_reserve_ratio[idx] < 1) {
7135 sysctl_lowmem_reserve_ratio[idx] = 0;
7136 lower_zone->lowmem_reserve[j] = 0;
7137 } else {
7138 lower_zone->lowmem_reserve[j] =
7139 managed_pages / sysctl_lowmem_reserve_ratio[idx];
7140 }
7141 managed_pages += lower_zone->managed_pages;
7142 }
7143 }
7144 }
7145
7146 /* update totalreserve_pages */
7147 calculate_totalreserve_pages();
7148}
7149
7150static void __setup_per_zone_wmarks(void)
7151{
7152 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
7153 unsigned long lowmem_pages = 0;
7154 struct zone *zone;
7155 unsigned long flags;
7156
7157 /* Calculate total number of !ZONE_HIGHMEM pages */
7158 for_each_zone(zone) {
7159 if (!is_highmem(zone))
7160 lowmem_pages += zone->managed_pages;
7161 }
7162
7163 for_each_zone(zone) {
7164 u64 tmp;
7165
7166 spin_lock_irqsave(&zone->lock, flags);
7167 tmp = (u64)pages_min * zone->managed_pages;
7168 do_div(tmp, lowmem_pages);
7169 if (is_highmem(zone)) {
7170 /*
7171 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7172 * need highmem pages, so cap pages_min to a small
7173 * value here.
7174 *
7175 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7176 * deltas control asynch page reclaim, and so should
7177 * not be capped for highmem.
7178 */
7179 unsigned long min_pages;
7180
7181 min_pages = zone->managed_pages / 1024;
7182 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
7183 zone->watermark[WMARK_MIN] = min_pages;
7184 } else {
7185 /*
7186 * If it's a lowmem zone, reserve a number of pages
7187 * proportionate to the zone's size.
7188 */
7189 zone->watermark[WMARK_MIN] = tmp;
7190 }
7191
7192 /*
7193 * Set the kswapd watermarks distance according to the
7194 * scale factor in proportion to available memory, but
7195 * ensure a minimum size on small systems.
7196 */
7197 tmp = max_t(u64, tmp >> 2,
7198 mult_frac(zone->managed_pages,
7199 watermark_scale_factor, 10000));
7200
7201 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
7202 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
7203
7204 spin_unlock_irqrestore(&zone->lock, flags);
7205 }
7206
7207 /* update totalreserve_pages */
7208 calculate_totalreserve_pages();
7209}
7210
7211/**
7212 * setup_per_zone_wmarks - called when min_free_kbytes changes
7213 * or when memory is hot-{added|removed}
7214 *
7215 * Ensures that the watermark[min,low,high] values for each zone are set
7216 * correctly with respect to min_free_kbytes.
7217 */
7218void setup_per_zone_wmarks(void)
7219{
7220 static DEFINE_SPINLOCK(lock);
7221
7222 spin_lock(&lock);
7223 __setup_per_zone_wmarks();
7224 spin_unlock(&lock);
7225}
7226
7227/*
7228 * Initialise min_free_kbytes.
7229 *
7230 * For small machines we want it small (128k min). For large machines
7231 * we want it large (64MB max). But it is not linear, because network
7232 * bandwidth does not increase linearly with machine size. We use
7233 *
7234 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7235 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7236 *
7237 * which yields
7238 *
7239 * 16MB: 512k
7240 * 32MB: 724k
7241 * 64MB: 1024k
7242 * 128MB: 1448k
7243 * 256MB: 2048k
7244 * 512MB: 2896k
7245 * 1024MB: 4096k
7246 * 2048MB: 5792k
7247 * 4096MB: 8192k
7248 * 8192MB: 11584k
7249 * 16384MB: 16384k
7250 */
7251int __meminit init_per_zone_wmark_min(void)
7252{
7253 unsigned long lowmem_kbytes;
7254 int new_min_free_kbytes;
7255
7256 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
7257 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
7258
7259 if (new_min_free_kbytes > user_min_free_kbytes) {
7260 min_free_kbytes = new_min_free_kbytes;
7261 if (min_free_kbytes < 128)
7262 min_free_kbytes = 128;
7263 if (min_free_kbytes > 65536)
7264 min_free_kbytes = 65536;
7265 } else {
7266 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7267 new_min_free_kbytes, user_min_free_kbytes);
7268 }
7269 setup_per_zone_wmarks();
7270 refresh_zone_stat_thresholds();
7271 setup_per_zone_lowmem_reserve();
7272
7273#ifdef CONFIG_NUMA
7274 setup_min_unmapped_ratio();
7275 setup_min_slab_ratio();
7276#endif
7277
7278 return 0;
7279}
7280core_initcall(init_per_zone_wmark_min)
7281
7282/*
7283 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7284 * that we can call two helper functions whenever min_free_kbytes
7285 * changes.
7286 */
7287int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
7288 void __user *buffer, size_t *length, loff_t *ppos)
7289{
7290 int rc;
7291
7292 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7293 if (rc)
7294 return rc;
7295
7296 if (write) {
7297 user_min_free_kbytes = min_free_kbytes;
7298 setup_per_zone_wmarks();
7299 }
7300 return 0;
7301}
7302
7303int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
7304 void __user *buffer, size_t *length, loff_t *ppos)
7305{
7306 int rc;
7307
7308 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7309 if (rc)
7310 return rc;
7311
7312 if (write)
7313 setup_per_zone_wmarks();
7314
7315 return 0;
7316}
7317
7318#ifdef CONFIG_NUMA
7319static void setup_min_unmapped_ratio(void)
7320{
7321 pg_data_t *pgdat;
7322 struct zone *zone;
7323
7324 for_each_online_pgdat(pgdat)
7325 pgdat->min_unmapped_pages = 0;
7326
7327 for_each_zone(zone)
7328 zone->zone_pgdat->min_unmapped_pages += (zone->managed_pages *
7329 sysctl_min_unmapped_ratio) / 100;
7330}
7331
7332
7333int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
7334 void __user *buffer, size_t *length, loff_t *ppos)
7335{
7336 int rc;
7337
7338 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7339 if (rc)
7340 return rc;
7341
7342 setup_min_unmapped_ratio();
7343
7344 return 0;
7345}
7346
7347static void setup_min_slab_ratio(void)
7348{
7349 pg_data_t *pgdat;
7350 struct zone *zone;
7351
7352 for_each_online_pgdat(pgdat)
7353 pgdat->min_slab_pages = 0;
7354
7355 for_each_zone(zone)
7356 zone->zone_pgdat->min_slab_pages += (zone->managed_pages *
7357 sysctl_min_slab_ratio) / 100;
7358}
7359
7360int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
7361 void __user *buffer, size_t *length, loff_t *ppos)
7362{
7363 int rc;
7364
7365 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7366 if (rc)
7367 return rc;
7368
7369 setup_min_slab_ratio();
7370
7371 return 0;
7372}
7373#endif
7374
7375/*
7376 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
7377 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
7378 * whenever sysctl_lowmem_reserve_ratio changes.
7379 *
7380 * The reserve ratio obviously has absolutely no relation with the
7381 * minimum watermarks. The lowmem reserve ratio can only make sense
7382 * if in function of the boot time zone sizes.
7383 */
7384int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
7385 void __user *buffer, size_t *length, loff_t *ppos)
7386{
7387 proc_dointvec_minmax(table, write, buffer, length, ppos);
7388 setup_per_zone_lowmem_reserve();
7389 return 0;
7390}
7391
7392/*
7393 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
7394 * cpu. It is the fraction of total pages in each zone that a hot per cpu
7395 * pagelist can have before it gets flushed back to buddy allocator.
7396 */
7397int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
7398 void __user *buffer, size_t *length, loff_t *ppos)
7399{
7400 struct zone *zone;
7401 int old_percpu_pagelist_fraction;
7402 int ret;
7403
7404 mutex_lock(&pcp_batch_high_lock);
7405 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
7406
7407 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
7408 if (!write || ret < 0)
7409 goto out;
7410
7411 /* Sanity checking to avoid pcp imbalance */
7412 if (percpu_pagelist_fraction &&
7413 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
7414 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
7415 ret = -EINVAL;
7416 goto out;
7417 }
7418
7419 /* No change? */
7420 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
7421 goto out;
7422
7423 for_each_populated_zone(zone) {
7424 unsigned int cpu;
7425
7426 for_each_possible_cpu(cpu)
7427 pageset_set_high_and_batch(zone,
7428 per_cpu_ptr(zone->pageset, cpu));
7429 }
7430out:
7431 mutex_unlock(&pcp_batch_high_lock);
7432 return ret;
7433}
7434
7435#ifdef CONFIG_NUMA
7436int hashdist = HASHDIST_DEFAULT;
7437
7438static int __init set_hashdist(char *str)
7439{
7440 if (!str)
7441 return 0;
7442 hashdist = simple_strtoul(str, &str, 0);
7443 return 1;
7444}
7445__setup("hashdist=", set_hashdist);
7446#endif
7447
7448#ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
7449/*
7450 * Returns the number of pages that arch has reserved but
7451 * is not known to alloc_large_system_hash().
7452 */
7453static unsigned long __init arch_reserved_kernel_pages(void)
7454{
7455 return 0;
7456}
7457#endif
7458
7459/*
7460 * Adaptive scale is meant to reduce sizes of hash tables on large memory
7461 * machines. As memory size is increased the scale is also increased but at
7462 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
7463 * quadruples the scale is increased by one, which means the size of hash table
7464 * only doubles, instead of quadrupling as well.
7465 * Because 32-bit systems cannot have large physical memory, where this scaling
7466 * makes sense, it is disabled on such platforms.
7467 */
7468#if __BITS_PER_LONG > 32
7469#define ADAPT_SCALE_BASE (64ul << 30)
7470#define ADAPT_SCALE_SHIFT 2
7471#define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
7472#endif
7473
7474/*
7475 * allocate a large system hash table from bootmem
7476 * - it is assumed that the hash table must contain an exact power-of-2
7477 * quantity of entries
7478 * - limit is the number of hash buckets, not the total allocation size
7479 */
7480void *__init alloc_large_system_hash(const char *tablename,
7481 unsigned long bucketsize,
7482 unsigned long numentries,
7483 int scale,
7484 int flags,
7485 unsigned int *_hash_shift,
7486 unsigned int *_hash_mask,
7487 unsigned long low_limit,
7488 unsigned long high_limit)
7489{
7490 unsigned long long max = high_limit;
7491 unsigned long log2qty, size;
7492 void *table = NULL;
7493 gfp_t gfp_flags;
7494
7495 /* allow the kernel cmdline to have a say */
7496 if (!numentries) {
7497 /* round applicable memory size up to nearest megabyte */
7498 numentries = nr_kernel_pages;
7499 numentries -= arch_reserved_kernel_pages();
7500
7501 /* It isn't necessary when PAGE_SIZE >= 1MB */
7502 if (PAGE_SHIFT < 20)
7503 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
7504
7505#if __BITS_PER_LONG > 32
7506 if (!high_limit) {
7507 unsigned long adapt;
7508
7509 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
7510 adapt <<= ADAPT_SCALE_SHIFT)
7511 scale++;
7512 }
7513#endif
7514
7515 /* limit to 1 bucket per 2^scale bytes of low memory */
7516 if (scale > PAGE_SHIFT)
7517 numentries >>= (scale - PAGE_SHIFT);
7518 else
7519 numentries <<= (PAGE_SHIFT - scale);
7520
7521 /* Make sure we've got at least a 0-order allocation.. */
7522 if (unlikely(flags & HASH_SMALL)) {
7523 /* Makes no sense without HASH_EARLY */
7524 WARN_ON(!(flags & HASH_EARLY));
7525 if (!(numentries >> *_hash_shift)) {
7526 numentries = 1UL << *_hash_shift;
7527 BUG_ON(!numentries);
7528 }
7529 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
7530 numentries = PAGE_SIZE / bucketsize;
7531 }
7532 numentries = roundup_pow_of_two(numentries);
7533
7534 /* limit allocation size to 1/16 total memory by default */
7535 if (max == 0) {
7536 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
7537 do_div(max, bucketsize);
7538 }
7539 max = min(max, 0x80000000ULL);
7540
7541 if (numentries < low_limit)
7542 numentries = low_limit;
7543 if (numentries > max)
7544 numentries = max;
7545
7546 log2qty = ilog2(numentries);
7547
7548 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
7549 do {
7550 size = bucketsize << log2qty;
7551 if (flags & HASH_EARLY) {
7552 if (flags & HASH_ZERO)
7553 table = memblock_virt_alloc_nopanic(size, 0);
7554 else
7555 table = memblock_virt_alloc_raw(size, 0);
7556 } else if (hashdist) {
7557 table = __vmalloc(size, gfp_flags, PAGE_KERNEL);
7558 } else {
7559 /*
7560 * If bucketsize is not a power-of-two, we may free
7561 * some pages at the end of hash table which
7562 * alloc_pages_exact() automatically does
7563 */
7564 if (get_order(size) < MAX_ORDER) {
7565 table = alloc_pages_exact(size, gfp_flags);
7566 kmemleak_alloc(table, size, 1, gfp_flags);
7567 }
7568 }
7569 } while (!table && size > PAGE_SIZE && --log2qty);
7570
7571 if (!table)
7572 panic("Failed to allocate %s hash table\n", tablename);
7573
7574 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
7575 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size);
7576
7577 if (_hash_shift)
7578 *_hash_shift = log2qty;
7579 if (_hash_mask)
7580 *_hash_mask = (1 << log2qty) - 1;
7581
7582 return table;
7583}
7584
7585/*
7586 * This function checks whether pageblock includes unmovable pages or not.
7587 * If @count is not zero, it is okay to include less @count unmovable pages
7588 *
7589 * PageLRU check without isolation or lru_lock could race so that
7590 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
7591 * check without lock_page also may miss some movable non-lru pages at
7592 * race condition. So you can't expect this function should be exact.
7593 */
7594bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
7595 int migratetype,
7596 bool skip_hwpoisoned_pages)
7597{
7598 unsigned long pfn, iter, found;
7599
7600 /*
7601 * TODO we could make this much more efficient by not checking every
7602 * page in the range if we know all of them are in MOVABLE_ZONE and
7603 * that the movable zone guarantees that pages are migratable but
7604 * the later is not the case right now unfortunatelly. E.g. movablecore
7605 * can still lead to having bootmem allocations in zone_movable.
7606 */
7607
7608 /*
7609 * CMA allocations (alloc_contig_range) really need to mark isolate
7610 * CMA pageblocks even when they are not movable in fact so consider
7611 * them movable here.
7612 */
7613 if (is_migrate_cma(migratetype) &&
7614 is_migrate_cma(get_pageblock_migratetype(page)))
7615 return false;
7616
7617 pfn = page_to_pfn(page);
7618 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
7619 unsigned long check = pfn + iter;
7620
7621 if (!pfn_valid_within(check))
7622 continue;
7623
7624 page = pfn_to_page(check);
7625
7626 if (PageReserved(page))
7627 goto unmovable;
7628
7629 /*
7630 * Hugepages are not in LRU lists, but they're movable.
7631 * We need not scan over tail pages bacause we don't
7632 * handle each tail page individually in migration.
7633 */
7634 if (PageHuge(page)) {
7635 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
7636 continue;
7637 }
7638
7639 /*
7640 * We can't use page_count without pin a page
7641 * because another CPU can free compound page.
7642 * This check already skips compound tails of THP
7643 * because their page->_refcount is zero at all time.
7644 */
7645 if (!page_ref_count(page)) {
7646 if (PageBuddy(page))
7647 iter += (1 << page_order(page)) - 1;
7648 continue;
7649 }
7650
7651 /*
7652 * The HWPoisoned page may be not in buddy system, and
7653 * page_count() is not 0.
7654 */
7655 if (skip_hwpoisoned_pages && PageHWPoison(page))
7656 continue;
7657
7658 if (__PageMovable(page))
7659 continue;
7660
7661 if (!PageLRU(page))
7662 found++;
7663 /*
7664 * If there are RECLAIMABLE pages, we need to check
7665 * it. But now, memory offline itself doesn't call
7666 * shrink_node_slabs() and it still to be fixed.
7667 */
7668 /*
7669 * If the page is not RAM, page_count()should be 0.
7670 * we don't need more check. This is an _used_ not-movable page.
7671 *
7672 * The problematic thing here is PG_reserved pages. PG_reserved
7673 * is set to both of a memory hole page and a _used_ kernel
7674 * page at boot.
7675 */
7676 if (found > count)
7677 goto unmovable;
7678 }
7679 return false;
7680unmovable:
7681 WARN_ON_ONCE(zone_idx(zone) == ZONE_MOVABLE);
7682 return true;
7683}
7684
7685bool is_pageblock_removable_nolock(struct page *page)
7686{
7687 struct zone *zone;
7688 unsigned long pfn;
7689
7690 /*
7691 * We have to be careful here because we are iterating over memory
7692 * sections which are not zone aware so we might end up outside of
7693 * the zone but still within the section.
7694 * We have to take care about the node as well. If the node is offline
7695 * its NODE_DATA will be NULL - see page_zone.
7696 */
7697 if (!node_online(page_to_nid(page)))
7698 return false;
7699
7700 zone = page_zone(page);
7701 pfn = page_to_pfn(page);
7702 if (!zone_spans_pfn(zone, pfn))
7703 return false;
7704
7705 return !has_unmovable_pages(zone, page, 0, MIGRATE_MOVABLE, true);
7706}
7707
7708#if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
7709
7710static unsigned long pfn_max_align_down(unsigned long pfn)
7711{
7712 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
7713 pageblock_nr_pages) - 1);
7714}
7715
7716static unsigned long pfn_max_align_up(unsigned long pfn)
7717{
7718 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
7719 pageblock_nr_pages));
7720}
7721
7722/* [start, end) must belong to a single zone. */
7723static int __alloc_contig_migrate_range(struct compact_control *cc,
7724 unsigned long start, unsigned long end)
7725{
7726 /* This function is based on compact_zone() from compaction.c. */
7727 unsigned long nr_reclaimed;
7728 unsigned long pfn = start;
7729 unsigned int tries = 0;
7730 int ret = 0;
7731
7732 migrate_prep();
7733
7734 while (pfn < end || !list_empty(&cc->migratepages)) {
7735 if (fatal_signal_pending(current)) {
7736 ret = -EINTR;
7737 break;
7738 }
7739
7740 if (list_empty(&cc->migratepages)) {
7741 cc->nr_migratepages = 0;
7742 pfn = isolate_migratepages_range(cc, pfn, end);
7743 if (!pfn) {
7744 ret = -EINTR;
7745 break;
7746 }
7747 tries = 0;
7748 } else if (++tries == 5) {
7749 ret = ret < 0 ? ret : -EBUSY;
7750 break;
7751 }
7752
7753 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
7754 &cc->migratepages);
7755 cc->nr_migratepages -= nr_reclaimed;
7756
7757 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
7758 NULL, 0, cc->mode, MR_CONTIG_RANGE);
7759 }
7760 if (ret < 0) {
7761 putback_movable_pages(&cc->migratepages);
7762 return ret;
7763 }
7764 return 0;
7765}
7766
7767/**
7768 * alloc_contig_range() -- tries to allocate given range of pages
7769 * @start: start PFN to allocate
7770 * @end: one-past-the-last PFN to allocate
7771 * @migratetype: migratetype of the underlaying pageblocks (either
7772 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
7773 * in range must have the same migratetype and it must
7774 * be either of the two.
7775 * @gfp_mask: GFP mask to use during compaction
7776 *
7777 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
7778 * aligned. The PFN range must belong to a single zone.
7779 *
7780 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
7781 * pageblocks in the range. Once isolated, the pageblocks should not
7782 * be modified by others.
7783 *
7784 * Returns zero on success or negative error code. On success all
7785 * pages which PFN is in [start, end) are allocated for the caller and
7786 * need to be freed with free_contig_range().
7787 */
7788int alloc_contig_range(unsigned long start, unsigned long end,
7789 unsigned migratetype, gfp_t gfp_mask)
7790{
7791 unsigned long outer_start, outer_end;
7792 unsigned int order;
7793 int ret = 0;
7794
7795 struct compact_control cc = {
7796 .nr_migratepages = 0,
7797 .order = -1,
7798 .zone = page_zone(pfn_to_page(start)),
7799 .mode = MIGRATE_SYNC,
7800 .ignore_skip_hint = true,
7801 .no_set_skip_hint = true,
7802 .gfp_mask = current_gfp_context(gfp_mask),
7803 };
7804 INIT_LIST_HEAD(&cc.migratepages);
7805
7806 /*
7807 * What we do here is we mark all pageblocks in range as
7808 * MIGRATE_ISOLATE. Because pageblock and max order pages may
7809 * have different sizes, and due to the way page allocator
7810 * work, we align the range to biggest of the two pages so
7811 * that page allocator won't try to merge buddies from
7812 * different pageblocks and change MIGRATE_ISOLATE to some
7813 * other migration type.
7814 *
7815 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
7816 * migrate the pages from an unaligned range (ie. pages that
7817 * we are interested in). This will put all the pages in
7818 * range back to page allocator as MIGRATE_ISOLATE.
7819 *
7820 * When this is done, we take the pages in range from page
7821 * allocator removing them from the buddy system. This way
7822 * page allocator will never consider using them.
7823 *
7824 * This lets us mark the pageblocks back as
7825 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
7826 * aligned range but not in the unaligned, original range are
7827 * put back to page allocator so that buddy can use them.
7828 */
7829
7830 ret = start_isolate_page_range(pfn_max_align_down(start),
7831 pfn_max_align_up(end), migratetype,
7832 false);
7833 if (ret)
7834 return ret;
7835
7836 /*
7837 * In case of -EBUSY, we'd like to know which page causes problem.
7838 * So, just fall through. test_pages_isolated() has a tracepoint
7839 * which will report the busy page.
7840 *
7841 * It is possible that busy pages could become available before
7842 * the call to test_pages_isolated, and the range will actually be
7843 * allocated. So, if we fall through be sure to clear ret so that
7844 * -EBUSY is not accidentally used or returned to caller.
7845 */
7846 ret = __alloc_contig_migrate_range(&cc, start, end);
7847 if (ret && ret != -EBUSY)
7848 goto done;
7849 ret =0;
7850
7851 /*
7852 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
7853 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
7854 * more, all pages in [start, end) are free in page allocator.
7855 * What we are going to do is to allocate all pages from
7856 * [start, end) (that is remove them from page allocator).
7857 *
7858 * The only problem is that pages at the beginning and at the
7859 * end of interesting range may be not aligned with pages that
7860 * page allocator holds, ie. they can be part of higher order
7861 * pages. Because of this, we reserve the bigger range and
7862 * once this is done free the pages we are not interested in.
7863 *
7864 * We don't have to hold zone->lock here because the pages are
7865 * isolated thus they won't get removed from buddy.
7866 */
7867
7868 lru_add_drain_all();
7869 drain_all_pages(cc.zone);
7870
7871 order = 0;
7872 outer_start = start;
7873 while (!PageBuddy(pfn_to_page(outer_start))) {
7874 if (++order >= MAX_ORDER) {
7875 outer_start = start;
7876 break;
7877 }
7878 outer_start &= ~0UL << order;
7879 }
7880
7881 if (outer_start != start) {
7882 order = page_order(pfn_to_page(outer_start));
7883
7884 /*
7885 * outer_start page could be small order buddy page and
7886 * it doesn't include start page. Adjust outer_start
7887 * in this case to report failed page properly
7888 * on tracepoint in test_pages_isolated()
7889 */
7890 if (outer_start + (1UL << order) <= start)
7891 outer_start = start;
7892 }
7893
7894 /* Make sure the range is really isolated. */
7895 if (test_pages_isolated(outer_start, end, false)) {
7896 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
7897 __func__, outer_start, end);
7898 ret = -EBUSY;
7899 goto done;
7900 }
7901
7902 /* Grab isolated pages from freelists. */
7903 outer_end = isolate_freepages_range(&cc, outer_start, end);
7904 if (!outer_end) {
7905 ret = -EBUSY;
7906 goto done;
7907 }
7908
7909 /* Free head and tail (if any) */
7910 if (start != outer_start)
7911 free_contig_range(outer_start, start - outer_start);
7912 if (end != outer_end)
7913 free_contig_range(end, outer_end - end);
7914
7915done:
7916 undo_isolate_page_range(pfn_max_align_down(start),
7917 pfn_max_align_up(end), migratetype);
7918 return ret;
7919}
7920
7921void free_contig_range(unsigned long pfn, unsigned nr_pages)
7922{
7923 unsigned int count = 0;
7924
7925 for (; nr_pages--; pfn++) {
7926 struct page *page = pfn_to_page(pfn);
7927
7928 count += page_count(page) != 1;
7929 __free_page(page);
7930 }
7931 WARN(count != 0, "%d pages are still in use!\n", count);
7932}
7933#endif
7934
7935#ifdef CONFIG_MEMORY_HOTPLUG
7936/*
7937 * The zone indicated has a new number of managed_pages; batch sizes and percpu
7938 * page high values need to be recalulated.
7939 */
7940void __meminit zone_pcp_update(struct zone *zone)
7941{
7942 unsigned cpu;
7943 mutex_lock(&pcp_batch_high_lock);
7944 for_each_possible_cpu(cpu)
7945 pageset_set_high_and_batch(zone,
7946 per_cpu_ptr(zone->pageset, cpu));
7947 mutex_unlock(&pcp_batch_high_lock);
7948}
7949#endif
7950
7951void zone_pcp_reset(struct zone *zone)
7952{
7953 unsigned long flags;
7954 int cpu;
7955 struct per_cpu_pageset *pset;
7956
7957 /* avoid races with drain_pages() */
7958 local_irq_save(flags);
7959 if (zone->pageset != &boot_pageset) {
7960 for_each_online_cpu(cpu) {
7961 pset = per_cpu_ptr(zone->pageset, cpu);
7962 drain_zonestat(zone, pset);
7963 }
7964 free_percpu(zone->pageset);
7965 zone->pageset = &boot_pageset;
7966 }
7967 local_irq_restore(flags);
7968}
7969
7970#ifdef CONFIG_MEMORY_HOTREMOVE
7971/*
7972 * All pages in the range must be in a single zone and isolated
7973 * before calling this.
7974 */
7975void
7976__offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
7977{
7978 struct page *page;
7979 struct zone *zone;
7980 unsigned int order, i;
7981 unsigned long pfn;
7982 unsigned long flags;
7983 /* find the first valid pfn */
7984 for (pfn = start_pfn; pfn < end_pfn; pfn++)
7985 if (pfn_valid(pfn))
7986 break;
7987 if (pfn == end_pfn)
7988 return;
7989 offline_mem_sections(pfn, end_pfn);
7990 zone = page_zone(pfn_to_page(pfn));
7991 spin_lock_irqsave(&zone->lock, flags);
7992 pfn = start_pfn;
7993 while (pfn < end_pfn) {
7994 if (!pfn_valid(pfn)) {
7995 pfn++;
7996 continue;
7997 }
7998 page = pfn_to_page(pfn);
7999 /*
8000 * The HWPoisoned page may be not in buddy system, and
8001 * page_count() is not 0.
8002 */
8003 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
8004 pfn++;
8005 SetPageReserved(page);
8006 continue;
8007 }
8008
8009 BUG_ON(page_count(page));
8010 BUG_ON(!PageBuddy(page));
8011 order = page_order(page);
8012#ifdef CONFIG_DEBUG_VM
8013 pr_info("remove from free list %lx %d %lx\n",
8014 pfn, 1 << order, end_pfn);
8015#endif
8016 list_del(&page->lru);
8017 rmv_page_order(page);
8018 zone->free_area[order].nr_free--;
8019 for (i = 0; i < (1 << order); i++)
8020 SetPageReserved((page+i));
8021 pfn += (1 << order);
8022 }
8023 spin_unlock_irqrestore(&zone->lock, flags);
8024}
8025#endif
8026
8027bool is_free_buddy_page(struct page *page)
8028{
8029 struct zone *zone = page_zone(page);
8030 unsigned long pfn = page_to_pfn(page);
8031 unsigned long flags;
8032 unsigned int order;
8033
8034 spin_lock_irqsave(&zone->lock, flags);
8035 for (order = 0; order < MAX_ORDER; order++) {
8036 struct page *page_head = page - (pfn & ((1 << order) - 1));
8037
8038 if (PageBuddy(page_head) && page_order(page_head) >= order)
8039 break;
8040 }
8041 spin_unlock_irqrestore(&zone->lock, flags);
8042
8043 return order < MAX_ORDER;
8044}
1/*
2 * linux/mm/page_alloc.c
3 *
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
6 *
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
15 */
16
17#include <linux/stddef.h>
18#include <linux/mm.h>
19#include <linux/swap.h>
20#include <linux/interrupt.h>
21#include <linux/pagemap.h>
22#include <linux/jiffies.h>
23#include <linux/bootmem.h>
24#include <linux/memblock.h>
25#include <linux/compiler.h>
26#include <linux/kernel.h>
27#include <linux/kmemcheck.h>
28#include <linux/kasan.h>
29#include <linux/module.h>
30#include <linux/suspend.h>
31#include <linux/pagevec.h>
32#include <linux/blkdev.h>
33#include <linux/slab.h>
34#include <linux/ratelimit.h>
35#include <linux/oom.h>
36#include <linux/notifier.h>
37#include <linux/topology.h>
38#include <linux/sysctl.h>
39#include <linux/cpu.h>
40#include <linux/cpuset.h>
41#include <linux/memory_hotplug.h>
42#include <linux/nodemask.h>
43#include <linux/vmalloc.h>
44#include <linux/vmstat.h>
45#include <linux/mempolicy.h>
46#include <linux/memremap.h>
47#include <linux/stop_machine.h>
48#include <linux/sort.h>
49#include <linux/pfn.h>
50#include <linux/backing-dev.h>
51#include <linux/fault-inject.h>
52#include <linux/page-isolation.h>
53#include <linux/page_ext.h>
54#include <linux/debugobjects.h>
55#include <linux/kmemleak.h>
56#include <linux/compaction.h>
57#include <trace/events/kmem.h>
58#include <linux/prefetch.h>
59#include <linux/mm_inline.h>
60#include <linux/migrate.h>
61#include <linux/page_ext.h>
62#include <linux/hugetlb.h>
63#include <linux/sched/rt.h>
64#include <linux/page_owner.h>
65#include <linux/kthread.h>
66#include <linux/memcontrol.h>
67
68#include <asm/sections.h>
69#include <asm/tlbflush.h>
70#include <asm/div64.h>
71#include "internal.h"
72
73/* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
74static DEFINE_MUTEX(pcp_batch_high_lock);
75#define MIN_PERCPU_PAGELIST_FRACTION (8)
76
77#ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
78DEFINE_PER_CPU(int, numa_node);
79EXPORT_PER_CPU_SYMBOL(numa_node);
80#endif
81
82#ifdef CONFIG_HAVE_MEMORYLESS_NODES
83/*
84 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
85 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
86 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
87 * defined in <linux/topology.h>.
88 */
89DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
90EXPORT_PER_CPU_SYMBOL(_numa_mem_);
91int _node_numa_mem_[MAX_NUMNODES];
92#endif
93
94#ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
95volatile unsigned long latent_entropy __latent_entropy;
96EXPORT_SYMBOL(latent_entropy);
97#endif
98
99/*
100 * Array of node states.
101 */
102nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
103 [N_POSSIBLE] = NODE_MASK_ALL,
104 [N_ONLINE] = { { [0] = 1UL } },
105#ifndef CONFIG_NUMA
106 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
107#ifdef CONFIG_HIGHMEM
108 [N_HIGH_MEMORY] = { { [0] = 1UL } },
109#endif
110#ifdef CONFIG_MOVABLE_NODE
111 [N_MEMORY] = { { [0] = 1UL } },
112#endif
113 [N_CPU] = { { [0] = 1UL } },
114#endif /* NUMA */
115};
116EXPORT_SYMBOL(node_states);
117
118/* Protect totalram_pages and zone->managed_pages */
119static DEFINE_SPINLOCK(managed_page_count_lock);
120
121unsigned long totalram_pages __read_mostly;
122unsigned long totalreserve_pages __read_mostly;
123unsigned long totalcma_pages __read_mostly;
124
125int percpu_pagelist_fraction;
126gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
127
128/*
129 * A cached value of the page's pageblock's migratetype, used when the page is
130 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
131 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
132 * Also the migratetype set in the page does not necessarily match the pcplist
133 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
134 * other index - this ensures that it will be put on the correct CMA freelist.
135 */
136static inline int get_pcppage_migratetype(struct page *page)
137{
138 return page->index;
139}
140
141static inline void set_pcppage_migratetype(struct page *page, int migratetype)
142{
143 page->index = migratetype;
144}
145
146#ifdef CONFIG_PM_SLEEP
147/*
148 * The following functions are used by the suspend/hibernate code to temporarily
149 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
150 * while devices are suspended. To avoid races with the suspend/hibernate code,
151 * they should always be called with pm_mutex held (gfp_allowed_mask also should
152 * only be modified with pm_mutex held, unless the suspend/hibernate code is
153 * guaranteed not to run in parallel with that modification).
154 */
155
156static gfp_t saved_gfp_mask;
157
158void pm_restore_gfp_mask(void)
159{
160 WARN_ON(!mutex_is_locked(&pm_mutex));
161 if (saved_gfp_mask) {
162 gfp_allowed_mask = saved_gfp_mask;
163 saved_gfp_mask = 0;
164 }
165}
166
167void pm_restrict_gfp_mask(void)
168{
169 WARN_ON(!mutex_is_locked(&pm_mutex));
170 WARN_ON(saved_gfp_mask);
171 saved_gfp_mask = gfp_allowed_mask;
172 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
173}
174
175bool pm_suspended_storage(void)
176{
177 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
178 return false;
179 return true;
180}
181#endif /* CONFIG_PM_SLEEP */
182
183#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
184unsigned int pageblock_order __read_mostly;
185#endif
186
187static void __free_pages_ok(struct page *page, unsigned int order);
188
189/*
190 * results with 256, 32 in the lowmem_reserve sysctl:
191 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
192 * 1G machine -> (16M dma, 784M normal, 224M high)
193 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
194 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
195 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
196 *
197 * TBD: should special case ZONE_DMA32 machines here - in those we normally
198 * don't need any ZONE_NORMAL reservation
199 */
200int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
201#ifdef CONFIG_ZONE_DMA
202 256,
203#endif
204#ifdef CONFIG_ZONE_DMA32
205 256,
206#endif
207#ifdef CONFIG_HIGHMEM
208 32,
209#endif
210 32,
211};
212
213EXPORT_SYMBOL(totalram_pages);
214
215static char * const zone_names[MAX_NR_ZONES] = {
216#ifdef CONFIG_ZONE_DMA
217 "DMA",
218#endif
219#ifdef CONFIG_ZONE_DMA32
220 "DMA32",
221#endif
222 "Normal",
223#ifdef CONFIG_HIGHMEM
224 "HighMem",
225#endif
226 "Movable",
227#ifdef CONFIG_ZONE_DEVICE
228 "Device",
229#endif
230};
231
232char * const migratetype_names[MIGRATE_TYPES] = {
233 "Unmovable",
234 "Movable",
235 "Reclaimable",
236 "HighAtomic",
237#ifdef CONFIG_CMA
238 "CMA",
239#endif
240#ifdef CONFIG_MEMORY_ISOLATION
241 "Isolate",
242#endif
243};
244
245compound_page_dtor * const compound_page_dtors[] = {
246 NULL,
247 free_compound_page,
248#ifdef CONFIG_HUGETLB_PAGE
249 free_huge_page,
250#endif
251#ifdef CONFIG_TRANSPARENT_HUGEPAGE
252 free_transhuge_page,
253#endif
254};
255
256int min_free_kbytes = 1024;
257int user_min_free_kbytes = -1;
258int watermark_scale_factor = 10;
259
260static unsigned long __meminitdata nr_kernel_pages;
261static unsigned long __meminitdata nr_all_pages;
262static unsigned long __meminitdata dma_reserve;
263
264#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
265static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
266static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
267static unsigned long __initdata required_kernelcore;
268static unsigned long __initdata required_movablecore;
269static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
270static bool mirrored_kernelcore;
271
272/* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
273int movable_zone;
274EXPORT_SYMBOL(movable_zone);
275#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
276
277#if MAX_NUMNODES > 1
278int nr_node_ids __read_mostly = MAX_NUMNODES;
279int nr_online_nodes __read_mostly = 1;
280EXPORT_SYMBOL(nr_node_ids);
281EXPORT_SYMBOL(nr_online_nodes);
282#endif
283
284int page_group_by_mobility_disabled __read_mostly;
285
286#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
287static inline void reset_deferred_meminit(pg_data_t *pgdat)
288{
289 pgdat->first_deferred_pfn = ULONG_MAX;
290}
291
292/* Returns true if the struct page for the pfn is uninitialised */
293static inline bool __meminit early_page_uninitialised(unsigned long pfn)
294{
295 int nid = early_pfn_to_nid(pfn);
296
297 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
298 return true;
299
300 return false;
301}
302
303/*
304 * Returns false when the remaining initialisation should be deferred until
305 * later in the boot cycle when it can be parallelised.
306 */
307static inline bool update_defer_init(pg_data_t *pgdat,
308 unsigned long pfn, unsigned long zone_end,
309 unsigned long *nr_initialised)
310{
311 unsigned long max_initialise;
312
313 /* Always populate low zones for address-contrained allocations */
314 if (zone_end < pgdat_end_pfn(pgdat))
315 return true;
316 /*
317 * Initialise at least 2G of a node but also take into account that
318 * two large system hashes that can take up 1GB for 0.25TB/node.
319 */
320 max_initialise = max(2UL << (30 - PAGE_SHIFT),
321 (pgdat->node_spanned_pages >> 8));
322
323 (*nr_initialised)++;
324 if ((*nr_initialised > max_initialise) &&
325 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
326 pgdat->first_deferred_pfn = pfn;
327 return false;
328 }
329
330 return true;
331}
332#else
333static inline void reset_deferred_meminit(pg_data_t *pgdat)
334{
335}
336
337static inline bool early_page_uninitialised(unsigned long pfn)
338{
339 return false;
340}
341
342static inline bool update_defer_init(pg_data_t *pgdat,
343 unsigned long pfn, unsigned long zone_end,
344 unsigned long *nr_initialised)
345{
346 return true;
347}
348#endif
349
350/* Return a pointer to the bitmap storing bits affecting a block of pages */
351static inline unsigned long *get_pageblock_bitmap(struct page *page,
352 unsigned long pfn)
353{
354#ifdef CONFIG_SPARSEMEM
355 return __pfn_to_section(pfn)->pageblock_flags;
356#else
357 return page_zone(page)->pageblock_flags;
358#endif /* CONFIG_SPARSEMEM */
359}
360
361static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
362{
363#ifdef CONFIG_SPARSEMEM
364 pfn &= (PAGES_PER_SECTION-1);
365 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
366#else
367 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
368 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
369#endif /* CONFIG_SPARSEMEM */
370}
371
372/**
373 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
374 * @page: The page within the block of interest
375 * @pfn: The target page frame number
376 * @end_bitidx: The last bit of interest to retrieve
377 * @mask: mask of bits that the caller is interested in
378 *
379 * Return: pageblock_bits flags
380 */
381static __always_inline unsigned long __get_pfnblock_flags_mask(struct page *page,
382 unsigned long pfn,
383 unsigned long end_bitidx,
384 unsigned long mask)
385{
386 unsigned long *bitmap;
387 unsigned long bitidx, word_bitidx;
388 unsigned long word;
389
390 bitmap = get_pageblock_bitmap(page, pfn);
391 bitidx = pfn_to_bitidx(page, pfn);
392 word_bitidx = bitidx / BITS_PER_LONG;
393 bitidx &= (BITS_PER_LONG-1);
394
395 word = bitmap[word_bitidx];
396 bitidx += end_bitidx;
397 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
398}
399
400unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
401 unsigned long end_bitidx,
402 unsigned long mask)
403{
404 return __get_pfnblock_flags_mask(page, pfn, end_bitidx, mask);
405}
406
407static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
408{
409 return __get_pfnblock_flags_mask(page, pfn, PB_migrate_end, MIGRATETYPE_MASK);
410}
411
412/**
413 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
414 * @page: The page within the block of interest
415 * @flags: The flags to set
416 * @pfn: The target page frame number
417 * @end_bitidx: The last bit of interest
418 * @mask: mask of bits that the caller is interested in
419 */
420void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
421 unsigned long pfn,
422 unsigned long end_bitidx,
423 unsigned long mask)
424{
425 unsigned long *bitmap;
426 unsigned long bitidx, word_bitidx;
427 unsigned long old_word, word;
428
429 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
430
431 bitmap = get_pageblock_bitmap(page, pfn);
432 bitidx = pfn_to_bitidx(page, pfn);
433 word_bitidx = bitidx / BITS_PER_LONG;
434 bitidx &= (BITS_PER_LONG-1);
435
436 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
437
438 bitidx += end_bitidx;
439 mask <<= (BITS_PER_LONG - bitidx - 1);
440 flags <<= (BITS_PER_LONG - bitidx - 1);
441
442 word = READ_ONCE(bitmap[word_bitidx]);
443 for (;;) {
444 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
445 if (word == old_word)
446 break;
447 word = old_word;
448 }
449}
450
451void set_pageblock_migratetype(struct page *page, int migratetype)
452{
453 if (unlikely(page_group_by_mobility_disabled &&
454 migratetype < MIGRATE_PCPTYPES))
455 migratetype = MIGRATE_UNMOVABLE;
456
457 set_pageblock_flags_group(page, (unsigned long)migratetype,
458 PB_migrate, PB_migrate_end);
459}
460
461#ifdef CONFIG_DEBUG_VM
462static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
463{
464 int ret = 0;
465 unsigned seq;
466 unsigned long pfn = page_to_pfn(page);
467 unsigned long sp, start_pfn;
468
469 do {
470 seq = zone_span_seqbegin(zone);
471 start_pfn = zone->zone_start_pfn;
472 sp = zone->spanned_pages;
473 if (!zone_spans_pfn(zone, pfn))
474 ret = 1;
475 } while (zone_span_seqretry(zone, seq));
476
477 if (ret)
478 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
479 pfn, zone_to_nid(zone), zone->name,
480 start_pfn, start_pfn + sp);
481
482 return ret;
483}
484
485static int page_is_consistent(struct zone *zone, struct page *page)
486{
487 if (!pfn_valid_within(page_to_pfn(page)))
488 return 0;
489 if (zone != page_zone(page))
490 return 0;
491
492 return 1;
493}
494/*
495 * Temporary debugging check for pages not lying within a given zone.
496 */
497static int bad_range(struct zone *zone, struct page *page)
498{
499 if (page_outside_zone_boundaries(zone, page))
500 return 1;
501 if (!page_is_consistent(zone, page))
502 return 1;
503
504 return 0;
505}
506#else
507static inline int bad_range(struct zone *zone, struct page *page)
508{
509 return 0;
510}
511#endif
512
513static void bad_page(struct page *page, const char *reason,
514 unsigned long bad_flags)
515{
516 static unsigned long resume;
517 static unsigned long nr_shown;
518 static unsigned long nr_unshown;
519
520 /*
521 * Allow a burst of 60 reports, then keep quiet for that minute;
522 * or allow a steady drip of one report per second.
523 */
524 if (nr_shown == 60) {
525 if (time_before(jiffies, resume)) {
526 nr_unshown++;
527 goto out;
528 }
529 if (nr_unshown) {
530 pr_alert(
531 "BUG: Bad page state: %lu messages suppressed\n",
532 nr_unshown);
533 nr_unshown = 0;
534 }
535 nr_shown = 0;
536 }
537 if (nr_shown++ == 0)
538 resume = jiffies + 60 * HZ;
539
540 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
541 current->comm, page_to_pfn(page));
542 __dump_page(page, reason);
543 bad_flags &= page->flags;
544 if (bad_flags)
545 pr_alert("bad because of flags: %#lx(%pGp)\n",
546 bad_flags, &bad_flags);
547 dump_page_owner(page);
548
549 print_modules();
550 dump_stack();
551out:
552 /* Leave bad fields for debug, except PageBuddy could make trouble */
553 page_mapcount_reset(page); /* remove PageBuddy */
554 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
555}
556
557/*
558 * Higher-order pages are called "compound pages". They are structured thusly:
559 *
560 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
561 *
562 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
563 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
564 *
565 * The first tail page's ->compound_dtor holds the offset in array of compound
566 * page destructors. See compound_page_dtors.
567 *
568 * The first tail page's ->compound_order holds the order of allocation.
569 * This usage means that zero-order pages may not be compound.
570 */
571
572void free_compound_page(struct page *page)
573{
574 __free_pages_ok(page, compound_order(page));
575}
576
577void prep_compound_page(struct page *page, unsigned int order)
578{
579 int i;
580 int nr_pages = 1 << order;
581
582 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
583 set_compound_order(page, order);
584 __SetPageHead(page);
585 for (i = 1; i < nr_pages; i++) {
586 struct page *p = page + i;
587 set_page_count(p, 0);
588 p->mapping = TAIL_MAPPING;
589 set_compound_head(p, page);
590 }
591 atomic_set(compound_mapcount_ptr(page), -1);
592}
593
594#ifdef CONFIG_DEBUG_PAGEALLOC
595unsigned int _debug_guardpage_minorder;
596bool _debug_pagealloc_enabled __read_mostly
597 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
598EXPORT_SYMBOL(_debug_pagealloc_enabled);
599bool _debug_guardpage_enabled __read_mostly;
600
601static int __init early_debug_pagealloc(char *buf)
602{
603 if (!buf)
604 return -EINVAL;
605 return kstrtobool(buf, &_debug_pagealloc_enabled);
606}
607early_param("debug_pagealloc", early_debug_pagealloc);
608
609static bool need_debug_guardpage(void)
610{
611 /* If we don't use debug_pagealloc, we don't need guard page */
612 if (!debug_pagealloc_enabled())
613 return false;
614
615 if (!debug_guardpage_minorder())
616 return false;
617
618 return true;
619}
620
621static void init_debug_guardpage(void)
622{
623 if (!debug_pagealloc_enabled())
624 return;
625
626 if (!debug_guardpage_minorder())
627 return;
628
629 _debug_guardpage_enabled = true;
630}
631
632struct page_ext_operations debug_guardpage_ops = {
633 .need = need_debug_guardpage,
634 .init = init_debug_guardpage,
635};
636
637static int __init debug_guardpage_minorder_setup(char *buf)
638{
639 unsigned long res;
640
641 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
642 pr_err("Bad debug_guardpage_minorder value\n");
643 return 0;
644 }
645 _debug_guardpage_minorder = res;
646 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
647 return 0;
648}
649early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
650
651static inline bool set_page_guard(struct zone *zone, struct page *page,
652 unsigned int order, int migratetype)
653{
654 struct page_ext *page_ext;
655
656 if (!debug_guardpage_enabled())
657 return false;
658
659 if (order >= debug_guardpage_minorder())
660 return false;
661
662 page_ext = lookup_page_ext(page);
663 if (unlikely(!page_ext))
664 return false;
665
666 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
667
668 INIT_LIST_HEAD(&page->lru);
669 set_page_private(page, order);
670 /* Guard pages are not available for any usage */
671 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
672
673 return true;
674}
675
676static inline void clear_page_guard(struct zone *zone, struct page *page,
677 unsigned int order, int migratetype)
678{
679 struct page_ext *page_ext;
680
681 if (!debug_guardpage_enabled())
682 return;
683
684 page_ext = lookup_page_ext(page);
685 if (unlikely(!page_ext))
686 return;
687
688 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
689
690 set_page_private(page, 0);
691 if (!is_migrate_isolate(migratetype))
692 __mod_zone_freepage_state(zone, (1 << order), migratetype);
693}
694#else
695struct page_ext_operations debug_guardpage_ops;
696static inline bool set_page_guard(struct zone *zone, struct page *page,
697 unsigned int order, int migratetype) { return false; }
698static inline void clear_page_guard(struct zone *zone, struct page *page,
699 unsigned int order, int migratetype) {}
700#endif
701
702static inline void set_page_order(struct page *page, unsigned int order)
703{
704 set_page_private(page, order);
705 __SetPageBuddy(page);
706}
707
708static inline void rmv_page_order(struct page *page)
709{
710 __ClearPageBuddy(page);
711 set_page_private(page, 0);
712}
713
714/*
715 * This function checks whether a page is free && is the buddy
716 * we can do coalesce a page and its buddy if
717 * (a) the buddy is not in a hole &&
718 * (b) the buddy is in the buddy system &&
719 * (c) a page and its buddy have the same order &&
720 * (d) a page and its buddy are in the same zone.
721 *
722 * For recording whether a page is in the buddy system, we set ->_mapcount
723 * PAGE_BUDDY_MAPCOUNT_VALUE.
724 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
725 * serialized by zone->lock.
726 *
727 * For recording page's order, we use page_private(page).
728 */
729static inline int page_is_buddy(struct page *page, struct page *buddy,
730 unsigned int order)
731{
732 if (!pfn_valid_within(page_to_pfn(buddy)))
733 return 0;
734
735 if (page_is_guard(buddy) && page_order(buddy) == order) {
736 if (page_zone_id(page) != page_zone_id(buddy))
737 return 0;
738
739 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
740
741 return 1;
742 }
743
744 if (PageBuddy(buddy) && page_order(buddy) == order) {
745 /*
746 * zone check is done late to avoid uselessly
747 * calculating zone/node ids for pages that could
748 * never merge.
749 */
750 if (page_zone_id(page) != page_zone_id(buddy))
751 return 0;
752
753 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
754
755 return 1;
756 }
757 return 0;
758}
759
760/*
761 * Freeing function for a buddy system allocator.
762 *
763 * The concept of a buddy system is to maintain direct-mapped table
764 * (containing bit values) for memory blocks of various "orders".
765 * The bottom level table contains the map for the smallest allocatable
766 * units of memory (here, pages), and each level above it describes
767 * pairs of units from the levels below, hence, "buddies".
768 * At a high level, all that happens here is marking the table entry
769 * at the bottom level available, and propagating the changes upward
770 * as necessary, plus some accounting needed to play nicely with other
771 * parts of the VM system.
772 * At each level, we keep a list of pages, which are heads of continuous
773 * free pages of length of (1 << order) and marked with _mapcount
774 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
775 * field.
776 * So when we are allocating or freeing one, we can derive the state of the
777 * other. That is, if we allocate a small block, and both were
778 * free, the remainder of the region must be split into blocks.
779 * If a block is freed, and its buddy is also free, then this
780 * triggers coalescing into a block of larger size.
781 *
782 * -- nyc
783 */
784
785static inline void __free_one_page(struct page *page,
786 unsigned long pfn,
787 struct zone *zone, unsigned int order,
788 int migratetype)
789{
790 unsigned long page_idx;
791 unsigned long combined_idx;
792 unsigned long uninitialized_var(buddy_idx);
793 struct page *buddy;
794 unsigned int max_order;
795
796 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
797
798 VM_BUG_ON(!zone_is_initialized(zone));
799 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
800
801 VM_BUG_ON(migratetype == -1);
802 if (likely(!is_migrate_isolate(migratetype)))
803 __mod_zone_freepage_state(zone, 1 << order, migratetype);
804
805 page_idx = pfn & ((1 << MAX_ORDER) - 1);
806
807 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page);
808 VM_BUG_ON_PAGE(bad_range(zone, page), page);
809
810continue_merging:
811 while (order < max_order - 1) {
812 buddy_idx = __find_buddy_index(page_idx, order);
813 buddy = page + (buddy_idx - page_idx);
814 if (!page_is_buddy(page, buddy, order))
815 goto done_merging;
816 /*
817 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
818 * merge with it and move up one order.
819 */
820 if (page_is_guard(buddy)) {
821 clear_page_guard(zone, buddy, order, migratetype);
822 } else {
823 list_del(&buddy->lru);
824 zone->free_area[order].nr_free--;
825 rmv_page_order(buddy);
826 }
827 combined_idx = buddy_idx & page_idx;
828 page = page + (combined_idx - page_idx);
829 page_idx = combined_idx;
830 order++;
831 }
832 if (max_order < MAX_ORDER) {
833 /* If we are here, it means order is >= pageblock_order.
834 * We want to prevent merge between freepages on isolate
835 * pageblock and normal pageblock. Without this, pageblock
836 * isolation could cause incorrect freepage or CMA accounting.
837 *
838 * We don't want to hit this code for the more frequent
839 * low-order merging.
840 */
841 if (unlikely(has_isolate_pageblock(zone))) {
842 int buddy_mt;
843
844 buddy_idx = __find_buddy_index(page_idx, order);
845 buddy = page + (buddy_idx - page_idx);
846 buddy_mt = get_pageblock_migratetype(buddy);
847
848 if (migratetype != buddy_mt
849 && (is_migrate_isolate(migratetype) ||
850 is_migrate_isolate(buddy_mt)))
851 goto done_merging;
852 }
853 max_order++;
854 goto continue_merging;
855 }
856
857done_merging:
858 set_page_order(page, order);
859
860 /*
861 * If this is not the largest possible page, check if the buddy
862 * of the next-highest order is free. If it is, it's possible
863 * that pages are being freed that will coalesce soon. In case,
864 * that is happening, add the free page to the tail of the list
865 * so it's less likely to be used soon and more likely to be merged
866 * as a higher order page
867 */
868 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
869 struct page *higher_page, *higher_buddy;
870 combined_idx = buddy_idx & page_idx;
871 higher_page = page + (combined_idx - page_idx);
872 buddy_idx = __find_buddy_index(combined_idx, order + 1);
873 higher_buddy = higher_page + (buddy_idx - combined_idx);
874 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
875 list_add_tail(&page->lru,
876 &zone->free_area[order].free_list[migratetype]);
877 goto out;
878 }
879 }
880
881 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
882out:
883 zone->free_area[order].nr_free++;
884}
885
886/*
887 * A bad page could be due to a number of fields. Instead of multiple branches,
888 * try and check multiple fields with one check. The caller must do a detailed
889 * check if necessary.
890 */
891static inline bool page_expected_state(struct page *page,
892 unsigned long check_flags)
893{
894 if (unlikely(atomic_read(&page->_mapcount) != -1))
895 return false;
896
897 if (unlikely((unsigned long)page->mapping |
898 page_ref_count(page) |
899#ifdef CONFIG_MEMCG
900 (unsigned long)page->mem_cgroup |
901#endif
902 (page->flags & check_flags)))
903 return false;
904
905 return true;
906}
907
908static void free_pages_check_bad(struct page *page)
909{
910 const char *bad_reason;
911 unsigned long bad_flags;
912
913 bad_reason = NULL;
914 bad_flags = 0;
915
916 if (unlikely(atomic_read(&page->_mapcount) != -1))
917 bad_reason = "nonzero mapcount";
918 if (unlikely(page->mapping != NULL))
919 bad_reason = "non-NULL mapping";
920 if (unlikely(page_ref_count(page) != 0))
921 bad_reason = "nonzero _refcount";
922 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
923 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
924 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
925 }
926#ifdef CONFIG_MEMCG
927 if (unlikely(page->mem_cgroup))
928 bad_reason = "page still charged to cgroup";
929#endif
930 bad_page(page, bad_reason, bad_flags);
931}
932
933static inline int free_pages_check(struct page *page)
934{
935 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
936 return 0;
937
938 /* Something has gone sideways, find it */
939 free_pages_check_bad(page);
940 return 1;
941}
942
943static int free_tail_pages_check(struct page *head_page, struct page *page)
944{
945 int ret = 1;
946
947 /*
948 * We rely page->lru.next never has bit 0 set, unless the page
949 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
950 */
951 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
952
953 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
954 ret = 0;
955 goto out;
956 }
957 switch (page - head_page) {
958 case 1:
959 /* the first tail page: ->mapping is compound_mapcount() */
960 if (unlikely(compound_mapcount(page))) {
961 bad_page(page, "nonzero compound_mapcount", 0);
962 goto out;
963 }
964 break;
965 case 2:
966 /*
967 * the second tail page: ->mapping is
968 * page_deferred_list().next -- ignore value.
969 */
970 break;
971 default:
972 if (page->mapping != TAIL_MAPPING) {
973 bad_page(page, "corrupted mapping in tail page", 0);
974 goto out;
975 }
976 break;
977 }
978 if (unlikely(!PageTail(page))) {
979 bad_page(page, "PageTail not set", 0);
980 goto out;
981 }
982 if (unlikely(compound_head(page) != head_page)) {
983 bad_page(page, "compound_head not consistent", 0);
984 goto out;
985 }
986 ret = 0;
987out:
988 page->mapping = NULL;
989 clear_compound_head(page);
990 return ret;
991}
992
993static __always_inline bool free_pages_prepare(struct page *page,
994 unsigned int order, bool check_free)
995{
996 int bad = 0;
997
998 VM_BUG_ON_PAGE(PageTail(page), page);
999
1000 trace_mm_page_free(page, order);
1001 kmemcheck_free_shadow(page, order);
1002
1003 /*
1004 * Check tail pages before head page information is cleared to
1005 * avoid checking PageCompound for order-0 pages.
1006 */
1007 if (unlikely(order)) {
1008 bool compound = PageCompound(page);
1009 int i;
1010
1011 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1012
1013 if (compound)
1014 ClearPageDoubleMap(page);
1015 for (i = 1; i < (1 << order); i++) {
1016 if (compound)
1017 bad += free_tail_pages_check(page, page + i);
1018 if (unlikely(free_pages_check(page + i))) {
1019 bad++;
1020 continue;
1021 }
1022 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1023 }
1024 }
1025 if (PageMappingFlags(page))
1026 page->mapping = NULL;
1027 if (memcg_kmem_enabled() && PageKmemcg(page))
1028 memcg_kmem_uncharge(page, order);
1029 if (check_free)
1030 bad += free_pages_check(page);
1031 if (bad)
1032 return false;
1033
1034 page_cpupid_reset_last(page);
1035 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1036 reset_page_owner(page, order);
1037
1038 if (!PageHighMem(page)) {
1039 debug_check_no_locks_freed(page_address(page),
1040 PAGE_SIZE << order);
1041 debug_check_no_obj_freed(page_address(page),
1042 PAGE_SIZE << order);
1043 }
1044 arch_free_page(page, order);
1045 kernel_poison_pages(page, 1 << order, 0);
1046 kernel_map_pages(page, 1 << order, 0);
1047 kasan_free_pages(page, order);
1048
1049 return true;
1050}
1051
1052#ifdef CONFIG_DEBUG_VM
1053static inline bool free_pcp_prepare(struct page *page)
1054{
1055 return free_pages_prepare(page, 0, true);
1056}
1057
1058static inline bool bulkfree_pcp_prepare(struct page *page)
1059{
1060 return false;
1061}
1062#else
1063static bool free_pcp_prepare(struct page *page)
1064{
1065 return free_pages_prepare(page, 0, false);
1066}
1067
1068static bool bulkfree_pcp_prepare(struct page *page)
1069{
1070 return free_pages_check(page);
1071}
1072#endif /* CONFIG_DEBUG_VM */
1073
1074/*
1075 * Frees a number of pages from the PCP lists
1076 * Assumes all pages on list are in same zone, and of same order.
1077 * count is the number of pages to free.
1078 *
1079 * If the zone was previously in an "all pages pinned" state then look to
1080 * see if this freeing clears that state.
1081 *
1082 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1083 * pinned" detection logic.
1084 */
1085static void free_pcppages_bulk(struct zone *zone, int count,
1086 struct per_cpu_pages *pcp)
1087{
1088 int migratetype = 0;
1089 int batch_free = 0;
1090 unsigned long nr_scanned;
1091 bool isolated_pageblocks;
1092
1093 spin_lock(&zone->lock);
1094 isolated_pageblocks = has_isolate_pageblock(zone);
1095 nr_scanned = node_page_state(zone->zone_pgdat, NR_PAGES_SCANNED);
1096 if (nr_scanned)
1097 __mod_node_page_state(zone->zone_pgdat, NR_PAGES_SCANNED, -nr_scanned);
1098
1099 while (count) {
1100 struct page *page;
1101 struct list_head *list;
1102
1103 /*
1104 * Remove pages from lists in a round-robin fashion. A
1105 * batch_free count is maintained that is incremented when an
1106 * empty list is encountered. This is so more pages are freed
1107 * off fuller lists instead of spinning excessively around empty
1108 * lists
1109 */
1110 do {
1111 batch_free++;
1112 if (++migratetype == MIGRATE_PCPTYPES)
1113 migratetype = 0;
1114 list = &pcp->lists[migratetype];
1115 } while (list_empty(list));
1116
1117 /* This is the only non-empty list. Free them all. */
1118 if (batch_free == MIGRATE_PCPTYPES)
1119 batch_free = count;
1120
1121 do {
1122 int mt; /* migratetype of the to-be-freed page */
1123
1124 page = list_last_entry(list, struct page, lru);
1125 /* must delete as __free_one_page list manipulates */
1126 list_del(&page->lru);
1127
1128 mt = get_pcppage_migratetype(page);
1129 /* MIGRATE_ISOLATE page should not go to pcplists */
1130 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1131 /* Pageblock could have been isolated meanwhile */
1132 if (unlikely(isolated_pageblocks))
1133 mt = get_pageblock_migratetype(page);
1134
1135 if (bulkfree_pcp_prepare(page))
1136 continue;
1137
1138 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
1139 trace_mm_page_pcpu_drain(page, 0, mt);
1140 } while (--count && --batch_free && !list_empty(list));
1141 }
1142 spin_unlock(&zone->lock);
1143}
1144
1145static void free_one_page(struct zone *zone,
1146 struct page *page, unsigned long pfn,
1147 unsigned int order,
1148 int migratetype)
1149{
1150 unsigned long nr_scanned;
1151 spin_lock(&zone->lock);
1152 nr_scanned = node_page_state(zone->zone_pgdat, NR_PAGES_SCANNED);
1153 if (nr_scanned)
1154 __mod_node_page_state(zone->zone_pgdat, NR_PAGES_SCANNED, -nr_scanned);
1155
1156 if (unlikely(has_isolate_pageblock(zone) ||
1157 is_migrate_isolate(migratetype))) {
1158 migratetype = get_pfnblock_migratetype(page, pfn);
1159 }
1160 __free_one_page(page, pfn, zone, order, migratetype);
1161 spin_unlock(&zone->lock);
1162}
1163
1164static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1165 unsigned long zone, int nid)
1166{
1167 set_page_links(page, zone, nid, pfn);
1168 init_page_count(page);
1169 page_mapcount_reset(page);
1170 page_cpupid_reset_last(page);
1171
1172 INIT_LIST_HEAD(&page->lru);
1173#ifdef WANT_PAGE_VIRTUAL
1174 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1175 if (!is_highmem_idx(zone))
1176 set_page_address(page, __va(pfn << PAGE_SHIFT));
1177#endif
1178}
1179
1180static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
1181 int nid)
1182{
1183 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
1184}
1185
1186#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1187static void init_reserved_page(unsigned long pfn)
1188{
1189 pg_data_t *pgdat;
1190 int nid, zid;
1191
1192 if (!early_page_uninitialised(pfn))
1193 return;
1194
1195 nid = early_pfn_to_nid(pfn);
1196 pgdat = NODE_DATA(nid);
1197
1198 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1199 struct zone *zone = &pgdat->node_zones[zid];
1200
1201 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1202 break;
1203 }
1204 __init_single_pfn(pfn, zid, nid);
1205}
1206#else
1207static inline void init_reserved_page(unsigned long pfn)
1208{
1209}
1210#endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1211
1212/*
1213 * Initialised pages do not have PageReserved set. This function is
1214 * called for each range allocated by the bootmem allocator and
1215 * marks the pages PageReserved. The remaining valid pages are later
1216 * sent to the buddy page allocator.
1217 */
1218void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1219{
1220 unsigned long start_pfn = PFN_DOWN(start);
1221 unsigned long end_pfn = PFN_UP(end);
1222
1223 for (; start_pfn < end_pfn; start_pfn++) {
1224 if (pfn_valid(start_pfn)) {
1225 struct page *page = pfn_to_page(start_pfn);
1226
1227 init_reserved_page(start_pfn);
1228
1229 /* Avoid false-positive PageTail() */
1230 INIT_LIST_HEAD(&page->lru);
1231
1232 SetPageReserved(page);
1233 }
1234 }
1235}
1236
1237static void __free_pages_ok(struct page *page, unsigned int order)
1238{
1239 unsigned long flags;
1240 int migratetype;
1241 unsigned long pfn = page_to_pfn(page);
1242
1243 if (!free_pages_prepare(page, order, true))
1244 return;
1245
1246 migratetype = get_pfnblock_migratetype(page, pfn);
1247 local_irq_save(flags);
1248 __count_vm_events(PGFREE, 1 << order);
1249 free_one_page(page_zone(page), page, pfn, order, migratetype);
1250 local_irq_restore(flags);
1251}
1252
1253static void __init __free_pages_boot_core(struct page *page, unsigned int order)
1254{
1255 unsigned int nr_pages = 1 << order;
1256 struct page *p = page;
1257 unsigned int loop;
1258
1259 prefetchw(p);
1260 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1261 prefetchw(p + 1);
1262 __ClearPageReserved(p);
1263 set_page_count(p, 0);
1264 }
1265 __ClearPageReserved(p);
1266 set_page_count(p, 0);
1267
1268 page_zone(page)->managed_pages += nr_pages;
1269 set_page_refcounted(page);
1270 __free_pages(page, order);
1271}
1272
1273#if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1274 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1275
1276static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1277
1278int __meminit early_pfn_to_nid(unsigned long pfn)
1279{
1280 static DEFINE_SPINLOCK(early_pfn_lock);
1281 int nid;
1282
1283 spin_lock(&early_pfn_lock);
1284 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1285 if (nid < 0)
1286 nid = first_online_node;
1287 spin_unlock(&early_pfn_lock);
1288
1289 return nid;
1290}
1291#endif
1292
1293#ifdef CONFIG_NODES_SPAN_OTHER_NODES
1294static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1295 struct mminit_pfnnid_cache *state)
1296{
1297 int nid;
1298
1299 nid = __early_pfn_to_nid(pfn, state);
1300 if (nid >= 0 && nid != node)
1301 return false;
1302 return true;
1303}
1304
1305/* Only safe to use early in boot when initialisation is single-threaded */
1306static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1307{
1308 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1309}
1310
1311#else
1312
1313static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1314{
1315 return true;
1316}
1317static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1318 struct mminit_pfnnid_cache *state)
1319{
1320 return true;
1321}
1322#endif
1323
1324
1325void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1326 unsigned int order)
1327{
1328 if (early_page_uninitialised(pfn))
1329 return;
1330 return __free_pages_boot_core(page, order);
1331}
1332
1333/*
1334 * Check that the whole (or subset of) a pageblock given by the interval of
1335 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1336 * with the migration of free compaction scanner. The scanners then need to
1337 * use only pfn_valid_within() check for arches that allow holes within
1338 * pageblocks.
1339 *
1340 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1341 *
1342 * It's possible on some configurations to have a setup like node0 node1 node0
1343 * i.e. it's possible that all pages within a zones range of pages do not
1344 * belong to a single zone. We assume that a border between node0 and node1
1345 * can occur within a single pageblock, but not a node0 node1 node0
1346 * interleaving within a single pageblock. It is therefore sufficient to check
1347 * the first and last page of a pageblock and avoid checking each individual
1348 * page in a pageblock.
1349 */
1350struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1351 unsigned long end_pfn, struct zone *zone)
1352{
1353 struct page *start_page;
1354 struct page *end_page;
1355
1356 /* end_pfn is one past the range we are checking */
1357 end_pfn--;
1358
1359 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1360 return NULL;
1361
1362 start_page = pfn_to_page(start_pfn);
1363
1364 if (page_zone(start_page) != zone)
1365 return NULL;
1366
1367 end_page = pfn_to_page(end_pfn);
1368
1369 /* This gives a shorter code than deriving page_zone(end_page) */
1370 if (page_zone_id(start_page) != page_zone_id(end_page))
1371 return NULL;
1372
1373 return start_page;
1374}
1375
1376void set_zone_contiguous(struct zone *zone)
1377{
1378 unsigned long block_start_pfn = zone->zone_start_pfn;
1379 unsigned long block_end_pfn;
1380
1381 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1382 for (; block_start_pfn < zone_end_pfn(zone);
1383 block_start_pfn = block_end_pfn,
1384 block_end_pfn += pageblock_nr_pages) {
1385
1386 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1387
1388 if (!__pageblock_pfn_to_page(block_start_pfn,
1389 block_end_pfn, zone))
1390 return;
1391 }
1392
1393 /* We confirm that there is no hole */
1394 zone->contiguous = true;
1395}
1396
1397void clear_zone_contiguous(struct zone *zone)
1398{
1399 zone->contiguous = false;
1400}
1401
1402#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1403static void __init deferred_free_range(struct page *page,
1404 unsigned long pfn, int nr_pages)
1405{
1406 int i;
1407
1408 if (!page)
1409 return;
1410
1411 /* Free a large naturally-aligned chunk if possible */
1412 if (nr_pages == pageblock_nr_pages &&
1413 (pfn & (pageblock_nr_pages - 1)) == 0) {
1414 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1415 __free_pages_boot_core(page, pageblock_order);
1416 return;
1417 }
1418
1419 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1420 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1421 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1422 __free_pages_boot_core(page, 0);
1423 }
1424}
1425
1426/* Completion tracking for deferred_init_memmap() threads */
1427static atomic_t pgdat_init_n_undone __initdata;
1428static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1429
1430static inline void __init pgdat_init_report_one_done(void)
1431{
1432 if (atomic_dec_and_test(&pgdat_init_n_undone))
1433 complete(&pgdat_init_all_done_comp);
1434}
1435
1436/* Initialise remaining memory on a node */
1437static int __init deferred_init_memmap(void *data)
1438{
1439 pg_data_t *pgdat = data;
1440 int nid = pgdat->node_id;
1441 struct mminit_pfnnid_cache nid_init_state = { };
1442 unsigned long start = jiffies;
1443 unsigned long nr_pages = 0;
1444 unsigned long walk_start, walk_end;
1445 int i, zid;
1446 struct zone *zone;
1447 unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1448 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1449
1450 if (first_init_pfn == ULONG_MAX) {
1451 pgdat_init_report_one_done();
1452 return 0;
1453 }
1454
1455 /* Bind memory initialisation thread to a local node if possible */
1456 if (!cpumask_empty(cpumask))
1457 set_cpus_allowed_ptr(current, cpumask);
1458
1459 /* Sanity check boundaries */
1460 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1461 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1462 pgdat->first_deferred_pfn = ULONG_MAX;
1463
1464 /* Only the highest zone is deferred so find it */
1465 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1466 zone = pgdat->node_zones + zid;
1467 if (first_init_pfn < zone_end_pfn(zone))
1468 break;
1469 }
1470
1471 for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) {
1472 unsigned long pfn, end_pfn;
1473 struct page *page = NULL;
1474 struct page *free_base_page = NULL;
1475 unsigned long free_base_pfn = 0;
1476 int nr_to_free = 0;
1477
1478 end_pfn = min(walk_end, zone_end_pfn(zone));
1479 pfn = first_init_pfn;
1480 if (pfn < walk_start)
1481 pfn = walk_start;
1482 if (pfn < zone->zone_start_pfn)
1483 pfn = zone->zone_start_pfn;
1484
1485 for (; pfn < end_pfn; pfn++) {
1486 if (!pfn_valid_within(pfn))
1487 goto free_range;
1488
1489 /*
1490 * Ensure pfn_valid is checked every
1491 * pageblock_nr_pages for memory holes
1492 */
1493 if ((pfn & (pageblock_nr_pages - 1)) == 0) {
1494 if (!pfn_valid(pfn)) {
1495 page = NULL;
1496 goto free_range;
1497 }
1498 }
1499
1500 if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1501 page = NULL;
1502 goto free_range;
1503 }
1504
1505 /* Minimise pfn page lookups and scheduler checks */
1506 if (page && (pfn & (pageblock_nr_pages - 1)) != 0) {
1507 page++;
1508 } else {
1509 nr_pages += nr_to_free;
1510 deferred_free_range(free_base_page,
1511 free_base_pfn, nr_to_free);
1512 free_base_page = NULL;
1513 free_base_pfn = nr_to_free = 0;
1514
1515 page = pfn_to_page(pfn);
1516 cond_resched();
1517 }
1518
1519 if (page->flags) {
1520 VM_BUG_ON(page_zone(page) != zone);
1521 goto free_range;
1522 }
1523
1524 __init_single_page(page, pfn, zid, nid);
1525 if (!free_base_page) {
1526 free_base_page = page;
1527 free_base_pfn = pfn;
1528 nr_to_free = 0;
1529 }
1530 nr_to_free++;
1531
1532 /* Where possible, batch up pages for a single free */
1533 continue;
1534free_range:
1535 /* Free the current block of pages to allocator */
1536 nr_pages += nr_to_free;
1537 deferred_free_range(free_base_page, free_base_pfn,
1538 nr_to_free);
1539 free_base_page = NULL;
1540 free_base_pfn = nr_to_free = 0;
1541 }
1542 /* Free the last block of pages to allocator */
1543 nr_pages += nr_to_free;
1544 deferred_free_range(free_base_page, free_base_pfn, nr_to_free);
1545
1546 first_init_pfn = max(end_pfn, first_init_pfn);
1547 }
1548
1549 /* Sanity check that the next zone really is unpopulated */
1550 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1551
1552 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1553 jiffies_to_msecs(jiffies - start));
1554
1555 pgdat_init_report_one_done();
1556 return 0;
1557}
1558#endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1559
1560void __init page_alloc_init_late(void)
1561{
1562 struct zone *zone;
1563
1564#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1565 int nid;
1566
1567 /* There will be num_node_state(N_MEMORY) threads */
1568 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1569 for_each_node_state(nid, N_MEMORY) {
1570 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1571 }
1572
1573 /* Block until all are initialised */
1574 wait_for_completion(&pgdat_init_all_done_comp);
1575
1576 /* Reinit limits that are based on free pages after the kernel is up */
1577 files_maxfiles_init();
1578#endif
1579
1580 for_each_populated_zone(zone)
1581 set_zone_contiguous(zone);
1582}
1583
1584#ifdef CONFIG_CMA
1585/* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1586void __init init_cma_reserved_pageblock(struct page *page)
1587{
1588 unsigned i = pageblock_nr_pages;
1589 struct page *p = page;
1590
1591 do {
1592 __ClearPageReserved(p);
1593 set_page_count(p, 0);
1594 } while (++p, --i);
1595
1596 set_pageblock_migratetype(page, MIGRATE_CMA);
1597
1598 if (pageblock_order >= MAX_ORDER) {
1599 i = pageblock_nr_pages;
1600 p = page;
1601 do {
1602 set_page_refcounted(p);
1603 __free_pages(p, MAX_ORDER - 1);
1604 p += MAX_ORDER_NR_PAGES;
1605 } while (i -= MAX_ORDER_NR_PAGES);
1606 } else {
1607 set_page_refcounted(page);
1608 __free_pages(page, pageblock_order);
1609 }
1610
1611 adjust_managed_page_count(page, pageblock_nr_pages);
1612}
1613#endif
1614
1615/*
1616 * The order of subdivision here is critical for the IO subsystem.
1617 * Please do not alter this order without good reasons and regression
1618 * testing. Specifically, as large blocks of memory are subdivided,
1619 * the order in which smaller blocks are delivered depends on the order
1620 * they're subdivided in this function. This is the primary factor
1621 * influencing the order in which pages are delivered to the IO
1622 * subsystem according to empirical testing, and this is also justified
1623 * by considering the behavior of a buddy system containing a single
1624 * large block of memory acted on by a series of small allocations.
1625 * This behavior is a critical factor in sglist merging's success.
1626 *
1627 * -- nyc
1628 */
1629static inline void expand(struct zone *zone, struct page *page,
1630 int low, int high, struct free_area *area,
1631 int migratetype)
1632{
1633 unsigned long size = 1 << high;
1634
1635 while (high > low) {
1636 area--;
1637 high--;
1638 size >>= 1;
1639 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1640
1641 /*
1642 * Mark as guard pages (or page), that will allow to
1643 * merge back to allocator when buddy will be freed.
1644 * Corresponding page table entries will not be touched,
1645 * pages will stay not present in virtual address space
1646 */
1647 if (set_page_guard(zone, &page[size], high, migratetype))
1648 continue;
1649
1650 list_add(&page[size].lru, &area->free_list[migratetype]);
1651 area->nr_free++;
1652 set_page_order(&page[size], high);
1653 }
1654}
1655
1656static void check_new_page_bad(struct page *page)
1657{
1658 const char *bad_reason = NULL;
1659 unsigned long bad_flags = 0;
1660
1661 if (unlikely(atomic_read(&page->_mapcount) != -1))
1662 bad_reason = "nonzero mapcount";
1663 if (unlikely(page->mapping != NULL))
1664 bad_reason = "non-NULL mapping";
1665 if (unlikely(page_ref_count(page) != 0))
1666 bad_reason = "nonzero _count";
1667 if (unlikely(page->flags & __PG_HWPOISON)) {
1668 bad_reason = "HWPoisoned (hardware-corrupted)";
1669 bad_flags = __PG_HWPOISON;
1670 /* Don't complain about hwpoisoned pages */
1671 page_mapcount_reset(page); /* remove PageBuddy */
1672 return;
1673 }
1674 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1675 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1676 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1677 }
1678#ifdef CONFIG_MEMCG
1679 if (unlikely(page->mem_cgroup))
1680 bad_reason = "page still charged to cgroup";
1681#endif
1682 bad_page(page, bad_reason, bad_flags);
1683}
1684
1685/*
1686 * This page is about to be returned from the page allocator
1687 */
1688static inline int check_new_page(struct page *page)
1689{
1690 if (likely(page_expected_state(page,
1691 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
1692 return 0;
1693
1694 check_new_page_bad(page);
1695 return 1;
1696}
1697
1698static inline bool free_pages_prezeroed(bool poisoned)
1699{
1700 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
1701 page_poisoning_enabled() && poisoned;
1702}
1703
1704#ifdef CONFIG_DEBUG_VM
1705static bool check_pcp_refill(struct page *page)
1706{
1707 return false;
1708}
1709
1710static bool check_new_pcp(struct page *page)
1711{
1712 return check_new_page(page);
1713}
1714#else
1715static bool check_pcp_refill(struct page *page)
1716{
1717 return check_new_page(page);
1718}
1719static bool check_new_pcp(struct page *page)
1720{
1721 return false;
1722}
1723#endif /* CONFIG_DEBUG_VM */
1724
1725static bool check_new_pages(struct page *page, unsigned int order)
1726{
1727 int i;
1728 for (i = 0; i < (1 << order); i++) {
1729 struct page *p = page + i;
1730
1731 if (unlikely(check_new_page(p)))
1732 return true;
1733 }
1734
1735 return false;
1736}
1737
1738inline void post_alloc_hook(struct page *page, unsigned int order,
1739 gfp_t gfp_flags)
1740{
1741 set_page_private(page, 0);
1742 set_page_refcounted(page);
1743
1744 arch_alloc_page(page, order);
1745 kernel_map_pages(page, 1 << order, 1);
1746 kernel_poison_pages(page, 1 << order, 1);
1747 kasan_alloc_pages(page, order);
1748 set_page_owner(page, order, gfp_flags);
1749}
1750
1751static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1752 unsigned int alloc_flags)
1753{
1754 int i;
1755 bool poisoned = true;
1756
1757 for (i = 0; i < (1 << order); i++) {
1758 struct page *p = page + i;
1759 if (poisoned)
1760 poisoned &= page_is_poisoned(p);
1761 }
1762
1763 post_alloc_hook(page, order, gfp_flags);
1764
1765 if (!free_pages_prezeroed(poisoned) && (gfp_flags & __GFP_ZERO))
1766 for (i = 0; i < (1 << order); i++)
1767 clear_highpage(page + i);
1768
1769 if (order && (gfp_flags & __GFP_COMP))
1770 prep_compound_page(page, order);
1771
1772 /*
1773 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1774 * allocate the page. The expectation is that the caller is taking
1775 * steps that will free more memory. The caller should avoid the page
1776 * being used for !PFMEMALLOC purposes.
1777 */
1778 if (alloc_flags & ALLOC_NO_WATERMARKS)
1779 set_page_pfmemalloc(page);
1780 else
1781 clear_page_pfmemalloc(page);
1782}
1783
1784/*
1785 * Go through the free lists for the given migratetype and remove
1786 * the smallest available page from the freelists
1787 */
1788static inline
1789struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1790 int migratetype)
1791{
1792 unsigned int current_order;
1793 struct free_area *area;
1794 struct page *page;
1795
1796 /* Find a page of the appropriate size in the preferred list */
1797 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1798 area = &(zone->free_area[current_order]);
1799 page = list_first_entry_or_null(&area->free_list[migratetype],
1800 struct page, lru);
1801 if (!page)
1802 continue;
1803 list_del(&page->lru);
1804 rmv_page_order(page);
1805 area->nr_free--;
1806 expand(zone, page, order, current_order, area, migratetype);
1807 set_pcppage_migratetype(page, migratetype);
1808 return page;
1809 }
1810
1811 return NULL;
1812}
1813
1814
1815/*
1816 * This array describes the order lists are fallen back to when
1817 * the free lists for the desirable migrate type are depleted
1818 */
1819static int fallbacks[MIGRATE_TYPES][4] = {
1820 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1821 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1822 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1823#ifdef CONFIG_CMA
1824 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1825#endif
1826#ifdef CONFIG_MEMORY_ISOLATION
1827 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1828#endif
1829};
1830
1831#ifdef CONFIG_CMA
1832static struct page *__rmqueue_cma_fallback(struct zone *zone,
1833 unsigned int order)
1834{
1835 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1836}
1837#else
1838static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1839 unsigned int order) { return NULL; }
1840#endif
1841
1842/*
1843 * Move the free pages in a range to the free lists of the requested type.
1844 * Note that start_page and end_pages are not aligned on a pageblock
1845 * boundary. If alignment is required, use move_freepages_block()
1846 */
1847int move_freepages(struct zone *zone,
1848 struct page *start_page, struct page *end_page,
1849 int migratetype)
1850{
1851 struct page *page;
1852 unsigned int order;
1853 int pages_moved = 0;
1854
1855#ifndef CONFIG_HOLES_IN_ZONE
1856 /*
1857 * page_zone is not safe to call in this context when
1858 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1859 * anyway as we check zone boundaries in move_freepages_block().
1860 * Remove at a later date when no bug reports exist related to
1861 * grouping pages by mobility
1862 */
1863 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1864#endif
1865
1866 for (page = start_page; page <= end_page;) {
1867 if (!pfn_valid_within(page_to_pfn(page))) {
1868 page++;
1869 continue;
1870 }
1871
1872 /* Make sure we are not inadvertently changing nodes */
1873 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1874
1875 if (!PageBuddy(page)) {
1876 page++;
1877 continue;
1878 }
1879
1880 order = page_order(page);
1881 list_move(&page->lru,
1882 &zone->free_area[order].free_list[migratetype]);
1883 page += 1 << order;
1884 pages_moved += 1 << order;
1885 }
1886
1887 return pages_moved;
1888}
1889
1890int move_freepages_block(struct zone *zone, struct page *page,
1891 int migratetype)
1892{
1893 unsigned long start_pfn, end_pfn;
1894 struct page *start_page, *end_page;
1895
1896 start_pfn = page_to_pfn(page);
1897 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1898 start_page = pfn_to_page(start_pfn);
1899 end_page = start_page + pageblock_nr_pages - 1;
1900 end_pfn = start_pfn + pageblock_nr_pages - 1;
1901
1902 /* Do not cross zone boundaries */
1903 if (!zone_spans_pfn(zone, start_pfn))
1904 start_page = page;
1905 if (!zone_spans_pfn(zone, end_pfn))
1906 return 0;
1907
1908 return move_freepages(zone, start_page, end_page, migratetype);
1909}
1910
1911static void change_pageblock_range(struct page *pageblock_page,
1912 int start_order, int migratetype)
1913{
1914 int nr_pageblocks = 1 << (start_order - pageblock_order);
1915
1916 while (nr_pageblocks--) {
1917 set_pageblock_migratetype(pageblock_page, migratetype);
1918 pageblock_page += pageblock_nr_pages;
1919 }
1920}
1921
1922/*
1923 * When we are falling back to another migratetype during allocation, try to
1924 * steal extra free pages from the same pageblocks to satisfy further
1925 * allocations, instead of polluting multiple pageblocks.
1926 *
1927 * If we are stealing a relatively large buddy page, it is likely there will
1928 * be more free pages in the pageblock, so try to steal them all. For
1929 * reclaimable and unmovable allocations, we steal regardless of page size,
1930 * as fragmentation caused by those allocations polluting movable pageblocks
1931 * is worse than movable allocations stealing from unmovable and reclaimable
1932 * pageblocks.
1933 */
1934static bool can_steal_fallback(unsigned int order, int start_mt)
1935{
1936 /*
1937 * Leaving this order check is intended, although there is
1938 * relaxed order check in next check. The reason is that
1939 * we can actually steal whole pageblock if this condition met,
1940 * but, below check doesn't guarantee it and that is just heuristic
1941 * so could be changed anytime.
1942 */
1943 if (order >= pageblock_order)
1944 return true;
1945
1946 if (order >= pageblock_order / 2 ||
1947 start_mt == MIGRATE_RECLAIMABLE ||
1948 start_mt == MIGRATE_UNMOVABLE ||
1949 page_group_by_mobility_disabled)
1950 return true;
1951
1952 return false;
1953}
1954
1955/*
1956 * This function implements actual steal behaviour. If order is large enough,
1957 * we can steal whole pageblock. If not, we first move freepages in this
1958 * pageblock and check whether half of pages are moved or not. If half of
1959 * pages are moved, we can change migratetype of pageblock and permanently
1960 * use it's pages as requested migratetype in the future.
1961 */
1962static void steal_suitable_fallback(struct zone *zone, struct page *page,
1963 int start_type)
1964{
1965 unsigned int current_order = page_order(page);
1966 int pages;
1967
1968 /* Take ownership for orders >= pageblock_order */
1969 if (current_order >= pageblock_order) {
1970 change_pageblock_range(page, current_order, start_type);
1971 return;
1972 }
1973
1974 pages = move_freepages_block(zone, page, start_type);
1975
1976 /* Claim the whole block if over half of it is free */
1977 if (pages >= (1 << (pageblock_order-1)) ||
1978 page_group_by_mobility_disabled)
1979 set_pageblock_migratetype(page, start_type);
1980}
1981
1982/*
1983 * Check whether there is a suitable fallback freepage with requested order.
1984 * If only_stealable is true, this function returns fallback_mt only if
1985 * we can steal other freepages all together. This would help to reduce
1986 * fragmentation due to mixed migratetype pages in one pageblock.
1987 */
1988int find_suitable_fallback(struct free_area *area, unsigned int order,
1989 int migratetype, bool only_stealable, bool *can_steal)
1990{
1991 int i;
1992 int fallback_mt;
1993
1994 if (area->nr_free == 0)
1995 return -1;
1996
1997 *can_steal = false;
1998 for (i = 0;; i++) {
1999 fallback_mt = fallbacks[migratetype][i];
2000 if (fallback_mt == MIGRATE_TYPES)
2001 break;
2002
2003 if (list_empty(&area->free_list[fallback_mt]))
2004 continue;
2005
2006 if (can_steal_fallback(order, migratetype))
2007 *can_steal = true;
2008
2009 if (!only_stealable)
2010 return fallback_mt;
2011
2012 if (*can_steal)
2013 return fallback_mt;
2014 }
2015
2016 return -1;
2017}
2018
2019/*
2020 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2021 * there are no empty page blocks that contain a page with a suitable order
2022 */
2023static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2024 unsigned int alloc_order)
2025{
2026 int mt;
2027 unsigned long max_managed, flags;
2028
2029 /*
2030 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2031 * Check is race-prone but harmless.
2032 */
2033 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
2034 if (zone->nr_reserved_highatomic >= max_managed)
2035 return;
2036
2037 spin_lock_irqsave(&zone->lock, flags);
2038
2039 /* Recheck the nr_reserved_highatomic limit under the lock */
2040 if (zone->nr_reserved_highatomic >= max_managed)
2041 goto out_unlock;
2042
2043 /* Yoink! */
2044 mt = get_pageblock_migratetype(page);
2045 if (mt != MIGRATE_HIGHATOMIC &&
2046 !is_migrate_isolate(mt) && !is_migrate_cma(mt)) {
2047 zone->nr_reserved_highatomic += pageblock_nr_pages;
2048 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2049 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC);
2050 }
2051
2052out_unlock:
2053 spin_unlock_irqrestore(&zone->lock, flags);
2054}
2055
2056/*
2057 * Used when an allocation is about to fail under memory pressure. This
2058 * potentially hurts the reliability of high-order allocations when under
2059 * intense memory pressure but failed atomic allocations should be easier
2060 * to recover from than an OOM.
2061 *
2062 * If @force is true, try to unreserve a pageblock even though highatomic
2063 * pageblock is exhausted.
2064 */
2065static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2066 bool force)
2067{
2068 struct zonelist *zonelist = ac->zonelist;
2069 unsigned long flags;
2070 struct zoneref *z;
2071 struct zone *zone;
2072 struct page *page;
2073 int order;
2074 bool ret;
2075
2076 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2077 ac->nodemask) {
2078 /*
2079 * Preserve at least one pageblock unless memory pressure
2080 * is really high.
2081 */
2082 if (!force && zone->nr_reserved_highatomic <=
2083 pageblock_nr_pages)
2084 continue;
2085
2086 spin_lock_irqsave(&zone->lock, flags);
2087 for (order = 0; order < MAX_ORDER; order++) {
2088 struct free_area *area = &(zone->free_area[order]);
2089
2090 page = list_first_entry_or_null(
2091 &area->free_list[MIGRATE_HIGHATOMIC],
2092 struct page, lru);
2093 if (!page)
2094 continue;
2095
2096 /*
2097 * In page freeing path, migratetype change is racy so
2098 * we can counter several free pages in a pageblock
2099 * in this loop althoug we changed the pageblock type
2100 * from highatomic to ac->migratetype. So we should
2101 * adjust the count once.
2102 */
2103 if (get_pageblock_migratetype(page) ==
2104 MIGRATE_HIGHATOMIC) {
2105 /*
2106 * It should never happen but changes to
2107 * locking could inadvertently allow a per-cpu
2108 * drain to add pages to MIGRATE_HIGHATOMIC
2109 * while unreserving so be safe and watch for
2110 * underflows.
2111 */
2112 zone->nr_reserved_highatomic -= min(
2113 pageblock_nr_pages,
2114 zone->nr_reserved_highatomic);
2115 }
2116
2117 /*
2118 * Convert to ac->migratetype and avoid the normal
2119 * pageblock stealing heuristics. Minimally, the caller
2120 * is doing the work and needs the pages. More
2121 * importantly, if the block was always converted to
2122 * MIGRATE_UNMOVABLE or another type then the number
2123 * of pageblocks that cannot be completely freed
2124 * may increase.
2125 */
2126 set_pageblock_migratetype(page, ac->migratetype);
2127 ret = move_freepages_block(zone, page, ac->migratetype);
2128 if (ret) {
2129 spin_unlock_irqrestore(&zone->lock, flags);
2130 return ret;
2131 }
2132 }
2133 spin_unlock_irqrestore(&zone->lock, flags);
2134 }
2135
2136 return false;
2137}
2138
2139/* Remove an element from the buddy allocator from the fallback list */
2140static inline struct page *
2141__rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
2142{
2143 struct free_area *area;
2144 unsigned int current_order;
2145 struct page *page;
2146 int fallback_mt;
2147 bool can_steal;
2148
2149 /* Find the largest possible block of pages in the other list */
2150 for (current_order = MAX_ORDER-1;
2151 current_order >= order && current_order <= MAX_ORDER-1;
2152 --current_order) {
2153 area = &(zone->free_area[current_order]);
2154 fallback_mt = find_suitable_fallback(area, current_order,
2155 start_migratetype, false, &can_steal);
2156 if (fallback_mt == -1)
2157 continue;
2158
2159 page = list_first_entry(&area->free_list[fallback_mt],
2160 struct page, lru);
2161 if (can_steal &&
2162 get_pageblock_migratetype(page) != MIGRATE_HIGHATOMIC)
2163 steal_suitable_fallback(zone, page, start_migratetype);
2164
2165 /* Remove the page from the freelists */
2166 area->nr_free--;
2167 list_del(&page->lru);
2168 rmv_page_order(page);
2169
2170 expand(zone, page, order, current_order, area,
2171 start_migratetype);
2172 /*
2173 * The pcppage_migratetype may differ from pageblock's
2174 * migratetype depending on the decisions in
2175 * find_suitable_fallback(). This is OK as long as it does not
2176 * differ for MIGRATE_CMA pageblocks. Those can be used as
2177 * fallback only via special __rmqueue_cma_fallback() function
2178 */
2179 set_pcppage_migratetype(page, start_migratetype);
2180
2181 trace_mm_page_alloc_extfrag(page, order, current_order,
2182 start_migratetype, fallback_mt);
2183
2184 return page;
2185 }
2186
2187 return NULL;
2188}
2189
2190/*
2191 * Do the hard work of removing an element from the buddy allocator.
2192 * Call me with the zone->lock already held.
2193 */
2194static struct page *__rmqueue(struct zone *zone, unsigned int order,
2195 int migratetype)
2196{
2197 struct page *page;
2198
2199 page = __rmqueue_smallest(zone, order, migratetype);
2200 if (unlikely(!page)) {
2201 if (migratetype == MIGRATE_MOVABLE)
2202 page = __rmqueue_cma_fallback(zone, order);
2203
2204 if (!page)
2205 page = __rmqueue_fallback(zone, order, migratetype);
2206 }
2207
2208 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2209 return page;
2210}
2211
2212/*
2213 * Obtain a specified number of elements from the buddy allocator, all under
2214 * a single hold of the lock, for efficiency. Add them to the supplied list.
2215 * Returns the number of new pages which were placed at *list.
2216 */
2217static int rmqueue_bulk(struct zone *zone, unsigned int order,
2218 unsigned long count, struct list_head *list,
2219 int migratetype, bool cold)
2220{
2221 int i, alloced = 0;
2222
2223 spin_lock(&zone->lock);
2224 for (i = 0; i < count; ++i) {
2225 struct page *page = __rmqueue(zone, order, migratetype);
2226 if (unlikely(page == NULL))
2227 break;
2228
2229 if (unlikely(check_pcp_refill(page)))
2230 continue;
2231
2232 /*
2233 * Split buddy pages returned by expand() are received here
2234 * in physical page order. The page is added to the callers and
2235 * list and the list head then moves forward. From the callers
2236 * perspective, the linked list is ordered by page number in
2237 * some conditions. This is useful for IO devices that can
2238 * merge IO requests if the physical pages are ordered
2239 * properly.
2240 */
2241 if (likely(!cold))
2242 list_add(&page->lru, list);
2243 else
2244 list_add_tail(&page->lru, list);
2245 list = &page->lru;
2246 alloced++;
2247 if (is_migrate_cma(get_pcppage_migratetype(page)))
2248 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2249 -(1 << order));
2250 }
2251
2252 /*
2253 * i pages were removed from the buddy list even if some leak due
2254 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2255 * on i. Do not confuse with 'alloced' which is the number of
2256 * pages added to the pcp list.
2257 */
2258 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2259 spin_unlock(&zone->lock);
2260 return alloced;
2261}
2262
2263#ifdef CONFIG_NUMA
2264/*
2265 * Called from the vmstat counter updater to drain pagesets of this
2266 * currently executing processor on remote nodes after they have
2267 * expired.
2268 *
2269 * Note that this function must be called with the thread pinned to
2270 * a single processor.
2271 */
2272void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2273{
2274 unsigned long flags;
2275 int to_drain, batch;
2276
2277 local_irq_save(flags);
2278 batch = READ_ONCE(pcp->batch);
2279 to_drain = min(pcp->count, batch);
2280 if (to_drain > 0) {
2281 free_pcppages_bulk(zone, to_drain, pcp);
2282 pcp->count -= to_drain;
2283 }
2284 local_irq_restore(flags);
2285}
2286#endif
2287
2288/*
2289 * Drain pcplists of the indicated processor and zone.
2290 *
2291 * The processor must either be the current processor and the
2292 * thread pinned to the current processor or a processor that
2293 * is not online.
2294 */
2295static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2296{
2297 unsigned long flags;
2298 struct per_cpu_pageset *pset;
2299 struct per_cpu_pages *pcp;
2300
2301 local_irq_save(flags);
2302 pset = per_cpu_ptr(zone->pageset, cpu);
2303
2304 pcp = &pset->pcp;
2305 if (pcp->count) {
2306 free_pcppages_bulk(zone, pcp->count, pcp);
2307 pcp->count = 0;
2308 }
2309 local_irq_restore(flags);
2310}
2311
2312/*
2313 * Drain pcplists of all zones on the indicated processor.
2314 *
2315 * The processor must either be the current processor and the
2316 * thread pinned to the current processor or a processor that
2317 * is not online.
2318 */
2319static void drain_pages(unsigned int cpu)
2320{
2321 struct zone *zone;
2322
2323 for_each_populated_zone(zone) {
2324 drain_pages_zone(cpu, zone);
2325 }
2326}
2327
2328/*
2329 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2330 *
2331 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2332 * the single zone's pages.
2333 */
2334void drain_local_pages(struct zone *zone)
2335{
2336 int cpu = smp_processor_id();
2337
2338 if (zone)
2339 drain_pages_zone(cpu, zone);
2340 else
2341 drain_pages(cpu);
2342}
2343
2344/*
2345 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2346 *
2347 * When zone parameter is non-NULL, spill just the single zone's pages.
2348 *
2349 * Note that this code is protected against sending an IPI to an offline
2350 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
2351 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
2352 * nothing keeps CPUs from showing up after we populated the cpumask and
2353 * before the call to on_each_cpu_mask().
2354 */
2355void drain_all_pages(struct zone *zone)
2356{
2357 int cpu;
2358
2359 /*
2360 * Allocate in the BSS so we wont require allocation in
2361 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2362 */
2363 static cpumask_t cpus_with_pcps;
2364
2365 /*
2366 * We don't care about racing with CPU hotplug event
2367 * as offline notification will cause the notified
2368 * cpu to drain that CPU pcps and on_each_cpu_mask
2369 * disables preemption as part of its processing
2370 */
2371 for_each_online_cpu(cpu) {
2372 struct per_cpu_pageset *pcp;
2373 struct zone *z;
2374 bool has_pcps = false;
2375
2376 if (zone) {
2377 pcp = per_cpu_ptr(zone->pageset, cpu);
2378 if (pcp->pcp.count)
2379 has_pcps = true;
2380 } else {
2381 for_each_populated_zone(z) {
2382 pcp = per_cpu_ptr(z->pageset, cpu);
2383 if (pcp->pcp.count) {
2384 has_pcps = true;
2385 break;
2386 }
2387 }
2388 }
2389
2390 if (has_pcps)
2391 cpumask_set_cpu(cpu, &cpus_with_pcps);
2392 else
2393 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2394 }
2395 on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages,
2396 zone, 1);
2397}
2398
2399#ifdef CONFIG_HIBERNATION
2400
2401void mark_free_pages(struct zone *zone)
2402{
2403 unsigned long pfn, max_zone_pfn;
2404 unsigned long flags;
2405 unsigned int order, t;
2406 struct page *page;
2407
2408 if (zone_is_empty(zone))
2409 return;
2410
2411 spin_lock_irqsave(&zone->lock, flags);
2412
2413 max_zone_pfn = zone_end_pfn(zone);
2414 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2415 if (pfn_valid(pfn)) {
2416 page = pfn_to_page(pfn);
2417
2418 if (page_zone(page) != zone)
2419 continue;
2420
2421 if (!swsusp_page_is_forbidden(page))
2422 swsusp_unset_page_free(page);
2423 }
2424
2425 for_each_migratetype_order(order, t) {
2426 list_for_each_entry(page,
2427 &zone->free_area[order].free_list[t], lru) {
2428 unsigned long i;
2429
2430 pfn = page_to_pfn(page);
2431 for (i = 0; i < (1UL << order); i++)
2432 swsusp_set_page_free(pfn_to_page(pfn + i));
2433 }
2434 }
2435 spin_unlock_irqrestore(&zone->lock, flags);
2436}
2437#endif /* CONFIG_PM */
2438
2439/*
2440 * Free a 0-order page
2441 * cold == true ? free a cold page : free a hot page
2442 */
2443void free_hot_cold_page(struct page *page, bool cold)
2444{
2445 struct zone *zone = page_zone(page);
2446 struct per_cpu_pages *pcp;
2447 unsigned long flags;
2448 unsigned long pfn = page_to_pfn(page);
2449 int migratetype;
2450
2451 if (!free_pcp_prepare(page))
2452 return;
2453
2454 migratetype = get_pfnblock_migratetype(page, pfn);
2455 set_pcppage_migratetype(page, migratetype);
2456 local_irq_save(flags);
2457 __count_vm_event(PGFREE);
2458
2459 /*
2460 * We only track unmovable, reclaimable and movable on pcp lists.
2461 * Free ISOLATE pages back to the allocator because they are being
2462 * offlined but treat RESERVE as movable pages so we can get those
2463 * areas back if necessary. Otherwise, we may have to free
2464 * excessively into the page allocator
2465 */
2466 if (migratetype >= MIGRATE_PCPTYPES) {
2467 if (unlikely(is_migrate_isolate(migratetype))) {
2468 free_one_page(zone, page, pfn, 0, migratetype);
2469 goto out;
2470 }
2471 migratetype = MIGRATE_MOVABLE;
2472 }
2473
2474 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2475 if (!cold)
2476 list_add(&page->lru, &pcp->lists[migratetype]);
2477 else
2478 list_add_tail(&page->lru, &pcp->lists[migratetype]);
2479 pcp->count++;
2480 if (pcp->count >= pcp->high) {
2481 unsigned long batch = READ_ONCE(pcp->batch);
2482 free_pcppages_bulk(zone, batch, pcp);
2483 pcp->count -= batch;
2484 }
2485
2486out:
2487 local_irq_restore(flags);
2488}
2489
2490/*
2491 * Free a list of 0-order pages
2492 */
2493void free_hot_cold_page_list(struct list_head *list, bool cold)
2494{
2495 struct page *page, *next;
2496
2497 list_for_each_entry_safe(page, next, list, lru) {
2498 trace_mm_page_free_batched(page, cold);
2499 free_hot_cold_page(page, cold);
2500 }
2501}
2502
2503/*
2504 * split_page takes a non-compound higher-order page, and splits it into
2505 * n (1<<order) sub-pages: page[0..n]
2506 * Each sub-page must be freed individually.
2507 *
2508 * Note: this is probably too low level an operation for use in drivers.
2509 * Please consult with lkml before using this in your driver.
2510 */
2511void split_page(struct page *page, unsigned int order)
2512{
2513 int i;
2514
2515 VM_BUG_ON_PAGE(PageCompound(page), page);
2516 VM_BUG_ON_PAGE(!page_count(page), page);
2517
2518#ifdef CONFIG_KMEMCHECK
2519 /*
2520 * Split shadow pages too, because free(page[0]) would
2521 * otherwise free the whole shadow.
2522 */
2523 if (kmemcheck_page_is_tracked(page))
2524 split_page(virt_to_page(page[0].shadow), order);
2525#endif
2526
2527 for (i = 1; i < (1 << order); i++)
2528 set_page_refcounted(page + i);
2529 split_page_owner(page, order);
2530}
2531EXPORT_SYMBOL_GPL(split_page);
2532
2533int __isolate_free_page(struct page *page, unsigned int order)
2534{
2535 unsigned long watermark;
2536 struct zone *zone;
2537 int mt;
2538
2539 BUG_ON(!PageBuddy(page));
2540
2541 zone = page_zone(page);
2542 mt = get_pageblock_migratetype(page);
2543
2544 if (!is_migrate_isolate(mt)) {
2545 /*
2546 * Obey watermarks as if the page was being allocated. We can
2547 * emulate a high-order watermark check with a raised order-0
2548 * watermark, because we already know our high-order page
2549 * exists.
2550 */
2551 watermark = min_wmark_pages(zone) + (1UL << order);
2552 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
2553 return 0;
2554
2555 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2556 }
2557
2558 /* Remove page from free list */
2559 list_del(&page->lru);
2560 zone->free_area[order].nr_free--;
2561 rmv_page_order(page);
2562
2563 /*
2564 * Set the pageblock if the isolated page is at least half of a
2565 * pageblock
2566 */
2567 if (order >= pageblock_order - 1) {
2568 struct page *endpage = page + (1 << order) - 1;
2569 for (; page < endpage; page += pageblock_nr_pages) {
2570 int mt = get_pageblock_migratetype(page);
2571 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
2572 && mt != MIGRATE_HIGHATOMIC)
2573 set_pageblock_migratetype(page,
2574 MIGRATE_MOVABLE);
2575 }
2576 }
2577
2578
2579 return 1UL << order;
2580}
2581
2582/*
2583 * Update NUMA hit/miss statistics
2584 *
2585 * Must be called with interrupts disabled.
2586 */
2587static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
2588{
2589#ifdef CONFIG_NUMA
2590 enum zone_stat_item local_stat = NUMA_LOCAL;
2591
2592 if (z->node != numa_node_id())
2593 local_stat = NUMA_OTHER;
2594
2595 if (z->node == preferred_zone->node)
2596 __inc_zone_state(z, NUMA_HIT);
2597 else {
2598 __inc_zone_state(z, NUMA_MISS);
2599 __inc_zone_state(preferred_zone, NUMA_FOREIGN);
2600 }
2601 __inc_zone_state(z, local_stat);
2602#endif
2603}
2604
2605/*
2606 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2607 */
2608static inline
2609struct page *buffered_rmqueue(struct zone *preferred_zone,
2610 struct zone *zone, unsigned int order,
2611 gfp_t gfp_flags, unsigned int alloc_flags,
2612 int migratetype)
2613{
2614 unsigned long flags;
2615 struct page *page;
2616 bool cold = ((gfp_flags & __GFP_COLD) != 0);
2617
2618 if (likely(order == 0)) {
2619 struct per_cpu_pages *pcp;
2620 struct list_head *list;
2621
2622 local_irq_save(flags);
2623 do {
2624 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2625 list = &pcp->lists[migratetype];
2626 if (list_empty(list)) {
2627 pcp->count += rmqueue_bulk(zone, 0,
2628 pcp->batch, list,
2629 migratetype, cold);
2630 if (unlikely(list_empty(list)))
2631 goto failed;
2632 }
2633
2634 if (cold)
2635 page = list_last_entry(list, struct page, lru);
2636 else
2637 page = list_first_entry(list, struct page, lru);
2638
2639 list_del(&page->lru);
2640 pcp->count--;
2641
2642 } while (check_new_pcp(page));
2643 } else {
2644 /*
2645 * We most definitely don't want callers attempting to
2646 * allocate greater than order-1 page units with __GFP_NOFAIL.
2647 */
2648 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
2649 spin_lock_irqsave(&zone->lock, flags);
2650
2651 do {
2652 page = NULL;
2653 if (alloc_flags & ALLOC_HARDER) {
2654 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2655 if (page)
2656 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2657 }
2658 if (!page)
2659 page = __rmqueue(zone, order, migratetype);
2660 } while (page && check_new_pages(page, order));
2661 spin_unlock(&zone->lock);
2662 if (!page)
2663 goto failed;
2664 __mod_zone_freepage_state(zone, -(1 << order),
2665 get_pcppage_migratetype(page));
2666 }
2667
2668 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2669 zone_statistics(preferred_zone, zone);
2670 local_irq_restore(flags);
2671
2672 VM_BUG_ON_PAGE(bad_range(zone, page), page);
2673 return page;
2674
2675failed:
2676 local_irq_restore(flags);
2677 return NULL;
2678}
2679
2680#ifdef CONFIG_FAIL_PAGE_ALLOC
2681
2682static struct {
2683 struct fault_attr attr;
2684
2685 bool ignore_gfp_highmem;
2686 bool ignore_gfp_reclaim;
2687 u32 min_order;
2688} fail_page_alloc = {
2689 .attr = FAULT_ATTR_INITIALIZER,
2690 .ignore_gfp_reclaim = true,
2691 .ignore_gfp_highmem = true,
2692 .min_order = 1,
2693};
2694
2695static int __init setup_fail_page_alloc(char *str)
2696{
2697 return setup_fault_attr(&fail_page_alloc.attr, str);
2698}
2699__setup("fail_page_alloc=", setup_fail_page_alloc);
2700
2701static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2702{
2703 if (order < fail_page_alloc.min_order)
2704 return false;
2705 if (gfp_mask & __GFP_NOFAIL)
2706 return false;
2707 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2708 return false;
2709 if (fail_page_alloc.ignore_gfp_reclaim &&
2710 (gfp_mask & __GFP_DIRECT_RECLAIM))
2711 return false;
2712
2713 return should_fail(&fail_page_alloc.attr, 1 << order);
2714}
2715
2716#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2717
2718static int __init fail_page_alloc_debugfs(void)
2719{
2720 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2721 struct dentry *dir;
2722
2723 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2724 &fail_page_alloc.attr);
2725 if (IS_ERR(dir))
2726 return PTR_ERR(dir);
2727
2728 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2729 &fail_page_alloc.ignore_gfp_reclaim))
2730 goto fail;
2731 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2732 &fail_page_alloc.ignore_gfp_highmem))
2733 goto fail;
2734 if (!debugfs_create_u32("min-order", mode, dir,
2735 &fail_page_alloc.min_order))
2736 goto fail;
2737
2738 return 0;
2739fail:
2740 debugfs_remove_recursive(dir);
2741
2742 return -ENOMEM;
2743}
2744
2745late_initcall(fail_page_alloc_debugfs);
2746
2747#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2748
2749#else /* CONFIG_FAIL_PAGE_ALLOC */
2750
2751static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2752{
2753 return false;
2754}
2755
2756#endif /* CONFIG_FAIL_PAGE_ALLOC */
2757
2758/*
2759 * Return true if free base pages are above 'mark'. For high-order checks it
2760 * will return true of the order-0 watermark is reached and there is at least
2761 * one free page of a suitable size. Checking now avoids taking the zone lock
2762 * to check in the allocation paths if no pages are free.
2763 */
2764bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2765 int classzone_idx, unsigned int alloc_flags,
2766 long free_pages)
2767{
2768 long min = mark;
2769 int o;
2770 const bool alloc_harder = (alloc_flags & ALLOC_HARDER);
2771
2772 /* free_pages may go negative - that's OK */
2773 free_pages -= (1 << order) - 1;
2774
2775 if (alloc_flags & ALLOC_HIGH)
2776 min -= min / 2;
2777
2778 /*
2779 * If the caller does not have rights to ALLOC_HARDER then subtract
2780 * the high-atomic reserves. This will over-estimate the size of the
2781 * atomic reserve but it avoids a search.
2782 */
2783 if (likely(!alloc_harder))
2784 free_pages -= z->nr_reserved_highatomic;
2785 else
2786 min -= min / 4;
2787
2788#ifdef CONFIG_CMA
2789 /* If allocation can't use CMA areas don't use free CMA pages */
2790 if (!(alloc_flags & ALLOC_CMA))
2791 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
2792#endif
2793
2794 /*
2795 * Check watermarks for an order-0 allocation request. If these
2796 * are not met, then a high-order request also cannot go ahead
2797 * even if a suitable page happened to be free.
2798 */
2799 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
2800 return false;
2801
2802 /* If this is an order-0 request then the watermark is fine */
2803 if (!order)
2804 return true;
2805
2806 /* For a high-order request, check at least one suitable page is free */
2807 for (o = order; o < MAX_ORDER; o++) {
2808 struct free_area *area = &z->free_area[o];
2809 int mt;
2810
2811 if (!area->nr_free)
2812 continue;
2813
2814 if (alloc_harder)
2815 return true;
2816
2817 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
2818 if (!list_empty(&area->free_list[mt]))
2819 return true;
2820 }
2821
2822#ifdef CONFIG_CMA
2823 if ((alloc_flags & ALLOC_CMA) &&
2824 !list_empty(&area->free_list[MIGRATE_CMA])) {
2825 return true;
2826 }
2827#endif
2828 }
2829 return false;
2830}
2831
2832bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2833 int classzone_idx, unsigned int alloc_flags)
2834{
2835 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2836 zone_page_state(z, NR_FREE_PAGES));
2837}
2838
2839static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
2840 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
2841{
2842 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2843 long cma_pages = 0;
2844
2845#ifdef CONFIG_CMA
2846 /* If allocation can't use CMA areas don't use free CMA pages */
2847 if (!(alloc_flags & ALLOC_CMA))
2848 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
2849#endif
2850
2851 /*
2852 * Fast check for order-0 only. If this fails then the reserves
2853 * need to be calculated. There is a corner case where the check
2854 * passes but only the high-order atomic reserve are free. If
2855 * the caller is !atomic then it'll uselessly search the free
2856 * list. That corner case is then slower but it is harmless.
2857 */
2858 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
2859 return true;
2860
2861 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2862 free_pages);
2863}
2864
2865bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
2866 unsigned long mark, int classzone_idx)
2867{
2868 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2869
2870 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
2871 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
2872
2873 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
2874 free_pages);
2875}
2876
2877#ifdef CONFIG_NUMA
2878static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2879{
2880 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
2881 RECLAIM_DISTANCE;
2882}
2883#else /* CONFIG_NUMA */
2884static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2885{
2886 return true;
2887}
2888#endif /* CONFIG_NUMA */
2889
2890/*
2891 * get_page_from_freelist goes through the zonelist trying to allocate
2892 * a page.
2893 */
2894static struct page *
2895get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
2896 const struct alloc_context *ac)
2897{
2898 struct zoneref *z = ac->preferred_zoneref;
2899 struct zone *zone;
2900 struct pglist_data *last_pgdat_dirty_limit = NULL;
2901
2902 /*
2903 * Scan zonelist, looking for a zone with enough free.
2904 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2905 */
2906 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
2907 ac->nodemask) {
2908 struct page *page;
2909 unsigned long mark;
2910
2911 if (cpusets_enabled() &&
2912 (alloc_flags & ALLOC_CPUSET) &&
2913 !__cpuset_zone_allowed(zone, gfp_mask))
2914 continue;
2915 /*
2916 * When allocating a page cache page for writing, we
2917 * want to get it from a node that is within its dirty
2918 * limit, such that no single node holds more than its
2919 * proportional share of globally allowed dirty pages.
2920 * The dirty limits take into account the node's
2921 * lowmem reserves and high watermark so that kswapd
2922 * should be able to balance it without having to
2923 * write pages from its LRU list.
2924 *
2925 * XXX: For now, allow allocations to potentially
2926 * exceed the per-node dirty limit in the slowpath
2927 * (spread_dirty_pages unset) before going into reclaim,
2928 * which is important when on a NUMA setup the allowed
2929 * nodes are together not big enough to reach the
2930 * global limit. The proper fix for these situations
2931 * will require awareness of nodes in the
2932 * dirty-throttling and the flusher threads.
2933 */
2934 if (ac->spread_dirty_pages) {
2935 if (last_pgdat_dirty_limit == zone->zone_pgdat)
2936 continue;
2937
2938 if (!node_dirty_ok(zone->zone_pgdat)) {
2939 last_pgdat_dirty_limit = zone->zone_pgdat;
2940 continue;
2941 }
2942 }
2943
2944 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2945 if (!zone_watermark_fast(zone, order, mark,
2946 ac_classzone_idx(ac), alloc_flags)) {
2947 int ret;
2948
2949 /* Checked here to keep the fast path fast */
2950 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2951 if (alloc_flags & ALLOC_NO_WATERMARKS)
2952 goto try_this_zone;
2953
2954 if (node_reclaim_mode == 0 ||
2955 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
2956 continue;
2957
2958 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
2959 switch (ret) {
2960 case NODE_RECLAIM_NOSCAN:
2961 /* did not scan */
2962 continue;
2963 case NODE_RECLAIM_FULL:
2964 /* scanned but unreclaimable */
2965 continue;
2966 default:
2967 /* did we reclaim enough */
2968 if (zone_watermark_ok(zone, order, mark,
2969 ac_classzone_idx(ac), alloc_flags))
2970 goto try_this_zone;
2971
2972 continue;
2973 }
2974 }
2975
2976try_this_zone:
2977 page = buffered_rmqueue(ac->preferred_zoneref->zone, zone, order,
2978 gfp_mask, alloc_flags, ac->migratetype);
2979 if (page) {
2980 prep_new_page(page, order, gfp_mask, alloc_flags);
2981
2982 /*
2983 * If this is a high-order atomic allocation then check
2984 * if the pageblock should be reserved for the future
2985 */
2986 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
2987 reserve_highatomic_pageblock(page, zone, order);
2988
2989 return page;
2990 }
2991 }
2992
2993 return NULL;
2994}
2995
2996/*
2997 * Large machines with many possible nodes should not always dump per-node
2998 * meminfo in irq context.
2999 */
3000static inline bool should_suppress_show_mem(void)
3001{
3002 bool ret = false;
3003
3004#if NODES_SHIFT > 8
3005 ret = in_interrupt();
3006#endif
3007 return ret;
3008}
3009
3010static DEFINE_RATELIMIT_STATE(nopage_rs,
3011 DEFAULT_RATELIMIT_INTERVAL,
3012 DEFAULT_RATELIMIT_BURST);
3013
3014void warn_alloc(gfp_t gfp_mask, const char *fmt, ...)
3015{
3016 unsigned int filter = SHOW_MEM_FILTER_NODES;
3017 struct va_format vaf;
3018 va_list args;
3019
3020 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
3021 debug_guardpage_minorder() > 0)
3022 return;
3023
3024 /*
3025 * This documents exceptions given to allocations in certain
3026 * contexts that are allowed to allocate outside current's set
3027 * of allowed nodes.
3028 */
3029 if (!(gfp_mask & __GFP_NOMEMALLOC))
3030 if (test_thread_flag(TIF_MEMDIE) ||
3031 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3032 filter &= ~SHOW_MEM_FILTER_NODES;
3033 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3034 filter &= ~SHOW_MEM_FILTER_NODES;
3035
3036 pr_warn("%s: ", current->comm);
3037
3038 va_start(args, fmt);
3039 vaf.fmt = fmt;
3040 vaf.va = &args;
3041 pr_cont("%pV", &vaf);
3042 va_end(args);
3043
3044 pr_cont(", mode:%#x(%pGg)\n", gfp_mask, &gfp_mask);
3045
3046 dump_stack();
3047 if (!should_suppress_show_mem())
3048 show_mem(filter);
3049}
3050
3051static inline struct page *
3052__alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3053 const struct alloc_context *ac, unsigned long *did_some_progress)
3054{
3055 struct oom_control oc = {
3056 .zonelist = ac->zonelist,
3057 .nodemask = ac->nodemask,
3058 .memcg = NULL,
3059 .gfp_mask = gfp_mask,
3060 .order = order,
3061 };
3062 struct page *page;
3063
3064 *did_some_progress = 0;
3065
3066 /*
3067 * Acquire the oom lock. If that fails, somebody else is
3068 * making progress for us.
3069 */
3070 if (!mutex_trylock(&oom_lock)) {
3071 *did_some_progress = 1;
3072 schedule_timeout_uninterruptible(1);
3073 return NULL;
3074 }
3075
3076 /*
3077 * Go through the zonelist yet one more time, keep very high watermark
3078 * here, this is only to catch a parallel oom killing, we must fail if
3079 * we're still under heavy pressure.
3080 */
3081 page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order,
3082 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3083 if (page)
3084 goto out;
3085
3086 if (!(gfp_mask & __GFP_NOFAIL)) {
3087 /* Coredumps can quickly deplete all memory reserves */
3088 if (current->flags & PF_DUMPCORE)
3089 goto out;
3090 /* The OOM killer will not help higher order allocs */
3091 if (order > PAGE_ALLOC_COSTLY_ORDER)
3092 goto out;
3093 /* The OOM killer does not needlessly kill tasks for lowmem */
3094 if (ac->high_zoneidx < ZONE_NORMAL)
3095 goto out;
3096 if (pm_suspended_storage())
3097 goto out;
3098 /*
3099 * XXX: GFP_NOFS allocations should rather fail than rely on
3100 * other request to make a forward progress.
3101 * We are in an unfortunate situation where out_of_memory cannot
3102 * do much for this context but let's try it to at least get
3103 * access to memory reserved if the current task is killed (see
3104 * out_of_memory). Once filesystems are ready to handle allocation
3105 * failures more gracefully we should just bail out here.
3106 */
3107
3108 /* The OOM killer may not free memory on a specific node */
3109 if (gfp_mask & __GFP_THISNODE)
3110 goto out;
3111 }
3112 /* Exhausted what can be done so it's blamo time */
3113 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3114 *did_some_progress = 1;
3115
3116 if (gfp_mask & __GFP_NOFAIL) {
3117 page = get_page_from_freelist(gfp_mask, order,
3118 ALLOC_NO_WATERMARKS|ALLOC_CPUSET, ac);
3119 /*
3120 * fallback to ignore cpuset restriction if our nodes
3121 * are depleted
3122 */
3123 if (!page)
3124 page = get_page_from_freelist(gfp_mask, order,
3125 ALLOC_NO_WATERMARKS, ac);
3126 }
3127 }
3128out:
3129 mutex_unlock(&oom_lock);
3130 return page;
3131}
3132
3133/*
3134 * Maximum number of compaction retries wit a progress before OOM
3135 * killer is consider as the only way to move forward.
3136 */
3137#define MAX_COMPACT_RETRIES 16
3138
3139#ifdef CONFIG_COMPACTION
3140/* Try memory compaction for high-order allocations before reclaim */
3141static struct page *
3142__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3143 unsigned int alloc_flags, const struct alloc_context *ac,
3144 enum compact_priority prio, enum compact_result *compact_result)
3145{
3146 struct page *page;
3147
3148 if (!order)
3149 return NULL;
3150
3151 current->flags |= PF_MEMALLOC;
3152 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3153 prio);
3154 current->flags &= ~PF_MEMALLOC;
3155
3156 if (*compact_result <= COMPACT_INACTIVE)
3157 return NULL;
3158
3159 /*
3160 * At least in one zone compaction wasn't deferred or skipped, so let's
3161 * count a compaction stall
3162 */
3163 count_vm_event(COMPACTSTALL);
3164
3165 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3166
3167 if (page) {
3168 struct zone *zone = page_zone(page);
3169
3170 zone->compact_blockskip_flush = false;
3171 compaction_defer_reset(zone, order, true);
3172 count_vm_event(COMPACTSUCCESS);
3173 return page;
3174 }
3175
3176 /*
3177 * It's bad if compaction run occurs and fails. The most likely reason
3178 * is that pages exist, but not enough to satisfy watermarks.
3179 */
3180 count_vm_event(COMPACTFAIL);
3181
3182 cond_resched();
3183
3184 return NULL;
3185}
3186
3187static inline bool
3188should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3189 enum compact_result compact_result,
3190 enum compact_priority *compact_priority,
3191 int *compaction_retries)
3192{
3193 int max_retries = MAX_COMPACT_RETRIES;
3194 int min_priority;
3195
3196 if (!order)
3197 return false;
3198
3199 if (compaction_made_progress(compact_result))
3200 (*compaction_retries)++;
3201
3202 /*
3203 * compaction considers all the zone as desperately out of memory
3204 * so it doesn't really make much sense to retry except when the
3205 * failure could be caused by insufficient priority
3206 */
3207 if (compaction_failed(compact_result))
3208 goto check_priority;
3209
3210 /*
3211 * make sure the compaction wasn't deferred or didn't bail out early
3212 * due to locks contention before we declare that we should give up.
3213 * But do not retry if the given zonelist is not suitable for
3214 * compaction.
3215 */
3216 if (compaction_withdrawn(compact_result))
3217 return compaction_zonelist_suitable(ac, order, alloc_flags);
3218
3219 /*
3220 * !costly requests are much more important than __GFP_REPEAT
3221 * costly ones because they are de facto nofail and invoke OOM
3222 * killer to move on while costly can fail and users are ready
3223 * to cope with that. 1/4 retries is rather arbitrary but we
3224 * would need much more detailed feedback from compaction to
3225 * make a better decision.
3226 */
3227 if (order > PAGE_ALLOC_COSTLY_ORDER)
3228 max_retries /= 4;
3229 if (*compaction_retries <= max_retries)
3230 return true;
3231
3232 /*
3233 * Make sure there are attempts at the highest priority if we exhausted
3234 * all retries or failed at the lower priorities.
3235 */
3236check_priority:
3237 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
3238 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
3239 if (*compact_priority > min_priority) {
3240 (*compact_priority)--;
3241 *compaction_retries = 0;
3242 return true;
3243 }
3244 return false;
3245}
3246#else
3247static inline struct page *
3248__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3249 unsigned int alloc_flags, const struct alloc_context *ac,
3250 enum compact_priority prio, enum compact_result *compact_result)
3251{
3252 *compact_result = COMPACT_SKIPPED;
3253 return NULL;
3254}
3255
3256static inline bool
3257should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
3258 enum compact_result compact_result,
3259 enum compact_priority *compact_priority,
3260 int *compaction_retries)
3261{
3262 struct zone *zone;
3263 struct zoneref *z;
3264
3265 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
3266 return false;
3267
3268 /*
3269 * There are setups with compaction disabled which would prefer to loop
3270 * inside the allocator rather than hit the oom killer prematurely.
3271 * Let's give them a good hope and keep retrying while the order-0
3272 * watermarks are OK.
3273 */
3274 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3275 ac->nodemask) {
3276 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
3277 ac_classzone_idx(ac), alloc_flags))
3278 return true;
3279 }
3280 return false;
3281}
3282#endif /* CONFIG_COMPACTION */
3283
3284/* Perform direct synchronous page reclaim */
3285static int
3286__perform_reclaim(gfp_t gfp_mask, unsigned int order,
3287 const struct alloc_context *ac)
3288{
3289 struct reclaim_state reclaim_state;
3290 int progress;
3291
3292 cond_resched();
3293
3294 /* We now go into synchronous reclaim */
3295 cpuset_memory_pressure_bump();
3296 current->flags |= PF_MEMALLOC;
3297 lockdep_set_current_reclaim_state(gfp_mask);
3298 reclaim_state.reclaimed_slab = 0;
3299 current->reclaim_state = &reclaim_state;
3300
3301 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
3302 ac->nodemask);
3303
3304 current->reclaim_state = NULL;
3305 lockdep_clear_current_reclaim_state();
3306 current->flags &= ~PF_MEMALLOC;
3307
3308 cond_resched();
3309
3310 return progress;
3311}
3312
3313/* The really slow allocator path where we enter direct reclaim */
3314static inline struct page *
3315__alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
3316 unsigned int alloc_flags, const struct alloc_context *ac,
3317 unsigned long *did_some_progress)
3318{
3319 struct page *page = NULL;
3320 bool drained = false;
3321
3322 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
3323 if (unlikely(!(*did_some_progress)))
3324 return NULL;
3325
3326retry:
3327 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3328
3329 /*
3330 * If an allocation failed after direct reclaim, it could be because
3331 * pages are pinned on the per-cpu lists or in high alloc reserves.
3332 * Shrink them them and try again
3333 */
3334 if (!page && !drained) {
3335 unreserve_highatomic_pageblock(ac, false);
3336 drain_all_pages(NULL);
3337 drained = true;
3338 goto retry;
3339 }
3340
3341 return page;
3342}
3343
3344static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
3345{
3346 struct zoneref *z;
3347 struct zone *zone;
3348 pg_data_t *last_pgdat = NULL;
3349
3350 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
3351 ac->high_zoneidx, ac->nodemask) {
3352 if (last_pgdat != zone->zone_pgdat)
3353 wakeup_kswapd(zone, order, ac->high_zoneidx);
3354 last_pgdat = zone->zone_pgdat;
3355 }
3356}
3357
3358static inline unsigned int
3359gfp_to_alloc_flags(gfp_t gfp_mask)
3360{
3361 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
3362
3363 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3364 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
3365
3366 /*
3367 * The caller may dip into page reserves a bit more if the caller
3368 * cannot run direct reclaim, or if the caller has realtime scheduling
3369 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3370 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3371 */
3372 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
3373
3374 if (gfp_mask & __GFP_ATOMIC) {
3375 /*
3376 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3377 * if it can't schedule.
3378 */
3379 if (!(gfp_mask & __GFP_NOMEMALLOC))
3380 alloc_flags |= ALLOC_HARDER;
3381 /*
3382 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3383 * comment for __cpuset_node_allowed().
3384 */
3385 alloc_flags &= ~ALLOC_CPUSET;
3386 } else if (unlikely(rt_task(current)) && !in_interrupt())
3387 alloc_flags |= ALLOC_HARDER;
3388
3389#ifdef CONFIG_CMA
3390 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3391 alloc_flags |= ALLOC_CMA;
3392#endif
3393 return alloc_flags;
3394}
3395
3396bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
3397{
3398 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
3399 return false;
3400
3401 if (gfp_mask & __GFP_MEMALLOC)
3402 return true;
3403 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
3404 return true;
3405 if (!in_interrupt() &&
3406 ((current->flags & PF_MEMALLOC) ||
3407 unlikely(test_thread_flag(TIF_MEMDIE))))
3408 return true;
3409
3410 return false;
3411}
3412
3413/*
3414 * Maximum number of reclaim retries without any progress before OOM killer
3415 * is consider as the only way to move forward.
3416 */
3417#define MAX_RECLAIM_RETRIES 16
3418
3419/*
3420 * Checks whether it makes sense to retry the reclaim to make a forward progress
3421 * for the given allocation request.
3422 * The reclaim feedback represented by did_some_progress (any progress during
3423 * the last reclaim round) and no_progress_loops (number of reclaim rounds without
3424 * any progress in a row) is considered as well as the reclaimable pages on the
3425 * applicable zone list (with a backoff mechanism which is a function of
3426 * no_progress_loops).
3427 *
3428 * Returns true if a retry is viable or false to enter the oom path.
3429 */
3430static inline bool
3431should_reclaim_retry(gfp_t gfp_mask, unsigned order,
3432 struct alloc_context *ac, int alloc_flags,
3433 bool did_some_progress, int *no_progress_loops)
3434{
3435 struct zone *zone;
3436 struct zoneref *z;
3437
3438 /*
3439 * Costly allocations might have made a progress but this doesn't mean
3440 * their order will become available due to high fragmentation so
3441 * always increment the no progress counter for them
3442 */
3443 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
3444 *no_progress_loops = 0;
3445 else
3446 (*no_progress_loops)++;
3447
3448 /*
3449 * Make sure we converge to OOM if we cannot make any progress
3450 * several times in the row.
3451 */
3452 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
3453 /* Before OOM, exhaust highatomic_reserve */
3454 return unreserve_highatomic_pageblock(ac, true);
3455 }
3456
3457 /*
3458 * Keep reclaiming pages while there is a chance this will lead
3459 * somewhere. If none of the target zones can satisfy our allocation
3460 * request even if all reclaimable pages are considered then we are
3461 * screwed and have to go OOM.
3462 */
3463 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3464 ac->nodemask) {
3465 unsigned long available;
3466 unsigned long reclaimable;
3467
3468 available = reclaimable = zone_reclaimable_pages(zone);
3469 available -= DIV_ROUND_UP((*no_progress_loops) * available,
3470 MAX_RECLAIM_RETRIES);
3471 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
3472
3473 /*
3474 * Would the allocation succeed if we reclaimed the whole
3475 * available?
3476 */
3477 if (__zone_watermark_ok(zone, order, min_wmark_pages(zone),
3478 ac_classzone_idx(ac), alloc_flags, available)) {
3479 /*
3480 * If we didn't make any progress and have a lot of
3481 * dirty + writeback pages then we should wait for
3482 * an IO to complete to slow down the reclaim and
3483 * prevent from pre mature OOM
3484 */
3485 if (!did_some_progress) {
3486 unsigned long write_pending;
3487
3488 write_pending = zone_page_state_snapshot(zone,
3489 NR_ZONE_WRITE_PENDING);
3490
3491 if (2 * write_pending > reclaimable) {
3492 congestion_wait(BLK_RW_ASYNC, HZ/10);
3493 return true;
3494 }
3495 }
3496
3497 /*
3498 * Memory allocation/reclaim might be called from a WQ
3499 * context and the current implementation of the WQ
3500 * concurrency control doesn't recognize that
3501 * a particular WQ is congested if the worker thread is
3502 * looping without ever sleeping. Therefore we have to
3503 * do a short sleep here rather than calling
3504 * cond_resched().
3505 */
3506 if (current->flags & PF_WQ_WORKER)
3507 schedule_timeout_uninterruptible(1);
3508 else
3509 cond_resched();
3510
3511 return true;
3512 }
3513 }
3514
3515 return false;
3516}
3517
3518static inline struct page *
3519__alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
3520 struct alloc_context *ac)
3521{
3522 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
3523 struct page *page = NULL;
3524 unsigned int alloc_flags;
3525 unsigned long did_some_progress;
3526 enum compact_priority compact_priority;
3527 enum compact_result compact_result;
3528 int compaction_retries;
3529 int no_progress_loops;
3530 unsigned long alloc_start = jiffies;
3531 unsigned int stall_timeout = 10 * HZ;
3532 unsigned int cpuset_mems_cookie;
3533
3534 /*
3535 * In the slowpath, we sanity check order to avoid ever trying to
3536 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3537 * be using allocators in order of preference for an area that is
3538 * too large.
3539 */
3540 if (order >= MAX_ORDER) {
3541 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
3542 return NULL;
3543 }
3544
3545 /*
3546 * We also sanity check to catch abuse of atomic reserves being used by
3547 * callers that are not in atomic context.
3548 */
3549 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
3550 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
3551 gfp_mask &= ~__GFP_ATOMIC;
3552
3553retry_cpuset:
3554 compaction_retries = 0;
3555 no_progress_loops = 0;
3556 compact_priority = DEF_COMPACT_PRIORITY;
3557 cpuset_mems_cookie = read_mems_allowed_begin();
3558 /*
3559 * We need to recalculate the starting point for the zonelist iterator
3560 * because we might have used different nodemask in the fast path, or
3561 * there was a cpuset modification and we are retrying - otherwise we
3562 * could end up iterating over non-eligible zones endlessly.
3563 */
3564 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
3565 ac->high_zoneidx, ac->nodemask);
3566 if (!ac->preferred_zoneref->zone)
3567 goto nopage;
3568
3569
3570 /*
3571 * The fast path uses conservative alloc_flags to succeed only until
3572 * kswapd needs to be woken up, and to avoid the cost of setting up
3573 * alloc_flags precisely. So we do that now.
3574 */
3575 alloc_flags = gfp_to_alloc_flags(gfp_mask);
3576
3577 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3578 wake_all_kswapds(order, ac);
3579
3580 /*
3581 * The adjusted alloc_flags might result in immediate success, so try
3582 * that first
3583 */
3584 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3585 if (page)
3586 goto got_pg;
3587
3588 /*
3589 * For costly allocations, try direct compaction first, as it's likely
3590 * that we have enough base pages and don't need to reclaim. Don't try
3591 * that for allocations that are allowed to ignore watermarks, as the
3592 * ALLOC_NO_WATERMARKS attempt didn't yet happen.
3593 */
3594 if (can_direct_reclaim && order > PAGE_ALLOC_COSTLY_ORDER &&
3595 !gfp_pfmemalloc_allowed(gfp_mask)) {
3596 page = __alloc_pages_direct_compact(gfp_mask, order,
3597 alloc_flags, ac,
3598 INIT_COMPACT_PRIORITY,
3599 &compact_result);
3600 if (page)
3601 goto got_pg;
3602
3603 /*
3604 * Checks for costly allocations with __GFP_NORETRY, which
3605 * includes THP page fault allocations
3606 */
3607 if (gfp_mask & __GFP_NORETRY) {
3608 /*
3609 * If compaction is deferred for high-order allocations,
3610 * it is because sync compaction recently failed. If
3611 * this is the case and the caller requested a THP
3612 * allocation, we do not want to heavily disrupt the
3613 * system, so we fail the allocation instead of entering
3614 * direct reclaim.
3615 */
3616 if (compact_result == COMPACT_DEFERRED)
3617 goto nopage;
3618
3619 /*
3620 * Looks like reclaim/compaction is worth trying, but
3621 * sync compaction could be very expensive, so keep
3622 * using async compaction.
3623 */
3624 compact_priority = INIT_COMPACT_PRIORITY;
3625 }
3626 }
3627
3628retry:
3629 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
3630 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3631 wake_all_kswapds(order, ac);
3632
3633 if (gfp_pfmemalloc_allowed(gfp_mask))
3634 alloc_flags = ALLOC_NO_WATERMARKS;
3635
3636 /*
3637 * Reset the zonelist iterators if memory policies can be ignored.
3638 * These allocations are high priority and system rather than user
3639 * orientated.
3640 */
3641 if (!(alloc_flags & ALLOC_CPUSET) || (alloc_flags & ALLOC_NO_WATERMARKS)) {
3642 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
3643 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
3644 ac->high_zoneidx, ac->nodemask);
3645 }
3646
3647 /* Attempt with potentially adjusted zonelist and alloc_flags */
3648 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3649 if (page)
3650 goto got_pg;
3651
3652 /* Caller is not willing to reclaim, we can't balance anything */
3653 if (!can_direct_reclaim) {
3654 /*
3655 * All existing users of the __GFP_NOFAIL are blockable, so warn
3656 * of any new users that actually allow this type of allocation
3657 * to fail.
3658 */
3659 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
3660 goto nopage;
3661 }
3662
3663 /* Avoid recursion of direct reclaim */
3664 if (current->flags & PF_MEMALLOC) {
3665 /*
3666 * __GFP_NOFAIL request from this context is rather bizarre
3667 * because we cannot reclaim anything and only can loop waiting
3668 * for somebody to do a work for us.
3669 */
3670 if (WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3671 cond_resched();
3672 goto retry;
3673 }
3674 goto nopage;
3675 }
3676
3677 /* Avoid allocations with no watermarks from looping endlessly */
3678 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
3679 goto nopage;
3680
3681
3682 /* Try direct reclaim and then allocating */
3683 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
3684 &did_some_progress);
3685 if (page)
3686 goto got_pg;
3687
3688 /* Try direct compaction and then allocating */
3689 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
3690 compact_priority, &compact_result);
3691 if (page)
3692 goto got_pg;
3693
3694 /* Do not loop if specifically requested */
3695 if (gfp_mask & __GFP_NORETRY)
3696 goto nopage;
3697
3698 /*
3699 * Do not retry costly high order allocations unless they are
3700 * __GFP_REPEAT
3701 */
3702 if (order > PAGE_ALLOC_COSTLY_ORDER && !(gfp_mask & __GFP_REPEAT))
3703 goto nopage;
3704
3705 /* Make sure we know about allocations which stall for too long */
3706 if (time_after(jiffies, alloc_start + stall_timeout)) {
3707 warn_alloc(gfp_mask,
3708 "page allocation stalls for %ums, order:%u",
3709 jiffies_to_msecs(jiffies-alloc_start), order);
3710 stall_timeout += 10 * HZ;
3711 }
3712
3713 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
3714 did_some_progress > 0, &no_progress_loops))
3715 goto retry;
3716
3717 /*
3718 * It doesn't make any sense to retry for the compaction if the order-0
3719 * reclaim is not able to make any progress because the current
3720 * implementation of the compaction depends on the sufficient amount
3721 * of free memory (see __compaction_suitable)
3722 */
3723 if (did_some_progress > 0 &&
3724 should_compact_retry(ac, order, alloc_flags,
3725 compact_result, &compact_priority,
3726 &compaction_retries))
3727 goto retry;
3728
3729 /*
3730 * It's possible we raced with cpuset update so the OOM would be
3731 * premature (see below the nopage: label for full explanation).
3732 */
3733 if (read_mems_allowed_retry(cpuset_mems_cookie))
3734 goto retry_cpuset;
3735
3736 /* Reclaim has failed us, start killing things */
3737 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
3738 if (page)
3739 goto got_pg;
3740
3741 /* Retry as long as the OOM killer is making progress */
3742 if (did_some_progress) {
3743 no_progress_loops = 0;
3744 goto retry;
3745 }
3746
3747nopage:
3748 /*
3749 * When updating a task's mems_allowed or mempolicy nodemask, it is
3750 * possible to race with parallel threads in such a way that our
3751 * allocation can fail while the mask is being updated. If we are about
3752 * to fail, check if the cpuset changed during allocation and if so,
3753 * retry.
3754 */
3755 if (read_mems_allowed_retry(cpuset_mems_cookie))
3756 goto retry_cpuset;
3757
3758 warn_alloc(gfp_mask,
3759 "page allocation failure: order:%u", order);
3760got_pg:
3761 return page;
3762}
3763
3764/*
3765 * This is the 'heart' of the zoned buddy allocator.
3766 */
3767struct page *
3768__alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
3769 struct zonelist *zonelist, nodemask_t *nodemask)
3770{
3771 struct page *page;
3772 unsigned int alloc_flags = ALLOC_WMARK_LOW;
3773 gfp_t alloc_mask = gfp_mask; /* The gfp_t that was actually used for allocation */
3774 struct alloc_context ac = {
3775 .high_zoneidx = gfp_zone(gfp_mask),
3776 .zonelist = zonelist,
3777 .nodemask = nodemask,
3778 .migratetype = gfpflags_to_migratetype(gfp_mask),
3779 };
3780
3781 if (cpusets_enabled()) {
3782 alloc_mask |= __GFP_HARDWALL;
3783 alloc_flags |= ALLOC_CPUSET;
3784 if (!ac.nodemask)
3785 ac.nodemask = &cpuset_current_mems_allowed;
3786 }
3787
3788 gfp_mask &= gfp_allowed_mask;
3789
3790 lockdep_trace_alloc(gfp_mask);
3791
3792 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
3793
3794 if (should_fail_alloc_page(gfp_mask, order))
3795 return NULL;
3796
3797 /*
3798 * Check the zones suitable for the gfp_mask contain at least one
3799 * valid zone. It's possible to have an empty zonelist as a result
3800 * of __GFP_THISNODE and a memoryless node
3801 */
3802 if (unlikely(!zonelist->_zonerefs->zone))
3803 return NULL;
3804
3805 if (IS_ENABLED(CONFIG_CMA) && ac.migratetype == MIGRATE_MOVABLE)
3806 alloc_flags |= ALLOC_CMA;
3807
3808 /* Dirty zone balancing only done in the fast path */
3809 ac.spread_dirty_pages = (gfp_mask & __GFP_WRITE);
3810
3811 /*
3812 * The preferred zone is used for statistics but crucially it is
3813 * also used as the starting point for the zonelist iterator. It
3814 * may get reset for allocations that ignore memory policies.
3815 */
3816 ac.preferred_zoneref = first_zones_zonelist(ac.zonelist,
3817 ac.high_zoneidx, ac.nodemask);
3818 if (!ac.preferred_zoneref->zone) {
3819 page = NULL;
3820 /*
3821 * This might be due to race with cpuset_current_mems_allowed
3822 * update, so make sure we retry with original nodemask in the
3823 * slow path.
3824 */
3825 goto no_zone;
3826 }
3827
3828 /* First allocation attempt */
3829 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
3830 if (likely(page))
3831 goto out;
3832
3833no_zone:
3834 /*
3835 * Runtime PM, block IO and its error handling path can deadlock
3836 * because I/O on the device might not complete.
3837 */
3838 alloc_mask = memalloc_noio_flags(gfp_mask);
3839 ac.spread_dirty_pages = false;
3840
3841 /*
3842 * Restore the original nodemask if it was potentially replaced with
3843 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
3844 */
3845 if (unlikely(ac.nodemask != nodemask))
3846 ac.nodemask = nodemask;
3847
3848 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
3849
3850out:
3851 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
3852 unlikely(memcg_kmem_charge(page, gfp_mask, order) != 0)) {
3853 __free_pages(page, order);
3854 page = NULL;
3855 }
3856
3857 if (kmemcheck_enabled && page)
3858 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
3859
3860 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
3861
3862 return page;
3863}
3864EXPORT_SYMBOL(__alloc_pages_nodemask);
3865
3866/*
3867 * Common helper functions.
3868 */
3869unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
3870{
3871 struct page *page;
3872
3873 /*
3874 * __get_free_pages() returns a 32-bit address, which cannot represent
3875 * a highmem page
3876 */
3877 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
3878
3879 page = alloc_pages(gfp_mask, order);
3880 if (!page)
3881 return 0;
3882 return (unsigned long) page_address(page);
3883}
3884EXPORT_SYMBOL(__get_free_pages);
3885
3886unsigned long get_zeroed_page(gfp_t gfp_mask)
3887{
3888 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
3889}
3890EXPORT_SYMBOL(get_zeroed_page);
3891
3892void __free_pages(struct page *page, unsigned int order)
3893{
3894 if (put_page_testzero(page)) {
3895 if (order == 0)
3896 free_hot_cold_page(page, false);
3897 else
3898 __free_pages_ok(page, order);
3899 }
3900}
3901
3902EXPORT_SYMBOL(__free_pages);
3903
3904void free_pages(unsigned long addr, unsigned int order)
3905{
3906 if (addr != 0) {
3907 VM_BUG_ON(!virt_addr_valid((void *)addr));
3908 __free_pages(virt_to_page((void *)addr), order);
3909 }
3910}
3911
3912EXPORT_SYMBOL(free_pages);
3913
3914/*
3915 * Page Fragment:
3916 * An arbitrary-length arbitrary-offset area of memory which resides
3917 * within a 0 or higher order page. Multiple fragments within that page
3918 * are individually refcounted, in the page's reference counter.
3919 *
3920 * The page_frag functions below provide a simple allocation framework for
3921 * page fragments. This is used by the network stack and network device
3922 * drivers to provide a backing region of memory for use as either an
3923 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3924 */
3925static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
3926 gfp_t gfp_mask)
3927{
3928 struct page *page = NULL;
3929 gfp_t gfp = gfp_mask;
3930
3931#if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3932 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
3933 __GFP_NOMEMALLOC;
3934 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
3935 PAGE_FRAG_CACHE_MAX_ORDER);
3936 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
3937#endif
3938 if (unlikely(!page))
3939 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
3940
3941 nc->va = page ? page_address(page) : NULL;
3942
3943 return page;
3944}
3945
3946void __page_frag_cache_drain(struct page *page, unsigned int count)
3947{
3948 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
3949
3950 if (page_ref_sub_and_test(page, count)) {
3951 unsigned int order = compound_order(page);
3952
3953 if (order == 0)
3954 free_hot_cold_page(page, false);
3955 else
3956 __free_pages_ok(page, order);
3957 }
3958}
3959EXPORT_SYMBOL(__page_frag_cache_drain);
3960
3961void *page_frag_alloc(struct page_frag_cache *nc,
3962 unsigned int fragsz, gfp_t gfp_mask)
3963{
3964 unsigned int size = PAGE_SIZE;
3965 struct page *page;
3966 int offset;
3967
3968 if (unlikely(!nc->va)) {
3969refill:
3970 page = __page_frag_cache_refill(nc, gfp_mask);
3971 if (!page)
3972 return NULL;
3973
3974#if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3975 /* if size can vary use size else just use PAGE_SIZE */
3976 size = nc->size;
3977#endif
3978 /* Even if we own the page, we do not use atomic_set().
3979 * This would break get_page_unless_zero() users.
3980 */
3981 page_ref_add(page, size - 1);
3982
3983 /* reset page count bias and offset to start of new frag */
3984 nc->pfmemalloc = page_is_pfmemalloc(page);
3985 nc->pagecnt_bias = size;
3986 nc->offset = size;
3987 }
3988
3989 offset = nc->offset - fragsz;
3990 if (unlikely(offset < 0)) {
3991 page = virt_to_page(nc->va);
3992
3993 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
3994 goto refill;
3995
3996#if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3997 /* if size can vary use size else just use PAGE_SIZE */
3998 size = nc->size;
3999#endif
4000 /* OK, page count is 0, we can safely set it */
4001 set_page_count(page, size);
4002
4003 /* reset page count bias and offset to start of new frag */
4004 nc->pagecnt_bias = size;
4005 offset = size - fragsz;
4006 }
4007
4008 nc->pagecnt_bias--;
4009 nc->offset = offset;
4010
4011 return nc->va + offset;
4012}
4013EXPORT_SYMBOL(page_frag_alloc);
4014
4015/*
4016 * Frees a page fragment allocated out of either a compound or order 0 page.
4017 */
4018void page_frag_free(void *addr)
4019{
4020 struct page *page = virt_to_head_page(addr);
4021
4022 if (unlikely(put_page_testzero(page)))
4023 __free_pages_ok(page, compound_order(page));
4024}
4025EXPORT_SYMBOL(page_frag_free);
4026
4027static void *make_alloc_exact(unsigned long addr, unsigned int order,
4028 size_t size)
4029{
4030 if (addr) {
4031 unsigned long alloc_end = addr + (PAGE_SIZE << order);
4032 unsigned long used = addr + PAGE_ALIGN(size);
4033
4034 split_page(virt_to_page((void *)addr), order);
4035 while (used < alloc_end) {
4036 free_page(used);
4037 used += PAGE_SIZE;
4038 }
4039 }
4040 return (void *)addr;
4041}
4042
4043/**
4044 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4045 * @size: the number of bytes to allocate
4046 * @gfp_mask: GFP flags for the allocation
4047 *
4048 * This function is similar to alloc_pages(), except that it allocates the
4049 * minimum number of pages to satisfy the request. alloc_pages() can only
4050 * allocate memory in power-of-two pages.
4051 *
4052 * This function is also limited by MAX_ORDER.
4053 *
4054 * Memory allocated by this function must be released by free_pages_exact().
4055 */
4056void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
4057{
4058 unsigned int order = get_order(size);
4059 unsigned long addr;
4060
4061 addr = __get_free_pages(gfp_mask, order);
4062 return make_alloc_exact(addr, order, size);
4063}
4064EXPORT_SYMBOL(alloc_pages_exact);
4065
4066/**
4067 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4068 * pages on a node.
4069 * @nid: the preferred node ID where memory should be allocated
4070 * @size: the number of bytes to allocate
4071 * @gfp_mask: GFP flags for the allocation
4072 *
4073 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4074 * back.
4075 */
4076void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
4077{
4078 unsigned int order = get_order(size);
4079 struct page *p = alloc_pages_node(nid, gfp_mask, order);
4080 if (!p)
4081 return NULL;
4082 return make_alloc_exact((unsigned long)page_address(p), order, size);
4083}
4084
4085/**
4086 * free_pages_exact - release memory allocated via alloc_pages_exact()
4087 * @virt: the value returned by alloc_pages_exact.
4088 * @size: size of allocation, same value as passed to alloc_pages_exact().
4089 *
4090 * Release the memory allocated by a previous call to alloc_pages_exact.
4091 */
4092void free_pages_exact(void *virt, size_t size)
4093{
4094 unsigned long addr = (unsigned long)virt;
4095 unsigned long end = addr + PAGE_ALIGN(size);
4096
4097 while (addr < end) {
4098 free_page(addr);
4099 addr += PAGE_SIZE;
4100 }
4101}
4102EXPORT_SYMBOL(free_pages_exact);
4103
4104/**
4105 * nr_free_zone_pages - count number of pages beyond high watermark
4106 * @offset: The zone index of the highest zone
4107 *
4108 * nr_free_zone_pages() counts the number of counts pages which are beyond the
4109 * high watermark within all zones at or below a given zone index. For each
4110 * zone, the number of pages is calculated as:
4111 * managed_pages - high_pages
4112 */
4113static unsigned long nr_free_zone_pages(int offset)
4114{
4115 struct zoneref *z;
4116 struct zone *zone;
4117
4118 /* Just pick one node, since fallback list is circular */
4119 unsigned long sum = 0;
4120
4121 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
4122
4123 for_each_zone_zonelist(zone, z, zonelist, offset) {
4124 unsigned long size = zone->managed_pages;
4125 unsigned long high = high_wmark_pages(zone);
4126 if (size > high)
4127 sum += size - high;
4128 }
4129
4130 return sum;
4131}
4132
4133/**
4134 * nr_free_buffer_pages - count number of pages beyond high watermark
4135 *
4136 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4137 * watermark within ZONE_DMA and ZONE_NORMAL.
4138 */
4139unsigned long nr_free_buffer_pages(void)
4140{
4141 return nr_free_zone_pages(gfp_zone(GFP_USER));
4142}
4143EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
4144
4145/**
4146 * nr_free_pagecache_pages - count number of pages beyond high watermark
4147 *
4148 * nr_free_pagecache_pages() counts the number of pages which are beyond the
4149 * high watermark within all zones.
4150 */
4151unsigned long nr_free_pagecache_pages(void)
4152{
4153 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
4154}
4155
4156static inline void show_node(struct zone *zone)
4157{
4158 if (IS_ENABLED(CONFIG_NUMA))
4159 printk("Node %d ", zone_to_nid(zone));
4160}
4161
4162long si_mem_available(void)
4163{
4164 long available;
4165 unsigned long pagecache;
4166 unsigned long wmark_low = 0;
4167 unsigned long pages[NR_LRU_LISTS];
4168 struct zone *zone;
4169 int lru;
4170
4171 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
4172 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
4173
4174 for_each_zone(zone)
4175 wmark_low += zone->watermark[WMARK_LOW];
4176
4177 /*
4178 * Estimate the amount of memory available for userspace allocations,
4179 * without causing swapping.
4180 */
4181 available = global_page_state(NR_FREE_PAGES) - totalreserve_pages;
4182
4183 /*
4184 * Not all the page cache can be freed, otherwise the system will
4185 * start swapping. Assume at least half of the page cache, or the
4186 * low watermark worth of cache, needs to stay.
4187 */
4188 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
4189 pagecache -= min(pagecache / 2, wmark_low);
4190 available += pagecache;
4191
4192 /*
4193 * Part of the reclaimable slab consists of items that are in use,
4194 * and cannot be freed. Cap this estimate at the low watermark.
4195 */
4196 available += global_page_state(NR_SLAB_RECLAIMABLE) -
4197 min(global_page_state(NR_SLAB_RECLAIMABLE) / 2, wmark_low);
4198
4199 if (available < 0)
4200 available = 0;
4201 return available;
4202}
4203EXPORT_SYMBOL_GPL(si_mem_available);
4204
4205void si_meminfo(struct sysinfo *val)
4206{
4207 val->totalram = totalram_pages;
4208 val->sharedram = global_node_page_state(NR_SHMEM);
4209 val->freeram = global_page_state(NR_FREE_PAGES);
4210 val->bufferram = nr_blockdev_pages();
4211 val->totalhigh = totalhigh_pages;
4212 val->freehigh = nr_free_highpages();
4213 val->mem_unit = PAGE_SIZE;
4214}
4215
4216EXPORT_SYMBOL(si_meminfo);
4217
4218#ifdef CONFIG_NUMA
4219void si_meminfo_node(struct sysinfo *val, int nid)
4220{
4221 int zone_type; /* needs to be signed */
4222 unsigned long managed_pages = 0;
4223 unsigned long managed_highpages = 0;
4224 unsigned long free_highpages = 0;
4225 pg_data_t *pgdat = NODE_DATA(nid);
4226
4227 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
4228 managed_pages += pgdat->node_zones[zone_type].managed_pages;
4229 val->totalram = managed_pages;
4230 val->sharedram = node_page_state(pgdat, NR_SHMEM);
4231 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
4232#ifdef CONFIG_HIGHMEM
4233 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
4234 struct zone *zone = &pgdat->node_zones[zone_type];
4235
4236 if (is_highmem(zone)) {
4237 managed_highpages += zone->managed_pages;
4238 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
4239 }
4240 }
4241 val->totalhigh = managed_highpages;
4242 val->freehigh = free_highpages;
4243#else
4244 val->totalhigh = managed_highpages;
4245 val->freehigh = free_highpages;
4246#endif
4247 val->mem_unit = PAGE_SIZE;
4248}
4249#endif
4250
4251/*
4252 * Determine whether the node should be displayed or not, depending on whether
4253 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
4254 */
4255bool skip_free_areas_node(unsigned int flags, int nid)
4256{
4257 bool ret = false;
4258 unsigned int cpuset_mems_cookie;
4259
4260 if (!(flags & SHOW_MEM_FILTER_NODES))
4261 goto out;
4262
4263 do {
4264 cpuset_mems_cookie = read_mems_allowed_begin();
4265 ret = !node_isset(nid, cpuset_current_mems_allowed);
4266 } while (read_mems_allowed_retry(cpuset_mems_cookie));
4267out:
4268 return ret;
4269}
4270
4271#define K(x) ((x) << (PAGE_SHIFT-10))
4272
4273static void show_migration_types(unsigned char type)
4274{
4275 static const char types[MIGRATE_TYPES] = {
4276 [MIGRATE_UNMOVABLE] = 'U',
4277 [MIGRATE_MOVABLE] = 'M',
4278 [MIGRATE_RECLAIMABLE] = 'E',
4279 [MIGRATE_HIGHATOMIC] = 'H',
4280#ifdef CONFIG_CMA
4281 [MIGRATE_CMA] = 'C',
4282#endif
4283#ifdef CONFIG_MEMORY_ISOLATION
4284 [MIGRATE_ISOLATE] = 'I',
4285#endif
4286 };
4287 char tmp[MIGRATE_TYPES + 1];
4288 char *p = tmp;
4289 int i;
4290
4291 for (i = 0; i < MIGRATE_TYPES; i++) {
4292 if (type & (1 << i))
4293 *p++ = types[i];
4294 }
4295
4296 *p = '\0';
4297 printk(KERN_CONT "(%s) ", tmp);
4298}
4299
4300/*
4301 * Show free area list (used inside shift_scroll-lock stuff)
4302 * We also calculate the percentage fragmentation. We do this by counting the
4303 * memory on each free list with the exception of the first item on the list.
4304 *
4305 * Bits in @filter:
4306 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
4307 * cpuset.
4308 */
4309void show_free_areas(unsigned int filter)
4310{
4311 unsigned long free_pcp = 0;
4312 int cpu;
4313 struct zone *zone;
4314 pg_data_t *pgdat;
4315
4316 for_each_populated_zone(zone) {
4317 if (skip_free_areas_node(filter, zone_to_nid(zone)))
4318 continue;
4319
4320 for_each_online_cpu(cpu)
4321 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4322 }
4323
4324 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
4325 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
4326 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
4327 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
4328 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
4329 " free:%lu free_pcp:%lu free_cma:%lu\n",
4330 global_node_page_state(NR_ACTIVE_ANON),
4331 global_node_page_state(NR_INACTIVE_ANON),
4332 global_node_page_state(NR_ISOLATED_ANON),
4333 global_node_page_state(NR_ACTIVE_FILE),
4334 global_node_page_state(NR_INACTIVE_FILE),
4335 global_node_page_state(NR_ISOLATED_FILE),
4336 global_node_page_state(NR_UNEVICTABLE),
4337 global_node_page_state(NR_FILE_DIRTY),
4338 global_node_page_state(NR_WRITEBACK),
4339 global_node_page_state(NR_UNSTABLE_NFS),
4340 global_page_state(NR_SLAB_RECLAIMABLE),
4341 global_page_state(NR_SLAB_UNRECLAIMABLE),
4342 global_node_page_state(NR_FILE_MAPPED),
4343 global_node_page_state(NR_SHMEM),
4344 global_page_state(NR_PAGETABLE),
4345 global_page_state(NR_BOUNCE),
4346 global_page_state(NR_FREE_PAGES),
4347 free_pcp,
4348 global_page_state(NR_FREE_CMA_PAGES));
4349
4350 for_each_online_pgdat(pgdat) {
4351 printk("Node %d"
4352 " active_anon:%lukB"
4353 " inactive_anon:%lukB"
4354 " active_file:%lukB"
4355 " inactive_file:%lukB"
4356 " unevictable:%lukB"
4357 " isolated(anon):%lukB"
4358 " isolated(file):%lukB"
4359 " mapped:%lukB"
4360 " dirty:%lukB"
4361 " writeback:%lukB"
4362 " shmem:%lukB"
4363#ifdef CONFIG_TRANSPARENT_HUGEPAGE
4364 " shmem_thp: %lukB"
4365 " shmem_pmdmapped: %lukB"
4366 " anon_thp: %lukB"
4367#endif
4368 " writeback_tmp:%lukB"
4369 " unstable:%lukB"
4370 " pages_scanned:%lu"
4371 " all_unreclaimable? %s"
4372 "\n",
4373 pgdat->node_id,
4374 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
4375 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
4376 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
4377 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
4378 K(node_page_state(pgdat, NR_UNEVICTABLE)),
4379 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
4380 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
4381 K(node_page_state(pgdat, NR_FILE_MAPPED)),
4382 K(node_page_state(pgdat, NR_FILE_DIRTY)),
4383 K(node_page_state(pgdat, NR_WRITEBACK)),
4384 K(node_page_state(pgdat, NR_SHMEM)),
4385#ifdef CONFIG_TRANSPARENT_HUGEPAGE
4386 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
4387 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
4388 * HPAGE_PMD_NR),
4389 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
4390#endif
4391 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
4392 K(node_page_state(pgdat, NR_UNSTABLE_NFS)),
4393 node_page_state(pgdat, NR_PAGES_SCANNED),
4394 !pgdat_reclaimable(pgdat) ? "yes" : "no");
4395 }
4396
4397 for_each_populated_zone(zone) {
4398 int i;
4399
4400 if (skip_free_areas_node(filter, zone_to_nid(zone)))
4401 continue;
4402
4403 free_pcp = 0;
4404 for_each_online_cpu(cpu)
4405 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4406
4407 show_node(zone);
4408 printk(KERN_CONT
4409 "%s"
4410 " free:%lukB"
4411 " min:%lukB"
4412 " low:%lukB"
4413 " high:%lukB"
4414 " active_anon:%lukB"
4415 " inactive_anon:%lukB"
4416 " active_file:%lukB"
4417 " inactive_file:%lukB"
4418 " unevictable:%lukB"
4419 " writepending:%lukB"
4420 " present:%lukB"
4421 " managed:%lukB"
4422 " mlocked:%lukB"
4423 " slab_reclaimable:%lukB"
4424 " slab_unreclaimable:%lukB"
4425 " kernel_stack:%lukB"
4426 " pagetables:%lukB"
4427 " bounce:%lukB"
4428 " free_pcp:%lukB"
4429 " local_pcp:%ukB"
4430 " free_cma:%lukB"
4431 "\n",
4432 zone->name,
4433 K(zone_page_state(zone, NR_FREE_PAGES)),
4434 K(min_wmark_pages(zone)),
4435 K(low_wmark_pages(zone)),
4436 K(high_wmark_pages(zone)),
4437 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
4438 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
4439 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
4440 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
4441 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
4442 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
4443 K(zone->present_pages),
4444 K(zone->managed_pages),
4445 K(zone_page_state(zone, NR_MLOCK)),
4446 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
4447 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
4448 zone_page_state(zone, NR_KERNEL_STACK_KB),
4449 K(zone_page_state(zone, NR_PAGETABLE)),
4450 K(zone_page_state(zone, NR_BOUNCE)),
4451 K(free_pcp),
4452 K(this_cpu_read(zone->pageset->pcp.count)),
4453 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
4454 printk("lowmem_reserve[]:");
4455 for (i = 0; i < MAX_NR_ZONES; i++)
4456 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
4457 printk(KERN_CONT "\n");
4458 }
4459
4460 for_each_populated_zone(zone) {
4461 unsigned int order;
4462 unsigned long nr[MAX_ORDER], flags, total = 0;
4463 unsigned char types[MAX_ORDER];
4464
4465 if (skip_free_areas_node(filter, zone_to_nid(zone)))
4466 continue;
4467 show_node(zone);
4468 printk(KERN_CONT "%s: ", zone->name);
4469
4470 spin_lock_irqsave(&zone->lock, flags);
4471 for (order = 0; order < MAX_ORDER; order++) {
4472 struct free_area *area = &zone->free_area[order];
4473 int type;
4474
4475 nr[order] = area->nr_free;
4476 total += nr[order] << order;
4477
4478 types[order] = 0;
4479 for (type = 0; type < MIGRATE_TYPES; type++) {
4480 if (!list_empty(&area->free_list[type]))
4481 types[order] |= 1 << type;
4482 }
4483 }
4484 spin_unlock_irqrestore(&zone->lock, flags);
4485 for (order = 0; order < MAX_ORDER; order++) {
4486 printk(KERN_CONT "%lu*%lukB ",
4487 nr[order], K(1UL) << order);
4488 if (nr[order])
4489 show_migration_types(types[order]);
4490 }
4491 printk(KERN_CONT "= %lukB\n", K(total));
4492 }
4493
4494 hugetlb_show_meminfo();
4495
4496 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
4497
4498 show_swap_cache_info();
4499}
4500
4501static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
4502{
4503 zoneref->zone = zone;
4504 zoneref->zone_idx = zone_idx(zone);
4505}
4506
4507/*
4508 * Builds allocation fallback zone lists.
4509 *
4510 * Add all populated zones of a node to the zonelist.
4511 */
4512static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
4513 int nr_zones)
4514{
4515 struct zone *zone;
4516 enum zone_type zone_type = MAX_NR_ZONES;
4517
4518 do {
4519 zone_type--;
4520 zone = pgdat->node_zones + zone_type;
4521 if (managed_zone(zone)) {
4522 zoneref_set_zone(zone,
4523 &zonelist->_zonerefs[nr_zones++]);
4524 check_highest_zone(zone_type);
4525 }
4526 } while (zone_type);
4527
4528 return nr_zones;
4529}
4530
4531
4532/*
4533 * zonelist_order:
4534 * 0 = automatic detection of better ordering.
4535 * 1 = order by ([node] distance, -zonetype)
4536 * 2 = order by (-zonetype, [node] distance)
4537 *
4538 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
4539 * the same zonelist. So only NUMA can configure this param.
4540 */
4541#define ZONELIST_ORDER_DEFAULT 0
4542#define ZONELIST_ORDER_NODE 1
4543#define ZONELIST_ORDER_ZONE 2
4544
4545/* zonelist order in the kernel.
4546 * set_zonelist_order() will set this to NODE or ZONE.
4547 */
4548static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
4549static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
4550
4551
4552#ifdef CONFIG_NUMA
4553/* The value user specified ....changed by config */
4554static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
4555/* string for sysctl */
4556#define NUMA_ZONELIST_ORDER_LEN 16
4557char numa_zonelist_order[16] = "default";
4558
4559/*
4560 * interface for configure zonelist ordering.
4561 * command line option "numa_zonelist_order"
4562 * = "[dD]efault - default, automatic configuration.
4563 * = "[nN]ode - order by node locality, then by zone within node
4564 * = "[zZ]one - order by zone, then by locality within zone
4565 */
4566
4567static int __parse_numa_zonelist_order(char *s)
4568{
4569 if (*s == 'd' || *s == 'D') {
4570 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
4571 } else if (*s == 'n' || *s == 'N') {
4572 user_zonelist_order = ZONELIST_ORDER_NODE;
4573 } else if (*s == 'z' || *s == 'Z') {
4574 user_zonelist_order = ZONELIST_ORDER_ZONE;
4575 } else {
4576 pr_warn("Ignoring invalid numa_zonelist_order value: %s\n", s);
4577 return -EINVAL;
4578 }
4579 return 0;
4580}
4581
4582static __init int setup_numa_zonelist_order(char *s)
4583{
4584 int ret;
4585
4586 if (!s)
4587 return 0;
4588
4589 ret = __parse_numa_zonelist_order(s);
4590 if (ret == 0)
4591 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
4592
4593 return ret;
4594}
4595early_param("numa_zonelist_order", setup_numa_zonelist_order);
4596
4597/*
4598 * sysctl handler for numa_zonelist_order
4599 */
4600int numa_zonelist_order_handler(struct ctl_table *table, int write,
4601 void __user *buffer, size_t *length,
4602 loff_t *ppos)
4603{
4604 char saved_string[NUMA_ZONELIST_ORDER_LEN];
4605 int ret;
4606 static DEFINE_MUTEX(zl_order_mutex);
4607
4608 mutex_lock(&zl_order_mutex);
4609 if (write) {
4610 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
4611 ret = -EINVAL;
4612 goto out;
4613 }
4614 strcpy(saved_string, (char *)table->data);
4615 }
4616 ret = proc_dostring(table, write, buffer, length, ppos);
4617 if (ret)
4618 goto out;
4619 if (write) {
4620 int oldval = user_zonelist_order;
4621
4622 ret = __parse_numa_zonelist_order((char *)table->data);
4623 if (ret) {
4624 /*
4625 * bogus value. restore saved string
4626 */
4627 strncpy((char *)table->data, saved_string,
4628 NUMA_ZONELIST_ORDER_LEN);
4629 user_zonelist_order = oldval;
4630 } else if (oldval != user_zonelist_order) {
4631 mutex_lock(&zonelists_mutex);
4632 build_all_zonelists(NULL, NULL);
4633 mutex_unlock(&zonelists_mutex);
4634 }
4635 }
4636out:
4637 mutex_unlock(&zl_order_mutex);
4638 return ret;
4639}
4640
4641
4642#define MAX_NODE_LOAD (nr_online_nodes)
4643static int node_load[MAX_NUMNODES];
4644
4645/**
4646 * find_next_best_node - find the next node that should appear in a given node's fallback list
4647 * @node: node whose fallback list we're appending
4648 * @used_node_mask: nodemask_t of already used nodes
4649 *
4650 * We use a number of factors to determine which is the next node that should
4651 * appear on a given node's fallback list. The node should not have appeared
4652 * already in @node's fallback list, and it should be the next closest node
4653 * according to the distance array (which contains arbitrary distance values
4654 * from each node to each node in the system), and should also prefer nodes
4655 * with no CPUs, since presumably they'll have very little allocation pressure
4656 * on them otherwise.
4657 * It returns -1 if no node is found.
4658 */
4659static int find_next_best_node(int node, nodemask_t *used_node_mask)
4660{
4661 int n, val;
4662 int min_val = INT_MAX;
4663 int best_node = NUMA_NO_NODE;
4664 const struct cpumask *tmp = cpumask_of_node(0);
4665
4666 /* Use the local node if we haven't already */
4667 if (!node_isset(node, *used_node_mask)) {
4668 node_set(node, *used_node_mask);
4669 return node;
4670 }
4671
4672 for_each_node_state(n, N_MEMORY) {
4673
4674 /* Don't want a node to appear more than once */
4675 if (node_isset(n, *used_node_mask))
4676 continue;
4677
4678 /* Use the distance array to find the distance */
4679 val = node_distance(node, n);
4680
4681 /* Penalize nodes under us ("prefer the next node") */
4682 val += (n < node);
4683
4684 /* Give preference to headless and unused nodes */
4685 tmp = cpumask_of_node(n);
4686 if (!cpumask_empty(tmp))
4687 val += PENALTY_FOR_NODE_WITH_CPUS;
4688
4689 /* Slight preference for less loaded node */
4690 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
4691 val += node_load[n];
4692
4693 if (val < min_val) {
4694 min_val = val;
4695 best_node = n;
4696 }
4697 }
4698
4699 if (best_node >= 0)
4700 node_set(best_node, *used_node_mask);
4701
4702 return best_node;
4703}
4704
4705
4706/*
4707 * Build zonelists ordered by node and zones within node.
4708 * This results in maximum locality--normal zone overflows into local
4709 * DMA zone, if any--but risks exhausting DMA zone.
4710 */
4711static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
4712{
4713 int j;
4714 struct zonelist *zonelist;
4715
4716 zonelist = &pgdat->node_zonelists[ZONELIST_FALLBACK];
4717 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
4718 ;
4719 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4720 zonelist->_zonerefs[j].zone = NULL;
4721 zonelist->_zonerefs[j].zone_idx = 0;
4722}
4723
4724/*
4725 * Build gfp_thisnode zonelists
4726 */
4727static void build_thisnode_zonelists(pg_data_t *pgdat)
4728{
4729 int j;
4730 struct zonelist *zonelist;
4731
4732 zonelist = &pgdat->node_zonelists[ZONELIST_NOFALLBACK];
4733 j = build_zonelists_node(pgdat, zonelist, 0);
4734 zonelist->_zonerefs[j].zone = NULL;
4735 zonelist->_zonerefs[j].zone_idx = 0;
4736}
4737
4738/*
4739 * Build zonelists ordered by zone and nodes within zones.
4740 * This results in conserving DMA zone[s] until all Normal memory is
4741 * exhausted, but results in overflowing to remote node while memory
4742 * may still exist in local DMA zone.
4743 */
4744static int node_order[MAX_NUMNODES];
4745
4746static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
4747{
4748 int pos, j, node;
4749 int zone_type; /* needs to be signed */
4750 struct zone *z;
4751 struct zonelist *zonelist;
4752
4753 zonelist = &pgdat->node_zonelists[ZONELIST_FALLBACK];
4754 pos = 0;
4755 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
4756 for (j = 0; j < nr_nodes; j++) {
4757 node = node_order[j];
4758 z = &NODE_DATA(node)->node_zones[zone_type];
4759 if (managed_zone(z)) {
4760 zoneref_set_zone(z,
4761 &zonelist->_zonerefs[pos++]);
4762 check_highest_zone(zone_type);
4763 }
4764 }
4765 }
4766 zonelist->_zonerefs[pos].zone = NULL;
4767 zonelist->_zonerefs[pos].zone_idx = 0;
4768}
4769
4770#if defined(CONFIG_64BIT)
4771/*
4772 * Devices that require DMA32/DMA are relatively rare and do not justify a
4773 * penalty to every machine in case the specialised case applies. Default
4774 * to Node-ordering on 64-bit NUMA machines
4775 */
4776static int default_zonelist_order(void)
4777{
4778 return ZONELIST_ORDER_NODE;
4779}
4780#else
4781/*
4782 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4783 * by the kernel. If processes running on node 0 deplete the low memory zone
4784 * then reclaim will occur more frequency increasing stalls and potentially
4785 * be easier to OOM if a large percentage of the zone is under writeback or
4786 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4787 * Hence, default to zone ordering on 32-bit.
4788 */
4789static int default_zonelist_order(void)
4790{
4791 return ZONELIST_ORDER_ZONE;
4792}
4793#endif /* CONFIG_64BIT */
4794
4795static void set_zonelist_order(void)
4796{
4797 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
4798 current_zonelist_order = default_zonelist_order();
4799 else
4800 current_zonelist_order = user_zonelist_order;
4801}
4802
4803static void build_zonelists(pg_data_t *pgdat)
4804{
4805 int i, node, load;
4806 nodemask_t used_mask;
4807 int local_node, prev_node;
4808 struct zonelist *zonelist;
4809 unsigned int order = current_zonelist_order;
4810
4811 /* initialize zonelists */
4812 for (i = 0; i < MAX_ZONELISTS; i++) {
4813 zonelist = pgdat->node_zonelists + i;
4814 zonelist->_zonerefs[0].zone = NULL;
4815 zonelist->_zonerefs[0].zone_idx = 0;
4816 }
4817
4818 /* NUMA-aware ordering of nodes */
4819 local_node = pgdat->node_id;
4820 load = nr_online_nodes;
4821 prev_node = local_node;
4822 nodes_clear(used_mask);
4823
4824 memset(node_order, 0, sizeof(node_order));
4825 i = 0;
4826
4827 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
4828 /*
4829 * We don't want to pressure a particular node.
4830 * So adding penalty to the first node in same
4831 * distance group to make it round-robin.
4832 */
4833 if (node_distance(local_node, node) !=
4834 node_distance(local_node, prev_node))
4835 node_load[node] = load;
4836
4837 prev_node = node;
4838 load--;
4839 if (order == ZONELIST_ORDER_NODE)
4840 build_zonelists_in_node_order(pgdat, node);
4841 else
4842 node_order[i++] = node; /* remember order */
4843 }
4844
4845 if (order == ZONELIST_ORDER_ZONE) {
4846 /* calculate node order -- i.e., DMA last! */
4847 build_zonelists_in_zone_order(pgdat, i);
4848 }
4849
4850 build_thisnode_zonelists(pgdat);
4851}
4852
4853#ifdef CONFIG_HAVE_MEMORYLESS_NODES
4854/*
4855 * Return node id of node used for "local" allocations.
4856 * I.e., first node id of first zone in arg node's generic zonelist.
4857 * Used for initializing percpu 'numa_mem', which is used primarily
4858 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4859 */
4860int local_memory_node(int node)
4861{
4862 struct zoneref *z;
4863
4864 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
4865 gfp_zone(GFP_KERNEL),
4866 NULL);
4867 return z->zone->node;
4868}
4869#endif
4870
4871static void setup_min_unmapped_ratio(void);
4872static void setup_min_slab_ratio(void);
4873#else /* CONFIG_NUMA */
4874
4875static void set_zonelist_order(void)
4876{
4877 current_zonelist_order = ZONELIST_ORDER_ZONE;
4878}
4879
4880static void build_zonelists(pg_data_t *pgdat)
4881{
4882 int node, local_node;
4883 enum zone_type j;
4884 struct zonelist *zonelist;
4885
4886 local_node = pgdat->node_id;
4887
4888 zonelist = &pgdat->node_zonelists[ZONELIST_FALLBACK];
4889 j = build_zonelists_node(pgdat, zonelist, 0);
4890
4891 /*
4892 * Now we build the zonelist so that it contains the zones
4893 * of all the other nodes.
4894 * We don't want to pressure a particular node, so when
4895 * building the zones for node N, we make sure that the
4896 * zones coming right after the local ones are those from
4897 * node N+1 (modulo N)
4898 */
4899 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
4900 if (!node_online(node))
4901 continue;
4902 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4903 }
4904 for (node = 0; node < local_node; node++) {
4905 if (!node_online(node))
4906 continue;
4907 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4908 }
4909
4910 zonelist->_zonerefs[j].zone = NULL;
4911 zonelist->_zonerefs[j].zone_idx = 0;
4912}
4913
4914#endif /* CONFIG_NUMA */
4915
4916/*
4917 * Boot pageset table. One per cpu which is going to be used for all
4918 * zones and all nodes. The parameters will be set in such a way
4919 * that an item put on a list will immediately be handed over to
4920 * the buddy list. This is safe since pageset manipulation is done
4921 * with interrupts disabled.
4922 *
4923 * The boot_pagesets must be kept even after bootup is complete for
4924 * unused processors and/or zones. They do play a role for bootstrapping
4925 * hotplugged processors.
4926 *
4927 * zoneinfo_show() and maybe other functions do
4928 * not check if the processor is online before following the pageset pointer.
4929 * Other parts of the kernel may not check if the zone is available.
4930 */
4931static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
4932static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
4933static void setup_zone_pageset(struct zone *zone);
4934
4935/*
4936 * Global mutex to protect against size modification of zonelists
4937 * as well as to serialize pageset setup for the new populated zone.
4938 */
4939DEFINE_MUTEX(zonelists_mutex);
4940
4941/* return values int ....just for stop_machine() */
4942static int __build_all_zonelists(void *data)
4943{
4944 int nid;
4945 int cpu;
4946 pg_data_t *self = data;
4947
4948#ifdef CONFIG_NUMA
4949 memset(node_load, 0, sizeof(node_load));
4950#endif
4951
4952 if (self && !node_online(self->node_id)) {
4953 build_zonelists(self);
4954 }
4955
4956 for_each_online_node(nid) {
4957 pg_data_t *pgdat = NODE_DATA(nid);
4958
4959 build_zonelists(pgdat);
4960 }
4961
4962 /*
4963 * Initialize the boot_pagesets that are going to be used
4964 * for bootstrapping processors. The real pagesets for
4965 * each zone will be allocated later when the per cpu
4966 * allocator is available.
4967 *
4968 * boot_pagesets are used also for bootstrapping offline
4969 * cpus if the system is already booted because the pagesets
4970 * are needed to initialize allocators on a specific cpu too.
4971 * F.e. the percpu allocator needs the page allocator which
4972 * needs the percpu allocator in order to allocate its pagesets
4973 * (a chicken-egg dilemma).
4974 */
4975 for_each_possible_cpu(cpu) {
4976 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
4977
4978#ifdef CONFIG_HAVE_MEMORYLESS_NODES
4979 /*
4980 * We now know the "local memory node" for each node--
4981 * i.e., the node of the first zone in the generic zonelist.
4982 * Set up numa_mem percpu variable for on-line cpus. During
4983 * boot, only the boot cpu should be on-line; we'll init the
4984 * secondary cpus' numa_mem as they come on-line. During
4985 * node/memory hotplug, we'll fixup all on-line cpus.
4986 */
4987 if (cpu_online(cpu))
4988 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
4989#endif
4990 }
4991
4992 return 0;
4993}
4994
4995static noinline void __init
4996build_all_zonelists_init(void)
4997{
4998 __build_all_zonelists(NULL);
4999 mminit_verify_zonelist();
5000 cpuset_init_current_mems_allowed();
5001}
5002
5003/*
5004 * Called with zonelists_mutex held always
5005 * unless system_state == SYSTEM_BOOTING.
5006 *
5007 * __ref due to (1) call of __meminit annotated setup_zone_pageset
5008 * [we're only called with non-NULL zone through __meminit paths] and
5009 * (2) call of __init annotated helper build_all_zonelists_init
5010 * [protected by SYSTEM_BOOTING].
5011 */
5012void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
5013{
5014 set_zonelist_order();
5015
5016 if (system_state == SYSTEM_BOOTING) {
5017 build_all_zonelists_init();
5018 } else {
5019#ifdef CONFIG_MEMORY_HOTPLUG
5020 if (zone)
5021 setup_zone_pageset(zone);
5022#endif
5023 /* we have to stop all cpus to guarantee there is no user
5024 of zonelist */
5025 stop_machine(__build_all_zonelists, pgdat, NULL);
5026 /* cpuset refresh routine should be here */
5027 }
5028 vm_total_pages = nr_free_pagecache_pages();
5029 /*
5030 * Disable grouping by mobility if the number of pages in the
5031 * system is too low to allow the mechanism to work. It would be
5032 * more accurate, but expensive to check per-zone. This check is
5033 * made on memory-hotadd so a system can start with mobility
5034 * disabled and enable it later
5035 */
5036 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5037 page_group_by_mobility_disabled = 1;
5038 else
5039 page_group_by_mobility_disabled = 0;
5040
5041 pr_info("Built %i zonelists in %s order, mobility grouping %s. Total pages: %ld\n",
5042 nr_online_nodes,
5043 zonelist_order_name[current_zonelist_order],
5044 page_group_by_mobility_disabled ? "off" : "on",
5045 vm_total_pages);
5046#ifdef CONFIG_NUMA
5047 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5048#endif
5049}
5050
5051/*
5052 * Initially all pages are reserved - free ones are freed
5053 * up by free_all_bootmem() once the early boot process is
5054 * done. Non-atomic initialization, single-pass.
5055 */
5056void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5057 unsigned long start_pfn, enum memmap_context context)
5058{
5059 struct vmem_altmap *altmap = to_vmem_altmap(__pfn_to_phys(start_pfn));
5060 unsigned long end_pfn = start_pfn + size;
5061 pg_data_t *pgdat = NODE_DATA(nid);
5062 unsigned long pfn;
5063 unsigned long nr_initialised = 0;
5064#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5065 struct memblock_region *r = NULL, *tmp;
5066#endif
5067
5068 if (highest_memmap_pfn < end_pfn - 1)
5069 highest_memmap_pfn = end_pfn - 1;
5070
5071 /*
5072 * Honor reservation requested by the driver for this ZONE_DEVICE
5073 * memory
5074 */
5075 if (altmap && start_pfn == altmap->base_pfn)
5076 start_pfn += altmap->reserve;
5077
5078 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5079 /*
5080 * There can be holes in boot-time mem_map[]s handed to this
5081 * function. They do not exist on hotplugged memory.
5082 */
5083 if (context != MEMMAP_EARLY)
5084 goto not_early;
5085
5086 if (!early_pfn_valid(pfn))
5087 continue;
5088 if (!early_pfn_in_nid(pfn, nid))
5089 continue;
5090 if (!update_defer_init(pgdat, pfn, end_pfn, &nr_initialised))
5091 break;
5092
5093#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5094 /*
5095 * Check given memblock attribute by firmware which can affect
5096 * kernel memory layout. If zone==ZONE_MOVABLE but memory is
5097 * mirrored, it's an overlapped memmap init. skip it.
5098 */
5099 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5100 if (!r || pfn >= memblock_region_memory_end_pfn(r)) {
5101 for_each_memblock(memory, tmp)
5102 if (pfn < memblock_region_memory_end_pfn(tmp))
5103 break;
5104 r = tmp;
5105 }
5106 if (pfn >= memblock_region_memory_base_pfn(r) &&
5107 memblock_is_mirror(r)) {
5108 /* already initialized as NORMAL */
5109 pfn = memblock_region_memory_end_pfn(r);
5110 continue;
5111 }
5112 }
5113#endif
5114
5115not_early:
5116 /*
5117 * Mark the block movable so that blocks are reserved for
5118 * movable at startup. This will force kernel allocations
5119 * to reserve their blocks rather than leaking throughout
5120 * the address space during boot when many long-lived
5121 * kernel allocations are made.
5122 *
5123 * bitmap is created for zone's valid pfn range. but memmap
5124 * can be created for invalid pages (for alignment)
5125 * check here not to call set_pageblock_migratetype() against
5126 * pfn out of zone.
5127 */
5128 if (!(pfn & (pageblock_nr_pages - 1))) {
5129 struct page *page = pfn_to_page(pfn);
5130
5131 __init_single_page(page, pfn, zone, nid);
5132 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5133 } else {
5134 __init_single_pfn(pfn, zone, nid);
5135 }
5136 }
5137}
5138
5139static void __meminit zone_init_free_lists(struct zone *zone)
5140{
5141 unsigned int order, t;
5142 for_each_migratetype_order(order, t) {
5143 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
5144 zone->free_area[order].nr_free = 0;
5145 }
5146}
5147
5148#ifndef __HAVE_ARCH_MEMMAP_INIT
5149#define memmap_init(size, nid, zone, start_pfn) \
5150 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
5151#endif
5152
5153static int zone_batchsize(struct zone *zone)
5154{
5155#ifdef CONFIG_MMU
5156 int batch;
5157
5158 /*
5159 * The per-cpu-pages pools are set to around 1000th of the
5160 * size of the zone. But no more than 1/2 of a meg.
5161 *
5162 * OK, so we don't know how big the cache is. So guess.
5163 */
5164 batch = zone->managed_pages / 1024;
5165 if (batch * PAGE_SIZE > 512 * 1024)
5166 batch = (512 * 1024) / PAGE_SIZE;
5167 batch /= 4; /* We effectively *= 4 below */
5168 if (batch < 1)
5169 batch = 1;
5170
5171 /*
5172 * Clamp the batch to a 2^n - 1 value. Having a power
5173 * of 2 value was found to be more likely to have
5174 * suboptimal cache aliasing properties in some cases.
5175 *
5176 * For example if 2 tasks are alternately allocating
5177 * batches of pages, one task can end up with a lot
5178 * of pages of one half of the possible page colors
5179 * and the other with pages of the other colors.
5180 */
5181 batch = rounddown_pow_of_two(batch + batch/2) - 1;
5182
5183 return batch;
5184
5185#else
5186 /* The deferral and batching of frees should be suppressed under NOMMU
5187 * conditions.
5188 *
5189 * The problem is that NOMMU needs to be able to allocate large chunks
5190 * of contiguous memory as there's no hardware page translation to
5191 * assemble apparent contiguous memory from discontiguous pages.
5192 *
5193 * Queueing large contiguous runs of pages for batching, however,
5194 * causes the pages to actually be freed in smaller chunks. As there
5195 * can be a significant delay between the individual batches being
5196 * recycled, this leads to the once large chunks of space being
5197 * fragmented and becoming unavailable for high-order allocations.
5198 */
5199 return 0;
5200#endif
5201}
5202
5203/*
5204 * pcp->high and pcp->batch values are related and dependent on one another:
5205 * ->batch must never be higher then ->high.
5206 * The following function updates them in a safe manner without read side
5207 * locking.
5208 *
5209 * Any new users of pcp->batch and pcp->high should ensure they can cope with
5210 * those fields changing asynchronously (acording the the above rule).
5211 *
5212 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5213 * outside of boot time (or some other assurance that no concurrent updaters
5214 * exist).
5215 */
5216static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
5217 unsigned long batch)
5218{
5219 /* start with a fail safe value for batch */
5220 pcp->batch = 1;
5221 smp_wmb();
5222
5223 /* Update high, then batch, in order */
5224 pcp->high = high;
5225 smp_wmb();
5226
5227 pcp->batch = batch;
5228}
5229
5230/* a companion to pageset_set_high() */
5231static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
5232{
5233 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
5234}
5235
5236static void pageset_init(struct per_cpu_pageset *p)
5237{
5238 struct per_cpu_pages *pcp;
5239 int migratetype;
5240
5241 memset(p, 0, sizeof(*p));
5242
5243 pcp = &p->pcp;
5244 pcp->count = 0;
5245 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
5246 INIT_LIST_HEAD(&pcp->lists[migratetype]);
5247}
5248
5249static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
5250{
5251 pageset_init(p);
5252 pageset_set_batch(p, batch);
5253}
5254
5255/*
5256 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5257 * to the value high for the pageset p.
5258 */
5259static void pageset_set_high(struct per_cpu_pageset *p,
5260 unsigned long high)
5261{
5262 unsigned long batch = max(1UL, high / 4);
5263 if ((high / 4) > (PAGE_SHIFT * 8))
5264 batch = PAGE_SHIFT * 8;
5265
5266 pageset_update(&p->pcp, high, batch);
5267}
5268
5269static void pageset_set_high_and_batch(struct zone *zone,
5270 struct per_cpu_pageset *pcp)
5271{
5272 if (percpu_pagelist_fraction)
5273 pageset_set_high(pcp,
5274 (zone->managed_pages /
5275 percpu_pagelist_fraction));
5276 else
5277 pageset_set_batch(pcp, zone_batchsize(zone));
5278}
5279
5280static void __meminit zone_pageset_init(struct zone *zone, int cpu)
5281{
5282 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
5283
5284 pageset_init(pcp);
5285 pageset_set_high_and_batch(zone, pcp);
5286}
5287
5288static void __meminit setup_zone_pageset(struct zone *zone)
5289{
5290 int cpu;
5291 zone->pageset = alloc_percpu(struct per_cpu_pageset);
5292 for_each_possible_cpu(cpu)
5293 zone_pageset_init(zone, cpu);
5294}
5295
5296/*
5297 * Allocate per cpu pagesets and initialize them.
5298 * Before this call only boot pagesets were available.
5299 */
5300void __init setup_per_cpu_pageset(void)
5301{
5302 struct pglist_data *pgdat;
5303 struct zone *zone;
5304
5305 for_each_populated_zone(zone)
5306 setup_zone_pageset(zone);
5307
5308 for_each_online_pgdat(pgdat)
5309 pgdat->per_cpu_nodestats =
5310 alloc_percpu(struct per_cpu_nodestat);
5311}
5312
5313static __meminit void zone_pcp_init(struct zone *zone)
5314{
5315 /*
5316 * per cpu subsystem is not up at this point. The following code
5317 * relies on the ability of the linker to provide the
5318 * offset of a (static) per cpu variable into the per cpu area.
5319 */
5320 zone->pageset = &boot_pageset;
5321
5322 if (populated_zone(zone))
5323 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
5324 zone->name, zone->present_pages,
5325 zone_batchsize(zone));
5326}
5327
5328int __meminit init_currently_empty_zone(struct zone *zone,
5329 unsigned long zone_start_pfn,
5330 unsigned long size)
5331{
5332 struct pglist_data *pgdat = zone->zone_pgdat;
5333
5334 pgdat->nr_zones = zone_idx(zone) + 1;
5335
5336 zone->zone_start_pfn = zone_start_pfn;
5337
5338 mminit_dprintk(MMINIT_TRACE, "memmap_init",
5339 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5340 pgdat->node_id,
5341 (unsigned long)zone_idx(zone),
5342 zone_start_pfn, (zone_start_pfn + size));
5343
5344 zone_init_free_lists(zone);
5345 zone->initialized = 1;
5346
5347 return 0;
5348}
5349
5350#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5351#ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5352
5353/*
5354 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5355 */
5356int __meminit __early_pfn_to_nid(unsigned long pfn,
5357 struct mminit_pfnnid_cache *state)
5358{
5359 unsigned long start_pfn, end_pfn;
5360 int nid;
5361
5362 if (state->last_start <= pfn && pfn < state->last_end)
5363 return state->last_nid;
5364
5365 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
5366 if (nid != -1) {
5367 state->last_start = start_pfn;
5368 state->last_end = end_pfn;
5369 state->last_nid = nid;
5370 }
5371
5372 return nid;
5373}
5374#endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5375
5376/**
5377 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5378 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5379 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5380 *
5381 * If an architecture guarantees that all ranges registered contain no holes
5382 * and may be freed, this this function may be used instead of calling
5383 * memblock_free_early_nid() manually.
5384 */
5385void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
5386{
5387 unsigned long start_pfn, end_pfn;
5388 int i, this_nid;
5389
5390 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
5391 start_pfn = min(start_pfn, max_low_pfn);
5392 end_pfn = min(end_pfn, max_low_pfn);
5393
5394 if (start_pfn < end_pfn)
5395 memblock_free_early_nid(PFN_PHYS(start_pfn),
5396 (end_pfn - start_pfn) << PAGE_SHIFT,
5397 this_nid);
5398 }
5399}
5400
5401/**
5402 * sparse_memory_present_with_active_regions - Call memory_present for each active range
5403 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
5404 *
5405 * If an architecture guarantees that all ranges registered contain no holes and may
5406 * be freed, this function may be used instead of calling memory_present() manually.
5407 */
5408void __init sparse_memory_present_with_active_regions(int nid)
5409{
5410 unsigned long start_pfn, end_pfn;
5411 int i, this_nid;
5412
5413 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
5414 memory_present(this_nid, start_pfn, end_pfn);
5415}
5416
5417/**
5418 * get_pfn_range_for_nid - Return the start and end page frames for a node
5419 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5420 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5421 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5422 *
5423 * It returns the start and end page frame of a node based on information
5424 * provided by memblock_set_node(). If called for a node
5425 * with no available memory, a warning is printed and the start and end
5426 * PFNs will be 0.
5427 */
5428void __meminit get_pfn_range_for_nid(unsigned int nid,
5429 unsigned long *start_pfn, unsigned long *end_pfn)
5430{
5431 unsigned long this_start_pfn, this_end_pfn;
5432 int i;
5433
5434 *start_pfn = -1UL;
5435 *end_pfn = 0;
5436
5437 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
5438 *start_pfn = min(*start_pfn, this_start_pfn);
5439 *end_pfn = max(*end_pfn, this_end_pfn);
5440 }
5441
5442 if (*start_pfn == -1UL)
5443 *start_pfn = 0;
5444}
5445
5446/*
5447 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5448 * assumption is made that zones within a node are ordered in monotonic
5449 * increasing memory addresses so that the "highest" populated zone is used
5450 */
5451static void __init find_usable_zone_for_movable(void)
5452{
5453 int zone_index;
5454 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
5455 if (zone_index == ZONE_MOVABLE)
5456 continue;
5457
5458 if (arch_zone_highest_possible_pfn[zone_index] >
5459 arch_zone_lowest_possible_pfn[zone_index])
5460 break;
5461 }
5462
5463 VM_BUG_ON(zone_index == -1);
5464 movable_zone = zone_index;
5465}
5466
5467/*
5468 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5469 * because it is sized independent of architecture. Unlike the other zones,
5470 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5471 * in each node depending on the size of each node and how evenly kernelcore
5472 * is distributed. This helper function adjusts the zone ranges
5473 * provided by the architecture for a given node by using the end of the
5474 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5475 * zones within a node are in order of monotonic increases memory addresses
5476 */
5477static void __meminit adjust_zone_range_for_zone_movable(int nid,
5478 unsigned long zone_type,
5479 unsigned long node_start_pfn,
5480 unsigned long node_end_pfn,
5481 unsigned long *zone_start_pfn,
5482 unsigned long *zone_end_pfn)
5483{
5484 /* Only adjust if ZONE_MOVABLE is on this node */
5485 if (zone_movable_pfn[nid]) {
5486 /* Size ZONE_MOVABLE */
5487 if (zone_type == ZONE_MOVABLE) {
5488 *zone_start_pfn = zone_movable_pfn[nid];
5489 *zone_end_pfn = min(node_end_pfn,
5490 arch_zone_highest_possible_pfn[movable_zone]);
5491
5492 /* Adjust for ZONE_MOVABLE starting within this range */
5493 } else if (!mirrored_kernelcore &&
5494 *zone_start_pfn < zone_movable_pfn[nid] &&
5495 *zone_end_pfn > zone_movable_pfn[nid]) {
5496 *zone_end_pfn = zone_movable_pfn[nid];
5497
5498 /* Check if this whole range is within ZONE_MOVABLE */
5499 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
5500 *zone_start_pfn = *zone_end_pfn;
5501 }
5502}
5503
5504/*
5505 * Return the number of pages a zone spans in a node, including holes
5506 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5507 */
5508static unsigned long __meminit zone_spanned_pages_in_node(int nid,
5509 unsigned long zone_type,
5510 unsigned long node_start_pfn,
5511 unsigned long node_end_pfn,
5512 unsigned long *zone_start_pfn,
5513 unsigned long *zone_end_pfn,
5514 unsigned long *ignored)
5515{
5516 /* When hotadd a new node from cpu_up(), the node should be empty */
5517 if (!node_start_pfn && !node_end_pfn)
5518 return 0;
5519
5520 /* Get the start and end of the zone */
5521 *zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
5522 *zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
5523 adjust_zone_range_for_zone_movable(nid, zone_type,
5524 node_start_pfn, node_end_pfn,
5525 zone_start_pfn, zone_end_pfn);
5526
5527 /* Check that this node has pages within the zone's required range */
5528 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
5529 return 0;
5530
5531 /* Move the zone boundaries inside the node if necessary */
5532 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
5533 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
5534
5535 /* Return the spanned pages */
5536 return *zone_end_pfn - *zone_start_pfn;
5537}
5538
5539/*
5540 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5541 * then all holes in the requested range will be accounted for.
5542 */
5543unsigned long __meminit __absent_pages_in_range(int nid,
5544 unsigned long range_start_pfn,
5545 unsigned long range_end_pfn)
5546{
5547 unsigned long nr_absent = range_end_pfn - range_start_pfn;
5548 unsigned long start_pfn, end_pfn;
5549 int i;
5550
5551 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5552 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
5553 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
5554 nr_absent -= end_pfn - start_pfn;
5555 }
5556 return nr_absent;
5557}
5558
5559/**
5560 * absent_pages_in_range - Return number of page frames in holes within a range
5561 * @start_pfn: The start PFN to start searching for holes
5562 * @end_pfn: The end PFN to stop searching for holes
5563 *
5564 * It returns the number of pages frames in memory holes within a range.
5565 */
5566unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5567 unsigned long end_pfn)
5568{
5569 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5570}
5571
5572/* Return the number of page frames in holes in a zone on a node */
5573static unsigned long __meminit zone_absent_pages_in_node(int nid,
5574 unsigned long zone_type,
5575 unsigned long node_start_pfn,
5576 unsigned long node_end_pfn,
5577 unsigned long *ignored)
5578{
5579 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5580 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5581 unsigned long zone_start_pfn, zone_end_pfn;
5582 unsigned long nr_absent;
5583
5584 /* When hotadd a new node from cpu_up(), the node should be empty */
5585 if (!node_start_pfn && !node_end_pfn)
5586 return 0;
5587
5588 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5589 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5590
5591 adjust_zone_range_for_zone_movable(nid, zone_type,
5592 node_start_pfn, node_end_pfn,
5593 &zone_start_pfn, &zone_end_pfn);
5594 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5595
5596 /*
5597 * ZONE_MOVABLE handling.
5598 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
5599 * and vice versa.
5600 */
5601 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
5602 unsigned long start_pfn, end_pfn;
5603 struct memblock_region *r;
5604
5605 for_each_memblock(memory, r) {
5606 start_pfn = clamp(memblock_region_memory_base_pfn(r),
5607 zone_start_pfn, zone_end_pfn);
5608 end_pfn = clamp(memblock_region_memory_end_pfn(r),
5609 zone_start_pfn, zone_end_pfn);
5610
5611 if (zone_type == ZONE_MOVABLE &&
5612 memblock_is_mirror(r))
5613 nr_absent += end_pfn - start_pfn;
5614
5615 if (zone_type == ZONE_NORMAL &&
5616 !memblock_is_mirror(r))
5617 nr_absent += end_pfn - start_pfn;
5618 }
5619 }
5620
5621 return nr_absent;
5622}
5623
5624#else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5625static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5626 unsigned long zone_type,
5627 unsigned long node_start_pfn,
5628 unsigned long node_end_pfn,
5629 unsigned long *zone_start_pfn,
5630 unsigned long *zone_end_pfn,
5631 unsigned long *zones_size)
5632{
5633 unsigned int zone;
5634
5635 *zone_start_pfn = node_start_pfn;
5636 for (zone = 0; zone < zone_type; zone++)
5637 *zone_start_pfn += zones_size[zone];
5638
5639 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
5640
5641 return zones_size[zone_type];
5642}
5643
5644static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
5645 unsigned long zone_type,
5646 unsigned long node_start_pfn,
5647 unsigned long node_end_pfn,
5648 unsigned long *zholes_size)
5649{
5650 if (!zholes_size)
5651 return 0;
5652
5653 return zholes_size[zone_type];
5654}
5655
5656#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5657
5658static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5659 unsigned long node_start_pfn,
5660 unsigned long node_end_pfn,
5661 unsigned long *zones_size,
5662 unsigned long *zholes_size)
5663{
5664 unsigned long realtotalpages = 0, totalpages = 0;
5665 enum zone_type i;
5666
5667 for (i = 0; i < MAX_NR_ZONES; i++) {
5668 struct zone *zone = pgdat->node_zones + i;
5669 unsigned long zone_start_pfn, zone_end_pfn;
5670 unsigned long size, real_size;
5671
5672 size = zone_spanned_pages_in_node(pgdat->node_id, i,
5673 node_start_pfn,
5674 node_end_pfn,
5675 &zone_start_pfn,
5676 &zone_end_pfn,
5677 zones_size);
5678 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5679 node_start_pfn, node_end_pfn,
5680 zholes_size);
5681 if (size)
5682 zone->zone_start_pfn = zone_start_pfn;
5683 else
5684 zone->zone_start_pfn = 0;
5685 zone->spanned_pages = size;
5686 zone->present_pages = real_size;
5687
5688 totalpages += size;
5689 realtotalpages += real_size;
5690 }
5691
5692 pgdat->node_spanned_pages = totalpages;
5693 pgdat->node_present_pages = realtotalpages;
5694 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5695 realtotalpages);
5696}
5697
5698#ifndef CONFIG_SPARSEMEM
5699/*
5700 * Calculate the size of the zone->blockflags rounded to an unsigned long
5701 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5702 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5703 * round what is now in bits to nearest long in bits, then return it in
5704 * bytes.
5705 */
5706static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
5707{
5708 unsigned long usemapsize;
5709
5710 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
5711 usemapsize = roundup(zonesize, pageblock_nr_pages);
5712 usemapsize = usemapsize >> pageblock_order;
5713 usemapsize *= NR_PAGEBLOCK_BITS;
5714 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
5715
5716 return usemapsize / 8;
5717}
5718
5719static void __init setup_usemap(struct pglist_data *pgdat,
5720 struct zone *zone,
5721 unsigned long zone_start_pfn,
5722 unsigned long zonesize)
5723{
5724 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
5725 zone->pageblock_flags = NULL;
5726 if (usemapsize)
5727 zone->pageblock_flags =
5728 memblock_virt_alloc_node_nopanic(usemapsize,
5729 pgdat->node_id);
5730}
5731#else
5732static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
5733 unsigned long zone_start_pfn, unsigned long zonesize) {}
5734#endif /* CONFIG_SPARSEMEM */
5735
5736#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5737
5738/* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5739void __paginginit set_pageblock_order(void)
5740{
5741 unsigned int order;
5742
5743 /* Check that pageblock_nr_pages has not already been setup */
5744 if (pageblock_order)
5745 return;
5746
5747 if (HPAGE_SHIFT > PAGE_SHIFT)
5748 order = HUGETLB_PAGE_ORDER;
5749 else
5750 order = MAX_ORDER - 1;
5751
5752 /*
5753 * Assume the largest contiguous order of interest is a huge page.
5754 * This value may be variable depending on boot parameters on IA64 and
5755 * powerpc.
5756 */
5757 pageblock_order = order;
5758}
5759#else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5760
5761/*
5762 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5763 * is unused as pageblock_order is set at compile-time. See
5764 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5765 * the kernel config
5766 */
5767void __paginginit set_pageblock_order(void)
5768{
5769}
5770
5771#endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5772
5773static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
5774 unsigned long present_pages)
5775{
5776 unsigned long pages = spanned_pages;
5777
5778 /*
5779 * Provide a more accurate estimation if there are holes within
5780 * the zone and SPARSEMEM is in use. If there are holes within the
5781 * zone, each populated memory region may cost us one or two extra
5782 * memmap pages due to alignment because memmap pages for each
5783 * populated regions may not naturally algined on page boundary.
5784 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5785 */
5786 if (spanned_pages > present_pages + (present_pages >> 4) &&
5787 IS_ENABLED(CONFIG_SPARSEMEM))
5788 pages = present_pages;
5789
5790 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
5791}
5792
5793/*
5794 * Set up the zone data structures:
5795 * - mark all pages reserved
5796 * - mark all memory queues empty
5797 * - clear the memory bitmaps
5798 *
5799 * NOTE: pgdat should get zeroed by caller.
5800 */
5801static void __paginginit free_area_init_core(struct pglist_data *pgdat)
5802{
5803 enum zone_type j;
5804 int nid = pgdat->node_id;
5805 int ret;
5806
5807 pgdat_resize_init(pgdat);
5808#ifdef CONFIG_NUMA_BALANCING
5809 spin_lock_init(&pgdat->numabalancing_migrate_lock);
5810 pgdat->numabalancing_migrate_nr_pages = 0;
5811 pgdat->numabalancing_migrate_next_window = jiffies;
5812#endif
5813#ifdef CONFIG_TRANSPARENT_HUGEPAGE
5814 spin_lock_init(&pgdat->split_queue_lock);
5815 INIT_LIST_HEAD(&pgdat->split_queue);
5816 pgdat->split_queue_len = 0;
5817#endif
5818 init_waitqueue_head(&pgdat->kswapd_wait);
5819 init_waitqueue_head(&pgdat->pfmemalloc_wait);
5820#ifdef CONFIG_COMPACTION
5821 init_waitqueue_head(&pgdat->kcompactd_wait);
5822#endif
5823 pgdat_page_ext_init(pgdat);
5824 spin_lock_init(&pgdat->lru_lock);
5825 lruvec_init(node_lruvec(pgdat));
5826
5827 for (j = 0; j < MAX_NR_ZONES; j++) {
5828 struct zone *zone = pgdat->node_zones + j;
5829 unsigned long size, realsize, freesize, memmap_pages;
5830 unsigned long zone_start_pfn = zone->zone_start_pfn;
5831
5832 size = zone->spanned_pages;
5833 realsize = freesize = zone->present_pages;
5834
5835 /*
5836 * Adjust freesize so that it accounts for how much memory
5837 * is used by this zone for memmap. This affects the watermark
5838 * and per-cpu initialisations
5839 */
5840 memmap_pages = calc_memmap_size(size, realsize);
5841 if (!is_highmem_idx(j)) {
5842 if (freesize >= memmap_pages) {
5843 freesize -= memmap_pages;
5844 if (memmap_pages)
5845 printk(KERN_DEBUG
5846 " %s zone: %lu pages used for memmap\n",
5847 zone_names[j], memmap_pages);
5848 } else
5849 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
5850 zone_names[j], memmap_pages, freesize);
5851 }
5852
5853 /* Account for reserved pages */
5854 if (j == 0 && freesize > dma_reserve) {
5855 freesize -= dma_reserve;
5856 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
5857 zone_names[0], dma_reserve);
5858 }
5859
5860 if (!is_highmem_idx(j))
5861 nr_kernel_pages += freesize;
5862 /* Charge for highmem memmap if there are enough kernel pages */
5863 else if (nr_kernel_pages > memmap_pages * 2)
5864 nr_kernel_pages -= memmap_pages;
5865 nr_all_pages += freesize;
5866
5867 /*
5868 * Set an approximate value for lowmem here, it will be adjusted
5869 * when the bootmem allocator frees pages into the buddy system.
5870 * And all highmem pages will be managed by the buddy system.
5871 */
5872 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
5873#ifdef CONFIG_NUMA
5874 zone->node = nid;
5875#endif
5876 zone->name = zone_names[j];
5877 zone->zone_pgdat = pgdat;
5878 spin_lock_init(&zone->lock);
5879 zone_seqlock_init(zone);
5880 zone_pcp_init(zone);
5881
5882 if (!size)
5883 continue;
5884
5885 set_pageblock_order();
5886 setup_usemap(pgdat, zone, zone_start_pfn, size);
5887 ret = init_currently_empty_zone(zone, zone_start_pfn, size);
5888 BUG_ON(ret);
5889 memmap_init(size, nid, j, zone_start_pfn);
5890 }
5891}
5892
5893static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
5894{
5895 unsigned long __maybe_unused start = 0;
5896 unsigned long __maybe_unused offset = 0;
5897
5898 /* Skip empty nodes */
5899 if (!pgdat->node_spanned_pages)
5900 return;
5901
5902#ifdef CONFIG_FLAT_NODE_MEM_MAP
5903 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
5904 offset = pgdat->node_start_pfn - start;
5905 /* ia64 gets its own node_mem_map, before this, without bootmem */
5906 if (!pgdat->node_mem_map) {
5907 unsigned long size, end;
5908 struct page *map;
5909
5910 /*
5911 * The zone's endpoints aren't required to be MAX_ORDER
5912 * aligned but the node_mem_map endpoints must be in order
5913 * for the buddy allocator to function correctly.
5914 */
5915 end = pgdat_end_pfn(pgdat);
5916 end = ALIGN(end, MAX_ORDER_NR_PAGES);
5917 size = (end - start) * sizeof(struct page);
5918 map = alloc_remap(pgdat->node_id, size);
5919 if (!map)
5920 map = memblock_virt_alloc_node_nopanic(size,
5921 pgdat->node_id);
5922 pgdat->node_mem_map = map + offset;
5923 }
5924#ifndef CONFIG_NEED_MULTIPLE_NODES
5925 /*
5926 * With no DISCONTIG, the global mem_map is just set as node 0's
5927 */
5928 if (pgdat == NODE_DATA(0)) {
5929 mem_map = NODE_DATA(0)->node_mem_map;
5930#if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
5931 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
5932 mem_map -= offset;
5933#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5934 }
5935#endif
5936#endif /* CONFIG_FLAT_NODE_MEM_MAP */
5937}
5938
5939void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
5940 unsigned long node_start_pfn, unsigned long *zholes_size)
5941{
5942 pg_data_t *pgdat = NODE_DATA(nid);
5943 unsigned long start_pfn = 0;
5944 unsigned long end_pfn = 0;
5945
5946 /* pg_data_t should be reset to zero when it's allocated */
5947 WARN_ON(pgdat->nr_zones || pgdat->kswapd_classzone_idx);
5948
5949 reset_deferred_meminit(pgdat);
5950 pgdat->node_id = nid;
5951 pgdat->node_start_pfn = node_start_pfn;
5952 pgdat->per_cpu_nodestats = NULL;
5953#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5954 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
5955 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
5956 (u64)start_pfn << PAGE_SHIFT,
5957 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
5958#else
5959 start_pfn = node_start_pfn;
5960#endif
5961 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
5962 zones_size, zholes_size);
5963
5964 alloc_node_mem_map(pgdat);
5965#ifdef CONFIG_FLAT_NODE_MEM_MAP
5966 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5967 nid, (unsigned long)pgdat,
5968 (unsigned long)pgdat->node_mem_map);
5969#endif
5970
5971 free_area_init_core(pgdat);
5972}
5973
5974#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5975
5976#if MAX_NUMNODES > 1
5977/*
5978 * Figure out the number of possible node ids.
5979 */
5980void __init setup_nr_node_ids(void)
5981{
5982 unsigned int highest;
5983
5984 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
5985 nr_node_ids = highest + 1;
5986}
5987#endif
5988
5989/**
5990 * node_map_pfn_alignment - determine the maximum internode alignment
5991 *
5992 * This function should be called after node map is populated and sorted.
5993 * It calculates the maximum power of two alignment which can distinguish
5994 * all the nodes.
5995 *
5996 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5997 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5998 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5999 * shifted, 1GiB is enough and this function will indicate so.
6000 *
6001 * This is used to test whether pfn -> nid mapping of the chosen memory
6002 * model has fine enough granularity to avoid incorrect mapping for the
6003 * populated node map.
6004 *
6005 * Returns the determined alignment in pfn's. 0 if there is no alignment
6006 * requirement (single node).
6007 */
6008unsigned long __init node_map_pfn_alignment(void)
6009{
6010 unsigned long accl_mask = 0, last_end = 0;
6011 unsigned long start, end, mask;
6012 int last_nid = -1;
6013 int i, nid;
6014
6015 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
6016 if (!start || last_nid < 0 || last_nid == nid) {
6017 last_nid = nid;
6018 last_end = end;
6019 continue;
6020 }
6021
6022 /*
6023 * Start with a mask granular enough to pin-point to the
6024 * start pfn and tick off bits one-by-one until it becomes
6025 * too coarse to separate the current node from the last.
6026 */
6027 mask = ~((1 << __ffs(start)) - 1);
6028 while (mask && last_end <= (start & (mask << 1)))
6029 mask <<= 1;
6030
6031 /* accumulate all internode masks */
6032 accl_mask |= mask;
6033 }
6034
6035 /* convert mask to number of pages */
6036 return ~accl_mask + 1;
6037}
6038
6039/* Find the lowest pfn for a node */
6040static unsigned long __init find_min_pfn_for_node(int nid)
6041{
6042 unsigned long min_pfn = ULONG_MAX;
6043 unsigned long start_pfn;
6044 int i;
6045
6046 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
6047 min_pfn = min(min_pfn, start_pfn);
6048
6049 if (min_pfn == ULONG_MAX) {
6050 pr_warn("Could not find start_pfn for node %d\n", nid);
6051 return 0;
6052 }
6053
6054 return min_pfn;
6055}
6056
6057/**
6058 * find_min_pfn_with_active_regions - Find the minimum PFN registered
6059 *
6060 * It returns the minimum PFN based on information provided via
6061 * memblock_set_node().
6062 */
6063unsigned long __init find_min_pfn_with_active_regions(void)
6064{
6065 return find_min_pfn_for_node(MAX_NUMNODES);
6066}
6067
6068/*
6069 * early_calculate_totalpages()
6070 * Sum pages in active regions for movable zone.
6071 * Populate N_MEMORY for calculating usable_nodes.
6072 */
6073static unsigned long __init early_calculate_totalpages(void)
6074{
6075 unsigned long totalpages = 0;
6076 unsigned long start_pfn, end_pfn;
6077 int i, nid;
6078
6079 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6080 unsigned long pages = end_pfn - start_pfn;
6081
6082 totalpages += pages;
6083 if (pages)
6084 node_set_state(nid, N_MEMORY);
6085 }
6086 return totalpages;
6087}
6088
6089/*
6090 * Find the PFN the Movable zone begins in each node. Kernel memory
6091 * is spread evenly between nodes as long as the nodes have enough
6092 * memory. When they don't, some nodes will have more kernelcore than
6093 * others
6094 */
6095static void __init find_zone_movable_pfns_for_nodes(void)
6096{
6097 int i, nid;
6098 unsigned long usable_startpfn;
6099 unsigned long kernelcore_node, kernelcore_remaining;
6100 /* save the state before borrow the nodemask */
6101 nodemask_t saved_node_state = node_states[N_MEMORY];
6102 unsigned long totalpages = early_calculate_totalpages();
6103 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
6104 struct memblock_region *r;
6105
6106 /* Need to find movable_zone earlier when movable_node is specified. */
6107 find_usable_zone_for_movable();
6108
6109 /*
6110 * If movable_node is specified, ignore kernelcore and movablecore
6111 * options.
6112 */
6113 if (movable_node_is_enabled()) {
6114 for_each_memblock(memory, r) {
6115 if (!memblock_is_hotpluggable(r))
6116 continue;
6117
6118 nid = r->nid;
6119
6120 usable_startpfn = PFN_DOWN(r->base);
6121 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6122 min(usable_startpfn, zone_movable_pfn[nid]) :
6123 usable_startpfn;
6124 }
6125
6126 goto out2;
6127 }
6128
6129 /*
6130 * If kernelcore=mirror is specified, ignore movablecore option
6131 */
6132 if (mirrored_kernelcore) {
6133 bool mem_below_4gb_not_mirrored = false;
6134
6135 for_each_memblock(memory, r) {
6136 if (memblock_is_mirror(r))
6137 continue;
6138
6139 nid = r->nid;
6140
6141 usable_startpfn = memblock_region_memory_base_pfn(r);
6142
6143 if (usable_startpfn < 0x100000) {
6144 mem_below_4gb_not_mirrored = true;
6145 continue;
6146 }
6147
6148 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6149 min(usable_startpfn, zone_movable_pfn[nid]) :
6150 usable_startpfn;
6151 }
6152
6153 if (mem_below_4gb_not_mirrored)
6154 pr_warn("This configuration results in unmirrored kernel memory.");
6155
6156 goto out2;
6157 }
6158
6159 /*
6160 * If movablecore=nn[KMG] was specified, calculate what size of
6161 * kernelcore that corresponds so that memory usable for
6162 * any allocation type is evenly spread. If both kernelcore
6163 * and movablecore are specified, then the value of kernelcore
6164 * will be used for required_kernelcore if it's greater than
6165 * what movablecore would have allowed.
6166 */
6167 if (required_movablecore) {
6168 unsigned long corepages;
6169
6170 /*
6171 * Round-up so that ZONE_MOVABLE is at least as large as what
6172 * was requested by the user
6173 */
6174 required_movablecore =
6175 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
6176 required_movablecore = min(totalpages, required_movablecore);
6177 corepages = totalpages - required_movablecore;
6178
6179 required_kernelcore = max(required_kernelcore, corepages);
6180 }
6181
6182 /*
6183 * If kernelcore was not specified or kernelcore size is larger
6184 * than totalpages, there is no ZONE_MOVABLE.
6185 */
6186 if (!required_kernelcore || required_kernelcore >= totalpages)
6187 goto out;
6188
6189 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
6190 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
6191
6192restart:
6193 /* Spread kernelcore memory as evenly as possible throughout nodes */
6194 kernelcore_node = required_kernelcore / usable_nodes;
6195 for_each_node_state(nid, N_MEMORY) {
6196 unsigned long start_pfn, end_pfn;
6197
6198 /*
6199 * Recalculate kernelcore_node if the division per node
6200 * now exceeds what is necessary to satisfy the requested
6201 * amount of memory for the kernel
6202 */
6203 if (required_kernelcore < kernelcore_node)
6204 kernelcore_node = required_kernelcore / usable_nodes;
6205
6206 /*
6207 * As the map is walked, we track how much memory is usable
6208 * by the kernel using kernelcore_remaining. When it is
6209 * 0, the rest of the node is usable by ZONE_MOVABLE
6210 */
6211 kernelcore_remaining = kernelcore_node;
6212
6213 /* Go through each range of PFNs within this node */
6214 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6215 unsigned long size_pages;
6216
6217 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
6218 if (start_pfn >= end_pfn)
6219 continue;
6220
6221 /* Account for what is only usable for kernelcore */
6222 if (start_pfn < usable_startpfn) {
6223 unsigned long kernel_pages;
6224 kernel_pages = min(end_pfn, usable_startpfn)
6225 - start_pfn;
6226
6227 kernelcore_remaining -= min(kernel_pages,
6228 kernelcore_remaining);
6229 required_kernelcore -= min(kernel_pages,
6230 required_kernelcore);
6231
6232 /* Continue if range is now fully accounted */
6233 if (end_pfn <= usable_startpfn) {
6234
6235 /*
6236 * Push zone_movable_pfn to the end so
6237 * that if we have to rebalance
6238 * kernelcore across nodes, we will
6239 * not double account here
6240 */
6241 zone_movable_pfn[nid] = end_pfn;
6242 continue;
6243 }
6244 start_pfn = usable_startpfn;
6245 }
6246
6247 /*
6248 * The usable PFN range for ZONE_MOVABLE is from
6249 * start_pfn->end_pfn. Calculate size_pages as the
6250 * number of pages used as kernelcore
6251 */
6252 size_pages = end_pfn - start_pfn;
6253 if (size_pages > kernelcore_remaining)
6254 size_pages = kernelcore_remaining;
6255 zone_movable_pfn[nid] = start_pfn + size_pages;
6256
6257 /*
6258 * Some kernelcore has been met, update counts and
6259 * break if the kernelcore for this node has been
6260 * satisfied
6261 */
6262 required_kernelcore -= min(required_kernelcore,
6263 size_pages);
6264 kernelcore_remaining -= size_pages;
6265 if (!kernelcore_remaining)
6266 break;
6267 }
6268 }
6269
6270 /*
6271 * If there is still required_kernelcore, we do another pass with one
6272 * less node in the count. This will push zone_movable_pfn[nid] further
6273 * along on the nodes that still have memory until kernelcore is
6274 * satisfied
6275 */
6276 usable_nodes--;
6277 if (usable_nodes && required_kernelcore > usable_nodes)
6278 goto restart;
6279
6280out2:
6281 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
6282 for (nid = 0; nid < MAX_NUMNODES; nid++)
6283 zone_movable_pfn[nid] =
6284 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
6285
6286out:
6287 /* restore the node_state */
6288 node_states[N_MEMORY] = saved_node_state;
6289}
6290
6291/* Any regular or high memory on that node ? */
6292static void check_for_memory(pg_data_t *pgdat, int nid)
6293{
6294 enum zone_type zone_type;
6295
6296 if (N_MEMORY == N_NORMAL_MEMORY)
6297 return;
6298
6299 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
6300 struct zone *zone = &pgdat->node_zones[zone_type];
6301 if (populated_zone(zone)) {
6302 node_set_state(nid, N_HIGH_MEMORY);
6303 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
6304 zone_type <= ZONE_NORMAL)
6305 node_set_state(nid, N_NORMAL_MEMORY);
6306 break;
6307 }
6308 }
6309}
6310
6311/**
6312 * free_area_init_nodes - Initialise all pg_data_t and zone data
6313 * @max_zone_pfn: an array of max PFNs for each zone
6314 *
6315 * This will call free_area_init_node() for each active node in the system.
6316 * Using the page ranges provided by memblock_set_node(), the size of each
6317 * zone in each node and their holes is calculated. If the maximum PFN
6318 * between two adjacent zones match, it is assumed that the zone is empty.
6319 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
6320 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
6321 * starts where the previous one ended. For example, ZONE_DMA32 starts
6322 * at arch_max_dma_pfn.
6323 */
6324void __init free_area_init_nodes(unsigned long *max_zone_pfn)
6325{
6326 unsigned long start_pfn, end_pfn;
6327 int i, nid;
6328
6329 /* Record where the zone boundaries are */
6330 memset(arch_zone_lowest_possible_pfn, 0,
6331 sizeof(arch_zone_lowest_possible_pfn));
6332 memset(arch_zone_highest_possible_pfn, 0,
6333 sizeof(arch_zone_highest_possible_pfn));
6334
6335 start_pfn = find_min_pfn_with_active_regions();
6336
6337 for (i = 0; i < MAX_NR_ZONES; i++) {
6338 if (i == ZONE_MOVABLE)
6339 continue;
6340
6341 end_pfn = max(max_zone_pfn[i], start_pfn);
6342 arch_zone_lowest_possible_pfn[i] = start_pfn;
6343 arch_zone_highest_possible_pfn[i] = end_pfn;
6344
6345 start_pfn = end_pfn;
6346 }
6347 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
6348 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
6349
6350 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
6351 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
6352 find_zone_movable_pfns_for_nodes();
6353
6354 /* Print out the zone ranges */
6355 pr_info("Zone ranges:\n");
6356 for (i = 0; i < MAX_NR_ZONES; i++) {
6357 if (i == ZONE_MOVABLE)
6358 continue;
6359 pr_info(" %-8s ", zone_names[i]);
6360 if (arch_zone_lowest_possible_pfn[i] ==
6361 arch_zone_highest_possible_pfn[i])
6362 pr_cont("empty\n");
6363 else
6364 pr_cont("[mem %#018Lx-%#018Lx]\n",
6365 (u64)arch_zone_lowest_possible_pfn[i]
6366 << PAGE_SHIFT,
6367 ((u64)arch_zone_highest_possible_pfn[i]
6368 << PAGE_SHIFT) - 1);
6369 }
6370
6371 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
6372 pr_info("Movable zone start for each node\n");
6373 for (i = 0; i < MAX_NUMNODES; i++) {
6374 if (zone_movable_pfn[i])
6375 pr_info(" Node %d: %#018Lx\n", i,
6376 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
6377 }
6378
6379 /* Print out the early node map */
6380 pr_info("Early memory node ranges\n");
6381 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
6382 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
6383 (u64)start_pfn << PAGE_SHIFT,
6384 ((u64)end_pfn << PAGE_SHIFT) - 1);
6385
6386 /* Initialise every node */
6387 mminit_verify_pageflags_layout();
6388 setup_nr_node_ids();
6389 for_each_online_node(nid) {
6390 pg_data_t *pgdat = NODE_DATA(nid);
6391 free_area_init_node(nid, NULL,
6392 find_min_pfn_for_node(nid), NULL);
6393
6394 /* Any memory on that node */
6395 if (pgdat->node_present_pages)
6396 node_set_state(nid, N_MEMORY);
6397 check_for_memory(pgdat, nid);
6398 }
6399}
6400
6401static int __init cmdline_parse_core(char *p, unsigned long *core)
6402{
6403 unsigned long long coremem;
6404 if (!p)
6405 return -EINVAL;
6406
6407 coremem = memparse(p, &p);
6408 *core = coremem >> PAGE_SHIFT;
6409
6410 /* Paranoid check that UL is enough for the coremem value */
6411 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
6412
6413 return 0;
6414}
6415
6416/*
6417 * kernelcore=size sets the amount of memory for use for allocations that
6418 * cannot be reclaimed or migrated.
6419 */
6420static int __init cmdline_parse_kernelcore(char *p)
6421{
6422 /* parse kernelcore=mirror */
6423 if (parse_option_str(p, "mirror")) {
6424 mirrored_kernelcore = true;
6425 return 0;
6426 }
6427
6428 return cmdline_parse_core(p, &required_kernelcore);
6429}
6430
6431/*
6432 * movablecore=size sets the amount of memory for use for allocations that
6433 * can be reclaimed or migrated.
6434 */
6435static int __init cmdline_parse_movablecore(char *p)
6436{
6437 return cmdline_parse_core(p, &required_movablecore);
6438}
6439
6440early_param("kernelcore", cmdline_parse_kernelcore);
6441early_param("movablecore", cmdline_parse_movablecore);
6442
6443#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6444
6445void adjust_managed_page_count(struct page *page, long count)
6446{
6447 spin_lock(&managed_page_count_lock);
6448 page_zone(page)->managed_pages += count;
6449 totalram_pages += count;
6450#ifdef CONFIG_HIGHMEM
6451 if (PageHighMem(page))
6452 totalhigh_pages += count;
6453#endif
6454 spin_unlock(&managed_page_count_lock);
6455}
6456EXPORT_SYMBOL(adjust_managed_page_count);
6457
6458unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
6459{
6460 void *pos;
6461 unsigned long pages = 0;
6462
6463 start = (void *)PAGE_ALIGN((unsigned long)start);
6464 end = (void *)((unsigned long)end & PAGE_MASK);
6465 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
6466 if ((unsigned int)poison <= 0xFF)
6467 memset(pos, poison, PAGE_SIZE);
6468 free_reserved_page(virt_to_page(pos));
6469 }
6470
6471 if (pages && s)
6472 pr_info("Freeing %s memory: %ldK\n",
6473 s, pages << (PAGE_SHIFT - 10));
6474
6475 return pages;
6476}
6477EXPORT_SYMBOL(free_reserved_area);
6478
6479#ifdef CONFIG_HIGHMEM
6480void free_highmem_page(struct page *page)
6481{
6482 __free_reserved_page(page);
6483 totalram_pages++;
6484 page_zone(page)->managed_pages++;
6485 totalhigh_pages++;
6486}
6487#endif
6488
6489
6490void __init mem_init_print_info(const char *str)
6491{
6492 unsigned long physpages, codesize, datasize, rosize, bss_size;
6493 unsigned long init_code_size, init_data_size;
6494
6495 physpages = get_num_physpages();
6496 codesize = _etext - _stext;
6497 datasize = _edata - _sdata;
6498 rosize = __end_rodata - __start_rodata;
6499 bss_size = __bss_stop - __bss_start;
6500 init_data_size = __init_end - __init_begin;
6501 init_code_size = _einittext - _sinittext;
6502
6503 /*
6504 * Detect special cases and adjust section sizes accordingly:
6505 * 1) .init.* may be embedded into .data sections
6506 * 2) .init.text.* may be out of [__init_begin, __init_end],
6507 * please refer to arch/tile/kernel/vmlinux.lds.S.
6508 * 3) .rodata.* may be embedded into .text or .data sections.
6509 */
6510#define adj_init_size(start, end, size, pos, adj) \
6511 do { \
6512 if (start <= pos && pos < end && size > adj) \
6513 size -= adj; \
6514 } while (0)
6515
6516 adj_init_size(__init_begin, __init_end, init_data_size,
6517 _sinittext, init_code_size);
6518 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
6519 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
6520 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
6521 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
6522
6523#undef adj_init_size
6524
6525 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
6526#ifdef CONFIG_HIGHMEM
6527 ", %luK highmem"
6528#endif
6529 "%s%s)\n",
6530 nr_free_pages() << (PAGE_SHIFT - 10),
6531 physpages << (PAGE_SHIFT - 10),
6532 codesize >> 10, datasize >> 10, rosize >> 10,
6533 (init_data_size + init_code_size) >> 10, bss_size >> 10,
6534 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT - 10),
6535 totalcma_pages << (PAGE_SHIFT - 10),
6536#ifdef CONFIG_HIGHMEM
6537 totalhigh_pages << (PAGE_SHIFT - 10),
6538#endif
6539 str ? ", " : "", str ? str : "");
6540}
6541
6542/**
6543 * set_dma_reserve - set the specified number of pages reserved in the first zone
6544 * @new_dma_reserve: The number of pages to mark reserved
6545 *
6546 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
6547 * In the DMA zone, a significant percentage may be consumed by kernel image
6548 * and other unfreeable allocations which can skew the watermarks badly. This
6549 * function may optionally be used to account for unfreeable pages in the
6550 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
6551 * smaller per-cpu batchsize.
6552 */
6553void __init set_dma_reserve(unsigned long new_dma_reserve)
6554{
6555 dma_reserve = new_dma_reserve;
6556}
6557
6558void __init free_area_init(unsigned long *zones_size)
6559{
6560 free_area_init_node(0, zones_size,
6561 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
6562}
6563
6564static int page_alloc_cpu_dead(unsigned int cpu)
6565{
6566
6567 lru_add_drain_cpu(cpu);
6568 drain_pages(cpu);
6569
6570 /*
6571 * Spill the event counters of the dead processor
6572 * into the current processors event counters.
6573 * This artificially elevates the count of the current
6574 * processor.
6575 */
6576 vm_events_fold_cpu(cpu);
6577
6578 /*
6579 * Zero the differential counters of the dead processor
6580 * so that the vm statistics are consistent.
6581 *
6582 * This is only okay since the processor is dead and cannot
6583 * race with what we are doing.
6584 */
6585 cpu_vm_stats_fold(cpu);
6586 return 0;
6587}
6588
6589void __init page_alloc_init(void)
6590{
6591 int ret;
6592
6593 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
6594 "mm/page_alloc:dead", NULL,
6595 page_alloc_cpu_dead);
6596 WARN_ON(ret < 0);
6597}
6598
6599/*
6600 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6601 * or min_free_kbytes changes.
6602 */
6603static void calculate_totalreserve_pages(void)
6604{
6605 struct pglist_data *pgdat;
6606 unsigned long reserve_pages = 0;
6607 enum zone_type i, j;
6608
6609 for_each_online_pgdat(pgdat) {
6610
6611 pgdat->totalreserve_pages = 0;
6612
6613 for (i = 0; i < MAX_NR_ZONES; i++) {
6614 struct zone *zone = pgdat->node_zones + i;
6615 long max = 0;
6616
6617 /* Find valid and maximum lowmem_reserve in the zone */
6618 for (j = i; j < MAX_NR_ZONES; j++) {
6619 if (zone->lowmem_reserve[j] > max)
6620 max = zone->lowmem_reserve[j];
6621 }
6622
6623 /* we treat the high watermark as reserved pages. */
6624 max += high_wmark_pages(zone);
6625
6626 if (max > zone->managed_pages)
6627 max = zone->managed_pages;
6628
6629 pgdat->totalreserve_pages += max;
6630
6631 reserve_pages += max;
6632 }
6633 }
6634 totalreserve_pages = reserve_pages;
6635}
6636
6637/*
6638 * setup_per_zone_lowmem_reserve - called whenever
6639 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6640 * has a correct pages reserved value, so an adequate number of
6641 * pages are left in the zone after a successful __alloc_pages().
6642 */
6643static void setup_per_zone_lowmem_reserve(void)
6644{
6645 struct pglist_data *pgdat;
6646 enum zone_type j, idx;
6647
6648 for_each_online_pgdat(pgdat) {
6649 for (j = 0; j < MAX_NR_ZONES; j++) {
6650 struct zone *zone = pgdat->node_zones + j;
6651 unsigned long managed_pages = zone->managed_pages;
6652
6653 zone->lowmem_reserve[j] = 0;
6654
6655 idx = j;
6656 while (idx) {
6657 struct zone *lower_zone;
6658
6659 idx--;
6660
6661 if (sysctl_lowmem_reserve_ratio[idx] < 1)
6662 sysctl_lowmem_reserve_ratio[idx] = 1;
6663
6664 lower_zone = pgdat->node_zones + idx;
6665 lower_zone->lowmem_reserve[j] = managed_pages /
6666 sysctl_lowmem_reserve_ratio[idx];
6667 managed_pages += lower_zone->managed_pages;
6668 }
6669 }
6670 }
6671
6672 /* update totalreserve_pages */
6673 calculate_totalreserve_pages();
6674}
6675
6676static void __setup_per_zone_wmarks(void)
6677{
6678 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6679 unsigned long lowmem_pages = 0;
6680 struct zone *zone;
6681 unsigned long flags;
6682
6683 /* Calculate total number of !ZONE_HIGHMEM pages */
6684 for_each_zone(zone) {
6685 if (!is_highmem(zone))
6686 lowmem_pages += zone->managed_pages;
6687 }
6688
6689 for_each_zone(zone) {
6690 u64 tmp;
6691
6692 spin_lock_irqsave(&zone->lock, flags);
6693 tmp = (u64)pages_min * zone->managed_pages;
6694 do_div(tmp, lowmem_pages);
6695 if (is_highmem(zone)) {
6696 /*
6697 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6698 * need highmem pages, so cap pages_min to a small
6699 * value here.
6700 *
6701 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6702 * deltas control asynch page reclaim, and so should
6703 * not be capped for highmem.
6704 */
6705 unsigned long min_pages;
6706
6707 min_pages = zone->managed_pages / 1024;
6708 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
6709 zone->watermark[WMARK_MIN] = min_pages;
6710 } else {
6711 /*
6712 * If it's a lowmem zone, reserve a number of pages
6713 * proportionate to the zone's size.
6714 */
6715 zone->watermark[WMARK_MIN] = tmp;
6716 }
6717
6718 /*
6719 * Set the kswapd watermarks distance according to the
6720 * scale factor in proportion to available memory, but
6721 * ensure a minimum size on small systems.
6722 */
6723 tmp = max_t(u64, tmp >> 2,
6724 mult_frac(zone->managed_pages,
6725 watermark_scale_factor, 10000));
6726
6727 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
6728 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
6729
6730 spin_unlock_irqrestore(&zone->lock, flags);
6731 }
6732
6733 /* update totalreserve_pages */
6734 calculate_totalreserve_pages();
6735}
6736
6737/**
6738 * setup_per_zone_wmarks - called when min_free_kbytes changes
6739 * or when memory is hot-{added|removed}
6740 *
6741 * Ensures that the watermark[min,low,high] values for each zone are set
6742 * correctly with respect to min_free_kbytes.
6743 */
6744void setup_per_zone_wmarks(void)
6745{
6746 mutex_lock(&zonelists_mutex);
6747 __setup_per_zone_wmarks();
6748 mutex_unlock(&zonelists_mutex);
6749}
6750
6751/*
6752 * Initialise min_free_kbytes.
6753 *
6754 * For small machines we want it small (128k min). For large machines
6755 * we want it large (64MB max). But it is not linear, because network
6756 * bandwidth does not increase linearly with machine size. We use
6757 *
6758 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6759 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6760 *
6761 * which yields
6762 *
6763 * 16MB: 512k
6764 * 32MB: 724k
6765 * 64MB: 1024k
6766 * 128MB: 1448k
6767 * 256MB: 2048k
6768 * 512MB: 2896k
6769 * 1024MB: 4096k
6770 * 2048MB: 5792k
6771 * 4096MB: 8192k
6772 * 8192MB: 11584k
6773 * 16384MB: 16384k
6774 */
6775int __meminit init_per_zone_wmark_min(void)
6776{
6777 unsigned long lowmem_kbytes;
6778 int new_min_free_kbytes;
6779
6780 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
6781 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
6782
6783 if (new_min_free_kbytes > user_min_free_kbytes) {
6784 min_free_kbytes = new_min_free_kbytes;
6785 if (min_free_kbytes < 128)
6786 min_free_kbytes = 128;
6787 if (min_free_kbytes > 65536)
6788 min_free_kbytes = 65536;
6789 } else {
6790 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6791 new_min_free_kbytes, user_min_free_kbytes);
6792 }
6793 setup_per_zone_wmarks();
6794 refresh_zone_stat_thresholds();
6795 setup_per_zone_lowmem_reserve();
6796
6797#ifdef CONFIG_NUMA
6798 setup_min_unmapped_ratio();
6799 setup_min_slab_ratio();
6800#endif
6801
6802 return 0;
6803}
6804core_initcall(init_per_zone_wmark_min)
6805
6806/*
6807 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6808 * that we can call two helper functions whenever min_free_kbytes
6809 * changes.
6810 */
6811int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
6812 void __user *buffer, size_t *length, loff_t *ppos)
6813{
6814 int rc;
6815
6816 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6817 if (rc)
6818 return rc;
6819
6820 if (write) {
6821 user_min_free_kbytes = min_free_kbytes;
6822 setup_per_zone_wmarks();
6823 }
6824 return 0;
6825}
6826
6827int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
6828 void __user *buffer, size_t *length, loff_t *ppos)
6829{
6830 int rc;
6831
6832 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6833 if (rc)
6834 return rc;
6835
6836 if (write)
6837 setup_per_zone_wmarks();
6838
6839 return 0;
6840}
6841
6842#ifdef CONFIG_NUMA
6843static void setup_min_unmapped_ratio(void)
6844{
6845 pg_data_t *pgdat;
6846 struct zone *zone;
6847
6848 for_each_online_pgdat(pgdat)
6849 pgdat->min_unmapped_pages = 0;
6850
6851 for_each_zone(zone)
6852 zone->zone_pgdat->min_unmapped_pages += (zone->managed_pages *
6853 sysctl_min_unmapped_ratio) / 100;
6854}
6855
6856
6857int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
6858 void __user *buffer, size_t *length, loff_t *ppos)
6859{
6860 int rc;
6861
6862 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6863 if (rc)
6864 return rc;
6865
6866 setup_min_unmapped_ratio();
6867
6868 return 0;
6869}
6870
6871static void setup_min_slab_ratio(void)
6872{
6873 pg_data_t *pgdat;
6874 struct zone *zone;
6875
6876 for_each_online_pgdat(pgdat)
6877 pgdat->min_slab_pages = 0;
6878
6879 for_each_zone(zone)
6880 zone->zone_pgdat->min_slab_pages += (zone->managed_pages *
6881 sysctl_min_slab_ratio) / 100;
6882}
6883
6884int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
6885 void __user *buffer, size_t *length, loff_t *ppos)
6886{
6887 int rc;
6888
6889 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6890 if (rc)
6891 return rc;
6892
6893 setup_min_slab_ratio();
6894
6895 return 0;
6896}
6897#endif
6898
6899/*
6900 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6901 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6902 * whenever sysctl_lowmem_reserve_ratio changes.
6903 *
6904 * The reserve ratio obviously has absolutely no relation with the
6905 * minimum watermarks. The lowmem reserve ratio can only make sense
6906 * if in function of the boot time zone sizes.
6907 */
6908int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
6909 void __user *buffer, size_t *length, loff_t *ppos)
6910{
6911 proc_dointvec_minmax(table, write, buffer, length, ppos);
6912 setup_per_zone_lowmem_reserve();
6913 return 0;
6914}
6915
6916/*
6917 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6918 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6919 * pagelist can have before it gets flushed back to buddy allocator.
6920 */
6921int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
6922 void __user *buffer, size_t *length, loff_t *ppos)
6923{
6924 struct zone *zone;
6925 int old_percpu_pagelist_fraction;
6926 int ret;
6927
6928 mutex_lock(&pcp_batch_high_lock);
6929 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
6930
6931 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
6932 if (!write || ret < 0)
6933 goto out;
6934
6935 /* Sanity checking to avoid pcp imbalance */
6936 if (percpu_pagelist_fraction &&
6937 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
6938 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
6939 ret = -EINVAL;
6940 goto out;
6941 }
6942
6943 /* No change? */
6944 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
6945 goto out;
6946
6947 for_each_populated_zone(zone) {
6948 unsigned int cpu;
6949
6950 for_each_possible_cpu(cpu)
6951 pageset_set_high_and_batch(zone,
6952 per_cpu_ptr(zone->pageset, cpu));
6953 }
6954out:
6955 mutex_unlock(&pcp_batch_high_lock);
6956 return ret;
6957}
6958
6959#ifdef CONFIG_NUMA
6960int hashdist = HASHDIST_DEFAULT;
6961
6962static int __init set_hashdist(char *str)
6963{
6964 if (!str)
6965 return 0;
6966 hashdist = simple_strtoul(str, &str, 0);
6967 return 1;
6968}
6969__setup("hashdist=", set_hashdist);
6970#endif
6971
6972#ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
6973/*
6974 * Returns the number of pages that arch has reserved but
6975 * is not known to alloc_large_system_hash().
6976 */
6977static unsigned long __init arch_reserved_kernel_pages(void)
6978{
6979 return 0;
6980}
6981#endif
6982
6983/*
6984 * allocate a large system hash table from bootmem
6985 * - it is assumed that the hash table must contain an exact power-of-2
6986 * quantity of entries
6987 * - limit is the number of hash buckets, not the total allocation size
6988 */
6989void *__init alloc_large_system_hash(const char *tablename,
6990 unsigned long bucketsize,
6991 unsigned long numentries,
6992 int scale,
6993 int flags,
6994 unsigned int *_hash_shift,
6995 unsigned int *_hash_mask,
6996 unsigned long low_limit,
6997 unsigned long high_limit)
6998{
6999 unsigned long long max = high_limit;
7000 unsigned long log2qty, size;
7001 void *table = NULL;
7002
7003 /* allow the kernel cmdline to have a say */
7004 if (!numentries) {
7005 /* round applicable memory size up to nearest megabyte */
7006 numentries = nr_kernel_pages;
7007 numentries -= arch_reserved_kernel_pages();
7008
7009 /* It isn't necessary when PAGE_SIZE >= 1MB */
7010 if (PAGE_SHIFT < 20)
7011 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
7012
7013 /* limit to 1 bucket per 2^scale bytes of low memory */
7014 if (scale > PAGE_SHIFT)
7015 numentries >>= (scale - PAGE_SHIFT);
7016 else
7017 numentries <<= (PAGE_SHIFT - scale);
7018
7019 /* Make sure we've got at least a 0-order allocation.. */
7020 if (unlikely(flags & HASH_SMALL)) {
7021 /* Makes no sense without HASH_EARLY */
7022 WARN_ON(!(flags & HASH_EARLY));
7023 if (!(numentries >> *_hash_shift)) {
7024 numentries = 1UL << *_hash_shift;
7025 BUG_ON(!numentries);
7026 }
7027 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
7028 numentries = PAGE_SIZE / bucketsize;
7029 }
7030 numentries = roundup_pow_of_two(numentries);
7031
7032 /* limit allocation size to 1/16 total memory by default */
7033 if (max == 0) {
7034 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
7035 do_div(max, bucketsize);
7036 }
7037 max = min(max, 0x80000000ULL);
7038
7039 if (numentries < low_limit)
7040 numentries = low_limit;
7041 if (numentries > max)
7042 numentries = max;
7043
7044 log2qty = ilog2(numentries);
7045
7046 do {
7047 size = bucketsize << log2qty;
7048 if (flags & HASH_EARLY)
7049 table = memblock_virt_alloc_nopanic(size, 0);
7050 else if (hashdist)
7051 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
7052 else {
7053 /*
7054 * If bucketsize is not a power-of-two, we may free
7055 * some pages at the end of hash table which
7056 * alloc_pages_exact() automatically does
7057 */
7058 if (get_order(size) < MAX_ORDER) {
7059 table = alloc_pages_exact(size, GFP_ATOMIC);
7060 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
7061 }
7062 }
7063 } while (!table && size > PAGE_SIZE && --log2qty);
7064
7065 if (!table)
7066 panic("Failed to allocate %s hash table\n", tablename);
7067
7068 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
7069 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size);
7070
7071 if (_hash_shift)
7072 *_hash_shift = log2qty;
7073 if (_hash_mask)
7074 *_hash_mask = (1 << log2qty) - 1;
7075
7076 return table;
7077}
7078
7079/*
7080 * This function checks whether pageblock includes unmovable pages or not.
7081 * If @count is not zero, it is okay to include less @count unmovable pages
7082 *
7083 * PageLRU check without isolation or lru_lock could race so that
7084 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
7085 * expect this function should be exact.
7086 */
7087bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
7088 bool skip_hwpoisoned_pages)
7089{
7090 unsigned long pfn, iter, found;
7091 int mt;
7092
7093 /*
7094 * For avoiding noise data, lru_add_drain_all() should be called
7095 * If ZONE_MOVABLE, the zone never contains unmovable pages
7096 */
7097 if (zone_idx(zone) == ZONE_MOVABLE)
7098 return false;
7099 mt = get_pageblock_migratetype(page);
7100 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
7101 return false;
7102
7103 pfn = page_to_pfn(page);
7104 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
7105 unsigned long check = pfn + iter;
7106
7107 if (!pfn_valid_within(check))
7108 continue;
7109
7110 page = pfn_to_page(check);
7111
7112 /*
7113 * Hugepages are not in LRU lists, but they're movable.
7114 * We need not scan over tail pages bacause we don't
7115 * handle each tail page individually in migration.
7116 */
7117 if (PageHuge(page)) {
7118 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
7119 continue;
7120 }
7121
7122 /*
7123 * We can't use page_count without pin a page
7124 * because another CPU can free compound page.
7125 * This check already skips compound tails of THP
7126 * because their page->_refcount is zero at all time.
7127 */
7128 if (!page_ref_count(page)) {
7129 if (PageBuddy(page))
7130 iter += (1 << page_order(page)) - 1;
7131 continue;
7132 }
7133
7134 /*
7135 * The HWPoisoned page may be not in buddy system, and
7136 * page_count() is not 0.
7137 */
7138 if (skip_hwpoisoned_pages && PageHWPoison(page))
7139 continue;
7140
7141 if (!PageLRU(page))
7142 found++;
7143 /*
7144 * If there are RECLAIMABLE pages, we need to check
7145 * it. But now, memory offline itself doesn't call
7146 * shrink_node_slabs() and it still to be fixed.
7147 */
7148 /*
7149 * If the page is not RAM, page_count()should be 0.
7150 * we don't need more check. This is an _used_ not-movable page.
7151 *
7152 * The problematic thing here is PG_reserved pages. PG_reserved
7153 * is set to both of a memory hole page and a _used_ kernel
7154 * page at boot.
7155 */
7156 if (found > count)
7157 return true;
7158 }
7159 return false;
7160}
7161
7162bool is_pageblock_removable_nolock(struct page *page)
7163{
7164 struct zone *zone;
7165 unsigned long pfn;
7166
7167 /*
7168 * We have to be careful here because we are iterating over memory
7169 * sections which are not zone aware so we might end up outside of
7170 * the zone but still within the section.
7171 * We have to take care about the node as well. If the node is offline
7172 * its NODE_DATA will be NULL - see page_zone.
7173 */
7174 if (!node_online(page_to_nid(page)))
7175 return false;
7176
7177 zone = page_zone(page);
7178 pfn = page_to_pfn(page);
7179 if (!zone_spans_pfn(zone, pfn))
7180 return false;
7181
7182 return !has_unmovable_pages(zone, page, 0, true);
7183}
7184
7185#if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
7186
7187static unsigned long pfn_max_align_down(unsigned long pfn)
7188{
7189 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
7190 pageblock_nr_pages) - 1);
7191}
7192
7193static unsigned long pfn_max_align_up(unsigned long pfn)
7194{
7195 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
7196 pageblock_nr_pages));
7197}
7198
7199/* [start, end) must belong to a single zone. */
7200static int __alloc_contig_migrate_range(struct compact_control *cc,
7201 unsigned long start, unsigned long end)
7202{
7203 /* This function is based on compact_zone() from compaction.c. */
7204 unsigned long nr_reclaimed;
7205 unsigned long pfn = start;
7206 unsigned int tries = 0;
7207 int ret = 0;
7208
7209 migrate_prep();
7210
7211 while (pfn < end || !list_empty(&cc->migratepages)) {
7212 if (fatal_signal_pending(current)) {
7213 ret = -EINTR;
7214 break;
7215 }
7216
7217 if (list_empty(&cc->migratepages)) {
7218 cc->nr_migratepages = 0;
7219 pfn = isolate_migratepages_range(cc, pfn, end);
7220 if (!pfn) {
7221 ret = -EINTR;
7222 break;
7223 }
7224 tries = 0;
7225 } else if (++tries == 5) {
7226 ret = ret < 0 ? ret : -EBUSY;
7227 break;
7228 }
7229
7230 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
7231 &cc->migratepages);
7232 cc->nr_migratepages -= nr_reclaimed;
7233
7234 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
7235 NULL, 0, cc->mode, MR_CMA);
7236 }
7237 if (ret < 0) {
7238 putback_movable_pages(&cc->migratepages);
7239 return ret;
7240 }
7241 return 0;
7242}
7243
7244/**
7245 * alloc_contig_range() -- tries to allocate given range of pages
7246 * @start: start PFN to allocate
7247 * @end: one-past-the-last PFN to allocate
7248 * @migratetype: migratetype of the underlaying pageblocks (either
7249 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
7250 * in range must have the same migratetype and it must
7251 * be either of the two.
7252 *
7253 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
7254 * aligned, however it's the caller's responsibility to guarantee that
7255 * we are the only thread that changes migrate type of pageblocks the
7256 * pages fall in.
7257 *
7258 * The PFN range must belong to a single zone.
7259 *
7260 * Returns zero on success or negative error code. On success all
7261 * pages which PFN is in [start, end) are allocated for the caller and
7262 * need to be freed with free_contig_range().
7263 */
7264int alloc_contig_range(unsigned long start, unsigned long end,
7265 unsigned migratetype)
7266{
7267 unsigned long outer_start, outer_end;
7268 unsigned int order;
7269 int ret = 0;
7270
7271 struct compact_control cc = {
7272 .nr_migratepages = 0,
7273 .order = -1,
7274 .zone = page_zone(pfn_to_page(start)),
7275 .mode = MIGRATE_SYNC,
7276 .ignore_skip_hint = true,
7277 .gfp_mask = GFP_KERNEL,
7278 };
7279 INIT_LIST_HEAD(&cc.migratepages);
7280
7281 /*
7282 * What we do here is we mark all pageblocks in range as
7283 * MIGRATE_ISOLATE. Because pageblock and max order pages may
7284 * have different sizes, and due to the way page allocator
7285 * work, we align the range to biggest of the two pages so
7286 * that page allocator won't try to merge buddies from
7287 * different pageblocks and change MIGRATE_ISOLATE to some
7288 * other migration type.
7289 *
7290 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
7291 * migrate the pages from an unaligned range (ie. pages that
7292 * we are interested in). This will put all the pages in
7293 * range back to page allocator as MIGRATE_ISOLATE.
7294 *
7295 * When this is done, we take the pages in range from page
7296 * allocator removing them from the buddy system. This way
7297 * page allocator will never consider using them.
7298 *
7299 * This lets us mark the pageblocks back as
7300 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
7301 * aligned range but not in the unaligned, original range are
7302 * put back to page allocator so that buddy can use them.
7303 */
7304
7305 ret = start_isolate_page_range(pfn_max_align_down(start),
7306 pfn_max_align_up(end), migratetype,
7307 false);
7308 if (ret)
7309 return ret;
7310
7311 /*
7312 * In case of -EBUSY, we'd like to know which page causes problem.
7313 * So, just fall through. We will check it in test_pages_isolated().
7314 */
7315 ret = __alloc_contig_migrate_range(&cc, start, end);
7316 if (ret && ret != -EBUSY)
7317 goto done;
7318
7319 /*
7320 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
7321 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
7322 * more, all pages in [start, end) are free in page allocator.
7323 * What we are going to do is to allocate all pages from
7324 * [start, end) (that is remove them from page allocator).
7325 *
7326 * The only problem is that pages at the beginning and at the
7327 * end of interesting range may be not aligned with pages that
7328 * page allocator holds, ie. they can be part of higher order
7329 * pages. Because of this, we reserve the bigger range and
7330 * once this is done free the pages we are not interested in.
7331 *
7332 * We don't have to hold zone->lock here because the pages are
7333 * isolated thus they won't get removed from buddy.
7334 */
7335
7336 lru_add_drain_all();
7337 drain_all_pages(cc.zone);
7338
7339 order = 0;
7340 outer_start = start;
7341 while (!PageBuddy(pfn_to_page(outer_start))) {
7342 if (++order >= MAX_ORDER) {
7343 outer_start = start;
7344 break;
7345 }
7346 outer_start &= ~0UL << order;
7347 }
7348
7349 if (outer_start != start) {
7350 order = page_order(pfn_to_page(outer_start));
7351
7352 /*
7353 * outer_start page could be small order buddy page and
7354 * it doesn't include start page. Adjust outer_start
7355 * in this case to report failed page properly
7356 * on tracepoint in test_pages_isolated()
7357 */
7358 if (outer_start + (1UL << order) <= start)
7359 outer_start = start;
7360 }
7361
7362 /* Make sure the range is really isolated. */
7363 if (test_pages_isolated(outer_start, end, false)) {
7364 pr_info("%s: [%lx, %lx) PFNs busy\n",
7365 __func__, outer_start, end);
7366 ret = -EBUSY;
7367 goto done;
7368 }
7369
7370 /* Grab isolated pages from freelists. */
7371 outer_end = isolate_freepages_range(&cc, outer_start, end);
7372 if (!outer_end) {
7373 ret = -EBUSY;
7374 goto done;
7375 }
7376
7377 /* Free head and tail (if any) */
7378 if (start != outer_start)
7379 free_contig_range(outer_start, start - outer_start);
7380 if (end != outer_end)
7381 free_contig_range(end, outer_end - end);
7382
7383done:
7384 undo_isolate_page_range(pfn_max_align_down(start),
7385 pfn_max_align_up(end), migratetype);
7386 return ret;
7387}
7388
7389void free_contig_range(unsigned long pfn, unsigned nr_pages)
7390{
7391 unsigned int count = 0;
7392
7393 for (; nr_pages--; pfn++) {
7394 struct page *page = pfn_to_page(pfn);
7395
7396 count += page_count(page) != 1;
7397 __free_page(page);
7398 }
7399 WARN(count != 0, "%d pages are still in use!\n", count);
7400}
7401#endif
7402
7403#ifdef CONFIG_MEMORY_HOTPLUG
7404/*
7405 * The zone indicated has a new number of managed_pages; batch sizes and percpu
7406 * page high values need to be recalulated.
7407 */
7408void __meminit zone_pcp_update(struct zone *zone)
7409{
7410 unsigned cpu;
7411 mutex_lock(&pcp_batch_high_lock);
7412 for_each_possible_cpu(cpu)
7413 pageset_set_high_and_batch(zone,
7414 per_cpu_ptr(zone->pageset, cpu));
7415 mutex_unlock(&pcp_batch_high_lock);
7416}
7417#endif
7418
7419void zone_pcp_reset(struct zone *zone)
7420{
7421 unsigned long flags;
7422 int cpu;
7423 struct per_cpu_pageset *pset;
7424
7425 /* avoid races with drain_pages() */
7426 local_irq_save(flags);
7427 if (zone->pageset != &boot_pageset) {
7428 for_each_online_cpu(cpu) {
7429 pset = per_cpu_ptr(zone->pageset, cpu);
7430 drain_zonestat(zone, pset);
7431 }
7432 free_percpu(zone->pageset);
7433 zone->pageset = &boot_pageset;
7434 }
7435 local_irq_restore(flags);
7436}
7437
7438#ifdef CONFIG_MEMORY_HOTREMOVE
7439/*
7440 * All pages in the range must be in a single zone and isolated
7441 * before calling this.
7442 */
7443void
7444__offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
7445{
7446 struct page *page;
7447 struct zone *zone;
7448 unsigned int order, i;
7449 unsigned long pfn;
7450 unsigned long flags;
7451 /* find the first valid pfn */
7452 for (pfn = start_pfn; pfn < end_pfn; pfn++)
7453 if (pfn_valid(pfn))
7454 break;
7455 if (pfn == end_pfn)
7456 return;
7457 zone = page_zone(pfn_to_page(pfn));
7458 spin_lock_irqsave(&zone->lock, flags);
7459 pfn = start_pfn;
7460 while (pfn < end_pfn) {
7461 if (!pfn_valid(pfn)) {
7462 pfn++;
7463 continue;
7464 }
7465 page = pfn_to_page(pfn);
7466 /*
7467 * The HWPoisoned page may be not in buddy system, and
7468 * page_count() is not 0.
7469 */
7470 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
7471 pfn++;
7472 SetPageReserved(page);
7473 continue;
7474 }
7475
7476 BUG_ON(page_count(page));
7477 BUG_ON(!PageBuddy(page));
7478 order = page_order(page);
7479#ifdef CONFIG_DEBUG_VM
7480 pr_info("remove from free list %lx %d %lx\n",
7481 pfn, 1 << order, end_pfn);
7482#endif
7483 list_del(&page->lru);
7484 rmv_page_order(page);
7485 zone->free_area[order].nr_free--;
7486 for (i = 0; i < (1 << order); i++)
7487 SetPageReserved((page+i));
7488 pfn += (1 << order);
7489 }
7490 spin_unlock_irqrestore(&zone->lock, flags);
7491}
7492#endif
7493
7494bool is_free_buddy_page(struct page *page)
7495{
7496 struct zone *zone = page_zone(page);
7497 unsigned long pfn = page_to_pfn(page);
7498 unsigned long flags;
7499 unsigned int order;
7500
7501 spin_lock_irqsave(&zone->lock, flags);
7502 for (order = 0; order < MAX_ORDER; order++) {
7503 struct page *page_head = page - (pfn & ((1 << order) - 1));
7504
7505 if (PageBuddy(page_head) && page_order(page_head) >= order)
7506 break;
7507 }
7508 spin_unlock_irqrestore(&zone->lock, flags);
7509
7510 return order < MAX_ORDER;
7511}