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/highmem.h>
20#include <linux/swap.h>
21#include <linux/interrupt.h>
22#include <linux/pagemap.h>
23#include <linux/jiffies.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/topology.h>
36#include <linux/sysctl.h>
37#include <linux/cpu.h>
38#include <linux/cpuset.h>
39#include <linux/memory_hotplug.h>
40#include <linux/nodemask.h>
41#include <linux/vmalloc.h>
42#include <linux/vmstat.h>
43#include <linux/mempolicy.h>
44#include <linux/memremap.h>
45#include <linux/stop_machine.h>
46#include <linux/sort.h>
47#include <linux/pfn.h>
48#include <linux/backing-dev.h>
49#include <linux/fault-inject.h>
50#include <linux/page-isolation.h>
51#include <linux/page_ext.h>
52#include <linux/debugobjects.h>
53#include <linux/kmemleak.h>
54#include <linux/compaction.h>
55#include <trace/events/kmem.h>
56#include <trace/events/oom.h>
57#include <linux/prefetch.h>
58#include <linux/mm_inline.h>
59#include <linux/migrate.h>
60#include <linux/hugetlb.h>
61#include <linux/sched/rt.h>
62#include <linux/sched/mm.h>
63#include <linux/page_owner.h>
64#include <linux/kthread.h>
65#include <linux/memcontrol.h>
66#include <linux/ftrace.h>
67#include <linux/lockdep.h>
68#include <linux/nmi.h>
69#include <linux/psi.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 */
100struct pcpu_drain {
101 struct zone *zone;
102 struct work_struct work;
103};
104DEFINE_MUTEX(pcpu_drain_mutex);
105DEFINE_PER_CPU(struct pcpu_drain, pcpu_drain);
106
107#ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
108volatile unsigned long latent_entropy __latent_entropy;
109EXPORT_SYMBOL(latent_entropy);
110#endif
111
112/*
113 * Array of node states.
114 */
115nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
116 [N_POSSIBLE] = NODE_MASK_ALL,
117 [N_ONLINE] = { { [0] = 1UL } },
118#ifndef CONFIG_NUMA
119 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
120#ifdef CONFIG_HIGHMEM
121 [N_HIGH_MEMORY] = { { [0] = 1UL } },
122#endif
123 [N_MEMORY] = { { [0] = 1UL } },
124 [N_CPU] = { { [0] = 1UL } },
125#endif /* NUMA */
126};
127EXPORT_SYMBOL(node_states);
128
129atomic_long_t _totalram_pages __read_mostly;
130EXPORT_SYMBOL(_totalram_pages);
131unsigned long totalreserve_pages __read_mostly;
132unsigned long totalcma_pages __read_mostly;
133
134int percpu_pagelist_fraction;
135gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
136
137/*
138 * A cached value of the page's pageblock's migratetype, used when the page is
139 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
140 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
141 * Also the migratetype set in the page does not necessarily match the pcplist
142 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
143 * other index - this ensures that it will be put on the correct CMA freelist.
144 */
145static inline int get_pcppage_migratetype(struct page *page)
146{
147 return page->index;
148}
149
150static inline void set_pcppage_migratetype(struct page *page, int migratetype)
151{
152 page->index = migratetype;
153}
154
155#ifdef CONFIG_PM_SLEEP
156/*
157 * The following functions are used by the suspend/hibernate code to temporarily
158 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
159 * while devices are suspended. To avoid races with the suspend/hibernate code,
160 * they should always be called with system_transition_mutex held
161 * (gfp_allowed_mask also should only be modified with system_transition_mutex
162 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
163 * with that modification).
164 */
165
166static gfp_t saved_gfp_mask;
167
168void pm_restore_gfp_mask(void)
169{
170 WARN_ON(!mutex_is_locked(&system_transition_mutex));
171 if (saved_gfp_mask) {
172 gfp_allowed_mask = saved_gfp_mask;
173 saved_gfp_mask = 0;
174 }
175}
176
177void pm_restrict_gfp_mask(void)
178{
179 WARN_ON(!mutex_is_locked(&system_transition_mutex));
180 WARN_ON(saved_gfp_mask);
181 saved_gfp_mask = gfp_allowed_mask;
182 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
183}
184
185bool pm_suspended_storage(void)
186{
187 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
188 return false;
189 return true;
190}
191#endif /* CONFIG_PM_SLEEP */
192
193#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
194unsigned int pageblock_order __read_mostly;
195#endif
196
197static void __free_pages_ok(struct page *page, unsigned int order);
198
199/*
200 * results with 256, 32 in the lowmem_reserve sysctl:
201 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
202 * 1G machine -> (16M dma, 784M normal, 224M high)
203 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
204 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
205 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
206 *
207 * TBD: should special case ZONE_DMA32 machines here - in those we normally
208 * don't need any ZONE_NORMAL reservation
209 */
210int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
211#ifdef CONFIG_ZONE_DMA
212 [ZONE_DMA] = 256,
213#endif
214#ifdef CONFIG_ZONE_DMA32
215 [ZONE_DMA32] = 256,
216#endif
217 [ZONE_NORMAL] = 32,
218#ifdef CONFIG_HIGHMEM
219 [ZONE_HIGHMEM] = 0,
220#endif
221 [ZONE_MOVABLE] = 0,
222};
223
224EXPORT_SYMBOL(totalram_pages);
225
226static char * const zone_names[MAX_NR_ZONES] = {
227#ifdef CONFIG_ZONE_DMA
228 "DMA",
229#endif
230#ifdef CONFIG_ZONE_DMA32
231 "DMA32",
232#endif
233 "Normal",
234#ifdef CONFIG_HIGHMEM
235 "HighMem",
236#endif
237 "Movable",
238#ifdef CONFIG_ZONE_DEVICE
239 "Device",
240#endif
241};
242
243const char * const migratetype_names[MIGRATE_TYPES] = {
244 "Unmovable",
245 "Movable",
246 "Reclaimable",
247 "HighAtomic",
248#ifdef CONFIG_CMA
249 "CMA",
250#endif
251#ifdef CONFIG_MEMORY_ISOLATION
252 "Isolate",
253#endif
254};
255
256compound_page_dtor * const compound_page_dtors[] = {
257 NULL,
258 free_compound_page,
259#ifdef CONFIG_HUGETLB_PAGE
260 free_huge_page,
261#endif
262#ifdef CONFIG_TRANSPARENT_HUGEPAGE
263 free_transhuge_page,
264#endif
265};
266
267int min_free_kbytes = 1024;
268int user_min_free_kbytes = -1;
269int watermark_boost_factor __read_mostly = 15000;
270int watermark_scale_factor = 10;
271
272static unsigned long nr_kernel_pages __initdata;
273static unsigned long nr_all_pages __initdata;
274static unsigned long dma_reserve __initdata;
275
276#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
277static unsigned long arch_zone_lowest_possible_pfn[MAX_NR_ZONES] __initdata;
278static unsigned long arch_zone_highest_possible_pfn[MAX_NR_ZONES] __initdata;
279static unsigned long required_kernelcore __initdata;
280static unsigned long required_kernelcore_percent __initdata;
281static unsigned long required_movablecore __initdata;
282static unsigned long required_movablecore_percent __initdata;
283static unsigned long zone_movable_pfn[MAX_NUMNODES] __initdata;
284static bool mirrored_kernelcore __meminitdata;
285
286/* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
287int movable_zone;
288EXPORT_SYMBOL(movable_zone);
289#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
290
291#if MAX_NUMNODES > 1
292unsigned int nr_node_ids __read_mostly = MAX_NUMNODES;
293unsigned int nr_online_nodes __read_mostly = 1;
294EXPORT_SYMBOL(nr_node_ids);
295EXPORT_SYMBOL(nr_online_nodes);
296#endif
297
298int page_group_by_mobility_disabled __read_mostly;
299
300#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
301/*
302 * During boot we initialize deferred pages on-demand, as needed, but once
303 * page_alloc_init_late() has finished, the deferred pages are all initialized,
304 * and we can permanently disable that path.
305 */
306static DEFINE_STATIC_KEY_TRUE(deferred_pages);
307
308/*
309 * Calling kasan_free_pages() only after deferred memory initialization
310 * has completed. Poisoning pages during deferred memory init will greatly
311 * lengthen the process and cause problem in large memory systems as the
312 * deferred pages initialization is done with interrupt disabled.
313 *
314 * Assuming that there will be no reference to those newly initialized
315 * pages before they are ever allocated, this should have no effect on
316 * KASAN memory tracking as the poison will be properly inserted at page
317 * allocation time. The only corner case is when pages are allocated by
318 * on-demand allocation and then freed again before the deferred pages
319 * initialization is done, but this is not likely to happen.
320 */
321static inline void kasan_free_nondeferred_pages(struct page *page, int order)
322{
323 if (!static_branch_unlikely(&deferred_pages))
324 kasan_free_pages(page, order);
325}
326
327/* Returns true if the struct page for the pfn is uninitialised */
328static inline bool __meminit early_page_uninitialised(unsigned long pfn)
329{
330 int nid = early_pfn_to_nid(pfn);
331
332 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
333 return true;
334
335 return false;
336}
337
338/*
339 * Returns true when the remaining initialisation should be deferred until
340 * later in the boot cycle when it can be parallelised.
341 */
342static bool __meminit
343defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
344{
345 static unsigned long prev_end_pfn, nr_initialised;
346
347 /*
348 * prev_end_pfn static that contains the end of previous zone
349 * No need to protect because called very early in boot before smp_init.
350 */
351 if (prev_end_pfn != end_pfn) {
352 prev_end_pfn = end_pfn;
353 nr_initialised = 0;
354 }
355
356 /* Always populate low zones for address-constrained allocations */
357 if (end_pfn < pgdat_end_pfn(NODE_DATA(nid)))
358 return false;
359
360 /*
361 * We start only with one section of pages, more pages are added as
362 * needed until the rest of deferred pages are initialized.
363 */
364 nr_initialised++;
365 if ((nr_initialised > PAGES_PER_SECTION) &&
366 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
367 NODE_DATA(nid)->first_deferred_pfn = pfn;
368 return true;
369 }
370 return false;
371}
372#else
373#define kasan_free_nondeferred_pages(p, o) kasan_free_pages(p, o)
374
375static inline bool early_page_uninitialised(unsigned long pfn)
376{
377 return false;
378}
379
380static inline bool defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
381{
382 return false;
383}
384#endif
385
386/* Return a pointer to the bitmap storing bits affecting a block of pages */
387static inline unsigned long *get_pageblock_bitmap(struct page *page,
388 unsigned long pfn)
389{
390#ifdef CONFIG_SPARSEMEM
391 return __pfn_to_section(pfn)->pageblock_flags;
392#else
393 return page_zone(page)->pageblock_flags;
394#endif /* CONFIG_SPARSEMEM */
395}
396
397static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
398{
399#ifdef CONFIG_SPARSEMEM
400 pfn &= (PAGES_PER_SECTION-1);
401 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
402#else
403 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
404 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
405#endif /* CONFIG_SPARSEMEM */
406}
407
408/**
409 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
410 * @page: The page within the block of interest
411 * @pfn: The target page frame number
412 * @end_bitidx: The last bit of interest to retrieve
413 * @mask: mask of bits that the caller is interested in
414 *
415 * Return: pageblock_bits flags
416 */
417static __always_inline unsigned long __get_pfnblock_flags_mask(struct page *page,
418 unsigned long pfn,
419 unsigned long end_bitidx,
420 unsigned long mask)
421{
422 unsigned long *bitmap;
423 unsigned long bitidx, word_bitidx;
424 unsigned long word;
425
426 bitmap = get_pageblock_bitmap(page, pfn);
427 bitidx = pfn_to_bitidx(page, pfn);
428 word_bitidx = bitidx / BITS_PER_LONG;
429 bitidx &= (BITS_PER_LONG-1);
430
431 word = bitmap[word_bitidx];
432 bitidx += end_bitidx;
433 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
434}
435
436unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
437 unsigned long end_bitidx,
438 unsigned long mask)
439{
440 return __get_pfnblock_flags_mask(page, pfn, end_bitidx, mask);
441}
442
443static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
444{
445 return __get_pfnblock_flags_mask(page, pfn, PB_migrate_end, MIGRATETYPE_MASK);
446}
447
448/**
449 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
450 * @page: The page within the block of interest
451 * @flags: The flags to set
452 * @pfn: The target page frame number
453 * @end_bitidx: The last bit of interest
454 * @mask: mask of bits that the caller is interested in
455 */
456void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
457 unsigned long pfn,
458 unsigned long end_bitidx,
459 unsigned long mask)
460{
461 unsigned long *bitmap;
462 unsigned long bitidx, word_bitidx;
463 unsigned long old_word, word;
464
465 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
466 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
467
468 bitmap = get_pageblock_bitmap(page, pfn);
469 bitidx = pfn_to_bitidx(page, pfn);
470 word_bitidx = bitidx / BITS_PER_LONG;
471 bitidx &= (BITS_PER_LONG-1);
472
473 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
474
475 bitidx += end_bitidx;
476 mask <<= (BITS_PER_LONG - bitidx - 1);
477 flags <<= (BITS_PER_LONG - bitidx - 1);
478
479 word = READ_ONCE(bitmap[word_bitidx]);
480 for (;;) {
481 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
482 if (word == old_word)
483 break;
484 word = old_word;
485 }
486}
487
488void set_pageblock_migratetype(struct page *page, int migratetype)
489{
490 if (unlikely(page_group_by_mobility_disabled &&
491 migratetype < MIGRATE_PCPTYPES))
492 migratetype = MIGRATE_UNMOVABLE;
493
494 set_pageblock_flags_group(page, (unsigned long)migratetype,
495 PB_migrate, PB_migrate_end);
496}
497
498#ifdef CONFIG_DEBUG_VM
499static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
500{
501 int ret = 0;
502 unsigned seq;
503 unsigned long pfn = page_to_pfn(page);
504 unsigned long sp, start_pfn;
505
506 do {
507 seq = zone_span_seqbegin(zone);
508 start_pfn = zone->zone_start_pfn;
509 sp = zone->spanned_pages;
510 if (!zone_spans_pfn(zone, pfn))
511 ret = 1;
512 } while (zone_span_seqretry(zone, seq));
513
514 if (ret)
515 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
516 pfn, zone_to_nid(zone), zone->name,
517 start_pfn, start_pfn + sp);
518
519 return ret;
520}
521
522static int page_is_consistent(struct zone *zone, struct page *page)
523{
524 if (!pfn_valid_within(page_to_pfn(page)))
525 return 0;
526 if (zone != page_zone(page))
527 return 0;
528
529 return 1;
530}
531/*
532 * Temporary debugging check for pages not lying within a given zone.
533 */
534static int __maybe_unused bad_range(struct zone *zone, struct page *page)
535{
536 if (page_outside_zone_boundaries(zone, page))
537 return 1;
538 if (!page_is_consistent(zone, page))
539 return 1;
540
541 return 0;
542}
543#else
544static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
545{
546 return 0;
547}
548#endif
549
550static void bad_page(struct page *page, const char *reason,
551 unsigned long bad_flags)
552{
553 static unsigned long resume;
554 static unsigned long nr_shown;
555 static unsigned long nr_unshown;
556
557 /*
558 * Allow a burst of 60 reports, then keep quiet for that minute;
559 * or allow a steady drip of one report per second.
560 */
561 if (nr_shown == 60) {
562 if (time_before(jiffies, resume)) {
563 nr_unshown++;
564 goto out;
565 }
566 if (nr_unshown) {
567 pr_alert(
568 "BUG: Bad page state: %lu messages suppressed\n",
569 nr_unshown);
570 nr_unshown = 0;
571 }
572 nr_shown = 0;
573 }
574 if (nr_shown++ == 0)
575 resume = jiffies + 60 * HZ;
576
577 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
578 current->comm, page_to_pfn(page));
579 __dump_page(page, reason);
580 bad_flags &= page->flags;
581 if (bad_flags)
582 pr_alert("bad because of flags: %#lx(%pGp)\n",
583 bad_flags, &bad_flags);
584 dump_page_owner(page);
585
586 print_modules();
587 dump_stack();
588out:
589 /* Leave bad fields for debug, except PageBuddy could make trouble */
590 page_mapcount_reset(page); /* remove PageBuddy */
591 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
592}
593
594/*
595 * Higher-order pages are called "compound pages". They are structured thusly:
596 *
597 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
598 *
599 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
600 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
601 *
602 * The first tail page's ->compound_dtor holds the offset in array of compound
603 * page destructors. See compound_page_dtors.
604 *
605 * The first tail page's ->compound_order holds the order of allocation.
606 * This usage means that zero-order pages may not be compound.
607 */
608
609void free_compound_page(struct page *page)
610{
611 __free_pages_ok(page, compound_order(page));
612}
613
614void prep_compound_page(struct page *page, unsigned int order)
615{
616 int i;
617 int nr_pages = 1 << order;
618
619 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
620 set_compound_order(page, order);
621 __SetPageHead(page);
622 for (i = 1; i < nr_pages; i++) {
623 struct page *p = page + i;
624 set_page_count(p, 0);
625 p->mapping = TAIL_MAPPING;
626 set_compound_head(p, page);
627 }
628 atomic_set(compound_mapcount_ptr(page), -1);
629}
630
631#ifdef CONFIG_DEBUG_PAGEALLOC
632unsigned int _debug_guardpage_minorder;
633bool _debug_pagealloc_enabled __read_mostly
634 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
635EXPORT_SYMBOL(_debug_pagealloc_enabled);
636bool _debug_guardpage_enabled __read_mostly;
637
638static int __init early_debug_pagealloc(char *buf)
639{
640 if (!buf)
641 return -EINVAL;
642 return kstrtobool(buf, &_debug_pagealloc_enabled);
643}
644early_param("debug_pagealloc", early_debug_pagealloc);
645
646static bool need_debug_guardpage(void)
647{
648 /* If we don't use debug_pagealloc, we don't need guard page */
649 if (!debug_pagealloc_enabled())
650 return false;
651
652 if (!debug_guardpage_minorder())
653 return false;
654
655 return true;
656}
657
658static void init_debug_guardpage(void)
659{
660 if (!debug_pagealloc_enabled())
661 return;
662
663 if (!debug_guardpage_minorder())
664 return;
665
666 _debug_guardpage_enabled = true;
667}
668
669struct page_ext_operations debug_guardpage_ops = {
670 .need = need_debug_guardpage,
671 .init = init_debug_guardpage,
672};
673
674static int __init debug_guardpage_minorder_setup(char *buf)
675{
676 unsigned long res;
677
678 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
679 pr_err("Bad debug_guardpage_minorder value\n");
680 return 0;
681 }
682 _debug_guardpage_minorder = res;
683 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
684 return 0;
685}
686early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
687
688static inline bool set_page_guard(struct zone *zone, struct page *page,
689 unsigned int order, int migratetype)
690{
691 struct page_ext *page_ext;
692
693 if (!debug_guardpage_enabled())
694 return false;
695
696 if (order >= debug_guardpage_minorder())
697 return false;
698
699 page_ext = lookup_page_ext(page);
700 if (unlikely(!page_ext))
701 return false;
702
703 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
704
705 INIT_LIST_HEAD(&page->lru);
706 set_page_private(page, order);
707 /* Guard pages are not available for any usage */
708 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
709
710 return true;
711}
712
713static inline void clear_page_guard(struct zone *zone, struct page *page,
714 unsigned int order, int migratetype)
715{
716 struct page_ext *page_ext;
717
718 if (!debug_guardpage_enabled())
719 return;
720
721 page_ext = lookup_page_ext(page);
722 if (unlikely(!page_ext))
723 return;
724
725 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
726
727 set_page_private(page, 0);
728 if (!is_migrate_isolate(migratetype))
729 __mod_zone_freepage_state(zone, (1 << order), migratetype);
730}
731#else
732struct page_ext_operations debug_guardpage_ops;
733static inline bool set_page_guard(struct zone *zone, struct page *page,
734 unsigned int order, int migratetype) { return false; }
735static inline void clear_page_guard(struct zone *zone, struct page *page,
736 unsigned int order, int migratetype) {}
737#endif
738
739static inline void set_page_order(struct page *page, unsigned int order)
740{
741 set_page_private(page, order);
742 __SetPageBuddy(page);
743}
744
745static inline void rmv_page_order(struct page *page)
746{
747 __ClearPageBuddy(page);
748 set_page_private(page, 0);
749}
750
751/*
752 * This function checks whether a page is free && is the buddy
753 * we can coalesce a page and its buddy if
754 * (a) the buddy is not in a hole (check before calling!) &&
755 * (b) the buddy is in the buddy system &&
756 * (c) a page and its buddy have the same order &&
757 * (d) a page and its buddy are in the same zone.
758 *
759 * For recording whether a page is in the buddy system, we set PageBuddy.
760 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
761 *
762 * For recording page's order, we use page_private(page).
763 */
764static inline int page_is_buddy(struct page *page, struct page *buddy,
765 unsigned int order)
766{
767 if (page_is_guard(buddy) && page_order(buddy) == order) {
768 if (page_zone_id(page) != page_zone_id(buddy))
769 return 0;
770
771 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
772
773 return 1;
774 }
775
776 if (PageBuddy(buddy) && page_order(buddy) == order) {
777 /*
778 * zone check is done late to avoid uselessly
779 * calculating zone/node ids for pages that could
780 * never merge.
781 */
782 if (page_zone_id(page) != page_zone_id(buddy))
783 return 0;
784
785 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
786
787 return 1;
788 }
789 return 0;
790}
791
792#ifdef CONFIG_COMPACTION
793static inline struct capture_control *task_capc(struct zone *zone)
794{
795 struct capture_control *capc = current->capture_control;
796
797 return capc &&
798 !(current->flags & PF_KTHREAD) &&
799 !capc->page &&
800 capc->cc->zone == zone &&
801 capc->cc->direct_compaction ? capc : NULL;
802}
803
804static inline bool
805compaction_capture(struct capture_control *capc, struct page *page,
806 int order, int migratetype)
807{
808 if (!capc || order != capc->cc->order)
809 return false;
810
811 /* Do not accidentally pollute CMA or isolated regions*/
812 if (is_migrate_cma(migratetype) ||
813 is_migrate_isolate(migratetype))
814 return false;
815
816 /*
817 * Do not let lower order allocations polluate a movable pageblock.
818 * This might let an unmovable request use a reclaimable pageblock
819 * and vice-versa but no more than normal fallback logic which can
820 * have trouble finding a high-order free page.
821 */
822 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
823 return false;
824
825 capc->page = page;
826 return true;
827}
828
829#else
830static inline struct capture_control *task_capc(struct zone *zone)
831{
832 return NULL;
833}
834
835static inline bool
836compaction_capture(struct capture_control *capc, struct page *page,
837 int order, int migratetype)
838{
839 return false;
840}
841#endif /* CONFIG_COMPACTION */
842
843/*
844 * Freeing function for a buddy system allocator.
845 *
846 * The concept of a buddy system is to maintain direct-mapped table
847 * (containing bit values) for memory blocks of various "orders".
848 * The bottom level table contains the map for the smallest allocatable
849 * units of memory (here, pages), and each level above it describes
850 * pairs of units from the levels below, hence, "buddies".
851 * At a high level, all that happens here is marking the table entry
852 * at the bottom level available, and propagating the changes upward
853 * as necessary, plus some accounting needed to play nicely with other
854 * parts of the VM system.
855 * At each level, we keep a list of pages, which are heads of continuous
856 * free pages of length of (1 << order) and marked with PageBuddy.
857 * Page's order is recorded in page_private(page) field.
858 * So when we are allocating or freeing one, we can derive the state of the
859 * other. That is, if we allocate a small block, and both were
860 * free, the remainder of the region must be split into blocks.
861 * If a block is freed, and its buddy is also free, then this
862 * triggers coalescing into a block of larger size.
863 *
864 * -- nyc
865 */
866
867static inline void __free_one_page(struct page *page,
868 unsigned long pfn,
869 struct zone *zone, unsigned int order,
870 int migratetype)
871{
872 unsigned long combined_pfn;
873 unsigned long uninitialized_var(buddy_pfn);
874 struct page *buddy;
875 unsigned int max_order;
876 struct capture_control *capc = task_capc(zone);
877
878 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
879
880 VM_BUG_ON(!zone_is_initialized(zone));
881 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
882
883 VM_BUG_ON(migratetype == -1);
884 if (likely(!is_migrate_isolate(migratetype)))
885 __mod_zone_freepage_state(zone, 1 << order, migratetype);
886
887 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
888 VM_BUG_ON_PAGE(bad_range(zone, page), page);
889
890continue_merging:
891 while (order < max_order - 1) {
892 if (compaction_capture(capc, page, order, migratetype)) {
893 __mod_zone_freepage_state(zone, -(1 << order),
894 migratetype);
895 return;
896 }
897 buddy_pfn = __find_buddy_pfn(pfn, order);
898 buddy = page + (buddy_pfn - pfn);
899
900 if (!pfn_valid_within(buddy_pfn))
901 goto done_merging;
902 if (!page_is_buddy(page, buddy, order))
903 goto done_merging;
904 /*
905 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
906 * merge with it and move up one order.
907 */
908 if (page_is_guard(buddy)) {
909 clear_page_guard(zone, buddy, order, migratetype);
910 } else {
911 list_del(&buddy->lru);
912 zone->free_area[order].nr_free--;
913 rmv_page_order(buddy);
914 }
915 combined_pfn = buddy_pfn & pfn;
916 page = page + (combined_pfn - pfn);
917 pfn = combined_pfn;
918 order++;
919 }
920 if (max_order < MAX_ORDER) {
921 /* If we are here, it means order is >= pageblock_order.
922 * We want to prevent merge between freepages on isolate
923 * pageblock and normal pageblock. Without this, pageblock
924 * isolation could cause incorrect freepage or CMA accounting.
925 *
926 * We don't want to hit this code for the more frequent
927 * low-order merging.
928 */
929 if (unlikely(has_isolate_pageblock(zone))) {
930 int buddy_mt;
931
932 buddy_pfn = __find_buddy_pfn(pfn, order);
933 buddy = page + (buddy_pfn - pfn);
934 buddy_mt = get_pageblock_migratetype(buddy);
935
936 if (migratetype != buddy_mt
937 && (is_migrate_isolate(migratetype) ||
938 is_migrate_isolate(buddy_mt)))
939 goto done_merging;
940 }
941 max_order++;
942 goto continue_merging;
943 }
944
945done_merging:
946 set_page_order(page, order);
947
948 /*
949 * If this is not the largest possible page, check if the buddy
950 * of the next-highest order is free. If it is, it's possible
951 * that pages are being freed that will coalesce soon. In case,
952 * that is happening, add the free page to the tail of the list
953 * so it's less likely to be used soon and more likely to be merged
954 * as a higher order page
955 */
956 if ((order < MAX_ORDER-2) && pfn_valid_within(buddy_pfn)) {
957 struct page *higher_page, *higher_buddy;
958 combined_pfn = buddy_pfn & pfn;
959 higher_page = page + (combined_pfn - pfn);
960 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
961 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
962 if (pfn_valid_within(buddy_pfn) &&
963 page_is_buddy(higher_page, higher_buddy, order + 1)) {
964 list_add_tail(&page->lru,
965 &zone->free_area[order].free_list[migratetype]);
966 goto out;
967 }
968 }
969
970 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
971out:
972 zone->free_area[order].nr_free++;
973}
974
975/*
976 * A bad page could be due to a number of fields. Instead of multiple branches,
977 * try and check multiple fields with one check. The caller must do a detailed
978 * check if necessary.
979 */
980static inline bool page_expected_state(struct page *page,
981 unsigned long check_flags)
982{
983 if (unlikely(atomic_read(&page->_mapcount) != -1))
984 return false;
985
986 if (unlikely((unsigned long)page->mapping |
987 page_ref_count(page) |
988#ifdef CONFIG_MEMCG
989 (unsigned long)page->mem_cgroup |
990#endif
991 (page->flags & check_flags)))
992 return false;
993
994 return true;
995}
996
997static void free_pages_check_bad(struct page *page)
998{
999 const char *bad_reason;
1000 unsigned long bad_flags;
1001
1002 bad_reason = NULL;
1003 bad_flags = 0;
1004
1005 if (unlikely(atomic_read(&page->_mapcount) != -1))
1006 bad_reason = "nonzero mapcount";
1007 if (unlikely(page->mapping != NULL))
1008 bad_reason = "non-NULL mapping";
1009 if (unlikely(page_ref_count(page) != 0))
1010 bad_reason = "nonzero _refcount";
1011 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
1012 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1013 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
1014 }
1015#ifdef CONFIG_MEMCG
1016 if (unlikely(page->mem_cgroup))
1017 bad_reason = "page still charged to cgroup";
1018#endif
1019 bad_page(page, bad_reason, bad_flags);
1020}
1021
1022static inline int free_pages_check(struct page *page)
1023{
1024 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1025 return 0;
1026
1027 /* Something has gone sideways, find it */
1028 free_pages_check_bad(page);
1029 return 1;
1030}
1031
1032static int free_tail_pages_check(struct page *head_page, struct page *page)
1033{
1034 int ret = 1;
1035
1036 /*
1037 * We rely page->lru.next never has bit 0 set, unless the page
1038 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1039 */
1040 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1041
1042 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
1043 ret = 0;
1044 goto out;
1045 }
1046 switch (page - head_page) {
1047 case 1:
1048 /* the first tail page: ->mapping may be compound_mapcount() */
1049 if (unlikely(compound_mapcount(page))) {
1050 bad_page(page, "nonzero compound_mapcount", 0);
1051 goto out;
1052 }
1053 break;
1054 case 2:
1055 /*
1056 * the second tail page: ->mapping is
1057 * deferred_list.next -- ignore value.
1058 */
1059 break;
1060 default:
1061 if (page->mapping != TAIL_MAPPING) {
1062 bad_page(page, "corrupted mapping in tail page", 0);
1063 goto out;
1064 }
1065 break;
1066 }
1067 if (unlikely(!PageTail(page))) {
1068 bad_page(page, "PageTail not set", 0);
1069 goto out;
1070 }
1071 if (unlikely(compound_head(page) != head_page)) {
1072 bad_page(page, "compound_head not consistent", 0);
1073 goto out;
1074 }
1075 ret = 0;
1076out:
1077 page->mapping = NULL;
1078 clear_compound_head(page);
1079 return ret;
1080}
1081
1082static __always_inline bool free_pages_prepare(struct page *page,
1083 unsigned int order, bool check_free)
1084{
1085 int bad = 0;
1086
1087 VM_BUG_ON_PAGE(PageTail(page), page);
1088
1089 trace_mm_page_free(page, order);
1090
1091 /*
1092 * Check tail pages before head page information is cleared to
1093 * avoid checking PageCompound for order-0 pages.
1094 */
1095 if (unlikely(order)) {
1096 bool compound = PageCompound(page);
1097 int i;
1098
1099 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1100
1101 if (compound)
1102 ClearPageDoubleMap(page);
1103 for (i = 1; i < (1 << order); i++) {
1104 if (compound)
1105 bad += free_tail_pages_check(page, page + i);
1106 if (unlikely(free_pages_check(page + i))) {
1107 bad++;
1108 continue;
1109 }
1110 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1111 }
1112 }
1113 if (PageMappingFlags(page))
1114 page->mapping = NULL;
1115 if (memcg_kmem_enabled() && PageKmemcg(page))
1116 __memcg_kmem_uncharge(page, order);
1117 if (check_free)
1118 bad += free_pages_check(page);
1119 if (bad)
1120 return false;
1121
1122 page_cpupid_reset_last(page);
1123 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1124 reset_page_owner(page, order);
1125
1126 if (!PageHighMem(page)) {
1127 debug_check_no_locks_freed(page_address(page),
1128 PAGE_SIZE << order);
1129 debug_check_no_obj_freed(page_address(page),
1130 PAGE_SIZE << order);
1131 }
1132 arch_free_page(page, order);
1133 kernel_poison_pages(page, 1 << order, 0);
1134 kernel_map_pages(page, 1 << order, 0);
1135 kasan_free_nondeferred_pages(page, order);
1136
1137 return true;
1138}
1139
1140#ifdef CONFIG_DEBUG_VM
1141static inline bool free_pcp_prepare(struct page *page)
1142{
1143 return free_pages_prepare(page, 0, true);
1144}
1145
1146static inline bool bulkfree_pcp_prepare(struct page *page)
1147{
1148 return false;
1149}
1150#else
1151static bool free_pcp_prepare(struct page *page)
1152{
1153 return free_pages_prepare(page, 0, false);
1154}
1155
1156static bool bulkfree_pcp_prepare(struct page *page)
1157{
1158 return free_pages_check(page);
1159}
1160#endif /* CONFIG_DEBUG_VM */
1161
1162static inline void prefetch_buddy(struct page *page)
1163{
1164 unsigned long pfn = page_to_pfn(page);
1165 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1166 struct page *buddy = page + (buddy_pfn - pfn);
1167
1168 prefetch(buddy);
1169}
1170
1171/*
1172 * Frees a number of pages from the PCP lists
1173 * Assumes all pages on list are in same zone, and of same order.
1174 * count is the number of pages to free.
1175 *
1176 * If the zone was previously in an "all pages pinned" state then look to
1177 * see if this freeing clears that state.
1178 *
1179 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1180 * pinned" detection logic.
1181 */
1182static void free_pcppages_bulk(struct zone *zone, int count,
1183 struct per_cpu_pages *pcp)
1184{
1185 int migratetype = 0;
1186 int batch_free = 0;
1187 int prefetch_nr = 0;
1188 bool isolated_pageblocks;
1189 struct page *page, *tmp;
1190 LIST_HEAD(head);
1191
1192 while (count) {
1193 struct list_head *list;
1194
1195 /*
1196 * Remove pages from lists in a round-robin fashion. A
1197 * batch_free count is maintained that is incremented when an
1198 * empty list is encountered. This is so more pages are freed
1199 * off fuller lists instead of spinning excessively around empty
1200 * lists
1201 */
1202 do {
1203 batch_free++;
1204 if (++migratetype == MIGRATE_PCPTYPES)
1205 migratetype = 0;
1206 list = &pcp->lists[migratetype];
1207 } while (list_empty(list));
1208
1209 /* This is the only non-empty list. Free them all. */
1210 if (batch_free == MIGRATE_PCPTYPES)
1211 batch_free = count;
1212
1213 do {
1214 page = list_last_entry(list, struct page, lru);
1215 /* must delete to avoid corrupting pcp list */
1216 list_del(&page->lru);
1217 pcp->count--;
1218
1219 if (bulkfree_pcp_prepare(page))
1220 continue;
1221
1222 list_add_tail(&page->lru, &head);
1223
1224 /*
1225 * We are going to put the page back to the global
1226 * pool, prefetch its buddy to speed up later access
1227 * under zone->lock. It is believed the overhead of
1228 * an additional test and calculating buddy_pfn here
1229 * can be offset by reduced memory latency later. To
1230 * avoid excessive prefetching due to large count, only
1231 * prefetch buddy for the first pcp->batch nr of pages.
1232 */
1233 if (prefetch_nr++ < pcp->batch)
1234 prefetch_buddy(page);
1235 } while (--count && --batch_free && !list_empty(list));
1236 }
1237
1238 spin_lock(&zone->lock);
1239 isolated_pageblocks = has_isolate_pageblock(zone);
1240
1241 /*
1242 * Use safe version since after __free_one_page(),
1243 * page->lru.next will not point to original list.
1244 */
1245 list_for_each_entry_safe(page, tmp, &head, lru) {
1246 int mt = get_pcppage_migratetype(page);
1247 /* MIGRATE_ISOLATE page should not go to pcplists */
1248 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1249 /* Pageblock could have been isolated meanwhile */
1250 if (unlikely(isolated_pageblocks))
1251 mt = get_pageblock_migratetype(page);
1252
1253 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
1254 trace_mm_page_pcpu_drain(page, 0, mt);
1255 }
1256 spin_unlock(&zone->lock);
1257}
1258
1259static void free_one_page(struct zone *zone,
1260 struct page *page, unsigned long pfn,
1261 unsigned int order,
1262 int migratetype)
1263{
1264 spin_lock(&zone->lock);
1265 if (unlikely(has_isolate_pageblock(zone) ||
1266 is_migrate_isolate(migratetype))) {
1267 migratetype = get_pfnblock_migratetype(page, pfn);
1268 }
1269 __free_one_page(page, pfn, zone, order, migratetype);
1270 spin_unlock(&zone->lock);
1271}
1272
1273static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1274 unsigned long zone, int nid)
1275{
1276 mm_zero_struct_page(page);
1277 set_page_links(page, zone, nid, pfn);
1278 init_page_count(page);
1279 page_mapcount_reset(page);
1280 page_cpupid_reset_last(page);
1281 page_kasan_tag_reset(page);
1282
1283 INIT_LIST_HEAD(&page->lru);
1284#ifdef WANT_PAGE_VIRTUAL
1285 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1286 if (!is_highmem_idx(zone))
1287 set_page_address(page, __va(pfn << PAGE_SHIFT));
1288#endif
1289}
1290
1291#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1292static void __meminit init_reserved_page(unsigned long pfn)
1293{
1294 pg_data_t *pgdat;
1295 int nid, zid;
1296
1297 if (!early_page_uninitialised(pfn))
1298 return;
1299
1300 nid = early_pfn_to_nid(pfn);
1301 pgdat = NODE_DATA(nid);
1302
1303 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1304 struct zone *zone = &pgdat->node_zones[zid];
1305
1306 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1307 break;
1308 }
1309 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1310}
1311#else
1312static inline void init_reserved_page(unsigned long pfn)
1313{
1314}
1315#endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1316
1317/*
1318 * Initialised pages do not have PageReserved set. This function is
1319 * called for each range allocated by the bootmem allocator and
1320 * marks the pages PageReserved. The remaining valid pages are later
1321 * sent to the buddy page allocator.
1322 */
1323void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1324{
1325 unsigned long start_pfn = PFN_DOWN(start);
1326 unsigned long end_pfn = PFN_UP(end);
1327
1328 for (; start_pfn < end_pfn; start_pfn++) {
1329 if (pfn_valid(start_pfn)) {
1330 struct page *page = pfn_to_page(start_pfn);
1331
1332 init_reserved_page(start_pfn);
1333
1334 /* Avoid false-positive PageTail() */
1335 INIT_LIST_HEAD(&page->lru);
1336
1337 /*
1338 * no need for atomic set_bit because the struct
1339 * page is not visible yet so nobody should
1340 * access it yet.
1341 */
1342 __SetPageReserved(page);
1343 }
1344 }
1345}
1346
1347static void __free_pages_ok(struct page *page, unsigned int order)
1348{
1349 unsigned long flags;
1350 int migratetype;
1351 unsigned long pfn = page_to_pfn(page);
1352
1353 if (!free_pages_prepare(page, order, true))
1354 return;
1355
1356 migratetype = get_pfnblock_migratetype(page, pfn);
1357 local_irq_save(flags);
1358 __count_vm_events(PGFREE, 1 << order);
1359 free_one_page(page_zone(page), page, pfn, order, migratetype);
1360 local_irq_restore(flags);
1361}
1362
1363void __free_pages_core(struct page *page, unsigned int order)
1364{
1365 unsigned int nr_pages = 1 << order;
1366 struct page *p = page;
1367 unsigned int loop;
1368
1369 prefetchw(p);
1370 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1371 prefetchw(p + 1);
1372 __ClearPageReserved(p);
1373 set_page_count(p, 0);
1374 }
1375 __ClearPageReserved(p);
1376 set_page_count(p, 0);
1377
1378 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1379 set_page_refcounted(page);
1380 __free_pages(page, order);
1381}
1382
1383#if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1384 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1385
1386static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1387
1388int __meminit early_pfn_to_nid(unsigned long pfn)
1389{
1390 static DEFINE_SPINLOCK(early_pfn_lock);
1391 int nid;
1392
1393 spin_lock(&early_pfn_lock);
1394 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1395 if (nid < 0)
1396 nid = first_online_node;
1397 spin_unlock(&early_pfn_lock);
1398
1399 return nid;
1400}
1401#endif
1402
1403#ifdef CONFIG_NODES_SPAN_OTHER_NODES
1404static inline bool __meminit __maybe_unused
1405meminit_pfn_in_nid(unsigned long pfn, int node,
1406 struct mminit_pfnnid_cache *state)
1407{
1408 int nid;
1409
1410 nid = __early_pfn_to_nid(pfn, state);
1411 if (nid >= 0 && nid != node)
1412 return false;
1413 return true;
1414}
1415
1416/* Only safe to use early in boot when initialisation is single-threaded */
1417static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1418{
1419 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1420}
1421
1422#else
1423
1424static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1425{
1426 return true;
1427}
1428static inline bool __meminit __maybe_unused
1429meminit_pfn_in_nid(unsigned long pfn, int node,
1430 struct mminit_pfnnid_cache *state)
1431{
1432 return true;
1433}
1434#endif
1435
1436
1437void __init memblock_free_pages(struct page *page, unsigned long pfn,
1438 unsigned int order)
1439{
1440 if (early_page_uninitialised(pfn))
1441 return;
1442 __free_pages_core(page, order);
1443}
1444
1445/*
1446 * Check that the whole (or subset of) a pageblock given by the interval of
1447 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1448 * with the migration of free compaction scanner. The scanners then need to
1449 * use only pfn_valid_within() check for arches that allow holes within
1450 * pageblocks.
1451 *
1452 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1453 *
1454 * It's possible on some configurations to have a setup like node0 node1 node0
1455 * i.e. it's possible that all pages within a zones range of pages do not
1456 * belong to a single zone. We assume that a border between node0 and node1
1457 * can occur within a single pageblock, but not a node0 node1 node0
1458 * interleaving within a single pageblock. It is therefore sufficient to check
1459 * the first and last page of a pageblock and avoid checking each individual
1460 * page in a pageblock.
1461 */
1462struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1463 unsigned long end_pfn, struct zone *zone)
1464{
1465 struct page *start_page;
1466 struct page *end_page;
1467
1468 /* end_pfn is one past the range we are checking */
1469 end_pfn--;
1470
1471 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1472 return NULL;
1473
1474 start_page = pfn_to_online_page(start_pfn);
1475 if (!start_page)
1476 return NULL;
1477
1478 if (page_zone(start_page) != zone)
1479 return NULL;
1480
1481 end_page = pfn_to_page(end_pfn);
1482
1483 /* This gives a shorter code than deriving page_zone(end_page) */
1484 if (page_zone_id(start_page) != page_zone_id(end_page))
1485 return NULL;
1486
1487 return start_page;
1488}
1489
1490void set_zone_contiguous(struct zone *zone)
1491{
1492 unsigned long block_start_pfn = zone->zone_start_pfn;
1493 unsigned long block_end_pfn;
1494
1495 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1496 for (; block_start_pfn < zone_end_pfn(zone);
1497 block_start_pfn = block_end_pfn,
1498 block_end_pfn += pageblock_nr_pages) {
1499
1500 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1501
1502 if (!__pageblock_pfn_to_page(block_start_pfn,
1503 block_end_pfn, zone))
1504 return;
1505 }
1506
1507 /* We confirm that there is no hole */
1508 zone->contiguous = true;
1509}
1510
1511void clear_zone_contiguous(struct zone *zone)
1512{
1513 zone->contiguous = false;
1514}
1515
1516#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1517static void __init deferred_free_range(unsigned long pfn,
1518 unsigned long nr_pages)
1519{
1520 struct page *page;
1521 unsigned long i;
1522
1523 if (!nr_pages)
1524 return;
1525
1526 page = pfn_to_page(pfn);
1527
1528 /* Free a large naturally-aligned chunk if possible */
1529 if (nr_pages == pageblock_nr_pages &&
1530 (pfn & (pageblock_nr_pages - 1)) == 0) {
1531 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1532 __free_pages_core(page, pageblock_order);
1533 return;
1534 }
1535
1536 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1537 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1538 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1539 __free_pages_core(page, 0);
1540 }
1541}
1542
1543/* Completion tracking for deferred_init_memmap() threads */
1544static atomic_t pgdat_init_n_undone __initdata;
1545static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1546
1547static inline void __init pgdat_init_report_one_done(void)
1548{
1549 if (atomic_dec_and_test(&pgdat_init_n_undone))
1550 complete(&pgdat_init_all_done_comp);
1551}
1552
1553/*
1554 * Returns true if page needs to be initialized or freed to buddy allocator.
1555 *
1556 * First we check if pfn is valid on architectures where it is possible to have
1557 * holes within pageblock_nr_pages. On systems where it is not possible, this
1558 * function is optimized out.
1559 *
1560 * Then, we check if a current large page is valid by only checking the validity
1561 * of the head pfn.
1562 *
1563 * Finally, meminit_pfn_in_nid is checked on systems where pfns can interleave
1564 * within a node: a pfn is between start and end of a node, but does not belong
1565 * to this memory node.
1566 */
1567static inline bool __init
1568deferred_pfn_valid(int nid, unsigned long pfn,
1569 struct mminit_pfnnid_cache *nid_init_state)
1570{
1571 if (!pfn_valid_within(pfn))
1572 return false;
1573 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1574 return false;
1575 if (!meminit_pfn_in_nid(pfn, nid, nid_init_state))
1576 return false;
1577 return true;
1578}
1579
1580/*
1581 * Free pages to buddy allocator. Try to free aligned pages in
1582 * pageblock_nr_pages sizes.
1583 */
1584static void __init deferred_free_pages(int nid, int zid, unsigned long pfn,
1585 unsigned long end_pfn)
1586{
1587 struct mminit_pfnnid_cache nid_init_state = { };
1588 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1589 unsigned long nr_free = 0;
1590
1591 for (; pfn < end_pfn; pfn++) {
1592 if (!deferred_pfn_valid(nid, pfn, &nid_init_state)) {
1593 deferred_free_range(pfn - nr_free, nr_free);
1594 nr_free = 0;
1595 } else if (!(pfn & nr_pgmask)) {
1596 deferred_free_range(pfn - nr_free, nr_free);
1597 nr_free = 1;
1598 touch_nmi_watchdog();
1599 } else {
1600 nr_free++;
1601 }
1602 }
1603 /* Free the last block of pages to allocator */
1604 deferred_free_range(pfn - nr_free, nr_free);
1605}
1606
1607/*
1608 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1609 * by performing it only once every pageblock_nr_pages.
1610 * Return number of pages initialized.
1611 */
1612static unsigned long __init deferred_init_pages(int nid, int zid,
1613 unsigned long pfn,
1614 unsigned long end_pfn)
1615{
1616 struct mminit_pfnnid_cache nid_init_state = { };
1617 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1618 unsigned long nr_pages = 0;
1619 struct page *page = NULL;
1620
1621 for (; pfn < end_pfn; pfn++) {
1622 if (!deferred_pfn_valid(nid, pfn, &nid_init_state)) {
1623 page = NULL;
1624 continue;
1625 } else if (!page || !(pfn & nr_pgmask)) {
1626 page = pfn_to_page(pfn);
1627 touch_nmi_watchdog();
1628 } else {
1629 page++;
1630 }
1631 __init_single_page(page, pfn, zid, nid);
1632 nr_pages++;
1633 }
1634 return (nr_pages);
1635}
1636
1637/* Initialise remaining memory on a node */
1638static int __init deferred_init_memmap(void *data)
1639{
1640 pg_data_t *pgdat = data;
1641 int nid = pgdat->node_id;
1642 unsigned long start = jiffies;
1643 unsigned long nr_pages = 0;
1644 unsigned long spfn, epfn, first_init_pfn, flags;
1645 phys_addr_t spa, epa;
1646 int zid;
1647 struct zone *zone;
1648 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1649 u64 i;
1650
1651 /* Bind memory initialisation thread to a local node if possible */
1652 if (!cpumask_empty(cpumask))
1653 set_cpus_allowed_ptr(current, cpumask);
1654
1655 pgdat_resize_lock(pgdat, &flags);
1656 first_init_pfn = pgdat->first_deferred_pfn;
1657 if (first_init_pfn == ULONG_MAX) {
1658 pgdat_resize_unlock(pgdat, &flags);
1659 pgdat_init_report_one_done();
1660 return 0;
1661 }
1662
1663 /* Sanity check boundaries */
1664 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1665 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1666 pgdat->first_deferred_pfn = ULONG_MAX;
1667
1668 /* Only the highest zone is deferred so find it */
1669 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1670 zone = pgdat->node_zones + zid;
1671 if (first_init_pfn < zone_end_pfn(zone))
1672 break;
1673 }
1674 first_init_pfn = max(zone->zone_start_pfn, first_init_pfn);
1675
1676 /*
1677 * Initialize and free pages. We do it in two loops: first we initialize
1678 * struct page, than free to buddy allocator, because while we are
1679 * freeing pages we can access pages that are ahead (computing buddy
1680 * page in __free_one_page()).
1681 */
1682 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1683 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1684 epfn = min_t(unsigned long, zone_end_pfn(zone), PFN_DOWN(epa));
1685 nr_pages += deferred_init_pages(nid, zid, spfn, epfn);
1686 }
1687 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1688 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1689 epfn = min_t(unsigned long, zone_end_pfn(zone), PFN_DOWN(epa));
1690 deferred_free_pages(nid, zid, spfn, epfn);
1691 }
1692 pgdat_resize_unlock(pgdat, &flags);
1693
1694 /* Sanity check that the next zone really is unpopulated */
1695 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1696
1697 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1698 jiffies_to_msecs(jiffies - start));
1699
1700 pgdat_init_report_one_done();
1701 return 0;
1702}
1703
1704/*
1705 * If this zone has deferred pages, try to grow it by initializing enough
1706 * deferred pages to satisfy the allocation specified by order, rounded up to
1707 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
1708 * of SECTION_SIZE bytes by initializing struct pages in increments of
1709 * PAGES_PER_SECTION * sizeof(struct page) bytes.
1710 *
1711 * Return true when zone was grown, otherwise return false. We return true even
1712 * when we grow less than requested, to let the caller decide if there are
1713 * enough pages to satisfy the allocation.
1714 *
1715 * Note: We use noinline because this function is needed only during boot, and
1716 * it is called from a __ref function _deferred_grow_zone. This way we are
1717 * making sure that it is not inlined into permanent text section.
1718 */
1719static noinline bool __init
1720deferred_grow_zone(struct zone *zone, unsigned int order)
1721{
1722 int zid = zone_idx(zone);
1723 int nid = zone_to_nid(zone);
1724 pg_data_t *pgdat = NODE_DATA(nid);
1725 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
1726 unsigned long nr_pages = 0;
1727 unsigned long first_init_pfn, spfn, epfn, t, flags;
1728 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
1729 phys_addr_t spa, epa;
1730 u64 i;
1731
1732 /* Only the last zone may have deferred pages */
1733 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
1734 return false;
1735
1736 pgdat_resize_lock(pgdat, &flags);
1737
1738 /*
1739 * If deferred pages have been initialized while we were waiting for
1740 * the lock, return true, as the zone was grown. The caller will retry
1741 * this zone. We won't return to this function since the caller also
1742 * has this static branch.
1743 */
1744 if (!static_branch_unlikely(&deferred_pages)) {
1745 pgdat_resize_unlock(pgdat, &flags);
1746 return true;
1747 }
1748
1749 /*
1750 * If someone grew this zone while we were waiting for spinlock, return
1751 * true, as there might be enough pages already.
1752 */
1753 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
1754 pgdat_resize_unlock(pgdat, &flags);
1755 return true;
1756 }
1757
1758 first_init_pfn = max(zone->zone_start_pfn, first_deferred_pfn);
1759
1760 if (first_init_pfn >= pgdat_end_pfn(pgdat)) {
1761 pgdat_resize_unlock(pgdat, &flags);
1762 return false;
1763 }
1764
1765 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1766 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1767 epfn = min_t(unsigned long, zone_end_pfn(zone), PFN_DOWN(epa));
1768
1769 while (spfn < epfn && nr_pages < nr_pages_needed) {
1770 t = ALIGN(spfn + PAGES_PER_SECTION, PAGES_PER_SECTION);
1771 first_deferred_pfn = min(t, epfn);
1772 nr_pages += deferred_init_pages(nid, zid, spfn,
1773 first_deferred_pfn);
1774 spfn = first_deferred_pfn;
1775 }
1776
1777 if (nr_pages >= nr_pages_needed)
1778 break;
1779 }
1780
1781 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1782 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1783 epfn = min_t(unsigned long, first_deferred_pfn, PFN_DOWN(epa));
1784 deferred_free_pages(nid, zid, spfn, epfn);
1785
1786 if (first_deferred_pfn == epfn)
1787 break;
1788 }
1789 pgdat->first_deferred_pfn = first_deferred_pfn;
1790 pgdat_resize_unlock(pgdat, &flags);
1791
1792 return nr_pages > 0;
1793}
1794
1795/*
1796 * deferred_grow_zone() is __init, but it is called from
1797 * get_page_from_freelist() during early boot until deferred_pages permanently
1798 * disables this call. This is why we have refdata wrapper to avoid warning,
1799 * and to ensure that the function body gets unloaded.
1800 */
1801static bool __ref
1802_deferred_grow_zone(struct zone *zone, unsigned int order)
1803{
1804 return deferred_grow_zone(zone, order);
1805}
1806
1807#endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1808
1809void __init page_alloc_init_late(void)
1810{
1811 struct zone *zone;
1812
1813#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1814 int nid;
1815
1816 /* There will be num_node_state(N_MEMORY) threads */
1817 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1818 for_each_node_state(nid, N_MEMORY) {
1819 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1820 }
1821
1822 /* Block until all are initialised */
1823 wait_for_completion(&pgdat_init_all_done_comp);
1824
1825 /*
1826 * We initialized the rest of the deferred pages. Permanently disable
1827 * on-demand struct page initialization.
1828 */
1829 static_branch_disable(&deferred_pages);
1830
1831 /* Reinit limits that are based on free pages after the kernel is up */
1832 files_maxfiles_init();
1833#endif
1834#ifdef CONFIG_ARCH_DISCARD_MEMBLOCK
1835 /* Discard memblock private memory */
1836 memblock_discard();
1837#endif
1838
1839 for_each_populated_zone(zone)
1840 set_zone_contiguous(zone);
1841}
1842
1843#ifdef CONFIG_CMA
1844/* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1845void __init init_cma_reserved_pageblock(struct page *page)
1846{
1847 unsigned i = pageblock_nr_pages;
1848 struct page *p = page;
1849
1850 do {
1851 __ClearPageReserved(p);
1852 set_page_count(p, 0);
1853 } while (++p, --i);
1854
1855 set_pageblock_migratetype(page, MIGRATE_CMA);
1856
1857 if (pageblock_order >= MAX_ORDER) {
1858 i = pageblock_nr_pages;
1859 p = page;
1860 do {
1861 set_page_refcounted(p);
1862 __free_pages(p, MAX_ORDER - 1);
1863 p += MAX_ORDER_NR_PAGES;
1864 } while (i -= MAX_ORDER_NR_PAGES);
1865 } else {
1866 set_page_refcounted(page);
1867 __free_pages(page, pageblock_order);
1868 }
1869
1870 adjust_managed_page_count(page, pageblock_nr_pages);
1871}
1872#endif
1873
1874/*
1875 * The order of subdivision here is critical for the IO subsystem.
1876 * Please do not alter this order without good reasons and regression
1877 * testing. Specifically, as large blocks of memory are subdivided,
1878 * the order in which smaller blocks are delivered depends on the order
1879 * they're subdivided in this function. This is the primary factor
1880 * influencing the order in which pages are delivered to the IO
1881 * subsystem according to empirical testing, and this is also justified
1882 * by considering the behavior of a buddy system containing a single
1883 * large block of memory acted on by a series of small allocations.
1884 * This behavior is a critical factor in sglist merging's success.
1885 *
1886 * -- nyc
1887 */
1888static inline void expand(struct zone *zone, struct page *page,
1889 int low, int high, struct free_area *area,
1890 int migratetype)
1891{
1892 unsigned long size = 1 << high;
1893
1894 while (high > low) {
1895 area--;
1896 high--;
1897 size >>= 1;
1898 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1899
1900 /*
1901 * Mark as guard pages (or page), that will allow to
1902 * merge back to allocator when buddy will be freed.
1903 * Corresponding page table entries will not be touched,
1904 * pages will stay not present in virtual address space
1905 */
1906 if (set_page_guard(zone, &page[size], high, migratetype))
1907 continue;
1908
1909 list_add(&page[size].lru, &area->free_list[migratetype]);
1910 area->nr_free++;
1911 set_page_order(&page[size], high);
1912 }
1913}
1914
1915static void check_new_page_bad(struct page *page)
1916{
1917 const char *bad_reason = NULL;
1918 unsigned long bad_flags = 0;
1919
1920 if (unlikely(atomic_read(&page->_mapcount) != -1))
1921 bad_reason = "nonzero mapcount";
1922 if (unlikely(page->mapping != NULL))
1923 bad_reason = "non-NULL mapping";
1924 if (unlikely(page_ref_count(page) != 0))
1925 bad_reason = "nonzero _count";
1926 if (unlikely(page->flags & __PG_HWPOISON)) {
1927 bad_reason = "HWPoisoned (hardware-corrupted)";
1928 bad_flags = __PG_HWPOISON;
1929 /* Don't complain about hwpoisoned pages */
1930 page_mapcount_reset(page); /* remove PageBuddy */
1931 return;
1932 }
1933 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1934 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1935 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1936 }
1937#ifdef CONFIG_MEMCG
1938 if (unlikely(page->mem_cgroup))
1939 bad_reason = "page still charged to cgroup";
1940#endif
1941 bad_page(page, bad_reason, bad_flags);
1942}
1943
1944/*
1945 * This page is about to be returned from the page allocator
1946 */
1947static inline int check_new_page(struct page *page)
1948{
1949 if (likely(page_expected_state(page,
1950 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
1951 return 0;
1952
1953 check_new_page_bad(page);
1954 return 1;
1955}
1956
1957static inline bool free_pages_prezeroed(void)
1958{
1959 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
1960 page_poisoning_enabled();
1961}
1962
1963#ifdef CONFIG_DEBUG_VM
1964static bool check_pcp_refill(struct page *page)
1965{
1966 return false;
1967}
1968
1969static bool check_new_pcp(struct page *page)
1970{
1971 return check_new_page(page);
1972}
1973#else
1974static bool check_pcp_refill(struct page *page)
1975{
1976 return check_new_page(page);
1977}
1978static bool check_new_pcp(struct page *page)
1979{
1980 return false;
1981}
1982#endif /* CONFIG_DEBUG_VM */
1983
1984static bool check_new_pages(struct page *page, unsigned int order)
1985{
1986 int i;
1987 for (i = 0; i < (1 << order); i++) {
1988 struct page *p = page + i;
1989
1990 if (unlikely(check_new_page(p)))
1991 return true;
1992 }
1993
1994 return false;
1995}
1996
1997inline void post_alloc_hook(struct page *page, unsigned int order,
1998 gfp_t gfp_flags)
1999{
2000 set_page_private(page, 0);
2001 set_page_refcounted(page);
2002
2003 arch_alloc_page(page, order);
2004 kernel_map_pages(page, 1 << order, 1);
2005 kasan_alloc_pages(page, order);
2006 kernel_poison_pages(page, 1 << order, 1);
2007 set_page_owner(page, order, gfp_flags);
2008}
2009
2010static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2011 unsigned int alloc_flags)
2012{
2013 int i;
2014
2015 post_alloc_hook(page, order, gfp_flags);
2016
2017 if (!free_pages_prezeroed() && (gfp_flags & __GFP_ZERO))
2018 for (i = 0; i < (1 << order); i++)
2019 clear_highpage(page + i);
2020
2021 if (order && (gfp_flags & __GFP_COMP))
2022 prep_compound_page(page, order);
2023
2024 /*
2025 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2026 * allocate the page. The expectation is that the caller is taking
2027 * steps that will free more memory. The caller should avoid the page
2028 * being used for !PFMEMALLOC purposes.
2029 */
2030 if (alloc_flags & ALLOC_NO_WATERMARKS)
2031 set_page_pfmemalloc(page);
2032 else
2033 clear_page_pfmemalloc(page);
2034}
2035
2036/*
2037 * Go through the free lists for the given migratetype and remove
2038 * the smallest available page from the freelists
2039 */
2040static __always_inline
2041struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2042 int migratetype)
2043{
2044 unsigned int current_order;
2045 struct free_area *area;
2046 struct page *page;
2047
2048 /* Find a page of the appropriate size in the preferred list */
2049 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2050 area = &(zone->free_area[current_order]);
2051 page = list_first_entry_or_null(&area->free_list[migratetype],
2052 struct page, lru);
2053 if (!page)
2054 continue;
2055 list_del(&page->lru);
2056 rmv_page_order(page);
2057 area->nr_free--;
2058 expand(zone, page, order, current_order, area, migratetype);
2059 set_pcppage_migratetype(page, migratetype);
2060 return page;
2061 }
2062
2063 return NULL;
2064}
2065
2066
2067/*
2068 * This array describes the order lists are fallen back to when
2069 * the free lists for the desirable migrate type are depleted
2070 */
2071static int fallbacks[MIGRATE_TYPES][4] = {
2072 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2073 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2074 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2075#ifdef CONFIG_CMA
2076 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
2077#endif
2078#ifdef CONFIG_MEMORY_ISOLATION
2079 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
2080#endif
2081};
2082
2083#ifdef CONFIG_CMA
2084static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2085 unsigned int order)
2086{
2087 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2088}
2089#else
2090static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2091 unsigned int order) { return NULL; }
2092#endif
2093
2094/*
2095 * Move the free pages in a range to the free lists of the requested type.
2096 * Note that start_page and end_pages are not aligned on a pageblock
2097 * boundary. If alignment is required, use move_freepages_block()
2098 */
2099static int move_freepages(struct zone *zone,
2100 struct page *start_page, struct page *end_page,
2101 int migratetype, int *num_movable)
2102{
2103 struct page *page;
2104 unsigned int order;
2105 int pages_moved = 0;
2106
2107#ifndef CONFIG_HOLES_IN_ZONE
2108 /*
2109 * page_zone is not safe to call in this context when
2110 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
2111 * anyway as we check zone boundaries in move_freepages_block().
2112 * Remove at a later date when no bug reports exist related to
2113 * grouping pages by mobility
2114 */
2115 VM_BUG_ON(pfn_valid(page_to_pfn(start_page)) &&
2116 pfn_valid(page_to_pfn(end_page)) &&
2117 page_zone(start_page) != page_zone(end_page));
2118#endif
2119 for (page = start_page; page <= end_page;) {
2120 if (!pfn_valid_within(page_to_pfn(page))) {
2121 page++;
2122 continue;
2123 }
2124
2125 /* Make sure we are not inadvertently changing nodes */
2126 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2127
2128 if (!PageBuddy(page)) {
2129 /*
2130 * We assume that pages that could be isolated for
2131 * migration are movable. But we don't actually try
2132 * isolating, as that would be expensive.
2133 */
2134 if (num_movable &&
2135 (PageLRU(page) || __PageMovable(page)))
2136 (*num_movable)++;
2137
2138 page++;
2139 continue;
2140 }
2141
2142 order = page_order(page);
2143 list_move(&page->lru,
2144 &zone->free_area[order].free_list[migratetype]);
2145 page += 1 << order;
2146 pages_moved += 1 << order;
2147 }
2148
2149 return pages_moved;
2150}
2151
2152int move_freepages_block(struct zone *zone, struct page *page,
2153 int migratetype, int *num_movable)
2154{
2155 unsigned long start_pfn, end_pfn;
2156 struct page *start_page, *end_page;
2157
2158 if (num_movable)
2159 *num_movable = 0;
2160
2161 start_pfn = page_to_pfn(page);
2162 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
2163 start_page = pfn_to_page(start_pfn);
2164 end_page = start_page + pageblock_nr_pages - 1;
2165 end_pfn = start_pfn + pageblock_nr_pages - 1;
2166
2167 /* Do not cross zone boundaries */
2168 if (!zone_spans_pfn(zone, start_pfn))
2169 start_page = page;
2170 if (!zone_spans_pfn(zone, end_pfn))
2171 return 0;
2172
2173 return move_freepages(zone, start_page, end_page, migratetype,
2174 num_movable);
2175}
2176
2177static void change_pageblock_range(struct page *pageblock_page,
2178 int start_order, int migratetype)
2179{
2180 int nr_pageblocks = 1 << (start_order - pageblock_order);
2181
2182 while (nr_pageblocks--) {
2183 set_pageblock_migratetype(pageblock_page, migratetype);
2184 pageblock_page += pageblock_nr_pages;
2185 }
2186}
2187
2188/*
2189 * When we are falling back to another migratetype during allocation, try to
2190 * steal extra free pages from the same pageblocks to satisfy further
2191 * allocations, instead of polluting multiple pageblocks.
2192 *
2193 * If we are stealing a relatively large buddy page, it is likely there will
2194 * be more free pages in the pageblock, so try to steal them all. For
2195 * reclaimable and unmovable allocations, we steal regardless of page size,
2196 * as fragmentation caused by those allocations polluting movable pageblocks
2197 * is worse than movable allocations stealing from unmovable and reclaimable
2198 * pageblocks.
2199 */
2200static bool can_steal_fallback(unsigned int order, int start_mt)
2201{
2202 /*
2203 * Leaving this order check is intended, although there is
2204 * relaxed order check in next check. The reason is that
2205 * we can actually steal whole pageblock if this condition met,
2206 * but, below check doesn't guarantee it and that is just heuristic
2207 * so could be changed anytime.
2208 */
2209 if (order >= pageblock_order)
2210 return true;
2211
2212 if (order >= pageblock_order / 2 ||
2213 start_mt == MIGRATE_RECLAIMABLE ||
2214 start_mt == MIGRATE_UNMOVABLE ||
2215 page_group_by_mobility_disabled)
2216 return true;
2217
2218 return false;
2219}
2220
2221static inline void boost_watermark(struct zone *zone)
2222{
2223 unsigned long max_boost;
2224
2225 if (!watermark_boost_factor)
2226 return;
2227
2228 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2229 watermark_boost_factor, 10000);
2230
2231 /*
2232 * high watermark may be uninitialised if fragmentation occurs
2233 * very early in boot so do not boost. We do not fall
2234 * through and boost by pageblock_nr_pages as failing
2235 * allocations that early means that reclaim is not going
2236 * to help and it may even be impossible to reclaim the
2237 * boosted watermark resulting in a hang.
2238 */
2239 if (!max_boost)
2240 return;
2241
2242 max_boost = max(pageblock_nr_pages, max_boost);
2243
2244 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2245 max_boost);
2246}
2247
2248/*
2249 * This function implements actual steal behaviour. If order is large enough,
2250 * we can steal whole pageblock. If not, we first move freepages in this
2251 * pageblock to our migratetype and determine how many already-allocated pages
2252 * are there in the pageblock with a compatible migratetype. If at least half
2253 * of pages are free or compatible, we can change migratetype of the pageblock
2254 * itself, so pages freed in the future will be put on the correct free list.
2255 */
2256static void steal_suitable_fallback(struct zone *zone, struct page *page,
2257 unsigned int alloc_flags, int start_type, bool whole_block)
2258{
2259 unsigned int current_order = page_order(page);
2260 struct free_area *area;
2261 int free_pages, movable_pages, alike_pages;
2262 int old_block_type;
2263
2264 old_block_type = get_pageblock_migratetype(page);
2265
2266 /*
2267 * This can happen due to races and we want to prevent broken
2268 * highatomic accounting.
2269 */
2270 if (is_migrate_highatomic(old_block_type))
2271 goto single_page;
2272
2273 /* Take ownership for orders >= pageblock_order */
2274 if (current_order >= pageblock_order) {
2275 change_pageblock_range(page, current_order, start_type);
2276 goto single_page;
2277 }
2278
2279 /*
2280 * Boost watermarks to increase reclaim pressure to reduce the
2281 * likelihood of future fallbacks. Wake kswapd now as the node
2282 * may be balanced overall and kswapd will not wake naturally.
2283 */
2284 boost_watermark(zone);
2285 if (alloc_flags & ALLOC_KSWAPD)
2286 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2287
2288 /* We are not allowed to try stealing from the whole block */
2289 if (!whole_block)
2290 goto single_page;
2291
2292 free_pages = move_freepages_block(zone, page, start_type,
2293 &movable_pages);
2294 /*
2295 * Determine how many pages are compatible with our allocation.
2296 * For movable allocation, it's the number of movable pages which
2297 * we just obtained. For other types it's a bit more tricky.
2298 */
2299 if (start_type == MIGRATE_MOVABLE) {
2300 alike_pages = movable_pages;
2301 } else {
2302 /*
2303 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2304 * to MOVABLE pageblock, consider all non-movable pages as
2305 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2306 * vice versa, be conservative since we can't distinguish the
2307 * exact migratetype of non-movable pages.
2308 */
2309 if (old_block_type == MIGRATE_MOVABLE)
2310 alike_pages = pageblock_nr_pages
2311 - (free_pages + movable_pages);
2312 else
2313 alike_pages = 0;
2314 }
2315
2316 /* moving whole block can fail due to zone boundary conditions */
2317 if (!free_pages)
2318 goto single_page;
2319
2320 /*
2321 * If a sufficient number of pages in the block are either free or of
2322 * comparable migratability as our allocation, claim the whole block.
2323 */
2324 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2325 page_group_by_mobility_disabled)
2326 set_pageblock_migratetype(page, start_type);
2327
2328 return;
2329
2330single_page:
2331 area = &zone->free_area[current_order];
2332 list_move(&page->lru, &area->free_list[start_type]);
2333}
2334
2335/*
2336 * Check whether there is a suitable fallback freepage with requested order.
2337 * If only_stealable is true, this function returns fallback_mt only if
2338 * we can steal other freepages all together. This would help to reduce
2339 * fragmentation due to mixed migratetype pages in one pageblock.
2340 */
2341int find_suitable_fallback(struct free_area *area, unsigned int order,
2342 int migratetype, bool only_stealable, bool *can_steal)
2343{
2344 int i;
2345 int fallback_mt;
2346
2347 if (area->nr_free == 0)
2348 return -1;
2349
2350 *can_steal = false;
2351 for (i = 0;; i++) {
2352 fallback_mt = fallbacks[migratetype][i];
2353 if (fallback_mt == MIGRATE_TYPES)
2354 break;
2355
2356 if (list_empty(&area->free_list[fallback_mt]))
2357 continue;
2358
2359 if (can_steal_fallback(order, migratetype))
2360 *can_steal = true;
2361
2362 if (!only_stealable)
2363 return fallback_mt;
2364
2365 if (*can_steal)
2366 return fallback_mt;
2367 }
2368
2369 return -1;
2370}
2371
2372/*
2373 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2374 * there are no empty page blocks that contain a page with a suitable order
2375 */
2376static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2377 unsigned int alloc_order)
2378{
2379 int mt;
2380 unsigned long max_managed, flags;
2381
2382 /*
2383 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2384 * Check is race-prone but harmless.
2385 */
2386 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2387 if (zone->nr_reserved_highatomic >= max_managed)
2388 return;
2389
2390 spin_lock_irqsave(&zone->lock, flags);
2391
2392 /* Recheck the nr_reserved_highatomic limit under the lock */
2393 if (zone->nr_reserved_highatomic >= max_managed)
2394 goto out_unlock;
2395
2396 /* Yoink! */
2397 mt = get_pageblock_migratetype(page);
2398 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2399 && !is_migrate_cma(mt)) {
2400 zone->nr_reserved_highatomic += pageblock_nr_pages;
2401 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2402 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2403 }
2404
2405out_unlock:
2406 spin_unlock_irqrestore(&zone->lock, flags);
2407}
2408
2409/*
2410 * Used when an allocation is about to fail under memory pressure. This
2411 * potentially hurts the reliability of high-order allocations when under
2412 * intense memory pressure but failed atomic allocations should be easier
2413 * to recover from than an OOM.
2414 *
2415 * If @force is true, try to unreserve a pageblock even though highatomic
2416 * pageblock is exhausted.
2417 */
2418static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2419 bool force)
2420{
2421 struct zonelist *zonelist = ac->zonelist;
2422 unsigned long flags;
2423 struct zoneref *z;
2424 struct zone *zone;
2425 struct page *page;
2426 int order;
2427 bool ret;
2428
2429 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2430 ac->nodemask) {
2431 /*
2432 * Preserve at least one pageblock unless memory pressure
2433 * is really high.
2434 */
2435 if (!force && zone->nr_reserved_highatomic <=
2436 pageblock_nr_pages)
2437 continue;
2438
2439 spin_lock_irqsave(&zone->lock, flags);
2440 for (order = 0; order < MAX_ORDER; order++) {
2441 struct free_area *area = &(zone->free_area[order]);
2442
2443 page = list_first_entry_or_null(
2444 &area->free_list[MIGRATE_HIGHATOMIC],
2445 struct page, lru);
2446 if (!page)
2447 continue;
2448
2449 /*
2450 * In page freeing path, migratetype change is racy so
2451 * we can counter several free pages in a pageblock
2452 * in this loop althoug we changed the pageblock type
2453 * from highatomic to ac->migratetype. So we should
2454 * adjust the count once.
2455 */
2456 if (is_migrate_highatomic_page(page)) {
2457 /*
2458 * It should never happen but changes to
2459 * locking could inadvertently allow a per-cpu
2460 * drain to add pages to MIGRATE_HIGHATOMIC
2461 * while unreserving so be safe and watch for
2462 * underflows.
2463 */
2464 zone->nr_reserved_highatomic -= min(
2465 pageblock_nr_pages,
2466 zone->nr_reserved_highatomic);
2467 }
2468
2469 /*
2470 * Convert to ac->migratetype and avoid the normal
2471 * pageblock stealing heuristics. Minimally, the caller
2472 * is doing the work and needs the pages. More
2473 * importantly, if the block was always converted to
2474 * MIGRATE_UNMOVABLE or another type then the number
2475 * of pageblocks that cannot be completely freed
2476 * may increase.
2477 */
2478 set_pageblock_migratetype(page, ac->migratetype);
2479 ret = move_freepages_block(zone, page, ac->migratetype,
2480 NULL);
2481 if (ret) {
2482 spin_unlock_irqrestore(&zone->lock, flags);
2483 return ret;
2484 }
2485 }
2486 spin_unlock_irqrestore(&zone->lock, flags);
2487 }
2488
2489 return false;
2490}
2491
2492/*
2493 * Try finding a free buddy page on the fallback list and put it on the free
2494 * list of requested migratetype, possibly along with other pages from the same
2495 * block, depending on fragmentation avoidance heuristics. Returns true if
2496 * fallback was found so that __rmqueue_smallest() can grab it.
2497 *
2498 * The use of signed ints for order and current_order is a deliberate
2499 * deviation from the rest of this file, to make the for loop
2500 * condition simpler.
2501 */
2502static __always_inline bool
2503__rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2504 unsigned int alloc_flags)
2505{
2506 struct free_area *area;
2507 int current_order;
2508 int min_order = order;
2509 struct page *page;
2510 int fallback_mt;
2511 bool can_steal;
2512
2513 /*
2514 * Do not steal pages from freelists belonging to other pageblocks
2515 * i.e. orders < pageblock_order. If there are no local zones free,
2516 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2517 */
2518 if (alloc_flags & ALLOC_NOFRAGMENT)
2519 min_order = pageblock_order;
2520
2521 /*
2522 * Find the largest available free page in the other list. This roughly
2523 * approximates finding the pageblock with the most free pages, which
2524 * would be too costly to do exactly.
2525 */
2526 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2527 --current_order) {
2528 area = &(zone->free_area[current_order]);
2529 fallback_mt = find_suitable_fallback(area, current_order,
2530 start_migratetype, false, &can_steal);
2531 if (fallback_mt == -1)
2532 continue;
2533
2534 /*
2535 * We cannot steal all free pages from the pageblock and the
2536 * requested migratetype is movable. In that case it's better to
2537 * steal and split the smallest available page instead of the
2538 * largest available page, because even if the next movable
2539 * allocation falls back into a different pageblock than this
2540 * one, it won't cause permanent fragmentation.
2541 */
2542 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2543 && current_order > order)
2544 goto find_smallest;
2545
2546 goto do_steal;
2547 }
2548
2549 return false;
2550
2551find_smallest:
2552 for (current_order = order; current_order < MAX_ORDER;
2553 current_order++) {
2554 area = &(zone->free_area[current_order]);
2555 fallback_mt = find_suitable_fallback(area, current_order,
2556 start_migratetype, false, &can_steal);
2557 if (fallback_mt != -1)
2558 break;
2559 }
2560
2561 /*
2562 * This should not happen - we already found a suitable fallback
2563 * when looking for the largest page.
2564 */
2565 VM_BUG_ON(current_order == MAX_ORDER);
2566
2567do_steal:
2568 page = list_first_entry(&area->free_list[fallback_mt],
2569 struct page, lru);
2570
2571 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2572 can_steal);
2573
2574 trace_mm_page_alloc_extfrag(page, order, current_order,
2575 start_migratetype, fallback_mt);
2576
2577 return true;
2578
2579}
2580
2581/*
2582 * Do the hard work of removing an element from the buddy allocator.
2583 * Call me with the zone->lock already held.
2584 */
2585static __always_inline struct page *
2586__rmqueue(struct zone *zone, unsigned int order, int migratetype,
2587 unsigned int alloc_flags)
2588{
2589 struct page *page;
2590
2591retry:
2592 page = __rmqueue_smallest(zone, order, migratetype);
2593 if (unlikely(!page)) {
2594 if (migratetype == MIGRATE_MOVABLE)
2595 page = __rmqueue_cma_fallback(zone, order);
2596
2597 if (!page && __rmqueue_fallback(zone, order, migratetype,
2598 alloc_flags))
2599 goto retry;
2600 }
2601
2602 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2603 return page;
2604}
2605
2606/*
2607 * Obtain a specified number of elements from the buddy allocator, all under
2608 * a single hold of the lock, for efficiency. Add them to the supplied list.
2609 * Returns the number of new pages which were placed at *list.
2610 */
2611static int rmqueue_bulk(struct zone *zone, unsigned int order,
2612 unsigned long count, struct list_head *list,
2613 int migratetype, unsigned int alloc_flags)
2614{
2615 int i, alloced = 0;
2616
2617 spin_lock(&zone->lock);
2618 for (i = 0; i < count; ++i) {
2619 struct page *page = __rmqueue(zone, order, migratetype,
2620 alloc_flags);
2621 if (unlikely(page == NULL))
2622 break;
2623
2624 if (unlikely(check_pcp_refill(page)))
2625 continue;
2626
2627 /*
2628 * Split buddy pages returned by expand() are received here in
2629 * physical page order. The page is added to the tail of
2630 * caller's list. From the callers perspective, the linked list
2631 * is ordered by page number under some conditions. This is
2632 * useful for IO devices that can forward direction from the
2633 * head, thus also in the physical page order. This is useful
2634 * for IO devices that can merge IO requests if the physical
2635 * pages are ordered properly.
2636 */
2637 list_add_tail(&page->lru, list);
2638 alloced++;
2639 if (is_migrate_cma(get_pcppage_migratetype(page)))
2640 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2641 -(1 << order));
2642 }
2643
2644 /*
2645 * i pages were removed from the buddy list even if some leak due
2646 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2647 * on i. Do not confuse with 'alloced' which is the number of
2648 * pages added to the pcp list.
2649 */
2650 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2651 spin_unlock(&zone->lock);
2652 return alloced;
2653}
2654
2655#ifdef CONFIG_NUMA
2656/*
2657 * Called from the vmstat counter updater to drain pagesets of this
2658 * currently executing processor on remote nodes after they have
2659 * expired.
2660 *
2661 * Note that this function must be called with the thread pinned to
2662 * a single processor.
2663 */
2664void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2665{
2666 unsigned long flags;
2667 int to_drain, batch;
2668
2669 local_irq_save(flags);
2670 batch = READ_ONCE(pcp->batch);
2671 to_drain = min(pcp->count, batch);
2672 if (to_drain > 0)
2673 free_pcppages_bulk(zone, to_drain, pcp);
2674 local_irq_restore(flags);
2675}
2676#endif
2677
2678/*
2679 * Drain pcplists of the indicated processor and zone.
2680 *
2681 * The processor must either be the current processor and the
2682 * thread pinned to the current processor or a processor that
2683 * is not online.
2684 */
2685static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2686{
2687 unsigned long flags;
2688 struct per_cpu_pageset *pset;
2689 struct per_cpu_pages *pcp;
2690
2691 local_irq_save(flags);
2692 pset = per_cpu_ptr(zone->pageset, cpu);
2693
2694 pcp = &pset->pcp;
2695 if (pcp->count)
2696 free_pcppages_bulk(zone, pcp->count, pcp);
2697 local_irq_restore(flags);
2698}
2699
2700/*
2701 * Drain pcplists of all zones on the indicated processor.
2702 *
2703 * The processor must either be the current processor and the
2704 * thread pinned to the current processor or a processor that
2705 * is not online.
2706 */
2707static void drain_pages(unsigned int cpu)
2708{
2709 struct zone *zone;
2710
2711 for_each_populated_zone(zone) {
2712 drain_pages_zone(cpu, zone);
2713 }
2714}
2715
2716/*
2717 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2718 *
2719 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2720 * the single zone's pages.
2721 */
2722void drain_local_pages(struct zone *zone)
2723{
2724 int cpu = smp_processor_id();
2725
2726 if (zone)
2727 drain_pages_zone(cpu, zone);
2728 else
2729 drain_pages(cpu);
2730}
2731
2732static void drain_local_pages_wq(struct work_struct *work)
2733{
2734 struct pcpu_drain *drain;
2735
2736 drain = container_of(work, struct pcpu_drain, work);
2737
2738 /*
2739 * drain_all_pages doesn't use proper cpu hotplug protection so
2740 * we can race with cpu offline when the WQ can move this from
2741 * a cpu pinned worker to an unbound one. We can operate on a different
2742 * cpu which is allright but we also have to make sure to not move to
2743 * a different one.
2744 */
2745 preempt_disable();
2746 drain_local_pages(drain->zone);
2747 preempt_enable();
2748}
2749
2750/*
2751 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2752 *
2753 * When zone parameter is non-NULL, spill just the single zone's pages.
2754 *
2755 * Note that this can be extremely slow as the draining happens in a workqueue.
2756 */
2757void drain_all_pages(struct zone *zone)
2758{
2759 int cpu;
2760
2761 /*
2762 * Allocate in the BSS so we wont require allocation in
2763 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2764 */
2765 static cpumask_t cpus_with_pcps;
2766
2767 /*
2768 * Make sure nobody triggers this path before mm_percpu_wq is fully
2769 * initialized.
2770 */
2771 if (WARN_ON_ONCE(!mm_percpu_wq))
2772 return;
2773
2774 /*
2775 * Do not drain if one is already in progress unless it's specific to
2776 * a zone. Such callers are primarily CMA and memory hotplug and need
2777 * the drain to be complete when the call returns.
2778 */
2779 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2780 if (!zone)
2781 return;
2782 mutex_lock(&pcpu_drain_mutex);
2783 }
2784
2785 /*
2786 * We don't care about racing with CPU hotplug event
2787 * as offline notification will cause the notified
2788 * cpu to drain that CPU pcps and on_each_cpu_mask
2789 * disables preemption as part of its processing
2790 */
2791 for_each_online_cpu(cpu) {
2792 struct per_cpu_pageset *pcp;
2793 struct zone *z;
2794 bool has_pcps = false;
2795
2796 if (zone) {
2797 pcp = per_cpu_ptr(zone->pageset, cpu);
2798 if (pcp->pcp.count)
2799 has_pcps = true;
2800 } else {
2801 for_each_populated_zone(z) {
2802 pcp = per_cpu_ptr(z->pageset, cpu);
2803 if (pcp->pcp.count) {
2804 has_pcps = true;
2805 break;
2806 }
2807 }
2808 }
2809
2810 if (has_pcps)
2811 cpumask_set_cpu(cpu, &cpus_with_pcps);
2812 else
2813 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2814 }
2815
2816 for_each_cpu(cpu, &cpus_with_pcps) {
2817 struct pcpu_drain *drain = per_cpu_ptr(&pcpu_drain, cpu);
2818
2819 drain->zone = zone;
2820 INIT_WORK(&drain->work, drain_local_pages_wq);
2821 queue_work_on(cpu, mm_percpu_wq, &drain->work);
2822 }
2823 for_each_cpu(cpu, &cpus_with_pcps)
2824 flush_work(&per_cpu_ptr(&pcpu_drain, cpu)->work);
2825
2826 mutex_unlock(&pcpu_drain_mutex);
2827}
2828
2829#ifdef CONFIG_HIBERNATION
2830
2831/*
2832 * Touch the watchdog for every WD_PAGE_COUNT pages.
2833 */
2834#define WD_PAGE_COUNT (128*1024)
2835
2836void mark_free_pages(struct zone *zone)
2837{
2838 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
2839 unsigned long flags;
2840 unsigned int order, t;
2841 struct page *page;
2842
2843 if (zone_is_empty(zone))
2844 return;
2845
2846 spin_lock_irqsave(&zone->lock, flags);
2847
2848 max_zone_pfn = zone_end_pfn(zone);
2849 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2850 if (pfn_valid(pfn)) {
2851 page = pfn_to_page(pfn);
2852
2853 if (!--page_count) {
2854 touch_nmi_watchdog();
2855 page_count = WD_PAGE_COUNT;
2856 }
2857
2858 if (page_zone(page) != zone)
2859 continue;
2860
2861 if (!swsusp_page_is_forbidden(page))
2862 swsusp_unset_page_free(page);
2863 }
2864
2865 for_each_migratetype_order(order, t) {
2866 list_for_each_entry(page,
2867 &zone->free_area[order].free_list[t], lru) {
2868 unsigned long i;
2869
2870 pfn = page_to_pfn(page);
2871 for (i = 0; i < (1UL << order); i++) {
2872 if (!--page_count) {
2873 touch_nmi_watchdog();
2874 page_count = WD_PAGE_COUNT;
2875 }
2876 swsusp_set_page_free(pfn_to_page(pfn + i));
2877 }
2878 }
2879 }
2880 spin_unlock_irqrestore(&zone->lock, flags);
2881}
2882#endif /* CONFIG_PM */
2883
2884static bool free_unref_page_prepare(struct page *page, unsigned long pfn)
2885{
2886 int migratetype;
2887
2888 if (!free_pcp_prepare(page))
2889 return false;
2890
2891 migratetype = get_pfnblock_migratetype(page, pfn);
2892 set_pcppage_migratetype(page, migratetype);
2893 return true;
2894}
2895
2896static void free_unref_page_commit(struct page *page, unsigned long pfn)
2897{
2898 struct zone *zone = page_zone(page);
2899 struct per_cpu_pages *pcp;
2900 int migratetype;
2901
2902 migratetype = get_pcppage_migratetype(page);
2903 __count_vm_event(PGFREE);
2904
2905 /*
2906 * We only track unmovable, reclaimable and movable on pcp lists.
2907 * Free ISOLATE pages back to the allocator because they are being
2908 * offlined but treat HIGHATOMIC as movable pages so we can get those
2909 * areas back if necessary. Otherwise, we may have to free
2910 * excessively into the page allocator
2911 */
2912 if (migratetype >= MIGRATE_PCPTYPES) {
2913 if (unlikely(is_migrate_isolate(migratetype))) {
2914 free_one_page(zone, page, pfn, 0, migratetype);
2915 return;
2916 }
2917 migratetype = MIGRATE_MOVABLE;
2918 }
2919
2920 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2921 list_add(&page->lru, &pcp->lists[migratetype]);
2922 pcp->count++;
2923 if (pcp->count >= pcp->high) {
2924 unsigned long batch = READ_ONCE(pcp->batch);
2925 free_pcppages_bulk(zone, batch, pcp);
2926 }
2927}
2928
2929/*
2930 * Free a 0-order page
2931 */
2932void free_unref_page(struct page *page)
2933{
2934 unsigned long flags;
2935 unsigned long pfn = page_to_pfn(page);
2936
2937 if (!free_unref_page_prepare(page, pfn))
2938 return;
2939
2940 local_irq_save(flags);
2941 free_unref_page_commit(page, pfn);
2942 local_irq_restore(flags);
2943}
2944
2945/*
2946 * Free a list of 0-order pages
2947 */
2948void free_unref_page_list(struct list_head *list)
2949{
2950 struct page *page, *next;
2951 unsigned long flags, pfn;
2952 int batch_count = 0;
2953
2954 /* Prepare pages for freeing */
2955 list_for_each_entry_safe(page, next, list, lru) {
2956 pfn = page_to_pfn(page);
2957 if (!free_unref_page_prepare(page, pfn))
2958 list_del(&page->lru);
2959 set_page_private(page, pfn);
2960 }
2961
2962 local_irq_save(flags);
2963 list_for_each_entry_safe(page, next, list, lru) {
2964 unsigned long pfn = page_private(page);
2965
2966 set_page_private(page, 0);
2967 trace_mm_page_free_batched(page);
2968 free_unref_page_commit(page, pfn);
2969
2970 /*
2971 * Guard against excessive IRQ disabled times when we get
2972 * a large list of pages to free.
2973 */
2974 if (++batch_count == SWAP_CLUSTER_MAX) {
2975 local_irq_restore(flags);
2976 batch_count = 0;
2977 local_irq_save(flags);
2978 }
2979 }
2980 local_irq_restore(flags);
2981}
2982
2983/*
2984 * split_page takes a non-compound higher-order page, and splits it into
2985 * n (1<<order) sub-pages: page[0..n]
2986 * Each sub-page must be freed individually.
2987 *
2988 * Note: this is probably too low level an operation for use in drivers.
2989 * Please consult with lkml before using this in your driver.
2990 */
2991void split_page(struct page *page, unsigned int order)
2992{
2993 int i;
2994
2995 VM_BUG_ON_PAGE(PageCompound(page), page);
2996 VM_BUG_ON_PAGE(!page_count(page), page);
2997
2998 for (i = 1; i < (1 << order); i++)
2999 set_page_refcounted(page + i);
3000 split_page_owner(page, order);
3001}
3002EXPORT_SYMBOL_GPL(split_page);
3003
3004int __isolate_free_page(struct page *page, unsigned int order)
3005{
3006 unsigned long watermark;
3007 struct zone *zone;
3008 int mt;
3009
3010 BUG_ON(!PageBuddy(page));
3011
3012 zone = page_zone(page);
3013 mt = get_pageblock_migratetype(page);
3014
3015 if (!is_migrate_isolate(mt)) {
3016 /*
3017 * Obey watermarks as if the page was being allocated. We can
3018 * emulate a high-order watermark check with a raised order-0
3019 * watermark, because we already know our high-order page
3020 * exists.
3021 */
3022 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3023 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3024 return 0;
3025
3026 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3027 }
3028
3029 /* Remove page from free list */
3030 list_del(&page->lru);
3031 zone->free_area[order].nr_free--;
3032 rmv_page_order(page);
3033
3034 /*
3035 * Set the pageblock if the isolated page is at least half of a
3036 * pageblock
3037 */
3038 if (order >= pageblock_order - 1) {
3039 struct page *endpage = page + (1 << order) - 1;
3040 for (; page < endpage; page += pageblock_nr_pages) {
3041 int mt = get_pageblock_migratetype(page);
3042 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
3043 && !is_migrate_highatomic(mt))
3044 set_pageblock_migratetype(page,
3045 MIGRATE_MOVABLE);
3046 }
3047 }
3048
3049
3050 return 1UL << order;
3051}
3052
3053/*
3054 * Update NUMA hit/miss statistics
3055 *
3056 * Must be called with interrupts disabled.
3057 */
3058static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
3059{
3060#ifdef CONFIG_NUMA
3061 enum numa_stat_item local_stat = NUMA_LOCAL;
3062
3063 /* skip numa counters update if numa stats is disabled */
3064 if (!static_branch_likely(&vm_numa_stat_key))
3065 return;
3066
3067 if (zone_to_nid(z) != numa_node_id())
3068 local_stat = NUMA_OTHER;
3069
3070 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3071 __inc_numa_state(z, NUMA_HIT);
3072 else {
3073 __inc_numa_state(z, NUMA_MISS);
3074 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
3075 }
3076 __inc_numa_state(z, local_stat);
3077#endif
3078}
3079
3080/* Remove page from the per-cpu list, caller must protect the list */
3081static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
3082 unsigned int alloc_flags,
3083 struct per_cpu_pages *pcp,
3084 struct list_head *list)
3085{
3086 struct page *page;
3087
3088 do {
3089 if (list_empty(list)) {
3090 pcp->count += rmqueue_bulk(zone, 0,
3091 pcp->batch, list,
3092 migratetype, alloc_flags);
3093 if (unlikely(list_empty(list)))
3094 return NULL;
3095 }
3096
3097 page = list_first_entry(list, struct page, lru);
3098 list_del(&page->lru);
3099 pcp->count--;
3100 } while (check_new_pcp(page));
3101
3102 return page;
3103}
3104
3105/* Lock and remove page from the per-cpu list */
3106static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3107 struct zone *zone, unsigned int order,
3108 gfp_t gfp_flags, int migratetype,
3109 unsigned int alloc_flags)
3110{
3111 struct per_cpu_pages *pcp;
3112 struct list_head *list;
3113 struct page *page;
3114 unsigned long flags;
3115
3116 local_irq_save(flags);
3117 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3118 list = &pcp->lists[migratetype];
3119 page = __rmqueue_pcplist(zone, migratetype, alloc_flags, pcp, list);
3120 if (page) {
3121 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3122 zone_statistics(preferred_zone, zone);
3123 }
3124 local_irq_restore(flags);
3125 return page;
3126}
3127
3128/*
3129 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3130 */
3131static inline
3132struct page *rmqueue(struct zone *preferred_zone,
3133 struct zone *zone, unsigned int order,
3134 gfp_t gfp_flags, unsigned int alloc_flags,
3135 int migratetype)
3136{
3137 unsigned long flags;
3138 struct page *page;
3139
3140 if (likely(order == 0)) {
3141 page = rmqueue_pcplist(preferred_zone, zone, order,
3142 gfp_flags, migratetype, alloc_flags);
3143 goto out;
3144 }
3145
3146 /*
3147 * We most definitely don't want callers attempting to
3148 * allocate greater than order-1 page units with __GFP_NOFAIL.
3149 */
3150 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3151 spin_lock_irqsave(&zone->lock, flags);
3152
3153 do {
3154 page = NULL;
3155 if (alloc_flags & ALLOC_HARDER) {
3156 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3157 if (page)
3158 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3159 }
3160 if (!page)
3161 page = __rmqueue(zone, order, migratetype, alloc_flags);
3162 } while (page && check_new_pages(page, order));
3163 spin_unlock(&zone->lock);
3164 if (!page)
3165 goto failed;
3166 __mod_zone_freepage_state(zone, -(1 << order),
3167 get_pcppage_migratetype(page));
3168
3169 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3170 zone_statistics(preferred_zone, zone);
3171 local_irq_restore(flags);
3172
3173out:
3174 /* Separate test+clear to avoid unnecessary atomics */
3175 if (test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags)) {
3176 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3177 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3178 }
3179
3180 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3181 return page;
3182
3183failed:
3184 local_irq_restore(flags);
3185 return NULL;
3186}
3187
3188#ifdef CONFIG_FAIL_PAGE_ALLOC
3189
3190static struct {
3191 struct fault_attr attr;
3192
3193 bool ignore_gfp_highmem;
3194 bool ignore_gfp_reclaim;
3195 u32 min_order;
3196} fail_page_alloc = {
3197 .attr = FAULT_ATTR_INITIALIZER,
3198 .ignore_gfp_reclaim = true,
3199 .ignore_gfp_highmem = true,
3200 .min_order = 1,
3201};
3202
3203static int __init setup_fail_page_alloc(char *str)
3204{
3205 return setup_fault_attr(&fail_page_alloc.attr, str);
3206}
3207__setup("fail_page_alloc=", setup_fail_page_alloc);
3208
3209static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3210{
3211 if (order < fail_page_alloc.min_order)
3212 return false;
3213 if (gfp_mask & __GFP_NOFAIL)
3214 return false;
3215 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3216 return false;
3217 if (fail_page_alloc.ignore_gfp_reclaim &&
3218 (gfp_mask & __GFP_DIRECT_RECLAIM))
3219 return false;
3220
3221 return should_fail(&fail_page_alloc.attr, 1 << order);
3222}
3223
3224#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3225
3226static int __init fail_page_alloc_debugfs(void)
3227{
3228 umode_t mode = S_IFREG | 0600;
3229 struct dentry *dir;
3230
3231 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3232 &fail_page_alloc.attr);
3233
3234 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3235 &fail_page_alloc.ignore_gfp_reclaim);
3236 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3237 &fail_page_alloc.ignore_gfp_highmem);
3238 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3239
3240 return 0;
3241}
3242
3243late_initcall(fail_page_alloc_debugfs);
3244
3245#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3246
3247#else /* CONFIG_FAIL_PAGE_ALLOC */
3248
3249static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3250{
3251 return false;
3252}
3253
3254#endif /* CONFIG_FAIL_PAGE_ALLOC */
3255
3256static noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3257{
3258 return __should_fail_alloc_page(gfp_mask, order);
3259}
3260ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3261
3262/*
3263 * Return true if free base pages are above 'mark'. For high-order checks it
3264 * will return true of the order-0 watermark is reached and there is at least
3265 * one free page of a suitable size. Checking now avoids taking the zone lock
3266 * to check in the allocation paths if no pages are free.
3267 */
3268bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3269 int classzone_idx, unsigned int alloc_flags,
3270 long free_pages)
3271{
3272 long min = mark;
3273 int o;
3274 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3275
3276 /* free_pages may go negative - that's OK */
3277 free_pages -= (1 << order) - 1;
3278
3279 if (alloc_flags & ALLOC_HIGH)
3280 min -= min / 2;
3281
3282 /*
3283 * If the caller does not have rights to ALLOC_HARDER then subtract
3284 * the high-atomic reserves. This will over-estimate the size of the
3285 * atomic reserve but it avoids a search.
3286 */
3287 if (likely(!alloc_harder)) {
3288 free_pages -= z->nr_reserved_highatomic;
3289 } else {
3290 /*
3291 * OOM victims can try even harder than normal ALLOC_HARDER
3292 * users on the grounds that it's definitely going to be in
3293 * the exit path shortly and free memory. Any allocation it
3294 * makes during the free path will be small and short-lived.
3295 */
3296 if (alloc_flags & ALLOC_OOM)
3297 min -= min / 2;
3298 else
3299 min -= min / 4;
3300 }
3301
3302
3303#ifdef CONFIG_CMA
3304 /* If allocation can't use CMA areas don't use free CMA pages */
3305 if (!(alloc_flags & ALLOC_CMA))
3306 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
3307#endif
3308
3309 /*
3310 * Check watermarks for an order-0 allocation request. If these
3311 * are not met, then a high-order request also cannot go ahead
3312 * even if a suitable page happened to be free.
3313 */
3314 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
3315 return false;
3316
3317 /* If this is an order-0 request then the watermark is fine */
3318 if (!order)
3319 return true;
3320
3321 /* For a high-order request, check at least one suitable page is free */
3322 for (o = order; o < MAX_ORDER; o++) {
3323 struct free_area *area = &z->free_area[o];
3324 int mt;
3325
3326 if (!area->nr_free)
3327 continue;
3328
3329 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3330 if (!list_empty(&area->free_list[mt]))
3331 return true;
3332 }
3333
3334#ifdef CONFIG_CMA
3335 if ((alloc_flags & ALLOC_CMA) &&
3336 !list_empty(&area->free_list[MIGRATE_CMA])) {
3337 return true;
3338 }
3339#endif
3340 if (alloc_harder &&
3341 !list_empty(&area->free_list[MIGRATE_HIGHATOMIC]))
3342 return true;
3343 }
3344 return false;
3345}
3346
3347bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3348 int classzone_idx, unsigned int alloc_flags)
3349{
3350 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3351 zone_page_state(z, NR_FREE_PAGES));
3352}
3353
3354static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3355 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
3356{
3357 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3358 long cma_pages = 0;
3359
3360#ifdef CONFIG_CMA
3361 /* If allocation can't use CMA areas don't use free CMA pages */
3362 if (!(alloc_flags & ALLOC_CMA))
3363 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
3364#endif
3365
3366 /*
3367 * Fast check for order-0 only. If this fails then the reserves
3368 * need to be calculated. There is a corner case where the check
3369 * passes but only the high-order atomic reserve are free. If
3370 * the caller is !atomic then it'll uselessly search the free
3371 * list. That corner case is then slower but it is harmless.
3372 */
3373 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
3374 return true;
3375
3376 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3377 free_pages);
3378}
3379
3380bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3381 unsigned long mark, int classzone_idx)
3382{
3383 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3384
3385 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3386 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3387
3388 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
3389 free_pages);
3390}
3391
3392#ifdef CONFIG_NUMA
3393static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3394{
3395 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3396 RECLAIM_DISTANCE;
3397}
3398#else /* CONFIG_NUMA */
3399static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3400{
3401 return true;
3402}
3403#endif /* CONFIG_NUMA */
3404
3405/*
3406 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3407 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3408 * premature use of a lower zone may cause lowmem pressure problems that
3409 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3410 * probably too small. It only makes sense to spread allocations to avoid
3411 * fragmentation between the Normal and DMA32 zones.
3412 */
3413static inline unsigned int
3414alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3415{
3416 unsigned int alloc_flags = 0;
3417
3418 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3419 alloc_flags |= ALLOC_KSWAPD;
3420
3421#ifdef CONFIG_ZONE_DMA32
3422 if (zone_idx(zone) != ZONE_NORMAL)
3423 goto out;
3424
3425 /*
3426 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3427 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3428 * on UMA that if Normal is populated then so is DMA32.
3429 */
3430 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3431 if (nr_online_nodes > 1 && !populated_zone(--zone))
3432 goto out;
3433
3434out:
3435#endif /* CONFIG_ZONE_DMA32 */
3436 return alloc_flags;
3437}
3438
3439/*
3440 * get_page_from_freelist goes through the zonelist trying to allocate
3441 * a page.
3442 */
3443static struct page *
3444get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3445 const struct alloc_context *ac)
3446{
3447 struct zoneref *z;
3448 struct zone *zone;
3449 struct pglist_data *last_pgdat_dirty_limit = NULL;
3450 bool no_fallback;
3451
3452retry:
3453 /*
3454 * Scan zonelist, looking for a zone with enough free.
3455 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3456 */
3457 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3458 z = ac->preferred_zoneref;
3459 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3460 ac->nodemask) {
3461 struct page *page;
3462 unsigned long mark;
3463
3464 if (cpusets_enabled() &&
3465 (alloc_flags & ALLOC_CPUSET) &&
3466 !__cpuset_zone_allowed(zone, gfp_mask))
3467 continue;
3468 /*
3469 * When allocating a page cache page for writing, we
3470 * want to get it from a node that is within its dirty
3471 * limit, such that no single node holds more than its
3472 * proportional share of globally allowed dirty pages.
3473 * The dirty limits take into account the node's
3474 * lowmem reserves and high watermark so that kswapd
3475 * should be able to balance it without having to
3476 * write pages from its LRU list.
3477 *
3478 * XXX: For now, allow allocations to potentially
3479 * exceed the per-node dirty limit in the slowpath
3480 * (spread_dirty_pages unset) before going into reclaim,
3481 * which is important when on a NUMA setup the allowed
3482 * nodes are together not big enough to reach the
3483 * global limit. The proper fix for these situations
3484 * will require awareness of nodes in the
3485 * dirty-throttling and the flusher threads.
3486 */
3487 if (ac->spread_dirty_pages) {
3488 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3489 continue;
3490
3491 if (!node_dirty_ok(zone->zone_pgdat)) {
3492 last_pgdat_dirty_limit = zone->zone_pgdat;
3493 continue;
3494 }
3495 }
3496
3497 if (no_fallback && nr_online_nodes > 1 &&
3498 zone != ac->preferred_zoneref->zone) {
3499 int local_nid;
3500
3501 /*
3502 * If moving to a remote node, retry but allow
3503 * fragmenting fallbacks. Locality is more important
3504 * than fragmentation avoidance.
3505 */
3506 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
3507 if (zone_to_nid(zone) != local_nid) {
3508 alloc_flags &= ~ALLOC_NOFRAGMENT;
3509 goto retry;
3510 }
3511 }
3512
3513 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3514 if (!zone_watermark_fast(zone, order, mark,
3515 ac_classzone_idx(ac), alloc_flags)) {
3516 int ret;
3517
3518#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3519 /*
3520 * Watermark failed for this zone, but see if we can
3521 * grow this zone if it contains deferred pages.
3522 */
3523 if (static_branch_unlikely(&deferred_pages)) {
3524 if (_deferred_grow_zone(zone, order))
3525 goto try_this_zone;
3526 }
3527#endif
3528 /* Checked here to keep the fast path fast */
3529 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3530 if (alloc_flags & ALLOC_NO_WATERMARKS)
3531 goto try_this_zone;
3532
3533 if (node_reclaim_mode == 0 ||
3534 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3535 continue;
3536
3537 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3538 switch (ret) {
3539 case NODE_RECLAIM_NOSCAN:
3540 /* did not scan */
3541 continue;
3542 case NODE_RECLAIM_FULL:
3543 /* scanned but unreclaimable */
3544 continue;
3545 default:
3546 /* did we reclaim enough */
3547 if (zone_watermark_ok(zone, order, mark,
3548 ac_classzone_idx(ac), alloc_flags))
3549 goto try_this_zone;
3550
3551 continue;
3552 }
3553 }
3554
3555try_this_zone:
3556 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3557 gfp_mask, alloc_flags, ac->migratetype);
3558 if (page) {
3559 prep_new_page(page, order, gfp_mask, alloc_flags);
3560
3561 /*
3562 * If this is a high-order atomic allocation then check
3563 * if the pageblock should be reserved for the future
3564 */
3565 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3566 reserve_highatomic_pageblock(page, zone, order);
3567
3568 return page;
3569 } else {
3570#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3571 /* Try again if zone has deferred pages */
3572 if (static_branch_unlikely(&deferred_pages)) {
3573 if (_deferred_grow_zone(zone, order))
3574 goto try_this_zone;
3575 }
3576#endif
3577 }
3578 }
3579
3580 /*
3581 * It's possible on a UMA machine to get through all zones that are
3582 * fragmented. If avoiding fragmentation, reset and try again.
3583 */
3584 if (no_fallback) {
3585 alloc_flags &= ~ALLOC_NOFRAGMENT;
3586 goto retry;
3587 }
3588
3589 return NULL;
3590}
3591
3592static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3593{
3594 unsigned int filter = SHOW_MEM_FILTER_NODES;
3595 static DEFINE_RATELIMIT_STATE(show_mem_rs, HZ, 1);
3596
3597 if (!__ratelimit(&show_mem_rs))
3598 return;
3599
3600 /*
3601 * This documents exceptions given to allocations in certain
3602 * contexts that are allowed to allocate outside current's set
3603 * of allowed nodes.
3604 */
3605 if (!(gfp_mask & __GFP_NOMEMALLOC))
3606 if (tsk_is_oom_victim(current) ||
3607 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3608 filter &= ~SHOW_MEM_FILTER_NODES;
3609 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3610 filter &= ~SHOW_MEM_FILTER_NODES;
3611
3612 show_mem(filter, nodemask);
3613}
3614
3615void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3616{
3617 struct va_format vaf;
3618 va_list args;
3619 static DEFINE_RATELIMIT_STATE(nopage_rs, DEFAULT_RATELIMIT_INTERVAL,
3620 DEFAULT_RATELIMIT_BURST);
3621
3622 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
3623 return;
3624
3625 va_start(args, fmt);
3626 vaf.fmt = fmt;
3627 vaf.va = &args;
3628 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
3629 current->comm, &vaf, gfp_mask, &gfp_mask,
3630 nodemask_pr_args(nodemask));
3631 va_end(args);
3632
3633 cpuset_print_current_mems_allowed();
3634 pr_cont("\n");
3635 dump_stack();
3636 warn_alloc_show_mem(gfp_mask, nodemask);
3637}
3638
3639static inline struct page *
3640__alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3641 unsigned int alloc_flags,
3642 const struct alloc_context *ac)
3643{
3644 struct page *page;
3645
3646 page = get_page_from_freelist(gfp_mask, order,
3647 alloc_flags|ALLOC_CPUSET, ac);
3648 /*
3649 * fallback to ignore cpuset restriction if our nodes
3650 * are depleted
3651 */
3652 if (!page)
3653 page = get_page_from_freelist(gfp_mask, order,
3654 alloc_flags, ac);
3655
3656 return page;
3657}
3658
3659static inline struct page *
3660__alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3661 const struct alloc_context *ac, unsigned long *did_some_progress)
3662{
3663 struct oom_control oc = {
3664 .zonelist = ac->zonelist,
3665 .nodemask = ac->nodemask,
3666 .memcg = NULL,
3667 .gfp_mask = gfp_mask,
3668 .order = order,
3669 };
3670 struct page *page;
3671
3672 *did_some_progress = 0;
3673
3674 /*
3675 * Acquire the oom lock. If that fails, somebody else is
3676 * making progress for us.
3677 */
3678 if (!mutex_trylock(&oom_lock)) {
3679 *did_some_progress = 1;
3680 schedule_timeout_uninterruptible(1);
3681 return NULL;
3682 }
3683
3684 /*
3685 * Go through the zonelist yet one more time, keep very high watermark
3686 * here, this is only to catch a parallel oom killing, we must fail if
3687 * we're still under heavy pressure. But make sure that this reclaim
3688 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3689 * allocation which will never fail due to oom_lock already held.
3690 */
3691 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3692 ~__GFP_DIRECT_RECLAIM, order,
3693 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3694 if (page)
3695 goto out;
3696
3697 /* Coredumps can quickly deplete all memory reserves */
3698 if (current->flags & PF_DUMPCORE)
3699 goto out;
3700 /* The OOM killer will not help higher order allocs */
3701 if (order > PAGE_ALLOC_COSTLY_ORDER)
3702 goto out;
3703 /*
3704 * We have already exhausted all our reclaim opportunities without any
3705 * success so it is time to admit defeat. We will skip the OOM killer
3706 * because it is very likely that the caller has a more reasonable
3707 * fallback than shooting a random task.
3708 */
3709 if (gfp_mask & __GFP_RETRY_MAYFAIL)
3710 goto out;
3711 /* The OOM killer does not needlessly kill tasks for lowmem */
3712 if (ac->high_zoneidx < ZONE_NORMAL)
3713 goto out;
3714 if (pm_suspended_storage())
3715 goto out;
3716 /*
3717 * XXX: GFP_NOFS allocations should rather fail than rely on
3718 * other request to make a forward progress.
3719 * We are in an unfortunate situation where out_of_memory cannot
3720 * do much for this context but let's try it to at least get
3721 * access to memory reserved if the current task is killed (see
3722 * out_of_memory). Once filesystems are ready to handle allocation
3723 * failures more gracefully we should just bail out here.
3724 */
3725
3726 /* The OOM killer may not free memory on a specific node */
3727 if (gfp_mask & __GFP_THISNODE)
3728 goto out;
3729
3730 /* Exhausted what can be done so it's blame time */
3731 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3732 *did_some_progress = 1;
3733
3734 /*
3735 * Help non-failing allocations by giving them access to memory
3736 * reserves
3737 */
3738 if (gfp_mask & __GFP_NOFAIL)
3739 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3740 ALLOC_NO_WATERMARKS, ac);
3741 }
3742out:
3743 mutex_unlock(&oom_lock);
3744 return page;
3745}
3746
3747/*
3748 * Maximum number of compaction retries wit a progress before OOM
3749 * killer is consider as the only way to move forward.
3750 */
3751#define MAX_COMPACT_RETRIES 16
3752
3753#ifdef CONFIG_COMPACTION
3754/* Try memory compaction for high-order allocations before reclaim */
3755static struct page *
3756__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3757 unsigned int alloc_flags, const struct alloc_context *ac,
3758 enum compact_priority prio, enum compact_result *compact_result)
3759{
3760 struct page *page = NULL;
3761 unsigned long pflags;
3762 unsigned int noreclaim_flag;
3763
3764 if (!order)
3765 return NULL;
3766
3767 psi_memstall_enter(&pflags);
3768 noreclaim_flag = memalloc_noreclaim_save();
3769
3770 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3771 prio, &page);
3772
3773 memalloc_noreclaim_restore(noreclaim_flag);
3774 psi_memstall_leave(&pflags);
3775
3776 if (*compact_result <= COMPACT_INACTIVE) {
3777 WARN_ON_ONCE(page);
3778 return NULL;
3779 }
3780
3781 /*
3782 * At least in one zone compaction wasn't deferred or skipped, so let's
3783 * count a compaction stall
3784 */
3785 count_vm_event(COMPACTSTALL);
3786
3787 /* Prep a captured page if available */
3788 if (page)
3789 prep_new_page(page, order, gfp_mask, alloc_flags);
3790
3791 /* Try get a page from the freelist if available */
3792 if (!page)
3793 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3794
3795 if (page) {
3796 struct zone *zone = page_zone(page);
3797
3798 zone->compact_blockskip_flush = false;
3799 compaction_defer_reset(zone, order, true);
3800 count_vm_event(COMPACTSUCCESS);
3801 return page;
3802 }
3803
3804 /*
3805 * It's bad if compaction run occurs and fails. The most likely reason
3806 * is that pages exist, but not enough to satisfy watermarks.
3807 */
3808 count_vm_event(COMPACTFAIL);
3809
3810 cond_resched();
3811
3812 return NULL;
3813}
3814
3815static inline bool
3816should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3817 enum compact_result compact_result,
3818 enum compact_priority *compact_priority,
3819 int *compaction_retries)
3820{
3821 int max_retries = MAX_COMPACT_RETRIES;
3822 int min_priority;
3823 bool ret = false;
3824 int retries = *compaction_retries;
3825 enum compact_priority priority = *compact_priority;
3826
3827 if (!order)
3828 return false;
3829
3830 if (compaction_made_progress(compact_result))
3831 (*compaction_retries)++;
3832
3833 /*
3834 * compaction considers all the zone as desperately out of memory
3835 * so it doesn't really make much sense to retry except when the
3836 * failure could be caused by insufficient priority
3837 */
3838 if (compaction_failed(compact_result))
3839 goto check_priority;
3840
3841 /*
3842 * make sure the compaction wasn't deferred or didn't bail out early
3843 * due to locks contention before we declare that we should give up.
3844 * But do not retry if the given zonelist is not suitable for
3845 * compaction.
3846 */
3847 if (compaction_withdrawn(compact_result)) {
3848 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
3849 goto out;
3850 }
3851
3852 /*
3853 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
3854 * costly ones because they are de facto nofail and invoke OOM
3855 * killer to move on while costly can fail and users are ready
3856 * to cope with that. 1/4 retries is rather arbitrary but we
3857 * would need much more detailed feedback from compaction to
3858 * make a better decision.
3859 */
3860 if (order > PAGE_ALLOC_COSTLY_ORDER)
3861 max_retries /= 4;
3862 if (*compaction_retries <= max_retries) {
3863 ret = true;
3864 goto out;
3865 }
3866
3867 /*
3868 * Make sure there are attempts at the highest priority if we exhausted
3869 * all retries or failed at the lower priorities.
3870 */
3871check_priority:
3872 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
3873 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
3874
3875 if (*compact_priority > min_priority) {
3876 (*compact_priority)--;
3877 *compaction_retries = 0;
3878 ret = true;
3879 }
3880out:
3881 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
3882 return ret;
3883}
3884#else
3885static inline struct page *
3886__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3887 unsigned int alloc_flags, const struct alloc_context *ac,
3888 enum compact_priority prio, enum compact_result *compact_result)
3889{
3890 *compact_result = COMPACT_SKIPPED;
3891 return NULL;
3892}
3893
3894static inline bool
3895should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
3896 enum compact_result compact_result,
3897 enum compact_priority *compact_priority,
3898 int *compaction_retries)
3899{
3900 struct zone *zone;
3901 struct zoneref *z;
3902
3903 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
3904 return false;
3905
3906 /*
3907 * There are setups with compaction disabled which would prefer to loop
3908 * inside the allocator rather than hit the oom killer prematurely.
3909 * Let's give them a good hope and keep retrying while the order-0
3910 * watermarks are OK.
3911 */
3912 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3913 ac->nodemask) {
3914 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
3915 ac_classzone_idx(ac), alloc_flags))
3916 return true;
3917 }
3918 return false;
3919}
3920#endif /* CONFIG_COMPACTION */
3921
3922#ifdef CONFIG_LOCKDEP
3923static struct lockdep_map __fs_reclaim_map =
3924 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
3925
3926static bool __need_fs_reclaim(gfp_t gfp_mask)
3927{
3928 gfp_mask = current_gfp_context(gfp_mask);
3929
3930 /* no reclaim without waiting on it */
3931 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
3932 return false;
3933
3934 /* this guy won't enter reclaim */
3935 if (current->flags & PF_MEMALLOC)
3936 return false;
3937
3938 /* We're only interested __GFP_FS allocations for now */
3939 if (!(gfp_mask & __GFP_FS))
3940 return false;
3941
3942 if (gfp_mask & __GFP_NOLOCKDEP)
3943 return false;
3944
3945 return true;
3946}
3947
3948void __fs_reclaim_acquire(void)
3949{
3950 lock_map_acquire(&__fs_reclaim_map);
3951}
3952
3953void __fs_reclaim_release(void)
3954{
3955 lock_map_release(&__fs_reclaim_map);
3956}
3957
3958void fs_reclaim_acquire(gfp_t gfp_mask)
3959{
3960 if (__need_fs_reclaim(gfp_mask))
3961 __fs_reclaim_acquire();
3962}
3963EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
3964
3965void fs_reclaim_release(gfp_t gfp_mask)
3966{
3967 if (__need_fs_reclaim(gfp_mask))
3968 __fs_reclaim_release();
3969}
3970EXPORT_SYMBOL_GPL(fs_reclaim_release);
3971#endif
3972
3973/* Perform direct synchronous page reclaim */
3974static int
3975__perform_reclaim(gfp_t gfp_mask, unsigned int order,
3976 const struct alloc_context *ac)
3977{
3978 struct reclaim_state reclaim_state;
3979 int progress;
3980 unsigned int noreclaim_flag;
3981 unsigned long pflags;
3982
3983 cond_resched();
3984
3985 /* We now go into synchronous reclaim */
3986 cpuset_memory_pressure_bump();
3987 psi_memstall_enter(&pflags);
3988 fs_reclaim_acquire(gfp_mask);
3989 noreclaim_flag = memalloc_noreclaim_save();
3990 reclaim_state.reclaimed_slab = 0;
3991 current->reclaim_state = &reclaim_state;
3992
3993 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
3994 ac->nodemask);
3995
3996 current->reclaim_state = NULL;
3997 memalloc_noreclaim_restore(noreclaim_flag);
3998 fs_reclaim_release(gfp_mask);
3999 psi_memstall_leave(&pflags);
4000
4001 cond_resched();
4002
4003 return progress;
4004}
4005
4006/* The really slow allocator path where we enter direct reclaim */
4007static inline struct page *
4008__alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4009 unsigned int alloc_flags, const struct alloc_context *ac,
4010 unsigned long *did_some_progress)
4011{
4012 struct page *page = NULL;
4013 bool drained = false;
4014
4015 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4016 if (unlikely(!(*did_some_progress)))
4017 return NULL;
4018
4019retry:
4020 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4021
4022 /*
4023 * If an allocation failed after direct reclaim, it could be because
4024 * pages are pinned on the per-cpu lists or in high alloc reserves.
4025 * Shrink them them and try again
4026 */
4027 if (!page && !drained) {
4028 unreserve_highatomic_pageblock(ac, false);
4029 drain_all_pages(NULL);
4030 drained = true;
4031 goto retry;
4032 }
4033
4034 return page;
4035}
4036
4037static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4038 const struct alloc_context *ac)
4039{
4040 struct zoneref *z;
4041 struct zone *zone;