1 // SPDX-License-Identifier: GPL-2.0-only
3 * linux/mm/page_alloc.c
5 * Manages the free list, the system allocates free pages here.
6 * Note that kmalloc() lives in slab.c
8 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
9 * Swap reorganised 29.12.95, Stephen Tweedie
10 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
11 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
12 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
13 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
14 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
15 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
18 #include <linux/stddef.h>
20 #include <linux/highmem.h>
21 #include <linux/swap.h>
22 #include <linux/interrupt.h>
23 #include <linux/pagemap.h>
24 #include <linux/jiffies.h>
25 #include <linux/memblock.h>
26 #include <linux/compiler.h>
27 #include <linux/kernel.h>
28 #include <linux/kasan.h>
29 #include <linux/module.h>
30 #include <linux/suspend.h>
31 #include <linux/pagevec.h>
32 #include <linux/blkdev.h>
33 #include <linux/slab.h>
34 #include <linux/ratelimit.h>
35 #include <linux/oom.h>
36 #include <linux/topology.h>
37 #include <linux/sysctl.h>
38 #include <linux/cpu.h>
39 #include <linux/cpuset.h>
40 #include <linux/memory_hotplug.h>
41 #include <linux/nodemask.h>
42 #include <linux/vmalloc.h>
43 #include <linux/vmstat.h>
44 #include <linux/mempolicy.h>
45 #include <linux/memremap.h>
46 #include <linux/stop_machine.h>
47 #include <linux/random.h>
48 #include <linux/sort.h>
49 #include <linux/pfn.h>
50 #include <linux/backing-dev.h>
51 #include <linux/fault-inject.h>
52 #include <linux/page-isolation.h>
53 #include <linux/debugobjects.h>
54 #include <linux/kmemleak.h>
55 #include <linux/compaction.h>
56 #include <trace/events/kmem.h>
57 #include <trace/events/oom.h>
58 #include <linux/prefetch.h>
59 #include <linux/mm_inline.h>
60 #include <linux/mmu_notifier.h>
61 #include <linux/migrate.h>
62 #include <linux/hugetlb.h>
63 #include <linux/sched/rt.h>
64 #include <linux/sched/mm.h>
65 #include <linux/page_owner.h>
66 #include <linux/kthread.h>
67 #include <linux/memcontrol.h>
68 #include <linux/ftrace.h>
69 #include <linux/lockdep.h>
70 #include <linux/nmi.h>
71 #include <linux/psi.h>
72 #include <linux/padata.h>
73 #include <linux/khugepaged.h>
74 #include <linux/buffer_head.h>
76 #include <asm/sections.h>
77 #include <asm/tlbflush.h>
78 #include <asm/div64.h>
81 #include "page_reporting.h"
83 /* Free Page Internal flags: for internal, non-pcp variants of free_pages(). */
84 typedef int __bitwise fpi_t
;
86 /* No special request */
87 #define FPI_NONE ((__force fpi_t)0)
90 * Skip free page reporting notification for the (possibly merged) page.
91 * This does not hinder free page reporting from grabbing the page,
92 * reporting it and marking it "reported" - it only skips notifying
93 * the free page reporting infrastructure about a newly freed page. For
94 * example, used when temporarily pulling a page from a freelist and
95 * putting it back unmodified.
97 #define FPI_SKIP_REPORT_NOTIFY ((__force fpi_t)BIT(0))
100 * Place the (possibly merged) page to the tail of the freelist. Will ignore
101 * page shuffling (relevant code - e.g., memory onlining - is expected to
102 * shuffle the whole zone).
104 * Note: No code should rely on this flag for correctness - it's purely
105 * to allow for optimizations when handing back either fresh pages
106 * (memory onlining) or untouched pages (page isolation, free page
109 #define FPI_TO_TAIL ((__force fpi_t)BIT(1))
111 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
112 static DEFINE_MUTEX(pcp_batch_high_lock
);
113 #define MIN_PERCPU_PAGELIST_FRACTION (8)
115 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
116 DEFINE_PER_CPU(int, numa_node
);
117 EXPORT_PER_CPU_SYMBOL(numa_node
);
120 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key
);
122 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
124 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
125 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
126 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
127 * defined in <linux/topology.h>.
129 DEFINE_PER_CPU(int, _numa_mem_
); /* Kernel "local memory" node */
130 EXPORT_PER_CPU_SYMBOL(_numa_mem_
);
133 /* work_structs for global per-cpu drains */
136 struct work_struct work
;
138 static DEFINE_MUTEX(pcpu_drain_mutex
);
139 static DEFINE_PER_CPU(struct pcpu_drain
, pcpu_drain
);
141 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
142 volatile unsigned long latent_entropy __latent_entropy
;
143 EXPORT_SYMBOL(latent_entropy
);
147 * Array of node states.
149 nodemask_t node_states
[NR_NODE_STATES
] __read_mostly
= {
150 [N_POSSIBLE
] = NODE_MASK_ALL
,
151 [N_ONLINE
] = { { [0] = 1UL } },
153 [N_NORMAL_MEMORY
] = { { [0] = 1UL } },
154 #ifdef CONFIG_HIGHMEM
155 [N_HIGH_MEMORY
] = { { [0] = 1UL } },
157 [N_MEMORY
] = { { [0] = 1UL } },
158 [N_CPU
] = { { [0] = 1UL } },
161 EXPORT_SYMBOL(node_states
);
163 atomic_long_t _totalram_pages __read_mostly
;
164 EXPORT_SYMBOL(_totalram_pages
);
165 unsigned long totalreserve_pages __read_mostly
;
166 unsigned long totalcma_pages __read_mostly
;
168 int percpu_pagelist_fraction
;
169 gfp_t gfp_allowed_mask __read_mostly
= GFP_BOOT_MASK
;
170 DEFINE_STATIC_KEY_FALSE(init_on_alloc
);
171 EXPORT_SYMBOL(init_on_alloc
);
173 DEFINE_STATIC_KEY_FALSE(init_on_free
);
174 EXPORT_SYMBOL(init_on_free
);
176 static bool _init_on_alloc_enabled_early __read_mostly
177 = IS_ENABLED(CONFIG_INIT_ON_ALLOC_DEFAULT_ON
);
178 static int __init
early_init_on_alloc(char *buf
)
181 return kstrtobool(buf
, &_init_on_alloc_enabled_early
);
183 early_param("init_on_alloc", early_init_on_alloc
);
185 static bool _init_on_free_enabled_early __read_mostly
186 = IS_ENABLED(CONFIG_INIT_ON_FREE_DEFAULT_ON
);
187 static int __init
early_init_on_free(char *buf
)
189 return kstrtobool(buf
, &_init_on_free_enabled_early
);
191 early_param("init_on_free", early_init_on_free
);
194 * A cached value of the page's pageblock's migratetype, used when the page is
195 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
196 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
197 * Also the migratetype set in the page does not necessarily match the pcplist
198 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
199 * other index - this ensures that it will be put on the correct CMA freelist.
201 static inline int get_pcppage_migratetype(struct page
*page
)
206 static inline void set_pcppage_migratetype(struct page
*page
, int migratetype
)
208 page
->index
= migratetype
;
211 #ifdef CONFIG_PM_SLEEP
213 * The following functions are used by the suspend/hibernate code to temporarily
214 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
215 * while devices are suspended. To avoid races with the suspend/hibernate code,
216 * they should always be called with system_transition_mutex held
217 * (gfp_allowed_mask also should only be modified with system_transition_mutex
218 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
219 * with that modification).
222 static gfp_t saved_gfp_mask
;
224 void pm_restore_gfp_mask(void)
226 WARN_ON(!mutex_is_locked(&system_transition_mutex
));
227 if (saved_gfp_mask
) {
228 gfp_allowed_mask
= saved_gfp_mask
;
233 void pm_restrict_gfp_mask(void)
235 WARN_ON(!mutex_is_locked(&system_transition_mutex
));
236 WARN_ON(saved_gfp_mask
);
237 saved_gfp_mask
= gfp_allowed_mask
;
238 gfp_allowed_mask
&= ~(__GFP_IO
| __GFP_FS
);
241 bool pm_suspended_storage(void)
243 if ((gfp_allowed_mask
& (__GFP_IO
| __GFP_FS
)) == (__GFP_IO
| __GFP_FS
))
247 #endif /* CONFIG_PM_SLEEP */
249 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
250 unsigned int pageblock_order __read_mostly
;
253 static void __free_pages_ok(struct page
*page
, unsigned int order
,
257 * results with 256, 32 in the lowmem_reserve sysctl:
258 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
259 * 1G machine -> (16M dma, 784M normal, 224M high)
260 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
261 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
262 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
264 * TBD: should special case ZONE_DMA32 machines here - in those we normally
265 * don't need any ZONE_NORMAL reservation
267 int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
] = {
268 #ifdef CONFIG_ZONE_DMA
271 #ifdef CONFIG_ZONE_DMA32
275 #ifdef CONFIG_HIGHMEM
281 static char * const zone_names
[MAX_NR_ZONES
] = {
282 #ifdef CONFIG_ZONE_DMA
285 #ifdef CONFIG_ZONE_DMA32
289 #ifdef CONFIG_HIGHMEM
293 #ifdef CONFIG_ZONE_DEVICE
298 const char * const migratetype_names
[MIGRATE_TYPES
] = {
306 #ifdef CONFIG_MEMORY_ISOLATION
311 compound_page_dtor
* const compound_page_dtors
[NR_COMPOUND_DTORS
] = {
312 [NULL_COMPOUND_DTOR
] = NULL
,
313 [COMPOUND_PAGE_DTOR
] = free_compound_page
,
314 #ifdef CONFIG_HUGETLB_PAGE
315 [HUGETLB_PAGE_DTOR
] = free_huge_page
,
317 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
318 [TRANSHUGE_PAGE_DTOR
] = free_transhuge_page
,
322 int min_free_kbytes
= 1024;
323 int user_min_free_kbytes
= -1;
324 #ifdef CONFIG_DISCONTIGMEM
326 * DiscontigMem defines memory ranges as separate pg_data_t even if the ranges
327 * are not on separate NUMA nodes. Functionally this works but with
328 * watermark_boost_factor, it can reclaim prematurely as the ranges can be
329 * quite small. By default, do not boost watermarks on discontigmem as in
330 * many cases very high-order allocations like THP are likely to be
331 * unsupported and the premature reclaim offsets the advantage of long-term
332 * fragmentation avoidance.
334 int watermark_boost_factor __read_mostly
;
336 int watermark_boost_factor __read_mostly
= 15000;
338 int watermark_scale_factor
= 10;
340 static unsigned long nr_kernel_pages __initdata
;
341 static unsigned long nr_all_pages __initdata
;
342 static unsigned long dma_reserve __initdata
;
344 static unsigned long arch_zone_lowest_possible_pfn
[MAX_NR_ZONES
] __initdata
;
345 static unsigned long arch_zone_highest_possible_pfn
[MAX_NR_ZONES
] __initdata
;
346 static unsigned long required_kernelcore __initdata
;
347 static unsigned long required_kernelcore_percent __initdata
;
348 static unsigned long required_movablecore __initdata
;
349 static unsigned long required_movablecore_percent __initdata
;
350 static unsigned long zone_movable_pfn
[MAX_NUMNODES
] __initdata
;
351 static bool mirrored_kernelcore __meminitdata
;
353 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
355 EXPORT_SYMBOL(movable_zone
);
358 unsigned int nr_node_ids __read_mostly
= MAX_NUMNODES
;
359 unsigned int nr_online_nodes __read_mostly
= 1;
360 EXPORT_SYMBOL(nr_node_ids
);
361 EXPORT_SYMBOL(nr_online_nodes
);
364 int page_group_by_mobility_disabled __read_mostly
;
366 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
368 * During boot we initialize deferred pages on-demand, as needed, but once
369 * page_alloc_init_late() has finished, the deferred pages are all initialized,
370 * and we can permanently disable that path.
372 static DEFINE_STATIC_KEY_TRUE(deferred_pages
);
375 * Calling kasan_free_pages() only after deferred memory initialization
376 * has completed. Poisoning pages during deferred memory init will greatly
377 * lengthen the process and cause problem in large memory systems as the
378 * deferred pages initialization is done with interrupt disabled.
380 * Assuming that there will be no reference to those newly initialized
381 * pages before they are ever allocated, this should have no effect on
382 * KASAN memory tracking as the poison will be properly inserted at page
383 * allocation time. The only corner case is when pages are allocated by
384 * on-demand allocation and then freed again before the deferred pages
385 * initialization is done, but this is not likely to happen.
387 static inline void kasan_free_nondeferred_pages(struct page
*page
, int order
)
389 if (!static_branch_unlikely(&deferred_pages
))
390 kasan_free_pages(page
, order
);
393 /* Returns true if the struct page for the pfn is uninitialised */
394 static inline bool __meminit
early_page_uninitialised(unsigned long pfn
)
396 int nid
= early_pfn_to_nid(pfn
);
398 if (node_online(nid
) && pfn
>= NODE_DATA(nid
)->first_deferred_pfn
)
405 * Returns true when the remaining initialisation should be deferred until
406 * later in the boot cycle when it can be parallelised.
408 static bool __meminit
409 defer_init(int nid
, unsigned long pfn
, unsigned long end_pfn
)
411 static unsigned long prev_end_pfn
, nr_initialised
;
414 * prev_end_pfn static that contains the end of previous zone
415 * No need to protect because called very early in boot before smp_init.
417 if (prev_end_pfn
!= end_pfn
) {
418 prev_end_pfn
= end_pfn
;
422 /* Always populate low zones for address-constrained allocations */
423 if (end_pfn
< pgdat_end_pfn(NODE_DATA(nid
)))
426 if (NODE_DATA(nid
)->first_deferred_pfn
!= ULONG_MAX
)
429 * We start only with one section of pages, more pages are added as
430 * needed until the rest of deferred pages are initialized.
433 if ((nr_initialised
> PAGES_PER_SECTION
) &&
434 (pfn
& (PAGES_PER_SECTION
- 1)) == 0) {
435 NODE_DATA(nid
)->first_deferred_pfn
= pfn
;
441 #define kasan_free_nondeferred_pages(p, o) kasan_free_pages(p, o)
443 static inline bool early_page_uninitialised(unsigned long pfn
)
448 static inline bool defer_init(int nid
, unsigned long pfn
, unsigned long end_pfn
)
454 /* Return a pointer to the bitmap storing bits affecting a block of pages */
455 static inline unsigned long *get_pageblock_bitmap(struct page
*page
,
458 #ifdef CONFIG_SPARSEMEM
459 return section_to_usemap(__pfn_to_section(pfn
));
461 return page_zone(page
)->pageblock_flags
;
462 #endif /* CONFIG_SPARSEMEM */
465 static inline int pfn_to_bitidx(struct page
*page
, unsigned long pfn
)
467 #ifdef CONFIG_SPARSEMEM
468 pfn
&= (PAGES_PER_SECTION
-1);
470 pfn
= pfn
- round_down(page_zone(page
)->zone_start_pfn
, pageblock_nr_pages
);
471 #endif /* CONFIG_SPARSEMEM */
472 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
475 static __always_inline
476 unsigned long __get_pfnblock_flags_mask(struct page
*page
,
480 unsigned long *bitmap
;
481 unsigned long bitidx
, word_bitidx
;
484 bitmap
= get_pageblock_bitmap(page
, pfn
);
485 bitidx
= pfn_to_bitidx(page
, pfn
);
486 word_bitidx
= bitidx
/ BITS_PER_LONG
;
487 bitidx
&= (BITS_PER_LONG
-1);
489 word
= bitmap
[word_bitidx
];
490 return (word
>> bitidx
) & mask
;
494 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
495 * @page: The page within the block of interest
496 * @pfn: The target page frame number
497 * @mask: mask of bits that the caller is interested in
499 * Return: pageblock_bits flags
501 unsigned long get_pfnblock_flags_mask(struct page
*page
, unsigned long pfn
,
504 return __get_pfnblock_flags_mask(page
, pfn
, mask
);
507 static __always_inline
int get_pfnblock_migratetype(struct page
*page
, unsigned long pfn
)
509 return __get_pfnblock_flags_mask(page
, pfn
, MIGRATETYPE_MASK
);
513 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
514 * @page: The page within the block of interest
515 * @flags: The flags to set
516 * @pfn: The target page frame number
517 * @mask: mask of bits that the caller is interested in
519 void set_pfnblock_flags_mask(struct page
*page
, unsigned long flags
,
523 unsigned long *bitmap
;
524 unsigned long bitidx
, word_bitidx
;
525 unsigned long old_word
, word
;
527 BUILD_BUG_ON(NR_PAGEBLOCK_BITS
!= 4);
528 BUILD_BUG_ON(MIGRATE_TYPES
> (1 << PB_migratetype_bits
));
530 bitmap
= get_pageblock_bitmap(page
, pfn
);
531 bitidx
= pfn_to_bitidx(page
, pfn
);
532 word_bitidx
= bitidx
/ BITS_PER_LONG
;
533 bitidx
&= (BITS_PER_LONG
-1);
535 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page
), pfn
), page
);
540 word
= READ_ONCE(bitmap
[word_bitidx
]);
542 old_word
= cmpxchg(&bitmap
[word_bitidx
], word
, (word
& ~mask
) | flags
);
543 if (word
== old_word
)
549 void set_pageblock_migratetype(struct page
*page
, int migratetype
)
551 if (unlikely(page_group_by_mobility_disabled
&&
552 migratetype
< MIGRATE_PCPTYPES
))
553 migratetype
= MIGRATE_UNMOVABLE
;
555 set_pfnblock_flags_mask(page
, (unsigned long)migratetype
,
556 page_to_pfn(page
), MIGRATETYPE_MASK
);
559 #ifdef CONFIG_DEBUG_VM
560 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
564 unsigned long pfn
= page_to_pfn(page
);
565 unsigned long sp
, start_pfn
;
568 seq
= zone_span_seqbegin(zone
);
569 start_pfn
= zone
->zone_start_pfn
;
570 sp
= zone
->spanned_pages
;
571 if (!zone_spans_pfn(zone
, pfn
))
573 } while (zone_span_seqretry(zone
, seq
));
576 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
577 pfn
, zone_to_nid(zone
), zone
->name
,
578 start_pfn
, start_pfn
+ sp
);
583 static int page_is_consistent(struct zone
*zone
, struct page
*page
)
585 if (!pfn_valid_within(page_to_pfn(page
)))
587 if (zone
!= page_zone(page
))
593 * Temporary debugging check for pages not lying within a given zone.
595 static int __maybe_unused
bad_range(struct zone
*zone
, struct page
*page
)
597 if (page_outside_zone_boundaries(zone
, page
))
599 if (!page_is_consistent(zone
, page
))
605 static inline int __maybe_unused
bad_range(struct zone
*zone
, struct page
*page
)
611 static void bad_page(struct page
*page
, const char *reason
)
613 static unsigned long resume
;
614 static unsigned long nr_shown
;
615 static unsigned long nr_unshown
;
618 * Allow a burst of 60 reports, then keep quiet for that minute;
619 * or allow a steady drip of one report per second.
621 if (nr_shown
== 60) {
622 if (time_before(jiffies
, resume
)) {
628 "BUG: Bad page state: %lu messages suppressed\n",
635 resume
= jiffies
+ 60 * HZ
;
637 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
638 current
->comm
, page_to_pfn(page
));
639 __dump_page(page
, reason
);
640 dump_page_owner(page
);
645 /* Leave bad fields for debug, except PageBuddy could make trouble */
646 page_mapcount_reset(page
); /* remove PageBuddy */
647 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
651 * Higher-order pages are called "compound pages". They are structured thusly:
653 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
655 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
656 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
658 * The first tail page's ->compound_dtor holds the offset in array of compound
659 * page destructors. See compound_page_dtors.
661 * The first tail page's ->compound_order holds the order of allocation.
662 * This usage means that zero-order pages may not be compound.
665 void free_compound_page(struct page
*page
)
667 mem_cgroup_uncharge(page
);
668 __free_pages_ok(page
, compound_order(page
), FPI_NONE
);
671 void prep_compound_page(struct page
*page
, unsigned int order
)
674 int nr_pages
= 1 << order
;
677 for (i
= 1; i
< nr_pages
; i
++) {
678 struct page
*p
= page
+ i
;
679 set_page_count(p
, 0);
680 p
->mapping
= TAIL_MAPPING
;
681 set_compound_head(p
, page
);
684 set_compound_page_dtor(page
, COMPOUND_PAGE_DTOR
);
685 set_compound_order(page
, order
);
686 atomic_set(compound_mapcount_ptr(page
), -1);
687 if (hpage_pincount_available(page
))
688 atomic_set(compound_pincount_ptr(page
), 0);
691 #ifdef CONFIG_DEBUG_PAGEALLOC
692 unsigned int _debug_guardpage_minorder
;
694 bool _debug_pagealloc_enabled_early __read_mostly
695 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT
);
696 EXPORT_SYMBOL(_debug_pagealloc_enabled_early
);
697 DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled
);
698 EXPORT_SYMBOL(_debug_pagealloc_enabled
);
700 DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled
);
702 static int __init
early_debug_pagealloc(char *buf
)
704 return kstrtobool(buf
, &_debug_pagealloc_enabled_early
);
706 early_param("debug_pagealloc", early_debug_pagealloc
);
708 static int __init
debug_guardpage_minorder_setup(char *buf
)
712 if (kstrtoul(buf
, 10, &res
) < 0 || res
> MAX_ORDER
/ 2) {
713 pr_err("Bad debug_guardpage_minorder value\n");
716 _debug_guardpage_minorder
= res
;
717 pr_info("Setting debug_guardpage_minorder to %lu\n", res
);
720 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup
);
722 static inline bool set_page_guard(struct zone
*zone
, struct page
*page
,
723 unsigned int order
, int migratetype
)
725 if (!debug_guardpage_enabled())
728 if (order
>= debug_guardpage_minorder())
731 __SetPageGuard(page
);
732 INIT_LIST_HEAD(&page
->lru
);
733 set_page_private(page
, order
);
734 /* Guard pages are not available for any usage */
735 __mod_zone_freepage_state(zone
, -(1 << order
), migratetype
);
740 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
741 unsigned int order
, int migratetype
)
743 if (!debug_guardpage_enabled())
746 __ClearPageGuard(page
);
748 set_page_private(page
, 0);
749 if (!is_migrate_isolate(migratetype
))
750 __mod_zone_freepage_state(zone
, (1 << order
), migratetype
);
753 static inline bool set_page_guard(struct zone
*zone
, struct page
*page
,
754 unsigned int order
, int migratetype
) { return false; }
755 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
756 unsigned int order
, int migratetype
) {}
760 * Enable static keys related to various memory debugging and hardening options.
761 * Some override others, and depend on early params that are evaluated in the
762 * order of appearance. So we need to first gather the full picture of what was
763 * enabled, and then make decisions.
765 void init_mem_debugging_and_hardening(void)
767 if (_init_on_alloc_enabled_early
) {
768 if (page_poisoning_enabled())
769 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, "
770 "will take precedence over init_on_alloc\n");
772 static_branch_enable(&init_on_alloc
);
774 if (_init_on_free_enabled_early
) {
775 if (page_poisoning_enabled())
776 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, "
777 "will take precedence over init_on_free\n");
779 static_branch_enable(&init_on_free
);
782 #ifdef CONFIG_PAGE_POISONING
784 * Page poisoning is debug page alloc for some arches. If
785 * either of those options are enabled, enable poisoning.
787 if (page_poisoning_enabled() ||
788 (!IS_ENABLED(CONFIG_ARCH_SUPPORTS_DEBUG_PAGEALLOC
) &&
789 debug_pagealloc_enabled()))
790 static_branch_enable(&_page_poisoning_enabled
);
793 #ifdef CONFIG_DEBUG_PAGEALLOC
794 if (!debug_pagealloc_enabled())
797 static_branch_enable(&_debug_pagealloc_enabled
);
799 if (!debug_guardpage_minorder())
802 static_branch_enable(&_debug_guardpage_enabled
);
806 static inline void set_buddy_order(struct page
*page
, unsigned int order
)
808 set_page_private(page
, order
);
809 __SetPageBuddy(page
);
813 * This function checks whether a page is free && is the buddy
814 * we can coalesce a page and its buddy if
815 * (a) the buddy is not in a hole (check before calling!) &&
816 * (b) the buddy is in the buddy system &&
817 * (c) a page and its buddy have the same order &&
818 * (d) a page and its buddy are in the same zone.
820 * For recording whether a page is in the buddy system, we set PageBuddy.
821 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
823 * For recording page's order, we use page_private(page).
825 static inline bool page_is_buddy(struct page
*page
, struct page
*buddy
,
828 if (!page_is_guard(buddy
) && !PageBuddy(buddy
))
831 if (buddy_order(buddy
) != order
)
835 * zone check is done late to avoid uselessly calculating
836 * zone/node ids for pages that could never merge.
838 if (page_zone_id(page
) != page_zone_id(buddy
))
841 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
846 #ifdef CONFIG_COMPACTION
847 static inline struct capture_control
*task_capc(struct zone
*zone
)
849 struct capture_control
*capc
= current
->capture_control
;
851 return unlikely(capc
) &&
852 !(current
->flags
& PF_KTHREAD
) &&
854 capc
->cc
->zone
== zone
? capc
: NULL
;
858 compaction_capture(struct capture_control
*capc
, struct page
*page
,
859 int order
, int migratetype
)
861 if (!capc
|| order
!= capc
->cc
->order
)
864 /* Do not accidentally pollute CMA or isolated regions*/
865 if (is_migrate_cma(migratetype
) ||
866 is_migrate_isolate(migratetype
))
870 * Do not let lower order allocations polluate a movable pageblock.
871 * This might let an unmovable request use a reclaimable pageblock
872 * and vice-versa but no more than normal fallback logic which can
873 * have trouble finding a high-order free page.
875 if (order
< pageblock_order
&& migratetype
== MIGRATE_MOVABLE
)
883 static inline struct capture_control
*task_capc(struct zone
*zone
)
889 compaction_capture(struct capture_control
*capc
, struct page
*page
,
890 int order
, int migratetype
)
894 #endif /* CONFIG_COMPACTION */
896 /* Used for pages not on another list */
897 static inline void add_to_free_list(struct page
*page
, struct zone
*zone
,
898 unsigned int order
, int migratetype
)
900 struct free_area
*area
= &zone
->free_area
[order
];
902 list_add(&page
->lru
, &area
->free_list
[migratetype
]);
906 /* Used for pages not on another list */
907 static inline void add_to_free_list_tail(struct page
*page
, struct zone
*zone
,
908 unsigned int order
, int migratetype
)
910 struct free_area
*area
= &zone
->free_area
[order
];
912 list_add_tail(&page
->lru
, &area
->free_list
[migratetype
]);
917 * Used for pages which are on another list. Move the pages to the tail
918 * of the list - so the moved pages won't immediately be considered for
919 * allocation again (e.g., optimization for memory onlining).
921 static inline void move_to_free_list(struct page
*page
, struct zone
*zone
,
922 unsigned int order
, int migratetype
)
924 struct free_area
*area
= &zone
->free_area
[order
];
926 list_move_tail(&page
->lru
, &area
->free_list
[migratetype
]);
929 static inline void del_page_from_free_list(struct page
*page
, struct zone
*zone
,
932 /* clear reported state and update reported page count */
933 if (page_reported(page
))
934 __ClearPageReported(page
);
936 list_del(&page
->lru
);
937 __ClearPageBuddy(page
);
938 set_page_private(page
, 0);
939 zone
->free_area
[order
].nr_free
--;
943 * If this is not the largest possible page, check if the buddy
944 * of the next-highest order is free. If it is, it's possible
945 * that pages are being freed that will coalesce soon. In case,
946 * that is happening, add the free page to the tail of the list
947 * so it's less likely to be used soon and more likely to be merged
948 * as a higher order page
951 buddy_merge_likely(unsigned long pfn
, unsigned long buddy_pfn
,
952 struct page
*page
, unsigned int order
)
954 struct page
*higher_page
, *higher_buddy
;
955 unsigned long combined_pfn
;
957 if (order
>= MAX_ORDER
- 2)
960 if (!pfn_valid_within(buddy_pfn
))
963 combined_pfn
= buddy_pfn
& pfn
;
964 higher_page
= page
+ (combined_pfn
- pfn
);
965 buddy_pfn
= __find_buddy_pfn(combined_pfn
, order
+ 1);
966 higher_buddy
= higher_page
+ (buddy_pfn
- combined_pfn
);
968 return pfn_valid_within(buddy_pfn
) &&
969 page_is_buddy(higher_page
, higher_buddy
, order
+ 1);
973 * Freeing function for a buddy system allocator.
975 * The concept of a buddy system is to maintain direct-mapped table
976 * (containing bit values) for memory blocks of various "orders".
977 * The bottom level table contains the map for the smallest allocatable
978 * units of memory (here, pages), and each level above it describes
979 * pairs of units from the levels below, hence, "buddies".
980 * At a high level, all that happens here is marking the table entry
981 * at the bottom level available, and propagating the changes upward
982 * as necessary, plus some accounting needed to play nicely with other
983 * parts of the VM system.
984 * At each level, we keep a list of pages, which are heads of continuous
985 * free pages of length of (1 << order) and marked with PageBuddy.
986 * Page's order is recorded in page_private(page) field.
987 * So when we are allocating or freeing one, we can derive the state of the
988 * other. That is, if we allocate a small block, and both were
989 * free, the remainder of the region must be split into blocks.
990 * If a block is freed, and its buddy is also free, then this
991 * triggers coalescing into a block of larger size.
996 static inline void __free_one_page(struct page
*page
,
998 struct zone
*zone
, unsigned int order
,
999 int migratetype
, fpi_t fpi_flags
)
1001 struct capture_control
*capc
= task_capc(zone
);
1002 unsigned long buddy_pfn
;
1003 unsigned long combined_pfn
;
1004 unsigned int max_order
;
1008 max_order
= min_t(unsigned int, MAX_ORDER
- 1, pageblock_order
);
1010 VM_BUG_ON(!zone_is_initialized(zone
));
1011 VM_BUG_ON_PAGE(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
, page
);
1013 VM_BUG_ON(migratetype
== -1);
1014 if (likely(!is_migrate_isolate(migratetype
)))
1015 __mod_zone_freepage_state(zone
, 1 << order
, migratetype
);
1017 VM_BUG_ON_PAGE(pfn
& ((1 << order
) - 1), page
);
1018 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
1021 while (order
< max_order
) {
1022 if (compaction_capture(capc
, page
, order
, migratetype
)) {
1023 __mod_zone_freepage_state(zone
, -(1 << order
),
1027 buddy_pfn
= __find_buddy_pfn(pfn
, order
);
1028 buddy
= page
+ (buddy_pfn
- pfn
);
1030 if (!pfn_valid_within(buddy_pfn
))
1032 if (!page_is_buddy(page
, buddy
, order
))
1035 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
1036 * merge with it and move up one order.
1038 if (page_is_guard(buddy
))
1039 clear_page_guard(zone
, buddy
, order
, migratetype
);
1041 del_page_from_free_list(buddy
, zone
, order
);
1042 combined_pfn
= buddy_pfn
& pfn
;
1043 page
= page
+ (combined_pfn
- pfn
);
1047 if (order
< MAX_ORDER
- 1) {
1048 /* If we are here, it means order is >= pageblock_order.
1049 * We want to prevent merge between freepages on isolate
1050 * pageblock and normal pageblock. Without this, pageblock
1051 * isolation could cause incorrect freepage or CMA accounting.
1053 * We don't want to hit this code for the more frequent
1054 * low-order merging.
1056 if (unlikely(has_isolate_pageblock(zone
))) {
1059 buddy_pfn
= __find_buddy_pfn(pfn
, order
);
1060 buddy
= page
+ (buddy_pfn
- pfn
);
1061 buddy_mt
= get_pageblock_migratetype(buddy
);
1063 if (migratetype
!= buddy_mt
1064 && (is_migrate_isolate(migratetype
) ||
1065 is_migrate_isolate(buddy_mt
)))
1068 max_order
= order
+ 1;
1069 goto continue_merging
;
1073 set_buddy_order(page
, order
);
1075 if (fpi_flags
& FPI_TO_TAIL
)
1077 else if (is_shuffle_order(order
))
1078 to_tail
= shuffle_pick_tail();
1080 to_tail
= buddy_merge_likely(pfn
, buddy_pfn
, page
, order
);
1083 add_to_free_list_tail(page
, zone
, order
, migratetype
);
1085 add_to_free_list(page
, zone
, order
, migratetype
);
1087 /* Notify page reporting subsystem of freed page */
1088 if (!(fpi_flags
& FPI_SKIP_REPORT_NOTIFY
))
1089 page_reporting_notify_free(order
);
1093 * A bad page could be due to a number of fields. Instead of multiple branches,
1094 * try and check multiple fields with one check. The caller must do a detailed
1095 * check if necessary.
1097 static inline bool page_expected_state(struct page
*page
,
1098 unsigned long check_flags
)
1100 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
1103 if (unlikely((unsigned long)page
->mapping
|
1104 page_ref_count(page
) |
1106 (unsigned long)page_memcg(page
) |
1108 (page
->flags
& check_flags
)))
1114 static const char *page_bad_reason(struct page
*page
, unsigned long flags
)
1116 const char *bad_reason
= NULL
;
1118 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
1119 bad_reason
= "nonzero mapcount";
1120 if (unlikely(page
->mapping
!= NULL
))
1121 bad_reason
= "non-NULL mapping";
1122 if (unlikely(page_ref_count(page
) != 0))
1123 bad_reason
= "nonzero _refcount";
1124 if (unlikely(page
->flags
& flags
)) {
1125 if (flags
== PAGE_FLAGS_CHECK_AT_PREP
)
1126 bad_reason
= "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
1128 bad_reason
= "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1131 if (unlikely(page_memcg(page
)))
1132 bad_reason
= "page still charged to cgroup";
1137 static void check_free_page_bad(struct page
*page
)
1140 page_bad_reason(page
, PAGE_FLAGS_CHECK_AT_FREE
));
1143 static inline int check_free_page(struct page
*page
)
1145 if (likely(page_expected_state(page
, PAGE_FLAGS_CHECK_AT_FREE
)))
1148 /* Something has gone sideways, find it */
1149 check_free_page_bad(page
);
1153 static int free_tail_pages_check(struct page
*head_page
, struct page
*page
)
1158 * We rely page->lru.next never has bit 0 set, unless the page
1159 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1161 BUILD_BUG_ON((unsigned long)LIST_POISON1
& 1);
1163 if (!IS_ENABLED(CONFIG_DEBUG_VM
)) {
1167 switch (page
- head_page
) {
1169 /* the first tail page: ->mapping may be compound_mapcount() */
1170 if (unlikely(compound_mapcount(page
))) {
1171 bad_page(page
, "nonzero compound_mapcount");
1177 * the second tail page: ->mapping is
1178 * deferred_list.next -- ignore value.
1182 if (page
->mapping
!= TAIL_MAPPING
) {
1183 bad_page(page
, "corrupted mapping in tail page");
1188 if (unlikely(!PageTail(page
))) {
1189 bad_page(page
, "PageTail not set");
1192 if (unlikely(compound_head(page
) != head_page
)) {
1193 bad_page(page
, "compound_head not consistent");
1198 page
->mapping
= NULL
;
1199 clear_compound_head(page
);
1203 static void kernel_init_free_pages(struct page
*page
, int numpages
)
1207 /* s390's use of memset() could override KASAN redzones. */
1208 kasan_disable_current();
1209 for (i
= 0; i
< numpages
; i
++) {
1210 u8 tag
= page_kasan_tag(page
+ i
);
1211 page_kasan_tag_reset(page
+ i
);
1212 clear_highpage(page
+ i
);
1213 page_kasan_tag_set(page
+ i
, tag
);
1215 kasan_enable_current();
1218 static __always_inline
bool free_pages_prepare(struct page
*page
,
1219 unsigned int order
, bool check_free
)
1223 VM_BUG_ON_PAGE(PageTail(page
), page
);
1225 trace_mm_page_free(page
, order
);
1227 if (unlikely(PageHWPoison(page
)) && !order
) {
1229 * Do not let hwpoison pages hit pcplists/buddy
1230 * Untie memcg state and reset page's owner
1232 if (memcg_kmem_enabled() && PageMemcgKmem(page
))
1233 __memcg_kmem_uncharge_page(page
, order
);
1234 reset_page_owner(page
, order
);
1239 * Check tail pages before head page information is cleared to
1240 * avoid checking PageCompound for order-0 pages.
1242 if (unlikely(order
)) {
1243 bool compound
= PageCompound(page
);
1246 VM_BUG_ON_PAGE(compound
&& compound_order(page
) != order
, page
);
1249 ClearPageDoubleMap(page
);
1250 for (i
= 1; i
< (1 << order
); i
++) {
1252 bad
+= free_tail_pages_check(page
, page
+ i
);
1253 if (unlikely(check_free_page(page
+ i
))) {
1257 (page
+ i
)->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
1260 if (PageMappingFlags(page
))
1261 page
->mapping
= NULL
;
1262 if (memcg_kmem_enabled() && PageMemcgKmem(page
))
1263 __memcg_kmem_uncharge_page(page
, order
);
1265 bad
+= check_free_page(page
);
1269 page_cpupid_reset_last(page
);
1270 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
1271 reset_page_owner(page
, order
);
1273 if (!PageHighMem(page
)) {
1274 debug_check_no_locks_freed(page_address(page
),
1275 PAGE_SIZE
<< order
);
1276 debug_check_no_obj_freed(page_address(page
),
1277 PAGE_SIZE
<< order
);
1279 if (want_init_on_free())
1280 kernel_init_free_pages(page
, 1 << order
);
1282 kernel_poison_pages(page
, 1 << order
);
1285 * With hardware tag-based KASAN, memory tags must be set before the
1286 * page becomes unavailable via debug_pagealloc or arch_free_page.
1288 kasan_free_nondeferred_pages(page
, order
);
1291 * arch_free_page() can make the page's contents inaccessible. s390
1292 * does this. So nothing which can access the page's contents should
1293 * happen after this.
1295 arch_free_page(page
, order
);
1297 debug_pagealloc_unmap_pages(page
, 1 << order
);
1302 #ifdef CONFIG_DEBUG_VM
1304 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1305 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1306 * moved from pcp lists to free lists.
1308 static bool free_pcp_prepare(struct page
*page
)
1310 return free_pages_prepare(page
, 0, true);
1313 static bool bulkfree_pcp_prepare(struct page
*page
)
1315 if (debug_pagealloc_enabled_static())
1316 return check_free_page(page
);
1322 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1323 * moving from pcp lists to free list in order to reduce overhead. With
1324 * debug_pagealloc enabled, they are checked also immediately when being freed
1327 static bool free_pcp_prepare(struct page
*page
)
1329 if (debug_pagealloc_enabled_static())
1330 return free_pages_prepare(page
, 0, true);
1332 return free_pages_prepare(page
, 0, false);
1335 static bool bulkfree_pcp_prepare(struct page
*page
)
1337 return check_free_page(page
);
1339 #endif /* CONFIG_DEBUG_VM */
1341 static inline void prefetch_buddy(struct page
*page
)
1343 unsigned long pfn
= page_to_pfn(page
);
1344 unsigned long buddy_pfn
= __find_buddy_pfn(pfn
, 0);
1345 struct page
*buddy
= page
+ (buddy_pfn
- pfn
);
1351 * Frees a number of pages from the PCP lists
1352 * Assumes all pages on list are in same zone, and of same order.
1353 * count is the number of pages to free.
1355 * If the zone was previously in an "all pages pinned" state then look to
1356 * see if this freeing clears that state.
1358 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1359 * pinned" detection logic.
1361 static void free_pcppages_bulk(struct zone
*zone
, int count
,
1362 struct per_cpu_pages
*pcp
)
1364 int migratetype
= 0;
1366 int prefetch_nr
= READ_ONCE(pcp
->batch
);
1367 bool isolated_pageblocks
;
1368 struct page
*page
, *tmp
;
1372 * Ensure proper count is passed which otherwise would stuck in the
1373 * below while (list_empty(list)) loop.
1375 count
= min(pcp
->count
, count
);
1377 struct list_head
*list
;
1380 * Remove pages from lists in a round-robin fashion. A
1381 * batch_free count is maintained that is incremented when an
1382 * empty list is encountered. This is so more pages are freed
1383 * off fuller lists instead of spinning excessively around empty
1388 if (++migratetype
== MIGRATE_PCPTYPES
)
1390 list
= &pcp
->lists
[migratetype
];
1391 } while (list_empty(list
));
1393 /* This is the only non-empty list. Free them all. */
1394 if (batch_free
== MIGRATE_PCPTYPES
)
1398 page
= list_last_entry(list
, struct page
, lru
);
1399 /* must delete to avoid corrupting pcp list */
1400 list_del(&page
->lru
);
1403 if (bulkfree_pcp_prepare(page
))
1406 list_add_tail(&page
->lru
, &head
);
1409 * We are going to put the page back to the global
1410 * pool, prefetch its buddy to speed up later access
1411 * under zone->lock. It is believed the overhead of
1412 * an additional test and calculating buddy_pfn here
1413 * can be offset by reduced memory latency later. To
1414 * avoid excessive prefetching due to large count, only
1415 * prefetch buddy for the first pcp->batch nr of pages.
1418 prefetch_buddy(page
);
1421 } while (--count
&& --batch_free
&& !list_empty(list
));
1424 spin_lock(&zone
->lock
);
1425 isolated_pageblocks
= has_isolate_pageblock(zone
);
1428 * Use safe version since after __free_one_page(),
1429 * page->lru.next will not point to original list.
1431 list_for_each_entry_safe(page
, tmp
, &head
, lru
) {
1432 int mt
= get_pcppage_migratetype(page
);
1433 /* MIGRATE_ISOLATE page should not go to pcplists */
1434 VM_BUG_ON_PAGE(is_migrate_isolate(mt
), page
);
1435 /* Pageblock could have been isolated meanwhile */
1436 if (unlikely(isolated_pageblocks
))
1437 mt
= get_pageblock_migratetype(page
);
1439 __free_one_page(page
, page_to_pfn(page
), zone
, 0, mt
, FPI_NONE
);
1440 trace_mm_page_pcpu_drain(page
, 0, mt
);
1442 spin_unlock(&zone
->lock
);
1445 static void free_one_page(struct zone
*zone
,
1446 struct page
*page
, unsigned long pfn
,
1448 int migratetype
, fpi_t fpi_flags
)
1450 spin_lock(&zone
->lock
);
1451 if (unlikely(has_isolate_pageblock(zone
) ||
1452 is_migrate_isolate(migratetype
))) {
1453 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1455 __free_one_page(page
, pfn
, zone
, order
, migratetype
, fpi_flags
);
1456 spin_unlock(&zone
->lock
);
1459 static void __meminit
__init_single_page(struct page
*page
, unsigned long pfn
,
1460 unsigned long zone
, int nid
)
1462 mm_zero_struct_page(page
);
1463 set_page_links(page
, zone
, nid
, pfn
);
1464 init_page_count(page
);
1465 page_mapcount_reset(page
);
1466 page_cpupid_reset_last(page
);
1467 page_kasan_tag_reset(page
);
1469 INIT_LIST_HEAD(&page
->lru
);
1470 #ifdef WANT_PAGE_VIRTUAL
1471 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1472 if (!is_highmem_idx(zone
))
1473 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
1477 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1478 static void __meminit
init_reserved_page(unsigned long pfn
)
1483 if (!early_page_uninitialised(pfn
))
1486 nid
= early_pfn_to_nid(pfn
);
1487 pgdat
= NODE_DATA(nid
);
1489 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1490 struct zone
*zone
= &pgdat
->node_zones
[zid
];
1492 if (pfn
>= zone
->zone_start_pfn
&& pfn
< zone_end_pfn(zone
))
1495 __init_single_page(pfn_to_page(pfn
), pfn
, zid
, nid
);
1498 static inline void init_reserved_page(unsigned long pfn
)
1501 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1504 * Initialised pages do not have PageReserved set. This function is
1505 * called for each range allocated by the bootmem allocator and
1506 * marks the pages PageReserved. The remaining valid pages are later
1507 * sent to the buddy page allocator.
1509 void __meminit
reserve_bootmem_region(phys_addr_t start
, phys_addr_t end
)
1511 unsigned long start_pfn
= PFN_DOWN(start
);
1512 unsigned long end_pfn
= PFN_UP(end
);
1514 for (; start_pfn
< end_pfn
; start_pfn
++) {
1515 if (pfn_valid(start_pfn
)) {
1516 struct page
*page
= pfn_to_page(start_pfn
);
1518 init_reserved_page(start_pfn
);
1520 /* Avoid false-positive PageTail() */
1521 INIT_LIST_HEAD(&page
->lru
);
1524 * no need for atomic set_bit because the struct
1525 * page is not visible yet so nobody should
1528 __SetPageReserved(page
);
1533 static void __free_pages_ok(struct page
*page
, unsigned int order
,
1536 unsigned long flags
;
1538 unsigned long pfn
= page_to_pfn(page
);
1540 if (!free_pages_prepare(page
, order
, true))
1543 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1544 local_irq_save(flags
);
1545 __count_vm_events(PGFREE
, 1 << order
);
1546 free_one_page(page_zone(page
), page
, pfn
, order
, migratetype
,
1548 local_irq_restore(flags
);
1551 void __free_pages_core(struct page
*page
, unsigned int order
)
1553 unsigned int nr_pages
= 1 << order
;
1554 struct page
*p
= page
;
1558 * When initializing the memmap, __init_single_page() sets the refcount
1559 * of all pages to 1 ("allocated"/"not free"). We have to set the
1560 * refcount of all involved pages to 0.
1563 for (loop
= 0; loop
< (nr_pages
- 1); loop
++, p
++) {
1565 __ClearPageReserved(p
);
1566 set_page_count(p
, 0);
1568 __ClearPageReserved(p
);
1569 set_page_count(p
, 0);
1571 atomic_long_add(nr_pages
, &page_zone(page
)->managed_pages
);
1574 * Bypass PCP and place fresh pages right to the tail, primarily
1575 * relevant for memory onlining.
1577 __free_pages_ok(page
, order
, FPI_TO_TAIL
);
1580 #ifdef CONFIG_NEED_MULTIPLE_NODES
1583 * During memory init memblocks map pfns to nids. The search is expensive and
1584 * this caches recent lookups. The implementation of __early_pfn_to_nid
1585 * treats start/end as pfns.
1587 struct mminit_pfnnid_cache
{
1588 unsigned long last_start
;
1589 unsigned long last_end
;
1593 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata
;
1596 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
1598 static int __meminit
__early_pfn_to_nid(unsigned long pfn
,
1599 struct mminit_pfnnid_cache
*state
)
1601 unsigned long start_pfn
, end_pfn
;
1604 if (state
->last_start
<= pfn
&& pfn
< state
->last_end
)
1605 return state
->last_nid
;
1607 nid
= memblock_search_pfn_nid(pfn
, &start_pfn
, &end_pfn
);
1608 if (nid
!= NUMA_NO_NODE
) {
1609 state
->last_start
= start_pfn
;
1610 state
->last_end
= end_pfn
;
1611 state
->last_nid
= nid
;
1617 int __meminit
early_pfn_to_nid(unsigned long pfn
)
1619 static DEFINE_SPINLOCK(early_pfn_lock
);
1622 spin_lock(&early_pfn_lock
);
1623 nid
= __early_pfn_to_nid(pfn
, &early_pfnnid_cache
);
1625 nid
= first_online_node
;
1626 spin_unlock(&early_pfn_lock
);
1630 #endif /* CONFIG_NEED_MULTIPLE_NODES */
1632 void __init
memblock_free_pages(struct page
*page
, unsigned long pfn
,
1635 if (early_page_uninitialised(pfn
))
1637 __free_pages_core(page
, order
);
1641 * Check that the whole (or subset of) a pageblock given by the interval of
1642 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1643 * with the migration of free compaction scanner. The scanners then need to
1644 * use only pfn_valid_within() check for arches that allow holes within
1647 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1649 * It's possible on some configurations to have a setup like node0 node1 node0
1650 * i.e. it's possible that all pages within a zones range of pages do not
1651 * belong to a single zone. We assume that a border between node0 and node1
1652 * can occur within a single pageblock, but not a node0 node1 node0
1653 * interleaving within a single pageblock. It is therefore sufficient to check
1654 * the first and last page of a pageblock and avoid checking each individual
1655 * page in a pageblock.
1657 struct page
*__pageblock_pfn_to_page(unsigned long start_pfn
,
1658 unsigned long end_pfn
, struct zone
*zone
)
1660 struct page
*start_page
;
1661 struct page
*end_page
;
1663 /* end_pfn is one past the range we are checking */
1666 if (!pfn_valid(start_pfn
) || !pfn_valid(end_pfn
))
1669 start_page
= pfn_to_online_page(start_pfn
);
1673 if (page_zone(start_page
) != zone
)
1676 end_page
= pfn_to_page(end_pfn
);
1678 /* This gives a shorter code than deriving page_zone(end_page) */
1679 if (page_zone_id(start_page
) != page_zone_id(end_page
))
1685 void set_zone_contiguous(struct zone
*zone
)
1687 unsigned long block_start_pfn
= zone
->zone_start_pfn
;
1688 unsigned long block_end_pfn
;
1690 block_end_pfn
= ALIGN(block_start_pfn
+ 1, pageblock_nr_pages
);
1691 for (; block_start_pfn
< zone_end_pfn(zone
);
1692 block_start_pfn
= block_end_pfn
,
1693 block_end_pfn
+= pageblock_nr_pages
) {
1695 block_end_pfn
= min(block_end_pfn
, zone_end_pfn(zone
));
1697 if (!__pageblock_pfn_to_page(block_start_pfn
,
1698 block_end_pfn
, zone
))
1703 /* We confirm that there is no hole */
1704 zone
->contiguous
= true;
1707 void clear_zone_contiguous(struct zone
*zone
)
1709 zone
->contiguous
= false;
1712 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1713 static void __init
deferred_free_range(unsigned long pfn
,
1714 unsigned long nr_pages
)
1722 page
= pfn_to_page(pfn
);
1724 /* Free a large naturally-aligned chunk if possible */
1725 if (nr_pages
== pageblock_nr_pages
&&
1726 (pfn
& (pageblock_nr_pages
- 1)) == 0) {
1727 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
1728 __free_pages_core(page
, pageblock_order
);
1732 for (i
= 0; i
< nr_pages
; i
++, page
++, pfn
++) {
1733 if ((pfn
& (pageblock_nr_pages
- 1)) == 0)
1734 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
1735 __free_pages_core(page
, 0);
1739 /* Completion tracking for deferred_init_memmap() threads */
1740 static atomic_t pgdat_init_n_undone __initdata
;
1741 static __initdata
DECLARE_COMPLETION(pgdat_init_all_done_comp
);
1743 static inline void __init
pgdat_init_report_one_done(void)
1745 if (atomic_dec_and_test(&pgdat_init_n_undone
))
1746 complete(&pgdat_init_all_done_comp
);
1750 * Returns true if page needs to be initialized or freed to buddy allocator.
1752 * First we check if pfn is valid on architectures where it is possible to have
1753 * holes within pageblock_nr_pages. On systems where it is not possible, this
1754 * function is optimized out.
1756 * Then, we check if a current large page is valid by only checking the validity
1759 static inline bool __init
deferred_pfn_valid(unsigned long pfn
)
1761 if (!pfn_valid_within(pfn
))
1763 if (!(pfn
& (pageblock_nr_pages
- 1)) && !pfn_valid(pfn
))
1769 * Free pages to buddy allocator. Try to free aligned pages in
1770 * pageblock_nr_pages sizes.
1772 static void __init
deferred_free_pages(unsigned long pfn
,
1773 unsigned long end_pfn
)
1775 unsigned long nr_pgmask
= pageblock_nr_pages
- 1;
1776 unsigned long nr_free
= 0;
1778 for (; pfn
< end_pfn
; pfn
++) {
1779 if (!deferred_pfn_valid(pfn
)) {
1780 deferred_free_range(pfn
- nr_free
, nr_free
);
1782 } else if (!(pfn
& nr_pgmask
)) {
1783 deferred_free_range(pfn
- nr_free
, nr_free
);
1789 /* Free the last block of pages to allocator */
1790 deferred_free_range(pfn
- nr_free
, nr_free
);
1794 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1795 * by performing it only once every pageblock_nr_pages.
1796 * Return number of pages initialized.
1798 static unsigned long __init
deferred_init_pages(struct zone
*zone
,
1800 unsigned long end_pfn
)
1802 unsigned long nr_pgmask
= pageblock_nr_pages
- 1;
1803 int nid
= zone_to_nid(zone
);
1804 unsigned long nr_pages
= 0;
1805 int zid
= zone_idx(zone
);
1806 struct page
*page
= NULL
;
1808 for (; pfn
< end_pfn
; pfn
++) {
1809 if (!deferred_pfn_valid(pfn
)) {
1812 } else if (!page
|| !(pfn
& nr_pgmask
)) {
1813 page
= pfn_to_page(pfn
);
1817 __init_single_page(page
, pfn
, zid
, nid
);
1824 * This function is meant to pre-load the iterator for the zone init.
1825 * Specifically it walks through the ranges until we are caught up to the
1826 * first_init_pfn value and exits there. If we never encounter the value we
1827 * return false indicating there are no valid ranges left.
1830 deferred_init_mem_pfn_range_in_zone(u64
*i
, struct zone
*zone
,
1831 unsigned long *spfn
, unsigned long *epfn
,
1832 unsigned long first_init_pfn
)
1837 * Start out by walking through the ranges in this zone that have
1838 * already been initialized. We don't need to do anything with them
1839 * so we just need to flush them out of the system.
1841 for_each_free_mem_pfn_range_in_zone(j
, zone
, spfn
, epfn
) {
1842 if (*epfn
<= first_init_pfn
)
1844 if (*spfn
< first_init_pfn
)
1845 *spfn
= first_init_pfn
;
1854 * Initialize and free pages. We do it in two loops: first we initialize
1855 * struct page, then free to buddy allocator, because while we are
1856 * freeing pages we can access pages that are ahead (computing buddy
1857 * page in __free_one_page()).
1859 * In order to try and keep some memory in the cache we have the loop
1860 * broken along max page order boundaries. This way we will not cause
1861 * any issues with the buddy page computation.
1863 static unsigned long __init
1864 deferred_init_maxorder(u64
*i
, struct zone
*zone
, unsigned long *start_pfn
,
1865 unsigned long *end_pfn
)
1867 unsigned long mo_pfn
= ALIGN(*start_pfn
+ 1, MAX_ORDER_NR_PAGES
);
1868 unsigned long spfn
= *start_pfn
, epfn
= *end_pfn
;
1869 unsigned long nr_pages
= 0;
1872 /* First we loop through and initialize the page values */
1873 for_each_free_mem_pfn_range_in_zone_from(j
, zone
, start_pfn
, end_pfn
) {
1876 if (mo_pfn
<= *start_pfn
)
1879 t
= min(mo_pfn
, *end_pfn
);
1880 nr_pages
+= deferred_init_pages(zone
, *start_pfn
, t
);
1882 if (mo_pfn
< *end_pfn
) {
1883 *start_pfn
= mo_pfn
;
1888 /* Reset values and now loop through freeing pages as needed */
1891 for_each_free_mem_pfn_range_in_zone_from(j
, zone
, &spfn
, &epfn
) {
1897 t
= min(mo_pfn
, epfn
);
1898 deferred_free_pages(spfn
, t
);
1908 deferred_init_memmap_chunk(unsigned long start_pfn
, unsigned long end_pfn
,
1911 unsigned long spfn
, epfn
;
1912 struct zone
*zone
= arg
;
1915 deferred_init_mem_pfn_range_in_zone(&i
, zone
, &spfn
, &epfn
, start_pfn
);
1918 * Initialize and free pages in MAX_ORDER sized increments so that we
1919 * can avoid introducing any issues with the buddy allocator.
1921 while (spfn
< end_pfn
) {
1922 deferred_init_maxorder(&i
, zone
, &spfn
, &epfn
);
1927 /* An arch may override for more concurrency. */
1929 deferred_page_init_max_threads(const struct cpumask
*node_cpumask
)
1934 /* Initialise remaining memory on a node */
1935 static int __init
deferred_init_memmap(void *data
)
1937 pg_data_t
*pgdat
= data
;
1938 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
1939 unsigned long spfn
= 0, epfn
= 0;
1940 unsigned long first_init_pfn
, flags
;
1941 unsigned long start
= jiffies
;
1943 int zid
, max_threads
;
1946 /* Bind memory initialisation thread to a local node if possible */
1947 if (!cpumask_empty(cpumask
))
1948 set_cpus_allowed_ptr(current
, cpumask
);
1950 pgdat_resize_lock(pgdat
, &flags
);
1951 first_init_pfn
= pgdat
->first_deferred_pfn
;
1952 if (first_init_pfn
== ULONG_MAX
) {
1953 pgdat_resize_unlock(pgdat
, &flags
);
1954 pgdat_init_report_one_done();
1958 /* Sanity check boundaries */
1959 BUG_ON(pgdat
->first_deferred_pfn
< pgdat
->node_start_pfn
);
1960 BUG_ON(pgdat
->first_deferred_pfn
> pgdat_end_pfn(pgdat
));
1961 pgdat
->first_deferred_pfn
= ULONG_MAX
;
1964 * Once we unlock here, the zone cannot be grown anymore, thus if an
1965 * interrupt thread must allocate this early in boot, zone must be
1966 * pre-grown prior to start of deferred page initialization.
1968 pgdat_resize_unlock(pgdat
, &flags
);
1970 /* Only the highest zone is deferred so find it */
1971 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1972 zone
= pgdat
->node_zones
+ zid
;
1973 if (first_init_pfn
< zone_end_pfn(zone
))
1977 /* If the zone is empty somebody else may have cleared out the zone */
1978 if (!deferred_init_mem_pfn_range_in_zone(&i
, zone
, &spfn
, &epfn
,
1982 max_threads
= deferred_page_init_max_threads(cpumask
);
1984 while (spfn
< epfn
) {
1985 unsigned long epfn_align
= ALIGN(epfn
, PAGES_PER_SECTION
);
1986 struct padata_mt_job job
= {
1987 .thread_fn
= deferred_init_memmap_chunk
,
1990 .size
= epfn_align
- spfn
,
1991 .align
= PAGES_PER_SECTION
,
1992 .min_chunk
= PAGES_PER_SECTION
,
1993 .max_threads
= max_threads
,
1996 padata_do_multithreaded(&job
);
1997 deferred_init_mem_pfn_range_in_zone(&i
, zone
, &spfn
, &epfn
,
2001 /* Sanity check that the next zone really is unpopulated */
2002 WARN_ON(++zid
< MAX_NR_ZONES
&& populated_zone(++zone
));
2004 pr_info("node %d deferred pages initialised in %ums\n",
2005 pgdat
->node_id
, jiffies_to_msecs(jiffies
- start
));
2007 pgdat_init_report_one_done();
2012 * If this zone has deferred pages, try to grow it by initializing enough
2013 * deferred pages to satisfy the allocation specified by order, rounded up to
2014 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
2015 * of SECTION_SIZE bytes by initializing struct pages in increments of
2016 * PAGES_PER_SECTION * sizeof(struct page) bytes.
2018 * Return true when zone was grown, otherwise return false. We return true even
2019 * when we grow less than requested, to let the caller decide if there are
2020 * enough pages to satisfy the allocation.
2022 * Note: We use noinline because this function is needed only during boot, and
2023 * it is called from a __ref function _deferred_grow_zone. This way we are
2024 * making sure that it is not inlined into permanent text section.
2026 static noinline
bool __init
2027 deferred_grow_zone(struct zone
*zone
, unsigned int order
)
2029 unsigned long nr_pages_needed
= ALIGN(1 << order
, PAGES_PER_SECTION
);
2030 pg_data_t
*pgdat
= zone
->zone_pgdat
;
2031 unsigned long first_deferred_pfn
= pgdat
->first_deferred_pfn
;
2032 unsigned long spfn
, epfn
, flags
;
2033 unsigned long nr_pages
= 0;
2036 /* Only the last zone may have deferred pages */
2037 if (zone_end_pfn(zone
) != pgdat_end_pfn(pgdat
))
2040 pgdat_resize_lock(pgdat
, &flags
);
2043 * If someone grew this zone while we were waiting for spinlock, return
2044 * true, as there might be enough pages already.
2046 if (first_deferred_pfn
!= pgdat
->first_deferred_pfn
) {
2047 pgdat_resize_unlock(pgdat
, &flags
);
2051 /* If the zone is empty somebody else may have cleared out the zone */
2052 if (!deferred_init_mem_pfn_range_in_zone(&i
, zone
, &spfn
, &epfn
,
2053 first_deferred_pfn
)) {
2054 pgdat
->first_deferred_pfn
= ULONG_MAX
;
2055 pgdat_resize_unlock(pgdat
, &flags
);
2056 /* Retry only once. */
2057 return first_deferred_pfn
!= ULONG_MAX
;
2061 * Initialize and free pages in MAX_ORDER sized increments so
2062 * that we can avoid introducing any issues with the buddy
2065 while (spfn
< epfn
) {
2066 /* update our first deferred PFN for this section */
2067 first_deferred_pfn
= spfn
;
2069 nr_pages
+= deferred_init_maxorder(&i
, zone
, &spfn
, &epfn
);
2070 touch_nmi_watchdog();
2072 /* We should only stop along section boundaries */
2073 if ((first_deferred_pfn
^ spfn
) < PAGES_PER_SECTION
)
2076 /* If our quota has been met we can stop here */
2077 if (nr_pages
>= nr_pages_needed
)
2081 pgdat
->first_deferred_pfn
= spfn
;
2082 pgdat_resize_unlock(pgdat
, &flags
);
2084 return nr_pages
> 0;
2088 * deferred_grow_zone() is __init, but it is called from
2089 * get_page_from_freelist() during early boot until deferred_pages permanently
2090 * disables this call. This is why we have refdata wrapper to avoid warning,
2091 * and to ensure that the function body gets unloaded.
2094 _deferred_grow_zone(struct zone
*zone
, unsigned int order
)
2096 return deferred_grow_zone(zone
, order
);
2099 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
2101 void __init
page_alloc_init_late(void)
2106 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
2108 /* There will be num_node_state(N_MEMORY) threads */
2109 atomic_set(&pgdat_init_n_undone
, num_node_state(N_MEMORY
));
2110 for_each_node_state(nid
, N_MEMORY
) {
2111 kthread_run(deferred_init_memmap
, NODE_DATA(nid
), "pgdatinit%d", nid
);
2114 /* Block until all are initialised */
2115 wait_for_completion(&pgdat_init_all_done_comp
);
2118 * The number of managed pages has changed due to the initialisation
2119 * so the pcpu batch and high limits needs to be updated or the limits
2120 * will be artificially small.
2122 for_each_populated_zone(zone
)
2123 zone_pcp_update(zone
);
2126 * We initialized the rest of the deferred pages. Permanently disable
2127 * on-demand struct page initialization.
2129 static_branch_disable(&deferred_pages
);
2131 /* Reinit limits that are based on free pages after the kernel is up */
2132 files_maxfiles_init();
2137 /* Discard memblock private memory */
2140 for_each_node_state(nid
, N_MEMORY
)
2141 shuffle_free_memory(NODE_DATA(nid
));
2143 for_each_populated_zone(zone
)
2144 set_zone_contiguous(zone
);
2148 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
2149 void __init
init_cma_reserved_pageblock(struct page
*page
)
2151 unsigned i
= pageblock_nr_pages
;
2152 struct page
*p
= page
;
2155 __ClearPageReserved(p
);
2156 set_page_count(p
, 0);
2159 set_pageblock_migratetype(page
, MIGRATE_CMA
);
2161 if (pageblock_order
>= MAX_ORDER
) {
2162 i
= pageblock_nr_pages
;
2165 set_page_refcounted(p
);
2166 __free_pages(p
, MAX_ORDER
- 1);
2167 p
+= MAX_ORDER_NR_PAGES
;
2168 } while (i
-= MAX_ORDER_NR_PAGES
);
2170 set_page_refcounted(page
);
2171 __free_pages(page
, pageblock_order
);
2174 adjust_managed_page_count(page
, pageblock_nr_pages
);
2175 page_zone(page
)->cma_pages
+= pageblock_nr_pages
;
2180 * The order of subdivision here is critical for the IO subsystem.
2181 * Please do not alter this order without good reasons and regression
2182 * testing. Specifically, as large blocks of memory are subdivided,
2183 * the order in which smaller blocks are delivered depends on the order
2184 * they're subdivided in this function. This is the primary factor
2185 * influencing the order in which pages are delivered to the IO
2186 * subsystem according to empirical testing, and this is also justified
2187 * by considering the behavior of a buddy system containing a single
2188 * large block of memory acted on by a series of small allocations.
2189 * This behavior is a critical factor in sglist merging's success.
2193 static inline void expand(struct zone
*zone
, struct page
*page
,
2194 int low
, int high
, int migratetype
)
2196 unsigned long size
= 1 << high
;
2198 while (high
> low
) {
2201 VM_BUG_ON_PAGE(bad_range(zone
, &page
[size
]), &page
[size
]);
2204 * Mark as guard pages (or page), that will allow to
2205 * merge back to allocator when buddy will be freed.
2206 * Corresponding page table entries will not be touched,
2207 * pages will stay not present in virtual address space
2209 if (set_page_guard(zone
, &page
[size
], high
, migratetype
))
2212 add_to_free_list(&page
[size
], zone
, high
, migratetype
);
2213 set_buddy_order(&page
[size
], high
);
2217 static void check_new_page_bad(struct page
*page
)
2219 if (unlikely(page
->flags
& __PG_HWPOISON
)) {
2220 /* Don't complain about hwpoisoned pages */
2221 page_mapcount_reset(page
); /* remove PageBuddy */
2226 page_bad_reason(page
, PAGE_FLAGS_CHECK_AT_PREP
));
2230 * This page is about to be returned from the page allocator
2232 static inline int check_new_page(struct page
*page
)
2234 if (likely(page_expected_state(page
,
2235 PAGE_FLAGS_CHECK_AT_PREP
|__PG_HWPOISON
)))
2238 check_new_page_bad(page
);
2242 #ifdef CONFIG_DEBUG_VM
2244 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2245 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2246 * also checked when pcp lists are refilled from the free lists.
2248 static inline bool check_pcp_refill(struct page
*page
)
2250 if (debug_pagealloc_enabled_static())
2251 return check_new_page(page
);
2256 static inline bool check_new_pcp(struct page
*page
)
2258 return check_new_page(page
);
2262 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2263 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2264 * enabled, they are also checked when being allocated from the pcp lists.
2266 static inline bool check_pcp_refill(struct page
*page
)
2268 return check_new_page(page
);
2270 static inline bool check_new_pcp(struct page
*page
)
2272 if (debug_pagealloc_enabled_static())
2273 return check_new_page(page
);
2277 #endif /* CONFIG_DEBUG_VM */
2279 static bool check_new_pages(struct page
*page
, unsigned int order
)
2282 for (i
= 0; i
< (1 << order
); i
++) {
2283 struct page
*p
= page
+ i
;
2285 if (unlikely(check_new_page(p
)))
2292 inline void post_alloc_hook(struct page
*page
, unsigned int order
,
2295 set_page_private(page
, 0);
2296 set_page_refcounted(page
);
2298 arch_alloc_page(page
, order
);
2299 debug_pagealloc_map_pages(page
, 1 << order
);
2300 kasan_alloc_pages(page
, order
);
2301 kernel_unpoison_pages(page
, 1 << order
);
2302 set_page_owner(page
, order
, gfp_flags
);
2304 if (!want_init_on_free() && want_init_on_alloc(gfp_flags
))
2305 kernel_init_free_pages(page
, 1 << order
);
2308 static void prep_new_page(struct page
*page
, unsigned int order
, gfp_t gfp_flags
,
2309 unsigned int alloc_flags
)
2311 post_alloc_hook(page
, order
, gfp_flags
);
2313 if (order
&& (gfp_flags
& __GFP_COMP
))
2314 prep_compound_page(page
, order
);
2317 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2318 * allocate the page. The expectation is that the caller is taking
2319 * steps that will free more memory. The caller should avoid the page
2320 * being used for !PFMEMALLOC purposes.
2322 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
2323 set_page_pfmemalloc(page
);
2325 clear_page_pfmemalloc(page
);
2329 * Go through the free lists for the given migratetype and remove
2330 * the smallest available page from the freelists
2332 static __always_inline
2333 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
2336 unsigned int current_order
;
2337 struct free_area
*area
;
2340 /* Find a page of the appropriate size in the preferred list */
2341 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
2342 area
= &(zone
->free_area
[current_order
]);
2343 page
= get_page_from_free_area(area
, migratetype
);
2346 del_page_from_free_list(page
, zone
, current_order
);
2347 expand(zone
, page
, order
, current_order
, migratetype
);
2348 set_pcppage_migratetype(page
, migratetype
);
2357 * This array describes the order lists are fallen back to when
2358 * the free lists for the desirable migrate type are depleted
2360 static int fallbacks
[MIGRATE_TYPES
][3] = {
2361 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_TYPES
},
2362 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_TYPES
},
2363 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_TYPES
},
2365 [MIGRATE_CMA
] = { MIGRATE_TYPES
}, /* Never used */
2367 #ifdef CONFIG_MEMORY_ISOLATION
2368 [MIGRATE_ISOLATE
] = { MIGRATE_TYPES
}, /* Never used */
2373 static __always_inline
struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
2376 return __rmqueue_smallest(zone
, order
, MIGRATE_CMA
);
2379 static inline struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
2380 unsigned int order
) { return NULL
; }
2384 * Move the free pages in a range to the freelist tail of the requested type.
2385 * Note that start_page and end_pages are not aligned on a pageblock
2386 * boundary. If alignment is required, use move_freepages_block()
2388 static int move_freepages(struct zone
*zone
,
2389 struct page
*start_page
, struct page
*end_page
,
2390 int migratetype
, int *num_movable
)
2394 int pages_moved
= 0;
2396 for (page
= start_page
; page
<= end_page
;) {
2397 if (!pfn_valid_within(page_to_pfn(page
))) {
2402 if (!PageBuddy(page
)) {
2404 * We assume that pages that could be isolated for
2405 * migration are movable. But we don't actually try
2406 * isolating, as that would be expensive.
2409 (PageLRU(page
) || __PageMovable(page
)))
2416 /* Make sure we are not inadvertently changing nodes */
2417 VM_BUG_ON_PAGE(page_to_nid(page
) != zone_to_nid(zone
), page
);
2418 VM_BUG_ON_PAGE(page_zone(page
) != zone
, page
);
2420 order
= buddy_order(page
);
2421 move_to_free_list(page
, zone
, order
, migratetype
);
2423 pages_moved
+= 1 << order
;
2429 int move_freepages_block(struct zone
*zone
, struct page
*page
,
2430 int migratetype
, int *num_movable
)
2432 unsigned long start_pfn
, end_pfn
;
2433 struct page
*start_page
, *end_page
;
2438 start_pfn
= page_to_pfn(page
);
2439 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
2440 start_page
= pfn_to_page(start_pfn
);
2441 end_page
= start_page
+ pageblock_nr_pages
- 1;
2442 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
2444 /* Do not cross zone boundaries */
2445 if (!zone_spans_pfn(zone
, start_pfn
))
2447 if (!zone_spans_pfn(zone
, end_pfn
))
2450 return move_freepages(zone
, start_page
, end_page
, migratetype
,
2454 static void change_pageblock_range(struct page
*pageblock_page
,
2455 int start_order
, int migratetype
)
2457 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
2459 while (nr_pageblocks
--) {
2460 set_pageblock_migratetype(pageblock_page
, migratetype
);
2461 pageblock_page
+= pageblock_nr_pages
;
2466 * When we are falling back to another migratetype during allocation, try to
2467 * steal extra free pages from the same pageblocks to satisfy further
2468 * allocations, instead of polluting multiple pageblocks.
2470 * If we are stealing a relatively large buddy page, it is likely there will
2471 * be more free pages in the pageblock, so try to steal them all. For
2472 * reclaimable and unmovable allocations, we steal regardless of page size,
2473 * as fragmentation caused by those allocations polluting movable pageblocks
2474 * is worse than movable allocations stealing from unmovable and reclaimable
2477 static bool can_steal_fallback(unsigned int order
, int start_mt
)
2480 * Leaving this order check is intended, although there is
2481 * relaxed order check in next check. The reason is that
2482 * we can actually steal whole pageblock if this condition met,
2483 * but, below check doesn't guarantee it and that is just heuristic
2484 * so could be changed anytime.
2486 if (order
>= pageblock_order
)
2489 if (order
>= pageblock_order
/ 2 ||
2490 start_mt
== MIGRATE_RECLAIMABLE
||
2491 start_mt
== MIGRATE_UNMOVABLE
||
2492 page_group_by_mobility_disabled
)
2498 static inline bool boost_watermark(struct zone
*zone
)
2500 unsigned long max_boost
;
2502 if (!watermark_boost_factor
)
2505 * Don't bother in zones that are unlikely to produce results.
2506 * On small machines, including kdump capture kernels running
2507 * in a small area, boosting the watermark can cause an out of
2508 * memory situation immediately.
2510 if ((pageblock_nr_pages
* 4) > zone_managed_pages(zone
))
2513 max_boost
= mult_frac(zone
->_watermark
[WMARK_HIGH
],
2514 watermark_boost_factor
, 10000);
2517 * high watermark may be uninitialised if fragmentation occurs
2518 * very early in boot so do not boost. We do not fall
2519 * through and boost by pageblock_nr_pages as failing
2520 * allocations that early means that reclaim is not going
2521 * to help and it may even be impossible to reclaim the
2522 * boosted watermark resulting in a hang.
2527 max_boost
= max(pageblock_nr_pages
, max_boost
);
2529 zone
->watermark_boost
= min(zone
->watermark_boost
+ pageblock_nr_pages
,
2536 * This function implements actual steal behaviour. If order is large enough,
2537 * we can steal whole pageblock. If not, we first move freepages in this
2538 * pageblock to our migratetype and determine how many already-allocated pages
2539 * are there in the pageblock with a compatible migratetype. If at least half
2540 * of pages are free or compatible, we can change migratetype of the pageblock
2541 * itself, so pages freed in the future will be put on the correct free list.
2543 static void steal_suitable_fallback(struct zone
*zone
, struct page
*page
,
2544 unsigned int alloc_flags
, int start_type
, bool whole_block
)
2546 unsigned int current_order
= buddy_order(page
);
2547 int free_pages
, movable_pages
, alike_pages
;
2550 old_block_type
= get_pageblock_migratetype(page
);
2553 * This can happen due to races and we want to prevent broken
2554 * highatomic accounting.
2556 if (is_migrate_highatomic(old_block_type
))
2559 /* Take ownership for orders >= pageblock_order */
2560 if (current_order
>= pageblock_order
) {
2561 change_pageblock_range(page
, current_order
, start_type
);
2566 * Boost watermarks to increase reclaim pressure to reduce the
2567 * likelihood of future fallbacks. Wake kswapd now as the node
2568 * may be balanced overall and kswapd will not wake naturally.
2570 if (boost_watermark(zone
) && (alloc_flags
& ALLOC_KSWAPD
))
2571 set_bit(ZONE_BOOSTED_WATERMARK
, &zone
->flags
);
2573 /* We are not allowed to try stealing from the whole block */
2577 free_pages
= move_freepages_block(zone
, page
, start_type
,
2580 * Determine how many pages are compatible with our allocation.
2581 * For movable allocation, it's the number of movable pages which
2582 * we just obtained. For other types it's a bit more tricky.
2584 if (start_type
== MIGRATE_MOVABLE
) {
2585 alike_pages
= movable_pages
;
2588 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2589 * to MOVABLE pageblock, consider all non-movable pages as
2590 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2591 * vice versa, be conservative since we can't distinguish the
2592 * exact migratetype of non-movable pages.
2594 if (old_block_type
== MIGRATE_MOVABLE
)
2595 alike_pages
= pageblock_nr_pages
2596 - (free_pages
+ movable_pages
);
2601 /* moving whole block can fail due to zone boundary conditions */
2606 * If a sufficient number of pages in the block are either free or of
2607 * comparable migratability as our allocation, claim the whole block.
2609 if (free_pages
+ alike_pages
>= (1 << (pageblock_order
-1)) ||
2610 page_group_by_mobility_disabled
)
2611 set_pageblock_migratetype(page
, start_type
);
2616 move_to_free_list(page
, zone
, current_order
, start_type
);
2620 * Check whether there is a suitable fallback freepage with requested order.
2621 * If only_stealable is true, this function returns fallback_mt only if
2622 * we can steal other freepages all together. This would help to reduce
2623 * fragmentation due to mixed migratetype pages in one pageblock.
2625 int find_suitable_fallback(struct free_area
*area
, unsigned int order
,
2626 int migratetype
, bool only_stealable
, bool *can_steal
)
2631 if (area
->nr_free
== 0)
2636 fallback_mt
= fallbacks
[migratetype
][i
];
2637 if (fallback_mt
== MIGRATE_TYPES
)
2640 if (free_area_empty(area
, fallback_mt
))
2643 if (can_steal_fallback(order
, migratetype
))
2646 if (!only_stealable
)
2657 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2658 * there are no empty page blocks that contain a page with a suitable order
2660 static void reserve_highatomic_pageblock(struct page
*page
, struct zone
*zone
,
2661 unsigned int alloc_order
)
2664 unsigned long max_managed
, flags
;
2667 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2668 * Check is race-prone but harmless.
2670 max_managed
= (zone_managed_pages(zone
) / 100) + pageblock_nr_pages
;
2671 if (zone
->nr_reserved_highatomic
>= max_managed
)
2674 spin_lock_irqsave(&zone
->lock
, flags
);
2676 /* Recheck the nr_reserved_highatomic limit under the lock */
2677 if (zone
->nr_reserved_highatomic
>= max_managed
)
2681 mt
= get_pageblock_migratetype(page
);
2682 if (!is_migrate_highatomic(mt
) && !is_migrate_isolate(mt
)
2683 && !is_migrate_cma(mt
)) {
2684 zone
->nr_reserved_highatomic
+= pageblock_nr_pages
;
2685 set_pageblock_migratetype(page
, MIGRATE_HIGHATOMIC
);
2686 move_freepages_block(zone
, page
, MIGRATE_HIGHATOMIC
, NULL
);
2690 spin_unlock_irqrestore(&zone
->lock
, flags
);
2694 * Used when an allocation is about to fail under memory pressure. This
2695 * potentially hurts the reliability of high-order allocations when under
2696 * intense memory pressure but failed atomic allocations should be easier
2697 * to recover from than an OOM.
2699 * If @force is true, try to unreserve a pageblock even though highatomic
2700 * pageblock is exhausted.
2702 static bool unreserve_highatomic_pageblock(const struct alloc_context
*ac
,
2705 struct zonelist
*zonelist
= ac
->zonelist
;
2706 unsigned long flags
;
2713 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, ac
->highest_zoneidx
,
2716 * Preserve at least one pageblock unless memory pressure
2719 if (!force
&& zone
->nr_reserved_highatomic
<=
2723 spin_lock_irqsave(&zone
->lock
, flags
);
2724 for (order
= 0; order
< MAX_ORDER
; order
++) {
2725 struct free_area
*area
= &(zone
->free_area
[order
]);
2727 page
= get_page_from_free_area(area
, MIGRATE_HIGHATOMIC
);
2732 * In page freeing path, migratetype change is racy so
2733 * we can counter several free pages in a pageblock
2734 * in this loop althoug we changed the pageblock type
2735 * from highatomic to ac->migratetype. So we should
2736 * adjust the count once.
2738 if (is_migrate_highatomic_page(page
)) {
2740 * It should never happen but changes to
2741 * locking could inadvertently allow a per-cpu
2742 * drain to add pages to MIGRATE_HIGHATOMIC
2743 * while unreserving so be safe and watch for
2746 zone
->nr_reserved_highatomic
-= min(
2748 zone
->nr_reserved_highatomic
);
2752 * Convert to ac->migratetype and avoid the normal
2753 * pageblock stealing heuristics. Minimally, the caller
2754 * is doing the work and needs the pages. More
2755 * importantly, if the block was always converted to
2756 * MIGRATE_UNMOVABLE or another type then the number
2757 * of pageblocks that cannot be completely freed
2760 set_pageblock_migratetype(page
, ac
->migratetype
);
2761 ret
= move_freepages_block(zone
, page
, ac
->migratetype
,
2764 spin_unlock_irqrestore(&zone
->lock
, flags
);
2768 spin_unlock_irqrestore(&zone
->lock
, flags
);
2775 * Try finding a free buddy page on the fallback list and put it on the free
2776 * list of requested migratetype, possibly along with other pages from the same
2777 * block, depending on fragmentation avoidance heuristics. Returns true if
2778 * fallback was found so that __rmqueue_smallest() can grab it.
2780 * The use of signed ints for order and current_order is a deliberate
2781 * deviation from the rest of this file, to make the for loop
2782 * condition simpler.
2784 static __always_inline
bool
2785 __rmqueue_fallback(struct zone
*zone
, int order
, int start_migratetype
,
2786 unsigned int alloc_flags
)
2788 struct free_area
*area
;
2790 int min_order
= order
;
2796 * Do not steal pages from freelists belonging to other pageblocks
2797 * i.e. orders < pageblock_order. If there are no local zones free,
2798 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2800 if (alloc_flags
& ALLOC_NOFRAGMENT
)
2801 min_order
= pageblock_order
;
2804 * Find the largest available free page in the other list. This roughly
2805 * approximates finding the pageblock with the most free pages, which
2806 * would be too costly to do exactly.
2808 for (current_order
= MAX_ORDER
- 1; current_order
>= min_order
;
2810 area
= &(zone
->free_area
[current_order
]);
2811 fallback_mt
= find_suitable_fallback(area
, current_order
,
2812 start_migratetype
, false, &can_steal
);
2813 if (fallback_mt
== -1)
2817 * We cannot steal all free pages from the pageblock and the
2818 * requested migratetype is movable. In that case it's better to
2819 * steal and split the smallest available page instead of the
2820 * largest available page, because even if the next movable
2821 * allocation falls back into a different pageblock than this
2822 * one, it won't cause permanent fragmentation.
2824 if (!can_steal
&& start_migratetype
== MIGRATE_MOVABLE
2825 && current_order
> order
)
2834 for (current_order
= order
; current_order
< MAX_ORDER
;
2836 area
= &(zone
->free_area
[current_order
]);
2837 fallback_mt
= find_suitable_fallback(area
, current_order
,
2838 start_migratetype
, false, &can_steal
);
2839 if (fallback_mt
!= -1)
2844 * This should not happen - we already found a suitable fallback
2845 * when looking for the largest page.
2847 VM_BUG_ON(current_order
== MAX_ORDER
);
2850 page
= get_page_from_free_area(area
, fallback_mt
);
2852 steal_suitable_fallback(zone
, page
, alloc_flags
, start_migratetype
,
2855 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
2856 start_migratetype
, fallback_mt
);
2863 * Do the hard work of removing an element from the buddy allocator.
2864 * Call me with the zone->lock already held.
2866 static __always_inline
struct page
*
2867 __rmqueue(struct zone
*zone
, unsigned int order
, int migratetype
,
2868 unsigned int alloc_flags
)
2872 if (IS_ENABLED(CONFIG_CMA
)) {
2874 * Balance movable allocations between regular and CMA areas by
2875 * allocating from CMA when over half of the zone's free memory
2876 * is in the CMA area.
2878 if (alloc_flags
& ALLOC_CMA
&&
2879 zone_page_state(zone
, NR_FREE_CMA_PAGES
) >
2880 zone_page_state(zone
, NR_FREE_PAGES
) / 2) {
2881 page
= __rmqueue_cma_fallback(zone
, order
);
2887 page
= __rmqueue_smallest(zone
, order
, migratetype
);
2888 if (unlikely(!page
)) {
2889 if (alloc_flags
& ALLOC_CMA
)
2890 page
= __rmqueue_cma_fallback(zone
, order
);
2892 if (!page
&& __rmqueue_fallback(zone
, order
, migratetype
,
2898 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
2903 * Obtain a specified number of elements from the buddy allocator, all under
2904 * a single hold of the lock, for efficiency. Add them to the supplied list.
2905 * Returns the number of new pages which were placed at *list.
2907 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
2908 unsigned long count
, struct list_head
*list
,
2909 int migratetype
, unsigned int alloc_flags
)
2913 spin_lock(&zone
->lock
);
2914 for (i
= 0; i
< count
; ++i
) {
2915 struct page
*page
= __rmqueue(zone
, order
, migratetype
,
2917 if (unlikely(page
== NULL
))
2920 if (unlikely(check_pcp_refill(page
)))
2924 * Split buddy pages returned by expand() are received here in
2925 * physical page order. The page is added to the tail of
2926 * caller's list. From the callers perspective, the linked list
2927 * is ordered by page number under some conditions. This is
2928 * useful for IO devices that can forward direction from the
2929 * head, thus also in the physical page order. This is useful
2930 * for IO devices that can merge IO requests if the physical
2931 * pages are ordered properly.
2933 list_add_tail(&page
->lru
, list
);
2935 if (is_migrate_cma(get_pcppage_migratetype(page
)))
2936 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
,
2941 * i pages were removed from the buddy list even if some leak due
2942 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2943 * on i. Do not confuse with 'alloced' which is the number of
2944 * pages added to the pcp list.
2946 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
2947 spin_unlock(&zone
->lock
);
2953 * Called from the vmstat counter updater to drain pagesets of this
2954 * currently executing processor on remote nodes after they have
2957 * Note that this function must be called with the thread pinned to
2958 * a single processor.
2960 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
2962 unsigned long flags
;
2963 int to_drain
, batch
;
2965 local_irq_save(flags
);
2966 batch
= READ_ONCE(pcp
->batch
);
2967 to_drain
= min(pcp
->count
, batch
);
2969 free_pcppages_bulk(zone
, to_drain
, pcp
);
2970 local_irq_restore(flags
);
2975 * Drain pcplists of the indicated processor and zone.
2977 * The processor must either be the current processor and the
2978 * thread pinned to the current processor or a processor that
2981 static void drain_pages_zone(unsigned int cpu
, struct zone
*zone
)
2983 unsigned long flags
;
2984 struct per_cpu_pageset
*pset
;
2985 struct per_cpu_pages
*pcp
;
2987 local_irq_save(flags
);
2988 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
2992 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
2993 local_irq_restore(flags
);
2997 * Drain pcplists of all zones on the indicated processor.
2999 * The processor must either be the current processor and the
3000 * thread pinned to the current processor or a processor that
3003 static void drain_pages(unsigned int cpu
)
3007 for_each_populated_zone(zone
) {
3008 drain_pages_zone(cpu
, zone
);
3013 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
3015 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
3016 * the single zone's pages.
3018 void drain_local_pages(struct zone
*zone
)
3020 int cpu
= smp_processor_id();
3023 drain_pages_zone(cpu
, zone
);
3028 static void drain_local_pages_wq(struct work_struct
*work
)
3030 struct pcpu_drain
*drain
;
3032 drain
= container_of(work
, struct pcpu_drain
, work
);
3035 * drain_all_pages doesn't use proper cpu hotplug protection so
3036 * we can race with cpu offline when the WQ can move this from
3037 * a cpu pinned worker to an unbound one. We can operate on a different
3038 * cpu which is allright but we also have to make sure to not move to
3042 drain_local_pages(drain
->zone
);
3047 * The implementation of drain_all_pages(), exposing an extra parameter to
3048 * drain on all cpus.
3050 * drain_all_pages() is optimized to only execute on cpus where pcplists are
3051 * not empty. The check for non-emptiness can however race with a free to
3052 * pcplist that has not yet increased the pcp->count from 0 to 1. Callers
3053 * that need the guarantee that every CPU has drained can disable the
3054 * optimizing racy check.
3056 static void __drain_all_pages(struct zone
*zone
, bool force_all_cpus
)
3061 * Allocate in the BSS so we wont require allocation in
3062 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
3064 static cpumask_t cpus_with_pcps
;
3067 * Make sure nobody triggers this path before mm_percpu_wq is fully
3070 if (WARN_ON_ONCE(!mm_percpu_wq
))
3074 * Do not drain if one is already in progress unless it's specific to
3075 * a zone. Such callers are primarily CMA and memory hotplug and need
3076 * the drain to be complete when the call returns.
3078 if (unlikely(!mutex_trylock(&pcpu_drain_mutex
))) {
3081 mutex_lock(&pcpu_drain_mutex
);
3085 * We don't care about racing with CPU hotplug event
3086 * as offline notification will cause the notified
3087 * cpu to drain that CPU pcps and on_each_cpu_mask
3088 * disables preemption as part of its processing
3090 for_each_online_cpu(cpu
) {
3091 struct per_cpu_pageset
*pcp
;
3093 bool has_pcps
= false;
3095 if (force_all_cpus
) {
3097 * The pcp.count check is racy, some callers need a
3098 * guarantee that no cpu is missed.
3102 pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
3106 for_each_populated_zone(z
) {
3107 pcp
= per_cpu_ptr(z
->pageset
, cpu
);
3108 if (pcp
->pcp
.count
) {
3116 cpumask_set_cpu(cpu
, &cpus_with_pcps
);
3118 cpumask_clear_cpu(cpu
, &cpus_with_pcps
);
3121 for_each_cpu(cpu
, &cpus_with_pcps
) {
3122 struct pcpu_drain
*drain
= per_cpu_ptr(&pcpu_drain
, cpu
);
3125 INIT_WORK(&drain
->work
, drain_local_pages_wq
);
3126 queue_work_on(cpu
, mm_percpu_wq
, &drain
->work
);
3128 for_each_cpu(cpu
, &cpus_with_pcps
)
3129 flush_work(&per_cpu_ptr(&pcpu_drain
, cpu
)->work
);
3131 mutex_unlock(&pcpu_drain_mutex
);
3135 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
3137 * When zone parameter is non-NULL, spill just the single zone's pages.
3139 * Note that this can be extremely slow as the draining happens in a workqueue.
3141 void drain_all_pages(struct zone
*zone
)
3143 __drain_all_pages(zone
, false);
3146 #ifdef CONFIG_HIBERNATION
3149 * Touch the watchdog for every WD_PAGE_COUNT pages.
3151 #define WD_PAGE_COUNT (128*1024)
3153 void mark_free_pages(struct zone
*zone
)
3155 unsigned long pfn
, max_zone_pfn
, page_count
= WD_PAGE_COUNT
;
3156 unsigned long flags
;
3157 unsigned int order
, t
;
3160 if (zone_is_empty(zone
))
3163 spin_lock_irqsave(&zone
->lock
, flags
);
3165 max_zone_pfn
= zone_end_pfn(zone
);
3166 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
3167 if (pfn_valid(pfn
)) {
3168 page
= pfn_to_page(pfn
);
3170 if (!--page_count
) {
3171 touch_nmi_watchdog();
3172 page_count
= WD_PAGE_COUNT
;
3175 if (page_zone(page
) != zone
)
3178 if (!swsusp_page_is_forbidden(page
))
3179 swsusp_unset_page_free(page
);
3182 for_each_migratetype_order(order
, t
) {
3183 list_for_each_entry(page
,
3184 &zone
->free_area
[order
].free_list
[t
], lru
) {
3187 pfn
= page_to_pfn(page
);
3188 for (i
= 0; i
< (1UL << order
); i
++) {
3189 if (!--page_count
) {
3190 touch_nmi_watchdog();
3191 page_count
= WD_PAGE_COUNT
;
3193 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
3197 spin_unlock_irqrestore(&zone
->lock
, flags
);
3199 #endif /* CONFIG_PM */
3201 static bool free_unref_page_prepare(struct page
*page
, unsigned long pfn
)
3205 if (!free_pcp_prepare(page
))
3208 migratetype
= get_pfnblock_migratetype(page
, pfn
);
3209 set_pcppage_migratetype(page
, migratetype
);
3213 static void free_unref_page_commit(struct page
*page
, unsigned long pfn
)
3215 struct zone
*zone
= page_zone(page
);
3216 struct per_cpu_pages
*pcp
;
3219 migratetype
= get_pcppage_migratetype(page
);
3220 __count_vm_event(PGFREE
);
3223 * We only track unmovable, reclaimable and movable on pcp lists.
3224 * Free ISOLATE pages back to the allocator because they are being
3225 * offlined but treat HIGHATOMIC as movable pages so we can get those
3226 * areas back if necessary. Otherwise, we may have to free
3227 * excessively into the page allocator
3229 if (migratetype
>= MIGRATE_PCPTYPES
) {
3230 if (unlikely(is_migrate_isolate(migratetype
))) {
3231 free_one_page(zone
, page
, pfn
, 0, migratetype
,
3235 migratetype
= MIGRATE_MOVABLE
;
3238 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
3239 list_add(&page
->lru
, &pcp
->lists
[migratetype
]);
3241 if (pcp
->count
>= READ_ONCE(pcp
->high
))
3242 free_pcppages_bulk(zone
, READ_ONCE(pcp
->batch
), pcp
);
3246 * Free a 0-order page
3248 void free_unref_page(struct page
*page
)
3250 unsigned long flags
;
3251 unsigned long pfn
= page_to_pfn(page
);
3253 if (!free_unref_page_prepare(page
, pfn
))
3256 local_irq_save(flags
);
3257 free_unref_page_commit(page
, pfn
);
3258 local_irq_restore(flags
);
3262 * Free a list of 0-order pages
3264 void free_unref_page_list(struct list_head
*list
)
3266 struct page
*page
, *next
;
3267 unsigned long flags
, pfn
;
3268 int batch_count
= 0;
3270 /* Prepare pages for freeing */
3271 list_for_each_entry_safe(page
, next
, list
, lru
) {
3272 pfn
= page_to_pfn(page
);
3273 if (!free_unref_page_prepare(page
, pfn
))
3274 list_del(&page
->lru
);
3275 set_page_private(page
, pfn
);
3278 local_irq_save(flags
);
3279 list_for_each_entry_safe(page
, next
, list
, lru
) {
3280 unsigned long pfn
= page_private(page
);
3282 set_page_private(page
, 0);
3283 trace_mm_page_free_batched(page
);
3284 free_unref_page_commit(page
, pfn
);
3287 * Guard against excessive IRQ disabled times when we get
3288 * a large list of pages to free.
3290 if (++batch_count
== SWAP_CLUSTER_MAX
) {
3291 local_irq_restore(flags
);
3293 local_irq_save(flags
);
3296 local_irq_restore(flags
);
3300 * split_page takes a non-compound higher-order page, and splits it into
3301 * n (1<<order) sub-pages: page[0..n]
3302 * Each sub-page must be freed individually.
3304 * Note: this is probably too low level an operation for use in drivers.
3305 * Please consult with lkml before using this in your driver.
3307 void split_page(struct page
*page
, unsigned int order
)
3311 VM_BUG_ON_PAGE(PageCompound(page
), page
);
3312 VM_BUG_ON_PAGE(!page_count(page
), page
);
3314 for (i
= 1; i
< (1 << order
); i
++)
3315 set_page_refcounted(page
+ i
);
3316 split_page_owner(page
, 1 << order
);
3318 EXPORT_SYMBOL_GPL(split_page
);
3320 int __isolate_free_page(struct page
*page
, unsigned int order
)
3322 unsigned long watermark
;
3326 BUG_ON(!PageBuddy(page
));
3328 zone
= page_zone(page
);
3329 mt
= get_pageblock_migratetype(page
);
3331 if (!is_migrate_isolate(mt
)) {
3333 * Obey watermarks as if the page was being allocated. We can
3334 * emulate a high-order watermark check with a raised order-0
3335 * watermark, because we already know our high-order page
3338 watermark
= zone
->_watermark
[WMARK_MIN
] + (1UL << order
);
3339 if (!zone_watermark_ok(zone
, 0, watermark
, 0, ALLOC_CMA
))
3342 __mod_zone_freepage_state(zone
, -(1UL << order
), mt
);
3345 /* Remove page from free list */
3347 del_page_from_free_list(page
, zone
, order
);
3350 * Set the pageblock if the isolated page is at least half of a
3353 if (order
>= pageblock_order
- 1) {
3354 struct page
*endpage
= page
+ (1 << order
) - 1;
3355 for (; page
< endpage
; page
+= pageblock_nr_pages
) {
3356 int mt
= get_pageblock_migratetype(page
);
3357 if (!is_migrate_isolate(mt
) && !is_migrate_cma(mt
)
3358 && !is_migrate_highatomic(mt
))
3359 set_pageblock_migratetype(page
,
3365 return 1UL << order
;
3369 * __putback_isolated_page - Return a now-isolated page back where we got it
3370 * @page: Page that was isolated
3371 * @order: Order of the isolated page
3372 * @mt: The page's pageblock's migratetype
3374 * This function is meant to return a page pulled from the free lists via
3375 * __isolate_free_page back to the free lists they were pulled from.
3377 void __putback_isolated_page(struct page
*page
, unsigned int order
, int mt
)
3379 struct zone
*zone
= page_zone(page
);
3381 /* zone lock should be held when this function is called */
3382 lockdep_assert_held(&zone
->lock
);
3384 /* Return isolated page to tail of freelist. */
3385 __free_one_page(page
, page_to_pfn(page
), zone
, order
, mt
,
3386 FPI_SKIP_REPORT_NOTIFY
| FPI_TO_TAIL
);
3390 * Update NUMA hit/miss statistics
3392 * Must be called with interrupts disabled.
3394 static inline void zone_statistics(struct zone
*preferred_zone
, struct zone
*z
)
3397 enum numa_stat_item local_stat
= NUMA_LOCAL
;
3399 /* skip numa counters update if numa stats is disabled */
3400 if (!static_branch_likely(&vm_numa_stat_key
))
3403 if (zone_to_nid(z
) != numa_node_id())
3404 local_stat
= NUMA_OTHER
;
3406 if (zone_to_nid(z
) == zone_to_nid(preferred_zone
))
3407 __inc_numa_state(z
, NUMA_HIT
);
3409 __inc_numa_state(z
, NUMA_MISS
);
3410 __inc_numa_state(preferred_zone
, NUMA_FOREIGN
);
3412 __inc_numa_state(z
, local_stat
);
3416 /* Remove page from the per-cpu list, caller must protect the list */
3417 static struct page
*__rmqueue_pcplist(struct zone
*zone
, int migratetype
,
3418 unsigned int alloc_flags
,
3419 struct per_cpu_pages
*pcp
,
3420 struct list_head
*list
)
3425 if (list_empty(list
)) {
3426 pcp
->count
+= rmqueue_bulk(zone
, 0,
3427 READ_ONCE(pcp
->batch
), list
,
3428 migratetype
, alloc_flags
);
3429 if (unlikely(list_empty(list
)))
3433 page
= list_first_entry(list
, struct page
, lru
);
3434 list_del(&page
->lru
);
3436 } while (check_new_pcp(page
));
3441 /* Lock and remove page from the per-cpu list */
3442 static struct page
*rmqueue_pcplist(struct zone
*preferred_zone
,
3443 struct zone
*zone
, gfp_t gfp_flags
,
3444 int migratetype
, unsigned int alloc_flags
)
3446 struct per_cpu_pages
*pcp
;
3447 struct list_head
*list
;
3449 unsigned long flags
;
3451 local_irq_save(flags
);
3452 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
3453 list
= &pcp
->lists
[migratetype
];
3454 page
= __rmqueue_pcplist(zone
, migratetype
, alloc_flags
, pcp
, list
);
3456 __count_zid_vm_events(PGALLOC
, page_zonenum(page
), 1);
3457 zone_statistics(preferred_zone
, zone
);
3459 local_irq_restore(flags
);
3464 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3467 struct page
*rmqueue(struct zone
*preferred_zone
,
3468 struct zone
*zone
, unsigned int order
,
3469 gfp_t gfp_flags
, unsigned int alloc_flags
,
3472 unsigned long flags
;
3475 if (likely(order
== 0)) {
3477 * MIGRATE_MOVABLE pcplist could have the pages on CMA area and
3478 * we need to skip it when CMA area isn't allowed.
3480 if (!IS_ENABLED(CONFIG_CMA
) || alloc_flags
& ALLOC_CMA
||
3481 migratetype
!= MIGRATE_MOVABLE
) {
3482 page
= rmqueue_pcplist(preferred_zone
, zone
, gfp_flags
,
3483 migratetype
, alloc_flags
);
3489 * We most definitely don't want callers attempting to
3490 * allocate greater than order-1 page units with __GFP_NOFAIL.
3492 WARN_ON_ONCE((gfp_flags
& __GFP_NOFAIL
) && (order
> 1));
3493 spin_lock_irqsave(&zone
->lock
, flags
);
3498 * order-0 request can reach here when the pcplist is skipped
3499 * due to non-CMA allocation context. HIGHATOMIC area is
3500 * reserved for high-order atomic allocation, so order-0
3501 * request should skip it.
3503 if (order
> 0 && alloc_flags
& ALLOC_HARDER
) {
3504 page
= __rmqueue_smallest(zone
, order
, MIGRATE_HIGHATOMIC
);
3506 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
3509 page
= __rmqueue(zone
, order
, migratetype
, alloc_flags
);
3510 } while (page
&& check_new_pages(page
, order
));
3511 spin_unlock(&zone
->lock
);
3514 __mod_zone_freepage_state(zone
, -(1 << order
),
3515 get_pcppage_migratetype(page
));
3517 __count_zid_vm_events(PGALLOC
, page_zonenum(page
), 1 << order
);
3518 zone_statistics(preferred_zone
, zone
);
3519 local_irq_restore(flags
);
3522 /* Separate test+clear to avoid unnecessary atomics */
3523 if (test_bit(ZONE_BOOSTED_WATERMARK
, &zone
->flags
)) {
3524 clear_bit(ZONE_BOOSTED_WATERMARK
, &zone
->flags
);
3525 wakeup_kswapd(zone
, 0, 0, zone_idx(zone
));
3528 VM_BUG_ON_PAGE(page
&& bad_range(zone
, page
), page
);
3532 local_irq_restore(flags
);
3536 #ifdef CONFIG_FAIL_PAGE_ALLOC
3539 struct fault_attr attr
;
3541 bool ignore_gfp_highmem
;
3542 bool ignore_gfp_reclaim
;
3544 } fail_page_alloc
= {
3545 .attr
= FAULT_ATTR_INITIALIZER
,
3546 .ignore_gfp_reclaim
= true,
3547 .ignore_gfp_highmem
= true,
3551 static int __init
setup_fail_page_alloc(char *str
)
3553 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
3555 __setup("fail_page_alloc=", setup_fail_page_alloc
);
3557 static bool __should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
3559 if (order
< fail_page_alloc
.min_order
)
3561 if (gfp_mask
& __GFP_NOFAIL
)
3563 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
3565 if (fail_page_alloc
.ignore_gfp_reclaim
&&
3566 (gfp_mask
& __GFP_DIRECT_RECLAIM
))
3569 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
3572 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3574 static int __init
fail_page_alloc_debugfs(void)
3576 umode_t mode
= S_IFREG
| 0600;
3579 dir
= fault_create_debugfs_attr("fail_page_alloc", NULL
,
3580 &fail_page_alloc
.attr
);
3582 debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
3583 &fail_page_alloc
.ignore_gfp_reclaim
);
3584 debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
3585 &fail_page_alloc
.ignore_gfp_highmem
);
3586 debugfs_create_u32("min-order", mode
, dir
, &fail_page_alloc
.min_order
);
3591 late_initcall(fail_page_alloc_debugfs
);
3593 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3595 #else /* CONFIG_FAIL_PAGE_ALLOC */
3597 static inline bool __should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
3602 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3604 noinline
bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
3606 return __should_fail_alloc_page(gfp_mask
, order
);
3608 ALLOW_ERROR_INJECTION(should_fail_alloc_page
, TRUE
);
3610 static inline long __zone_watermark_unusable_free(struct zone
*z
,
3611 unsigned int order
, unsigned int alloc_flags
)
3613 const bool alloc_harder
= (alloc_flags
& (ALLOC_HARDER
|ALLOC_OOM
));
3614 long unusable_free
= (1 << order
) - 1;
3617 * If the caller does not have rights to ALLOC_HARDER then subtract
3618 * the high-atomic reserves. This will over-estimate the size of the
3619 * atomic reserve but it avoids a search.
3621 if (likely(!alloc_harder
))
3622 unusable_free
+= z
->nr_reserved_highatomic
;
3625 /* If allocation can't use CMA areas don't use free CMA pages */
3626 if (!(alloc_flags
& ALLOC_CMA
))
3627 unusable_free
+= zone_page_state(z
, NR_FREE_CMA_PAGES
);
3630 return unusable_free
;
3634 * Return true if free base pages are above 'mark'. For high-order checks it
3635 * will return true of the order-0 watermark is reached and there is at least
3636 * one free page of a suitable size. Checking now avoids taking the zone lock
3637 * to check in the allocation paths if no pages are free.
3639 bool __zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
3640 int highest_zoneidx
, unsigned int alloc_flags
,
3645 const bool alloc_harder
= (alloc_flags
& (ALLOC_HARDER
|ALLOC_OOM
));
3647 /* free_pages may go negative - that's OK */
3648 free_pages
-= __zone_watermark_unusable_free(z
, order
, alloc_flags
);
3650 if (alloc_flags
& ALLOC_HIGH
)
3653 if (unlikely(alloc_harder
)) {
3655 * OOM victims can try even harder than normal ALLOC_HARDER
3656 * users on the grounds that it's definitely going to be in
3657 * the exit path shortly and free memory. Any allocation it
3658 * makes during the free path will be small and short-lived.
3660 if (alloc_flags
& ALLOC_OOM
)
3667 * Check watermarks for an order-0 allocation request. If these
3668 * are not met, then a high-order request also cannot go ahead
3669 * even if a suitable page happened to be free.
3671 if (free_pages
<= min
+ z
->lowmem_reserve
[highest_zoneidx
])
3674 /* If this is an order-0 request then the watermark is fine */
3678 /* For a high-order request, check at least one suitable page is free */
3679 for (o
= order
; o
< MAX_ORDER
; o
++) {
3680 struct free_area
*area
= &z
->free_area
[o
];
3686 for (mt
= 0; mt
< MIGRATE_PCPTYPES
; mt
++) {
3687 if (!free_area_empty(area
, mt
))
3692 if ((alloc_flags
& ALLOC_CMA
) &&
3693 !free_area_empty(area
, MIGRATE_CMA
)) {
3697 if (alloc_harder
&& !free_area_empty(area
, MIGRATE_HIGHATOMIC
))
3703 bool zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
3704 int highest_zoneidx
, unsigned int alloc_flags
)
3706 return __zone_watermark_ok(z
, order
, mark
, highest_zoneidx
, alloc_flags
,
3707 zone_page_state(z
, NR_FREE_PAGES
));
3710 static inline bool zone_watermark_fast(struct zone
*z
, unsigned int order
,
3711 unsigned long mark
, int highest_zoneidx
,
3712 unsigned int alloc_flags
, gfp_t gfp_mask
)
3716 free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
3719 * Fast check for order-0 only. If this fails then the reserves
3720 * need to be calculated.
3725 fast_free
= free_pages
;
3726 fast_free
-= __zone_watermark_unusable_free(z
, 0, alloc_flags
);
3727 if (fast_free
> mark
+ z
->lowmem_reserve
[highest_zoneidx
])
3731 if (__zone_watermark_ok(z
, order
, mark
, highest_zoneidx
, alloc_flags
,
3735 * Ignore watermark boosting for GFP_ATOMIC order-0 allocations
3736 * when checking the min watermark. The min watermark is the
3737 * point where boosting is ignored so that kswapd is woken up
3738 * when below the low watermark.
3740 if (unlikely(!order
&& (gfp_mask
& __GFP_ATOMIC
) && z
->watermark_boost
3741 && ((alloc_flags
& ALLOC_WMARK_MASK
) == WMARK_MIN
))) {
3742 mark
= z
->_watermark
[WMARK_MIN
];
3743 return __zone_watermark_ok(z
, order
, mark
, highest_zoneidx
,
3744 alloc_flags
, free_pages
);
3750 bool zone_watermark_ok_safe(struct zone
*z
, unsigned int order
,
3751 unsigned long mark
, int highest_zoneidx
)
3753 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
3755 if (z
->percpu_drift_mark
&& free_pages
< z
->percpu_drift_mark
)
3756 free_pages
= zone_page_state_snapshot(z
, NR_FREE_PAGES
);
3758 return __zone_watermark_ok(z
, order
, mark
, highest_zoneidx
, 0,
3763 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
3765 return node_distance(zone_to_nid(local_zone
), zone_to_nid(zone
)) <=
3766 node_reclaim_distance
;
3768 #else /* CONFIG_NUMA */
3769 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
3773 #endif /* CONFIG_NUMA */
3776 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3777 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3778 * premature use of a lower zone may cause lowmem pressure problems that
3779 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3780 * probably too small. It only makes sense to spread allocations to avoid
3781 * fragmentation between the Normal and DMA32 zones.
3783 static inline unsigned int
3784 alloc_flags_nofragment(struct zone
*zone
, gfp_t gfp_mask
)
3786 unsigned int alloc_flags
;
3789 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3792 alloc_flags
= (__force
int) (gfp_mask
& __GFP_KSWAPD_RECLAIM
);
3794 #ifdef CONFIG_ZONE_DMA32
3798 if (zone_idx(zone
) != ZONE_NORMAL
)
3802 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3803 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3804 * on UMA that if Normal is populated then so is DMA32.
3806 BUILD_BUG_ON(ZONE_NORMAL
- ZONE_DMA32
!= 1);
3807 if (nr_online_nodes
> 1 && !populated_zone(--zone
))
3810 alloc_flags
|= ALLOC_NOFRAGMENT
;
3811 #endif /* CONFIG_ZONE_DMA32 */
3815 static inline unsigned int current_alloc_flags(gfp_t gfp_mask
,
3816 unsigned int alloc_flags
)
3819 unsigned int pflags
= current
->flags
;
3821 if (!(pflags
& PF_MEMALLOC_NOCMA
) &&
3822 gfp_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
3823 alloc_flags
|= ALLOC_CMA
;
3830 * get_page_from_freelist goes through the zonelist trying to allocate
3833 static struct page
*
3834 get_page_from_freelist(gfp_t gfp_mask
, unsigned int order
, int alloc_flags
,
3835 const struct alloc_context
*ac
)
3839 struct pglist_data
*last_pgdat_dirty_limit
= NULL
;
3844 * Scan zonelist, looking for a zone with enough free.
3845 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3847 no_fallback
= alloc_flags
& ALLOC_NOFRAGMENT
;
3848 z
= ac
->preferred_zoneref
;
3849 for_next_zone_zonelist_nodemask(zone
, z
, ac
->highest_zoneidx
,
3854 if (cpusets_enabled() &&
3855 (alloc_flags
& ALLOC_CPUSET
) &&
3856 !__cpuset_zone_allowed(zone
, gfp_mask
))
3859 * When allocating a page cache page for writing, we
3860 * want to get it from a node that is within its dirty
3861 * limit, such that no single node holds more than its
3862 * proportional share of globally allowed dirty pages.
3863 * The dirty limits take into account the node's
3864 * lowmem reserves and high watermark so that kswapd
3865 * should be able to balance it without having to
3866 * write pages from its LRU list.
3868 * XXX: For now, allow allocations to potentially
3869 * exceed the per-node dirty limit in the slowpath
3870 * (spread_dirty_pages unset) before going into reclaim,
3871 * which is important when on a NUMA setup the allowed
3872 * nodes are together not big enough to reach the
3873 * global limit. The proper fix for these situations
3874 * will require awareness of nodes in the
3875 * dirty-throttling and the flusher threads.
3877 if (ac
->spread_dirty_pages
) {
3878 if (last_pgdat_dirty_limit
== zone
->zone_pgdat
)
3881 if (!node_dirty_ok(zone
->zone_pgdat
)) {
3882 last_pgdat_dirty_limit
= zone
->zone_pgdat
;
3887 if (no_fallback
&& nr_online_nodes
> 1 &&
3888 zone
!= ac
->preferred_zoneref
->zone
) {
3892 * If moving to a remote node, retry but allow
3893 * fragmenting fallbacks. Locality is more important
3894 * than fragmentation avoidance.
3896 local_nid
= zone_to_nid(ac
->preferred_zoneref
->zone
);
3897 if (zone_to_nid(zone
) != local_nid
) {
3898 alloc_flags
&= ~ALLOC_NOFRAGMENT
;
3903 mark
= wmark_pages(zone
, alloc_flags
& ALLOC_WMARK_MASK
);
3904 if (!zone_watermark_fast(zone
, order
, mark
,
3905 ac
->highest_zoneidx
, alloc_flags
,
3909 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3911 * Watermark failed for this zone, but see if we can
3912 * grow this zone if it contains deferred pages.
3914 if (static_branch_unlikely(&deferred_pages
)) {
3915 if (_deferred_grow_zone(zone
, order
))
3919 /* Checked here to keep the fast path fast */
3920 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
3921 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
3924 if (node_reclaim_mode
== 0 ||
3925 !zone_allows_reclaim(ac
->preferred_zoneref
->zone
, zone
))
3928 ret
= node_reclaim(zone
->zone_pgdat
, gfp_mask
, order
);
3930 case NODE_RECLAIM_NOSCAN
:
3933 case NODE_RECLAIM_FULL
:
3934 /* scanned but unreclaimable */
3937 /* did we reclaim enough */
3938 if (zone_watermark_ok(zone
, order
, mark
,
3939 ac
->highest_zoneidx
, alloc_flags
))
3947 page
= rmqueue(ac
->preferred_zoneref
->zone
, zone
, order
,
3948 gfp_mask
, alloc_flags
, ac
->migratetype
);
3950 prep_new_page(page
, order
, gfp_mask
, alloc_flags
);
3953 * If this is a high-order atomic allocation then check
3954 * if the pageblock should be reserved for the future
3956 if (unlikely(order
&& (alloc_flags
& ALLOC_HARDER
)))
3957 reserve_highatomic_pageblock(page
, zone
, order
);
3961 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3962 /* Try again if zone has deferred pages */
3963 if (static_branch_unlikely(&deferred_pages
)) {
3964 if (_deferred_grow_zone(zone
, order
))
3972 * It's possible on a UMA machine to get through all zones that are
3973 * fragmented. If avoiding fragmentation, reset and try again.
3976 alloc_flags
&= ~ALLOC_NOFRAGMENT
;
3983 static void warn_alloc_show_mem(gfp_t gfp_mask
, nodemask_t
*nodemask
)
3985 unsigned int filter
= SHOW_MEM_FILTER_NODES
;
3988 * This documents exceptions given to allocations in certain
3989 * contexts that are allowed to allocate outside current's set
3992 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
3993 if (tsk_is_oom_victim(current
) ||
3994 (current
->flags
& (PF_MEMALLOC
| PF_EXITING
)))
3995 filter
&= ~SHOW_MEM_FILTER_NODES
;
3996 if (in_interrupt() || !(gfp_mask
& __GFP_DIRECT_RECLAIM
))
3997 filter
&= ~SHOW_MEM_FILTER_NODES
;
3999 show_mem(filter
, nodemask
);
4002 void warn_alloc(gfp_t gfp_mask
, nodemask_t
*nodemask
, const char *fmt
, ...)
4004 struct va_format vaf
;
4006 static DEFINE_RATELIMIT_STATE(nopage_rs
, 10*HZ
, 1);
4008 if ((gfp_mask
& __GFP_NOWARN
) || !__ratelimit(&nopage_rs
))
4011 va_start(args
, fmt
);
4014 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
4015 current
->comm
, &vaf
, gfp_mask
, &gfp_mask
,
4016 nodemask_pr_args(nodemask
));
4019 cpuset_print_current_mems_allowed();
4022 warn_alloc_show_mem(gfp_mask
, nodemask
);
4025 static inline struct page
*
4026 __alloc_pages_cpuset_fallback(gfp_t gfp_mask
, unsigned int order
,
4027 unsigned int alloc_flags
,
4028 const struct alloc_context
*ac
)
4032 page
= get_page_from_freelist(gfp_mask
, order
,
4033 alloc_flags
|ALLOC_CPUSET
, ac
);
4035 * fallback to ignore cpuset restriction if our nodes
4039 page
= get_page_from_freelist(gfp_mask
, order
,
4045 static inline struct page
*
4046 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
4047 const struct alloc_context
*ac
, unsigned long *did_some_progress
)
4049 struct oom_control oc
= {
4050 .zonelist
= ac
->zonelist
,
4051 .nodemask
= ac
->nodemask
,
4053 .gfp_mask
= gfp_mask
,
4058 *did_some_progress
= 0;
4061 * Acquire the oom lock. If that fails, somebody else is
4062 * making progress for us.
4064 if (!mutex_trylock(&oom_lock
)) {
4065 *did_some_progress
= 1;
4066 schedule_timeout_uninterruptible(1);
4071 * Go through the zonelist yet one more time, keep very high watermark
4072 * here, this is only to catch a parallel oom killing, we must fail if
4073 * we're still under heavy pressure. But make sure that this reclaim
4074 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
4075 * allocation which will never fail due to oom_lock already held.
4077 page
= get_page_from_freelist((gfp_mask
| __GFP_HARDWALL
) &
4078 ~__GFP_DIRECT_RECLAIM
, order
,
4079 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
, ac
);
4083 /* Coredumps can quickly deplete all memory reserves */
4084 if (current
->flags
& PF_DUMPCORE
)
4086 /* The OOM killer will not help higher order allocs */
4087 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
4090 * We have already exhausted all our reclaim opportunities without any
4091 * success so it is time to admit defeat. We will skip the OOM killer
4092 * because it is very likely that the caller has a more reasonable
4093 * fallback than shooting a random task.
4095 * The OOM killer may not free memory on a specific node.
4097 if (gfp_mask
& (__GFP_RETRY_MAYFAIL
| __GFP_THISNODE
))
4099 /* The OOM killer does not needlessly kill tasks for lowmem */
4100 if (ac
->highest_zoneidx
< ZONE_NORMAL
)
4102 if (pm_suspended_storage())
4105 * XXX: GFP_NOFS allocations should rather fail than rely on
4106 * other request to make a forward progress.
4107 * We are in an unfortunate situation where out_of_memory cannot
4108 * do much for this context but let's try it to at least get
4109 * access to memory reserved if the current task is killed (see
4110 * out_of_memory). Once filesystems are ready to handle allocation
4111 * failures more gracefully we should just bail out here.
4114 /* Exhausted what can be done so it's blame time */
4115 if (out_of_memory(&oc
) || WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
)) {
4116 *did_some_progress
= 1;
4119 * Help non-failing allocations by giving them access to memory
4122 if (gfp_mask
& __GFP_NOFAIL
)
4123 page
= __alloc_pages_cpuset_fallback(gfp_mask
, order
,
4124 ALLOC_NO_WATERMARKS
, ac
);
4127 mutex_unlock(&oom_lock
);
4132 * Maximum number of compaction retries wit a progress before OOM
4133 * killer is consider as the only way to move forward.
4135 #define MAX_COMPACT_RETRIES 16
4137 #ifdef CONFIG_COMPACTION
4138 /* Try memory compaction for high-order allocations before reclaim */
4139 static struct page
*
4140 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
4141 unsigned int alloc_flags
, const struct alloc_context
*ac
,
4142 enum compact_priority prio
, enum compact_result
*compact_result
)
4144 struct page
*page
= NULL
;
4145 unsigned long pflags
;
4146 unsigned int noreclaim_flag
;
4151 psi_memstall_enter(&pflags
);
4152 noreclaim_flag
= memalloc_noreclaim_save();
4154 *compact_result
= try_to_compact_pages(gfp_mask
, order
, alloc_flags
, ac
,
4157 memalloc_noreclaim_restore(noreclaim_flag
);
4158 psi_memstall_leave(&pflags
);
4161 * At least in one zone compaction wasn't deferred or skipped, so let's
4162 * count a compaction stall
4164 count_vm_event(COMPACTSTALL
);
4166 /* Prep a captured page if available */
4168 prep_new_page(page
, order
, gfp_mask
, alloc_flags
);
4170 /* Try get a page from the freelist if available */
4172 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
4175 struct zone
*zone
= page_zone(page
);
4177 zone
->compact_blockskip_flush
= false;
4178 compaction_defer_reset(zone
, order
, true);
4179 count_vm_event(COMPACTSUCCESS
);
4184 * It's bad if compaction run occurs and fails. The most likely reason
4185 * is that pages exist, but not enough to satisfy watermarks.
4187 count_vm_event(COMPACTFAIL
);
4195 should_compact_retry(struct alloc_context
*ac
, int order
, int alloc_flags
,
4196 enum compact_result compact_result
,
4197 enum compact_priority
*compact_priority
,
4198 int *compaction_retries
)
4200 int max_retries
= MAX_COMPACT_RETRIES
;
4203 int retries
= *compaction_retries
;
4204 enum compact_priority priority
= *compact_priority
;
4209 if (compaction_made_progress(compact_result
))
4210 (*compaction_retries
)++;
4213 * compaction considers all the zone as desperately out of memory
4214 * so it doesn't really make much sense to retry except when the
4215 * failure could be caused by insufficient priority
4217 if (compaction_failed(compact_result
))
4218 goto check_priority
;
4221 * compaction was skipped because there are not enough order-0 pages
4222 * to work with, so we retry only if it looks like reclaim can help.
4224 if (compaction_needs_reclaim(compact_result
)) {
4225 ret
= compaction_zonelist_suitable(ac
, order
, alloc_flags
);
4230 * make sure the compaction wasn't deferred or didn't bail out early
4231 * due to locks contention before we declare that we should give up.
4232 * But the next retry should use a higher priority if allowed, so
4233 * we don't just keep bailing out endlessly.
4235 if (compaction_withdrawn(compact_result
)) {
4236 goto check_priority
;
4240 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
4241 * costly ones because they are de facto nofail and invoke OOM
4242 * killer to move on while costly can fail and users are ready
4243 * to cope with that. 1/4 retries is rather arbitrary but we
4244 * would need much more detailed feedback from compaction to
4245 * make a better decision.
4247 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
4249 if (*compaction_retries
<= max_retries
) {
4255 * Make sure there are attempts at the highest priority if we exhausted
4256 * all retries or failed at the lower priorities.
4259 min_priority
= (order
> PAGE_ALLOC_COSTLY_ORDER
) ?
4260 MIN_COMPACT_COSTLY_PRIORITY
: MIN_COMPACT_PRIORITY
;
4262 if (*compact_priority
> min_priority
) {
4263 (*compact_priority
)--;
4264 *compaction_retries
= 0;
4268 trace_compact_retry(order
, priority
, compact_result
, retries
, max_retries
, ret
);
4272 static inline struct page
*
4273 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
4274 unsigned int alloc_flags
, const struct alloc_context
*ac
,
4275 enum compact_priority prio
, enum compact_result
*compact_result
)
4277 *compact_result
= COMPACT_SKIPPED
;
4282 should_compact_retry(struct alloc_context
*ac
, unsigned int order
, int alloc_flags
,
4283 enum compact_result compact_result
,
4284 enum compact_priority
*compact_priority
,
4285 int *compaction_retries
)
4290 if (!order
|| order
> PAGE_ALLOC_COSTLY_ORDER
)
4294 * There are setups with compaction disabled which would prefer to loop
4295 * inside the allocator rather than hit the oom killer prematurely.
4296 * Let's give them a good hope and keep retrying while the order-0
4297 * watermarks are OK.
4299 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
,
4300 ac
->highest_zoneidx
, ac
->nodemask
) {
4301 if (zone_watermark_ok(zone
, 0, min_wmark_pages(zone
),
4302 ac
->highest_zoneidx
, alloc_flags
))
4307 #endif /* CONFIG_COMPACTION */
4309 #ifdef CONFIG_LOCKDEP
4310 static struct lockdep_map __fs_reclaim_map
=
4311 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map
);
4313 static bool __need_reclaim(gfp_t gfp_mask
)
4315 /* no reclaim without waiting on it */
4316 if (!(gfp_mask
& __GFP_DIRECT_RECLAIM
))
4319 /* this guy won't enter reclaim */
4320 if (current
->flags
& PF_MEMALLOC
)
4323 if (gfp_mask
& __GFP_NOLOCKDEP
)
4329 void __fs_reclaim_acquire(void)
4331 lock_map_acquire(&__fs_reclaim_map
);
4334 void __fs_reclaim_release(void)
4336 lock_map_release(&__fs_reclaim_map
);
4339 void fs_reclaim_acquire(gfp_t gfp_mask
)
4341 gfp_mask
= current_gfp_context(gfp_mask
);
4343 if (__need_reclaim(gfp_mask
)) {
4344 if (gfp_mask
& __GFP_FS
)
4345 __fs_reclaim_acquire();
4347 #ifdef CONFIG_MMU_NOTIFIER
4348 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map
);
4349 lock_map_release(&__mmu_notifier_invalidate_range_start_map
);
4354 EXPORT_SYMBOL_GPL(fs_reclaim_acquire
);
4356 void fs_reclaim_release(gfp_t gfp_mask
)
4358 gfp_mask
= current_gfp_context(gfp_mask
);
4360 if (__need_reclaim(gfp_mask
)) {
4361 if (gfp_mask
& __GFP_FS
)
4362 __fs_reclaim_release();
4365 EXPORT_SYMBOL_GPL(fs_reclaim_release
);
4368 /* Perform direct synchronous page reclaim */
4369 static unsigned long
4370 __perform_reclaim(gfp_t gfp_mask
, unsigned int order
,
4371 const struct alloc_context
*ac
)
4373 unsigned int noreclaim_flag
;
4374 unsigned long pflags
, progress
;
4378 /* We now go into synchronous reclaim */
4379 cpuset_memory_pressure_bump();
4380 psi_memstall_enter(&pflags
);
4381 fs_reclaim_acquire(gfp_mask
);
4382 noreclaim_flag
= memalloc_noreclaim_save();
4384 progress
= try_to_free_pages(ac
->zonelist
, order
, gfp_mask
,
4387 memalloc_noreclaim_restore(noreclaim_flag
);
4388 fs_reclaim_release(gfp_mask
);
4389 psi_memstall_leave(&pflags
);
4396 /* The really slow allocator path where we enter direct reclaim */
4397 static inline struct page
*
4398 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
4399 unsigned int alloc_flags
, const struct alloc_context
*ac
,
4400 unsigned long *did_some_progress
)
4402 struct page
*page
= NULL
;
4403 bool drained
= false;
4405 *did_some_progress
= __perform_reclaim(gfp_mask
, order
, ac
);
4406 if (unlikely(!(*did_some_progress
)))
4410 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
4413 * If an allocation failed after direct reclaim, it could be because
4414 * pages are pinned on the per-cpu lists or in high alloc reserves.
4415 * Shrink them and try again
4417 if (!page
&& !drained
) {
4418 unreserve_highatomic_pageblock(ac
, false);
4419 drain_all_pages(NULL
);
4427 static void wake_all_kswapds(unsigned int order
, gfp_t gfp_mask
,
4428 const struct alloc_context
*ac
)
4432 pg_data_t
*last_pgdat
= NULL
;
4433 enum zone_type highest_zoneidx
= ac
->highest_zoneidx
;
4435 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, highest_zoneidx
,
4437 if (last_pgdat
!= zone
->zone_pgdat
)
4438 wakeup_kswapd(zone
, gfp_mask
, order
, highest_zoneidx
);
4439 last_pgdat
= zone
->zone_pgdat
;
4443 static inline unsigned int
4444 gfp_to_alloc_flags(gfp_t gfp_mask
)
4446 unsigned int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
4449 * __GFP_HIGH is assumed to be the same as ALLOC_HIGH
4450 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4451 * to save two branches.
4453 BUILD_BUG_ON(__GFP_HIGH
!= (__force gfp_t
) ALLOC_HIGH
);
4454 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM
!= (__force gfp_t
) ALLOC_KSWAPD
);
4457 * The caller may dip into page reserves a bit more if the caller
4458 * cannot run direct reclaim, or if the caller has realtime scheduling
4459 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4460 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4462 alloc_flags
|= (__force
int)
4463 (gfp_mask
& (__GFP_HIGH
| __GFP_KSWAPD_RECLAIM
));
4465 if (gfp_mask
& __GFP_ATOMIC
) {
4467 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4468 * if it can't schedule.
4470 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
4471 alloc_flags
|= ALLOC_HARDER
;
4473 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4474 * comment for __cpuset_node_allowed().
4476 alloc_flags
&= ~ALLOC_CPUSET
;
4477 } else if (unlikely(rt_task(current
)) && !in_interrupt())
4478 alloc_flags
|= ALLOC_HARDER
;
4480 alloc_flags
= current_alloc_flags(gfp_mask
, alloc_flags
);
4485 static bool oom_reserves_allowed(struct task_struct
*tsk
)
4487 if (!tsk_is_oom_victim(tsk
))
4491 * !MMU doesn't have oom reaper so give access to memory reserves
4492 * only to the thread with TIF_MEMDIE set
4494 if (!IS_ENABLED(CONFIG_MMU
) && !test_thread_flag(TIF_MEMDIE
))
4501 * Distinguish requests which really need access to full memory
4502 * reserves from oom victims which can live with a portion of it
4504 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask
)
4506 if (unlikely(gfp_mask
& __GFP_NOMEMALLOC
))
4508 if (gfp_mask
& __GFP_MEMALLOC
)
4509 return ALLOC_NO_WATERMARKS
;
4510 if (in_serving_softirq() && (current
->flags
& PF_MEMALLOC
))
4511 return ALLOC_NO_WATERMARKS
;
4512 if (!in_interrupt()) {
4513 if (current
->flags
& PF_MEMALLOC
)
4514 return ALLOC_NO_WATERMARKS
;
4515 else if (oom_reserves_allowed(current
))
4522 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask
)
4524 return !!__gfp_pfmemalloc_flags(gfp_mask
);
4528 * Checks whether it makes sense to retry the reclaim to make a forward progress
4529 * for the given allocation request.
4531 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4532 * without success, or when we couldn't even meet the watermark if we
4533 * reclaimed all remaining pages on the LRU lists.
4535 * Returns true if a retry is viable or false to enter the oom path.
4538 should_reclaim_retry(gfp_t gfp_mask
, unsigned order
,
4539 struct alloc_context
*ac
, int alloc_flags
,
4540 bool did_some_progress
, int *no_progress_loops
)
4547 * Costly allocations might have made a progress but this doesn't mean
4548 * their order will become available due to high fragmentation so
4549 * always increment the no progress counter for them
4551 if (did_some_progress
&& order
<= PAGE_ALLOC_COSTLY_ORDER
)
4552 *no_progress_loops
= 0;
4554 (*no_progress_loops
)++;
4557 * Make sure we converge to OOM if we cannot make any progress
4558 * several times in the row.
4560 if (*no_progress_loops
> MAX_RECLAIM_RETRIES
) {
4561 /* Before OOM, exhaust highatomic_reserve */
4562 return unreserve_highatomic_pageblock(ac
, true);
4566 * Keep reclaiming pages while there is a chance this will lead
4567 * somewhere. If none of the target zones can satisfy our allocation
4568 * request even if all reclaimable pages are considered then we are
4569 * screwed and have to go OOM.
4571 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
,
4572 ac
->highest_zoneidx
, ac
->nodemask
) {
4573 unsigned long available
;
4574 unsigned long reclaimable
;
4575 unsigned long min_wmark
= min_wmark_pages(zone
);
4578 available
= reclaimable
= zone_reclaimable_pages(zone
);
4579 available
+= zone_page_state_snapshot(zone
, NR_FREE_PAGES
);
4582 * Would the allocation succeed if we reclaimed all
4583 * reclaimable pages?
4585 wmark
= __zone_watermark_ok(zone
, order
, min_wmark
,
4586 ac
->highest_zoneidx
, alloc_flags
, available
);
4587 trace_reclaim_retry_zone(z
, order
, reclaimable
,
4588 available
, min_wmark
, *no_progress_loops
, wmark
);
4591 * If we didn't make any progress and have a lot of
4592 * dirty + writeback pages then we should wait for
4593 * an IO to complete to slow down the reclaim and
4594 * prevent from pre mature OOM
4596 if (!did_some_progress
) {
4597 unsigned long write_pending
;
4599 write_pending
= zone_page_state_snapshot(zone
,
4600 NR_ZONE_WRITE_PENDING
);
4602 if (2 * write_pending
> reclaimable
) {
4603 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
4615 * Memory allocation/reclaim might be called from a WQ context and the
4616 * current implementation of the WQ concurrency control doesn't
4617 * recognize that a particular WQ is congested if the worker thread is
4618 * looping without ever sleeping. Therefore we have to do a short sleep
4619 * here rather than calling cond_resched().
4621 if (current
->flags
& PF_WQ_WORKER
)
4622 schedule_timeout_uninterruptible(1);
4629 check_retry_cpuset(int cpuset_mems_cookie
, struct alloc_context
*ac
)
4632 * It's possible that cpuset's mems_allowed and the nodemask from
4633 * mempolicy don't intersect. This should be normally dealt with by
4634 * policy_nodemask(), but it's possible to race with cpuset update in
4635 * such a way the check therein was true, and then it became false
4636 * before we got our cpuset_mems_cookie here.
4637 * This assumes that for all allocations, ac->nodemask can come only
4638 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4639 * when it does not intersect with the cpuset restrictions) or the
4640 * caller can deal with a violated nodemask.
4642 if (cpusets_enabled() && ac
->nodemask
&&
4643 !cpuset_nodemask_valid_mems_allowed(ac
->nodemask
)) {
4644 ac
->nodemask
= NULL
;
4649 * When updating a task's mems_allowed or mempolicy nodemask, it is
4650 * possible to race with parallel threads in such a way that our
4651 * allocation can fail while the mask is being updated. If we are about
4652 * to fail, check if the cpuset changed during allocation and if so,
4655 if (read_mems_allowed_retry(cpuset_mems_cookie
))
4661 static inline struct page
*
4662 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
4663 struct alloc_context
*ac
)
4665 bool can_direct_reclaim
= gfp_mask
& __GFP_DIRECT_RECLAIM
;
4666 const bool costly_order
= order
> PAGE_ALLOC_COSTLY_ORDER
;
4667 struct page
*page
= NULL
;
4668 unsigned int alloc_flags
;
4669 unsigned long did_some_progress
;
4670 enum compact_priority compact_priority
;
4671 enum compact_result compact_result
;
4672 int compaction_retries
;
4673 int no_progress_loops
;
4674 unsigned int cpuset_mems_cookie
;
4678 * We also sanity check to catch abuse of atomic reserves being used by
4679 * callers that are not in atomic context.
4681 if (WARN_ON_ONCE((gfp_mask
& (__GFP_ATOMIC
|__GFP_DIRECT_RECLAIM
)) ==
4682 (__GFP_ATOMIC
|__GFP_DIRECT_RECLAIM
)))
4683 gfp_mask
&= ~__GFP_ATOMIC
;
4686 compaction_retries
= 0;
4687 no_progress_loops
= 0;
4688 compact_priority
= DEF_COMPACT_PRIORITY
;
4689 cpuset_mems_cookie
= read_mems_allowed_begin();
4692 * The fast path uses conservative alloc_flags to succeed only until
4693 * kswapd needs to be woken up, and to avoid the cost of setting up
4694 * alloc_flags precisely. So we do that now.
4696 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
4699 * We need to recalculate the starting point for the zonelist iterator
4700 * because we might have used different nodemask in the fast path, or
4701 * there was a cpuset modification and we are retrying - otherwise we
4702 * could end up iterating over non-eligible zones endlessly.
4704 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
4705 ac
->highest_zoneidx
, ac
->nodemask
);
4706 if (!ac
->preferred_zoneref
->zone
)
4709 if (alloc_flags
& ALLOC_KSWAPD
)
4710 wake_all_kswapds(order
, gfp_mask
, ac
);
4713 * The adjusted alloc_flags might result in immediate success, so try
4716 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
4721 * For costly allocations, try direct compaction first, as it's likely
4722 * that we have enough base pages and don't need to reclaim. For non-
4723 * movable high-order allocations, do that as well, as compaction will
4724 * try prevent permanent fragmentation by migrating from blocks of the
4726 * Don't try this for allocations that are allowed to ignore
4727 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4729 if (can_direct_reclaim
&&
4731 (order
> 0 && ac
->migratetype
!= MIGRATE_MOVABLE
))
4732 && !gfp_pfmemalloc_allowed(gfp_mask
)) {
4733 page
= __alloc_pages_direct_compact(gfp_mask
, order
,
4735 INIT_COMPACT_PRIORITY
,
4741 * Checks for costly allocations with __GFP_NORETRY, which
4742 * includes some THP page fault allocations
4744 if (costly_order
&& (gfp_mask
& __GFP_NORETRY
)) {
4746 * If allocating entire pageblock(s) and compaction
4747 * failed because all zones are below low watermarks
4748 * or is prohibited because it recently failed at this
4749 * order, fail immediately unless the allocator has
4750 * requested compaction and reclaim retry.
4753 * - potentially very expensive because zones are far
4754 * below their low watermarks or this is part of very
4755 * bursty high order allocations,
4756 * - not guaranteed to help because isolate_freepages()
4757 * may not iterate over freed pages as part of its
4759 * - unlikely to make entire pageblocks free on its
4762 if (compact_result
== COMPACT_SKIPPED
||
4763 compact_result
== COMPACT_DEFERRED
)
4767 * Looks like reclaim/compaction is worth trying, but
4768 * sync compaction could be very expensive, so keep
4769 * using async compaction.
4771 compact_priority
= INIT_COMPACT_PRIORITY
;
4776 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4777 if (alloc_flags
& ALLOC_KSWAPD
)
4778 wake_all_kswapds(order
, gfp_mask
, ac
);
4780 reserve_flags
= __gfp_pfmemalloc_flags(gfp_mask
);
4782 alloc_flags
= current_alloc_flags(gfp_mask
, reserve_flags
);
4785 * Reset the nodemask and zonelist iterators if memory policies can be
4786 * ignored. These allocations are high priority and system rather than
4789 if (!(alloc_flags
& ALLOC_CPUSET
) || reserve_flags
) {
4790 ac
->nodemask
= NULL
;
4791 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
4792 ac
->highest_zoneidx
, ac
->nodemask
);
4795 /* Attempt with potentially adjusted zonelist and alloc_flags */
4796 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
4800 /* Caller is not willing to reclaim, we can't balance anything */
4801 if (!can_direct_reclaim
)
4804 /* Avoid recursion of direct reclaim */
4805 if (current
->flags
& PF_MEMALLOC
)
4808 /* Try direct reclaim and then allocating */
4809 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
, alloc_flags
, ac
,
4810 &did_some_progress
);
4814 /* Try direct compaction and then allocating */
4815 page
= __alloc_pages_direct_compact(gfp_mask
, order
, alloc_flags
, ac
,
4816 compact_priority
, &compact_result
);
4820 /* Do not loop if specifically requested */
4821 if (gfp_mask
& __GFP_NORETRY
)
4825 * Do not retry costly high order allocations unless they are
4826 * __GFP_RETRY_MAYFAIL
4828 if (costly_order
&& !(gfp_mask
& __GFP_RETRY_MAYFAIL
))
4831 if (should_reclaim_retry(gfp_mask
, order
, ac
, alloc_flags
,
4832 did_some_progress
> 0, &no_progress_loops
))
4836 * It doesn't make any sense to retry for the compaction if the order-0
4837 * reclaim is not able to make any progress because the current
4838 * implementation of the compaction depends on the sufficient amount
4839 * of free memory (see __compaction_suitable)
4841 if (did_some_progress
> 0 &&
4842 should_compact_retry(ac
, order
, alloc_flags
,
4843 compact_result
, &compact_priority
,
4844 &compaction_retries
))
4848 /* Deal with possible cpuset update races before we start OOM killing */
4849 if (check_retry_cpuset(cpuset_mems_cookie
, ac
))
4852 /* Reclaim has failed us, start killing things */
4853 page
= __alloc_pages_may_oom(gfp_mask
, order
, ac
, &did_some_progress
);
4857 /* Avoid allocations with no watermarks from looping endlessly */
4858 if (tsk_is_oom_victim(current
) &&
4859 (alloc_flags
& ALLOC_OOM
||
4860 (gfp_mask
& __GFP_NOMEMALLOC
)))
4863 /* Retry as long as the OOM killer is making progress */
4864 if (did_some_progress
) {
4865 no_progress_loops
= 0;
4870 /* Deal with possible cpuset update races before we fail */
4871 if (check_retry_cpuset(cpuset_mems_cookie
, ac
))
4875 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4878 if (gfp_mask
& __GFP_NOFAIL
) {
4880 * All existing users of the __GFP_NOFAIL are blockable, so warn
4881 * of any new users that actually require GFP_NOWAIT
4883 if (WARN_ON_ONCE(!can_direct_reclaim
))
4887 * PF_MEMALLOC request from this context is rather bizarre
4888 * because we cannot reclaim anything and only can loop waiting
4889 * for somebody to do a work for us
4891 WARN_ON_ONCE(current
->flags
& PF_MEMALLOC
);
4894 * non failing costly orders are a hard requirement which we
4895 * are not prepared for much so let's warn about these users
4896 * so that we can identify them and convert them to something
4899 WARN_ON_ONCE(order
> PAGE_ALLOC_COSTLY_ORDER
);
4902 * Help non-failing allocations by giving them access to memory
4903 * reserves but do not use ALLOC_NO_WATERMARKS because this
4904 * could deplete whole memory reserves which would just make
4905 * the situation worse
4907 page
= __alloc_pages_cpuset_fallback(gfp_mask
, order
, ALLOC_HARDER
, ac
);
4915 warn_alloc(gfp_mask
, ac
->nodemask
,
4916 "page allocation failure: order:%u", order
);
4921 static inline bool prepare_alloc_pages(gfp_t gfp_mask
, unsigned int order
,
4922 int preferred_nid
, nodemask_t
*nodemask
,
4923 struct alloc_context
*ac
, gfp_t
*alloc_mask
,
4924 unsigned int *alloc_flags
)
4926 ac
->highest_zoneidx
= gfp_zone(gfp_mask
);
4927 ac
->zonelist
= node_zonelist(preferred_nid
, gfp_mask
);
4928 ac
->nodemask
= nodemask
;
4929 ac
->migratetype
= gfp_migratetype(gfp_mask
);
4931 if (cpusets_enabled()) {
4932 *alloc_mask
|= __GFP_HARDWALL
;
4934 * When we are in the interrupt context, it is irrelevant
4935 * to the current task context. It means that any node ok.
4937 if (!in_interrupt() && !ac
->nodemask
)
4938 ac
->nodemask
= &cpuset_current_mems_allowed
;
4940 *alloc_flags
|= ALLOC_CPUSET
;
4943 fs_reclaim_acquire(gfp_mask
);
4944 fs_reclaim_release(gfp_mask
);
4946 might_sleep_if(gfp_mask
& __GFP_DIRECT_RECLAIM
);
4948 if (should_fail_alloc_page(gfp_mask
, order
))
4951 *alloc_flags
= current_alloc_flags(gfp_mask
, *alloc_flags
);
4953 /* Dirty zone balancing only done in the fast path */
4954 ac
->spread_dirty_pages
= (gfp_mask
& __GFP_WRITE
);
4957 * The preferred zone is used for statistics but crucially it is
4958 * also used as the starting point for the zonelist iterator. It
4959 * may get reset for allocations that ignore memory policies.
4961 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
4962 ac
->highest_zoneidx
, ac
->nodemask
);
4968 * This is the 'heart' of the zoned buddy allocator.
4971 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
, int preferred_nid
,
4972 nodemask_t
*nodemask
)
4975 unsigned int alloc_flags
= ALLOC_WMARK_LOW
;
4976 gfp_t alloc_mask
; /* The gfp_t that was actually used for allocation */
4977 struct alloc_context ac
= { };
4980 * There are several places where we assume that the order value is sane
4981 * so bail out early if the request is out of bound.
4983 if (unlikely(order
>= MAX_ORDER
)) {
4984 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
4988 gfp_mask
&= gfp_allowed_mask
;
4989 alloc_mask
= gfp_mask
;
4990 if (!prepare_alloc_pages(gfp_mask
, order
, preferred_nid
, nodemask
, &ac
, &alloc_mask
, &alloc_flags
))
4994 * Forbid the first pass from falling back to types that fragment
4995 * memory until all local zones are considered.
4997 alloc_flags
|= alloc_flags_nofragment(ac
.preferred_zoneref
->zone
, gfp_mask
);
4999 /* First allocation attempt */
5000 page
= get_page_from_freelist(alloc_mask
, order
, alloc_flags
, &ac
);
5005 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
5006 * resp. GFP_NOIO which has to be inherited for all allocation requests
5007 * from a particular context which has been marked by
5008 * memalloc_no{fs,io}_{save,restore}.
5010 alloc_mask
= current_gfp_context(gfp_mask
);
5011 ac
.spread_dirty_pages
= false;
5014 * Restore the original nodemask if it was potentially replaced with
5015 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
5017 ac
.nodemask
= nodemask
;
5019 page
= __alloc_pages_slowpath(alloc_mask
, order
, &ac
);
5022 if (memcg_kmem_enabled() && (gfp_mask
& __GFP_ACCOUNT
) && page
&&
5023 unlikely(__memcg_kmem_charge_page(page
, gfp_mask
, order
) != 0)) {
5024 __free_pages(page
, order
);
5028 trace_mm_page_alloc(page
, order
, alloc_mask
, ac
.migratetype
);
5032 EXPORT_SYMBOL(__alloc_pages_nodemask
);
5035 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
5036 * address cannot represent highmem pages. Use alloc_pages and then kmap if
5037 * you need to access high mem.
5039 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
5043 page
= alloc_pages(gfp_mask
& ~__GFP_HIGHMEM
, order
);
5046 return (unsigned long) page_address(page
);
5048 EXPORT_SYMBOL(__get_free_pages
);
5050 unsigned long get_zeroed_page(gfp_t gfp_mask
)
5052 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
5054 EXPORT_SYMBOL(get_zeroed_page
);
5056 static inline void free_the_page(struct page
*page
, unsigned int order
)
5058 if (order
== 0) /* Via pcp? */
5059 free_unref_page(page
);
5061 __free_pages_ok(page
, order
, FPI_NONE
);
5065 * __free_pages - Free pages allocated with alloc_pages().
5066 * @page: The page pointer returned from alloc_pages().
5067 * @order: The order of the allocation.
5069 * This function can free multi-page allocations that are not compound
5070 * pages. It does not check that the @order passed in matches that of
5071 * the allocation, so it is easy to leak memory. Freeing more memory
5072 * than was allocated will probably emit a warning.
5074 * If the last reference to this page is speculative, it will be released
5075 * by put_page() which only frees the first page of a non-compound
5076 * allocation. To prevent the remaining pages from being leaked, we free
5077 * the subsequent pages here. If you want to use the page's reference
5078 * count to decide when to free the allocation, you should allocate a
5079 * compound page, and use put_page() instead of __free_pages().
5081 * Context: May be called in interrupt context or while holding a normal
5082 * spinlock, but not in NMI context or while holding a raw spinlock.
5084 void __free_pages(struct page
*page
, unsigned int order
)
5086 if (put_page_testzero(page
))
5087 free_the_page(page
, order
);
5088 else if (!PageHead(page
))
5090 free_the_page(page
+ (1 << order
), order
);
5092 EXPORT_SYMBOL(__free_pages
);
5094 void free_pages(unsigned long addr
, unsigned int order
)
5097 VM_BUG_ON(!virt_addr_valid((void *)addr
));
5098 __free_pages(virt_to_page((void *)addr
), order
);
5102 EXPORT_SYMBOL(free_pages
);
5106 * An arbitrary-length arbitrary-offset area of memory which resides
5107 * within a 0 or higher order page. Multiple fragments within that page
5108 * are individually refcounted, in the page's reference counter.
5110 * The page_frag functions below provide a simple allocation framework for
5111 * page fragments. This is used by the network stack and network device
5112 * drivers to provide a backing region of memory for use as either an
5113 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
5115 static struct page
*__page_frag_cache_refill(struct page_frag_cache
*nc
,
5118 struct page
*page
= NULL
;
5119 gfp_t gfp
= gfp_mask
;
5121 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5122 gfp_mask
|= __GFP_COMP
| __GFP_NOWARN
| __GFP_NORETRY
|
5124 page
= alloc_pages_node(NUMA_NO_NODE
, gfp_mask
,
5125 PAGE_FRAG_CACHE_MAX_ORDER
);
5126 nc
->size
= page
? PAGE_FRAG_CACHE_MAX_SIZE
: PAGE_SIZE
;
5128 if (unlikely(!page
))
5129 page
= alloc_pages_node(NUMA_NO_NODE
, gfp
, 0);
5131 nc
->va
= page
? page_address(page
) : NULL
;
5136 void __page_frag_cache_drain(struct page
*page
, unsigned int count
)
5138 VM_BUG_ON_PAGE(page_ref_count(page
) == 0, page
);
5140 if (page_ref_sub_and_test(page
, count
))
5141 free_the_page(page
, compound_order(page
));
5143 EXPORT_SYMBOL(__page_frag_cache_drain
);
5145 void *page_frag_alloc_align(struct page_frag_cache
*nc
,
5146 unsigned int fragsz
, gfp_t gfp_mask
,
5147 unsigned int align_mask
)
5149 unsigned int size
= PAGE_SIZE
;
5153 if (unlikely(!nc
->va
)) {
5155 page
= __page_frag_cache_refill(nc
, gfp_mask
);
5159 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5160 /* if size can vary use size else just use PAGE_SIZE */
5163 /* Even if we own the page, we do not use atomic_set().
5164 * This would break get_page_unless_zero() users.
5166 page_ref_add(page
, PAGE_FRAG_CACHE_MAX_SIZE
);
5168 /* reset page count bias and offset to start of new frag */
5169 nc
->pfmemalloc
= page_is_pfmemalloc(page
);
5170 nc
->pagecnt_bias
= PAGE_FRAG_CACHE_MAX_SIZE
+ 1;
5174 offset
= nc
->offset
- fragsz
;
5175 if (unlikely(offset
< 0)) {
5176 page
= virt_to_page(nc
->va
);
5178 if (!page_ref_sub_and_test(page
, nc
->pagecnt_bias
))
5181 if (unlikely(nc
->pfmemalloc
)) {
5182 free_the_page(page
, compound_order(page
));
5186 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5187 /* if size can vary use size else just use PAGE_SIZE */
5190 /* OK, page count is 0, we can safely set it */
5191 set_page_count(page
, PAGE_FRAG_CACHE_MAX_SIZE
+ 1);
5193 /* reset page count bias and offset to start of new frag */
5194 nc
->pagecnt_bias
= PAGE_FRAG_CACHE_MAX_SIZE
+ 1;
5195 offset
= size
- fragsz
;
5199 offset
&= align_mask
;
5200 nc
->offset
= offset
;
5202 return nc
->va
+ offset
;
5204 EXPORT_SYMBOL(page_frag_alloc_align
);
5207 * Frees a page fragment allocated out of either a compound or order 0 page.
5209 void page_frag_free(void *addr
)
5211 struct page
*page
= virt_to_head_page(addr
);
5213 if (unlikely(put_page_testzero(page
)))
5214 free_the_page(page
, compound_order(page
));
5216 EXPORT_SYMBOL(page_frag_free
);
5218 static void *make_alloc_exact(unsigned long addr
, unsigned int order
,
5222 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
5223 unsigned long used
= addr
+ PAGE_ALIGN(size
);
5225 split_page(virt_to_page((void *)addr
), order
);
5226 while (used
< alloc_end
) {
5231 return (void *)addr
;
5235 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
5236 * @size: the number of bytes to allocate
5237 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5239 * This function is similar to alloc_pages(), except that it allocates the
5240 * minimum number of pages to satisfy the request. alloc_pages() can only
5241 * allocate memory in power-of-two pages.
5243 * This function is also limited by MAX_ORDER.
5245 * Memory allocated by this function must be released by free_pages_exact().
5247 * Return: pointer to the allocated area or %NULL in case of error.
5249 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
5251 unsigned int order
= get_order(size
);
5254 if (WARN_ON_ONCE(gfp_mask
& __GFP_COMP
))
5255 gfp_mask
&= ~__GFP_COMP
;
5257 addr
= __get_free_pages(gfp_mask
, order
);
5258 return make_alloc_exact(addr
, order
, size
);
5260 EXPORT_SYMBOL(alloc_pages_exact
);
5263 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5265 * @nid: the preferred node ID where memory should be allocated
5266 * @size: the number of bytes to allocate
5267 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5269 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5272 * Return: pointer to the allocated area or %NULL in case of error.
5274 void * __meminit
alloc_pages_exact_nid(int nid
, size_t size
, gfp_t gfp_mask
)
5276 unsigned int order
= get_order(size
);
5279 if (WARN_ON_ONCE(gfp_mask
& __GFP_COMP
))
5280 gfp_mask
&= ~__GFP_COMP
;
5282 p
= alloc_pages_node(nid
, gfp_mask
, order
);
5285 return make_alloc_exact((unsigned long)page_address(p
), order
, size
);
5289 * free_pages_exact - release memory allocated via alloc_pages_exact()
5290 * @virt: the value returned by alloc_pages_exact.
5291 * @size: size of allocation, same value as passed to alloc_pages_exact().
5293 * Release the memory allocated by a previous call to alloc_pages_exact.
5295 void free_pages_exact(void *virt
, size_t size
)
5297 unsigned long addr
= (unsigned long)virt
;
5298 unsigned long end
= addr
+ PAGE_ALIGN(size
);
5300 while (addr
< end
) {
5305 EXPORT_SYMBOL(free_pages_exact
);
5308 * nr_free_zone_pages - count number of pages beyond high watermark
5309 * @offset: The zone index of the highest zone
5311 * nr_free_zone_pages() counts the number of pages which are beyond the
5312 * high watermark within all zones at or below a given zone index. For each
5313 * zone, the number of pages is calculated as:
5315 * nr_free_zone_pages = managed_pages - high_pages
5317 * Return: number of pages beyond high watermark.
5319 static unsigned long nr_free_zone_pages(int offset
)
5324 /* Just pick one node, since fallback list is circular */
5325 unsigned long sum
= 0;
5327 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
5329 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
5330 unsigned long size
= zone_managed_pages(zone
);
5331 unsigned long high
= high_wmark_pages(zone
);
5340 * nr_free_buffer_pages - count number of pages beyond high watermark
5342 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5343 * watermark within ZONE_DMA and ZONE_NORMAL.
5345 * Return: number of pages beyond high watermark within ZONE_DMA and
5348 unsigned long nr_free_buffer_pages(void)
5350 return nr_free_zone_pages(gfp_zone(GFP_USER
));
5352 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
5354 static inline void show_node(struct zone
*zone
)
5356 if (IS_ENABLED(CONFIG_NUMA
))
5357 printk("Node %d ", zone_to_nid(zone
));
5360 long si_mem_available(void)
5363 unsigned long pagecache
;
5364 unsigned long wmark_low
= 0;
5365 unsigned long pages
[NR_LRU_LISTS
];
5366 unsigned long reclaimable
;
5370 for (lru
= LRU_BASE
; lru
< NR_LRU_LISTS
; lru
++)
5371 pages
[lru
] = global_node_page_state(NR_LRU_BASE
+ lru
);
5374 wmark_low
+= low_wmark_pages(zone
);
5377 * Estimate the amount of memory available for userspace allocations,
5378 * without causing swapping.
5380 available
= global_zone_page_state(NR_FREE_PAGES
) - totalreserve_pages
;
5383 * Not all the page cache can be freed, otherwise the system will
5384 * start swapping. Assume at least half of the page cache, or the
5385 * low watermark worth of cache, needs to stay.
5387 pagecache
= pages
[LRU_ACTIVE_FILE
] + pages
[LRU_INACTIVE_FILE
];
5388 pagecache
-= min(pagecache
/ 2, wmark_low
);
5389 available
+= pagecache
;
5392 * Part of the reclaimable slab and other kernel memory consists of
5393 * items that are in use, and cannot be freed. Cap this estimate at the
5396 reclaimable
= global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B
) +
5397 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE
);
5398 available
+= reclaimable
- min(reclaimable
/ 2, wmark_low
);
5404 EXPORT_SYMBOL_GPL(si_mem_available
);
5406 void si_meminfo(struct sysinfo
*val
)
5408 val
->totalram
= totalram_pages();
5409 val
->sharedram
= global_node_page_state(NR_SHMEM
);
5410 val
->freeram
= global_zone_page_state(NR_FREE_PAGES
);
5411 val
->bufferram
= nr_blockdev_pages();
5412 val
->totalhigh
= totalhigh_pages();
5413 val
->freehigh
= nr_free_highpages();
5414 val
->mem_unit
= PAGE_SIZE
;
5417 EXPORT_SYMBOL(si_meminfo
);
5420 void si_meminfo_node(struct sysinfo
*val
, int nid
)
5422 int zone_type
; /* needs to be signed */
5423 unsigned long managed_pages
= 0;
5424 unsigned long managed_highpages
= 0;
5425 unsigned long free_highpages
= 0;
5426 pg_data_t
*pgdat
= NODE_DATA(nid
);
5428 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++)
5429 managed_pages
+= zone_managed_pages(&pgdat
->node_zones
[zone_type
]);
5430 val
->totalram
= managed_pages
;
5431 val
->sharedram
= node_page_state(pgdat
, NR_SHMEM
);
5432 val
->freeram
= sum_zone_node_page_state(nid
, NR_FREE_PAGES
);
5433 #ifdef CONFIG_HIGHMEM
5434 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
5435 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
5437 if (is_highmem(zone
)) {
5438 managed_highpages
+= zone_managed_pages(zone
);
5439 free_highpages
+= zone_page_state(zone
, NR_FREE_PAGES
);
5442 val
->totalhigh
= managed_highpages
;
5443 val
->freehigh
= free_highpages
;
5445 val
->totalhigh
= managed_highpages
;
5446 val
->freehigh
= free_highpages
;
5448 val
->mem_unit
= PAGE_SIZE
;
5453 * Determine whether the node should be displayed or not, depending on whether
5454 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5456 static bool show_mem_node_skip(unsigned int flags
, int nid
, nodemask_t
*nodemask
)
5458 if (!(flags
& SHOW_MEM_FILTER_NODES
))
5462 * no node mask - aka implicit memory numa policy. Do not bother with
5463 * the synchronization - read_mems_allowed_begin - because we do not
5464 * have to be precise here.
5467 nodemask
= &cpuset_current_mems_allowed
;
5469 return !node_isset(nid
, *nodemask
);
5472 #define K(x) ((x) << (PAGE_SHIFT-10))
5474 static void show_migration_types(unsigned char type
)
5476 static const char types
[MIGRATE_TYPES
] = {
5477 [MIGRATE_UNMOVABLE
] = 'U',
5478 [MIGRATE_MOVABLE
] = 'M',
5479 [MIGRATE_RECLAIMABLE
] = 'E',
5480 [MIGRATE_HIGHATOMIC
] = 'H',
5482 [MIGRATE_CMA
] = 'C',
5484 #ifdef CONFIG_MEMORY_ISOLATION
5485 [MIGRATE_ISOLATE
] = 'I',
5488 char tmp
[MIGRATE_TYPES
+ 1];
5492 for (i
= 0; i
< MIGRATE_TYPES
; i
++) {
5493 if (type
& (1 << i
))
5498 printk(KERN_CONT
"(%s) ", tmp
);
5502 * Show free area list (used inside shift_scroll-lock stuff)
5503 * We also calculate the percentage fragmentation. We do this by counting the
5504 * memory on each free list with the exception of the first item on the list.
5507 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5510 void show_free_areas(unsigned int filter
, nodemask_t
*nodemask
)
5512 unsigned long free_pcp
= 0;
5517 for_each_populated_zone(zone
) {
5518 if (show_mem_node_skip(filter
, zone_to_nid(zone
), nodemask
))
5521 for_each_online_cpu(cpu
)
5522 free_pcp
+= per_cpu_ptr(zone
->pageset
, cpu
)->pcp
.count
;
5525 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5526 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5527 " unevictable:%lu dirty:%lu writeback:%lu\n"
5528 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5529 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5530 " free:%lu free_pcp:%lu free_cma:%lu\n",
5531 global_node_page_state(NR_ACTIVE_ANON
),
5532 global_node_page_state(NR_INACTIVE_ANON
),
5533 global_node_page_state(NR_ISOLATED_ANON
),
5534 global_node_page_state(NR_ACTIVE_FILE
),
5535 global_node_page_state(NR_INACTIVE_FILE
),
5536 global_node_page_state(NR_ISOLATED_FILE
),
5537 global_node_page_state(NR_UNEVICTABLE
),
5538 global_node_page_state(NR_FILE_DIRTY
),
5539 global_node_page_state(NR_WRITEBACK
),
5540 global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B
),
5541 global_node_page_state_pages(NR_SLAB_UNRECLAIMABLE_B
),
5542 global_node_page_state(NR_FILE_MAPPED
),
5543 global_node_page_state(NR_SHMEM
),
5544 global_node_page_state(NR_PAGETABLE
),
5545 global_zone_page_state(NR_BOUNCE
),
5546 global_zone_page_state(NR_FREE_PAGES
),
5548 global_zone_page_state(NR_FREE_CMA_PAGES
));
5550 for_each_online_pgdat(pgdat
) {
5551 if (show_mem_node_skip(filter
, pgdat
->node_id
, nodemask
))
5555 " active_anon:%lukB"
5556 " inactive_anon:%lukB"
5557 " active_file:%lukB"
5558 " inactive_file:%lukB"
5559 " unevictable:%lukB"
5560 " isolated(anon):%lukB"
5561 " isolated(file):%lukB"
5566 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5568 " shmem_pmdmapped: %lukB"
5571 " writeback_tmp:%lukB"
5572 " kernel_stack:%lukB"
5573 #ifdef CONFIG_SHADOW_CALL_STACK
5574 " shadow_call_stack:%lukB"
5577 " all_unreclaimable? %s"
5580 K(node_page_state(pgdat
, NR_ACTIVE_ANON
)),
5581 K(node_page_state(pgdat
, NR_INACTIVE_ANON
)),
5582 K(node_page_state(pgdat
, NR_ACTIVE_FILE
)),
5583 K(node_page_state(pgdat
, NR_INACTIVE_FILE
)),
5584 K(node_page_state(pgdat
, NR_UNEVICTABLE
)),
5585 K(node_page_state(pgdat
, NR_ISOLATED_ANON
)),
5586 K(node_page_state(pgdat
, NR_ISOLATED_FILE
)),
5587 K(node_page_state(pgdat
, NR_FILE_MAPPED
)),
5588 K(node_page_state(pgdat
, NR_FILE_DIRTY
)),
5589 K(node_page_state(pgdat
, NR_WRITEBACK
)),
5590 K(node_page_state(pgdat
, NR_SHMEM
)),
5591 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5592 K(node_page_state(pgdat
, NR_SHMEM_THPS
)),
5593 K(node_page_state(pgdat
, NR_SHMEM_PMDMAPPED
)),
5594 K(node_page_state(pgdat
, NR_ANON_THPS
)),
5596 K(node_page_state(pgdat
, NR_WRITEBACK_TEMP
)),
5597 node_page_state(pgdat
, NR_KERNEL_STACK_KB
),
5598 #ifdef CONFIG_SHADOW_CALL_STACK
5599 node_page_state(pgdat
, NR_KERNEL_SCS_KB
),
5601 K(node_page_state(pgdat
, NR_PAGETABLE
)),
5602 pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
?
5606 for_each_populated_zone(zone
) {
5609 if (show_mem_node_skip(filter
, zone_to_nid(zone
), nodemask
))
5613 for_each_online_cpu(cpu
)
5614 free_pcp
+= per_cpu_ptr(zone
->pageset
, cpu
)->pcp
.count
;
5623 " reserved_highatomic:%luKB"
5624 " active_anon:%lukB"
5625 " inactive_anon:%lukB"
5626 " active_file:%lukB"
5627 " inactive_file:%lukB"
5628 " unevictable:%lukB"
5629 " writepending:%lukB"
5639 K(zone_page_state(zone
, NR_FREE_PAGES
)),
5640 K(min_wmark_pages(zone
)),
5641 K(low_wmark_pages(zone
)),
5642 K(high_wmark_pages(zone
)),
5643 K(zone
->nr_reserved_highatomic
),
5644 K(zone_page_state(zone
, NR_ZONE_ACTIVE_ANON
)),
5645 K(zone_page_state(zone
, NR_ZONE_INACTIVE_ANON
)),
5646 K(zone_page_state(zone
, NR_ZONE_ACTIVE_FILE
)),
5647 K(zone_page_state(zone
, NR_ZONE_INACTIVE_FILE
)),
5648 K(zone_page_state(zone
, NR_ZONE_UNEVICTABLE
)),
5649 K(zone_page_state(zone
, NR_ZONE_WRITE_PENDING
)),
5650 K(zone
->present_pages
),
5651 K(zone_managed_pages(zone
)),
5652 K(zone_page_state(zone
, NR_MLOCK
)),
5653 K(zone_page_state(zone
, NR_BOUNCE
)),
5655 K(this_cpu_read(zone
->pageset
->pcp
.count
)),
5656 K(zone_page_state(zone
, NR_FREE_CMA_PAGES
)));
5657 printk("lowmem_reserve[]:");
5658 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
5659 printk(KERN_CONT
" %ld", zone
->lowmem_reserve
[i
]);
5660 printk(KERN_CONT
"\n");
5663 for_each_populated_zone(zone
) {
5665 unsigned long nr
[MAX_ORDER
], flags
, total
= 0;
5666 unsigned char types
[MAX_ORDER
];
5668 if (show_mem_node_skip(filter
, zone_to_nid(zone
), nodemask
))
5671 printk(KERN_CONT
"%s: ", zone
->name
);
5673 spin_lock_irqsave(&zone
->lock
, flags
);
5674 for (order
= 0; order
< MAX_ORDER
; order
++) {
5675 struct free_area
*area
= &zone
->free_area
[order
];
5678 nr
[order
] = area
->nr_free
;
5679 total
+= nr
[order
] << order
;
5682 for (type
= 0; type
< MIGRATE_TYPES
; type
++) {
5683 if (!free_area_empty(area
, type
))
5684 types
[order
] |= 1 << type
;
5687 spin_unlock_irqrestore(&zone
->lock
, flags
);
5688 for (order
= 0; order
< MAX_ORDER
; order
++) {
5689 printk(KERN_CONT
"%lu*%lukB ",
5690 nr
[order
], K(1UL) << order
);
5692 show_migration_types(types
[order
]);
5694 printk(KERN_CONT
"= %lukB\n", K(total
));
5697 hugetlb_show_meminfo();
5699 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES
));
5701 show_swap_cache_info();
5704 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
5706 zoneref
->zone
= zone
;
5707 zoneref
->zone_idx
= zone_idx(zone
);
5711 * Builds allocation fallback zone lists.
5713 * Add all populated zones of a node to the zonelist.
5715 static int build_zonerefs_node(pg_data_t
*pgdat
, struct zoneref
*zonerefs
)
5718 enum zone_type zone_type
= MAX_NR_ZONES
;
5723 zone
= pgdat
->node_zones
+ zone_type
;
5724 if (managed_zone(zone
)) {
5725 zoneref_set_zone(zone
, &zonerefs
[nr_zones
++]);
5726 check_highest_zone(zone_type
);
5728 } while (zone_type
);
5735 static int __parse_numa_zonelist_order(char *s
)
5738 * We used to support different zonlists modes but they turned
5739 * out to be just not useful. Let's keep the warning in place
5740 * if somebody still use the cmd line parameter so that we do
5741 * not fail it silently
5743 if (!(*s
== 'd' || *s
== 'D' || *s
== 'n' || *s
== 'N')) {
5744 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s
);
5750 char numa_zonelist_order
[] = "Node";
5753 * sysctl handler for numa_zonelist_order
5755 int numa_zonelist_order_handler(struct ctl_table
*table
, int write
,
5756 void *buffer
, size_t *length
, loff_t
*ppos
)
5759 return __parse_numa_zonelist_order(buffer
);
5760 return proc_dostring(table
, write
, buffer
, length
, ppos
);
5764 #define MAX_NODE_LOAD (nr_online_nodes)
5765 static int node_load
[MAX_NUMNODES
];
5768 * find_next_best_node - find the next node that should appear in a given node's fallback list
5769 * @node: node whose fallback list we're appending
5770 * @used_node_mask: nodemask_t of already used nodes
5772 * We use a number of factors to determine which is the next node that should
5773 * appear on a given node's fallback list. The node should not have appeared
5774 * already in @node's fallback list, and it should be the next closest node
5775 * according to the distance array (which contains arbitrary distance values
5776 * from each node to each node in the system), and should also prefer nodes
5777 * with no CPUs, since presumably they'll have very little allocation pressure
5778 * on them otherwise.
5780 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5782 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
5785 int min_val
= INT_MAX
;
5786 int best_node
= NUMA_NO_NODE
;
5788 /* Use the local node if we haven't already */
5789 if (!node_isset(node
, *used_node_mask
)) {
5790 node_set(node
, *used_node_mask
);
5794 for_each_node_state(n
, N_MEMORY
) {
5796 /* Don't want a node to appear more than once */
5797 if (node_isset(n
, *used_node_mask
))
5800 /* Use the distance array to find the distance */
5801 val
= node_distance(node
, n
);
5803 /* Penalize nodes under us ("prefer the next node") */
5806 /* Give preference to headless and unused nodes */
5807 if (!cpumask_empty(cpumask_of_node(n
)))
5808 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
5810 /* Slight preference for less loaded node */
5811 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
5812 val
+= node_load
[n
];
5814 if (val
< min_val
) {
5821 node_set(best_node
, *used_node_mask
);
5828 * Build zonelists ordered by node and zones within node.
5829 * This results in maximum locality--normal zone overflows into local
5830 * DMA zone, if any--but risks exhausting DMA zone.
5832 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int *node_order
,
5835 struct zoneref
*zonerefs
;
5838 zonerefs
= pgdat
->node_zonelists
[ZONELIST_FALLBACK
]._zonerefs
;
5840 for (i
= 0; i
< nr_nodes
; i
++) {
5843 pg_data_t
*node
= NODE_DATA(node_order
[i
]);
5845 nr_zones
= build_zonerefs_node(node
, zonerefs
);
5846 zonerefs
+= nr_zones
;
5848 zonerefs
->zone
= NULL
;
5849 zonerefs
->zone_idx
= 0;
5853 * Build gfp_thisnode zonelists
5855 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
5857 struct zoneref
*zonerefs
;
5860 zonerefs
= pgdat
->node_zonelists
[ZONELIST_NOFALLBACK
]._zonerefs
;
5861 nr_zones
= build_zonerefs_node(pgdat
, zonerefs
);
5862 zonerefs
+= nr_zones
;
5863 zonerefs
->zone
= NULL
;
5864 zonerefs
->zone_idx
= 0;
5868 * Build zonelists ordered by zone and nodes within zones.
5869 * This results in conserving DMA zone[s] until all Normal memory is
5870 * exhausted, but results in overflowing to remote node while memory
5871 * may still exist in local DMA zone.
5874 static void build_zonelists(pg_data_t
*pgdat
)
5876 static int node_order
[MAX_NUMNODES
];
5877 int node
, load
, nr_nodes
= 0;
5878 nodemask_t used_mask
= NODE_MASK_NONE
;
5879 int local_node
, prev_node
;
5881 /* NUMA-aware ordering of nodes */
5882 local_node
= pgdat
->node_id
;
5883 load
= nr_online_nodes
;
5884 prev_node
= local_node
;
5886 memset(node_order
, 0, sizeof(node_order
));
5887 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
5889 * We don't want to pressure a particular node.
5890 * So adding penalty to the first node in same
5891 * distance group to make it round-robin.
5893 if (node_distance(local_node
, node
) !=
5894 node_distance(local_node
, prev_node
))
5895 node_load
[node
] = load
;
5897 node_order
[nr_nodes
++] = node
;
5902 build_zonelists_in_node_order(pgdat
, node_order
, nr_nodes
);
5903 build_thisnode_zonelists(pgdat
);
5906 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5908 * Return node id of node used for "local" allocations.
5909 * I.e., first node id of first zone in arg node's generic zonelist.
5910 * Used for initializing percpu 'numa_mem', which is used primarily
5911 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5913 int local_memory_node(int node
)
5917 z
= first_zones_zonelist(node_zonelist(node
, GFP_KERNEL
),
5918 gfp_zone(GFP_KERNEL
),
5920 return zone_to_nid(z
->zone
);
5924 static void setup_min_unmapped_ratio(void);
5925 static void setup_min_slab_ratio(void);
5926 #else /* CONFIG_NUMA */
5928 static void build_zonelists(pg_data_t
*pgdat
)
5930 int node
, local_node
;
5931 struct zoneref
*zonerefs
;
5934 local_node
= pgdat
->node_id
;
5936 zonerefs
= pgdat
->node_zonelists
[ZONELIST_FALLBACK
]._zonerefs
;
5937 nr_zones
= build_zonerefs_node(pgdat
, zonerefs
);
5938 zonerefs
+= nr_zones
;
5941 * Now we build the zonelist so that it contains the zones
5942 * of all the other nodes.
5943 * We don't want to pressure a particular node, so when
5944 * building the zones for node N, we make sure that the
5945 * zones coming right after the local ones are those from
5946 * node N+1 (modulo N)
5948 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
5949 if (!node_online(node
))
5951 nr_zones
= build_zonerefs_node(NODE_DATA(node
), zonerefs
);
5952 zonerefs
+= nr_zones
;
5954 for (node
= 0; node
< local_node
; node
++) {
5955 if (!node_online(node
))
5957 nr_zones
= build_zonerefs_node(NODE_DATA(node
), zonerefs
);
5958 zonerefs
+= nr_zones
;
5961 zonerefs
->zone
= NULL
;
5962 zonerefs
->zone_idx
= 0;
5965 #endif /* CONFIG_NUMA */
5968 * Boot pageset table. One per cpu which is going to be used for all
5969 * zones and all nodes. The parameters will be set in such a way
5970 * that an item put on a list will immediately be handed over to
5971 * the buddy list. This is safe since pageset manipulation is done
5972 * with interrupts disabled.
5974 * The boot_pagesets must be kept even after bootup is complete for
5975 * unused processors and/or zones. They do play a role for bootstrapping
5976 * hotplugged processors.
5978 * zoneinfo_show() and maybe other functions do
5979 * not check if the processor is online before following the pageset pointer.
5980 * Other parts of the kernel may not check if the zone is available.
5982 static void pageset_init(struct per_cpu_pageset
*p
);
5983 /* These effectively disable the pcplists in the boot pageset completely */
5984 #define BOOT_PAGESET_HIGH 0
5985 #define BOOT_PAGESET_BATCH 1
5986 static DEFINE_PER_CPU(struct per_cpu_pageset
, boot_pageset
);
5987 static DEFINE_PER_CPU(struct per_cpu_nodestat
, boot_nodestats
);
5989 static void __build_all_zonelists(void *data
)
5992 int __maybe_unused cpu
;
5993 pg_data_t
*self
= data
;
5994 static DEFINE_SPINLOCK(lock
);
5999 memset(node_load
, 0, sizeof(node_load
));
6003 * This node is hotadded and no memory is yet present. So just
6004 * building zonelists is fine - no need to touch other nodes.
6006 if (self
&& !node_online(self
->node_id
)) {
6007 build_zonelists(self
);
6009 for_each_online_node(nid
) {
6010 pg_data_t
*pgdat
= NODE_DATA(nid
);
6012 build_zonelists(pgdat
);
6015 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
6017 * We now know the "local memory node" for each node--
6018 * i.e., the node of the first zone in the generic zonelist.
6019 * Set up numa_mem percpu variable for on-line cpus. During
6020 * boot, only the boot cpu should be on-line; we'll init the
6021 * secondary cpus' numa_mem as they come on-line. During
6022 * node/memory hotplug, we'll fixup all on-line cpus.
6024 for_each_online_cpu(cpu
)
6025 set_cpu_numa_mem(cpu
, local_memory_node(cpu_to_node(cpu
)));
6032 static noinline
void __init
6033 build_all_zonelists_init(void)
6037 __build_all_zonelists(NULL
);
6040 * Initialize the boot_pagesets that are going to be used
6041 * for bootstrapping processors. The real pagesets for
6042 * each zone will be allocated later when the per cpu
6043 * allocator is available.
6045 * boot_pagesets are used also for bootstrapping offline
6046 * cpus if the system is already booted because the pagesets
6047 * are needed to initialize allocators on a specific cpu too.
6048 * F.e. the percpu allocator needs the page allocator which
6049 * needs the percpu allocator in order to allocate its pagesets
6050 * (a chicken-egg dilemma).
6052 for_each_possible_cpu(cpu
)
6053 pageset_init(&per_cpu(boot_pageset
, cpu
));
6055 mminit_verify_zonelist();
6056 cpuset_init_current_mems_allowed();
6060 * unless system_state == SYSTEM_BOOTING.
6062 * __ref due to call of __init annotated helper build_all_zonelists_init
6063 * [protected by SYSTEM_BOOTING].
6065 void __ref
build_all_zonelists(pg_data_t
*pgdat
)
6067 unsigned long vm_total_pages
;
6069 if (system_state
== SYSTEM_BOOTING
) {
6070 build_all_zonelists_init();
6072 __build_all_zonelists(pgdat
);
6073 /* cpuset refresh routine should be here */
6075 /* Get the number of free pages beyond high watermark in all zones. */
6076 vm_total_pages
= nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
6078 * Disable grouping by mobility if the number of pages in the
6079 * system is too low to allow the mechanism to work. It would be
6080 * more accurate, but expensive to check per-zone. This check is
6081 * made on memory-hotadd so a system can start with mobility
6082 * disabled and enable it later
6084 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
6085 page_group_by_mobility_disabled
= 1;
6087 page_group_by_mobility_disabled
= 0;
6089 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
6091 page_group_by_mobility_disabled
? "off" : "on",
6094 pr_info("Policy zone: %s\n", zone_names
[policy_zone
]);
6098 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
6099 static bool __meminit
6100 overlap_memmap_init(unsigned long zone
, unsigned long *pfn
)
6102 static struct memblock_region
*r
;
6104 if (mirrored_kernelcore
&& zone
== ZONE_MOVABLE
) {
6105 if (!r
|| *pfn
>= memblock_region_memory_end_pfn(r
)) {
6106 for_each_mem_region(r
) {
6107 if (*pfn
< memblock_region_memory_end_pfn(r
))
6111 if (*pfn
>= memblock_region_memory_base_pfn(r
) &&
6112 memblock_is_mirror(r
)) {
6113 *pfn
= memblock_region_memory_end_pfn(r
);
6121 * Initially all pages are reserved - free ones are freed
6122 * up by memblock_free_all() once the early boot process is
6123 * done. Non-atomic initialization, single-pass.
6125 * All aligned pageblocks are initialized to the specified migratetype
6126 * (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related
6127 * zone stats (e.g., nr_isolate_pageblock) are touched.
6129 void __meminit
memmap_init_range(unsigned long size
, int nid
, unsigned long zone
,
6130 unsigned long start_pfn
, unsigned long zone_end_pfn
,
6131 enum meminit_context context
,
6132 struct vmem_altmap
*altmap
, int migratetype
)
6134 unsigned long pfn
, end_pfn
= start_pfn
+ size
;
6137 if (highest_memmap_pfn
< end_pfn
- 1)
6138 highest_memmap_pfn
= end_pfn
- 1;
6140 #ifdef CONFIG_ZONE_DEVICE
6142 * Honor reservation requested by the driver for this ZONE_DEVICE
6143 * memory. We limit the total number of pages to initialize to just
6144 * those that might contain the memory mapping. We will defer the
6145 * ZONE_DEVICE page initialization until after we have released
6148 if (zone
== ZONE_DEVICE
) {
6152 if (start_pfn
== altmap
->base_pfn
)
6153 start_pfn
+= altmap
->reserve
;
6154 end_pfn
= altmap
->base_pfn
+ vmem_altmap_offset(altmap
);
6158 for (pfn
= start_pfn
; pfn
< end_pfn
; ) {
6160 * There can be holes in boot-time mem_map[]s handed to this
6161 * function. They do not exist on hotplugged memory.
6163 if (context
== MEMINIT_EARLY
) {
6164 if (overlap_memmap_init(zone
, &pfn
))
6166 if (defer_init(nid
, pfn
, zone_end_pfn
))
6170 page
= pfn_to_page(pfn
);
6171 __init_single_page(page
, pfn
, zone
, nid
);
6172 if (context
== MEMINIT_HOTPLUG
)
6173 __SetPageReserved(page
);
6176 * Usually, we want to mark the pageblock MIGRATE_MOVABLE,
6177 * such that unmovable allocations won't be scattered all
6178 * over the place during system boot.
6180 if (IS_ALIGNED(pfn
, pageblock_nr_pages
)) {
6181 set_pageblock_migratetype(page
, migratetype
);
6188 #ifdef CONFIG_ZONE_DEVICE
6189 void __ref
memmap_init_zone_device(struct zone
*zone
,
6190 unsigned long start_pfn
,
6191 unsigned long nr_pages
,
6192 struct dev_pagemap
*pgmap
)
6194 unsigned long pfn
, end_pfn
= start_pfn
+ nr_pages
;
6195 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
6196 struct vmem_altmap
*altmap
= pgmap_altmap(pgmap
);
6197 unsigned long zone_idx
= zone_idx(zone
);
6198 unsigned long start
= jiffies
;
6199 int nid
= pgdat
->node_id
;
6201 if (WARN_ON_ONCE(!pgmap
|| zone_idx(zone
) != ZONE_DEVICE
))
6205 * The call to memmap_init_zone should have already taken care
6206 * of the pages reserved for the memmap, so we can just jump to
6207 * the end of that region and start processing the device pages.
6210 start_pfn
= altmap
->base_pfn
+ vmem_altmap_offset(altmap
);
6211 nr_pages
= end_pfn
- start_pfn
;
6214 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
6215 struct page
*page
= pfn_to_page(pfn
);
6217 __init_single_page(page
, pfn
, zone_idx
, nid
);
6220 * Mark page reserved as it will need to wait for onlining
6221 * phase for it to be fully associated with a zone.
6223 * We can use the non-atomic __set_bit operation for setting
6224 * the flag as we are still initializing the pages.
6226 __SetPageReserved(page
);
6229 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
6230 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
6231 * ever freed or placed on a driver-private list.
6233 page
->pgmap
= pgmap
;
6234 page
->zone_device_data
= NULL
;
6237 * Mark the block movable so that blocks are reserved for
6238 * movable at startup. This will force kernel allocations
6239 * to reserve their blocks rather than leaking throughout
6240 * the address space during boot when many long-lived
6241 * kernel allocations are made.
6243 * Please note that MEMINIT_HOTPLUG path doesn't clear memmap
6244 * because this is done early in section_activate()
6246 if (IS_ALIGNED(pfn
, pageblock_nr_pages
)) {
6247 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
6252 pr_info("%s initialised %lu pages in %ums\n", __func__
,
6253 nr_pages
, jiffies_to_msecs(jiffies
- start
));
6257 static void __meminit
zone_init_free_lists(struct zone
*zone
)
6259 unsigned int order
, t
;
6260 for_each_migratetype_order(order
, t
) {
6261 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
6262 zone
->free_area
[order
].nr_free
= 0;
6266 #if !defined(CONFIG_FLAT_NODE_MEM_MAP)
6268 * Only struct pages that correspond to ranges defined by memblock.memory
6269 * are zeroed and initialized by going through __init_single_page() during
6270 * memmap_init_zone().
6272 * But, there could be struct pages that correspond to holes in
6273 * memblock.memory. This can happen because of the following reasons:
6274 * - physical memory bank size is not necessarily the exact multiple of the
6275 * arbitrary section size
6276 * - early reserved memory may not be listed in memblock.memory
6277 * - memory layouts defined with memmap= kernel parameter may not align
6278 * nicely with memmap sections
6280 * Explicitly initialize those struct pages so that:
6281 * - PG_Reserved is set
6282 * - zone and node links point to zone and node that span the page if the
6283 * hole is in the middle of a zone
6284 * - zone and node links point to adjacent zone/node if the hole falls on
6285 * the zone boundary; the pages in such holes will be prepended to the
6286 * zone/node above the hole except for the trailing pages in the last
6287 * section that will be appended to the zone/node below.
6289 static u64 __meminit
init_unavailable_range(unsigned long spfn
,
6296 for (pfn
= spfn
; pfn
< epfn
; pfn
++) {
6297 if (!pfn_valid(ALIGN_DOWN(pfn
, pageblock_nr_pages
))) {
6298 pfn
= ALIGN_DOWN(pfn
, pageblock_nr_pages
)
6299 + pageblock_nr_pages
- 1;
6302 __init_single_page(pfn_to_page(pfn
), pfn
, zone
, node
);
6303 __SetPageReserved(pfn_to_page(pfn
));
6310 static inline u64
init_unavailable_range(unsigned long spfn
, unsigned long epfn
,
6317 void __meminit __weak
memmap_init_zone(struct zone
*zone
)
6319 unsigned long zone_start_pfn
= zone
->zone_start_pfn
;
6320 unsigned long zone_end_pfn
= zone_start_pfn
+ zone
->spanned_pages
;
6321 int i
, nid
= zone_to_nid(zone
), zone_id
= zone_idx(zone
);
6322 static unsigned long hole_pfn
;
6323 unsigned long start_pfn
, end_pfn
;
6326 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
6327 start_pfn
= clamp(start_pfn
, zone_start_pfn
, zone_end_pfn
);
6328 end_pfn
= clamp(end_pfn
, zone_start_pfn
, zone_end_pfn
);
6330 if (end_pfn
> start_pfn
)
6331 memmap_init_range(end_pfn
- start_pfn
, nid
,
6332 zone_id
, start_pfn
, zone_end_pfn
,
6333 MEMINIT_EARLY
, NULL
, MIGRATE_MOVABLE
);
6335 if (hole_pfn
< start_pfn
)
6336 pgcnt
+= init_unavailable_range(hole_pfn
, start_pfn
,
6341 #ifdef CONFIG_SPARSEMEM
6343 * Initialize the hole in the range [zone_end_pfn, section_end].
6344 * If zone boundary falls in the middle of a section, this hole
6345 * will be re-initialized during the call to this function for the
6348 end_pfn
= round_up(zone_end_pfn
, PAGES_PER_SECTION
);
6349 if (hole_pfn
< end_pfn
)
6350 pgcnt
+= init_unavailable_range(hole_pfn
, end_pfn
,
6355 pr_info(" %s zone: %llu pages in unavailable ranges\n",
6359 static int zone_batchsize(struct zone
*zone
)
6365 * The per-cpu-pages pools are set to around 1000th of the
6368 batch
= zone_managed_pages(zone
) / 1024;
6369 /* But no more than a meg. */
6370 if (batch
* PAGE_SIZE
> 1024 * 1024)
6371 batch
= (1024 * 1024) / PAGE_SIZE
;
6372 batch
/= 4; /* We effectively *= 4 below */
6377 * Clamp the batch to a 2^n - 1 value. Having a power
6378 * of 2 value was found to be more likely to have
6379 * suboptimal cache aliasing properties in some cases.
6381 * For example if 2 tasks are alternately allocating
6382 * batches of pages, one task can end up with a lot
6383 * of pages of one half of the possible page colors
6384 * and the other with pages of the other colors.
6386 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
6391 /* The deferral and batching of frees should be suppressed under NOMMU
6394 * The problem is that NOMMU needs to be able to allocate large chunks
6395 * of contiguous memory as there's no hardware page translation to
6396 * assemble apparent contiguous memory from discontiguous pages.
6398 * Queueing large contiguous runs of pages for batching, however,
6399 * causes the pages to actually be freed in smaller chunks. As there
6400 * can be a significant delay between the individual batches being
6401 * recycled, this leads to the once large chunks of space being
6402 * fragmented and becoming unavailable for high-order allocations.
6409 * pcp->high and pcp->batch values are related and generally batch is lower
6410 * than high. They are also related to pcp->count such that count is lower
6411 * than high, and as soon as it reaches high, the pcplist is flushed.
6413 * However, guaranteeing these relations at all times would require e.g. write
6414 * barriers here but also careful usage of read barriers at the read side, and
6415 * thus be prone to error and bad for performance. Thus the update only prevents
6416 * store tearing. Any new users of pcp->batch and pcp->high should ensure they
6417 * can cope with those fields changing asynchronously, and fully trust only the
6418 * pcp->count field on the local CPU with interrupts disabled.
6420 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6421 * outside of boot time (or some other assurance that no concurrent updaters
6424 static void pageset_update(struct per_cpu_pages
*pcp
, unsigned long high
,
6425 unsigned long batch
)
6427 WRITE_ONCE(pcp
->batch
, batch
);
6428 WRITE_ONCE(pcp
->high
, high
);
6431 static void pageset_init(struct per_cpu_pageset
*p
)
6433 struct per_cpu_pages
*pcp
;
6436 memset(p
, 0, sizeof(*p
));
6439 for (migratetype
= 0; migratetype
< MIGRATE_PCPTYPES
; migratetype
++)
6440 INIT_LIST_HEAD(&pcp
->lists
[migratetype
]);
6443 * Set batch and high values safe for a boot pageset. A true percpu
6444 * pageset's initialization will update them subsequently. Here we don't
6445 * need to be as careful as pageset_update() as nobody can access the
6448 pcp
->high
= BOOT_PAGESET_HIGH
;
6449 pcp
->batch
= BOOT_PAGESET_BATCH
;
6452 static void __zone_set_pageset_high_and_batch(struct zone
*zone
, unsigned long high
,
6453 unsigned long batch
)
6455 struct per_cpu_pageset
*p
;
6458 for_each_possible_cpu(cpu
) {
6459 p
= per_cpu_ptr(zone
->pageset
, cpu
);
6460 pageset_update(&p
->pcp
, high
, batch
);
6465 * Calculate and set new high and batch values for all per-cpu pagesets of a
6466 * zone, based on the zone's size and the percpu_pagelist_fraction sysctl.
6468 static void zone_set_pageset_high_and_batch(struct zone
*zone
)
6470 unsigned long new_high
, new_batch
;
6472 if (percpu_pagelist_fraction
) {
6473 new_high
= zone_managed_pages(zone
) / percpu_pagelist_fraction
;
6474 new_batch
= max(1UL, new_high
/ 4);
6475 if ((new_high
/ 4) > (PAGE_SHIFT
* 8))
6476 new_batch
= PAGE_SHIFT
* 8;
6478 new_batch
= zone_batchsize(zone
);
6479 new_high
= 6 * new_batch
;
6480 new_batch
= max(1UL, 1 * new_batch
);
6483 if (zone
->pageset_high
== new_high
&&
6484 zone
->pageset_batch
== new_batch
)
6487 zone
->pageset_high
= new_high
;
6488 zone
->pageset_batch
= new_batch
;
6490 __zone_set_pageset_high_and_batch(zone
, new_high
, new_batch
);
6493 void __meminit
setup_zone_pageset(struct zone
*zone
)
6495 struct per_cpu_pageset
*p
;
6498 zone
->pageset
= alloc_percpu(struct per_cpu_pageset
);
6499 for_each_possible_cpu(cpu
) {
6500 p
= per_cpu_ptr(zone
->pageset
, cpu
);
6504 zone_set_pageset_high_and_batch(zone
);
6508 * Allocate per cpu pagesets and initialize them.
6509 * Before this call only boot pagesets were available.
6511 void __init
setup_per_cpu_pageset(void)
6513 struct pglist_data
*pgdat
;
6515 int __maybe_unused cpu
;
6517 for_each_populated_zone(zone
)
6518 setup_zone_pageset(zone
);
6522 * Unpopulated zones continue using the boot pagesets.
6523 * The numa stats for these pagesets need to be reset.
6524 * Otherwise, they will end up skewing the stats of
6525 * the nodes these zones are associated with.
6527 for_each_possible_cpu(cpu
) {
6528 struct per_cpu_pageset
*pcp
= &per_cpu(boot_pageset
, cpu
);
6529 memset(pcp
->vm_numa_stat_diff
, 0,
6530 sizeof(pcp
->vm_numa_stat_diff
));
6534 for_each_online_pgdat(pgdat
)
6535 pgdat
->per_cpu_nodestats
=
6536 alloc_percpu(struct per_cpu_nodestat
);
6539 static __meminit
void zone_pcp_init(struct zone
*zone
)
6542 * per cpu subsystem is not up at this point. The following code
6543 * relies on the ability of the linker to provide the
6544 * offset of a (static) per cpu variable into the per cpu area.
6546 zone
->pageset
= &boot_pageset
;
6547 zone
->pageset_high
= BOOT_PAGESET_HIGH
;
6548 zone
->pageset_batch
= BOOT_PAGESET_BATCH
;
6550 if (populated_zone(zone
))
6551 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%u\n",
6552 zone
->name
, zone
->present_pages
,
6553 zone_batchsize(zone
));
6556 void __meminit
init_currently_empty_zone(struct zone
*zone
,
6557 unsigned long zone_start_pfn
,
6560 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
6561 int zone_idx
= zone_idx(zone
) + 1;
6563 if (zone_idx
> pgdat
->nr_zones
)
6564 pgdat
->nr_zones
= zone_idx
;
6566 zone
->zone_start_pfn
= zone_start_pfn
;
6568 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
6569 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
6571 (unsigned long)zone_idx(zone
),
6572 zone_start_pfn
, (zone_start_pfn
+ size
));
6574 zone_init_free_lists(zone
);
6575 zone
->initialized
= 1;
6579 * get_pfn_range_for_nid - Return the start and end page frames for a node
6580 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
6581 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
6582 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
6584 * It returns the start and end page frame of a node based on information
6585 * provided by memblock_set_node(). If called for a node
6586 * with no available memory, a warning is printed and the start and end
6589 void __init
get_pfn_range_for_nid(unsigned int nid
,
6590 unsigned long *start_pfn
, unsigned long *end_pfn
)
6592 unsigned long this_start_pfn
, this_end_pfn
;
6598 for_each_mem_pfn_range(i
, nid
, &this_start_pfn
, &this_end_pfn
, NULL
) {
6599 *start_pfn
= min(*start_pfn
, this_start_pfn
);
6600 *end_pfn
= max(*end_pfn
, this_end_pfn
);
6603 if (*start_pfn
== -1UL)
6608 * This finds a zone that can be used for ZONE_MOVABLE pages. The
6609 * assumption is made that zones within a node are ordered in monotonic
6610 * increasing memory addresses so that the "highest" populated zone is used
6612 static void __init
find_usable_zone_for_movable(void)
6615 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
6616 if (zone_index
== ZONE_MOVABLE
)
6619 if (arch_zone_highest_possible_pfn
[zone_index
] >
6620 arch_zone_lowest_possible_pfn
[zone_index
])
6624 VM_BUG_ON(zone_index
== -1);
6625 movable_zone
= zone_index
;
6629 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
6630 * because it is sized independent of architecture. Unlike the other zones,
6631 * the starting point for ZONE_MOVABLE is not fixed. It may be different
6632 * in each node depending on the size of each node and how evenly kernelcore
6633 * is distributed. This helper function adjusts the zone ranges
6634 * provided by the architecture for a given node by using the end of the
6635 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
6636 * zones within a node are in order of monotonic increases memory addresses
6638 static void __init
adjust_zone_range_for_zone_movable(int nid
,
6639 unsigned long zone_type
,
6640 unsigned long node_start_pfn
,
6641 unsigned long node_end_pfn
,
6642 unsigned long *zone_start_pfn
,
6643 unsigned long *zone_end_pfn
)
6645 /* Only adjust if ZONE_MOVABLE is on this node */
6646 if (zone_movable_pfn
[nid
]) {
6647 /* Size ZONE_MOVABLE */
6648 if (zone_type
== ZONE_MOVABLE
) {
6649 *zone_start_pfn
= zone_movable_pfn
[nid
];
6650 *zone_end_pfn
= min(node_end_pfn
,
6651 arch_zone_highest_possible_pfn
[movable_zone
]);
6653 /* Adjust for ZONE_MOVABLE starting within this range */
6654 } else if (!mirrored_kernelcore
&&
6655 *zone_start_pfn
< zone_movable_pfn
[nid
] &&
6656 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
6657 *zone_end_pfn
= zone_movable_pfn
[nid
];
6659 /* Check if this whole range is within ZONE_MOVABLE */
6660 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
6661 *zone_start_pfn
= *zone_end_pfn
;
6666 * Return the number of pages a zone spans in a node, including holes
6667 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
6669 static unsigned long __init
zone_spanned_pages_in_node(int nid
,
6670 unsigned long zone_type
,
6671 unsigned long node_start_pfn
,
6672 unsigned long node_end_pfn
,
6673 unsigned long *zone_start_pfn
,
6674 unsigned long *zone_end_pfn
)
6676 unsigned long zone_low
= arch_zone_lowest_possible_pfn
[zone_type
];
6677 unsigned long zone_high
= arch_zone_highest_possible_pfn
[zone_type
];
6678 /* When hotadd a new node from cpu_up(), the node should be empty */
6679 if (!node_start_pfn
&& !node_end_pfn
)
6682 /* Get the start and end of the zone */
6683 *zone_start_pfn
= clamp(node_start_pfn
, zone_low
, zone_high
);
6684 *zone_end_pfn
= clamp(node_end_pfn
, zone_low
, zone_high
);
6685 adjust_zone_range_for_zone_movable(nid
, zone_type
,
6686 node_start_pfn
, node_end_pfn
,
6687 zone_start_pfn
, zone_end_pfn
);
6689 /* Check that this node has pages within the zone's required range */
6690 if (*zone_end_pfn
< node_start_pfn
|| *zone_start_pfn
> node_end_pfn
)
6693 /* Move the zone boundaries inside the node if necessary */
6694 *zone_end_pfn
= min(*zone_end_pfn
, node_end_pfn
);
6695 *zone_start_pfn
= max(*zone_start_pfn
, node_start_pfn
);
6697 /* Return the spanned pages */
6698 return *zone_end_pfn
- *zone_start_pfn
;
6702 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
6703 * then all holes in the requested range will be accounted for.
6705 unsigned long __init
__absent_pages_in_range(int nid
,
6706 unsigned long range_start_pfn
,
6707 unsigned long range_end_pfn
)
6709 unsigned long nr_absent
= range_end_pfn
- range_start_pfn
;
6710 unsigned long start_pfn
, end_pfn
;
6713 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
6714 start_pfn
= clamp(start_pfn
, range_start_pfn
, range_end_pfn
);
6715 end_pfn
= clamp(end_pfn
, range_start_pfn
, range_end_pfn
);
6716 nr_absent
-= end_pfn
- start_pfn
;
6722 * absent_pages_in_range - Return number of page frames in holes within a range
6723 * @start_pfn: The start PFN to start searching for holes
6724 * @end_pfn: The end PFN to stop searching for holes
6726 * Return: the number of pages frames in memory holes within a range.
6728 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
6729 unsigned long end_pfn
)
6731 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
6734 /* Return the number of page frames in holes in a zone on a node */
6735 static unsigned long __init
zone_absent_pages_in_node(int nid
,
6736 unsigned long zone_type
,
6737 unsigned long node_start_pfn
,
6738 unsigned long node_end_pfn
)
6740 unsigned long zone_low
= arch_zone_lowest_possible_pfn
[zone_type
];
6741 unsigned long zone_high
= arch_zone_highest_possible_pfn
[zone_type
];
6742 unsigned long zone_start_pfn
, zone_end_pfn
;
6743 unsigned long nr_absent
;
6745 /* When hotadd a new node from cpu_up(), the node should be empty */
6746 if (!node_start_pfn
&& !node_end_pfn
)
6749 zone_start_pfn
= clamp(node_start_pfn
, zone_low
, zone_high
);
6750 zone_end_pfn
= clamp(node_end_pfn
, zone_low
, zone_high
);
6752 adjust_zone_range_for_zone_movable(nid
, zone_type
,
6753 node_start_pfn
, node_end_pfn
,
6754 &zone_start_pfn
, &zone_end_pfn
);
6755 nr_absent
= __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
6758 * ZONE_MOVABLE handling.
6759 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6762 if (mirrored_kernelcore
&& zone_movable_pfn
[nid
]) {
6763 unsigned long start_pfn
, end_pfn
;
6764 struct memblock_region
*r
;
6766 for_each_mem_region(r
) {
6767 start_pfn
= clamp(memblock_region_memory_base_pfn(r
),
6768 zone_start_pfn
, zone_end_pfn
);
6769 end_pfn
= clamp(memblock_region_memory_end_pfn(r
),
6770 zone_start_pfn
, zone_end_pfn
);
6772 if (zone_type
== ZONE_MOVABLE
&&
6773 memblock_is_mirror(r
))
6774 nr_absent
+= end_pfn
- start_pfn
;
6776 if (zone_type
== ZONE_NORMAL
&&
6777 !memblock_is_mirror(r
))
6778 nr_absent
+= end_pfn
- start_pfn
;
6785 static void __init
calculate_node_totalpages(struct pglist_data
*pgdat
,
6786 unsigned long node_start_pfn
,
6787 unsigned long node_end_pfn
)
6789 unsigned long realtotalpages
= 0, totalpages
= 0;
6792 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
6793 struct zone
*zone
= pgdat
->node_zones
+ i
;
6794 unsigned long zone_start_pfn
, zone_end_pfn
;
6795 unsigned long spanned
, absent
;
6796 unsigned long size
, real_size
;
6798 spanned
= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
6803 absent
= zone_absent_pages_in_node(pgdat
->node_id
, i
,
6808 real_size
= size
- absent
;
6811 zone
->zone_start_pfn
= zone_start_pfn
;
6813 zone
->zone_start_pfn
= 0;
6814 zone
->spanned_pages
= size
;
6815 zone
->present_pages
= real_size
;
6818 realtotalpages
+= real_size
;
6821 pgdat
->node_spanned_pages
= totalpages
;
6822 pgdat
->node_present_pages
= realtotalpages
;
6823 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
6827 #ifndef CONFIG_SPARSEMEM
6829 * Calculate the size of the zone->blockflags rounded to an unsigned long
6830 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6831 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6832 * round what is now in bits to nearest long in bits, then return it in
6835 static unsigned long __init
usemap_size(unsigned long zone_start_pfn
, unsigned long zonesize
)
6837 unsigned long usemapsize
;
6839 zonesize
+= zone_start_pfn
& (pageblock_nr_pages
-1);
6840 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
6841 usemapsize
= usemapsize
>> pageblock_order
;
6842 usemapsize
*= NR_PAGEBLOCK_BITS
;
6843 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
6845 return usemapsize
/ 8;
6848 static void __ref
setup_usemap(struct zone
*zone
)
6850 unsigned long usemapsize
= usemap_size(zone
->zone_start_pfn
,
6851 zone
->spanned_pages
);
6852 zone
->pageblock_flags
= NULL
;
6854 zone
->pageblock_flags
=
6855 memblock_alloc_node(usemapsize
, SMP_CACHE_BYTES
,
6857 if (!zone
->pageblock_flags
)
6858 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
6859 usemapsize
, zone
->name
, zone_to_nid(zone
));
6863 static inline void setup_usemap(struct zone
*zone
) {}
6864 #endif /* CONFIG_SPARSEMEM */
6866 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6868 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6869 void __init
set_pageblock_order(void)
6873 /* Check that pageblock_nr_pages has not already been setup */
6874 if (pageblock_order
)
6877 if (HPAGE_SHIFT
> PAGE_SHIFT
)
6878 order
= HUGETLB_PAGE_ORDER
;
6880 order
= MAX_ORDER
- 1;
6883 * Assume the largest contiguous order of interest is a huge page.
6884 * This value may be variable depending on boot parameters on IA64 and
6887 pageblock_order
= order
;
6889 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6892 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6893 * is unused as pageblock_order is set at compile-time. See
6894 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6897 void __init
set_pageblock_order(void)
6901 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6903 static unsigned long __init
calc_memmap_size(unsigned long spanned_pages
,
6904 unsigned long present_pages
)
6906 unsigned long pages
= spanned_pages
;
6909 * Provide a more accurate estimation if there are holes within
6910 * the zone and SPARSEMEM is in use. If there are holes within the
6911 * zone, each populated memory region may cost us one or two extra
6912 * memmap pages due to alignment because memmap pages for each
6913 * populated regions may not be naturally aligned on page boundary.
6914 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6916 if (spanned_pages
> present_pages
+ (present_pages
>> 4) &&
6917 IS_ENABLED(CONFIG_SPARSEMEM
))
6918 pages
= present_pages
;
6920 return PAGE_ALIGN(pages
* sizeof(struct page
)) >> PAGE_SHIFT
;
6923 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6924 static void pgdat_init_split_queue(struct pglist_data
*pgdat
)
6926 struct deferred_split
*ds_queue
= &pgdat
->deferred_split_queue
;
6928 spin_lock_init(&ds_queue
->split_queue_lock
);
6929 INIT_LIST_HEAD(&ds_queue
->split_queue
);
6930 ds_queue
->split_queue_len
= 0;
6933 static void pgdat_init_split_queue(struct pglist_data
*pgdat
) {}
6936 #ifdef CONFIG_COMPACTION
6937 static void pgdat_init_kcompactd(struct pglist_data
*pgdat
)
6939 init_waitqueue_head(&pgdat
->kcompactd_wait
);
6942 static void pgdat_init_kcompactd(struct pglist_data
*pgdat
) {}
6945 static void __meminit
pgdat_init_internals(struct pglist_data
*pgdat
)
6947 pgdat_resize_init(pgdat
);
6949 pgdat_init_split_queue(pgdat
);
6950 pgdat_init_kcompactd(pgdat
);
6952 init_waitqueue_head(&pgdat
->kswapd_wait
);
6953 init_waitqueue_head(&pgdat
->pfmemalloc_wait
);
6955 pgdat_page_ext_init(pgdat
);
6956 lruvec_init(&pgdat
->__lruvec
);
6959 static void __meminit
zone_init_internals(struct zone
*zone
, enum zone_type idx
, int nid
,
6960 unsigned long remaining_pages
)
6962 atomic_long_set(&zone
->managed_pages
, remaining_pages
);
6963 zone_set_nid(zone
, nid
);
6964 zone
->name
= zone_names
[idx
];
6965 zone
->zone_pgdat
= NODE_DATA(nid
);
6966 spin_lock_init(&zone
->lock
);
6967 zone_seqlock_init(zone
);
6968 zone_pcp_init(zone
);
6972 * Set up the zone data structures
6973 * - init pgdat internals
6974 * - init all zones belonging to this node
6976 * NOTE: this function is only called during memory hotplug
6978 #ifdef CONFIG_MEMORY_HOTPLUG
6979 void __ref
free_area_init_core_hotplug(int nid
)
6982 pg_data_t
*pgdat
= NODE_DATA(nid
);
6984 pgdat_init_internals(pgdat
);
6985 for (z
= 0; z
< MAX_NR_ZONES
; z
++)
6986 zone_init_internals(&pgdat
->node_zones
[z
], z
, nid
, 0);
6991 * Set up the zone data structures:
6992 * - mark all pages reserved
6993 * - mark all memory queues empty
6994 * - clear the memory bitmaps
6996 * NOTE: pgdat should get zeroed by caller.
6997 * NOTE: this function is only called during early init.
6999 static void __init
free_area_init_core(struct pglist_data
*pgdat
)
7002 int nid
= pgdat
->node_id
;
7004 pgdat_init_internals(pgdat
);
7005 pgdat
->per_cpu_nodestats
= &boot_nodestats
;
7007 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
7008 struct zone
*zone
= pgdat
->node_zones
+ j
;
7009 unsigned long size
, freesize
, memmap_pages
;
7011 size
= zone
->spanned_pages
;
7012 freesize
= zone
->present_pages
;
7015 * Adjust freesize so that it accounts for how much memory
7016 * is used by this zone for memmap. This affects the watermark
7017 * and per-cpu initialisations
7019 memmap_pages
= calc_memmap_size(size
, freesize
);
7020 if (!is_highmem_idx(j
)) {
7021 if (freesize
>= memmap_pages
) {
7022 freesize
-= memmap_pages
;
7025 " %s zone: %lu pages used for memmap\n",
7026 zone_names
[j
], memmap_pages
);
7028 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
7029 zone_names
[j
], memmap_pages
, freesize
);
7032 /* Account for reserved pages */
7033 if (j
== 0 && freesize
> dma_reserve
) {
7034 freesize
-= dma_reserve
;
7035 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
7036 zone_names
[0], dma_reserve
);
7039 if (!is_highmem_idx(j
))
7040 nr_kernel_pages
+= freesize
;
7041 /* Charge for highmem memmap if there are enough kernel pages */
7042 else if (nr_kernel_pages
> memmap_pages
* 2)
7043 nr_kernel_pages
-= memmap_pages
;
7044 nr_all_pages
+= freesize
;
7047 * Set an approximate value for lowmem here, it will be adjusted
7048 * when the bootmem allocator frees pages into the buddy system.
7049 * And all highmem pages will be managed by the buddy system.
7051 zone_init_internals(zone
, j
, nid
, freesize
);
7056 set_pageblock_order();
7058 init_currently_empty_zone(zone
, zone
->zone_start_pfn
, size
);
7059 memmap_init_zone(zone
);
7063 #ifdef CONFIG_FLAT_NODE_MEM_MAP
7064 static void __ref
alloc_node_mem_map(struct pglist_data
*pgdat
)
7066 unsigned long __maybe_unused start
= 0;
7067 unsigned long __maybe_unused offset
= 0;
7069 /* Skip empty nodes */
7070 if (!pgdat
->node_spanned_pages
)
7073 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
7074 offset
= pgdat
->node_start_pfn
- start
;
7075 /* ia64 gets its own node_mem_map, before this, without bootmem */
7076 if (!pgdat
->node_mem_map
) {
7077 unsigned long size
, end
;
7081 * The zone's endpoints aren't required to be MAX_ORDER
7082 * aligned but the node_mem_map endpoints must be in order
7083 * for the buddy allocator to function correctly.
7085 end
= pgdat_end_pfn(pgdat
);
7086 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
7087 size
= (end
- start
) * sizeof(struct page
);
7088 map
= memblock_alloc_node(size
, SMP_CACHE_BYTES
,
7091 panic("Failed to allocate %ld bytes for node %d memory map\n",
7092 size
, pgdat
->node_id
);
7093 pgdat
->node_mem_map
= map
+ offset
;
7095 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
7096 __func__
, pgdat
->node_id
, (unsigned long)pgdat
,
7097 (unsigned long)pgdat
->node_mem_map
);
7098 #ifndef CONFIG_NEED_MULTIPLE_NODES
7100 * With no DISCONTIG, the global mem_map is just set as node 0's
7102 if (pgdat
== NODE_DATA(0)) {
7103 mem_map
= NODE_DATA(0)->node_mem_map
;
7104 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
7110 static void __ref
alloc_node_mem_map(struct pglist_data
*pgdat
) { }
7111 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
7113 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
7114 static inline void pgdat_set_deferred_range(pg_data_t
*pgdat
)
7116 pgdat
->first_deferred_pfn
= ULONG_MAX
;
7119 static inline void pgdat_set_deferred_range(pg_data_t
*pgdat
) {}
7122 static void __init
free_area_init_node(int nid
)
7124 pg_data_t
*pgdat
= NODE_DATA(nid
);
7125 unsigned long start_pfn
= 0;
7126 unsigned long end_pfn
= 0;
7128 /* pg_data_t should be reset to zero when it's allocated */
7129 WARN_ON(pgdat
->nr_zones
|| pgdat
->kswapd_highest_zoneidx
);
7131 get_pfn_range_for_nid(nid
, &start_pfn
, &end_pfn
);
7133 pgdat
->node_id
= nid
;
7134 pgdat
->node_start_pfn
= start_pfn
;
7135 pgdat
->per_cpu_nodestats
= NULL
;
7137 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid
,
7138 (u64
)start_pfn
<< PAGE_SHIFT
,
7139 end_pfn
? ((u64
)end_pfn
<< PAGE_SHIFT
) - 1 : 0);
7140 calculate_node_totalpages(pgdat
, start_pfn
, end_pfn
);
7142 alloc_node_mem_map(pgdat
);
7143 pgdat_set_deferred_range(pgdat
);
7145 free_area_init_core(pgdat
);
7148 void __init
free_area_init_memoryless_node(int nid
)
7150 free_area_init_node(nid
);
7153 #if MAX_NUMNODES > 1
7155 * Figure out the number of possible node ids.
7157 void __init
setup_nr_node_ids(void)
7159 unsigned int highest
;
7161 highest
= find_last_bit(node_possible_map
.bits
, MAX_NUMNODES
);
7162 nr_node_ids
= highest
+ 1;
7167 * node_map_pfn_alignment - determine the maximum internode alignment
7169 * This function should be called after node map is populated and sorted.
7170 * It calculates the maximum power of two alignment which can distinguish
7173 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7174 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7175 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7176 * shifted, 1GiB is enough and this function will indicate so.
7178 * This is used to test whether pfn -> nid mapping of the chosen memory
7179 * model has fine enough granularity to avoid incorrect mapping for the
7180 * populated node map.
7182 * Return: the determined alignment in pfn's. 0 if there is no alignment
7183 * requirement (single node).
7185 unsigned long __init
node_map_pfn_alignment(void)
7187 unsigned long accl_mask
= 0, last_end
= 0;
7188 unsigned long start
, end
, mask
;
7189 int last_nid
= NUMA_NO_NODE
;
7192 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start
, &end
, &nid
) {
7193 if (!start
|| last_nid
< 0 || last_nid
== nid
) {
7200 * Start with a mask granular enough to pin-point to the
7201 * start pfn and tick off bits one-by-one until it becomes
7202 * too coarse to separate the current node from the last.
7204 mask
= ~((1 << __ffs(start
)) - 1);
7205 while (mask
&& last_end
<= (start
& (mask
<< 1)))
7208 /* accumulate all internode masks */
7212 /* convert mask to number of pages */
7213 return ~accl_mask
+ 1;
7217 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7219 * Return: the minimum PFN based on information provided via
7220 * memblock_set_node().
7222 unsigned long __init
find_min_pfn_with_active_regions(void)
7224 return PHYS_PFN(memblock_start_of_DRAM());
7228 * early_calculate_totalpages()
7229 * Sum pages in active regions for movable zone.
7230 * Populate N_MEMORY for calculating usable_nodes.
7232 static unsigned long __init
early_calculate_totalpages(void)
7234 unsigned long totalpages
= 0;
7235 unsigned long start_pfn
, end_pfn
;
7238 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
7239 unsigned long pages
= end_pfn
- start_pfn
;
7241 totalpages
+= pages
;
7243 node_set_state(nid
, N_MEMORY
);
7249 * Find the PFN the Movable zone begins in each node. Kernel memory
7250 * is spread evenly between nodes as long as the nodes have enough
7251 * memory. When they don't, some nodes will have more kernelcore than
7254 static void __init
find_zone_movable_pfns_for_nodes(void)
7257 unsigned long usable_startpfn
;
7258 unsigned long kernelcore_node
, kernelcore_remaining
;
7259 /* save the state before borrow the nodemask */
7260 nodemask_t saved_node_state
= node_states
[N_MEMORY
];
7261 unsigned long totalpages
= early_calculate_totalpages();
7262 int usable_nodes
= nodes_weight(node_states
[N_MEMORY
]);
7263 struct memblock_region
*r
;
7265 /* Need to find movable_zone earlier when movable_node is specified. */
7266 find_usable_zone_for_movable();
7269 * If movable_node is specified, ignore kernelcore and movablecore
7272 if (movable_node_is_enabled()) {
7273 for_each_mem_region(r
) {
7274 if (!memblock_is_hotpluggable(r
))
7277 nid
= memblock_get_region_node(r
);
7279 usable_startpfn
= PFN_DOWN(r
->base
);
7280 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
7281 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
7289 * If kernelcore=mirror is specified, ignore movablecore option
7291 if (mirrored_kernelcore
) {
7292 bool mem_below_4gb_not_mirrored
= false;
7294 for_each_mem_region(r
) {
7295 if (memblock_is_mirror(r
))
7298 nid
= memblock_get_region_node(r
);
7300 usable_startpfn
= memblock_region_memory_base_pfn(r
);
7302 if (usable_startpfn
< 0x100000) {
7303 mem_below_4gb_not_mirrored
= true;
7307 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
7308 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
7312 if (mem_below_4gb_not_mirrored
)
7313 pr_warn("This configuration results in unmirrored kernel memory.\n");
7319 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7320 * amount of necessary memory.
7322 if (required_kernelcore_percent
)
7323 required_kernelcore
= (totalpages
* 100 * required_kernelcore_percent
) /
7325 if (required_movablecore_percent
)
7326 required_movablecore
= (totalpages
* 100 * required_movablecore_percent
) /
7330 * If movablecore= was specified, calculate what size of
7331 * kernelcore that corresponds so that memory usable for
7332 * any allocation type is evenly spread. If both kernelcore
7333 * and movablecore are specified, then the value of kernelcore
7334 * will be used for required_kernelcore if it's greater than
7335 * what movablecore would have allowed.
7337 if (required_movablecore
) {
7338 unsigned long corepages
;
7341 * Round-up so that ZONE_MOVABLE is at least as large as what
7342 * was requested by the user
7344 required_movablecore
=
7345 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
7346 required_movablecore
= min(totalpages
, required_movablecore
);
7347 corepages
= totalpages
- required_movablecore
;
7349 required_kernelcore
= max(required_kernelcore
, corepages
);
7353 * If kernelcore was not specified or kernelcore size is larger
7354 * than totalpages, there is no ZONE_MOVABLE.
7356 if (!required_kernelcore
|| required_kernelcore
>= totalpages
)
7359 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7360 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
7363 /* Spread kernelcore memory as evenly as possible throughout nodes */
7364 kernelcore_node
= required_kernelcore
/ usable_nodes
;
7365 for_each_node_state(nid
, N_MEMORY
) {
7366 unsigned long start_pfn
, end_pfn
;
7369 * Recalculate kernelcore_node if the division per node
7370 * now exceeds what is necessary to satisfy the requested
7371 * amount of memory for the kernel
7373 if (required_kernelcore
< kernelcore_node
)
7374 kernelcore_node
= required_kernelcore
/ usable_nodes
;
7377 * As the map is walked, we track how much memory is usable
7378 * by the kernel using kernelcore_remaining. When it is
7379 * 0, the rest of the node is usable by ZONE_MOVABLE
7381 kernelcore_remaining
= kernelcore_node
;
7383 /* Go through each range of PFNs within this node */
7384 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
7385 unsigned long size_pages
;
7387 start_pfn
= max(start_pfn
, zone_movable_pfn
[nid
]);
7388 if (start_pfn
>= end_pfn
)
7391 /* Account for what is only usable for kernelcore */
7392 if (start_pfn
< usable_startpfn
) {
7393 unsigned long kernel_pages
;
7394 kernel_pages
= min(end_pfn
, usable_startpfn
)
7397 kernelcore_remaining
-= min(kernel_pages
,
7398 kernelcore_remaining
);
7399 required_kernelcore
-= min(kernel_pages
,
7400 required_kernelcore
);
7402 /* Continue if range is now fully accounted */
7403 if (end_pfn
<= usable_startpfn
) {
7406 * Push zone_movable_pfn to the end so
7407 * that if we have to rebalance
7408 * kernelcore across nodes, we will
7409 * not double account here
7411 zone_movable_pfn
[nid
] = end_pfn
;
7414 start_pfn
= usable_startpfn
;
7418 * The usable PFN range for ZONE_MOVABLE is from
7419 * start_pfn->end_pfn. Calculate size_pages as the
7420 * number of pages used as kernelcore
7422 size_pages
= end_pfn
- start_pfn
;
7423 if (size_pages
> kernelcore_remaining
)
7424 size_pages
= kernelcore_remaining
;
7425 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
7428 * Some kernelcore has been met, update counts and
7429 * break if the kernelcore for this node has been
7432 required_kernelcore
-= min(required_kernelcore
,
7434 kernelcore_remaining
-= size_pages
;
7435 if (!kernelcore_remaining
)
7441 * If there is still required_kernelcore, we do another pass with one
7442 * less node in the count. This will push zone_movable_pfn[nid] further
7443 * along on the nodes that still have memory until kernelcore is
7447 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
7451 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7452 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
7453 zone_movable_pfn
[nid
] =
7454 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
7457 /* restore the node_state */
7458 node_states
[N_MEMORY
] = saved_node_state
;
7461 /* Any regular or high memory on that node ? */
7462 static void check_for_memory(pg_data_t
*pgdat
, int nid
)
7464 enum zone_type zone_type
;
7466 for (zone_type
= 0; zone_type
<= ZONE_MOVABLE
- 1; zone_type
++) {
7467 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
7468 if (populated_zone(zone
)) {
7469 if (IS_ENABLED(CONFIG_HIGHMEM
))
7470 node_set_state(nid
, N_HIGH_MEMORY
);
7471 if (zone_type
<= ZONE_NORMAL
)
7472 node_set_state(nid
, N_NORMAL_MEMORY
);
7479 * Some architecturs, e.g. ARC may have ZONE_HIGHMEM below ZONE_NORMAL. For
7480 * such cases we allow max_zone_pfn sorted in the descending order
7482 bool __weak
arch_has_descending_max_zone_pfns(void)
7488 * free_area_init - Initialise all pg_data_t and zone data
7489 * @max_zone_pfn: an array of max PFNs for each zone
7491 * This will call free_area_init_node() for each active node in the system.
7492 * Using the page ranges provided by memblock_set_node(), the size of each
7493 * zone in each node and their holes is calculated. If the maximum PFN
7494 * between two adjacent zones match, it is assumed that the zone is empty.
7495 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7496 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7497 * starts where the previous one ended. For example, ZONE_DMA32 starts
7498 * at arch_max_dma_pfn.
7500 void __init
free_area_init(unsigned long *max_zone_pfn
)
7502 unsigned long start_pfn
, end_pfn
;
7506 /* Record where the zone boundaries are */
7507 memset(arch_zone_lowest_possible_pfn
, 0,
7508 sizeof(arch_zone_lowest_possible_pfn
));
7509 memset(arch_zone_highest_possible_pfn
, 0,
7510 sizeof(arch_zone_highest_possible_pfn
));
7512 start_pfn
= find_min_pfn_with_active_regions();
7513 descending
= arch_has_descending_max_zone_pfns();
7515 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
7517 zone
= MAX_NR_ZONES
- i
- 1;
7521 if (zone
== ZONE_MOVABLE
)
7524 end_pfn
= max(max_zone_pfn
[zone
], start_pfn
);
7525 arch_zone_lowest_possible_pfn
[zone
] = start_pfn
;
7526 arch_zone_highest_possible_pfn
[zone
] = end_pfn
;
7528 start_pfn
= end_pfn
;
7531 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7532 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
7533 find_zone_movable_pfns_for_nodes();
7535 /* Print out the zone ranges */
7536 pr_info("Zone ranges:\n");
7537 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
7538 if (i
== ZONE_MOVABLE
)
7540 pr_info(" %-8s ", zone_names
[i
]);
7541 if (arch_zone_lowest_possible_pfn
[i
] ==
7542 arch_zone_highest_possible_pfn
[i
])
7545 pr_cont("[mem %#018Lx-%#018Lx]\n",
7546 (u64
)arch_zone_lowest_possible_pfn
[i
]
7548 ((u64
)arch_zone_highest_possible_pfn
[i
]
7549 << PAGE_SHIFT
) - 1);
7552 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
7553 pr_info("Movable zone start for each node\n");
7554 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
7555 if (zone_movable_pfn
[i
])
7556 pr_info(" Node %d: %#018Lx\n", i
,
7557 (u64
)zone_movable_pfn
[i
] << PAGE_SHIFT
);
7561 * Print out the early node map, and initialize the
7562 * subsection-map relative to active online memory ranges to
7563 * enable future "sub-section" extensions of the memory map.
7565 pr_info("Early memory node ranges\n");
7566 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
7567 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid
,
7568 (u64
)start_pfn
<< PAGE_SHIFT
,
7569 ((u64
)end_pfn
<< PAGE_SHIFT
) - 1);
7570 subsection_map_init(start_pfn
, end_pfn
- start_pfn
);
7573 /* Initialise every node */
7574 mminit_verify_pageflags_layout();
7575 setup_nr_node_ids();
7576 for_each_online_node(nid
) {
7577 pg_data_t
*pgdat
= NODE_DATA(nid
);
7578 free_area_init_node(nid
);
7580 /* Any memory on that node */
7581 if (pgdat
->node_present_pages
)
7582 node_set_state(nid
, N_MEMORY
);
7583 check_for_memory(pgdat
, nid
);
7587 static int __init
cmdline_parse_core(char *p
, unsigned long *core
,
7588 unsigned long *percent
)
7590 unsigned long long coremem
;
7596 /* Value may be a percentage of total memory, otherwise bytes */
7597 coremem
= simple_strtoull(p
, &endptr
, 0);
7598 if (*endptr
== '%') {
7599 /* Paranoid check for percent values greater than 100 */
7600 WARN_ON(coremem
> 100);
7604 coremem
= memparse(p
, &p
);
7605 /* Paranoid check that UL is enough for the coremem value */
7606 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
7608 *core
= coremem
>> PAGE_SHIFT
;
7615 * kernelcore=size sets the amount of memory for use for allocations that
7616 * cannot be reclaimed or migrated.
7618 static int __init
cmdline_parse_kernelcore(char *p
)
7620 /* parse kernelcore=mirror */
7621 if (parse_option_str(p
, "mirror")) {
7622 mirrored_kernelcore
= true;
7626 return cmdline_parse_core(p
, &required_kernelcore
,
7627 &required_kernelcore_percent
);
7631 * movablecore=size sets the amount of memory for use for allocations that
7632 * can be reclaimed or migrated.
7634 static int __init
cmdline_parse_movablecore(char *p
)
7636 return cmdline_parse_core(p
, &required_movablecore
,
7637 &required_movablecore_percent
);
7640 early_param("kernelcore", cmdline_parse_kernelcore
);
7641 early_param("movablecore", cmdline_parse_movablecore
);
7643 void adjust_managed_page_count(struct page
*page
, long count
)
7645 atomic_long_add(count
, &page_zone(page
)->managed_pages
);
7646 totalram_pages_add(count
);
7647 #ifdef CONFIG_HIGHMEM
7648 if (PageHighMem(page
))
7649 totalhigh_pages_add(count
);
7652 EXPORT_SYMBOL(adjust_managed_page_count
);
7654 unsigned long free_reserved_area(void *start
, void *end
, int poison
, const char *s
)
7657 unsigned long pages
= 0;
7659 start
= (void *)PAGE_ALIGN((unsigned long)start
);
7660 end
= (void *)((unsigned long)end
& PAGE_MASK
);
7661 for (pos
= start
; pos
< end
; pos
+= PAGE_SIZE
, pages
++) {
7662 struct page
*page
= virt_to_page(pos
);
7663 void *direct_map_addr
;
7666 * 'direct_map_addr' might be different from 'pos'
7667 * because some architectures' virt_to_page()
7668 * work with aliases. Getting the direct map
7669 * address ensures that we get a _writeable_
7670 * alias for the memset().
7672 direct_map_addr
= page_address(page
);
7674 * Perform a kasan-unchecked memset() since this memory
7675 * has not been initialized.
7677 direct_map_addr
= kasan_reset_tag(direct_map_addr
);
7678 if ((unsigned int)poison
<= 0xFF)
7679 memset(direct_map_addr
, poison
, PAGE_SIZE
);
7681 free_reserved_page(page
);
7685 pr_info("Freeing %s memory: %ldK\n",
7686 s
, pages
<< (PAGE_SHIFT
- 10));
7691 void __init
mem_init_print_info(const char *str
)
7693 unsigned long physpages
, codesize
, datasize
, rosize
, bss_size
;
7694 unsigned long init_code_size
, init_data_size
;
7696 physpages
= get_num_physpages();
7697 codesize
= _etext
- _stext
;
7698 datasize
= _edata
- _sdata
;
7699 rosize
= __end_rodata
- __start_rodata
;
7700 bss_size
= __bss_stop
- __bss_start
;
7701 init_data_size
= __init_end
- __init_begin
;
7702 init_code_size
= _einittext
- _sinittext
;
7705 * Detect special cases and adjust section sizes accordingly:
7706 * 1) .init.* may be embedded into .data sections
7707 * 2) .init.text.* may be out of [__init_begin, __init_end],
7708 * please refer to arch/tile/kernel/vmlinux.lds.S.
7709 * 3) .rodata.* may be embedded into .text or .data sections.
7711 #define adj_init_size(start, end, size, pos, adj) \
7713 if (start <= pos && pos < end && size > adj) \
7717 adj_init_size(__init_begin
, __init_end
, init_data_size
,
7718 _sinittext
, init_code_size
);
7719 adj_init_size(_stext
, _etext
, codesize
, _sinittext
, init_code_size
);
7720 adj_init_size(_sdata
, _edata
, datasize
, __init_begin
, init_data_size
);
7721 adj_init_size(_stext
, _etext
, codesize
, __start_rodata
, rosize
);
7722 adj_init_size(_sdata
, _edata
, datasize
, __start_rodata
, rosize
);
7724 #undef adj_init_size
7726 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
7727 #ifdef CONFIG_HIGHMEM
7731 nr_free_pages() << (PAGE_SHIFT
- 10),
7732 physpages
<< (PAGE_SHIFT
- 10),
7733 codesize
>> 10, datasize
>> 10, rosize
>> 10,
7734 (init_data_size
+ init_code_size
) >> 10, bss_size
>> 10,
7735 (physpages
- totalram_pages() - totalcma_pages
) << (PAGE_SHIFT
- 10),
7736 totalcma_pages
<< (PAGE_SHIFT
- 10),
7737 #ifdef CONFIG_HIGHMEM
7738 totalhigh_pages() << (PAGE_SHIFT
- 10),
7740 str
? ", " : "", str
? str
: "");
7744 * set_dma_reserve - set the specified number of pages reserved in the first zone
7745 * @new_dma_reserve: The number of pages to mark reserved
7747 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7748 * In the DMA zone, a significant percentage may be consumed by kernel image
7749 * and other unfreeable allocations which can skew the watermarks badly. This
7750 * function may optionally be used to account for unfreeable pages in the
7751 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7752 * smaller per-cpu batchsize.
7754 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
7756 dma_reserve
= new_dma_reserve
;
7759 static int page_alloc_cpu_dead(unsigned int cpu
)
7762 lru_add_drain_cpu(cpu
);
7766 * Spill the event counters of the dead processor
7767 * into the current processors event counters.
7768 * This artificially elevates the count of the current
7771 vm_events_fold_cpu(cpu
);
7774 * Zero the differential counters of the dead processor
7775 * so that the vm statistics are consistent.
7777 * This is only okay since the processor is dead and cannot
7778 * race with what we are doing.
7780 cpu_vm_stats_fold(cpu
);
7785 int hashdist
= HASHDIST_DEFAULT
;
7787 static int __init
set_hashdist(char *str
)
7791 hashdist
= simple_strtoul(str
, &str
, 0);
7794 __setup("hashdist=", set_hashdist
);
7797 void __init
page_alloc_init(void)
7802 if (num_node_state(N_MEMORY
) == 1)
7806 ret
= cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD
,
7807 "mm/page_alloc:dead", NULL
,
7808 page_alloc_cpu_dead
);
7813 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7814 * or min_free_kbytes changes.
7816 static void calculate_totalreserve_pages(void)
7818 struct pglist_data
*pgdat
;
7819 unsigned long reserve_pages
= 0;
7820 enum zone_type i
, j
;
7822 for_each_online_pgdat(pgdat
) {
7824 pgdat
->totalreserve_pages
= 0;
7826 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
7827 struct zone
*zone
= pgdat
->node_zones
+ i
;
7829 unsigned long managed_pages
= zone_managed_pages(zone
);
7831 /* Find valid and maximum lowmem_reserve in the zone */
7832 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
7833 if (zone
->lowmem_reserve
[j
] > max
)
7834 max
= zone
->lowmem_reserve
[j
];
7837 /* we treat the high watermark as reserved pages. */
7838 max
+= high_wmark_pages(zone
);
7840 if (max
> managed_pages
)
7841 max
= managed_pages
;
7843 pgdat
->totalreserve_pages
+= max
;
7845 reserve_pages
+= max
;
7848 totalreserve_pages
= reserve_pages
;
7852 * setup_per_zone_lowmem_reserve - called whenever
7853 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7854 * has a correct pages reserved value, so an adequate number of
7855 * pages are left in the zone after a successful __alloc_pages().
7857 static void setup_per_zone_lowmem_reserve(void)
7859 struct pglist_data
*pgdat
;
7860 enum zone_type i
, j
;
7862 for_each_online_pgdat(pgdat
) {
7863 for (i
= 0; i
< MAX_NR_ZONES
- 1; i
++) {
7864 struct zone
*zone
= &pgdat
->node_zones
[i
];
7865 int ratio
= sysctl_lowmem_reserve_ratio
[i
];
7866 bool clear
= !ratio
|| !zone_managed_pages(zone
);
7867 unsigned long managed_pages
= 0;
7869 for (j
= i
+ 1; j
< MAX_NR_ZONES
; j
++) {
7871 zone
->lowmem_reserve
[j
] = 0;
7873 struct zone
*upper_zone
= &pgdat
->node_zones
[j
];
7875 managed_pages
+= zone_managed_pages(upper_zone
);
7876 zone
->lowmem_reserve
[j
] = managed_pages
/ ratio
;
7882 /* update totalreserve_pages */
7883 calculate_totalreserve_pages();
7886 static void __setup_per_zone_wmarks(void)
7888 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
7889 unsigned long lowmem_pages
= 0;
7891 unsigned long flags
;
7893 /* Calculate total number of !ZONE_HIGHMEM pages */
7894 for_each_zone(zone
) {
7895 if (!is_highmem(zone
))
7896 lowmem_pages
+= zone_managed_pages(zone
);
7899 for_each_zone(zone
) {
7902 spin_lock_irqsave(&zone
->lock
, flags
);
7903 tmp
= (u64
)pages_min
* zone_managed_pages(zone
);
7904 do_div(tmp
, lowmem_pages
);
7905 if (is_highmem(zone
)) {
7907 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7908 * need highmem pages, so cap pages_min to a small
7911 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7912 * deltas control async page reclaim, and so should
7913 * not be capped for highmem.
7915 unsigned long min_pages
;
7917 min_pages
= zone_managed_pages(zone
) / 1024;
7918 min_pages
= clamp(min_pages
, SWAP_CLUSTER_MAX
, 128UL);
7919 zone
->_watermark
[WMARK_MIN
] = min_pages
;
7922 * If it's a lowmem zone, reserve a number of pages
7923 * proportionate to the zone's size.
7925 zone
->_watermark
[WMARK_MIN
] = tmp
;
7929 * Set the kswapd watermarks distance according to the
7930 * scale factor in proportion to available memory, but
7931 * ensure a minimum size on small systems.
7933 tmp
= max_t(u64
, tmp
>> 2,
7934 mult_frac(zone_managed_pages(zone
),
7935 watermark_scale_factor
, 10000));
7937 zone
->watermark_boost
= 0;
7938 zone
->_watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + tmp
;
7939 zone
->_watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + tmp
* 2;
7941 spin_unlock_irqrestore(&zone
->lock
, flags
);
7944 /* update totalreserve_pages */
7945 calculate_totalreserve_pages();
7949 * setup_per_zone_wmarks - called when min_free_kbytes changes
7950 * or when memory is hot-{added|removed}
7952 * Ensures that the watermark[min,low,high] values for each zone are set
7953 * correctly with respect to min_free_kbytes.
7955 void setup_per_zone_wmarks(void)
7957 static DEFINE_SPINLOCK(lock
);
7960 __setup_per_zone_wmarks();
7965 * Initialise min_free_kbytes.
7967 * For small machines we want it small (128k min). For large machines
7968 * we want it large (256MB max). But it is not linear, because network
7969 * bandwidth does not increase linearly with machine size. We use
7971 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7972 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7988 int __meminit
init_per_zone_wmark_min(void)
7990 unsigned long lowmem_kbytes
;
7991 int new_min_free_kbytes
;
7993 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
7994 new_min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
7996 if (new_min_free_kbytes
> user_min_free_kbytes
) {
7997 min_free_kbytes
= new_min_free_kbytes
;
7998 if (min_free_kbytes
< 128)
7999 min_free_kbytes
= 128;
8000 if (min_free_kbytes
> 262144)
8001 min_free_kbytes
= 262144;
8003 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
8004 new_min_free_kbytes
, user_min_free_kbytes
);
8006 setup_per_zone_wmarks();
8007 refresh_zone_stat_thresholds();
8008 setup_per_zone_lowmem_reserve();
8011 setup_min_unmapped_ratio();
8012 setup_min_slab_ratio();
8015 khugepaged_min_free_kbytes_update();
8019 postcore_initcall(init_per_zone_wmark_min
)
8022 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
8023 * that we can call two helper functions whenever min_free_kbytes
8026 int min_free_kbytes_sysctl_handler(struct ctl_table
*table
, int write
,
8027 void *buffer
, size_t *length
, loff_t
*ppos
)
8031 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
8036 user_min_free_kbytes
= min_free_kbytes
;
8037 setup_per_zone_wmarks();
8042 int watermark_scale_factor_sysctl_handler(struct ctl_table
*table
, int write
,
8043 void *buffer
, size_t *length
, loff_t
*ppos
)
8047 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
8052 setup_per_zone_wmarks();
8058 static void setup_min_unmapped_ratio(void)
8063 for_each_online_pgdat(pgdat
)
8064 pgdat
->min_unmapped_pages
= 0;
8067 zone
->zone_pgdat
->min_unmapped_pages
+= (zone_managed_pages(zone
) *
8068 sysctl_min_unmapped_ratio
) / 100;
8072 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
8073 void *buffer
, size_t *length
, loff_t
*ppos
)
8077 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
8081 setup_min_unmapped_ratio();
8086 static void setup_min_slab_ratio(void)
8091 for_each_online_pgdat(pgdat
)
8092 pgdat
->min_slab_pages
= 0;
8095 zone
->zone_pgdat
->min_slab_pages
+= (zone_managed_pages(zone
) *
8096 sysctl_min_slab_ratio
) / 100;
8099 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
8100 void *buffer
, size_t *length
, loff_t
*ppos
)
8104 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
8108 setup_min_slab_ratio();
8115 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
8116 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
8117 * whenever sysctl_lowmem_reserve_ratio changes.
8119 * The reserve ratio obviously has absolutely no relation with the
8120 * minimum watermarks. The lowmem reserve ratio can only make sense
8121 * if in function of the boot time zone sizes.
8123 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
8124 void *buffer
, size_t *length
, loff_t
*ppos
)
8128 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
8130 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
8131 if (sysctl_lowmem_reserve_ratio
[i
] < 1)
8132 sysctl_lowmem_reserve_ratio
[i
] = 0;
8135 setup_per_zone_lowmem_reserve();
8140 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
8141 * cpu. It is the fraction of total pages in each zone that a hot per cpu
8142 * pagelist can have before it gets flushed back to buddy allocator.
8144 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table
*table
, int write
,
8145 void *buffer
, size_t *length
, loff_t
*ppos
)
8148 int old_percpu_pagelist_fraction
;
8151 mutex_lock(&pcp_batch_high_lock
);
8152 old_percpu_pagelist_fraction
= percpu_pagelist_fraction
;
8154 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
8155 if (!write
|| ret
< 0)
8158 /* Sanity checking to avoid pcp imbalance */
8159 if (percpu_pagelist_fraction
&&
8160 percpu_pagelist_fraction
< MIN_PERCPU_PAGELIST_FRACTION
) {
8161 percpu_pagelist_fraction
= old_percpu_pagelist_fraction
;
8167 if (percpu_pagelist_fraction
== old_percpu_pagelist_fraction
)
8170 for_each_populated_zone(zone
)
8171 zone_set_pageset_high_and_batch(zone
);
8173 mutex_unlock(&pcp_batch_high_lock
);
8177 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8179 * Returns the number of pages that arch has reserved but
8180 * is not known to alloc_large_system_hash().
8182 static unsigned long __init
arch_reserved_kernel_pages(void)
8189 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8190 * machines. As memory size is increased the scale is also increased but at
8191 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8192 * quadruples the scale is increased by one, which means the size of hash table
8193 * only doubles, instead of quadrupling as well.
8194 * Because 32-bit systems cannot have large physical memory, where this scaling
8195 * makes sense, it is disabled on such platforms.
8197 #if __BITS_PER_LONG > 32
8198 #define ADAPT_SCALE_BASE (64ul << 30)
8199 #define ADAPT_SCALE_SHIFT 2
8200 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8204 * allocate a large system hash table from bootmem
8205 * - it is assumed that the hash table must contain an exact power-of-2
8206 * quantity of entries
8207 * - limit is the number of hash buckets, not the total allocation size
8209 void *__init
alloc_large_system_hash(const char *tablename
,
8210 unsigned long bucketsize
,
8211 unsigned long numentries
,
8214 unsigned int *_hash_shift
,
8215 unsigned int *_hash_mask
,
8216 unsigned long low_limit
,
8217 unsigned long high_limit
)
8219 unsigned long long max
= high_limit
;
8220 unsigned long log2qty
, size
;
8225 /* allow the kernel cmdline to have a say */
8227 /* round applicable memory size up to nearest megabyte */
8228 numentries
= nr_kernel_pages
;
8229 numentries
-= arch_reserved_kernel_pages();
8231 /* It isn't necessary when PAGE_SIZE >= 1MB */
8232 if (PAGE_SHIFT
< 20)
8233 numentries
= round_up(numentries
, (1<<20)/PAGE_SIZE
);
8235 #if __BITS_PER_LONG > 32
8237 unsigned long adapt
;
8239 for (adapt
= ADAPT_SCALE_NPAGES
; adapt
< numentries
;
8240 adapt
<<= ADAPT_SCALE_SHIFT
)
8245 /* limit to 1 bucket per 2^scale bytes of low memory */
8246 if (scale
> PAGE_SHIFT
)
8247 numentries
>>= (scale
- PAGE_SHIFT
);
8249 numentries
<<= (PAGE_SHIFT
- scale
);
8251 /* Make sure we've got at least a 0-order allocation.. */
8252 if (unlikely(flags
& HASH_SMALL
)) {
8253 /* Makes no sense without HASH_EARLY */
8254 WARN_ON(!(flags
& HASH_EARLY
));
8255 if (!(numentries
>> *_hash_shift
)) {
8256 numentries
= 1UL << *_hash_shift
;
8257 BUG_ON(!numentries
);
8259 } else if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
8260 numentries
= PAGE_SIZE
/ bucketsize
;
8262 numentries
= roundup_pow_of_two(numentries
);
8264 /* limit allocation size to 1/16 total memory by default */
8266 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
8267 do_div(max
, bucketsize
);
8269 max
= min(max
, 0x80000000ULL
);
8271 if (numentries
< low_limit
)
8272 numentries
= low_limit
;
8273 if (numentries
> max
)
8276 log2qty
= ilog2(numentries
);
8278 gfp_flags
= (flags
& HASH_ZERO
) ? GFP_ATOMIC
| __GFP_ZERO
: GFP_ATOMIC
;
8281 size
= bucketsize
<< log2qty
;
8282 if (flags
& HASH_EARLY
) {
8283 if (flags
& HASH_ZERO
)
8284 table
= memblock_alloc(size
, SMP_CACHE_BYTES
);
8286 table
= memblock_alloc_raw(size
,
8288 } else if (get_order(size
) >= MAX_ORDER
|| hashdist
) {
8289 table
= __vmalloc(size
, gfp_flags
);
8293 * If bucketsize is not a power-of-two, we may free
8294 * some pages at the end of hash table which
8295 * alloc_pages_exact() automatically does
8297 table
= alloc_pages_exact(size
, gfp_flags
);
8298 kmemleak_alloc(table
, size
, 1, gfp_flags
);
8300 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
8303 panic("Failed to allocate %s hash table\n", tablename
);
8305 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8306 tablename
, 1UL << log2qty
, ilog2(size
) - PAGE_SHIFT
, size
,
8307 virt
? "vmalloc" : "linear");
8310 *_hash_shift
= log2qty
;
8312 *_hash_mask
= (1 << log2qty
) - 1;
8318 * This function checks whether pageblock includes unmovable pages or not.
8320 * PageLRU check without isolation or lru_lock could race so that
8321 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8322 * check without lock_page also may miss some movable non-lru pages at
8323 * race condition. So you can't expect this function should be exact.
8325 * Returns a page without holding a reference. If the caller wants to
8326 * dereference that page (e.g., dumping), it has to make sure that it
8327 * cannot get removed (e.g., via memory unplug) concurrently.
8330 struct page
*has_unmovable_pages(struct zone
*zone
, struct page
*page
,
8331 int migratetype
, int flags
)
8333 unsigned long iter
= 0;
8334 unsigned long pfn
= page_to_pfn(page
);
8335 unsigned long offset
= pfn
% pageblock_nr_pages
;
8337 if (is_migrate_cma_page(page
)) {
8339 * CMA allocations (alloc_contig_range) really need to mark
8340 * isolate CMA pageblocks even when they are not movable in fact
8341 * so consider them movable here.
8343 if (is_migrate_cma(migratetype
))
8349 for (; iter
< pageblock_nr_pages
- offset
; iter
++) {
8350 if (!pfn_valid_within(pfn
+ iter
))
8353 page
= pfn_to_page(pfn
+ iter
);
8356 * Both, bootmem allocations and memory holes are marked
8357 * PG_reserved and are unmovable. We can even have unmovable
8358 * allocations inside ZONE_MOVABLE, for example when
8359 * specifying "movablecore".
8361 if (PageReserved(page
))
8365 * If the zone is movable and we have ruled out all reserved
8366 * pages then it should be reasonably safe to assume the rest
8369 if (zone_idx(zone
) == ZONE_MOVABLE
)
8373 * Hugepages are not in LRU lists, but they're movable.
8374 * THPs are on the LRU, but need to be counted as #small pages.
8375 * We need not scan over tail pages because we don't
8376 * handle each tail page individually in migration.
8378 if (PageHuge(page
) || PageTransCompound(page
)) {
8379 struct page
*head
= compound_head(page
);
8380 unsigned int skip_pages
;
8382 if (PageHuge(page
)) {
8383 if (!hugepage_migration_supported(page_hstate(head
)))
8385 } else if (!PageLRU(head
) && !__PageMovable(head
)) {
8389 skip_pages
= compound_nr(head
) - (page
- head
);
8390 iter
+= skip_pages
- 1;
8395 * We can't use page_count without pin a page
8396 * because another CPU can free compound page.
8397 * This check already skips compound tails of THP
8398 * because their page->_refcount is zero at all time.
8400 if (!page_ref_count(page
)) {
8401 if (PageBuddy(page
))
8402 iter
+= (1 << buddy_order(page
)) - 1;
8407 * The HWPoisoned page may be not in buddy system, and
8408 * page_count() is not 0.
8410 if ((flags
& MEMORY_OFFLINE
) && PageHWPoison(page
))
8414 * We treat all PageOffline() pages as movable when offlining
8415 * to give drivers a chance to decrement their reference count
8416 * in MEM_GOING_OFFLINE in order to indicate that these pages
8417 * can be offlined as there are no direct references anymore.
8418 * For actually unmovable PageOffline() where the driver does
8419 * not support this, we will fail later when trying to actually
8420 * move these pages that still have a reference count > 0.
8421 * (false negatives in this function only)
8423 if ((flags
& MEMORY_OFFLINE
) && PageOffline(page
))
8426 if (__PageMovable(page
) || PageLRU(page
))
8430 * If there are RECLAIMABLE pages, we need to check
8431 * it. But now, memory offline itself doesn't call
8432 * shrink_node_slabs() and it still to be fixed.
8439 #ifdef CONFIG_CONTIG_ALLOC
8440 static unsigned long pfn_max_align_down(unsigned long pfn
)
8442 return pfn
& ~(max_t(unsigned long, MAX_ORDER_NR_PAGES
,
8443 pageblock_nr_pages
) - 1);
8446 static unsigned long pfn_max_align_up(unsigned long pfn
)
8448 return ALIGN(pfn
, max_t(unsigned long, MAX_ORDER_NR_PAGES
,
8449 pageblock_nr_pages
));
8452 /* [start, end) must belong to a single zone. */
8453 static int __alloc_contig_migrate_range(struct compact_control
*cc
,
8454 unsigned long start
, unsigned long end
)
8456 /* This function is based on compact_zone() from compaction.c. */
8457 unsigned int nr_reclaimed
;
8458 unsigned long pfn
= start
;
8459 unsigned int tries
= 0;
8461 struct migration_target_control mtc
= {
8462 .nid
= zone_to_nid(cc
->zone
),
8463 .gfp_mask
= GFP_USER
| __GFP_MOVABLE
| __GFP_RETRY_MAYFAIL
,
8468 while (pfn
< end
|| !list_empty(&cc
->migratepages
)) {
8469 if (fatal_signal_pending(current
)) {
8474 if (list_empty(&cc
->migratepages
)) {
8475 cc
->nr_migratepages
= 0;
8476 pfn
= isolate_migratepages_range(cc
, pfn
, end
);
8482 } else if (++tries
== 5) {
8483 ret
= ret
< 0 ? ret
: -EBUSY
;
8487 nr_reclaimed
= reclaim_clean_pages_from_list(cc
->zone
,
8489 cc
->nr_migratepages
-= nr_reclaimed
;
8491 ret
= migrate_pages(&cc
->migratepages
, alloc_migration_target
,
8492 NULL
, (unsigned long)&mtc
, cc
->mode
, MR_CONTIG_RANGE
);
8495 putback_movable_pages(&cc
->migratepages
);
8502 * alloc_contig_range() -- tries to allocate given range of pages
8503 * @start: start PFN to allocate
8504 * @end: one-past-the-last PFN to allocate
8505 * @migratetype: migratetype of the underlaying pageblocks (either
8506 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
8507 * in range must have the same migratetype and it must
8508 * be either of the two.
8509 * @gfp_mask: GFP mask to use during compaction
8511 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
8512 * aligned. The PFN range must belong to a single zone.
8514 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
8515 * pageblocks in the range. Once isolated, the pageblocks should not
8516 * be modified by others.
8518 * Return: zero on success or negative error code. On success all
8519 * pages which PFN is in [start, end) are allocated for the caller and
8520 * need to be freed with free_contig_range().
8522 int alloc_contig_range(unsigned long start
, unsigned long end
,
8523 unsigned migratetype
, gfp_t gfp_mask
)
8525 unsigned long outer_start
, outer_end
;
8529 struct compact_control cc
= {
8530 .nr_migratepages
= 0,
8532 .zone
= page_zone(pfn_to_page(start
)),
8533 .mode
= MIGRATE_SYNC
,
8534 .ignore_skip_hint
= true,
8535 .no_set_skip_hint
= true,
8536 .gfp_mask
= current_gfp_context(gfp_mask
),
8537 .alloc_contig
= true,
8539 INIT_LIST_HEAD(&cc
.migratepages
);
8542 * What we do here is we mark all pageblocks in range as
8543 * MIGRATE_ISOLATE. Because pageblock and max order pages may
8544 * have different sizes, and due to the way page allocator
8545 * work, we align the range to biggest of the two pages so
8546 * that page allocator won't try to merge buddies from
8547 * different pageblocks and change MIGRATE_ISOLATE to some
8548 * other migration type.
8550 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
8551 * migrate the pages from an unaligned range (ie. pages that
8552 * we are interested in). This will put all the pages in
8553 * range back to page allocator as MIGRATE_ISOLATE.
8555 * When this is done, we take the pages in range from page
8556 * allocator removing them from the buddy system. This way
8557 * page allocator will never consider using them.
8559 * This lets us mark the pageblocks back as
8560 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
8561 * aligned range but not in the unaligned, original range are
8562 * put back to page allocator so that buddy can use them.
8565 ret
= start_isolate_page_range(pfn_max_align_down(start
),
8566 pfn_max_align_up(end
), migratetype
, 0);
8570 drain_all_pages(cc
.zone
);
8573 * In case of -EBUSY, we'd like to know which page causes problem.
8574 * So, just fall through. test_pages_isolated() has a tracepoint
8575 * which will report the busy page.
8577 * It is possible that busy pages could become available before
8578 * the call to test_pages_isolated, and the range will actually be
8579 * allocated. So, if we fall through be sure to clear ret so that
8580 * -EBUSY is not accidentally used or returned to caller.
8582 ret
= __alloc_contig_migrate_range(&cc
, start
, end
);
8583 if (ret
&& ret
!= -EBUSY
)
8588 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
8589 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
8590 * more, all pages in [start, end) are free in page allocator.
8591 * What we are going to do is to allocate all pages from
8592 * [start, end) (that is remove them from page allocator).
8594 * The only problem is that pages at the beginning and at the
8595 * end of interesting range may be not aligned with pages that
8596 * page allocator holds, ie. they can be part of higher order
8597 * pages. Because of this, we reserve the bigger range and
8598 * once this is done free the pages we are not interested in.
8600 * We don't have to hold zone->lock here because the pages are
8601 * isolated thus they won't get removed from buddy.
8604 lru_add_drain_all();
8607 outer_start
= start
;
8608 while (!PageBuddy(pfn_to_page(outer_start
))) {
8609 if (++order
>= MAX_ORDER
) {
8610 outer_start
= start
;
8613 outer_start
&= ~0UL << order
;
8616 if (outer_start
!= start
) {
8617 order
= buddy_order(pfn_to_page(outer_start
));
8620 * outer_start page could be small order buddy page and
8621 * it doesn't include start page. Adjust outer_start
8622 * in this case to report failed page properly
8623 * on tracepoint in test_pages_isolated()
8625 if (outer_start
+ (1UL << order
) <= start
)
8626 outer_start
= start
;
8629 /* Make sure the range is really isolated. */
8630 if (test_pages_isolated(outer_start
, end
, 0)) {
8631 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
8632 __func__
, outer_start
, end
);
8637 /* Grab isolated pages from freelists. */
8638 outer_end
= isolate_freepages_range(&cc
, outer_start
, end
);
8644 /* Free head and tail (if any) */
8645 if (start
!= outer_start
)
8646 free_contig_range(outer_start
, start
- outer_start
);
8647 if (end
!= outer_end
)
8648 free_contig_range(end
, outer_end
- end
);
8651 undo_isolate_page_range(pfn_max_align_down(start
),
8652 pfn_max_align_up(end
), migratetype
);
8655 EXPORT_SYMBOL(alloc_contig_range
);
8657 static int __alloc_contig_pages(unsigned long start_pfn
,
8658 unsigned long nr_pages
, gfp_t gfp_mask
)
8660 unsigned long end_pfn
= start_pfn
+ nr_pages
;
8662 return alloc_contig_range(start_pfn
, end_pfn
, MIGRATE_MOVABLE
,
8666 static bool pfn_range_valid_contig(struct zone
*z
, unsigned long start_pfn
,
8667 unsigned long nr_pages
)
8669 unsigned long i
, end_pfn
= start_pfn
+ nr_pages
;
8672 for (i
= start_pfn
; i
< end_pfn
; i
++) {
8673 page
= pfn_to_online_page(i
);
8677 if (page_zone(page
) != z
)
8680 if (PageReserved(page
))
8683 if (page_count(page
) > 0)
8692 static bool zone_spans_last_pfn(const struct zone
*zone
,
8693 unsigned long start_pfn
, unsigned long nr_pages
)
8695 unsigned long last_pfn
= start_pfn
+ nr_pages
- 1;
8697 return zone_spans_pfn(zone
, last_pfn
);
8701 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
8702 * @nr_pages: Number of contiguous pages to allocate
8703 * @gfp_mask: GFP mask to limit search and used during compaction
8705 * @nodemask: Mask for other possible nodes
8707 * This routine is a wrapper around alloc_contig_range(). It scans over zones
8708 * on an applicable zonelist to find a contiguous pfn range which can then be
8709 * tried for allocation with alloc_contig_range(). This routine is intended
8710 * for allocation requests which can not be fulfilled with the buddy allocator.
8712 * The allocated memory is always aligned to a page boundary. If nr_pages is a
8713 * power of two then the alignment is guaranteed to be to the given nr_pages
8714 * (e.g. 1GB request would be aligned to 1GB).
8716 * Allocated pages can be freed with free_contig_range() or by manually calling
8717 * __free_page() on each allocated page.
8719 * Return: pointer to contiguous pages on success, or NULL if not successful.
8721 struct page
*alloc_contig_pages(unsigned long nr_pages
, gfp_t gfp_mask
,
8722 int nid
, nodemask_t
*nodemask
)
8724 unsigned long ret
, pfn
, flags
;
8725 struct zonelist
*zonelist
;
8729 zonelist
= node_zonelist(nid
, gfp_mask
);
8730 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
8731 gfp_zone(gfp_mask
), nodemask
) {
8732 spin_lock_irqsave(&zone
->lock
, flags
);
8734 pfn
= ALIGN(zone
->zone_start_pfn
, nr_pages
);
8735 while (zone_spans_last_pfn(zone
, pfn
, nr_pages
)) {
8736 if (pfn_range_valid_contig(zone
, pfn
, nr_pages
)) {
8738 * We release the zone lock here because
8739 * alloc_contig_range() will also lock the zone
8740 * at some point. If there's an allocation
8741 * spinning on this lock, it may win the race
8742 * and cause alloc_contig_range() to fail...
8744 spin_unlock_irqrestore(&zone
->lock
, flags
);
8745 ret
= __alloc_contig_pages(pfn
, nr_pages
,
8748 return pfn_to_page(pfn
);
8749 spin_lock_irqsave(&zone
->lock
, flags
);
8753 spin_unlock_irqrestore(&zone
->lock
, flags
);
8757 #endif /* CONFIG_CONTIG_ALLOC */
8759 void free_contig_range(unsigned long pfn
, unsigned int nr_pages
)
8761 unsigned int count
= 0;
8763 for (; nr_pages
--; pfn
++) {
8764 struct page
*page
= pfn_to_page(pfn
);
8766 count
+= page_count(page
) != 1;
8769 WARN(count
!= 0, "%d pages are still in use!\n", count
);
8771 EXPORT_SYMBOL(free_contig_range
);
8774 * The zone indicated has a new number of managed_pages; batch sizes and percpu
8775 * page high values need to be recalulated.
8777 void __meminit
zone_pcp_update(struct zone
*zone
)
8779 mutex_lock(&pcp_batch_high_lock
);
8780 zone_set_pageset_high_and_batch(zone
);
8781 mutex_unlock(&pcp_batch_high_lock
);
8785 * Effectively disable pcplists for the zone by setting the high limit to 0
8786 * and draining all cpus. A concurrent page freeing on another CPU that's about
8787 * to put the page on pcplist will either finish before the drain and the page
8788 * will be drained, or observe the new high limit and skip the pcplist.
8790 * Must be paired with a call to zone_pcp_enable().
8792 void zone_pcp_disable(struct zone
*zone
)
8794 mutex_lock(&pcp_batch_high_lock
);
8795 __zone_set_pageset_high_and_batch(zone
, 0, 1);
8796 __drain_all_pages(zone
, true);
8799 void zone_pcp_enable(struct zone
*zone
)
8801 __zone_set_pageset_high_and_batch(zone
, zone
->pageset_high
, zone
->pageset_batch
);
8802 mutex_unlock(&pcp_batch_high_lock
);
8805 void zone_pcp_reset(struct zone
*zone
)
8807 unsigned long flags
;
8809 struct per_cpu_pageset
*pset
;
8811 /* avoid races with drain_pages() */
8812 local_irq_save(flags
);
8813 if (zone
->pageset
!= &boot_pageset
) {
8814 for_each_online_cpu(cpu
) {
8815 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
8816 drain_zonestat(zone
, pset
);
8818 free_percpu(zone
->pageset
);
8819 zone
->pageset
= &boot_pageset
;
8821 local_irq_restore(flags
);
8824 #ifdef CONFIG_MEMORY_HOTREMOVE
8826 * All pages in the range must be in a single zone, must not contain holes,
8827 * must span full sections, and must be isolated before calling this function.
8829 void __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
8831 unsigned long pfn
= start_pfn
;
8835 unsigned long flags
;
8837 offline_mem_sections(pfn
, end_pfn
);
8838 zone
= page_zone(pfn_to_page(pfn
));
8839 spin_lock_irqsave(&zone
->lock
, flags
);
8840 while (pfn
< end_pfn
) {
8841 page
= pfn_to_page(pfn
);
8843 * The HWPoisoned page may be not in buddy system, and
8844 * page_count() is not 0.
8846 if (unlikely(!PageBuddy(page
) && PageHWPoison(page
))) {
8851 * At this point all remaining PageOffline() pages have a
8852 * reference count of 0 and can simply be skipped.
8854 if (PageOffline(page
)) {
8855 BUG_ON(page_count(page
));
8856 BUG_ON(PageBuddy(page
));
8861 BUG_ON(page_count(page
));
8862 BUG_ON(!PageBuddy(page
));
8863 order
= buddy_order(page
);
8864 del_page_from_free_list(page
, zone
, order
);
8865 pfn
+= (1 << order
);
8867 spin_unlock_irqrestore(&zone
->lock
, flags
);
8871 bool is_free_buddy_page(struct page
*page
)
8873 struct zone
*zone
= page_zone(page
);
8874 unsigned long pfn
= page_to_pfn(page
);
8875 unsigned long flags
;
8878 spin_lock_irqsave(&zone
->lock
, flags
);
8879 for (order
= 0; order
< MAX_ORDER
; order
++) {
8880 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
8882 if (PageBuddy(page_head
) && buddy_order(page_head
) >= order
)
8885 spin_unlock_irqrestore(&zone
->lock
, flags
);
8887 return order
< MAX_ORDER
;
8890 #ifdef CONFIG_MEMORY_FAILURE
8892 * Break down a higher-order page in sub-pages, and keep our target out of
8895 static void break_down_buddy_pages(struct zone
*zone
, struct page
*page
,
8896 struct page
*target
, int low
, int high
,
8899 unsigned long size
= 1 << high
;
8900 struct page
*current_buddy
, *next_page
;
8902 while (high
> low
) {
8906 if (target
>= &page
[size
]) {
8907 next_page
= page
+ size
;
8908 current_buddy
= page
;
8911 current_buddy
= page
+ size
;
8914 if (set_page_guard(zone
, current_buddy
, high
, migratetype
))
8917 if (current_buddy
!= target
) {
8918 add_to_free_list(current_buddy
, zone
, high
, migratetype
);
8919 set_buddy_order(current_buddy
, high
);
8926 * Take a page that will be marked as poisoned off the buddy allocator.
8928 bool take_page_off_buddy(struct page
*page
)
8930 struct zone
*zone
= page_zone(page
);
8931 unsigned long pfn
= page_to_pfn(page
);
8932 unsigned long flags
;
8936 spin_lock_irqsave(&zone
->lock
, flags
);
8937 for (order
= 0; order
< MAX_ORDER
; order
++) {
8938 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
8939 int page_order
= buddy_order(page_head
);
8941 if (PageBuddy(page_head
) && page_order
>= order
) {
8942 unsigned long pfn_head
= page_to_pfn(page_head
);
8943 int migratetype
= get_pfnblock_migratetype(page_head
,
8946 del_page_from_free_list(page_head
, zone
, page_order
);
8947 break_down_buddy_pages(zone
, page_head
, page
, 0,
8948 page_order
, migratetype
);
8952 if (page_count(page_head
) > 0)
8955 spin_unlock_irqrestore(&zone
->lock
, flags
);