2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kasan.h>
28 #include <linux/module.h>
29 #include <linux/suspend.h>
30 #include <linux/pagevec.h>
31 #include <linux/blkdev.h>
32 #include <linux/slab.h>
33 #include <linux/ratelimit.h>
34 #include <linux/oom.h>
35 #include <linux/notifier.h>
36 #include <linux/topology.h>
37 #include <linux/sysctl.h>
38 #include <linux/cpu.h>
39 #include <linux/cpuset.h>
40 #include <linux/memory_hotplug.h>
41 #include <linux/nodemask.h>
42 #include <linux/vmalloc.h>
43 #include <linux/vmstat.h>
44 #include <linux/mempolicy.h>
45 #include <linux/memremap.h>
46 #include <linux/stop_machine.h>
47 #include <linux/sort.h>
48 #include <linux/pfn.h>
50 #include <linux/backing-dev.h>
51 #include <linux/fault-inject.h>
52 #include <linux/page-isolation.h>
53 #include <linux/page_ext.h>
54 #include <linux/debugobjects.h>
55 #include <linux/kmemleak.h>
56 #include <linux/compaction.h>
57 #include <trace/events/kmem.h>
58 #include <trace/events/oom.h>
59 #include <linux/prefetch.h>
60 #include <linux/mm_inline.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>
72 #include <asm/sections.h>
73 #include <asm/tlbflush.h>
74 #include <asm/div64.h>
77 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
78 static DEFINE_MUTEX(pcp_batch_high_lock
);
79 #define MIN_PERCPU_PAGELIST_FRACTION (8)
81 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
82 DEFINE_PER_CPU(int, numa_node
);
83 EXPORT_PER_CPU_SYMBOL(numa_node
);
86 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key
);
88 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
90 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
91 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
92 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
93 * defined in <linux/topology.h>.
95 DEFINE_PER_CPU(int, _numa_mem_
); /* Kernel "local memory" node */
96 EXPORT_PER_CPU_SYMBOL(_numa_mem_
);
97 int _node_numa_mem_
[MAX_NUMNODES
];
100 /* work_structs for global per-cpu drains */
101 DEFINE_MUTEX(pcpu_drain_mutex
);
102 DEFINE_PER_CPU(struct work_struct
, pcpu_drain
);
104 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
105 volatile unsigned long latent_entropy __latent_entropy
;
106 EXPORT_SYMBOL(latent_entropy
);
110 * Array of node states.
112 nodemask_t node_states
[NR_NODE_STATES
] __read_mostly
= {
113 [N_POSSIBLE
] = NODE_MASK_ALL
,
114 [N_ONLINE
] = { { [0] = 1UL } },
116 [N_NORMAL_MEMORY
] = { { [0] = 1UL } },
117 #ifdef CONFIG_HIGHMEM
118 [N_HIGH_MEMORY
] = { { [0] = 1UL } },
120 [N_MEMORY
] = { { [0] = 1UL } },
121 [N_CPU
] = { { [0] = 1UL } },
124 EXPORT_SYMBOL(node_states
);
126 /* Protect totalram_pages and zone->managed_pages */
127 static DEFINE_SPINLOCK(managed_page_count_lock
);
129 unsigned long totalram_pages __read_mostly
;
130 unsigned long totalreserve_pages __read_mostly
;
131 unsigned long totalcma_pages __read_mostly
;
133 int percpu_pagelist_fraction
;
134 gfp_t gfp_allowed_mask __read_mostly
= GFP_BOOT_MASK
;
137 * A cached value of the page's pageblock's migratetype, used when the page is
138 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
139 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
140 * Also the migratetype set in the page does not necessarily match the pcplist
141 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
142 * other index - this ensures that it will be put on the correct CMA freelist.
144 static inline int get_pcppage_migratetype(struct page
*page
)
149 static inline void set_pcppage_migratetype(struct page
*page
, int migratetype
)
151 page
->index
= migratetype
;
154 #ifdef CONFIG_PM_SLEEP
156 * The following functions are used by the suspend/hibernate code to temporarily
157 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
158 * while devices are suspended. To avoid races with the suspend/hibernate code,
159 * they should always be called with pm_mutex held (gfp_allowed_mask also should
160 * only be modified with pm_mutex held, unless the suspend/hibernate code is
161 * guaranteed not to run in parallel with that modification).
164 static gfp_t saved_gfp_mask
;
166 void pm_restore_gfp_mask(void)
168 WARN_ON(!mutex_is_locked(&pm_mutex
));
169 if (saved_gfp_mask
) {
170 gfp_allowed_mask
= saved_gfp_mask
;
175 void pm_restrict_gfp_mask(void)
177 WARN_ON(!mutex_is_locked(&pm_mutex
));
178 WARN_ON(saved_gfp_mask
);
179 saved_gfp_mask
= gfp_allowed_mask
;
180 gfp_allowed_mask
&= ~(__GFP_IO
| __GFP_FS
);
183 bool pm_suspended_storage(void)
185 if ((gfp_allowed_mask
& (__GFP_IO
| __GFP_FS
)) == (__GFP_IO
| __GFP_FS
))
189 #endif /* CONFIG_PM_SLEEP */
191 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
192 unsigned int pageblock_order __read_mostly
;
195 static void __free_pages_ok(struct page
*page
, unsigned int order
);
198 * results with 256, 32 in the lowmem_reserve sysctl:
199 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
200 * 1G machine -> (16M dma, 784M normal, 224M high)
201 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
202 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
203 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
205 * TBD: should special case ZONE_DMA32 machines here - in those we normally
206 * don't need any ZONE_NORMAL reservation
208 int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
-1] = {
209 #ifdef CONFIG_ZONE_DMA
212 #ifdef CONFIG_ZONE_DMA32
215 #ifdef CONFIG_HIGHMEM
221 EXPORT_SYMBOL(totalram_pages
);
223 static char * const zone_names
[MAX_NR_ZONES
] = {
224 #ifdef CONFIG_ZONE_DMA
227 #ifdef CONFIG_ZONE_DMA32
231 #ifdef CONFIG_HIGHMEM
235 #ifdef CONFIG_ZONE_DEVICE
240 char * const migratetype_names
[MIGRATE_TYPES
] = {
248 #ifdef CONFIG_MEMORY_ISOLATION
253 compound_page_dtor
* const compound_page_dtors
[] = {
256 #ifdef CONFIG_HUGETLB_PAGE
259 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
264 int min_free_kbytes
= 1024;
265 int user_min_free_kbytes
= -1;
266 int watermark_scale_factor
= 10;
268 static unsigned long __meminitdata nr_kernel_pages
;
269 static unsigned long __meminitdata nr_all_pages
;
270 static unsigned long __meminitdata dma_reserve
;
272 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
273 static unsigned long __meminitdata arch_zone_lowest_possible_pfn
[MAX_NR_ZONES
];
274 static unsigned long __meminitdata arch_zone_highest_possible_pfn
[MAX_NR_ZONES
];
275 static unsigned long __initdata required_kernelcore
;
276 static unsigned long __initdata required_movablecore
;
277 static unsigned long __meminitdata zone_movable_pfn
[MAX_NUMNODES
];
278 static bool mirrored_kernelcore
;
280 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
282 EXPORT_SYMBOL(movable_zone
);
283 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
286 int nr_node_ids __read_mostly
= MAX_NUMNODES
;
287 int nr_online_nodes __read_mostly
= 1;
288 EXPORT_SYMBOL(nr_node_ids
);
289 EXPORT_SYMBOL(nr_online_nodes
);
292 int page_group_by_mobility_disabled __read_mostly
;
294 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
297 * Determine how many pages need to be initialized durig early boot
298 * (non-deferred initialization).
299 * The value of first_deferred_pfn will be set later, once non-deferred pages
300 * are initialized, but for now set it ULONG_MAX.
302 static inline void reset_deferred_meminit(pg_data_t
*pgdat
)
304 phys_addr_t start_addr
, end_addr
;
305 unsigned long max_pgcnt
;
306 unsigned long reserved
;
309 * Initialise at least 2G of a node but also take into account that
310 * two large system hashes that can take up 1GB for 0.25TB/node.
312 max_pgcnt
= max(2UL << (30 - PAGE_SHIFT
),
313 (pgdat
->node_spanned_pages
>> 8));
316 * Compensate the all the memblock reservations (e.g. crash kernel)
317 * from the initial estimation to make sure we will initialize enough
320 start_addr
= PFN_PHYS(pgdat
->node_start_pfn
);
321 end_addr
= PFN_PHYS(pgdat
->node_start_pfn
+ max_pgcnt
);
322 reserved
= memblock_reserved_memory_within(start_addr
, end_addr
);
323 max_pgcnt
+= PHYS_PFN(reserved
);
325 pgdat
->static_init_pgcnt
= min(max_pgcnt
, pgdat
->node_spanned_pages
);
326 pgdat
->first_deferred_pfn
= ULONG_MAX
;
329 /* Returns true if the struct page for the pfn is uninitialised */
330 static inline bool __meminit
early_page_uninitialised(unsigned long pfn
)
332 int nid
= early_pfn_to_nid(pfn
);
334 if (node_online(nid
) && pfn
>= NODE_DATA(nid
)->first_deferred_pfn
)
341 * Returns false when the remaining initialisation should be deferred until
342 * later in the boot cycle when it can be parallelised.
344 static inline bool update_defer_init(pg_data_t
*pgdat
,
345 unsigned long pfn
, unsigned long zone_end
,
346 unsigned long *nr_initialised
)
348 /* Always populate low zones for address-contrained allocations */
349 if (zone_end
< pgdat_end_pfn(pgdat
))
351 /* Xen PV domains need page structures early */
355 if ((*nr_initialised
> pgdat
->static_init_pgcnt
) &&
356 (pfn
& (PAGES_PER_SECTION
- 1)) == 0) {
357 pgdat
->first_deferred_pfn
= pfn
;
364 static inline void reset_deferred_meminit(pg_data_t
*pgdat
)
368 static inline bool early_page_uninitialised(unsigned long pfn
)
373 static inline bool update_defer_init(pg_data_t
*pgdat
,
374 unsigned long pfn
, unsigned long zone_end
,
375 unsigned long *nr_initialised
)
381 /* Return a pointer to the bitmap storing bits affecting a block of pages */
382 static inline unsigned long *get_pageblock_bitmap(struct page
*page
,
385 #ifdef CONFIG_SPARSEMEM
386 return __pfn_to_section(pfn
)->pageblock_flags
;
388 return page_zone(page
)->pageblock_flags
;
389 #endif /* CONFIG_SPARSEMEM */
392 static inline int pfn_to_bitidx(struct page
*page
, unsigned long pfn
)
394 #ifdef CONFIG_SPARSEMEM
395 pfn
&= (PAGES_PER_SECTION
-1);
396 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
398 pfn
= pfn
- round_down(page_zone(page
)->zone_start_pfn
, pageblock_nr_pages
);
399 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
400 #endif /* CONFIG_SPARSEMEM */
404 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
405 * @page: The page within the block of interest
406 * @pfn: The target page frame number
407 * @end_bitidx: The last bit of interest to retrieve
408 * @mask: mask of bits that the caller is interested in
410 * Return: pageblock_bits flags
412 static __always_inline
unsigned long __get_pfnblock_flags_mask(struct page
*page
,
414 unsigned long end_bitidx
,
417 unsigned long *bitmap
;
418 unsigned long bitidx
, word_bitidx
;
421 bitmap
= get_pageblock_bitmap(page
, pfn
);
422 bitidx
= pfn_to_bitidx(page
, pfn
);
423 word_bitidx
= bitidx
/ BITS_PER_LONG
;
424 bitidx
&= (BITS_PER_LONG
-1);
426 word
= bitmap
[word_bitidx
];
427 bitidx
+= end_bitidx
;
428 return (word
>> (BITS_PER_LONG
- bitidx
- 1)) & mask
;
431 unsigned long get_pfnblock_flags_mask(struct page
*page
, unsigned long pfn
,
432 unsigned long end_bitidx
,
435 return __get_pfnblock_flags_mask(page
, pfn
, end_bitidx
, mask
);
438 static __always_inline
int get_pfnblock_migratetype(struct page
*page
, unsigned long pfn
)
440 return __get_pfnblock_flags_mask(page
, pfn
, PB_migrate_end
, MIGRATETYPE_MASK
);
444 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
445 * @page: The page within the block of interest
446 * @flags: The flags to set
447 * @pfn: The target page frame number
448 * @end_bitidx: The last bit of interest
449 * @mask: mask of bits that the caller is interested in
451 void set_pfnblock_flags_mask(struct page
*page
, unsigned long flags
,
453 unsigned long end_bitidx
,
456 unsigned long *bitmap
;
457 unsigned long bitidx
, word_bitidx
;
458 unsigned long old_word
, word
;
460 BUILD_BUG_ON(NR_PAGEBLOCK_BITS
!= 4);
462 bitmap
= get_pageblock_bitmap(page
, pfn
);
463 bitidx
= pfn_to_bitidx(page
, pfn
);
464 word_bitidx
= bitidx
/ BITS_PER_LONG
;
465 bitidx
&= (BITS_PER_LONG
-1);
467 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page
), pfn
), page
);
469 bitidx
+= end_bitidx
;
470 mask
<<= (BITS_PER_LONG
- bitidx
- 1);
471 flags
<<= (BITS_PER_LONG
- bitidx
- 1);
473 word
= READ_ONCE(bitmap
[word_bitidx
]);
475 old_word
= cmpxchg(&bitmap
[word_bitidx
], word
, (word
& ~mask
) | flags
);
476 if (word
== old_word
)
482 void set_pageblock_migratetype(struct page
*page
, int migratetype
)
484 if (unlikely(page_group_by_mobility_disabled
&&
485 migratetype
< MIGRATE_PCPTYPES
))
486 migratetype
= MIGRATE_UNMOVABLE
;
488 set_pageblock_flags_group(page
, (unsigned long)migratetype
,
489 PB_migrate
, PB_migrate_end
);
492 #ifdef CONFIG_DEBUG_VM
493 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
497 unsigned long pfn
= page_to_pfn(page
);
498 unsigned long sp
, start_pfn
;
501 seq
= zone_span_seqbegin(zone
);
502 start_pfn
= zone
->zone_start_pfn
;
503 sp
= zone
->spanned_pages
;
504 if (!zone_spans_pfn(zone
, pfn
))
506 } while (zone_span_seqretry(zone
, seq
));
509 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
510 pfn
, zone_to_nid(zone
), zone
->name
,
511 start_pfn
, start_pfn
+ sp
);
516 static int page_is_consistent(struct zone
*zone
, struct page
*page
)
518 if (!pfn_valid_within(page_to_pfn(page
)))
520 if (zone
!= page_zone(page
))
526 * Temporary debugging check for pages not lying within a given zone.
528 static int __maybe_unused
bad_range(struct zone
*zone
, struct page
*page
)
530 if (page_outside_zone_boundaries(zone
, page
))
532 if (!page_is_consistent(zone
, page
))
538 static inline int __maybe_unused
bad_range(struct zone
*zone
, struct page
*page
)
544 static void bad_page(struct page
*page
, const char *reason
,
545 unsigned long bad_flags
)
547 static unsigned long resume
;
548 static unsigned long nr_shown
;
549 static unsigned long nr_unshown
;
552 * Allow a burst of 60 reports, then keep quiet for that minute;
553 * or allow a steady drip of one report per second.
555 if (nr_shown
== 60) {
556 if (time_before(jiffies
, resume
)) {
562 "BUG: Bad page state: %lu messages suppressed\n",
569 resume
= jiffies
+ 60 * HZ
;
571 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
572 current
->comm
, page_to_pfn(page
));
573 __dump_page(page
, reason
);
574 bad_flags
&= page
->flags
;
576 pr_alert("bad because of flags: %#lx(%pGp)\n",
577 bad_flags
, &bad_flags
);
578 dump_page_owner(page
);
583 /* Leave bad fields for debug, except PageBuddy could make trouble */
584 page_mapcount_reset(page
); /* remove PageBuddy */
585 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
589 * Higher-order pages are called "compound pages". They are structured thusly:
591 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
593 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
594 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
596 * The first tail page's ->compound_dtor holds the offset in array of compound
597 * page destructors. See compound_page_dtors.
599 * The first tail page's ->compound_order holds the order of allocation.
600 * This usage means that zero-order pages may not be compound.
603 void free_compound_page(struct page
*page
)
605 __free_pages_ok(page
, compound_order(page
));
608 void prep_compound_page(struct page
*page
, unsigned int order
)
611 int nr_pages
= 1 << order
;
613 set_compound_page_dtor(page
, COMPOUND_PAGE_DTOR
);
614 set_compound_order(page
, order
);
616 for (i
= 1; i
< nr_pages
; i
++) {
617 struct page
*p
= page
+ i
;
618 set_page_count(p
, 0);
619 p
->mapping
= TAIL_MAPPING
;
620 set_compound_head(p
, page
);
622 atomic_set(compound_mapcount_ptr(page
), -1);
625 #ifdef CONFIG_DEBUG_PAGEALLOC
626 unsigned int _debug_guardpage_minorder
;
627 bool _debug_pagealloc_enabled __read_mostly
628 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT
);
629 EXPORT_SYMBOL(_debug_pagealloc_enabled
);
630 bool _debug_guardpage_enabled __read_mostly
;
632 static int __init
early_debug_pagealloc(char *buf
)
636 return kstrtobool(buf
, &_debug_pagealloc_enabled
);
638 early_param("debug_pagealloc", early_debug_pagealloc
);
640 static bool need_debug_guardpage(void)
642 /* If we don't use debug_pagealloc, we don't need guard page */
643 if (!debug_pagealloc_enabled())
646 if (!debug_guardpage_minorder())
652 static void init_debug_guardpage(void)
654 if (!debug_pagealloc_enabled())
657 if (!debug_guardpage_minorder())
660 _debug_guardpage_enabled
= true;
663 struct page_ext_operations debug_guardpage_ops
= {
664 .need
= need_debug_guardpage
,
665 .init
= init_debug_guardpage
,
668 static int __init
debug_guardpage_minorder_setup(char *buf
)
672 if (kstrtoul(buf
, 10, &res
) < 0 || res
> MAX_ORDER
/ 2) {
673 pr_err("Bad debug_guardpage_minorder value\n");
676 _debug_guardpage_minorder
= res
;
677 pr_info("Setting debug_guardpage_minorder to %lu\n", res
);
680 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup
);
682 static inline bool set_page_guard(struct zone
*zone
, struct page
*page
,
683 unsigned int order
, int migratetype
)
685 struct page_ext
*page_ext
;
687 if (!debug_guardpage_enabled())
690 if (order
>= debug_guardpage_minorder())
693 page_ext
= lookup_page_ext(page
);
694 if (unlikely(!page_ext
))
697 __set_bit(PAGE_EXT_DEBUG_GUARD
, &page_ext
->flags
);
699 INIT_LIST_HEAD(&page
->lru
);
700 set_page_private(page
, order
);
701 /* Guard pages are not available for any usage */
702 __mod_zone_freepage_state(zone
, -(1 << order
), migratetype
);
707 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
708 unsigned int order
, int migratetype
)
710 struct page_ext
*page_ext
;
712 if (!debug_guardpage_enabled())
715 page_ext
= lookup_page_ext(page
);
716 if (unlikely(!page_ext
))
719 __clear_bit(PAGE_EXT_DEBUG_GUARD
, &page_ext
->flags
);
721 set_page_private(page
, 0);
722 if (!is_migrate_isolate(migratetype
))
723 __mod_zone_freepage_state(zone
, (1 << order
), migratetype
);
726 struct page_ext_operations debug_guardpage_ops
;
727 static inline bool set_page_guard(struct zone
*zone
, struct page
*page
,
728 unsigned int order
, int migratetype
) { return false; }
729 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
730 unsigned int order
, int migratetype
) {}
733 static inline void set_page_order(struct page
*page
, unsigned int order
)
735 set_page_private(page
, order
);
736 __SetPageBuddy(page
);
739 static inline void rmv_page_order(struct page
*page
)
741 __ClearPageBuddy(page
);
742 set_page_private(page
, 0);
746 * This function checks whether a page is free && is the buddy
747 * we can do coalesce a page and its buddy if
748 * (a) the buddy is not in a hole (check before calling!) &&
749 * (b) the buddy is in the buddy system &&
750 * (c) a page and its buddy have the same order &&
751 * (d) a page and its buddy are in the same zone.
753 * For recording whether a page is in the buddy system, we set ->_mapcount
754 * PAGE_BUDDY_MAPCOUNT_VALUE.
755 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
756 * serialized by zone->lock.
758 * For recording page's order, we use page_private(page).
760 static inline int page_is_buddy(struct page
*page
, struct page
*buddy
,
763 if (page_is_guard(buddy
) && page_order(buddy
) == order
) {
764 if (page_zone_id(page
) != page_zone_id(buddy
))
767 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
772 if (PageBuddy(buddy
) && page_order(buddy
) == order
) {
774 * zone check is done late to avoid uselessly
775 * calculating zone/node ids for pages that could
778 if (page_zone_id(page
) != page_zone_id(buddy
))
781 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
789 * Freeing function for a buddy system allocator.
791 * The concept of a buddy system is to maintain direct-mapped table
792 * (containing bit values) for memory blocks of various "orders".
793 * The bottom level table contains the map for the smallest allocatable
794 * units of memory (here, pages), and each level above it describes
795 * pairs of units from the levels below, hence, "buddies".
796 * At a high level, all that happens here is marking the table entry
797 * at the bottom level available, and propagating the changes upward
798 * as necessary, plus some accounting needed to play nicely with other
799 * parts of the VM system.
800 * At each level, we keep a list of pages, which are heads of continuous
801 * free pages of length of (1 << order) and marked with _mapcount
802 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
804 * So when we are allocating or freeing one, we can derive the state of the
805 * other. That is, if we allocate a small block, and both were
806 * free, the remainder of the region must be split into blocks.
807 * If a block is freed, and its buddy is also free, then this
808 * triggers coalescing into a block of larger size.
813 static inline void __free_one_page(struct page
*page
,
815 struct zone
*zone
, unsigned int order
,
818 unsigned long combined_pfn
;
819 unsigned long uninitialized_var(buddy_pfn
);
821 unsigned int max_order
;
823 max_order
= min_t(unsigned int, MAX_ORDER
, pageblock_order
+ 1);
825 VM_BUG_ON(!zone_is_initialized(zone
));
826 VM_BUG_ON_PAGE(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
, page
);
828 VM_BUG_ON(migratetype
== -1);
829 if (likely(!is_migrate_isolate(migratetype
)))
830 __mod_zone_freepage_state(zone
, 1 << order
, migratetype
);
832 VM_BUG_ON_PAGE(pfn
& ((1 << order
) - 1), page
);
833 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
836 while (order
< max_order
- 1) {
837 buddy_pfn
= __find_buddy_pfn(pfn
, order
);
838 buddy
= page
+ (buddy_pfn
- pfn
);
840 if (!pfn_valid_within(buddy_pfn
))
842 if (!page_is_buddy(page
, buddy
, order
))
845 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
846 * merge with it and move up one order.
848 if (page_is_guard(buddy
)) {
849 clear_page_guard(zone
, buddy
, order
, migratetype
);
851 list_del(&buddy
->lru
);
852 zone
->free_area
[order
].nr_free
--;
853 rmv_page_order(buddy
);
855 combined_pfn
= buddy_pfn
& pfn
;
856 page
= page
+ (combined_pfn
- pfn
);
860 if (max_order
< MAX_ORDER
) {
861 /* If we are here, it means order is >= pageblock_order.
862 * We want to prevent merge between freepages on isolate
863 * pageblock and normal pageblock. Without this, pageblock
864 * isolation could cause incorrect freepage or CMA accounting.
866 * We don't want to hit this code for the more frequent
869 if (unlikely(has_isolate_pageblock(zone
))) {
872 buddy_pfn
= __find_buddy_pfn(pfn
, order
);
873 buddy
= page
+ (buddy_pfn
- pfn
);
874 buddy_mt
= get_pageblock_migratetype(buddy
);
876 if (migratetype
!= buddy_mt
877 && (is_migrate_isolate(migratetype
) ||
878 is_migrate_isolate(buddy_mt
)))
882 goto continue_merging
;
886 set_page_order(page
, order
);
889 * If this is not the largest possible page, check if the buddy
890 * of the next-highest order is free. If it is, it's possible
891 * that pages are being freed that will coalesce soon. In case,
892 * that is happening, add the free page to the tail of the list
893 * so it's less likely to be used soon and more likely to be merged
894 * as a higher order page
896 if ((order
< MAX_ORDER
-2) && pfn_valid_within(buddy_pfn
)) {
897 struct page
*higher_page
, *higher_buddy
;
898 combined_pfn
= buddy_pfn
& pfn
;
899 higher_page
= page
+ (combined_pfn
- pfn
);
900 buddy_pfn
= __find_buddy_pfn(combined_pfn
, order
+ 1);
901 higher_buddy
= higher_page
+ (buddy_pfn
- combined_pfn
);
902 if (pfn_valid_within(buddy_pfn
) &&
903 page_is_buddy(higher_page
, higher_buddy
, order
+ 1)) {
904 list_add_tail(&page
->lru
,
905 &zone
->free_area
[order
].free_list
[migratetype
]);
910 list_add(&page
->lru
, &zone
->free_area
[order
].free_list
[migratetype
]);
912 zone
->free_area
[order
].nr_free
++;
916 * A bad page could be due to a number of fields. Instead of multiple branches,
917 * try and check multiple fields with one check. The caller must do a detailed
918 * check if necessary.
920 static inline bool page_expected_state(struct page
*page
,
921 unsigned long check_flags
)
923 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
926 if (unlikely((unsigned long)page
->mapping
|
927 page_ref_count(page
) |
929 (unsigned long)page
->mem_cgroup
|
931 (page
->flags
& check_flags
)))
937 static void free_pages_check_bad(struct page
*page
)
939 const char *bad_reason
;
940 unsigned long bad_flags
;
945 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
946 bad_reason
= "nonzero mapcount";
947 if (unlikely(page
->mapping
!= NULL
))
948 bad_reason
= "non-NULL mapping";
949 if (unlikely(page_ref_count(page
) != 0))
950 bad_reason
= "nonzero _refcount";
951 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_FREE
)) {
952 bad_reason
= "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
953 bad_flags
= PAGE_FLAGS_CHECK_AT_FREE
;
956 if (unlikely(page
->mem_cgroup
))
957 bad_reason
= "page still charged to cgroup";
959 bad_page(page
, bad_reason
, bad_flags
);
962 static inline int free_pages_check(struct page
*page
)
964 if (likely(page_expected_state(page
, PAGE_FLAGS_CHECK_AT_FREE
)))
967 /* Something has gone sideways, find it */
968 free_pages_check_bad(page
);
972 static int free_tail_pages_check(struct page
*head_page
, struct page
*page
)
977 * We rely page->lru.next never has bit 0 set, unless the page
978 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
980 BUILD_BUG_ON((unsigned long)LIST_POISON1
& 1);
982 if (!IS_ENABLED(CONFIG_DEBUG_VM
)) {
986 switch (page
- head_page
) {
988 /* the first tail page: ->mapping is compound_mapcount() */
989 if (unlikely(compound_mapcount(page
))) {
990 bad_page(page
, "nonzero compound_mapcount", 0);
996 * the second tail page: ->mapping is
997 * page_deferred_list().next -- ignore value.
1001 if (page
->mapping
!= TAIL_MAPPING
) {
1002 bad_page(page
, "corrupted mapping in tail page", 0);
1007 if (unlikely(!PageTail(page
))) {
1008 bad_page(page
, "PageTail not set", 0);
1011 if (unlikely(compound_head(page
) != head_page
)) {
1012 bad_page(page
, "compound_head not consistent", 0);
1017 page
->mapping
= NULL
;
1018 clear_compound_head(page
);
1022 static __always_inline
bool free_pages_prepare(struct page
*page
,
1023 unsigned int order
, bool check_free
)
1027 VM_BUG_ON_PAGE(PageTail(page
), page
);
1029 trace_mm_page_free(page
, order
);
1032 * Check tail pages before head page information is cleared to
1033 * avoid checking PageCompound for order-0 pages.
1035 if (unlikely(order
)) {
1036 bool compound
= PageCompound(page
);
1039 VM_BUG_ON_PAGE(compound
&& compound_order(page
) != order
, page
);
1042 ClearPageDoubleMap(page
);
1043 for (i
= 1; i
< (1 << order
); i
++) {
1045 bad
+= free_tail_pages_check(page
, page
+ i
);
1046 if (unlikely(free_pages_check(page
+ i
))) {
1050 (page
+ i
)->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
1053 if (PageMappingFlags(page
))
1054 page
->mapping
= NULL
;
1055 if (memcg_kmem_enabled() && PageKmemcg(page
))
1056 memcg_kmem_uncharge(page
, order
);
1058 bad
+= free_pages_check(page
);
1062 page_cpupid_reset_last(page
);
1063 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
1064 reset_page_owner(page
, order
);
1066 if (!PageHighMem(page
)) {
1067 debug_check_no_locks_freed(page_address(page
),
1068 PAGE_SIZE
<< order
);
1069 debug_check_no_obj_freed(page_address(page
),
1070 PAGE_SIZE
<< order
);
1072 arch_free_page(page
, order
);
1073 kernel_poison_pages(page
, 1 << order
, 0);
1074 kernel_map_pages(page
, 1 << order
, 0);
1075 kasan_free_pages(page
, order
);
1080 #ifdef CONFIG_DEBUG_VM
1081 static inline bool free_pcp_prepare(struct page
*page
)
1083 return free_pages_prepare(page
, 0, true);
1086 static inline bool bulkfree_pcp_prepare(struct page
*page
)
1091 static bool free_pcp_prepare(struct page
*page
)
1093 return free_pages_prepare(page
, 0, false);
1096 static bool bulkfree_pcp_prepare(struct page
*page
)
1098 return free_pages_check(page
);
1100 #endif /* CONFIG_DEBUG_VM */
1103 * Frees a number of pages from the PCP lists
1104 * Assumes all pages on list are in same zone, and of same order.
1105 * count is the number of pages to free.
1107 * If the zone was previously in an "all pages pinned" state then look to
1108 * see if this freeing clears that state.
1110 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1111 * pinned" detection logic.
1113 static void free_pcppages_bulk(struct zone
*zone
, int count
,
1114 struct per_cpu_pages
*pcp
)
1116 int migratetype
= 0;
1118 bool isolated_pageblocks
;
1120 spin_lock(&zone
->lock
);
1121 isolated_pageblocks
= has_isolate_pageblock(zone
);
1125 struct list_head
*list
;
1128 * Remove pages from lists in a round-robin fashion. A
1129 * batch_free count is maintained that is incremented when an
1130 * empty list is encountered. This is so more pages are freed
1131 * off fuller lists instead of spinning excessively around empty
1136 if (++migratetype
== MIGRATE_PCPTYPES
)
1138 list
= &pcp
->lists
[migratetype
];
1139 } while (list_empty(list
));
1141 /* This is the only non-empty list. Free them all. */
1142 if (batch_free
== MIGRATE_PCPTYPES
)
1146 int mt
; /* migratetype of the to-be-freed page */
1148 page
= list_last_entry(list
, struct page
, lru
);
1149 /* must delete as __free_one_page list manipulates */
1150 list_del(&page
->lru
);
1152 mt
= get_pcppage_migratetype(page
);
1153 /* MIGRATE_ISOLATE page should not go to pcplists */
1154 VM_BUG_ON_PAGE(is_migrate_isolate(mt
), page
);
1155 /* Pageblock could have been isolated meanwhile */
1156 if (unlikely(isolated_pageblocks
))
1157 mt
= get_pageblock_migratetype(page
);
1159 if (bulkfree_pcp_prepare(page
))
1162 __free_one_page(page
, page_to_pfn(page
), zone
, 0, mt
);
1163 trace_mm_page_pcpu_drain(page
, 0, mt
);
1164 } while (--count
&& --batch_free
&& !list_empty(list
));
1166 spin_unlock(&zone
->lock
);
1169 static void free_one_page(struct zone
*zone
,
1170 struct page
*page
, unsigned long pfn
,
1174 spin_lock(&zone
->lock
);
1175 if (unlikely(has_isolate_pageblock(zone
) ||
1176 is_migrate_isolate(migratetype
))) {
1177 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1179 __free_one_page(page
, pfn
, zone
, order
, migratetype
);
1180 spin_unlock(&zone
->lock
);
1183 static void __meminit
__init_single_page(struct page
*page
, unsigned long pfn
,
1184 unsigned long zone
, int nid
, bool zero
)
1187 mm_zero_struct_page(page
);
1188 set_page_links(page
, zone
, nid
, pfn
);
1189 init_page_count(page
);
1190 page_mapcount_reset(page
);
1191 page_cpupid_reset_last(page
);
1193 INIT_LIST_HEAD(&page
->lru
);
1194 #ifdef WANT_PAGE_VIRTUAL
1195 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1196 if (!is_highmem_idx(zone
))
1197 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
1201 static void __meminit
__init_single_pfn(unsigned long pfn
, unsigned long zone
,
1204 return __init_single_page(pfn_to_page(pfn
), pfn
, zone
, nid
, zero
);
1207 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1208 static void __meminit
init_reserved_page(unsigned long pfn
)
1213 if (!early_page_uninitialised(pfn
))
1216 nid
= early_pfn_to_nid(pfn
);
1217 pgdat
= NODE_DATA(nid
);
1219 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1220 struct zone
*zone
= &pgdat
->node_zones
[zid
];
1222 if (pfn
>= zone
->zone_start_pfn
&& pfn
< zone_end_pfn(zone
))
1225 __init_single_pfn(pfn
, zid
, nid
, true);
1228 static inline void init_reserved_page(unsigned long pfn
)
1231 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1234 * Initialised pages do not have PageReserved set. This function is
1235 * called for each range allocated by the bootmem allocator and
1236 * marks the pages PageReserved. The remaining valid pages are later
1237 * sent to the buddy page allocator.
1239 void __meminit
reserve_bootmem_region(phys_addr_t start
, phys_addr_t end
)
1241 unsigned long start_pfn
= PFN_DOWN(start
);
1242 unsigned long end_pfn
= PFN_UP(end
);
1244 for (; start_pfn
< end_pfn
; start_pfn
++) {
1245 if (pfn_valid(start_pfn
)) {
1246 struct page
*page
= pfn_to_page(start_pfn
);
1248 init_reserved_page(start_pfn
);
1250 /* Avoid false-positive PageTail() */
1251 INIT_LIST_HEAD(&page
->lru
);
1253 SetPageReserved(page
);
1258 static void __free_pages_ok(struct page
*page
, unsigned int order
)
1260 unsigned long flags
;
1262 unsigned long pfn
= page_to_pfn(page
);
1264 if (!free_pages_prepare(page
, order
, true))
1267 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1268 local_irq_save(flags
);
1269 __count_vm_events(PGFREE
, 1 << order
);
1270 free_one_page(page_zone(page
), page
, pfn
, order
, migratetype
);
1271 local_irq_restore(flags
);
1274 static void __init
__free_pages_boot_core(struct page
*page
, unsigned int order
)
1276 unsigned int nr_pages
= 1 << order
;
1277 struct page
*p
= page
;
1281 for (loop
= 0; loop
< (nr_pages
- 1); loop
++, p
++) {
1283 __ClearPageReserved(p
);
1284 set_page_count(p
, 0);
1286 __ClearPageReserved(p
);
1287 set_page_count(p
, 0);
1289 page_zone(page
)->managed_pages
+= nr_pages
;
1290 set_page_refcounted(page
);
1291 __free_pages(page
, order
);
1294 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1295 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1297 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata
;
1299 int __meminit
early_pfn_to_nid(unsigned long pfn
)
1301 static DEFINE_SPINLOCK(early_pfn_lock
);
1304 spin_lock(&early_pfn_lock
);
1305 nid
= __early_pfn_to_nid(pfn
, &early_pfnnid_cache
);
1307 nid
= first_online_node
;
1308 spin_unlock(&early_pfn_lock
);
1314 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1315 static inline bool __meminit __maybe_unused
1316 meminit_pfn_in_nid(unsigned long pfn
, int node
,
1317 struct mminit_pfnnid_cache
*state
)
1321 nid
= __early_pfn_to_nid(pfn
, state
);
1322 if (nid
>= 0 && nid
!= node
)
1327 /* Only safe to use early in boot when initialisation is single-threaded */
1328 static inline bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
1330 return meminit_pfn_in_nid(pfn
, node
, &early_pfnnid_cache
);
1335 static inline bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
1339 static inline bool __meminit __maybe_unused
1340 meminit_pfn_in_nid(unsigned long pfn
, int node
,
1341 struct mminit_pfnnid_cache
*state
)
1348 void __init
__free_pages_bootmem(struct page
*page
, unsigned long pfn
,
1351 if (early_page_uninitialised(pfn
))
1353 return __free_pages_boot_core(page
, order
);
1357 * Check that the whole (or subset of) a pageblock given by the interval of
1358 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1359 * with the migration of free compaction scanner. The scanners then need to
1360 * use only pfn_valid_within() check for arches that allow holes within
1363 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1365 * It's possible on some configurations to have a setup like node0 node1 node0
1366 * i.e. it's possible that all pages within a zones range of pages do not
1367 * belong to a single zone. We assume that a border between node0 and node1
1368 * can occur within a single pageblock, but not a node0 node1 node0
1369 * interleaving within a single pageblock. It is therefore sufficient to check
1370 * the first and last page of a pageblock and avoid checking each individual
1371 * page in a pageblock.
1373 struct page
*__pageblock_pfn_to_page(unsigned long start_pfn
,
1374 unsigned long end_pfn
, struct zone
*zone
)
1376 struct page
*start_page
;
1377 struct page
*end_page
;
1379 /* end_pfn is one past the range we are checking */
1382 if (!pfn_valid(start_pfn
) || !pfn_valid(end_pfn
))
1385 start_page
= pfn_to_online_page(start_pfn
);
1389 if (page_zone(start_page
) != zone
)
1392 end_page
= pfn_to_page(end_pfn
);
1394 /* This gives a shorter code than deriving page_zone(end_page) */
1395 if (page_zone_id(start_page
) != page_zone_id(end_page
))
1401 void set_zone_contiguous(struct zone
*zone
)
1403 unsigned long block_start_pfn
= zone
->zone_start_pfn
;
1404 unsigned long block_end_pfn
;
1406 block_end_pfn
= ALIGN(block_start_pfn
+ 1, pageblock_nr_pages
);
1407 for (; block_start_pfn
< zone_end_pfn(zone
);
1408 block_start_pfn
= block_end_pfn
,
1409 block_end_pfn
+= pageblock_nr_pages
) {
1411 block_end_pfn
= min(block_end_pfn
, zone_end_pfn(zone
));
1413 if (!__pageblock_pfn_to_page(block_start_pfn
,
1414 block_end_pfn
, zone
))
1418 /* We confirm that there is no hole */
1419 zone
->contiguous
= true;
1422 void clear_zone_contiguous(struct zone
*zone
)
1424 zone
->contiguous
= false;
1427 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1428 static void __init
deferred_free_range(unsigned long pfn
,
1429 unsigned long nr_pages
)
1437 page
= pfn_to_page(pfn
);
1439 /* Free a large naturally-aligned chunk if possible */
1440 if (nr_pages
== pageblock_nr_pages
&&
1441 (pfn
& (pageblock_nr_pages
- 1)) == 0) {
1442 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
1443 __free_pages_boot_core(page
, pageblock_order
);
1447 for (i
= 0; i
< nr_pages
; i
++, page
++, pfn
++) {
1448 if ((pfn
& (pageblock_nr_pages
- 1)) == 0)
1449 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
1450 __free_pages_boot_core(page
, 0);
1454 /* Completion tracking for deferred_init_memmap() threads */
1455 static atomic_t pgdat_init_n_undone __initdata
;
1456 static __initdata
DECLARE_COMPLETION(pgdat_init_all_done_comp
);
1458 static inline void __init
pgdat_init_report_one_done(void)
1460 if (atomic_dec_and_test(&pgdat_init_n_undone
))
1461 complete(&pgdat_init_all_done_comp
);
1465 * Helper for deferred_init_range, free the given range, reset the counters, and
1466 * return number of pages freed.
1468 static inline unsigned long __init
__def_free(unsigned long *nr_free
,
1469 unsigned long *free_base_pfn
,
1472 unsigned long nr
= *nr_free
;
1474 deferred_free_range(*free_base_pfn
, nr
);
1482 static unsigned long __init
deferred_init_range(int nid
, int zid
,
1483 unsigned long start_pfn
,
1484 unsigned long end_pfn
)
1486 struct mminit_pfnnid_cache nid_init_state
= { };
1487 unsigned long nr_pgmask
= pageblock_nr_pages
- 1;
1488 unsigned long free_base_pfn
= 0;
1489 unsigned long nr_pages
= 0;
1490 unsigned long nr_free
= 0;
1491 struct page
*page
= NULL
;
1495 * First we check if pfn is valid on architectures where it is possible
1496 * to have holes within pageblock_nr_pages. On systems where it is not
1497 * possible, this function is optimized out.
1499 * Then, we check if a current large page is valid by only checking the
1500 * validity of the head pfn.
1502 * meminit_pfn_in_nid is checked on systems where pfns can interleave
1503 * within a node: a pfn is between start and end of a node, but does not
1504 * belong to this memory node.
1506 * Finally, we minimize pfn page lookups and scheduler checks by
1507 * performing it only once every pageblock_nr_pages.
1509 * We do it in two loops: first we initialize struct page, than free to
1510 * buddy allocator, becuse while we are freeing pages we can access
1511 * pages that are ahead (computing buddy page in __free_one_page()).
1513 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
1514 if (!pfn_valid_within(pfn
))
1516 if ((pfn
& nr_pgmask
) || pfn_valid(pfn
)) {
1517 if (meminit_pfn_in_nid(pfn
, nid
, &nid_init_state
)) {
1518 if (page
&& (pfn
& nr_pgmask
))
1521 page
= pfn_to_page(pfn
);
1522 __init_single_page(page
, pfn
, zid
, nid
, true);
1529 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
1530 if (!pfn_valid_within(pfn
)) {
1531 nr_pages
+= __def_free(&nr_free
, &free_base_pfn
, &page
);
1532 } else if (!(pfn
& nr_pgmask
) && !pfn_valid(pfn
)) {
1533 nr_pages
+= __def_free(&nr_free
, &free_base_pfn
, &page
);
1534 } else if (!meminit_pfn_in_nid(pfn
, nid
, &nid_init_state
)) {
1535 nr_pages
+= __def_free(&nr_free
, &free_base_pfn
, &page
);
1536 } else if (page
&& (pfn
& nr_pgmask
)) {
1540 nr_pages
+= __def_free(&nr_free
, &free_base_pfn
, &page
);
1541 page
= pfn_to_page(pfn
);
1542 free_base_pfn
= pfn
;
1547 /* Free the last block of pages to allocator */
1548 nr_pages
+= __def_free(&nr_free
, &free_base_pfn
, &page
);
1553 /* Initialise remaining memory on a node */
1554 static int __init
deferred_init_memmap(void *data
)
1556 pg_data_t
*pgdat
= data
;
1557 int nid
= pgdat
->node_id
;
1558 unsigned long start
= jiffies
;
1559 unsigned long nr_pages
= 0;
1560 unsigned long spfn
, epfn
;
1561 phys_addr_t spa
, epa
;
1564 unsigned long first_init_pfn
= pgdat
->first_deferred_pfn
;
1565 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
1568 if (first_init_pfn
== ULONG_MAX
) {
1569 pgdat_init_report_one_done();
1573 /* Bind memory initialisation thread to a local node if possible */
1574 if (!cpumask_empty(cpumask
))
1575 set_cpus_allowed_ptr(current
, cpumask
);
1577 /* Sanity check boundaries */
1578 BUG_ON(pgdat
->first_deferred_pfn
< pgdat
->node_start_pfn
);
1579 BUG_ON(pgdat
->first_deferred_pfn
> pgdat_end_pfn(pgdat
));
1580 pgdat
->first_deferred_pfn
= ULONG_MAX
;
1582 /* Only the highest zone is deferred so find it */
1583 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1584 zone
= pgdat
->node_zones
+ zid
;
1585 if (first_init_pfn
< zone_end_pfn(zone
))
1588 first_init_pfn
= max(zone
->zone_start_pfn
, first_init_pfn
);
1590 for_each_free_mem_range(i
, nid
, MEMBLOCK_NONE
, &spa
, &epa
, NULL
) {
1591 spfn
= max_t(unsigned long, first_init_pfn
, PFN_UP(spa
));
1592 epfn
= min_t(unsigned long, zone_end_pfn(zone
), PFN_DOWN(epa
));
1593 nr_pages
+= deferred_init_range(nid
, zid
, spfn
, epfn
);
1596 /* Sanity check that the next zone really is unpopulated */
1597 WARN_ON(++zid
< MAX_NR_ZONES
&& populated_zone(++zone
));
1599 pr_info("node %d initialised, %lu pages in %ums\n", nid
, nr_pages
,
1600 jiffies_to_msecs(jiffies
- start
));
1602 pgdat_init_report_one_done();
1605 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1607 void __init
page_alloc_init_late(void)
1611 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1614 /* There will be num_node_state(N_MEMORY) threads */
1615 atomic_set(&pgdat_init_n_undone
, num_node_state(N_MEMORY
));
1616 for_each_node_state(nid
, N_MEMORY
) {
1617 kthread_run(deferred_init_memmap
, NODE_DATA(nid
), "pgdatinit%d", nid
);
1620 /* Block until all are initialised */
1621 wait_for_completion(&pgdat_init_all_done_comp
);
1623 /* Reinit limits that are based on free pages after the kernel is up */
1624 files_maxfiles_init();
1626 #ifdef CONFIG_ARCH_DISCARD_MEMBLOCK
1627 /* Discard memblock private memory */
1631 for_each_populated_zone(zone
)
1632 set_zone_contiguous(zone
);
1636 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1637 void __init
init_cma_reserved_pageblock(struct page
*page
)
1639 unsigned i
= pageblock_nr_pages
;
1640 struct page
*p
= page
;
1643 __ClearPageReserved(p
);
1644 set_page_count(p
, 0);
1647 set_pageblock_migratetype(page
, MIGRATE_CMA
);
1649 if (pageblock_order
>= MAX_ORDER
) {
1650 i
= pageblock_nr_pages
;
1653 set_page_refcounted(p
);
1654 __free_pages(p
, MAX_ORDER
- 1);
1655 p
+= MAX_ORDER_NR_PAGES
;
1656 } while (i
-= MAX_ORDER_NR_PAGES
);
1658 set_page_refcounted(page
);
1659 __free_pages(page
, pageblock_order
);
1662 adjust_managed_page_count(page
, pageblock_nr_pages
);
1667 * The order of subdivision here is critical for the IO subsystem.
1668 * Please do not alter this order without good reasons and regression
1669 * testing. Specifically, as large blocks of memory are subdivided,
1670 * the order in which smaller blocks are delivered depends on the order
1671 * they're subdivided in this function. This is the primary factor
1672 * influencing the order in which pages are delivered to the IO
1673 * subsystem according to empirical testing, and this is also justified
1674 * by considering the behavior of a buddy system containing a single
1675 * large block of memory acted on by a series of small allocations.
1676 * This behavior is a critical factor in sglist merging's success.
1680 static inline void expand(struct zone
*zone
, struct page
*page
,
1681 int low
, int high
, struct free_area
*area
,
1684 unsigned long size
= 1 << high
;
1686 while (high
> low
) {
1690 VM_BUG_ON_PAGE(bad_range(zone
, &page
[size
]), &page
[size
]);
1693 * Mark as guard pages (or page), that will allow to
1694 * merge back to allocator when buddy will be freed.
1695 * Corresponding page table entries will not be touched,
1696 * pages will stay not present in virtual address space
1698 if (set_page_guard(zone
, &page
[size
], high
, migratetype
))
1701 list_add(&page
[size
].lru
, &area
->free_list
[migratetype
]);
1703 set_page_order(&page
[size
], high
);
1707 static void check_new_page_bad(struct page
*page
)
1709 const char *bad_reason
= NULL
;
1710 unsigned long bad_flags
= 0;
1712 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
1713 bad_reason
= "nonzero mapcount";
1714 if (unlikely(page
->mapping
!= NULL
))
1715 bad_reason
= "non-NULL mapping";
1716 if (unlikely(page_ref_count(page
) != 0))
1717 bad_reason
= "nonzero _count";
1718 if (unlikely(page
->flags
& __PG_HWPOISON
)) {
1719 bad_reason
= "HWPoisoned (hardware-corrupted)";
1720 bad_flags
= __PG_HWPOISON
;
1721 /* Don't complain about hwpoisoned pages */
1722 page_mapcount_reset(page
); /* remove PageBuddy */
1725 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)) {
1726 bad_reason
= "PAGE_FLAGS_CHECK_AT_PREP flag set";
1727 bad_flags
= PAGE_FLAGS_CHECK_AT_PREP
;
1730 if (unlikely(page
->mem_cgroup
))
1731 bad_reason
= "page still charged to cgroup";
1733 bad_page(page
, bad_reason
, bad_flags
);
1737 * This page is about to be returned from the page allocator
1739 static inline int check_new_page(struct page
*page
)
1741 if (likely(page_expected_state(page
,
1742 PAGE_FLAGS_CHECK_AT_PREP
|__PG_HWPOISON
)))
1745 check_new_page_bad(page
);
1749 static inline bool free_pages_prezeroed(void)
1751 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO
) &&
1752 page_poisoning_enabled();
1755 #ifdef CONFIG_DEBUG_VM
1756 static bool check_pcp_refill(struct page
*page
)
1761 static bool check_new_pcp(struct page
*page
)
1763 return check_new_page(page
);
1766 static bool check_pcp_refill(struct page
*page
)
1768 return check_new_page(page
);
1770 static bool check_new_pcp(struct page
*page
)
1774 #endif /* CONFIG_DEBUG_VM */
1776 static bool check_new_pages(struct page
*page
, unsigned int order
)
1779 for (i
= 0; i
< (1 << order
); i
++) {
1780 struct page
*p
= page
+ i
;
1782 if (unlikely(check_new_page(p
)))
1789 inline void post_alloc_hook(struct page
*page
, unsigned int order
,
1792 set_page_private(page
, 0);
1793 set_page_refcounted(page
);
1795 arch_alloc_page(page
, order
);
1796 kernel_map_pages(page
, 1 << order
, 1);
1797 kernel_poison_pages(page
, 1 << order
, 1);
1798 kasan_alloc_pages(page
, order
);
1799 set_page_owner(page
, order
, gfp_flags
);
1802 static void prep_new_page(struct page
*page
, unsigned int order
, gfp_t gfp_flags
,
1803 unsigned int alloc_flags
)
1807 post_alloc_hook(page
, order
, gfp_flags
);
1809 if (!free_pages_prezeroed() && (gfp_flags
& __GFP_ZERO
))
1810 for (i
= 0; i
< (1 << order
); i
++)
1811 clear_highpage(page
+ i
);
1813 if (order
&& (gfp_flags
& __GFP_COMP
))
1814 prep_compound_page(page
, order
);
1817 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1818 * allocate the page. The expectation is that the caller is taking
1819 * steps that will free more memory. The caller should avoid the page
1820 * being used for !PFMEMALLOC purposes.
1822 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
1823 set_page_pfmemalloc(page
);
1825 clear_page_pfmemalloc(page
);
1829 * Go through the free lists for the given migratetype and remove
1830 * the smallest available page from the freelists
1832 static __always_inline
1833 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
1836 unsigned int current_order
;
1837 struct free_area
*area
;
1840 /* Find a page of the appropriate size in the preferred list */
1841 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
1842 area
= &(zone
->free_area
[current_order
]);
1843 page
= list_first_entry_or_null(&area
->free_list
[migratetype
],
1847 list_del(&page
->lru
);
1848 rmv_page_order(page
);
1850 expand(zone
, page
, order
, current_order
, area
, migratetype
);
1851 set_pcppage_migratetype(page
, migratetype
);
1860 * This array describes the order lists are fallen back to when
1861 * the free lists for the desirable migrate type are depleted
1863 static int fallbacks
[MIGRATE_TYPES
][4] = {
1864 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_TYPES
},
1865 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_TYPES
},
1866 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_TYPES
},
1868 [MIGRATE_CMA
] = { MIGRATE_TYPES
}, /* Never used */
1870 #ifdef CONFIG_MEMORY_ISOLATION
1871 [MIGRATE_ISOLATE
] = { MIGRATE_TYPES
}, /* Never used */
1876 static __always_inline
struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
1879 return __rmqueue_smallest(zone
, order
, MIGRATE_CMA
);
1882 static inline struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
1883 unsigned int order
) { return NULL
; }
1887 * Move the free pages in a range to the free lists of the requested type.
1888 * Note that start_page and end_pages are not aligned on a pageblock
1889 * boundary. If alignment is required, use move_freepages_block()
1891 static int move_freepages(struct zone
*zone
,
1892 struct page
*start_page
, struct page
*end_page
,
1893 int migratetype
, int *num_movable
)
1897 int pages_moved
= 0;
1899 #ifndef CONFIG_HOLES_IN_ZONE
1901 * page_zone is not safe to call in this context when
1902 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1903 * anyway as we check zone boundaries in move_freepages_block().
1904 * Remove at a later date when no bug reports exist related to
1905 * grouping pages by mobility
1907 VM_BUG_ON(page_zone(start_page
) != page_zone(end_page
));
1913 for (page
= start_page
; page
<= end_page
;) {
1914 if (!pfn_valid_within(page_to_pfn(page
))) {
1919 /* Make sure we are not inadvertently changing nodes */
1920 VM_BUG_ON_PAGE(page_to_nid(page
) != zone_to_nid(zone
), page
);
1922 if (!PageBuddy(page
)) {
1924 * We assume that pages that could be isolated for
1925 * migration are movable. But we don't actually try
1926 * isolating, as that would be expensive.
1929 (PageLRU(page
) || __PageMovable(page
)))
1936 order
= page_order(page
);
1937 list_move(&page
->lru
,
1938 &zone
->free_area
[order
].free_list
[migratetype
]);
1940 pages_moved
+= 1 << order
;
1946 int move_freepages_block(struct zone
*zone
, struct page
*page
,
1947 int migratetype
, int *num_movable
)
1949 unsigned long start_pfn
, end_pfn
;
1950 struct page
*start_page
, *end_page
;
1952 start_pfn
= page_to_pfn(page
);
1953 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
1954 start_page
= pfn_to_page(start_pfn
);
1955 end_page
= start_page
+ pageblock_nr_pages
- 1;
1956 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
1958 /* Do not cross zone boundaries */
1959 if (!zone_spans_pfn(zone
, start_pfn
))
1961 if (!zone_spans_pfn(zone
, end_pfn
))
1964 return move_freepages(zone
, start_page
, end_page
, migratetype
,
1968 static void change_pageblock_range(struct page
*pageblock_page
,
1969 int start_order
, int migratetype
)
1971 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
1973 while (nr_pageblocks
--) {
1974 set_pageblock_migratetype(pageblock_page
, migratetype
);
1975 pageblock_page
+= pageblock_nr_pages
;
1980 * When we are falling back to another migratetype during allocation, try to
1981 * steal extra free pages from the same pageblocks to satisfy further
1982 * allocations, instead of polluting multiple pageblocks.
1984 * If we are stealing a relatively large buddy page, it is likely there will
1985 * be more free pages in the pageblock, so try to steal them all. For
1986 * reclaimable and unmovable allocations, we steal regardless of page size,
1987 * as fragmentation caused by those allocations polluting movable pageblocks
1988 * is worse than movable allocations stealing from unmovable and reclaimable
1991 static bool can_steal_fallback(unsigned int order
, int start_mt
)
1994 * Leaving this order check is intended, although there is
1995 * relaxed order check in next check. The reason is that
1996 * we can actually steal whole pageblock if this condition met,
1997 * but, below check doesn't guarantee it and that is just heuristic
1998 * so could be changed anytime.
2000 if (order
>= pageblock_order
)
2003 if (order
>= pageblock_order
/ 2 ||
2004 start_mt
== MIGRATE_RECLAIMABLE
||
2005 start_mt
== MIGRATE_UNMOVABLE
||
2006 page_group_by_mobility_disabled
)
2013 * This function implements actual steal behaviour. If order is large enough,
2014 * we can steal whole pageblock. If not, we first move freepages in this
2015 * pageblock to our migratetype and determine how many already-allocated pages
2016 * are there in the pageblock with a compatible migratetype. If at least half
2017 * of pages are free or compatible, we can change migratetype of the pageblock
2018 * itself, so pages freed in the future will be put on the correct free list.
2020 static void steal_suitable_fallback(struct zone
*zone
, struct page
*page
,
2021 int start_type
, bool whole_block
)
2023 unsigned int current_order
= page_order(page
);
2024 struct free_area
*area
;
2025 int free_pages
, movable_pages
, alike_pages
;
2028 old_block_type
= get_pageblock_migratetype(page
);
2031 * This can happen due to races and we want to prevent broken
2032 * highatomic accounting.
2034 if (is_migrate_highatomic(old_block_type
))
2037 /* Take ownership for orders >= pageblock_order */
2038 if (current_order
>= pageblock_order
) {
2039 change_pageblock_range(page
, current_order
, start_type
);
2043 /* We are not allowed to try stealing from the whole block */
2047 free_pages
= move_freepages_block(zone
, page
, start_type
,
2050 * Determine how many pages are compatible with our allocation.
2051 * For movable allocation, it's the number of movable pages which
2052 * we just obtained. For other types it's a bit more tricky.
2054 if (start_type
== MIGRATE_MOVABLE
) {
2055 alike_pages
= movable_pages
;
2058 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2059 * to MOVABLE pageblock, consider all non-movable pages as
2060 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2061 * vice versa, be conservative since we can't distinguish the
2062 * exact migratetype of non-movable pages.
2064 if (old_block_type
== MIGRATE_MOVABLE
)
2065 alike_pages
= pageblock_nr_pages
2066 - (free_pages
+ movable_pages
);
2071 /* moving whole block can fail due to zone boundary conditions */
2076 * If a sufficient number of pages in the block are either free or of
2077 * comparable migratability as our allocation, claim the whole block.
2079 if (free_pages
+ alike_pages
>= (1 << (pageblock_order
-1)) ||
2080 page_group_by_mobility_disabled
)
2081 set_pageblock_migratetype(page
, start_type
);
2086 area
= &zone
->free_area
[current_order
];
2087 list_move(&page
->lru
, &area
->free_list
[start_type
]);
2091 * Check whether there is a suitable fallback freepage with requested order.
2092 * If only_stealable is true, this function returns fallback_mt only if
2093 * we can steal other freepages all together. This would help to reduce
2094 * fragmentation due to mixed migratetype pages in one pageblock.
2096 int find_suitable_fallback(struct free_area
*area
, unsigned int order
,
2097 int migratetype
, bool only_stealable
, bool *can_steal
)
2102 if (area
->nr_free
== 0)
2107 fallback_mt
= fallbacks
[migratetype
][i
];
2108 if (fallback_mt
== MIGRATE_TYPES
)
2111 if (list_empty(&area
->free_list
[fallback_mt
]))
2114 if (can_steal_fallback(order
, migratetype
))
2117 if (!only_stealable
)
2128 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2129 * there are no empty page blocks that contain a page with a suitable order
2131 static void reserve_highatomic_pageblock(struct page
*page
, struct zone
*zone
,
2132 unsigned int alloc_order
)
2135 unsigned long max_managed
, flags
;
2138 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2139 * Check is race-prone but harmless.
2141 max_managed
= (zone
->managed_pages
/ 100) + pageblock_nr_pages
;
2142 if (zone
->nr_reserved_highatomic
>= max_managed
)
2145 spin_lock_irqsave(&zone
->lock
, flags
);
2147 /* Recheck the nr_reserved_highatomic limit under the lock */
2148 if (zone
->nr_reserved_highatomic
>= max_managed
)
2152 mt
= get_pageblock_migratetype(page
);
2153 if (!is_migrate_highatomic(mt
) && !is_migrate_isolate(mt
)
2154 && !is_migrate_cma(mt
)) {
2155 zone
->nr_reserved_highatomic
+= pageblock_nr_pages
;
2156 set_pageblock_migratetype(page
, MIGRATE_HIGHATOMIC
);
2157 move_freepages_block(zone
, page
, MIGRATE_HIGHATOMIC
, NULL
);
2161 spin_unlock_irqrestore(&zone
->lock
, flags
);
2165 * Used when an allocation is about to fail under memory pressure. This
2166 * potentially hurts the reliability of high-order allocations when under
2167 * intense memory pressure but failed atomic allocations should be easier
2168 * to recover from than an OOM.
2170 * If @force is true, try to unreserve a pageblock even though highatomic
2171 * pageblock is exhausted.
2173 static bool unreserve_highatomic_pageblock(const struct alloc_context
*ac
,
2176 struct zonelist
*zonelist
= ac
->zonelist
;
2177 unsigned long flags
;
2184 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, ac
->high_zoneidx
,
2187 * Preserve at least one pageblock unless memory pressure
2190 if (!force
&& zone
->nr_reserved_highatomic
<=
2194 spin_lock_irqsave(&zone
->lock
, flags
);
2195 for (order
= 0; order
< MAX_ORDER
; order
++) {
2196 struct free_area
*area
= &(zone
->free_area
[order
]);
2198 page
= list_first_entry_or_null(
2199 &area
->free_list
[MIGRATE_HIGHATOMIC
],
2205 * In page freeing path, migratetype change is racy so
2206 * we can counter several free pages in a pageblock
2207 * in this loop althoug we changed the pageblock type
2208 * from highatomic to ac->migratetype. So we should
2209 * adjust the count once.
2211 if (is_migrate_highatomic_page(page
)) {
2213 * It should never happen but changes to
2214 * locking could inadvertently allow a per-cpu
2215 * drain to add pages to MIGRATE_HIGHATOMIC
2216 * while unreserving so be safe and watch for
2219 zone
->nr_reserved_highatomic
-= min(
2221 zone
->nr_reserved_highatomic
);
2225 * Convert to ac->migratetype and avoid the normal
2226 * pageblock stealing heuristics. Minimally, the caller
2227 * is doing the work and needs the pages. More
2228 * importantly, if the block was always converted to
2229 * MIGRATE_UNMOVABLE or another type then the number
2230 * of pageblocks that cannot be completely freed
2233 set_pageblock_migratetype(page
, ac
->migratetype
);
2234 ret
= move_freepages_block(zone
, page
, ac
->migratetype
,
2237 spin_unlock_irqrestore(&zone
->lock
, flags
);
2241 spin_unlock_irqrestore(&zone
->lock
, flags
);
2248 * Try finding a free buddy page on the fallback list and put it on the free
2249 * list of requested migratetype, possibly along with other pages from the same
2250 * block, depending on fragmentation avoidance heuristics. Returns true if
2251 * fallback was found so that __rmqueue_smallest() can grab it.
2253 * The use of signed ints for order and current_order is a deliberate
2254 * deviation from the rest of this file, to make the for loop
2255 * condition simpler.
2257 static __always_inline
bool
2258 __rmqueue_fallback(struct zone
*zone
, int order
, int start_migratetype
)
2260 struct free_area
*area
;
2267 * Find the largest available free page in the other list. This roughly
2268 * approximates finding the pageblock with the most free pages, which
2269 * would be too costly to do exactly.
2271 for (current_order
= MAX_ORDER
- 1; current_order
>= order
;
2273 area
= &(zone
->free_area
[current_order
]);
2274 fallback_mt
= find_suitable_fallback(area
, current_order
,
2275 start_migratetype
, false, &can_steal
);
2276 if (fallback_mt
== -1)
2280 * We cannot steal all free pages from the pageblock and the
2281 * requested migratetype is movable. In that case it's better to
2282 * steal and split the smallest available page instead of the
2283 * largest available page, because even if the next movable
2284 * allocation falls back into a different pageblock than this
2285 * one, it won't cause permanent fragmentation.
2287 if (!can_steal
&& start_migratetype
== MIGRATE_MOVABLE
2288 && current_order
> order
)
2297 for (current_order
= order
; current_order
< MAX_ORDER
;
2299 area
= &(zone
->free_area
[current_order
]);
2300 fallback_mt
= find_suitable_fallback(area
, current_order
,
2301 start_migratetype
, false, &can_steal
);
2302 if (fallback_mt
!= -1)
2307 * This should not happen - we already found a suitable fallback
2308 * when looking for the largest page.
2310 VM_BUG_ON(current_order
== MAX_ORDER
);
2313 page
= list_first_entry(&area
->free_list
[fallback_mt
],
2316 steal_suitable_fallback(zone
, page
, start_migratetype
, can_steal
);
2318 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
2319 start_migratetype
, fallback_mt
);
2326 * Do the hard work of removing an element from the buddy allocator.
2327 * Call me with the zone->lock already held.
2329 static __always_inline
struct page
*
2330 __rmqueue(struct zone
*zone
, unsigned int order
, int migratetype
)
2335 page
= __rmqueue_smallest(zone
, order
, migratetype
);
2336 if (unlikely(!page
)) {
2337 if (migratetype
== MIGRATE_MOVABLE
)
2338 page
= __rmqueue_cma_fallback(zone
, order
);
2340 if (!page
&& __rmqueue_fallback(zone
, order
, migratetype
))
2344 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
2349 * Obtain a specified number of elements from the buddy allocator, all under
2350 * a single hold of the lock, for efficiency. Add them to the supplied list.
2351 * Returns the number of new pages which were placed at *list.
2353 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
2354 unsigned long count
, struct list_head
*list
,
2359 spin_lock(&zone
->lock
);
2360 for (i
= 0; i
< count
; ++i
) {
2361 struct page
*page
= __rmqueue(zone
, order
, migratetype
);
2362 if (unlikely(page
== NULL
))
2365 if (unlikely(check_pcp_refill(page
)))
2369 * Split buddy pages returned by expand() are received here in
2370 * physical page order. The page is added to the tail of
2371 * caller's list. From the callers perspective, the linked list
2372 * is ordered by page number under some conditions. This is
2373 * useful for IO devices that can forward direction from the
2374 * head, thus also in the physical page order. This is useful
2375 * for IO devices that can merge IO requests if the physical
2376 * pages are ordered properly.
2378 list_add_tail(&page
->lru
, list
);
2380 if (is_migrate_cma(get_pcppage_migratetype(page
)))
2381 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
,
2386 * i pages were removed from the buddy list even if some leak due
2387 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2388 * on i. Do not confuse with 'alloced' which is the number of
2389 * pages added to the pcp list.
2391 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
2392 spin_unlock(&zone
->lock
);
2398 * Called from the vmstat counter updater to drain pagesets of this
2399 * currently executing processor on remote nodes after they have
2402 * Note that this function must be called with the thread pinned to
2403 * a single processor.
2405 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
2407 unsigned long flags
;
2408 int to_drain
, batch
;
2410 local_irq_save(flags
);
2411 batch
= READ_ONCE(pcp
->batch
);
2412 to_drain
= min(pcp
->count
, batch
);
2414 free_pcppages_bulk(zone
, to_drain
, pcp
);
2415 pcp
->count
-= to_drain
;
2417 local_irq_restore(flags
);
2422 * Drain pcplists of the indicated processor and zone.
2424 * The processor must either be the current processor and the
2425 * thread pinned to the current processor or a processor that
2428 static void drain_pages_zone(unsigned int cpu
, struct zone
*zone
)
2430 unsigned long flags
;
2431 struct per_cpu_pageset
*pset
;
2432 struct per_cpu_pages
*pcp
;
2434 local_irq_save(flags
);
2435 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
2439 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
2442 local_irq_restore(flags
);
2446 * Drain pcplists of all zones on the indicated processor.
2448 * The processor must either be the current processor and the
2449 * thread pinned to the current processor or a processor that
2452 static void drain_pages(unsigned int cpu
)
2456 for_each_populated_zone(zone
) {
2457 drain_pages_zone(cpu
, zone
);
2462 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2464 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2465 * the single zone's pages.
2467 void drain_local_pages(struct zone
*zone
)
2469 int cpu
= smp_processor_id();
2472 drain_pages_zone(cpu
, zone
);
2477 static void drain_local_pages_wq(struct work_struct
*work
)
2480 * drain_all_pages doesn't use proper cpu hotplug protection so
2481 * we can race with cpu offline when the WQ can move this from
2482 * a cpu pinned worker to an unbound one. We can operate on a different
2483 * cpu which is allright but we also have to make sure to not move to
2487 drain_local_pages(NULL
);
2492 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2494 * When zone parameter is non-NULL, spill just the single zone's pages.
2496 * Note that this can be extremely slow as the draining happens in a workqueue.
2498 void drain_all_pages(struct zone
*zone
)
2503 * Allocate in the BSS so we wont require allocation in
2504 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2506 static cpumask_t cpus_with_pcps
;
2509 * Make sure nobody triggers this path before mm_percpu_wq is fully
2512 if (WARN_ON_ONCE(!mm_percpu_wq
))
2516 * Do not drain if one is already in progress unless it's specific to
2517 * a zone. Such callers are primarily CMA and memory hotplug and need
2518 * the drain to be complete when the call returns.
2520 if (unlikely(!mutex_trylock(&pcpu_drain_mutex
))) {
2523 mutex_lock(&pcpu_drain_mutex
);
2527 * We don't care about racing with CPU hotplug event
2528 * as offline notification will cause the notified
2529 * cpu to drain that CPU pcps and on_each_cpu_mask
2530 * disables preemption as part of its processing
2532 for_each_online_cpu(cpu
) {
2533 struct per_cpu_pageset
*pcp
;
2535 bool has_pcps
= false;
2538 pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
2542 for_each_populated_zone(z
) {
2543 pcp
= per_cpu_ptr(z
->pageset
, cpu
);
2544 if (pcp
->pcp
.count
) {
2552 cpumask_set_cpu(cpu
, &cpus_with_pcps
);
2554 cpumask_clear_cpu(cpu
, &cpus_with_pcps
);
2557 for_each_cpu(cpu
, &cpus_with_pcps
) {
2558 struct work_struct
*work
= per_cpu_ptr(&pcpu_drain
, cpu
);
2559 INIT_WORK(work
, drain_local_pages_wq
);
2560 queue_work_on(cpu
, mm_percpu_wq
, work
);
2562 for_each_cpu(cpu
, &cpus_with_pcps
)
2563 flush_work(per_cpu_ptr(&pcpu_drain
, cpu
));
2565 mutex_unlock(&pcpu_drain_mutex
);
2568 #ifdef CONFIG_HIBERNATION
2571 * Touch the watchdog for every WD_PAGE_COUNT pages.
2573 #define WD_PAGE_COUNT (128*1024)
2575 void mark_free_pages(struct zone
*zone
)
2577 unsigned long pfn
, max_zone_pfn
, page_count
= WD_PAGE_COUNT
;
2578 unsigned long flags
;
2579 unsigned int order
, t
;
2582 if (zone_is_empty(zone
))
2585 spin_lock_irqsave(&zone
->lock
, flags
);
2587 max_zone_pfn
= zone_end_pfn(zone
);
2588 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
2589 if (pfn_valid(pfn
)) {
2590 page
= pfn_to_page(pfn
);
2592 if (!--page_count
) {
2593 touch_nmi_watchdog();
2594 page_count
= WD_PAGE_COUNT
;
2597 if (page_zone(page
) != zone
)
2600 if (!swsusp_page_is_forbidden(page
))
2601 swsusp_unset_page_free(page
);
2604 for_each_migratetype_order(order
, t
) {
2605 list_for_each_entry(page
,
2606 &zone
->free_area
[order
].free_list
[t
], lru
) {
2609 pfn
= page_to_pfn(page
);
2610 for (i
= 0; i
< (1UL << order
); i
++) {
2611 if (!--page_count
) {
2612 touch_nmi_watchdog();
2613 page_count
= WD_PAGE_COUNT
;
2615 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
2619 spin_unlock_irqrestore(&zone
->lock
, flags
);
2621 #endif /* CONFIG_PM */
2623 static bool free_unref_page_prepare(struct page
*page
, unsigned long pfn
)
2627 if (!free_pcp_prepare(page
))
2630 migratetype
= get_pfnblock_migratetype(page
, pfn
);
2631 set_pcppage_migratetype(page
, migratetype
);
2635 static void free_unref_page_commit(struct page
*page
, unsigned long pfn
)
2637 struct zone
*zone
= page_zone(page
);
2638 struct per_cpu_pages
*pcp
;
2641 migratetype
= get_pcppage_migratetype(page
);
2642 __count_vm_event(PGFREE
);
2645 * We only track unmovable, reclaimable and movable on pcp lists.
2646 * Free ISOLATE pages back to the allocator because they are being
2647 * offlined but treat HIGHATOMIC as movable pages so we can get those
2648 * areas back if necessary. Otherwise, we may have to free
2649 * excessively into the page allocator
2651 if (migratetype
>= MIGRATE_PCPTYPES
) {
2652 if (unlikely(is_migrate_isolate(migratetype
))) {
2653 free_one_page(zone
, page
, pfn
, 0, migratetype
);
2656 migratetype
= MIGRATE_MOVABLE
;
2659 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
2660 list_add(&page
->lru
, &pcp
->lists
[migratetype
]);
2662 if (pcp
->count
>= pcp
->high
) {
2663 unsigned long batch
= READ_ONCE(pcp
->batch
);
2664 free_pcppages_bulk(zone
, batch
, pcp
);
2665 pcp
->count
-= batch
;
2670 * Free a 0-order page
2672 void free_unref_page(struct page
*page
)
2674 unsigned long flags
;
2675 unsigned long pfn
= page_to_pfn(page
);
2677 if (!free_unref_page_prepare(page
, pfn
))
2680 local_irq_save(flags
);
2681 free_unref_page_commit(page
, pfn
);
2682 local_irq_restore(flags
);
2686 * Free a list of 0-order pages
2688 void free_unref_page_list(struct list_head
*list
)
2690 struct page
*page
, *next
;
2691 unsigned long flags
, pfn
;
2692 int batch_count
= 0;
2694 /* Prepare pages for freeing */
2695 list_for_each_entry_safe(page
, next
, list
, lru
) {
2696 pfn
= page_to_pfn(page
);
2697 if (!free_unref_page_prepare(page
, pfn
))
2698 list_del(&page
->lru
);
2699 set_page_private(page
, pfn
);
2702 local_irq_save(flags
);
2703 list_for_each_entry_safe(page
, next
, list
, lru
) {
2704 unsigned long pfn
= page_private(page
);
2706 set_page_private(page
, 0);
2707 trace_mm_page_free_batched(page
);
2708 free_unref_page_commit(page
, pfn
);
2711 * Guard against excessive IRQ disabled times when we get
2712 * a large list of pages to free.
2714 if (++batch_count
== SWAP_CLUSTER_MAX
) {
2715 local_irq_restore(flags
);
2717 local_irq_save(flags
);
2720 local_irq_restore(flags
);
2724 * split_page takes a non-compound higher-order page, and splits it into
2725 * n (1<<order) sub-pages: page[0..n]
2726 * Each sub-page must be freed individually.
2728 * Note: this is probably too low level an operation for use in drivers.
2729 * Please consult with lkml before using this in your driver.
2731 void split_page(struct page
*page
, unsigned int order
)
2735 VM_BUG_ON_PAGE(PageCompound(page
), page
);
2736 VM_BUG_ON_PAGE(!page_count(page
), page
);
2738 for (i
= 1; i
< (1 << order
); i
++)
2739 set_page_refcounted(page
+ i
);
2740 split_page_owner(page
, order
);
2742 EXPORT_SYMBOL_GPL(split_page
);
2744 int __isolate_free_page(struct page
*page
, unsigned int order
)
2746 unsigned long watermark
;
2750 BUG_ON(!PageBuddy(page
));
2752 zone
= page_zone(page
);
2753 mt
= get_pageblock_migratetype(page
);
2755 if (!is_migrate_isolate(mt
)) {
2757 * Obey watermarks as if the page was being allocated. We can
2758 * emulate a high-order watermark check with a raised order-0
2759 * watermark, because we already know our high-order page
2762 watermark
= min_wmark_pages(zone
) + (1UL << order
);
2763 if (!zone_watermark_ok(zone
, 0, watermark
, 0, ALLOC_CMA
))
2766 __mod_zone_freepage_state(zone
, -(1UL << order
), mt
);
2769 /* Remove page from free list */
2770 list_del(&page
->lru
);
2771 zone
->free_area
[order
].nr_free
--;
2772 rmv_page_order(page
);
2775 * Set the pageblock if the isolated page is at least half of a
2778 if (order
>= pageblock_order
- 1) {
2779 struct page
*endpage
= page
+ (1 << order
) - 1;
2780 for (; page
< endpage
; page
+= pageblock_nr_pages
) {
2781 int mt
= get_pageblock_migratetype(page
);
2782 if (!is_migrate_isolate(mt
) && !is_migrate_cma(mt
)
2783 && !is_migrate_highatomic(mt
))
2784 set_pageblock_migratetype(page
,
2790 return 1UL << order
;
2794 * Update NUMA hit/miss statistics
2796 * Must be called with interrupts disabled.
2798 static inline void zone_statistics(struct zone
*preferred_zone
, struct zone
*z
)
2801 enum numa_stat_item local_stat
= NUMA_LOCAL
;
2803 /* skip numa counters update if numa stats is disabled */
2804 if (!static_branch_likely(&vm_numa_stat_key
))
2807 if (z
->node
!= numa_node_id())
2808 local_stat
= NUMA_OTHER
;
2810 if (z
->node
== preferred_zone
->node
)
2811 __inc_numa_state(z
, NUMA_HIT
);
2813 __inc_numa_state(z
, NUMA_MISS
);
2814 __inc_numa_state(preferred_zone
, NUMA_FOREIGN
);
2816 __inc_numa_state(z
, local_stat
);
2820 /* Remove page from the per-cpu list, caller must protect the list */
2821 static struct page
*__rmqueue_pcplist(struct zone
*zone
, int migratetype
,
2822 struct per_cpu_pages
*pcp
,
2823 struct list_head
*list
)
2828 if (list_empty(list
)) {
2829 pcp
->count
+= rmqueue_bulk(zone
, 0,
2832 if (unlikely(list_empty(list
)))
2836 page
= list_first_entry(list
, struct page
, lru
);
2837 list_del(&page
->lru
);
2839 } while (check_new_pcp(page
));
2844 /* Lock and remove page from the per-cpu list */
2845 static struct page
*rmqueue_pcplist(struct zone
*preferred_zone
,
2846 struct zone
*zone
, unsigned int order
,
2847 gfp_t gfp_flags
, int migratetype
)
2849 struct per_cpu_pages
*pcp
;
2850 struct list_head
*list
;
2852 unsigned long flags
;
2854 local_irq_save(flags
);
2855 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
2856 list
= &pcp
->lists
[migratetype
];
2857 page
= __rmqueue_pcplist(zone
, migratetype
, pcp
, list
);
2859 __count_zid_vm_events(PGALLOC
, page_zonenum(page
), 1 << order
);
2860 zone_statistics(preferred_zone
, zone
);
2862 local_irq_restore(flags
);
2867 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2870 struct page
*rmqueue(struct zone
*preferred_zone
,
2871 struct zone
*zone
, unsigned int order
,
2872 gfp_t gfp_flags
, unsigned int alloc_flags
,
2875 unsigned long flags
;
2878 if (likely(order
== 0)) {
2879 page
= rmqueue_pcplist(preferred_zone
, zone
, order
,
2880 gfp_flags
, migratetype
);
2885 * We most definitely don't want callers attempting to
2886 * allocate greater than order-1 page units with __GFP_NOFAIL.
2888 WARN_ON_ONCE((gfp_flags
& __GFP_NOFAIL
) && (order
> 1));
2889 spin_lock_irqsave(&zone
->lock
, flags
);
2893 if (alloc_flags
& ALLOC_HARDER
) {
2894 page
= __rmqueue_smallest(zone
, order
, MIGRATE_HIGHATOMIC
);
2896 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
2899 page
= __rmqueue(zone
, order
, migratetype
);
2900 } while (page
&& check_new_pages(page
, order
));
2901 spin_unlock(&zone
->lock
);
2904 __mod_zone_freepage_state(zone
, -(1 << order
),
2905 get_pcppage_migratetype(page
));
2907 __count_zid_vm_events(PGALLOC
, page_zonenum(page
), 1 << order
);
2908 zone_statistics(preferred_zone
, zone
);
2909 local_irq_restore(flags
);
2912 VM_BUG_ON_PAGE(page
&& bad_range(zone
, page
), page
);
2916 local_irq_restore(flags
);
2920 #ifdef CONFIG_FAIL_PAGE_ALLOC
2923 struct fault_attr attr
;
2925 bool ignore_gfp_highmem
;
2926 bool ignore_gfp_reclaim
;
2928 } fail_page_alloc
= {
2929 .attr
= FAULT_ATTR_INITIALIZER
,
2930 .ignore_gfp_reclaim
= true,
2931 .ignore_gfp_highmem
= true,
2935 static int __init
setup_fail_page_alloc(char *str
)
2937 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
2939 __setup("fail_page_alloc=", setup_fail_page_alloc
);
2941 static bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
2943 if (order
< fail_page_alloc
.min_order
)
2945 if (gfp_mask
& __GFP_NOFAIL
)
2947 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
2949 if (fail_page_alloc
.ignore_gfp_reclaim
&&
2950 (gfp_mask
& __GFP_DIRECT_RECLAIM
))
2953 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
2956 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2958 static int __init
fail_page_alloc_debugfs(void)
2960 umode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
2963 dir
= fault_create_debugfs_attr("fail_page_alloc", NULL
,
2964 &fail_page_alloc
.attr
);
2966 return PTR_ERR(dir
);
2968 if (!debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
2969 &fail_page_alloc
.ignore_gfp_reclaim
))
2971 if (!debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
2972 &fail_page_alloc
.ignore_gfp_highmem
))
2974 if (!debugfs_create_u32("min-order", mode
, dir
,
2975 &fail_page_alloc
.min_order
))
2980 debugfs_remove_recursive(dir
);
2985 late_initcall(fail_page_alloc_debugfs
);
2987 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2989 #else /* CONFIG_FAIL_PAGE_ALLOC */
2991 static inline bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
2996 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2999 * Return true if free base pages are above 'mark'. For high-order checks it
3000 * will return true of the order-0 watermark is reached and there is at least
3001 * one free page of a suitable size. Checking now avoids taking the zone lock
3002 * to check in the allocation paths if no pages are free.
3004 bool __zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
3005 int classzone_idx
, unsigned int alloc_flags
,
3010 const bool alloc_harder
= (alloc_flags
& (ALLOC_HARDER
|ALLOC_OOM
));
3012 /* free_pages may go negative - that's OK */
3013 free_pages
-= (1 << order
) - 1;
3015 if (alloc_flags
& ALLOC_HIGH
)
3019 * If the caller does not have rights to ALLOC_HARDER then subtract
3020 * the high-atomic reserves. This will over-estimate the size of the
3021 * atomic reserve but it avoids a search.
3023 if (likely(!alloc_harder
)) {
3024 free_pages
-= z
->nr_reserved_highatomic
;
3027 * OOM victims can try even harder than normal ALLOC_HARDER
3028 * users on the grounds that it's definitely going to be in
3029 * the exit path shortly and free memory. Any allocation it
3030 * makes during the free path will be small and short-lived.
3032 if (alloc_flags
& ALLOC_OOM
)
3040 /* If allocation can't use CMA areas don't use free CMA pages */
3041 if (!(alloc_flags
& ALLOC_CMA
))
3042 free_pages
-= zone_page_state(z
, NR_FREE_CMA_PAGES
);
3046 * Check watermarks for an order-0 allocation request. If these
3047 * are not met, then a high-order request also cannot go ahead
3048 * even if a suitable page happened to be free.
3050 if (free_pages
<= min
+ z
->lowmem_reserve
[classzone_idx
])
3053 /* If this is an order-0 request then the watermark is fine */
3057 /* For a high-order request, check at least one suitable page is free */
3058 for (o
= order
; o
< MAX_ORDER
; o
++) {
3059 struct free_area
*area
= &z
->free_area
[o
];
3065 for (mt
= 0; mt
< MIGRATE_PCPTYPES
; mt
++) {
3066 if (!list_empty(&area
->free_list
[mt
]))
3071 if ((alloc_flags
& ALLOC_CMA
) &&
3072 !list_empty(&area
->free_list
[MIGRATE_CMA
])) {
3077 !list_empty(&area
->free_list
[MIGRATE_HIGHATOMIC
]))
3083 bool zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
3084 int classzone_idx
, unsigned int alloc_flags
)
3086 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
3087 zone_page_state(z
, NR_FREE_PAGES
));
3090 static inline bool zone_watermark_fast(struct zone
*z
, unsigned int order
,
3091 unsigned long mark
, int classzone_idx
, unsigned int alloc_flags
)
3093 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
3097 /* If allocation can't use CMA areas don't use free CMA pages */
3098 if (!(alloc_flags
& ALLOC_CMA
))
3099 cma_pages
= zone_page_state(z
, NR_FREE_CMA_PAGES
);
3103 * Fast check for order-0 only. If this fails then the reserves
3104 * need to be calculated. There is a corner case where the check
3105 * passes but only the high-order atomic reserve are free. If
3106 * the caller is !atomic then it'll uselessly search the free
3107 * list. That corner case is then slower but it is harmless.
3109 if (!order
&& (free_pages
- cma_pages
) > mark
+ z
->lowmem_reserve
[classzone_idx
])
3112 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
3116 bool zone_watermark_ok_safe(struct zone
*z
, unsigned int order
,
3117 unsigned long mark
, int classzone_idx
)
3119 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
3121 if (z
->percpu_drift_mark
&& free_pages
< z
->percpu_drift_mark
)
3122 free_pages
= zone_page_state_snapshot(z
, NR_FREE_PAGES
);
3124 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, 0,
3129 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
3131 return node_distance(zone_to_nid(local_zone
), zone_to_nid(zone
)) <=
3134 #else /* CONFIG_NUMA */
3135 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
3139 #endif /* CONFIG_NUMA */
3142 * get_page_from_freelist goes through the zonelist trying to allocate
3145 static struct page
*
3146 get_page_from_freelist(gfp_t gfp_mask
, unsigned int order
, int alloc_flags
,
3147 const struct alloc_context
*ac
)
3149 struct zoneref
*z
= ac
->preferred_zoneref
;
3151 struct pglist_data
*last_pgdat_dirty_limit
= NULL
;
3154 * Scan zonelist, looking for a zone with enough free.
3155 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3157 for_next_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
3162 if (cpusets_enabled() &&
3163 (alloc_flags
& ALLOC_CPUSET
) &&
3164 !__cpuset_zone_allowed(zone
, gfp_mask
))
3167 * When allocating a page cache page for writing, we
3168 * want to get it from a node that is within its dirty
3169 * limit, such that no single node holds more than its
3170 * proportional share of globally allowed dirty pages.
3171 * The dirty limits take into account the node's
3172 * lowmem reserves and high watermark so that kswapd
3173 * should be able to balance it without having to
3174 * write pages from its LRU list.
3176 * XXX: For now, allow allocations to potentially
3177 * exceed the per-node dirty limit in the slowpath
3178 * (spread_dirty_pages unset) before going into reclaim,
3179 * which is important when on a NUMA setup the allowed
3180 * nodes are together not big enough to reach the
3181 * global limit. The proper fix for these situations
3182 * will require awareness of nodes in the
3183 * dirty-throttling and the flusher threads.
3185 if (ac
->spread_dirty_pages
) {
3186 if (last_pgdat_dirty_limit
== zone
->zone_pgdat
)
3189 if (!node_dirty_ok(zone
->zone_pgdat
)) {
3190 last_pgdat_dirty_limit
= zone
->zone_pgdat
;
3195 mark
= zone
->watermark
[alloc_flags
& ALLOC_WMARK_MASK
];
3196 if (!zone_watermark_fast(zone
, order
, mark
,
3197 ac_classzone_idx(ac
), alloc_flags
)) {
3200 /* Checked here to keep the fast path fast */
3201 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
3202 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
3205 if (node_reclaim_mode
== 0 ||
3206 !zone_allows_reclaim(ac
->preferred_zoneref
->zone
, zone
))
3209 ret
= node_reclaim(zone
->zone_pgdat
, gfp_mask
, order
);
3211 case NODE_RECLAIM_NOSCAN
:
3214 case NODE_RECLAIM_FULL
:
3215 /* scanned but unreclaimable */
3218 /* did we reclaim enough */
3219 if (zone_watermark_ok(zone
, order
, mark
,
3220 ac_classzone_idx(ac
), alloc_flags
))
3228 page
= rmqueue(ac
->preferred_zoneref
->zone
, zone
, order
,
3229 gfp_mask
, alloc_flags
, ac
->migratetype
);
3231 prep_new_page(page
, order
, gfp_mask
, alloc_flags
);
3234 * If this is a high-order atomic allocation then check
3235 * if the pageblock should be reserved for the future
3237 if (unlikely(order
&& (alloc_flags
& ALLOC_HARDER
)))
3238 reserve_highatomic_pageblock(page
, zone
, order
);
3248 * Large machines with many possible nodes should not always dump per-node
3249 * meminfo in irq context.
3251 static inline bool should_suppress_show_mem(void)
3256 ret
= in_interrupt();
3261 static void warn_alloc_show_mem(gfp_t gfp_mask
, nodemask_t
*nodemask
)
3263 unsigned int filter
= SHOW_MEM_FILTER_NODES
;
3264 static DEFINE_RATELIMIT_STATE(show_mem_rs
, HZ
, 1);
3266 if (should_suppress_show_mem() || !__ratelimit(&show_mem_rs
))
3270 * This documents exceptions given to allocations in certain
3271 * contexts that are allowed to allocate outside current's set
3274 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
3275 if (tsk_is_oom_victim(current
) ||
3276 (current
->flags
& (PF_MEMALLOC
| PF_EXITING
)))
3277 filter
&= ~SHOW_MEM_FILTER_NODES
;
3278 if (in_interrupt() || !(gfp_mask
& __GFP_DIRECT_RECLAIM
))
3279 filter
&= ~SHOW_MEM_FILTER_NODES
;
3281 show_mem(filter
, nodemask
);
3284 void warn_alloc(gfp_t gfp_mask
, nodemask_t
*nodemask
, const char *fmt
, ...)
3286 struct va_format vaf
;
3288 static DEFINE_RATELIMIT_STATE(nopage_rs
, DEFAULT_RATELIMIT_INTERVAL
,
3289 DEFAULT_RATELIMIT_BURST
);
3291 if ((gfp_mask
& __GFP_NOWARN
) || !__ratelimit(&nopage_rs
))
3294 va_start(args
, fmt
);
3297 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl\n",
3298 current
->comm
, &vaf
, gfp_mask
, &gfp_mask
,
3299 nodemask_pr_args(nodemask
));
3302 cpuset_print_current_mems_allowed();
3305 warn_alloc_show_mem(gfp_mask
, nodemask
);
3308 static inline struct page
*
3309 __alloc_pages_cpuset_fallback(gfp_t gfp_mask
, unsigned int order
,
3310 unsigned int alloc_flags
,
3311 const struct alloc_context
*ac
)
3315 page
= get_page_from_freelist(gfp_mask
, order
,
3316 alloc_flags
|ALLOC_CPUSET
, ac
);
3318 * fallback to ignore cpuset restriction if our nodes
3322 page
= get_page_from_freelist(gfp_mask
, order
,
3328 static inline struct page
*
3329 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
3330 const struct alloc_context
*ac
, unsigned long *did_some_progress
)
3332 struct oom_control oc
= {
3333 .zonelist
= ac
->zonelist
,
3334 .nodemask
= ac
->nodemask
,
3336 .gfp_mask
= gfp_mask
,
3341 *did_some_progress
= 0;
3344 * Acquire the oom lock. If that fails, somebody else is
3345 * making progress for us.
3347 if (!mutex_trylock(&oom_lock
)) {
3348 *did_some_progress
= 1;
3349 schedule_timeout_uninterruptible(1);
3354 * Go through the zonelist yet one more time, keep very high watermark
3355 * here, this is only to catch a parallel oom killing, we must fail if
3356 * we're still under heavy pressure. But make sure that this reclaim
3357 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3358 * allocation which will never fail due to oom_lock already held.
3360 page
= get_page_from_freelist((gfp_mask
| __GFP_HARDWALL
) &
3361 ~__GFP_DIRECT_RECLAIM
, order
,
3362 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
, ac
);
3366 /* Coredumps can quickly deplete all memory reserves */
3367 if (current
->flags
& PF_DUMPCORE
)
3369 /* The OOM killer will not help higher order allocs */
3370 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
3373 * We have already exhausted all our reclaim opportunities without any
3374 * success so it is time to admit defeat. We will skip the OOM killer
3375 * because it is very likely that the caller has a more reasonable
3376 * fallback than shooting a random task.
3378 if (gfp_mask
& __GFP_RETRY_MAYFAIL
)
3380 /* The OOM killer does not needlessly kill tasks for lowmem */
3381 if (ac
->high_zoneidx
< ZONE_NORMAL
)
3383 if (pm_suspended_storage())
3386 * XXX: GFP_NOFS allocations should rather fail than rely on
3387 * other request to make a forward progress.
3388 * We are in an unfortunate situation where out_of_memory cannot
3389 * do much for this context but let's try it to at least get
3390 * access to memory reserved if the current task is killed (see
3391 * out_of_memory). Once filesystems are ready to handle allocation
3392 * failures more gracefully we should just bail out here.
3395 /* The OOM killer may not free memory on a specific node */
3396 if (gfp_mask
& __GFP_THISNODE
)
3399 /* Exhausted what can be done so it's blamo time */
3400 if (out_of_memory(&oc
) || WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
)) {
3401 *did_some_progress
= 1;
3404 * Help non-failing allocations by giving them access to memory
3407 if (gfp_mask
& __GFP_NOFAIL
)
3408 page
= __alloc_pages_cpuset_fallback(gfp_mask
, order
,
3409 ALLOC_NO_WATERMARKS
, ac
);
3412 mutex_unlock(&oom_lock
);
3417 * Maximum number of compaction retries wit a progress before OOM
3418 * killer is consider as the only way to move forward.
3420 #define MAX_COMPACT_RETRIES 16
3422 #ifdef CONFIG_COMPACTION
3423 /* Try memory compaction for high-order allocations before reclaim */
3424 static struct page
*
3425 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
3426 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3427 enum compact_priority prio
, enum compact_result
*compact_result
)
3430 unsigned int noreclaim_flag
;
3435 noreclaim_flag
= memalloc_noreclaim_save();
3436 *compact_result
= try_to_compact_pages(gfp_mask
, order
, alloc_flags
, ac
,
3438 memalloc_noreclaim_restore(noreclaim_flag
);
3440 if (*compact_result
<= COMPACT_INACTIVE
)
3444 * At least in one zone compaction wasn't deferred or skipped, so let's
3445 * count a compaction stall
3447 count_vm_event(COMPACTSTALL
);
3449 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
3452 struct zone
*zone
= page_zone(page
);
3454 zone
->compact_blockskip_flush
= false;
3455 compaction_defer_reset(zone
, order
, true);
3456 count_vm_event(COMPACTSUCCESS
);
3461 * It's bad if compaction run occurs and fails. The most likely reason
3462 * is that pages exist, but not enough to satisfy watermarks.
3464 count_vm_event(COMPACTFAIL
);
3472 should_compact_retry(struct alloc_context
*ac
, int order
, int alloc_flags
,
3473 enum compact_result compact_result
,
3474 enum compact_priority
*compact_priority
,
3475 int *compaction_retries
)
3477 int max_retries
= MAX_COMPACT_RETRIES
;
3480 int retries
= *compaction_retries
;
3481 enum compact_priority priority
= *compact_priority
;
3486 if (compaction_made_progress(compact_result
))
3487 (*compaction_retries
)++;
3490 * compaction considers all the zone as desperately out of memory
3491 * so it doesn't really make much sense to retry except when the
3492 * failure could be caused by insufficient priority
3494 if (compaction_failed(compact_result
))
3495 goto check_priority
;
3498 * make sure the compaction wasn't deferred or didn't bail out early
3499 * due to locks contention before we declare that we should give up.
3500 * But do not retry if the given zonelist is not suitable for
3503 if (compaction_withdrawn(compact_result
)) {
3504 ret
= compaction_zonelist_suitable(ac
, order
, alloc_flags
);
3509 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
3510 * costly ones because they are de facto nofail and invoke OOM
3511 * killer to move on while costly can fail and users are ready
3512 * to cope with that. 1/4 retries is rather arbitrary but we
3513 * would need much more detailed feedback from compaction to
3514 * make a better decision.
3516 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
3518 if (*compaction_retries
<= max_retries
) {
3524 * Make sure there are attempts at the highest priority if we exhausted
3525 * all retries or failed at the lower priorities.
3528 min_priority
= (order
> PAGE_ALLOC_COSTLY_ORDER
) ?
3529 MIN_COMPACT_COSTLY_PRIORITY
: MIN_COMPACT_PRIORITY
;
3531 if (*compact_priority
> min_priority
) {
3532 (*compact_priority
)--;
3533 *compaction_retries
= 0;
3537 trace_compact_retry(order
, priority
, compact_result
, retries
, max_retries
, ret
);
3541 static inline struct page
*
3542 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
3543 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3544 enum compact_priority prio
, enum compact_result
*compact_result
)
3546 *compact_result
= COMPACT_SKIPPED
;
3551 should_compact_retry(struct alloc_context
*ac
, unsigned int order
, int alloc_flags
,
3552 enum compact_result compact_result
,
3553 enum compact_priority
*compact_priority
,
3554 int *compaction_retries
)
3559 if (!order
|| order
> PAGE_ALLOC_COSTLY_ORDER
)
3563 * There are setups with compaction disabled which would prefer to loop
3564 * inside the allocator rather than hit the oom killer prematurely.
3565 * Let's give them a good hope and keep retrying while the order-0
3566 * watermarks are OK.
3568 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
3570 if (zone_watermark_ok(zone
, 0, min_wmark_pages(zone
),
3571 ac_classzone_idx(ac
), alloc_flags
))
3576 #endif /* CONFIG_COMPACTION */
3578 #ifdef CONFIG_LOCKDEP
3579 struct lockdep_map __fs_reclaim_map
=
3580 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map
);
3582 static bool __need_fs_reclaim(gfp_t gfp_mask
)
3584 gfp_mask
= current_gfp_context(gfp_mask
);
3586 /* no reclaim without waiting on it */
3587 if (!(gfp_mask
& __GFP_DIRECT_RECLAIM
))
3590 /* this guy won't enter reclaim */
3591 if (current
->flags
& PF_MEMALLOC
)
3594 /* We're only interested __GFP_FS allocations for now */
3595 if (!(gfp_mask
& __GFP_FS
))
3598 if (gfp_mask
& __GFP_NOLOCKDEP
)
3604 void fs_reclaim_acquire(gfp_t gfp_mask
)
3606 if (__need_fs_reclaim(gfp_mask
))
3607 lock_map_acquire(&__fs_reclaim_map
);
3609 EXPORT_SYMBOL_GPL(fs_reclaim_acquire
);
3611 void fs_reclaim_release(gfp_t gfp_mask
)
3613 if (__need_fs_reclaim(gfp_mask
))
3614 lock_map_release(&__fs_reclaim_map
);
3616 EXPORT_SYMBOL_GPL(fs_reclaim_release
);
3619 /* Perform direct synchronous page reclaim */
3621 __perform_reclaim(gfp_t gfp_mask
, unsigned int order
,
3622 const struct alloc_context
*ac
)
3624 struct reclaim_state reclaim_state
;
3626 unsigned int noreclaim_flag
;
3630 /* We now go into synchronous reclaim */
3631 cpuset_memory_pressure_bump();
3632 noreclaim_flag
= memalloc_noreclaim_save();
3633 fs_reclaim_acquire(gfp_mask
);
3634 reclaim_state
.reclaimed_slab
= 0;
3635 current
->reclaim_state
= &reclaim_state
;
3637 progress
= try_to_free_pages(ac
->zonelist
, order
, gfp_mask
,
3640 current
->reclaim_state
= NULL
;
3641 fs_reclaim_release(gfp_mask
);
3642 memalloc_noreclaim_restore(noreclaim_flag
);
3649 /* The really slow allocator path where we enter direct reclaim */
3650 static inline struct page
*
3651 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
3652 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3653 unsigned long *did_some_progress
)
3655 struct page
*page
= NULL
;
3656 bool drained
= false;
3658 *did_some_progress
= __perform_reclaim(gfp_mask
, order
, ac
);
3659 if (unlikely(!(*did_some_progress
)))
3663 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
3666 * If an allocation failed after direct reclaim, it could be because
3667 * pages are pinned on the per-cpu lists or in high alloc reserves.
3668 * Shrink them them and try again
3670 if (!page
&& !drained
) {
3671 unreserve_highatomic_pageblock(ac
, false);
3672 drain_all_pages(NULL
);
3680 static void wake_all_kswapds(unsigned int order
, const struct alloc_context
*ac
)
3684 pg_data_t
*last_pgdat
= NULL
;
3686 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
,
3687 ac
->high_zoneidx
, ac
->nodemask
) {
3688 if (last_pgdat
!= zone
->zone_pgdat
)
3689 wakeup_kswapd(zone
, order
, ac
->high_zoneidx
);
3690 last_pgdat
= zone
->zone_pgdat
;
3694 static inline unsigned int
3695 gfp_to_alloc_flags(gfp_t gfp_mask
)
3697 unsigned int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
3699 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3700 BUILD_BUG_ON(__GFP_HIGH
!= (__force gfp_t
) ALLOC_HIGH
);
3703 * The caller may dip into page reserves a bit more if the caller
3704 * cannot run direct reclaim, or if the caller has realtime scheduling
3705 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3706 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3708 alloc_flags
|= (__force
int) (gfp_mask
& __GFP_HIGH
);
3710 if (gfp_mask
& __GFP_ATOMIC
) {
3712 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3713 * if it can't schedule.
3715 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
3716 alloc_flags
|= ALLOC_HARDER
;
3718 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3719 * comment for __cpuset_node_allowed().
3721 alloc_flags
&= ~ALLOC_CPUSET
;
3722 } else if (unlikely(rt_task(current
)) && !in_interrupt())
3723 alloc_flags
|= ALLOC_HARDER
;
3726 if (gfpflags_to_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
3727 alloc_flags
|= ALLOC_CMA
;
3732 static bool oom_reserves_allowed(struct task_struct
*tsk
)
3734 if (!tsk_is_oom_victim(tsk
))
3738 * !MMU doesn't have oom reaper so give access to memory reserves
3739 * only to the thread with TIF_MEMDIE set
3741 if (!IS_ENABLED(CONFIG_MMU
) && !test_thread_flag(TIF_MEMDIE
))
3748 * Distinguish requests which really need access to full memory
3749 * reserves from oom victims which can live with a portion of it
3751 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask
)
3753 if (unlikely(gfp_mask
& __GFP_NOMEMALLOC
))
3755 if (gfp_mask
& __GFP_MEMALLOC
)
3756 return ALLOC_NO_WATERMARKS
;
3757 if (in_serving_softirq() && (current
->flags
& PF_MEMALLOC
))
3758 return ALLOC_NO_WATERMARKS
;
3759 if (!in_interrupt()) {
3760 if (current
->flags
& PF_MEMALLOC
)
3761 return ALLOC_NO_WATERMARKS
;
3762 else if (oom_reserves_allowed(current
))
3769 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask
)
3771 return !!__gfp_pfmemalloc_flags(gfp_mask
);
3775 * Checks whether it makes sense to retry the reclaim to make a forward progress
3776 * for the given allocation request.
3778 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
3779 * without success, or when we couldn't even meet the watermark if we
3780 * reclaimed all remaining pages on the LRU lists.
3782 * Returns true if a retry is viable or false to enter the oom path.
3785 should_reclaim_retry(gfp_t gfp_mask
, unsigned order
,
3786 struct alloc_context
*ac
, int alloc_flags
,
3787 bool did_some_progress
, int *no_progress_loops
)
3793 * Costly allocations might have made a progress but this doesn't mean
3794 * their order will become available due to high fragmentation so
3795 * always increment the no progress counter for them
3797 if (did_some_progress
&& order
<= PAGE_ALLOC_COSTLY_ORDER
)
3798 *no_progress_loops
= 0;
3800 (*no_progress_loops
)++;
3803 * Make sure we converge to OOM if we cannot make any progress
3804 * several times in the row.
3806 if (*no_progress_loops
> MAX_RECLAIM_RETRIES
) {
3807 /* Before OOM, exhaust highatomic_reserve */
3808 return unreserve_highatomic_pageblock(ac
, true);
3812 * Keep reclaiming pages while there is a chance this will lead
3813 * somewhere. If none of the target zones can satisfy our allocation
3814 * request even if all reclaimable pages are considered then we are
3815 * screwed and have to go OOM.
3817 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
3819 unsigned long available
;
3820 unsigned long reclaimable
;
3821 unsigned long min_wmark
= min_wmark_pages(zone
);
3824 available
= reclaimable
= zone_reclaimable_pages(zone
);
3825 available
+= zone_page_state_snapshot(zone
, NR_FREE_PAGES
);
3828 * Would the allocation succeed if we reclaimed all
3829 * reclaimable pages?
3831 wmark
= __zone_watermark_ok(zone
, order
, min_wmark
,
3832 ac_classzone_idx(ac
), alloc_flags
, available
);
3833 trace_reclaim_retry_zone(z
, order
, reclaimable
,
3834 available
, min_wmark
, *no_progress_loops
, wmark
);
3837 * If we didn't make any progress and have a lot of
3838 * dirty + writeback pages then we should wait for
3839 * an IO to complete to slow down the reclaim and
3840 * prevent from pre mature OOM
3842 if (!did_some_progress
) {
3843 unsigned long write_pending
;
3845 write_pending
= zone_page_state_snapshot(zone
,
3846 NR_ZONE_WRITE_PENDING
);
3848 if (2 * write_pending
> reclaimable
) {
3849 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
3855 * Memory allocation/reclaim might be called from a WQ
3856 * context and the current implementation of the WQ
3857 * concurrency control doesn't recognize that
3858 * a particular WQ is congested if the worker thread is
3859 * looping without ever sleeping. Therefore we have to
3860 * do a short sleep here rather than calling
3863 if (current
->flags
& PF_WQ_WORKER
)
3864 schedule_timeout_uninterruptible(1);
3876 check_retry_cpuset(int cpuset_mems_cookie
, struct alloc_context
*ac
)
3879 * It's possible that cpuset's mems_allowed and the nodemask from
3880 * mempolicy don't intersect. This should be normally dealt with by
3881 * policy_nodemask(), but it's possible to race with cpuset update in
3882 * such a way the check therein was true, and then it became false
3883 * before we got our cpuset_mems_cookie here.
3884 * This assumes that for all allocations, ac->nodemask can come only
3885 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
3886 * when it does not intersect with the cpuset restrictions) or the
3887 * caller can deal with a violated nodemask.
3889 if (cpusets_enabled() && ac
->nodemask
&&
3890 !cpuset_nodemask_valid_mems_allowed(ac
->nodemask
)) {
3891 ac
->nodemask
= NULL
;
3896 * When updating a task's mems_allowed or mempolicy nodemask, it is
3897 * possible to race with parallel threads in such a way that our
3898 * allocation can fail while the mask is being updated. If we are about
3899 * to fail, check if the cpuset changed during allocation and if so,
3902 if (read_mems_allowed_retry(cpuset_mems_cookie
))
3908 static inline struct page
*
3909 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
3910 struct alloc_context
*ac
)
3912 bool can_direct_reclaim
= gfp_mask
& __GFP_DIRECT_RECLAIM
;
3913 const bool costly_order
= order
> PAGE_ALLOC_COSTLY_ORDER
;
3914 struct page
*page
= NULL
;
3915 unsigned int alloc_flags
;
3916 unsigned long did_some_progress
;
3917 enum compact_priority compact_priority
;
3918 enum compact_result compact_result
;
3919 int compaction_retries
;
3920 int no_progress_loops
;
3921 unsigned int cpuset_mems_cookie
;
3925 * In the slowpath, we sanity check order to avoid ever trying to
3926 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3927 * be using allocators in order of preference for an area that is
3930 if (order
>= MAX_ORDER
) {
3931 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
3936 * We also sanity check to catch abuse of atomic reserves being used by
3937 * callers that are not in atomic context.
3939 if (WARN_ON_ONCE((gfp_mask
& (__GFP_ATOMIC
|__GFP_DIRECT_RECLAIM
)) ==
3940 (__GFP_ATOMIC
|__GFP_DIRECT_RECLAIM
)))
3941 gfp_mask
&= ~__GFP_ATOMIC
;
3944 compaction_retries
= 0;
3945 no_progress_loops
= 0;
3946 compact_priority
= DEF_COMPACT_PRIORITY
;
3947 cpuset_mems_cookie
= read_mems_allowed_begin();
3950 * The fast path uses conservative alloc_flags to succeed only until
3951 * kswapd needs to be woken up, and to avoid the cost of setting up
3952 * alloc_flags precisely. So we do that now.
3954 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
3957 * We need to recalculate the starting point for the zonelist iterator
3958 * because we might have used different nodemask in the fast path, or
3959 * there was a cpuset modification and we are retrying - otherwise we
3960 * could end up iterating over non-eligible zones endlessly.
3962 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
3963 ac
->high_zoneidx
, ac
->nodemask
);
3964 if (!ac
->preferred_zoneref
->zone
)
3967 if (gfp_mask
& __GFP_KSWAPD_RECLAIM
)
3968 wake_all_kswapds(order
, ac
);
3971 * The adjusted alloc_flags might result in immediate success, so try
3974 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
3979 * For costly allocations, try direct compaction first, as it's likely
3980 * that we have enough base pages and don't need to reclaim. For non-
3981 * movable high-order allocations, do that as well, as compaction will
3982 * try prevent permanent fragmentation by migrating from blocks of the
3984 * Don't try this for allocations that are allowed to ignore
3985 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
3987 if (can_direct_reclaim
&&
3989 (order
> 0 && ac
->migratetype
!= MIGRATE_MOVABLE
))
3990 && !gfp_pfmemalloc_allowed(gfp_mask
)) {
3991 page
= __alloc_pages_direct_compact(gfp_mask
, order
,
3993 INIT_COMPACT_PRIORITY
,
3999 * Checks for costly allocations with __GFP_NORETRY, which
4000 * includes THP page fault allocations
4002 if (costly_order
&& (gfp_mask
& __GFP_NORETRY
)) {
4004 * If compaction is deferred for high-order allocations,
4005 * it is because sync compaction recently failed. If
4006 * this is the case and the caller requested a THP
4007 * allocation, we do not want to heavily disrupt the
4008 * system, so we fail the allocation instead of entering
4011 if (compact_result
== COMPACT_DEFERRED
)
4015 * Looks like reclaim/compaction is worth trying, but
4016 * sync compaction could be very expensive, so keep
4017 * using async compaction.
4019 compact_priority
= INIT_COMPACT_PRIORITY
;
4024 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4025 if (gfp_mask
& __GFP_KSWAPD_RECLAIM
)
4026 wake_all_kswapds(order
, ac
);
4028 reserve_flags
= __gfp_pfmemalloc_flags(gfp_mask
);
4030 alloc_flags
= reserve_flags
;
4033 * Reset the zonelist iterators if memory policies can be ignored.
4034 * These allocations are high priority and system rather than user
4037 if (!(alloc_flags
& ALLOC_CPUSET
) || reserve_flags
) {
4038 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
4039 ac
->high_zoneidx
, ac
->nodemask
);
4042 /* Attempt with potentially adjusted zonelist and alloc_flags */
4043 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
4047 /* Caller is not willing to reclaim, we can't balance anything */
4048 if (!can_direct_reclaim
)
4051 /* Avoid recursion of direct reclaim */
4052 if (current
->flags
& PF_MEMALLOC
)
4055 /* Try direct reclaim and then allocating */
4056 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
, alloc_flags
, ac
,
4057 &did_some_progress
);
4061 /* Try direct compaction and then allocating */
4062 page
= __alloc_pages_direct_compact(gfp_mask
, order
, alloc_flags
, ac
,
4063 compact_priority
, &compact_result
);
4067 /* Do not loop if specifically requested */
4068 if (gfp_mask
& __GFP_NORETRY
)
4072 * Do not retry costly high order allocations unless they are
4073 * __GFP_RETRY_MAYFAIL
4075 if (costly_order
&& !(gfp_mask
& __GFP_RETRY_MAYFAIL
))
4078 if (should_reclaim_retry(gfp_mask
, order
, ac
, alloc_flags
,
4079 did_some_progress
> 0, &no_progress_loops
))
4083 * It doesn't make any sense to retry for the compaction if the order-0
4084 * reclaim is not able to make any progress because the current
4085 * implementation of the compaction depends on the sufficient amount
4086 * of free memory (see __compaction_suitable)
4088 if (did_some_progress
> 0 &&
4089 should_compact_retry(ac
, order
, alloc_flags
,
4090 compact_result
, &compact_priority
,
4091 &compaction_retries
))
4095 /* Deal with possible cpuset update races before we start OOM killing */
4096 if (check_retry_cpuset(cpuset_mems_cookie
, ac
))
4099 /* Reclaim has failed us, start killing things */
4100 page
= __alloc_pages_may_oom(gfp_mask
, order
, ac
, &did_some_progress
);
4104 /* Avoid allocations with no watermarks from looping endlessly */
4105 if (tsk_is_oom_victim(current
) &&
4106 (alloc_flags
== ALLOC_OOM
||
4107 (gfp_mask
& __GFP_NOMEMALLOC
)))
4110 /* Retry as long as the OOM killer is making progress */
4111 if (did_some_progress
) {
4112 no_progress_loops
= 0;
4117 /* Deal with possible cpuset update races before we fail */
4118 if (check_retry_cpuset(cpuset_mems_cookie
, ac
))
4122 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4125 if (gfp_mask
& __GFP_NOFAIL
) {
4127 * All existing users of the __GFP_NOFAIL are blockable, so warn
4128 * of any new users that actually require GFP_NOWAIT
4130 if (WARN_ON_ONCE(!can_direct_reclaim
))
4134 * PF_MEMALLOC request from this context is rather bizarre
4135 * because we cannot reclaim anything and only can loop waiting
4136 * for somebody to do a work for us
4138 WARN_ON_ONCE(current
->flags
& PF_MEMALLOC
);
4141 * non failing costly orders are a hard requirement which we
4142 * are not prepared for much so let's warn about these users
4143 * so that we can identify them and convert them to something
4146 WARN_ON_ONCE(order
> PAGE_ALLOC_COSTLY_ORDER
);
4149 * Help non-failing allocations by giving them access to memory
4150 * reserves but do not use ALLOC_NO_WATERMARKS because this
4151 * could deplete whole memory reserves which would just make
4152 * the situation worse
4154 page
= __alloc_pages_cpuset_fallback(gfp_mask
, order
, ALLOC_HARDER
, ac
);
4162 warn_alloc(gfp_mask
, ac
->nodemask
,
4163 "page allocation failure: order:%u", order
);
4168 static inline bool prepare_alloc_pages(gfp_t gfp_mask
, unsigned int order
,
4169 int preferred_nid
, nodemask_t
*nodemask
,
4170 struct alloc_context
*ac
, gfp_t
*alloc_mask
,
4171 unsigned int *alloc_flags
)
4173 ac
->high_zoneidx
= gfp_zone(gfp_mask
);
4174 ac
->zonelist
= node_zonelist(preferred_nid
, gfp_mask
);
4175 ac
->nodemask
= nodemask
;
4176 ac
->migratetype
= gfpflags_to_migratetype(gfp_mask
);
4178 if (cpusets_enabled()) {
4179 *alloc_mask
|= __GFP_HARDWALL
;
4181 ac
->nodemask
= &cpuset_current_mems_allowed
;
4183 *alloc_flags
|= ALLOC_CPUSET
;
4186 fs_reclaim_acquire(gfp_mask
);
4187 fs_reclaim_release(gfp_mask
);
4189 might_sleep_if(gfp_mask
& __GFP_DIRECT_RECLAIM
);
4191 if (should_fail_alloc_page(gfp_mask
, order
))
4194 if (IS_ENABLED(CONFIG_CMA
) && ac
->migratetype
== MIGRATE_MOVABLE
)
4195 *alloc_flags
|= ALLOC_CMA
;
4200 /* Determine whether to spread dirty pages and what the first usable zone */
4201 static inline void finalise_ac(gfp_t gfp_mask
,
4202 unsigned int order
, struct alloc_context
*ac
)
4204 /* Dirty zone balancing only done in the fast path */
4205 ac
->spread_dirty_pages
= (gfp_mask
& __GFP_WRITE
);
4208 * The preferred zone is used for statistics but crucially it is
4209 * also used as the starting point for the zonelist iterator. It
4210 * may get reset for allocations that ignore memory policies.
4212 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
4213 ac
->high_zoneidx
, ac
->nodemask
);
4217 * This is the 'heart' of the zoned buddy allocator.
4220 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
, int preferred_nid
,
4221 nodemask_t
*nodemask
)
4224 unsigned int alloc_flags
= ALLOC_WMARK_LOW
;
4225 gfp_t alloc_mask
; /* The gfp_t that was actually used for allocation */
4226 struct alloc_context ac
= { };
4228 gfp_mask
&= gfp_allowed_mask
;
4229 alloc_mask
= gfp_mask
;
4230 if (!prepare_alloc_pages(gfp_mask
, order
, preferred_nid
, nodemask
, &ac
, &alloc_mask
, &alloc_flags
))
4233 finalise_ac(gfp_mask
, order
, &ac
);
4235 /* First allocation attempt */
4236 page
= get_page_from_freelist(alloc_mask
, order
, alloc_flags
, &ac
);
4241 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4242 * resp. GFP_NOIO which has to be inherited for all allocation requests
4243 * from a particular context which has been marked by
4244 * memalloc_no{fs,io}_{save,restore}.
4246 alloc_mask
= current_gfp_context(gfp_mask
);
4247 ac
.spread_dirty_pages
= false;
4250 * Restore the original nodemask if it was potentially replaced with
4251 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4253 if (unlikely(ac
.nodemask
!= nodemask
))
4254 ac
.nodemask
= nodemask
;
4256 page
= __alloc_pages_slowpath(alloc_mask
, order
, &ac
);
4259 if (memcg_kmem_enabled() && (gfp_mask
& __GFP_ACCOUNT
) && page
&&
4260 unlikely(memcg_kmem_charge(page
, gfp_mask
, order
) != 0)) {
4261 __free_pages(page
, order
);
4265 trace_mm_page_alloc(page
, order
, alloc_mask
, ac
.migratetype
);
4269 EXPORT_SYMBOL(__alloc_pages_nodemask
);
4272 * Common helper functions.
4274 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
4279 * __get_free_pages() returns a 32-bit address, which cannot represent
4282 VM_BUG_ON((gfp_mask
& __GFP_HIGHMEM
) != 0);
4284 page
= alloc_pages(gfp_mask
, order
);
4287 return (unsigned long) page_address(page
);
4289 EXPORT_SYMBOL(__get_free_pages
);
4291 unsigned long get_zeroed_page(gfp_t gfp_mask
)
4293 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
4295 EXPORT_SYMBOL(get_zeroed_page
);
4297 void __free_pages(struct page
*page
, unsigned int order
)
4299 if (put_page_testzero(page
)) {
4301 free_unref_page(page
);
4303 __free_pages_ok(page
, order
);
4307 EXPORT_SYMBOL(__free_pages
);
4309 void free_pages(unsigned long addr
, unsigned int order
)
4312 VM_BUG_ON(!virt_addr_valid((void *)addr
));
4313 __free_pages(virt_to_page((void *)addr
), order
);
4317 EXPORT_SYMBOL(free_pages
);
4321 * An arbitrary-length arbitrary-offset area of memory which resides
4322 * within a 0 or higher order page. Multiple fragments within that page
4323 * are individually refcounted, in the page's reference counter.
4325 * The page_frag functions below provide a simple allocation framework for
4326 * page fragments. This is used by the network stack and network device
4327 * drivers to provide a backing region of memory for use as either an
4328 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4330 static struct page
*__page_frag_cache_refill(struct page_frag_cache
*nc
,
4333 struct page
*page
= NULL
;
4334 gfp_t gfp
= gfp_mask
;
4336 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4337 gfp_mask
|= __GFP_COMP
| __GFP_NOWARN
| __GFP_NORETRY
|
4339 page
= alloc_pages_node(NUMA_NO_NODE
, gfp_mask
,
4340 PAGE_FRAG_CACHE_MAX_ORDER
);
4341 nc
->size
= page
? PAGE_FRAG_CACHE_MAX_SIZE
: PAGE_SIZE
;
4343 if (unlikely(!page
))
4344 page
= alloc_pages_node(NUMA_NO_NODE
, gfp
, 0);
4346 nc
->va
= page
? page_address(page
) : NULL
;
4351 void __page_frag_cache_drain(struct page
*page
, unsigned int count
)
4353 VM_BUG_ON_PAGE(page_ref_count(page
) == 0, page
);
4355 if (page_ref_sub_and_test(page
, count
)) {
4356 unsigned int order
= compound_order(page
);
4359 free_unref_page(page
);
4361 __free_pages_ok(page
, order
);
4364 EXPORT_SYMBOL(__page_frag_cache_drain
);
4366 void *page_frag_alloc(struct page_frag_cache
*nc
,
4367 unsigned int fragsz
, gfp_t gfp_mask
)
4369 unsigned int size
= PAGE_SIZE
;
4373 if (unlikely(!nc
->va
)) {
4375 page
= __page_frag_cache_refill(nc
, gfp_mask
);
4379 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4380 /* if size can vary use size else just use PAGE_SIZE */
4383 /* Even if we own the page, we do not use atomic_set().
4384 * This would break get_page_unless_zero() users.
4386 page_ref_add(page
, size
- 1);
4388 /* reset page count bias and offset to start of new frag */
4389 nc
->pfmemalloc
= page_is_pfmemalloc(page
);
4390 nc
->pagecnt_bias
= size
;
4394 offset
= nc
->offset
- fragsz
;
4395 if (unlikely(offset
< 0)) {
4396 page
= virt_to_page(nc
->va
);
4398 if (!page_ref_sub_and_test(page
, nc
->pagecnt_bias
))
4401 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4402 /* if size can vary use size else just use PAGE_SIZE */
4405 /* OK, page count is 0, we can safely set it */
4406 set_page_count(page
, size
);
4408 /* reset page count bias and offset to start of new frag */
4409 nc
->pagecnt_bias
= size
;
4410 offset
= size
- fragsz
;
4414 nc
->offset
= offset
;
4416 return nc
->va
+ offset
;
4418 EXPORT_SYMBOL(page_frag_alloc
);
4421 * Frees a page fragment allocated out of either a compound or order 0 page.
4423 void page_frag_free(void *addr
)
4425 struct page
*page
= virt_to_head_page(addr
);
4427 if (unlikely(put_page_testzero(page
)))
4428 __free_pages_ok(page
, compound_order(page
));
4430 EXPORT_SYMBOL(page_frag_free
);
4432 static void *make_alloc_exact(unsigned long addr
, unsigned int order
,
4436 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
4437 unsigned long used
= addr
+ PAGE_ALIGN(size
);
4439 split_page(virt_to_page((void *)addr
), order
);
4440 while (used
< alloc_end
) {
4445 return (void *)addr
;
4449 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4450 * @size: the number of bytes to allocate
4451 * @gfp_mask: GFP flags for the allocation
4453 * This function is similar to alloc_pages(), except that it allocates the
4454 * minimum number of pages to satisfy the request. alloc_pages() can only
4455 * allocate memory in power-of-two pages.
4457 * This function is also limited by MAX_ORDER.
4459 * Memory allocated by this function must be released by free_pages_exact().
4461 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
4463 unsigned int order
= get_order(size
);
4466 addr
= __get_free_pages(gfp_mask
, order
);
4467 return make_alloc_exact(addr
, order
, size
);
4469 EXPORT_SYMBOL(alloc_pages_exact
);
4472 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4474 * @nid: the preferred node ID where memory should be allocated
4475 * @size: the number of bytes to allocate
4476 * @gfp_mask: GFP flags for the allocation
4478 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4481 void * __meminit
alloc_pages_exact_nid(int nid
, size_t size
, gfp_t gfp_mask
)
4483 unsigned int order
= get_order(size
);
4484 struct page
*p
= alloc_pages_node(nid
, gfp_mask
, order
);
4487 return make_alloc_exact((unsigned long)page_address(p
), order
, size
);
4491 * free_pages_exact - release memory allocated via alloc_pages_exact()
4492 * @virt: the value returned by alloc_pages_exact.
4493 * @size: size of allocation, same value as passed to alloc_pages_exact().
4495 * Release the memory allocated by a previous call to alloc_pages_exact.
4497 void free_pages_exact(void *virt
, size_t size
)
4499 unsigned long addr
= (unsigned long)virt
;
4500 unsigned long end
= addr
+ PAGE_ALIGN(size
);
4502 while (addr
< end
) {
4507 EXPORT_SYMBOL(free_pages_exact
);
4510 * nr_free_zone_pages - count number of pages beyond high watermark
4511 * @offset: The zone index of the highest zone
4513 * nr_free_zone_pages() counts the number of counts pages which are beyond the
4514 * high watermark within all zones at or below a given zone index. For each
4515 * zone, the number of pages is calculated as:
4517 * nr_free_zone_pages = managed_pages - high_pages
4519 static unsigned long nr_free_zone_pages(int offset
)
4524 /* Just pick one node, since fallback list is circular */
4525 unsigned long sum
= 0;
4527 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
4529 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
4530 unsigned long size
= zone
->managed_pages
;
4531 unsigned long high
= high_wmark_pages(zone
);
4540 * nr_free_buffer_pages - count number of pages beyond high watermark
4542 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4543 * watermark within ZONE_DMA and ZONE_NORMAL.
4545 unsigned long nr_free_buffer_pages(void)
4547 return nr_free_zone_pages(gfp_zone(GFP_USER
));
4549 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
4552 * nr_free_pagecache_pages - count number of pages beyond high watermark
4554 * nr_free_pagecache_pages() counts the number of pages which are beyond the
4555 * high watermark within all zones.
4557 unsigned long nr_free_pagecache_pages(void)
4559 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
4562 static inline void show_node(struct zone
*zone
)
4564 if (IS_ENABLED(CONFIG_NUMA
))
4565 printk("Node %d ", zone_to_nid(zone
));
4568 long si_mem_available(void)
4571 unsigned long pagecache
;
4572 unsigned long wmark_low
= 0;
4573 unsigned long pages
[NR_LRU_LISTS
];
4577 for (lru
= LRU_BASE
; lru
< NR_LRU_LISTS
; lru
++)
4578 pages
[lru
] = global_node_page_state(NR_LRU_BASE
+ lru
);
4581 wmark_low
+= zone
->watermark
[WMARK_LOW
];
4584 * Estimate the amount of memory available for userspace allocations,
4585 * without causing swapping.
4587 available
= global_zone_page_state(NR_FREE_PAGES
) - totalreserve_pages
;
4590 * Not all the page cache can be freed, otherwise the system will
4591 * start swapping. Assume at least half of the page cache, or the
4592 * low watermark worth of cache, needs to stay.
4594 pagecache
= pages
[LRU_ACTIVE_FILE
] + pages
[LRU_INACTIVE_FILE
];
4595 pagecache
-= min(pagecache
/ 2, wmark_low
);
4596 available
+= pagecache
;
4599 * Part of the reclaimable slab consists of items that are in use,
4600 * and cannot be freed. Cap this estimate at the low watermark.
4602 available
+= global_node_page_state(NR_SLAB_RECLAIMABLE
) -
4603 min(global_node_page_state(NR_SLAB_RECLAIMABLE
) / 2,
4610 EXPORT_SYMBOL_GPL(si_mem_available
);
4612 void si_meminfo(struct sysinfo
*val
)
4614 val
->totalram
= totalram_pages
;
4615 val
->sharedram
= global_node_page_state(NR_SHMEM
);
4616 val
->freeram
= global_zone_page_state(NR_FREE_PAGES
);
4617 val
->bufferram
= nr_blockdev_pages();
4618 val
->totalhigh
= totalhigh_pages
;
4619 val
->freehigh
= nr_free_highpages();
4620 val
->mem_unit
= PAGE_SIZE
;
4623 EXPORT_SYMBOL(si_meminfo
);
4626 void si_meminfo_node(struct sysinfo
*val
, int nid
)
4628 int zone_type
; /* needs to be signed */
4629 unsigned long managed_pages
= 0;
4630 unsigned long managed_highpages
= 0;
4631 unsigned long free_highpages
= 0;
4632 pg_data_t
*pgdat
= NODE_DATA(nid
);
4634 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++)
4635 managed_pages
+= pgdat
->node_zones
[zone_type
].managed_pages
;
4636 val
->totalram
= managed_pages
;
4637 val
->sharedram
= node_page_state(pgdat
, NR_SHMEM
);
4638 val
->freeram
= sum_zone_node_page_state(nid
, NR_FREE_PAGES
);
4639 #ifdef CONFIG_HIGHMEM
4640 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
4641 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
4643 if (is_highmem(zone
)) {
4644 managed_highpages
+= zone
->managed_pages
;
4645 free_highpages
+= zone_page_state(zone
, NR_FREE_PAGES
);
4648 val
->totalhigh
= managed_highpages
;
4649 val
->freehigh
= free_highpages
;
4651 val
->totalhigh
= managed_highpages
;
4652 val
->freehigh
= free_highpages
;
4654 val
->mem_unit
= PAGE_SIZE
;
4659 * Determine whether the node should be displayed or not, depending on whether
4660 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
4662 static bool show_mem_node_skip(unsigned int flags
, int nid
, nodemask_t
*nodemask
)
4664 if (!(flags
& SHOW_MEM_FILTER_NODES
))
4668 * no node mask - aka implicit memory numa policy. Do not bother with
4669 * the synchronization - read_mems_allowed_begin - because we do not
4670 * have to be precise here.
4673 nodemask
= &cpuset_current_mems_allowed
;
4675 return !node_isset(nid
, *nodemask
);
4678 #define K(x) ((x) << (PAGE_SHIFT-10))
4680 static void show_migration_types(unsigned char type
)
4682 static const char types
[MIGRATE_TYPES
] = {
4683 [MIGRATE_UNMOVABLE
] = 'U',
4684 [MIGRATE_MOVABLE
] = 'M',
4685 [MIGRATE_RECLAIMABLE
] = 'E',
4686 [MIGRATE_HIGHATOMIC
] = 'H',
4688 [MIGRATE_CMA
] = 'C',
4690 #ifdef CONFIG_MEMORY_ISOLATION
4691 [MIGRATE_ISOLATE
] = 'I',
4694 char tmp
[MIGRATE_TYPES
+ 1];
4698 for (i
= 0; i
< MIGRATE_TYPES
; i
++) {
4699 if (type
& (1 << i
))
4704 printk(KERN_CONT
"(%s) ", tmp
);
4708 * Show free area list (used inside shift_scroll-lock stuff)
4709 * We also calculate the percentage fragmentation. We do this by counting the
4710 * memory on each free list with the exception of the first item on the list.
4713 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
4716 void show_free_areas(unsigned int filter
, nodemask_t
*nodemask
)
4718 unsigned long free_pcp
= 0;
4723 for_each_populated_zone(zone
) {
4724 if (show_mem_node_skip(filter
, zone_to_nid(zone
), nodemask
))
4727 for_each_online_cpu(cpu
)
4728 free_pcp
+= per_cpu_ptr(zone
->pageset
, cpu
)->pcp
.count
;
4731 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
4732 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
4733 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
4734 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
4735 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
4736 " free:%lu free_pcp:%lu free_cma:%lu\n",
4737 global_node_page_state(NR_ACTIVE_ANON
),
4738 global_node_page_state(NR_INACTIVE_ANON
),
4739 global_node_page_state(NR_ISOLATED_ANON
),
4740 global_node_page_state(NR_ACTIVE_FILE
),
4741 global_node_page_state(NR_INACTIVE_FILE
),
4742 global_node_page_state(NR_ISOLATED_FILE
),
4743 global_node_page_state(NR_UNEVICTABLE
),
4744 global_node_page_state(NR_FILE_DIRTY
),
4745 global_node_page_state(NR_WRITEBACK
),
4746 global_node_page_state(NR_UNSTABLE_NFS
),
4747 global_node_page_state(NR_SLAB_RECLAIMABLE
),
4748 global_node_page_state(NR_SLAB_UNRECLAIMABLE
),
4749 global_node_page_state(NR_FILE_MAPPED
),
4750 global_node_page_state(NR_SHMEM
),
4751 global_zone_page_state(NR_PAGETABLE
),
4752 global_zone_page_state(NR_BOUNCE
),
4753 global_zone_page_state(NR_FREE_PAGES
),
4755 global_zone_page_state(NR_FREE_CMA_PAGES
));
4757 for_each_online_pgdat(pgdat
) {
4758 if (show_mem_node_skip(filter
, pgdat
->node_id
, nodemask
))
4762 " active_anon:%lukB"
4763 " inactive_anon:%lukB"
4764 " active_file:%lukB"
4765 " inactive_file:%lukB"
4766 " unevictable:%lukB"
4767 " isolated(anon):%lukB"
4768 " isolated(file):%lukB"
4773 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4775 " shmem_pmdmapped: %lukB"
4778 " writeback_tmp:%lukB"
4780 " all_unreclaimable? %s"
4783 K(node_page_state(pgdat
, NR_ACTIVE_ANON
)),
4784 K(node_page_state(pgdat
, NR_INACTIVE_ANON
)),
4785 K(node_page_state(pgdat
, NR_ACTIVE_FILE
)),
4786 K(node_page_state(pgdat
, NR_INACTIVE_FILE
)),
4787 K(node_page_state(pgdat
, NR_UNEVICTABLE
)),
4788 K(node_page_state(pgdat
, NR_ISOLATED_ANON
)),
4789 K(node_page_state(pgdat
, NR_ISOLATED_FILE
)),
4790 K(node_page_state(pgdat
, NR_FILE_MAPPED
)),
4791 K(node_page_state(pgdat
, NR_FILE_DIRTY
)),
4792 K(node_page_state(pgdat
, NR_WRITEBACK
)),
4793 K(node_page_state(pgdat
, NR_SHMEM
)),
4794 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4795 K(node_page_state(pgdat
, NR_SHMEM_THPS
) * HPAGE_PMD_NR
),
4796 K(node_page_state(pgdat
, NR_SHMEM_PMDMAPPED
)
4798 K(node_page_state(pgdat
, NR_ANON_THPS
) * HPAGE_PMD_NR
),
4800 K(node_page_state(pgdat
, NR_WRITEBACK_TEMP
)),
4801 K(node_page_state(pgdat
, NR_UNSTABLE_NFS
)),
4802 pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
?
4806 for_each_populated_zone(zone
) {
4809 if (show_mem_node_skip(filter
, zone_to_nid(zone
), nodemask
))
4813 for_each_online_cpu(cpu
)
4814 free_pcp
+= per_cpu_ptr(zone
->pageset
, cpu
)->pcp
.count
;
4823 " active_anon:%lukB"
4824 " inactive_anon:%lukB"
4825 " active_file:%lukB"
4826 " inactive_file:%lukB"
4827 " unevictable:%lukB"
4828 " writepending:%lukB"
4832 " kernel_stack:%lukB"
4840 K(zone_page_state(zone
, NR_FREE_PAGES
)),
4841 K(min_wmark_pages(zone
)),
4842 K(low_wmark_pages(zone
)),
4843 K(high_wmark_pages(zone
)),
4844 K(zone_page_state(zone
, NR_ZONE_ACTIVE_ANON
)),
4845 K(zone_page_state(zone
, NR_ZONE_INACTIVE_ANON
)),
4846 K(zone_page_state(zone
, NR_ZONE_ACTIVE_FILE
)),
4847 K(zone_page_state(zone
, NR_ZONE_INACTIVE_FILE
)),
4848 K(zone_page_state(zone
, NR_ZONE_UNEVICTABLE
)),
4849 K(zone_page_state(zone
, NR_ZONE_WRITE_PENDING
)),
4850 K(zone
->present_pages
),
4851 K(zone
->managed_pages
),
4852 K(zone_page_state(zone
, NR_MLOCK
)),
4853 zone_page_state(zone
, NR_KERNEL_STACK_KB
),
4854 K(zone_page_state(zone
, NR_PAGETABLE
)),
4855 K(zone_page_state(zone
, NR_BOUNCE
)),
4857 K(this_cpu_read(zone
->pageset
->pcp
.count
)),
4858 K(zone_page_state(zone
, NR_FREE_CMA_PAGES
)));
4859 printk("lowmem_reserve[]:");
4860 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4861 printk(KERN_CONT
" %ld", zone
->lowmem_reserve
[i
]);
4862 printk(KERN_CONT
"\n");
4865 for_each_populated_zone(zone
) {
4867 unsigned long nr
[MAX_ORDER
], flags
, total
= 0;
4868 unsigned char types
[MAX_ORDER
];
4870 if (show_mem_node_skip(filter
, zone_to_nid(zone
), nodemask
))
4873 printk(KERN_CONT
"%s: ", zone
->name
);
4875 spin_lock_irqsave(&zone
->lock
, flags
);
4876 for (order
= 0; order
< MAX_ORDER
; order
++) {
4877 struct free_area
*area
= &zone
->free_area
[order
];
4880 nr
[order
] = area
->nr_free
;
4881 total
+= nr
[order
] << order
;
4884 for (type
= 0; type
< MIGRATE_TYPES
; type
++) {
4885 if (!list_empty(&area
->free_list
[type
]))
4886 types
[order
] |= 1 << type
;
4889 spin_unlock_irqrestore(&zone
->lock
, flags
);
4890 for (order
= 0; order
< MAX_ORDER
; order
++) {
4891 printk(KERN_CONT
"%lu*%lukB ",
4892 nr
[order
], K(1UL) << order
);
4894 show_migration_types(types
[order
]);
4896 printk(KERN_CONT
"= %lukB\n", K(total
));
4899 hugetlb_show_meminfo();
4901 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES
));
4903 show_swap_cache_info();
4906 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
4908 zoneref
->zone
= zone
;
4909 zoneref
->zone_idx
= zone_idx(zone
);
4913 * Builds allocation fallback zone lists.
4915 * Add all populated zones of a node to the zonelist.
4917 static int build_zonerefs_node(pg_data_t
*pgdat
, struct zoneref
*zonerefs
)
4920 enum zone_type zone_type
= MAX_NR_ZONES
;
4925 zone
= pgdat
->node_zones
+ zone_type
;
4926 if (managed_zone(zone
)) {
4927 zoneref_set_zone(zone
, &zonerefs
[nr_zones
++]);
4928 check_highest_zone(zone_type
);
4930 } while (zone_type
);
4937 static int __parse_numa_zonelist_order(char *s
)
4940 * We used to support different zonlists modes but they turned
4941 * out to be just not useful. Let's keep the warning in place
4942 * if somebody still use the cmd line parameter so that we do
4943 * not fail it silently
4945 if (!(*s
== 'd' || *s
== 'D' || *s
== 'n' || *s
== 'N')) {
4946 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s
);
4952 static __init
int setup_numa_zonelist_order(char *s
)
4957 return __parse_numa_zonelist_order(s
);
4959 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
4961 char numa_zonelist_order
[] = "Node";
4964 * sysctl handler for numa_zonelist_order
4966 int numa_zonelist_order_handler(struct ctl_table
*table
, int write
,
4967 void __user
*buffer
, size_t *length
,
4974 return proc_dostring(table
, write
, buffer
, length
, ppos
);
4975 str
= memdup_user_nul(buffer
, 16);
4977 return PTR_ERR(str
);
4979 ret
= __parse_numa_zonelist_order(str
);
4985 #define MAX_NODE_LOAD (nr_online_nodes)
4986 static int node_load
[MAX_NUMNODES
];
4989 * find_next_best_node - find the next node that should appear in a given node's fallback list
4990 * @node: node whose fallback list we're appending
4991 * @used_node_mask: nodemask_t of already used nodes
4993 * We use a number of factors to determine which is the next node that should
4994 * appear on a given node's fallback list. The node should not have appeared
4995 * already in @node's fallback list, and it should be the next closest node
4996 * according to the distance array (which contains arbitrary distance values
4997 * from each node to each node in the system), and should also prefer nodes
4998 * with no CPUs, since presumably they'll have very little allocation pressure
4999 * on them otherwise.
5000 * It returns -1 if no node is found.
5002 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
5005 int min_val
= INT_MAX
;
5006 int best_node
= NUMA_NO_NODE
;
5007 const struct cpumask
*tmp
= cpumask_of_node(0);
5009 /* Use the local node if we haven't already */
5010 if (!node_isset(node
, *used_node_mask
)) {
5011 node_set(node
, *used_node_mask
);
5015 for_each_node_state(n
, N_MEMORY
) {
5017 /* Don't want a node to appear more than once */
5018 if (node_isset(n
, *used_node_mask
))
5021 /* Use the distance array to find the distance */
5022 val
= node_distance(node
, n
);
5024 /* Penalize nodes under us ("prefer the next node") */
5027 /* Give preference to headless and unused nodes */
5028 tmp
= cpumask_of_node(n
);
5029 if (!cpumask_empty(tmp
))
5030 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
5032 /* Slight preference for less loaded node */
5033 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
5034 val
+= node_load
[n
];
5036 if (val
< min_val
) {
5043 node_set(best_node
, *used_node_mask
);
5050 * Build zonelists ordered by node and zones within node.
5051 * This results in maximum locality--normal zone overflows into local
5052 * DMA zone, if any--but risks exhausting DMA zone.
5054 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int *node_order
,
5057 struct zoneref
*zonerefs
;
5060 zonerefs
= pgdat
->node_zonelists
[ZONELIST_FALLBACK
]._zonerefs
;
5062 for (i
= 0; i
< nr_nodes
; i
++) {
5065 pg_data_t
*node
= NODE_DATA(node_order
[i
]);
5067 nr_zones
= build_zonerefs_node(node
, zonerefs
);
5068 zonerefs
+= nr_zones
;
5070 zonerefs
->zone
= NULL
;
5071 zonerefs
->zone_idx
= 0;
5075 * Build gfp_thisnode zonelists
5077 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
5079 struct zoneref
*zonerefs
;
5082 zonerefs
= pgdat
->node_zonelists
[ZONELIST_NOFALLBACK
]._zonerefs
;
5083 nr_zones
= build_zonerefs_node(pgdat
, zonerefs
);
5084 zonerefs
+= nr_zones
;
5085 zonerefs
->zone
= NULL
;
5086 zonerefs
->zone_idx
= 0;
5090 * Build zonelists ordered by zone and nodes within zones.
5091 * This results in conserving DMA zone[s] until all Normal memory is
5092 * exhausted, but results in overflowing to remote node while memory
5093 * may still exist in local DMA zone.
5096 static void build_zonelists(pg_data_t
*pgdat
)
5098 static int node_order
[MAX_NUMNODES
];
5099 int node
, load
, nr_nodes
= 0;
5100 nodemask_t used_mask
;
5101 int local_node
, prev_node
;
5103 /* NUMA-aware ordering of nodes */
5104 local_node
= pgdat
->node_id
;
5105 load
= nr_online_nodes
;
5106 prev_node
= local_node
;
5107 nodes_clear(used_mask
);
5109 memset(node_order
, 0, sizeof(node_order
));
5110 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
5112 * We don't want to pressure a particular node.
5113 * So adding penalty to the first node in same
5114 * distance group to make it round-robin.
5116 if (node_distance(local_node
, node
) !=
5117 node_distance(local_node
, prev_node
))
5118 node_load
[node
] = load
;
5120 node_order
[nr_nodes
++] = node
;
5125 build_zonelists_in_node_order(pgdat
, node_order
, nr_nodes
);
5126 build_thisnode_zonelists(pgdat
);
5129 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5131 * Return node id of node used for "local" allocations.
5132 * I.e., first node id of first zone in arg node's generic zonelist.
5133 * Used for initializing percpu 'numa_mem', which is used primarily
5134 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5136 int local_memory_node(int node
)
5140 z
= first_zones_zonelist(node_zonelist(node
, GFP_KERNEL
),
5141 gfp_zone(GFP_KERNEL
),
5143 return z
->zone
->node
;
5147 static void setup_min_unmapped_ratio(void);
5148 static void setup_min_slab_ratio(void);
5149 #else /* CONFIG_NUMA */
5151 static void build_zonelists(pg_data_t
*pgdat
)
5153 int node
, local_node
;
5154 struct zoneref
*zonerefs
;
5157 local_node
= pgdat
->node_id
;
5159 zonerefs
= pgdat
->node_zonelists
[ZONELIST_FALLBACK
]._zonerefs
;
5160 nr_zones
= build_zonerefs_node(pgdat
, zonerefs
);
5161 zonerefs
+= nr_zones
;
5164 * Now we build the zonelist so that it contains the zones
5165 * of all the other nodes.
5166 * We don't want to pressure a particular node, so when
5167 * building the zones for node N, we make sure that the
5168 * zones coming right after the local ones are those from
5169 * node N+1 (modulo N)
5171 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
5172 if (!node_online(node
))
5174 nr_zones
= build_zonerefs_node(NODE_DATA(node
), zonerefs
);
5175 zonerefs
+= nr_zones
;
5177 for (node
= 0; node
< local_node
; node
++) {
5178 if (!node_online(node
))
5180 nr_zones
= build_zonerefs_node(NODE_DATA(node
), zonerefs
);
5181 zonerefs
+= nr_zones
;
5184 zonerefs
->zone
= NULL
;
5185 zonerefs
->zone_idx
= 0;
5188 #endif /* CONFIG_NUMA */
5191 * Boot pageset table. One per cpu which is going to be used for all
5192 * zones and all nodes. The parameters will be set in such a way
5193 * that an item put on a list will immediately be handed over to
5194 * the buddy list. This is safe since pageset manipulation is done
5195 * with interrupts disabled.
5197 * The boot_pagesets must be kept even after bootup is complete for
5198 * unused processors and/or zones. They do play a role for bootstrapping
5199 * hotplugged processors.
5201 * zoneinfo_show() and maybe other functions do
5202 * not check if the processor is online before following the pageset pointer.
5203 * Other parts of the kernel may not check if the zone is available.
5205 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
);
5206 static DEFINE_PER_CPU(struct per_cpu_pageset
, boot_pageset
);
5207 static DEFINE_PER_CPU(struct per_cpu_nodestat
, boot_nodestats
);
5209 static void __build_all_zonelists(void *data
)
5212 int __maybe_unused cpu
;
5213 pg_data_t
*self
= data
;
5214 static DEFINE_SPINLOCK(lock
);
5219 memset(node_load
, 0, sizeof(node_load
));
5223 * This node is hotadded and no memory is yet present. So just
5224 * building zonelists is fine - no need to touch other nodes.
5226 if (self
&& !node_online(self
->node_id
)) {
5227 build_zonelists(self
);
5229 for_each_online_node(nid
) {
5230 pg_data_t
*pgdat
= NODE_DATA(nid
);
5232 build_zonelists(pgdat
);
5235 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5237 * We now know the "local memory node" for each node--
5238 * i.e., the node of the first zone in the generic zonelist.
5239 * Set up numa_mem percpu variable for on-line cpus. During
5240 * boot, only the boot cpu should be on-line; we'll init the
5241 * secondary cpus' numa_mem as they come on-line. During
5242 * node/memory hotplug, we'll fixup all on-line cpus.
5244 for_each_online_cpu(cpu
)
5245 set_cpu_numa_mem(cpu
, local_memory_node(cpu_to_node(cpu
)));
5252 static noinline
void __init
5253 build_all_zonelists_init(void)
5257 __build_all_zonelists(NULL
);
5260 * Initialize the boot_pagesets that are going to be used
5261 * for bootstrapping processors. The real pagesets for
5262 * each zone will be allocated later when the per cpu
5263 * allocator is available.
5265 * boot_pagesets are used also for bootstrapping offline
5266 * cpus if the system is already booted because the pagesets
5267 * are needed to initialize allocators on a specific cpu too.
5268 * F.e. the percpu allocator needs the page allocator which
5269 * needs the percpu allocator in order to allocate its pagesets
5270 * (a chicken-egg dilemma).
5272 for_each_possible_cpu(cpu
)
5273 setup_pageset(&per_cpu(boot_pageset
, cpu
), 0);
5275 mminit_verify_zonelist();
5276 cpuset_init_current_mems_allowed();
5280 * unless system_state == SYSTEM_BOOTING.
5282 * __ref due to call of __init annotated helper build_all_zonelists_init
5283 * [protected by SYSTEM_BOOTING].
5285 void __ref
build_all_zonelists(pg_data_t
*pgdat
)
5287 if (system_state
== SYSTEM_BOOTING
) {
5288 build_all_zonelists_init();
5290 __build_all_zonelists(pgdat
);
5291 /* cpuset refresh routine should be here */
5293 vm_total_pages
= nr_free_pagecache_pages();
5295 * Disable grouping by mobility if the number of pages in the
5296 * system is too low to allow the mechanism to work. It would be
5297 * more accurate, but expensive to check per-zone. This check is
5298 * made on memory-hotadd so a system can start with mobility
5299 * disabled and enable it later
5301 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
5302 page_group_by_mobility_disabled
= 1;
5304 page_group_by_mobility_disabled
= 0;
5306 pr_info("Built %i zonelists, mobility grouping %s. Total pages: %ld\n",
5308 page_group_by_mobility_disabled
? "off" : "on",
5311 pr_info("Policy zone: %s\n", zone_names
[policy_zone
]);
5316 * Initially all pages are reserved - free ones are freed
5317 * up by free_all_bootmem() once the early boot process is
5318 * done. Non-atomic initialization, single-pass.
5320 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
5321 unsigned long start_pfn
, enum memmap_context context
)
5323 struct vmem_altmap
*altmap
= to_vmem_altmap(__pfn_to_phys(start_pfn
));
5324 unsigned long end_pfn
= start_pfn
+ size
;
5325 pg_data_t
*pgdat
= NODE_DATA(nid
);
5327 unsigned long nr_initialised
= 0;
5328 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5329 struct memblock_region
*r
= NULL
, *tmp
;
5332 if (highest_memmap_pfn
< end_pfn
- 1)
5333 highest_memmap_pfn
= end_pfn
- 1;
5336 * Honor reservation requested by the driver for this ZONE_DEVICE
5339 if (altmap
&& start_pfn
== altmap
->base_pfn
)
5340 start_pfn
+= altmap
->reserve
;
5342 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
5344 * There can be holes in boot-time mem_map[]s handed to this
5345 * function. They do not exist on hotplugged memory.
5347 if (context
!= MEMMAP_EARLY
)
5350 if (!early_pfn_valid(pfn
))
5352 if (!early_pfn_in_nid(pfn
, nid
))
5354 if (!update_defer_init(pgdat
, pfn
, end_pfn
, &nr_initialised
))
5357 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5359 * Check given memblock attribute by firmware which can affect
5360 * kernel memory layout. If zone==ZONE_MOVABLE but memory is
5361 * mirrored, it's an overlapped memmap init. skip it.
5363 if (mirrored_kernelcore
&& zone
== ZONE_MOVABLE
) {
5364 if (!r
|| pfn
>= memblock_region_memory_end_pfn(r
)) {
5365 for_each_memblock(memory
, tmp
)
5366 if (pfn
< memblock_region_memory_end_pfn(tmp
))
5370 if (pfn
>= memblock_region_memory_base_pfn(r
) &&
5371 memblock_is_mirror(r
)) {
5372 /* already initialized as NORMAL */
5373 pfn
= memblock_region_memory_end_pfn(r
);
5381 * Mark the block movable so that blocks are reserved for
5382 * movable at startup. This will force kernel allocations
5383 * to reserve their blocks rather than leaking throughout
5384 * the address space during boot when many long-lived
5385 * kernel allocations are made.
5387 * bitmap is created for zone's valid pfn range. but memmap
5388 * can be created for invalid pages (for alignment)
5389 * check here not to call set_pageblock_migratetype() against
5392 * Please note that MEMMAP_HOTPLUG path doesn't clear memmap
5393 * because this is done early in sparse_add_one_section
5395 if (!(pfn
& (pageblock_nr_pages
- 1))) {
5396 struct page
*page
= pfn_to_page(pfn
);
5398 __init_single_page(page
, pfn
, zone
, nid
,
5399 context
!= MEMMAP_HOTPLUG
);
5400 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
5403 __init_single_pfn(pfn
, zone
, nid
,
5404 context
!= MEMMAP_HOTPLUG
);
5409 static void __meminit
zone_init_free_lists(struct zone
*zone
)
5411 unsigned int order
, t
;
5412 for_each_migratetype_order(order
, t
) {
5413 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
5414 zone
->free_area
[order
].nr_free
= 0;
5418 #ifndef __HAVE_ARCH_MEMMAP_INIT
5419 #define memmap_init(size, nid, zone, start_pfn) \
5420 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
5423 static int zone_batchsize(struct zone
*zone
)
5429 * The per-cpu-pages pools are set to around 1000th of the
5430 * size of the zone. But no more than 1/2 of a meg.
5432 * OK, so we don't know how big the cache is. So guess.
5434 batch
= zone
->managed_pages
/ 1024;
5435 if (batch
* PAGE_SIZE
> 512 * 1024)
5436 batch
= (512 * 1024) / PAGE_SIZE
;
5437 batch
/= 4; /* We effectively *= 4 below */
5442 * Clamp the batch to a 2^n - 1 value. Having a power
5443 * of 2 value was found to be more likely to have
5444 * suboptimal cache aliasing properties in some cases.
5446 * For example if 2 tasks are alternately allocating
5447 * batches of pages, one task can end up with a lot
5448 * of pages of one half of the possible page colors
5449 * and the other with pages of the other colors.
5451 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
5456 /* The deferral and batching of frees should be suppressed under NOMMU
5459 * The problem is that NOMMU needs to be able to allocate large chunks
5460 * of contiguous memory as there's no hardware page translation to
5461 * assemble apparent contiguous memory from discontiguous pages.
5463 * Queueing large contiguous runs of pages for batching, however,
5464 * causes the pages to actually be freed in smaller chunks. As there
5465 * can be a significant delay between the individual batches being
5466 * recycled, this leads to the once large chunks of space being
5467 * fragmented and becoming unavailable for high-order allocations.
5474 * pcp->high and pcp->batch values are related and dependent on one another:
5475 * ->batch must never be higher then ->high.
5476 * The following function updates them in a safe manner without read side
5479 * Any new users of pcp->batch and pcp->high should ensure they can cope with
5480 * those fields changing asynchronously (acording the the above rule).
5482 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5483 * outside of boot time (or some other assurance that no concurrent updaters
5486 static void pageset_update(struct per_cpu_pages
*pcp
, unsigned long high
,
5487 unsigned long batch
)
5489 /* start with a fail safe value for batch */
5493 /* Update high, then batch, in order */
5500 /* a companion to pageset_set_high() */
5501 static void pageset_set_batch(struct per_cpu_pageset
*p
, unsigned long batch
)
5503 pageset_update(&p
->pcp
, 6 * batch
, max(1UL, 1 * batch
));
5506 static void pageset_init(struct per_cpu_pageset
*p
)
5508 struct per_cpu_pages
*pcp
;
5511 memset(p
, 0, sizeof(*p
));
5515 for (migratetype
= 0; migratetype
< MIGRATE_PCPTYPES
; migratetype
++)
5516 INIT_LIST_HEAD(&pcp
->lists
[migratetype
]);
5519 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
5522 pageset_set_batch(p
, batch
);
5526 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5527 * to the value high for the pageset p.
5529 static void pageset_set_high(struct per_cpu_pageset
*p
,
5532 unsigned long batch
= max(1UL, high
/ 4);
5533 if ((high
/ 4) > (PAGE_SHIFT
* 8))
5534 batch
= PAGE_SHIFT
* 8;
5536 pageset_update(&p
->pcp
, high
, batch
);
5539 static void pageset_set_high_and_batch(struct zone
*zone
,
5540 struct per_cpu_pageset
*pcp
)
5542 if (percpu_pagelist_fraction
)
5543 pageset_set_high(pcp
,
5544 (zone
->managed_pages
/
5545 percpu_pagelist_fraction
));
5547 pageset_set_batch(pcp
, zone_batchsize(zone
));
5550 static void __meminit
zone_pageset_init(struct zone
*zone
, int cpu
)
5552 struct per_cpu_pageset
*pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
5555 pageset_set_high_and_batch(zone
, pcp
);
5558 void __meminit
setup_zone_pageset(struct zone
*zone
)
5561 zone
->pageset
= alloc_percpu(struct per_cpu_pageset
);
5562 for_each_possible_cpu(cpu
)
5563 zone_pageset_init(zone
, cpu
);
5567 * Allocate per cpu pagesets and initialize them.
5568 * Before this call only boot pagesets were available.
5570 void __init
setup_per_cpu_pageset(void)
5572 struct pglist_data
*pgdat
;
5575 for_each_populated_zone(zone
)
5576 setup_zone_pageset(zone
);
5578 for_each_online_pgdat(pgdat
)
5579 pgdat
->per_cpu_nodestats
=
5580 alloc_percpu(struct per_cpu_nodestat
);
5583 static __meminit
void zone_pcp_init(struct zone
*zone
)
5586 * per cpu subsystem is not up at this point. The following code
5587 * relies on the ability of the linker to provide the
5588 * offset of a (static) per cpu variable into the per cpu area.
5590 zone
->pageset
= &boot_pageset
;
5592 if (populated_zone(zone
))
5593 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%u\n",
5594 zone
->name
, zone
->present_pages
,
5595 zone_batchsize(zone
));
5598 void __meminit
init_currently_empty_zone(struct zone
*zone
,
5599 unsigned long zone_start_pfn
,
5602 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
5604 pgdat
->nr_zones
= zone_idx(zone
) + 1;
5606 zone
->zone_start_pfn
= zone_start_pfn
;
5608 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
5609 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5611 (unsigned long)zone_idx(zone
),
5612 zone_start_pfn
, (zone_start_pfn
+ size
));
5614 zone_init_free_lists(zone
);
5615 zone
->initialized
= 1;
5618 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5619 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5622 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5624 int __meminit
__early_pfn_to_nid(unsigned long pfn
,
5625 struct mminit_pfnnid_cache
*state
)
5627 unsigned long start_pfn
, end_pfn
;
5630 if (state
->last_start
<= pfn
&& pfn
< state
->last_end
)
5631 return state
->last_nid
;
5633 nid
= memblock_search_pfn_nid(pfn
, &start_pfn
, &end_pfn
);
5635 state
->last_start
= start_pfn
;
5636 state
->last_end
= end_pfn
;
5637 state
->last_nid
= nid
;
5642 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5645 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5646 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5647 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5649 * If an architecture guarantees that all ranges registered contain no holes
5650 * and may be freed, this this function may be used instead of calling
5651 * memblock_free_early_nid() manually.
5653 void __init
free_bootmem_with_active_regions(int nid
, unsigned long max_low_pfn
)
5655 unsigned long start_pfn
, end_pfn
;
5658 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
) {
5659 start_pfn
= min(start_pfn
, max_low_pfn
);
5660 end_pfn
= min(end_pfn
, max_low_pfn
);
5662 if (start_pfn
< end_pfn
)
5663 memblock_free_early_nid(PFN_PHYS(start_pfn
),
5664 (end_pfn
- start_pfn
) << PAGE_SHIFT
,
5670 * sparse_memory_present_with_active_regions - Call memory_present for each active range
5671 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
5673 * If an architecture guarantees that all ranges registered contain no holes and may
5674 * be freed, this function may be used instead of calling memory_present() manually.
5676 void __init
sparse_memory_present_with_active_regions(int nid
)
5678 unsigned long start_pfn
, end_pfn
;
5681 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
)
5682 memory_present(this_nid
, start_pfn
, end_pfn
);
5686 * get_pfn_range_for_nid - Return the start and end page frames for a node
5687 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5688 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5689 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5691 * It returns the start and end page frame of a node based on information
5692 * provided by memblock_set_node(). If called for a node
5693 * with no available memory, a warning is printed and the start and end
5696 void __meminit
get_pfn_range_for_nid(unsigned int nid
,
5697 unsigned long *start_pfn
, unsigned long *end_pfn
)
5699 unsigned long this_start_pfn
, this_end_pfn
;
5705 for_each_mem_pfn_range(i
, nid
, &this_start_pfn
, &this_end_pfn
, NULL
) {
5706 *start_pfn
= min(*start_pfn
, this_start_pfn
);
5707 *end_pfn
= max(*end_pfn
, this_end_pfn
);
5710 if (*start_pfn
== -1UL)
5715 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5716 * assumption is made that zones within a node are ordered in monotonic
5717 * increasing memory addresses so that the "highest" populated zone is used
5719 static void __init
find_usable_zone_for_movable(void)
5722 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
5723 if (zone_index
== ZONE_MOVABLE
)
5726 if (arch_zone_highest_possible_pfn
[zone_index
] >
5727 arch_zone_lowest_possible_pfn
[zone_index
])
5731 VM_BUG_ON(zone_index
== -1);
5732 movable_zone
= zone_index
;
5736 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5737 * because it is sized independent of architecture. Unlike the other zones,
5738 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5739 * in each node depending on the size of each node and how evenly kernelcore
5740 * is distributed. This helper function adjusts the zone ranges
5741 * provided by the architecture for a given node by using the end of the
5742 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5743 * zones within a node are in order of monotonic increases memory addresses
5745 static void __meminit
adjust_zone_range_for_zone_movable(int nid
,
5746 unsigned long zone_type
,
5747 unsigned long node_start_pfn
,
5748 unsigned long node_end_pfn
,
5749 unsigned long *zone_start_pfn
,
5750 unsigned long *zone_end_pfn
)
5752 /* Only adjust if ZONE_MOVABLE is on this node */
5753 if (zone_movable_pfn
[nid
]) {
5754 /* Size ZONE_MOVABLE */
5755 if (zone_type
== ZONE_MOVABLE
) {
5756 *zone_start_pfn
= zone_movable_pfn
[nid
];
5757 *zone_end_pfn
= min(node_end_pfn
,
5758 arch_zone_highest_possible_pfn
[movable_zone
]);
5760 /* Adjust for ZONE_MOVABLE starting within this range */
5761 } else if (!mirrored_kernelcore
&&
5762 *zone_start_pfn
< zone_movable_pfn
[nid
] &&
5763 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
5764 *zone_end_pfn
= zone_movable_pfn
[nid
];
5766 /* Check if this whole range is within ZONE_MOVABLE */
5767 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
5768 *zone_start_pfn
= *zone_end_pfn
;
5773 * Return the number of pages a zone spans in a node, including holes
5774 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5776 static unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
5777 unsigned long zone_type
,
5778 unsigned long node_start_pfn
,
5779 unsigned long node_end_pfn
,
5780 unsigned long *zone_start_pfn
,
5781 unsigned long *zone_end_pfn
,
5782 unsigned long *ignored
)
5784 /* When hotadd a new node from cpu_up(), the node should be empty */
5785 if (!node_start_pfn
&& !node_end_pfn
)
5788 /* Get the start and end of the zone */
5789 *zone_start_pfn
= arch_zone_lowest_possible_pfn
[zone_type
];
5790 *zone_end_pfn
= arch_zone_highest_possible_pfn
[zone_type
];
5791 adjust_zone_range_for_zone_movable(nid
, zone_type
,
5792 node_start_pfn
, node_end_pfn
,
5793 zone_start_pfn
, zone_end_pfn
);
5795 /* Check that this node has pages within the zone's required range */
5796 if (*zone_end_pfn
< node_start_pfn
|| *zone_start_pfn
> node_end_pfn
)
5799 /* Move the zone boundaries inside the node if necessary */
5800 *zone_end_pfn
= min(*zone_end_pfn
, node_end_pfn
);
5801 *zone_start_pfn
= max(*zone_start_pfn
, node_start_pfn
);
5803 /* Return the spanned pages */
5804 return *zone_end_pfn
- *zone_start_pfn
;
5808 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5809 * then all holes in the requested range will be accounted for.
5811 unsigned long __meminit
__absent_pages_in_range(int nid
,
5812 unsigned long range_start_pfn
,
5813 unsigned long range_end_pfn
)
5815 unsigned long nr_absent
= range_end_pfn
- range_start_pfn
;
5816 unsigned long start_pfn
, end_pfn
;
5819 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
5820 start_pfn
= clamp(start_pfn
, range_start_pfn
, range_end_pfn
);
5821 end_pfn
= clamp(end_pfn
, range_start_pfn
, range_end_pfn
);
5822 nr_absent
-= end_pfn
- start_pfn
;
5828 * absent_pages_in_range - Return number of page frames in holes within a range
5829 * @start_pfn: The start PFN to start searching for holes
5830 * @end_pfn: The end PFN to stop searching for holes
5832 * It returns the number of pages frames in memory holes within a range.
5834 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
5835 unsigned long end_pfn
)
5837 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
5840 /* Return the number of page frames in holes in a zone on a node */
5841 static unsigned long __meminit
zone_absent_pages_in_node(int nid
,
5842 unsigned long zone_type
,
5843 unsigned long node_start_pfn
,
5844 unsigned long node_end_pfn
,
5845 unsigned long *ignored
)
5847 unsigned long zone_low
= arch_zone_lowest_possible_pfn
[zone_type
];
5848 unsigned long zone_high
= arch_zone_highest_possible_pfn
[zone_type
];
5849 unsigned long zone_start_pfn
, zone_end_pfn
;
5850 unsigned long nr_absent
;
5852 /* When hotadd a new node from cpu_up(), the node should be empty */
5853 if (!node_start_pfn
&& !node_end_pfn
)
5856 zone_start_pfn
= clamp(node_start_pfn
, zone_low
, zone_high
);
5857 zone_end_pfn
= clamp(node_end_pfn
, zone_low
, zone_high
);
5859 adjust_zone_range_for_zone_movable(nid
, zone_type
,
5860 node_start_pfn
, node_end_pfn
,
5861 &zone_start_pfn
, &zone_end_pfn
);
5862 nr_absent
= __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
5865 * ZONE_MOVABLE handling.
5866 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
5869 if (mirrored_kernelcore
&& zone_movable_pfn
[nid
]) {
5870 unsigned long start_pfn
, end_pfn
;
5871 struct memblock_region
*r
;
5873 for_each_memblock(memory
, r
) {
5874 start_pfn
= clamp(memblock_region_memory_base_pfn(r
),
5875 zone_start_pfn
, zone_end_pfn
);
5876 end_pfn
= clamp(memblock_region_memory_end_pfn(r
),
5877 zone_start_pfn
, zone_end_pfn
);
5879 if (zone_type
== ZONE_MOVABLE
&&
5880 memblock_is_mirror(r
))
5881 nr_absent
+= end_pfn
- start_pfn
;
5883 if (zone_type
== ZONE_NORMAL
&&
5884 !memblock_is_mirror(r
))
5885 nr_absent
+= end_pfn
- start_pfn
;
5892 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5893 static inline unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
5894 unsigned long zone_type
,
5895 unsigned long node_start_pfn
,
5896 unsigned long node_end_pfn
,
5897 unsigned long *zone_start_pfn
,
5898 unsigned long *zone_end_pfn
,
5899 unsigned long *zones_size
)
5903 *zone_start_pfn
= node_start_pfn
;
5904 for (zone
= 0; zone
< zone_type
; zone
++)
5905 *zone_start_pfn
+= zones_size
[zone
];
5907 *zone_end_pfn
= *zone_start_pfn
+ zones_size
[zone_type
];
5909 return zones_size
[zone_type
];
5912 static inline unsigned long __meminit
zone_absent_pages_in_node(int nid
,
5913 unsigned long zone_type
,
5914 unsigned long node_start_pfn
,
5915 unsigned long node_end_pfn
,
5916 unsigned long *zholes_size
)
5921 return zholes_size
[zone_type
];
5924 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5926 static void __meminit
calculate_node_totalpages(struct pglist_data
*pgdat
,
5927 unsigned long node_start_pfn
,
5928 unsigned long node_end_pfn
,
5929 unsigned long *zones_size
,
5930 unsigned long *zholes_size
)
5932 unsigned long realtotalpages
= 0, totalpages
= 0;
5935 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5936 struct zone
*zone
= pgdat
->node_zones
+ i
;
5937 unsigned long zone_start_pfn
, zone_end_pfn
;
5938 unsigned long size
, real_size
;
5940 size
= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
5946 real_size
= size
- zone_absent_pages_in_node(pgdat
->node_id
, i
,
5947 node_start_pfn
, node_end_pfn
,
5950 zone
->zone_start_pfn
= zone_start_pfn
;
5952 zone
->zone_start_pfn
= 0;
5953 zone
->spanned_pages
= size
;
5954 zone
->present_pages
= real_size
;
5957 realtotalpages
+= real_size
;
5960 pgdat
->node_spanned_pages
= totalpages
;
5961 pgdat
->node_present_pages
= realtotalpages
;
5962 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
5966 #ifndef CONFIG_SPARSEMEM
5968 * Calculate the size of the zone->blockflags rounded to an unsigned long
5969 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5970 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5971 * round what is now in bits to nearest long in bits, then return it in
5974 static unsigned long __init
usemap_size(unsigned long zone_start_pfn
, unsigned long zonesize
)
5976 unsigned long usemapsize
;
5978 zonesize
+= zone_start_pfn
& (pageblock_nr_pages
-1);
5979 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
5980 usemapsize
= usemapsize
>> pageblock_order
;
5981 usemapsize
*= NR_PAGEBLOCK_BITS
;
5982 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
5984 return usemapsize
/ 8;
5987 static void __init
setup_usemap(struct pglist_data
*pgdat
,
5989 unsigned long zone_start_pfn
,
5990 unsigned long zonesize
)
5992 unsigned long usemapsize
= usemap_size(zone_start_pfn
, zonesize
);
5993 zone
->pageblock_flags
= NULL
;
5995 zone
->pageblock_flags
=
5996 memblock_virt_alloc_node_nopanic(usemapsize
,
6000 static inline void setup_usemap(struct pglist_data
*pgdat
, struct zone
*zone
,
6001 unsigned long zone_start_pfn
, unsigned long zonesize
) {}
6002 #endif /* CONFIG_SPARSEMEM */
6004 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6006 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6007 void __paginginit
set_pageblock_order(void)
6011 /* Check that pageblock_nr_pages has not already been setup */
6012 if (pageblock_order
)
6015 if (HPAGE_SHIFT
> PAGE_SHIFT
)
6016 order
= HUGETLB_PAGE_ORDER
;
6018 order
= MAX_ORDER
- 1;
6021 * Assume the largest contiguous order of interest is a huge page.
6022 * This value may be variable depending on boot parameters on IA64 and
6025 pageblock_order
= order
;
6027 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6030 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6031 * is unused as pageblock_order is set at compile-time. See
6032 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6035 void __paginginit
set_pageblock_order(void)
6039 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6041 static unsigned long __paginginit
calc_memmap_size(unsigned long spanned_pages
,
6042 unsigned long present_pages
)
6044 unsigned long pages
= spanned_pages
;
6047 * Provide a more accurate estimation if there are holes within
6048 * the zone and SPARSEMEM is in use. If there are holes within the
6049 * zone, each populated memory region may cost us one or two extra
6050 * memmap pages due to alignment because memmap pages for each
6051 * populated regions may not be naturally aligned on page boundary.
6052 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6054 if (spanned_pages
> present_pages
+ (present_pages
>> 4) &&
6055 IS_ENABLED(CONFIG_SPARSEMEM
))
6056 pages
= present_pages
;
6058 return PAGE_ALIGN(pages
* sizeof(struct page
)) >> PAGE_SHIFT
;
6062 * Set up the zone data structures:
6063 * - mark all pages reserved
6064 * - mark all memory queues empty
6065 * - clear the memory bitmaps
6067 * NOTE: pgdat should get zeroed by caller.
6069 static void __paginginit
free_area_init_core(struct pglist_data
*pgdat
)
6072 int nid
= pgdat
->node_id
;
6074 pgdat_resize_init(pgdat
);
6075 #ifdef CONFIG_NUMA_BALANCING
6076 spin_lock_init(&pgdat
->numabalancing_migrate_lock
);
6077 pgdat
->numabalancing_migrate_nr_pages
= 0;
6078 pgdat
->numabalancing_migrate_next_window
= jiffies
;
6080 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6081 spin_lock_init(&pgdat
->split_queue_lock
);
6082 INIT_LIST_HEAD(&pgdat
->split_queue
);
6083 pgdat
->split_queue_len
= 0;
6085 init_waitqueue_head(&pgdat
->kswapd_wait
);
6086 init_waitqueue_head(&pgdat
->pfmemalloc_wait
);
6087 #ifdef CONFIG_COMPACTION
6088 init_waitqueue_head(&pgdat
->kcompactd_wait
);
6090 pgdat_page_ext_init(pgdat
);
6091 spin_lock_init(&pgdat
->lru_lock
);
6092 lruvec_init(node_lruvec(pgdat
));
6094 pgdat
->per_cpu_nodestats
= &boot_nodestats
;
6096 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
6097 struct zone
*zone
= pgdat
->node_zones
+ j
;
6098 unsigned long size
, realsize
, freesize
, memmap_pages
;
6099 unsigned long zone_start_pfn
= zone
->zone_start_pfn
;
6101 size
= zone
->spanned_pages
;
6102 realsize
= freesize
= zone
->present_pages
;
6105 * Adjust freesize so that it accounts for how much memory
6106 * is used by this zone for memmap. This affects the watermark
6107 * and per-cpu initialisations
6109 memmap_pages
= calc_memmap_size(size
, realsize
);
6110 if (!is_highmem_idx(j
)) {
6111 if (freesize
>= memmap_pages
) {
6112 freesize
-= memmap_pages
;
6115 " %s zone: %lu pages used for memmap\n",
6116 zone_names
[j
], memmap_pages
);
6118 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6119 zone_names
[j
], memmap_pages
, freesize
);
6122 /* Account for reserved pages */
6123 if (j
== 0 && freesize
> dma_reserve
) {
6124 freesize
-= dma_reserve
;
6125 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
6126 zone_names
[0], dma_reserve
);
6129 if (!is_highmem_idx(j
))
6130 nr_kernel_pages
+= freesize
;
6131 /* Charge for highmem memmap if there are enough kernel pages */
6132 else if (nr_kernel_pages
> memmap_pages
* 2)
6133 nr_kernel_pages
-= memmap_pages
;
6134 nr_all_pages
+= freesize
;
6137 * Set an approximate value for lowmem here, it will be adjusted
6138 * when the bootmem allocator frees pages into the buddy system.
6139 * And all highmem pages will be managed by the buddy system.
6141 zone
->managed_pages
= is_highmem_idx(j
) ? realsize
: freesize
;
6145 zone
->name
= zone_names
[j
];
6146 zone
->zone_pgdat
= pgdat
;
6147 spin_lock_init(&zone
->lock
);
6148 zone_seqlock_init(zone
);
6149 zone_pcp_init(zone
);
6154 set_pageblock_order();
6155 setup_usemap(pgdat
, zone
, zone_start_pfn
, size
);
6156 init_currently_empty_zone(zone
, zone_start_pfn
, size
);
6157 memmap_init(size
, nid
, j
, zone_start_pfn
);
6161 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6162 static void __ref
alloc_node_mem_map(struct pglist_data
*pgdat
)
6164 unsigned long __maybe_unused start
= 0;
6165 unsigned long __maybe_unused offset
= 0;
6167 /* Skip empty nodes */
6168 if (!pgdat
->node_spanned_pages
)
6171 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
6172 offset
= pgdat
->node_start_pfn
- start
;
6173 /* ia64 gets its own node_mem_map, before this, without bootmem */
6174 if (!pgdat
->node_mem_map
) {
6175 unsigned long size
, end
;
6179 * The zone's endpoints aren't required to be MAX_ORDER
6180 * aligned but the node_mem_map endpoints must be in order
6181 * for the buddy allocator to function correctly.
6183 end
= pgdat_end_pfn(pgdat
);
6184 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
6185 size
= (end
- start
) * sizeof(struct page
);
6186 map
= alloc_remap(pgdat
->node_id
, size
);
6188 map
= memblock_virt_alloc_node_nopanic(size
,
6190 pgdat
->node_mem_map
= map
+ offset
;
6192 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
6193 __func__
, pgdat
->node_id
, (unsigned long)pgdat
,
6194 (unsigned long)pgdat
->node_mem_map
);
6195 #ifndef CONFIG_NEED_MULTIPLE_NODES
6197 * With no DISCONTIG, the global mem_map is just set as node 0's
6199 if (pgdat
== NODE_DATA(0)) {
6200 mem_map
= NODE_DATA(0)->node_mem_map
;
6201 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6202 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
6204 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6209 static void __ref
alloc_node_mem_map(struct pglist_data
*pgdat
) { }
6210 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6212 void __paginginit
free_area_init_node(int nid
, unsigned long *zones_size
,
6213 unsigned long node_start_pfn
, unsigned long *zholes_size
)
6215 pg_data_t
*pgdat
= NODE_DATA(nid
);
6216 unsigned long start_pfn
= 0;
6217 unsigned long end_pfn
= 0;
6219 /* pg_data_t should be reset to zero when it's allocated */
6220 WARN_ON(pgdat
->nr_zones
|| pgdat
->kswapd_classzone_idx
);
6222 pgdat
->node_id
= nid
;
6223 pgdat
->node_start_pfn
= node_start_pfn
;
6224 pgdat
->per_cpu_nodestats
= NULL
;
6225 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6226 get_pfn_range_for_nid(nid
, &start_pfn
, &end_pfn
);
6227 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid
,
6228 (u64
)start_pfn
<< PAGE_SHIFT
,
6229 end_pfn
? ((u64
)end_pfn
<< PAGE_SHIFT
) - 1 : 0);
6231 start_pfn
= node_start_pfn
;
6233 calculate_node_totalpages(pgdat
, start_pfn
, end_pfn
,
6234 zones_size
, zholes_size
);
6236 alloc_node_mem_map(pgdat
);
6238 reset_deferred_meminit(pgdat
);
6239 free_area_init_core(pgdat
);
6242 #if defined(CONFIG_HAVE_MEMBLOCK) && !defined(CONFIG_FLAT_NODE_MEM_MAP)
6244 * Only struct pages that are backed by physical memory are zeroed and
6245 * initialized by going through __init_single_page(). But, there are some
6246 * struct pages which are reserved in memblock allocator and their fields
6247 * may be accessed (for example page_to_pfn() on some configuration accesses
6248 * flags). We must explicitly zero those struct pages.
6250 void __paginginit
zero_resv_unavail(void)
6252 phys_addr_t start
, end
;
6257 * Loop through ranges that are reserved, but do not have reported
6258 * physical memory backing.
6261 for_each_resv_unavail_range(i
, &start
, &end
) {
6262 for (pfn
= PFN_DOWN(start
); pfn
< PFN_UP(end
); pfn
++) {
6263 if (!pfn_valid(ALIGN_DOWN(pfn
, pageblock_nr_pages
)))
6265 mm_zero_struct_page(pfn_to_page(pfn
));
6271 * Struct pages that do not have backing memory. This could be because
6272 * firmware is using some of this memory, or for some other reasons.
6273 * Once memblock is changed so such behaviour is not allowed: i.e.
6274 * list of "reserved" memory must be a subset of list of "memory", then
6275 * this code can be removed.
6278 pr_info("Reserved but unavailable: %lld pages", pgcnt
);
6280 #endif /* CONFIG_HAVE_MEMBLOCK && !CONFIG_FLAT_NODE_MEM_MAP */
6282 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6284 #if MAX_NUMNODES > 1
6286 * Figure out the number of possible node ids.
6288 void __init
setup_nr_node_ids(void)
6290 unsigned int highest
;
6292 highest
= find_last_bit(node_possible_map
.bits
, MAX_NUMNODES
);
6293 nr_node_ids
= highest
+ 1;
6298 * node_map_pfn_alignment - determine the maximum internode alignment
6300 * This function should be called after node map is populated and sorted.
6301 * It calculates the maximum power of two alignment which can distinguish
6304 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
6305 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
6306 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
6307 * shifted, 1GiB is enough and this function will indicate so.
6309 * This is used to test whether pfn -> nid mapping of the chosen memory
6310 * model has fine enough granularity to avoid incorrect mapping for the
6311 * populated node map.
6313 * Returns the determined alignment in pfn's. 0 if there is no alignment
6314 * requirement (single node).
6316 unsigned long __init
node_map_pfn_alignment(void)
6318 unsigned long accl_mask
= 0, last_end
= 0;
6319 unsigned long start
, end
, mask
;
6323 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start
, &end
, &nid
) {
6324 if (!start
|| last_nid
< 0 || last_nid
== nid
) {
6331 * Start with a mask granular enough to pin-point to the
6332 * start pfn and tick off bits one-by-one until it becomes
6333 * too coarse to separate the current node from the last.
6335 mask
= ~((1 << __ffs(start
)) - 1);
6336 while (mask
&& last_end
<= (start
& (mask
<< 1)))
6339 /* accumulate all internode masks */
6343 /* convert mask to number of pages */
6344 return ~accl_mask
+ 1;
6347 /* Find the lowest pfn for a node */
6348 static unsigned long __init
find_min_pfn_for_node(int nid
)
6350 unsigned long min_pfn
= ULONG_MAX
;
6351 unsigned long start_pfn
;
6354 for_each_mem_pfn_range(i
, nid
, &start_pfn
, NULL
, NULL
)
6355 min_pfn
= min(min_pfn
, start_pfn
);
6357 if (min_pfn
== ULONG_MAX
) {
6358 pr_warn("Could not find start_pfn for node %d\n", nid
);
6366 * find_min_pfn_with_active_regions - Find the minimum PFN registered
6368 * It returns the minimum PFN based on information provided via
6369 * memblock_set_node().
6371 unsigned long __init
find_min_pfn_with_active_regions(void)
6373 return find_min_pfn_for_node(MAX_NUMNODES
);
6377 * early_calculate_totalpages()
6378 * Sum pages in active regions for movable zone.
6379 * Populate N_MEMORY for calculating usable_nodes.
6381 static unsigned long __init
early_calculate_totalpages(void)
6383 unsigned long totalpages
= 0;
6384 unsigned long start_pfn
, end_pfn
;
6387 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
6388 unsigned long pages
= end_pfn
- start_pfn
;
6390 totalpages
+= pages
;
6392 node_set_state(nid
, N_MEMORY
);
6398 * Find the PFN the Movable zone begins in each node. Kernel memory
6399 * is spread evenly between nodes as long as the nodes have enough
6400 * memory. When they don't, some nodes will have more kernelcore than
6403 static void __init
find_zone_movable_pfns_for_nodes(void)
6406 unsigned long usable_startpfn
;
6407 unsigned long kernelcore_node
, kernelcore_remaining
;
6408 /* save the state before borrow the nodemask */
6409 nodemask_t saved_node_state
= node_states
[N_MEMORY
];
6410 unsigned long totalpages
= early_calculate_totalpages();
6411 int usable_nodes
= nodes_weight(node_states
[N_MEMORY
]);
6412 struct memblock_region
*r
;
6414 /* Need to find movable_zone earlier when movable_node is specified. */
6415 find_usable_zone_for_movable();
6418 * If movable_node is specified, ignore kernelcore and movablecore
6421 if (movable_node_is_enabled()) {
6422 for_each_memblock(memory
, r
) {
6423 if (!memblock_is_hotpluggable(r
))
6428 usable_startpfn
= PFN_DOWN(r
->base
);
6429 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
6430 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
6438 * If kernelcore=mirror is specified, ignore movablecore option
6440 if (mirrored_kernelcore
) {
6441 bool mem_below_4gb_not_mirrored
= false;
6443 for_each_memblock(memory
, r
) {
6444 if (memblock_is_mirror(r
))
6449 usable_startpfn
= memblock_region_memory_base_pfn(r
);
6451 if (usable_startpfn
< 0x100000) {
6452 mem_below_4gb_not_mirrored
= true;
6456 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
6457 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
6461 if (mem_below_4gb_not_mirrored
)
6462 pr_warn("This configuration results in unmirrored kernel memory.");
6468 * If movablecore=nn[KMG] was specified, calculate what size of
6469 * kernelcore that corresponds so that memory usable for
6470 * any allocation type is evenly spread. If both kernelcore
6471 * and movablecore are specified, then the value of kernelcore
6472 * will be used for required_kernelcore if it's greater than
6473 * what movablecore would have allowed.
6475 if (required_movablecore
) {
6476 unsigned long corepages
;
6479 * Round-up so that ZONE_MOVABLE is at least as large as what
6480 * was requested by the user
6482 required_movablecore
=
6483 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
6484 required_movablecore
= min(totalpages
, required_movablecore
);
6485 corepages
= totalpages
- required_movablecore
;
6487 required_kernelcore
= max(required_kernelcore
, corepages
);
6491 * If kernelcore was not specified or kernelcore size is larger
6492 * than totalpages, there is no ZONE_MOVABLE.
6494 if (!required_kernelcore
|| required_kernelcore
>= totalpages
)
6497 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
6498 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
6501 /* Spread kernelcore memory as evenly as possible throughout nodes */
6502 kernelcore_node
= required_kernelcore
/ usable_nodes
;
6503 for_each_node_state(nid
, N_MEMORY
) {
6504 unsigned long start_pfn
, end_pfn
;
6507 * Recalculate kernelcore_node if the division per node
6508 * now exceeds what is necessary to satisfy the requested
6509 * amount of memory for the kernel
6511 if (required_kernelcore
< kernelcore_node
)
6512 kernelcore_node
= required_kernelcore
/ usable_nodes
;
6515 * As the map is walked, we track how much memory is usable
6516 * by the kernel using kernelcore_remaining. When it is
6517 * 0, the rest of the node is usable by ZONE_MOVABLE
6519 kernelcore_remaining
= kernelcore_node
;
6521 /* Go through each range of PFNs within this node */
6522 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
6523 unsigned long size_pages
;
6525 start_pfn
= max(start_pfn
, zone_movable_pfn
[nid
]);
6526 if (start_pfn
>= end_pfn
)
6529 /* Account for what is only usable for kernelcore */
6530 if (start_pfn
< usable_startpfn
) {
6531 unsigned long kernel_pages
;
6532 kernel_pages
= min(end_pfn
, usable_startpfn
)
6535 kernelcore_remaining
-= min(kernel_pages
,
6536 kernelcore_remaining
);
6537 required_kernelcore
-= min(kernel_pages
,
6538 required_kernelcore
);
6540 /* Continue if range is now fully accounted */
6541 if (end_pfn
<= usable_startpfn
) {
6544 * Push zone_movable_pfn to the end so
6545 * that if we have to rebalance
6546 * kernelcore across nodes, we will
6547 * not double account here
6549 zone_movable_pfn
[nid
] = end_pfn
;
6552 start_pfn
= usable_startpfn
;
6556 * The usable PFN range for ZONE_MOVABLE is from
6557 * start_pfn->end_pfn. Calculate size_pages as the
6558 * number of pages used as kernelcore
6560 size_pages
= end_pfn
- start_pfn
;
6561 if (size_pages
> kernelcore_remaining
)
6562 size_pages
= kernelcore_remaining
;
6563 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
6566 * Some kernelcore has been met, update counts and
6567 * break if the kernelcore for this node has been
6570 required_kernelcore
-= min(required_kernelcore
,
6572 kernelcore_remaining
-= size_pages
;
6573 if (!kernelcore_remaining
)
6579 * If there is still required_kernelcore, we do another pass with one
6580 * less node in the count. This will push zone_movable_pfn[nid] further
6581 * along on the nodes that still have memory until kernelcore is
6585 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
6589 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
6590 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
6591 zone_movable_pfn
[nid
] =
6592 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
6595 /* restore the node_state */
6596 node_states
[N_MEMORY
] = saved_node_state
;
6599 /* Any regular or high memory on that node ? */
6600 static void check_for_memory(pg_data_t
*pgdat
, int nid
)
6602 enum zone_type zone_type
;
6604 if (N_MEMORY
== N_NORMAL_MEMORY
)
6607 for (zone_type
= 0; zone_type
<= ZONE_MOVABLE
- 1; zone_type
++) {
6608 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
6609 if (populated_zone(zone
)) {
6610 node_set_state(nid
, N_HIGH_MEMORY
);
6611 if (N_NORMAL_MEMORY
!= N_HIGH_MEMORY
&&
6612 zone_type
<= ZONE_NORMAL
)
6613 node_set_state(nid
, N_NORMAL_MEMORY
);
6620 * free_area_init_nodes - Initialise all pg_data_t and zone data
6621 * @max_zone_pfn: an array of max PFNs for each zone
6623 * This will call free_area_init_node() for each active node in the system.
6624 * Using the page ranges provided by memblock_set_node(), the size of each
6625 * zone in each node and their holes is calculated. If the maximum PFN
6626 * between two adjacent zones match, it is assumed that the zone is empty.
6627 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
6628 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
6629 * starts where the previous one ended. For example, ZONE_DMA32 starts
6630 * at arch_max_dma_pfn.
6632 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
6634 unsigned long start_pfn
, end_pfn
;
6637 /* Record where the zone boundaries are */
6638 memset(arch_zone_lowest_possible_pfn
, 0,
6639 sizeof(arch_zone_lowest_possible_pfn
));
6640 memset(arch_zone_highest_possible_pfn
, 0,
6641 sizeof(arch_zone_highest_possible_pfn
));
6643 start_pfn
= find_min_pfn_with_active_regions();
6645 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
6646 if (i
== ZONE_MOVABLE
)
6649 end_pfn
= max(max_zone_pfn
[i
], start_pfn
);
6650 arch_zone_lowest_possible_pfn
[i
] = start_pfn
;
6651 arch_zone_highest_possible_pfn
[i
] = end_pfn
;
6653 start_pfn
= end_pfn
;
6656 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
6657 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
6658 find_zone_movable_pfns_for_nodes();
6660 /* Print out the zone ranges */
6661 pr_info("Zone ranges:\n");
6662 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
6663 if (i
== ZONE_MOVABLE
)
6665 pr_info(" %-8s ", zone_names
[i
]);
6666 if (arch_zone_lowest_possible_pfn
[i
] ==
6667 arch_zone_highest_possible_pfn
[i
])
6670 pr_cont("[mem %#018Lx-%#018Lx]\n",
6671 (u64
)arch_zone_lowest_possible_pfn
[i
]
6673 ((u64
)arch_zone_highest_possible_pfn
[i
]
6674 << PAGE_SHIFT
) - 1);
6677 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
6678 pr_info("Movable zone start for each node\n");
6679 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6680 if (zone_movable_pfn
[i
])
6681 pr_info(" Node %d: %#018Lx\n", i
,
6682 (u64
)zone_movable_pfn
[i
] << PAGE_SHIFT
);
6685 /* Print out the early node map */
6686 pr_info("Early memory node ranges\n");
6687 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
)
6688 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid
,
6689 (u64
)start_pfn
<< PAGE_SHIFT
,
6690 ((u64
)end_pfn
<< PAGE_SHIFT
) - 1);
6692 /* Initialise every node */
6693 mminit_verify_pageflags_layout();
6694 setup_nr_node_ids();
6695 zero_resv_unavail();
6696 for_each_online_node(nid
) {
6697 pg_data_t
*pgdat
= NODE_DATA(nid
);
6698 free_area_init_node(nid
, NULL
,
6699 find_min_pfn_for_node(nid
), NULL
);
6701 /* Any memory on that node */
6702 if (pgdat
->node_present_pages
)
6703 node_set_state(nid
, N_MEMORY
);
6704 check_for_memory(pgdat
, nid
);
6708 static int __init
cmdline_parse_core(char *p
, unsigned long *core
)
6710 unsigned long long coremem
;
6714 coremem
= memparse(p
, &p
);
6715 *core
= coremem
>> PAGE_SHIFT
;
6717 /* Paranoid check that UL is enough for the coremem value */
6718 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
6724 * kernelcore=size sets the amount of memory for use for allocations that
6725 * cannot be reclaimed or migrated.
6727 static int __init
cmdline_parse_kernelcore(char *p
)
6729 /* parse kernelcore=mirror */
6730 if (parse_option_str(p
, "mirror")) {
6731 mirrored_kernelcore
= true;
6735 return cmdline_parse_core(p
, &required_kernelcore
);
6739 * movablecore=size sets the amount of memory for use for allocations that
6740 * can be reclaimed or migrated.
6742 static int __init
cmdline_parse_movablecore(char *p
)
6744 return cmdline_parse_core(p
, &required_movablecore
);
6747 early_param("kernelcore", cmdline_parse_kernelcore
);
6748 early_param("movablecore", cmdline_parse_movablecore
);
6750 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6752 void adjust_managed_page_count(struct page
*page
, long count
)
6754 spin_lock(&managed_page_count_lock
);
6755 page_zone(page
)->managed_pages
+= count
;
6756 totalram_pages
+= count
;
6757 #ifdef CONFIG_HIGHMEM
6758 if (PageHighMem(page
))
6759 totalhigh_pages
+= count
;
6761 spin_unlock(&managed_page_count_lock
);
6763 EXPORT_SYMBOL(adjust_managed_page_count
);
6765 unsigned long free_reserved_area(void *start
, void *end
, int poison
, char *s
)
6768 unsigned long pages
= 0;
6770 start
= (void *)PAGE_ALIGN((unsigned long)start
);
6771 end
= (void *)((unsigned long)end
& PAGE_MASK
);
6772 for (pos
= start
; pos
< end
; pos
+= PAGE_SIZE
, pages
++) {
6773 if ((unsigned int)poison
<= 0xFF)
6774 memset(pos
, poison
, PAGE_SIZE
);
6775 free_reserved_page(virt_to_page(pos
));
6779 pr_info("Freeing %s memory: %ldK\n",
6780 s
, pages
<< (PAGE_SHIFT
- 10));
6784 EXPORT_SYMBOL(free_reserved_area
);
6786 #ifdef CONFIG_HIGHMEM
6787 void free_highmem_page(struct page
*page
)
6789 __free_reserved_page(page
);
6791 page_zone(page
)->managed_pages
++;
6797 void __init
mem_init_print_info(const char *str
)
6799 unsigned long physpages
, codesize
, datasize
, rosize
, bss_size
;
6800 unsigned long init_code_size
, init_data_size
;
6802 physpages
= get_num_physpages();
6803 codesize
= _etext
- _stext
;
6804 datasize
= _edata
- _sdata
;
6805 rosize
= __end_rodata
- __start_rodata
;
6806 bss_size
= __bss_stop
- __bss_start
;
6807 init_data_size
= __init_end
- __init_begin
;
6808 init_code_size
= _einittext
- _sinittext
;
6811 * Detect special cases and adjust section sizes accordingly:
6812 * 1) .init.* may be embedded into .data sections
6813 * 2) .init.text.* may be out of [__init_begin, __init_end],
6814 * please refer to arch/tile/kernel/vmlinux.lds.S.
6815 * 3) .rodata.* may be embedded into .text or .data sections.
6817 #define adj_init_size(start, end, size, pos, adj) \
6819 if (start <= pos && pos < end && size > adj) \
6823 adj_init_size(__init_begin
, __init_end
, init_data_size
,
6824 _sinittext
, init_code_size
);
6825 adj_init_size(_stext
, _etext
, codesize
, _sinittext
, init_code_size
);
6826 adj_init_size(_sdata
, _edata
, datasize
, __init_begin
, init_data_size
);
6827 adj_init_size(_stext
, _etext
, codesize
, __start_rodata
, rosize
);
6828 adj_init_size(_sdata
, _edata
, datasize
, __start_rodata
, rosize
);
6830 #undef adj_init_size
6832 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
6833 #ifdef CONFIG_HIGHMEM
6837 nr_free_pages() << (PAGE_SHIFT
- 10),
6838 physpages
<< (PAGE_SHIFT
- 10),
6839 codesize
>> 10, datasize
>> 10, rosize
>> 10,
6840 (init_data_size
+ init_code_size
) >> 10, bss_size
>> 10,
6841 (physpages
- totalram_pages
- totalcma_pages
) << (PAGE_SHIFT
- 10),
6842 totalcma_pages
<< (PAGE_SHIFT
- 10),
6843 #ifdef CONFIG_HIGHMEM
6844 totalhigh_pages
<< (PAGE_SHIFT
- 10),
6846 str
? ", " : "", str
? str
: "");
6850 * set_dma_reserve - set the specified number of pages reserved in the first zone
6851 * @new_dma_reserve: The number of pages to mark reserved
6853 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
6854 * In the DMA zone, a significant percentage may be consumed by kernel image
6855 * and other unfreeable allocations which can skew the watermarks badly. This
6856 * function may optionally be used to account for unfreeable pages in the
6857 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
6858 * smaller per-cpu batchsize.
6860 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
6862 dma_reserve
= new_dma_reserve
;
6865 void __init
free_area_init(unsigned long *zones_size
)
6867 zero_resv_unavail();
6868 free_area_init_node(0, zones_size
,
6869 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
6872 static int page_alloc_cpu_dead(unsigned int cpu
)
6875 lru_add_drain_cpu(cpu
);
6879 * Spill the event counters of the dead processor
6880 * into the current processors event counters.
6881 * This artificially elevates the count of the current
6884 vm_events_fold_cpu(cpu
);
6887 * Zero the differential counters of the dead processor
6888 * so that the vm statistics are consistent.
6890 * This is only okay since the processor is dead and cannot
6891 * race with what we are doing.
6893 cpu_vm_stats_fold(cpu
);
6897 void __init
page_alloc_init(void)
6901 ret
= cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD
,
6902 "mm/page_alloc:dead", NULL
,
6903 page_alloc_cpu_dead
);
6908 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6909 * or min_free_kbytes changes.
6911 static void calculate_totalreserve_pages(void)
6913 struct pglist_data
*pgdat
;
6914 unsigned long reserve_pages
= 0;
6915 enum zone_type i
, j
;
6917 for_each_online_pgdat(pgdat
) {
6919 pgdat
->totalreserve_pages
= 0;
6921 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
6922 struct zone
*zone
= pgdat
->node_zones
+ i
;
6925 /* Find valid and maximum lowmem_reserve in the zone */
6926 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
6927 if (zone
->lowmem_reserve
[j
] > max
)
6928 max
= zone
->lowmem_reserve
[j
];
6931 /* we treat the high watermark as reserved pages. */
6932 max
+= high_wmark_pages(zone
);
6934 if (max
> zone
->managed_pages
)
6935 max
= zone
->managed_pages
;
6937 pgdat
->totalreserve_pages
+= max
;
6939 reserve_pages
+= max
;
6942 totalreserve_pages
= reserve_pages
;
6946 * setup_per_zone_lowmem_reserve - called whenever
6947 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6948 * has a correct pages reserved value, so an adequate number of
6949 * pages are left in the zone after a successful __alloc_pages().
6951 static void setup_per_zone_lowmem_reserve(void)
6953 struct pglist_data
*pgdat
;
6954 enum zone_type j
, idx
;
6956 for_each_online_pgdat(pgdat
) {
6957 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
6958 struct zone
*zone
= pgdat
->node_zones
+ j
;
6959 unsigned long managed_pages
= zone
->managed_pages
;
6961 zone
->lowmem_reserve
[j
] = 0;
6965 struct zone
*lower_zone
;
6969 if (sysctl_lowmem_reserve_ratio
[idx
] < 1)
6970 sysctl_lowmem_reserve_ratio
[idx
] = 1;
6972 lower_zone
= pgdat
->node_zones
+ idx
;
6973 lower_zone
->lowmem_reserve
[j
] = managed_pages
/
6974 sysctl_lowmem_reserve_ratio
[idx
];
6975 managed_pages
+= lower_zone
->managed_pages
;
6980 /* update totalreserve_pages */
6981 calculate_totalreserve_pages();
6984 static void __setup_per_zone_wmarks(void)
6986 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
6987 unsigned long lowmem_pages
= 0;
6989 unsigned long flags
;
6991 /* Calculate total number of !ZONE_HIGHMEM pages */
6992 for_each_zone(zone
) {
6993 if (!is_highmem(zone
))
6994 lowmem_pages
+= zone
->managed_pages
;
6997 for_each_zone(zone
) {
7000 spin_lock_irqsave(&zone
->lock
, flags
);
7001 tmp
= (u64
)pages_min
* zone
->managed_pages
;
7002 do_div(tmp
, lowmem_pages
);
7003 if (is_highmem(zone
)) {
7005 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7006 * need highmem pages, so cap pages_min to a small
7009 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7010 * deltas control asynch page reclaim, and so should
7011 * not be capped for highmem.
7013 unsigned long min_pages
;
7015 min_pages
= zone
->managed_pages
/ 1024;
7016 min_pages
= clamp(min_pages
, SWAP_CLUSTER_MAX
, 128UL);
7017 zone
->watermark
[WMARK_MIN
] = min_pages
;
7020 * If it's a lowmem zone, reserve a number of pages
7021 * proportionate to the zone's size.
7023 zone
->watermark
[WMARK_MIN
] = tmp
;
7027 * Set the kswapd watermarks distance according to the
7028 * scale factor in proportion to available memory, but
7029 * ensure a minimum size on small systems.
7031 tmp
= max_t(u64
, tmp
>> 2,
7032 mult_frac(zone
->managed_pages
,
7033 watermark_scale_factor
, 10000));
7035 zone
->watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + tmp
;
7036 zone
->watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + tmp
* 2;
7038 spin_unlock_irqrestore(&zone
->lock
, flags
);
7041 /* update totalreserve_pages */
7042 calculate_totalreserve_pages();
7046 * setup_per_zone_wmarks - called when min_free_kbytes changes
7047 * or when memory is hot-{added|removed}
7049 * Ensures that the watermark[min,low,high] values for each zone are set
7050 * correctly with respect to min_free_kbytes.
7052 void setup_per_zone_wmarks(void)
7054 static DEFINE_SPINLOCK(lock
);
7057 __setup_per_zone_wmarks();
7062 * Initialise min_free_kbytes.
7064 * For small machines we want it small (128k min). For large machines
7065 * we want it large (64MB max). But it is not linear, because network
7066 * bandwidth does not increase linearly with machine size. We use
7068 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7069 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7085 int __meminit
init_per_zone_wmark_min(void)
7087 unsigned long lowmem_kbytes
;
7088 int new_min_free_kbytes
;
7090 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
7091 new_min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
7093 if (new_min_free_kbytes
> user_min_free_kbytes
) {
7094 min_free_kbytes
= new_min_free_kbytes
;
7095 if (min_free_kbytes
< 128)
7096 min_free_kbytes
= 128;
7097 if (min_free_kbytes
> 65536)
7098 min_free_kbytes
= 65536;
7100 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7101 new_min_free_kbytes
, user_min_free_kbytes
);
7103 setup_per_zone_wmarks();
7104 refresh_zone_stat_thresholds();
7105 setup_per_zone_lowmem_reserve();
7108 setup_min_unmapped_ratio();
7109 setup_min_slab_ratio();
7114 core_initcall(init_per_zone_wmark_min
)
7117 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7118 * that we can call two helper functions whenever min_free_kbytes
7121 int min_free_kbytes_sysctl_handler(struct ctl_table
*table
, int write
,
7122 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7126 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7131 user_min_free_kbytes
= min_free_kbytes
;
7132 setup_per_zone_wmarks();
7137 int watermark_scale_factor_sysctl_handler(struct ctl_table
*table
, int write
,
7138 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7142 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7147 setup_per_zone_wmarks();
7153 static void setup_min_unmapped_ratio(void)
7158 for_each_online_pgdat(pgdat
)
7159 pgdat
->min_unmapped_pages
= 0;
7162 zone
->zone_pgdat
->min_unmapped_pages
+= (zone
->managed_pages
*
7163 sysctl_min_unmapped_ratio
) / 100;
7167 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
7168 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7172 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7176 setup_min_unmapped_ratio();
7181 static void setup_min_slab_ratio(void)
7186 for_each_online_pgdat(pgdat
)
7187 pgdat
->min_slab_pages
= 0;
7190 zone
->zone_pgdat
->min_slab_pages
+= (zone
->managed_pages
*
7191 sysctl_min_slab_ratio
) / 100;
7194 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
7195 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7199 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7203 setup_min_slab_ratio();
7210 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
7211 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
7212 * whenever sysctl_lowmem_reserve_ratio changes.
7214 * The reserve ratio obviously has absolutely no relation with the
7215 * minimum watermarks. The lowmem reserve ratio can only make sense
7216 * if in function of the boot time zone sizes.
7218 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
7219 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7221 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7222 setup_per_zone_lowmem_reserve();
7227 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
7228 * cpu. It is the fraction of total pages in each zone that a hot per cpu
7229 * pagelist can have before it gets flushed back to buddy allocator.
7231 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table
*table
, int write
,
7232 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7235 int old_percpu_pagelist_fraction
;
7238 mutex_lock(&pcp_batch_high_lock
);
7239 old_percpu_pagelist_fraction
= percpu_pagelist_fraction
;
7241 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7242 if (!write
|| ret
< 0)
7245 /* Sanity checking to avoid pcp imbalance */
7246 if (percpu_pagelist_fraction
&&
7247 percpu_pagelist_fraction
< MIN_PERCPU_PAGELIST_FRACTION
) {
7248 percpu_pagelist_fraction
= old_percpu_pagelist_fraction
;
7254 if (percpu_pagelist_fraction
== old_percpu_pagelist_fraction
)
7257 for_each_populated_zone(zone
) {
7260 for_each_possible_cpu(cpu
)
7261 pageset_set_high_and_batch(zone
,
7262 per_cpu_ptr(zone
->pageset
, cpu
));
7265 mutex_unlock(&pcp_batch_high_lock
);
7270 int hashdist
= HASHDIST_DEFAULT
;
7272 static int __init
set_hashdist(char *str
)
7276 hashdist
= simple_strtoul(str
, &str
, 0);
7279 __setup("hashdist=", set_hashdist
);
7282 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
7284 * Returns the number of pages that arch has reserved but
7285 * is not known to alloc_large_system_hash().
7287 static unsigned long __init
arch_reserved_kernel_pages(void)
7294 * Adaptive scale is meant to reduce sizes of hash tables on large memory
7295 * machines. As memory size is increased the scale is also increased but at
7296 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
7297 * quadruples the scale is increased by one, which means the size of hash table
7298 * only doubles, instead of quadrupling as well.
7299 * Because 32-bit systems cannot have large physical memory, where this scaling
7300 * makes sense, it is disabled on such platforms.
7302 #if __BITS_PER_LONG > 32
7303 #define ADAPT_SCALE_BASE (64ul << 30)
7304 #define ADAPT_SCALE_SHIFT 2
7305 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
7309 * allocate a large system hash table from bootmem
7310 * - it is assumed that the hash table must contain an exact power-of-2
7311 * quantity of entries
7312 * - limit is the number of hash buckets, not the total allocation size
7314 void *__init
alloc_large_system_hash(const char *tablename
,
7315 unsigned long bucketsize
,
7316 unsigned long numentries
,
7319 unsigned int *_hash_shift
,
7320 unsigned int *_hash_mask
,
7321 unsigned long low_limit
,
7322 unsigned long high_limit
)
7324 unsigned long long max
= high_limit
;
7325 unsigned long log2qty
, size
;
7329 /* allow the kernel cmdline to have a say */
7331 /* round applicable memory size up to nearest megabyte */
7332 numentries
= nr_kernel_pages
;
7333 numentries
-= arch_reserved_kernel_pages();
7335 /* It isn't necessary when PAGE_SIZE >= 1MB */
7336 if (PAGE_SHIFT
< 20)
7337 numentries
= round_up(numentries
, (1<<20)/PAGE_SIZE
);
7339 #if __BITS_PER_LONG > 32
7341 unsigned long adapt
;
7343 for (adapt
= ADAPT_SCALE_NPAGES
; adapt
< numentries
;
7344 adapt
<<= ADAPT_SCALE_SHIFT
)
7349 /* limit to 1 bucket per 2^scale bytes of low memory */
7350 if (scale
> PAGE_SHIFT
)
7351 numentries
>>= (scale
- PAGE_SHIFT
);
7353 numentries
<<= (PAGE_SHIFT
- scale
);
7355 /* Make sure we've got at least a 0-order allocation.. */
7356 if (unlikely(flags
& HASH_SMALL
)) {
7357 /* Makes no sense without HASH_EARLY */
7358 WARN_ON(!(flags
& HASH_EARLY
));
7359 if (!(numentries
>> *_hash_shift
)) {
7360 numentries
= 1UL << *_hash_shift
;
7361 BUG_ON(!numentries
);
7363 } else if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
7364 numentries
= PAGE_SIZE
/ bucketsize
;
7366 numentries
= roundup_pow_of_two(numentries
);
7368 /* limit allocation size to 1/16 total memory by default */
7370 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
7371 do_div(max
, bucketsize
);
7373 max
= min(max
, 0x80000000ULL
);
7375 if (numentries
< low_limit
)
7376 numentries
= low_limit
;
7377 if (numentries
> max
)
7380 log2qty
= ilog2(numentries
);
7382 gfp_flags
= (flags
& HASH_ZERO
) ? GFP_ATOMIC
| __GFP_ZERO
: GFP_ATOMIC
;
7384 size
= bucketsize
<< log2qty
;
7385 if (flags
& HASH_EARLY
) {
7386 if (flags
& HASH_ZERO
)
7387 table
= memblock_virt_alloc_nopanic(size
, 0);
7389 table
= memblock_virt_alloc_raw(size
, 0);
7390 } else if (hashdist
) {
7391 table
= __vmalloc(size
, gfp_flags
, PAGE_KERNEL
);
7394 * If bucketsize is not a power-of-two, we may free
7395 * some pages at the end of hash table which
7396 * alloc_pages_exact() automatically does
7398 if (get_order(size
) < MAX_ORDER
) {
7399 table
= alloc_pages_exact(size
, gfp_flags
);
7400 kmemleak_alloc(table
, size
, 1, gfp_flags
);
7403 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
7406 panic("Failed to allocate %s hash table\n", tablename
);
7408 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
7409 tablename
, 1UL << log2qty
, ilog2(size
) - PAGE_SHIFT
, size
);
7412 *_hash_shift
= log2qty
;
7414 *_hash_mask
= (1 << log2qty
) - 1;
7420 * This function checks whether pageblock includes unmovable pages or not.
7421 * If @count is not zero, it is okay to include less @count unmovable pages
7423 * PageLRU check without isolation or lru_lock could race so that
7424 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
7425 * check without lock_page also may miss some movable non-lru pages at
7426 * race condition. So you can't expect this function should be exact.
7428 bool has_unmovable_pages(struct zone
*zone
, struct page
*page
, int count
,
7430 bool skip_hwpoisoned_pages
)
7432 unsigned long pfn
, iter
, found
;
7435 * For avoiding noise data, lru_add_drain_all() should be called
7436 * If ZONE_MOVABLE, the zone never contains unmovable pages
7438 if (zone_idx(zone
) == ZONE_MOVABLE
)
7442 * CMA allocations (alloc_contig_range) really need to mark isolate
7443 * CMA pageblocks even when they are not movable in fact so consider
7444 * them movable here.
7446 if (is_migrate_cma(migratetype
) &&
7447 is_migrate_cma(get_pageblock_migratetype(page
)))
7450 pfn
= page_to_pfn(page
);
7451 for (found
= 0, iter
= 0; iter
< pageblock_nr_pages
; iter
++) {
7452 unsigned long check
= pfn
+ iter
;
7454 if (!pfn_valid_within(check
))
7457 page
= pfn_to_page(check
);
7459 if (PageReserved(page
))
7463 * Hugepages are not in LRU lists, but they're movable.
7464 * We need not scan over tail pages bacause we don't
7465 * handle each tail page individually in migration.
7467 if (PageHuge(page
)) {
7468 iter
= round_up(iter
+ 1, 1<<compound_order(page
)) - 1;
7473 * We can't use page_count without pin a page
7474 * because another CPU can free compound page.
7475 * This check already skips compound tails of THP
7476 * because their page->_refcount is zero at all time.
7478 if (!page_ref_count(page
)) {
7479 if (PageBuddy(page
))
7480 iter
+= (1 << page_order(page
)) - 1;
7485 * The HWPoisoned page may be not in buddy system, and
7486 * page_count() is not 0.
7488 if (skip_hwpoisoned_pages
&& PageHWPoison(page
))
7491 if (__PageMovable(page
))
7497 * If there are RECLAIMABLE pages, we need to check
7498 * it. But now, memory offline itself doesn't call
7499 * shrink_node_slabs() and it still to be fixed.
7502 * If the page is not RAM, page_count()should be 0.
7503 * we don't need more check. This is an _used_ not-movable page.
7505 * The problematic thing here is PG_reserved pages. PG_reserved
7506 * is set to both of a memory hole page and a _used_ kernel
7515 bool is_pageblock_removable_nolock(struct page
*page
)
7521 * We have to be careful here because we are iterating over memory
7522 * sections which are not zone aware so we might end up outside of
7523 * the zone but still within the section.
7524 * We have to take care about the node as well. If the node is offline
7525 * its NODE_DATA will be NULL - see page_zone.
7527 if (!node_online(page_to_nid(page
)))
7530 zone
= page_zone(page
);
7531 pfn
= page_to_pfn(page
);
7532 if (!zone_spans_pfn(zone
, pfn
))
7535 return !has_unmovable_pages(zone
, page
, 0, MIGRATE_MOVABLE
, true);
7538 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
7540 static unsigned long pfn_max_align_down(unsigned long pfn
)
7542 return pfn
& ~(max_t(unsigned long, MAX_ORDER_NR_PAGES
,
7543 pageblock_nr_pages
) - 1);
7546 static unsigned long pfn_max_align_up(unsigned long pfn
)
7548 return ALIGN(pfn
, max_t(unsigned long, MAX_ORDER_NR_PAGES
,
7549 pageblock_nr_pages
));
7552 /* [start, end) must belong to a single zone. */
7553 static int __alloc_contig_migrate_range(struct compact_control
*cc
,
7554 unsigned long start
, unsigned long end
)
7556 /* This function is based on compact_zone() from compaction.c. */
7557 unsigned long nr_reclaimed
;
7558 unsigned long pfn
= start
;
7559 unsigned int tries
= 0;
7564 while (pfn
< end
|| !list_empty(&cc
->migratepages
)) {
7565 if (fatal_signal_pending(current
)) {
7570 if (list_empty(&cc
->migratepages
)) {
7571 cc
->nr_migratepages
= 0;
7572 pfn
= isolate_migratepages_range(cc
, pfn
, end
);
7578 } else if (++tries
== 5) {
7579 ret
= ret
< 0 ? ret
: -EBUSY
;
7583 nr_reclaimed
= reclaim_clean_pages_from_list(cc
->zone
,
7585 cc
->nr_migratepages
-= nr_reclaimed
;
7587 ret
= migrate_pages(&cc
->migratepages
, alloc_migrate_target
,
7588 NULL
, 0, cc
->mode
, MR_CMA
);
7591 putback_movable_pages(&cc
->migratepages
);
7598 * alloc_contig_range() -- tries to allocate given range of pages
7599 * @start: start PFN to allocate
7600 * @end: one-past-the-last PFN to allocate
7601 * @migratetype: migratetype of the underlaying pageblocks (either
7602 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
7603 * in range must have the same migratetype and it must
7604 * be either of the two.
7605 * @gfp_mask: GFP mask to use during compaction
7607 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
7608 * aligned, however it's the caller's responsibility to guarantee that
7609 * we are the only thread that changes migrate type of pageblocks the
7612 * The PFN range must belong to a single zone.
7614 * Returns zero on success or negative error code. On success all
7615 * pages which PFN is in [start, end) are allocated for the caller and
7616 * need to be freed with free_contig_range().
7618 int alloc_contig_range(unsigned long start
, unsigned long end
,
7619 unsigned migratetype
, gfp_t gfp_mask
)
7621 unsigned long outer_start
, outer_end
;
7625 struct compact_control cc
= {
7626 .nr_migratepages
= 0,
7628 .zone
= page_zone(pfn_to_page(start
)),
7629 .mode
= MIGRATE_SYNC
,
7630 .ignore_skip_hint
= true,
7631 .no_set_skip_hint
= true,
7632 .gfp_mask
= current_gfp_context(gfp_mask
),
7634 INIT_LIST_HEAD(&cc
.migratepages
);
7637 * What we do here is we mark all pageblocks in range as
7638 * MIGRATE_ISOLATE. Because pageblock and max order pages may
7639 * have different sizes, and due to the way page allocator
7640 * work, we align the range to biggest of the two pages so
7641 * that page allocator won't try to merge buddies from
7642 * different pageblocks and change MIGRATE_ISOLATE to some
7643 * other migration type.
7645 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
7646 * migrate the pages from an unaligned range (ie. pages that
7647 * we are interested in). This will put all the pages in
7648 * range back to page allocator as MIGRATE_ISOLATE.
7650 * When this is done, we take the pages in range from page
7651 * allocator removing them from the buddy system. This way
7652 * page allocator will never consider using them.
7654 * This lets us mark the pageblocks back as
7655 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
7656 * aligned range but not in the unaligned, original range are
7657 * put back to page allocator so that buddy can use them.
7660 ret
= start_isolate_page_range(pfn_max_align_down(start
),
7661 pfn_max_align_up(end
), migratetype
,
7667 * In case of -EBUSY, we'd like to know which page causes problem.
7668 * So, just fall through. test_pages_isolated() has a tracepoint
7669 * which will report the busy page.
7671 * It is possible that busy pages could become available before
7672 * the call to test_pages_isolated, and the range will actually be
7673 * allocated. So, if we fall through be sure to clear ret so that
7674 * -EBUSY is not accidentally used or returned to caller.
7676 ret
= __alloc_contig_migrate_range(&cc
, start
, end
);
7677 if (ret
&& ret
!= -EBUSY
)
7682 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
7683 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
7684 * more, all pages in [start, end) are free in page allocator.
7685 * What we are going to do is to allocate all pages from
7686 * [start, end) (that is remove them from page allocator).
7688 * The only problem is that pages at the beginning and at the
7689 * end of interesting range may be not aligned with pages that
7690 * page allocator holds, ie. they can be part of higher order
7691 * pages. Because of this, we reserve the bigger range and
7692 * once this is done free the pages we are not interested in.
7694 * We don't have to hold zone->lock here because the pages are
7695 * isolated thus they won't get removed from buddy.
7698 lru_add_drain_all();
7699 drain_all_pages(cc
.zone
);
7702 outer_start
= start
;
7703 while (!PageBuddy(pfn_to_page(outer_start
))) {
7704 if (++order
>= MAX_ORDER
) {
7705 outer_start
= start
;
7708 outer_start
&= ~0UL << order
;
7711 if (outer_start
!= start
) {
7712 order
= page_order(pfn_to_page(outer_start
));
7715 * outer_start page could be small order buddy page and
7716 * it doesn't include start page. Adjust outer_start
7717 * in this case to report failed page properly
7718 * on tracepoint in test_pages_isolated()
7720 if (outer_start
+ (1UL << order
) <= start
)
7721 outer_start
= start
;
7724 /* Make sure the range is really isolated. */
7725 if (test_pages_isolated(outer_start
, end
, false)) {
7726 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
7727 __func__
, outer_start
, end
);
7732 /* Grab isolated pages from freelists. */
7733 outer_end
= isolate_freepages_range(&cc
, outer_start
, end
);
7739 /* Free head and tail (if any) */
7740 if (start
!= outer_start
)
7741 free_contig_range(outer_start
, start
- outer_start
);
7742 if (end
!= outer_end
)
7743 free_contig_range(end
, outer_end
- end
);
7746 undo_isolate_page_range(pfn_max_align_down(start
),
7747 pfn_max_align_up(end
), migratetype
);
7751 void free_contig_range(unsigned long pfn
, unsigned nr_pages
)
7753 unsigned int count
= 0;
7755 for (; nr_pages
--; pfn
++) {
7756 struct page
*page
= pfn_to_page(pfn
);
7758 count
+= page_count(page
) != 1;
7761 WARN(count
!= 0, "%d pages are still in use!\n", count
);
7765 #ifdef CONFIG_MEMORY_HOTPLUG
7767 * The zone indicated has a new number of managed_pages; batch sizes and percpu
7768 * page high values need to be recalulated.
7770 void __meminit
zone_pcp_update(struct zone
*zone
)
7773 mutex_lock(&pcp_batch_high_lock
);
7774 for_each_possible_cpu(cpu
)
7775 pageset_set_high_and_batch(zone
,
7776 per_cpu_ptr(zone
->pageset
, cpu
));
7777 mutex_unlock(&pcp_batch_high_lock
);
7781 void zone_pcp_reset(struct zone
*zone
)
7783 unsigned long flags
;
7785 struct per_cpu_pageset
*pset
;
7787 /* avoid races with drain_pages() */
7788 local_irq_save(flags
);
7789 if (zone
->pageset
!= &boot_pageset
) {
7790 for_each_online_cpu(cpu
) {
7791 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
7792 drain_zonestat(zone
, pset
);
7794 free_percpu(zone
->pageset
);
7795 zone
->pageset
= &boot_pageset
;
7797 local_irq_restore(flags
);
7800 #ifdef CONFIG_MEMORY_HOTREMOVE
7802 * All pages in the range must be in a single zone and isolated
7803 * before calling this.
7806 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
7810 unsigned int order
, i
;
7812 unsigned long flags
;
7813 /* find the first valid pfn */
7814 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
7819 offline_mem_sections(pfn
, end_pfn
);
7820 zone
= page_zone(pfn_to_page(pfn
));
7821 spin_lock_irqsave(&zone
->lock
, flags
);
7823 while (pfn
< end_pfn
) {
7824 if (!pfn_valid(pfn
)) {
7828 page
= pfn_to_page(pfn
);
7830 * The HWPoisoned page may be not in buddy system, and
7831 * page_count() is not 0.
7833 if (unlikely(!PageBuddy(page
) && PageHWPoison(page
))) {
7835 SetPageReserved(page
);
7839 BUG_ON(page_count(page
));
7840 BUG_ON(!PageBuddy(page
));
7841 order
= page_order(page
);
7842 #ifdef CONFIG_DEBUG_VM
7843 pr_info("remove from free list %lx %d %lx\n",
7844 pfn
, 1 << order
, end_pfn
);
7846 list_del(&page
->lru
);
7847 rmv_page_order(page
);
7848 zone
->free_area
[order
].nr_free
--;
7849 for (i
= 0; i
< (1 << order
); i
++)
7850 SetPageReserved((page
+i
));
7851 pfn
+= (1 << order
);
7853 spin_unlock_irqrestore(&zone
->lock
, flags
);
7857 bool is_free_buddy_page(struct page
*page
)
7859 struct zone
*zone
= page_zone(page
);
7860 unsigned long pfn
= page_to_pfn(page
);
7861 unsigned long flags
;
7864 spin_lock_irqsave(&zone
->lock
, flags
);
7865 for (order
= 0; order
< MAX_ORDER
; order
++) {
7866 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
7868 if (PageBuddy(page_head
) && page_order(page_head
) >= order
)
7871 spin_unlock_irqrestore(&zone
->lock
, flags
);
7873 return order
< MAX_ORDER
;