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/kmemcheck.h>
28 #include <linux/kasan.h>
29 #include <linux/module.h>
30 #include <linux/suspend.h>
31 #include <linux/pagevec.h>
32 #include <linux/blkdev.h>
33 #include <linux/slab.h>
34 #include <linux/ratelimit.h>
35 #include <linux/oom.h>
36 #include <linux/notifier.h>
37 #include <linux/topology.h>
38 #include <linux/sysctl.h>
39 #include <linux/cpu.h>
40 #include <linux/cpuset.h>
41 #include <linux/memory_hotplug.h>
42 #include <linux/nodemask.h>
43 #include <linux/vmalloc.h>
44 #include <linux/vmstat.h>
45 #include <linux/mempolicy.h>
46 #include <linux/memremap.h>
47 #include <linux/stop_machine.h>
48 #include <linux/sort.h>
49 #include <linux/pfn.h>
50 #include <linux/backing-dev.h>
51 #include <linux/fault-inject.h>
52 #include <linux/page-isolation.h>
53 #include <linux/page_ext.h>
54 #include <linux/debugobjects.h>
55 #include <linux/kmemleak.h>
56 #include <linux/compaction.h>
57 #include <trace/events/kmem.h>
58 #include <linux/prefetch.h>
59 #include <linux/mm_inline.h>
60 #include <linux/migrate.h>
61 #include <linux/page_ext.h>
62 #include <linux/hugetlb.h>
63 #include <linux/sched/rt.h>
64 #include <linux/page_owner.h>
65 #include <linux/kthread.h>
67 #include <asm/sections.h>
68 #include <asm/tlbflush.h>
69 #include <asm/div64.h>
72 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
73 static DEFINE_MUTEX(pcp_batch_high_lock
);
74 #define MIN_PERCPU_PAGELIST_FRACTION (8)
76 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
77 DEFINE_PER_CPU(int, numa_node
);
78 EXPORT_PER_CPU_SYMBOL(numa_node
);
81 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
83 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
84 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
85 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
86 * defined in <linux/topology.h>.
88 DEFINE_PER_CPU(int, _numa_mem_
); /* Kernel "local memory" node */
89 EXPORT_PER_CPU_SYMBOL(_numa_mem_
);
90 int _node_numa_mem_
[MAX_NUMNODES
];
94 * Array of node states.
96 nodemask_t node_states
[NR_NODE_STATES
] __read_mostly
= {
97 [N_POSSIBLE
] = NODE_MASK_ALL
,
98 [N_ONLINE
] = { { [0] = 1UL } },
100 [N_NORMAL_MEMORY
] = { { [0] = 1UL } },
101 #ifdef CONFIG_HIGHMEM
102 [N_HIGH_MEMORY
] = { { [0] = 1UL } },
104 #ifdef CONFIG_MOVABLE_NODE
105 [N_MEMORY
] = { { [0] = 1UL } },
107 [N_CPU
] = { { [0] = 1UL } },
110 EXPORT_SYMBOL(node_states
);
112 /* Protect totalram_pages and zone->managed_pages */
113 static DEFINE_SPINLOCK(managed_page_count_lock
);
115 unsigned long totalram_pages __read_mostly
;
116 unsigned long totalreserve_pages __read_mostly
;
117 unsigned long totalcma_pages __read_mostly
;
119 int percpu_pagelist_fraction
;
120 gfp_t gfp_allowed_mask __read_mostly
= GFP_BOOT_MASK
;
123 * A cached value of the page's pageblock's migratetype, used when the page is
124 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
125 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
126 * Also the migratetype set in the page does not necessarily match the pcplist
127 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
128 * other index - this ensures that it will be put on the correct CMA freelist.
130 static inline int get_pcppage_migratetype(struct page
*page
)
135 static inline void set_pcppage_migratetype(struct page
*page
, int migratetype
)
137 page
->index
= migratetype
;
140 #ifdef CONFIG_PM_SLEEP
142 * The following functions are used by the suspend/hibernate code to temporarily
143 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
144 * while devices are suspended. To avoid races with the suspend/hibernate code,
145 * they should always be called with pm_mutex held (gfp_allowed_mask also should
146 * only be modified with pm_mutex held, unless the suspend/hibernate code is
147 * guaranteed not to run in parallel with that modification).
150 static gfp_t saved_gfp_mask
;
152 void pm_restore_gfp_mask(void)
154 WARN_ON(!mutex_is_locked(&pm_mutex
));
155 if (saved_gfp_mask
) {
156 gfp_allowed_mask
= saved_gfp_mask
;
161 void pm_restrict_gfp_mask(void)
163 WARN_ON(!mutex_is_locked(&pm_mutex
));
164 WARN_ON(saved_gfp_mask
);
165 saved_gfp_mask
= gfp_allowed_mask
;
166 gfp_allowed_mask
&= ~(__GFP_IO
| __GFP_FS
);
169 bool pm_suspended_storage(void)
171 if ((gfp_allowed_mask
& (__GFP_IO
| __GFP_FS
)) == (__GFP_IO
| __GFP_FS
))
175 #endif /* CONFIG_PM_SLEEP */
177 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
178 unsigned int pageblock_order __read_mostly
;
181 static void __free_pages_ok(struct page
*page
, unsigned int order
);
184 * results with 256, 32 in the lowmem_reserve sysctl:
185 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
186 * 1G machine -> (16M dma, 784M normal, 224M high)
187 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
188 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
189 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
191 * TBD: should special case ZONE_DMA32 machines here - in those we normally
192 * don't need any ZONE_NORMAL reservation
194 int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
-1] = {
195 #ifdef CONFIG_ZONE_DMA
198 #ifdef CONFIG_ZONE_DMA32
201 #ifdef CONFIG_HIGHMEM
207 EXPORT_SYMBOL(totalram_pages
);
209 static char * const zone_names
[MAX_NR_ZONES
] = {
210 #ifdef CONFIG_ZONE_DMA
213 #ifdef CONFIG_ZONE_DMA32
217 #ifdef CONFIG_HIGHMEM
221 #ifdef CONFIG_ZONE_DEVICE
226 char * const migratetype_names
[MIGRATE_TYPES
] = {
234 #ifdef CONFIG_MEMORY_ISOLATION
239 compound_page_dtor
* const compound_page_dtors
[] = {
242 #ifdef CONFIG_HUGETLB_PAGE
245 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
250 int min_free_kbytes
= 1024;
251 int user_min_free_kbytes
= -1;
252 int watermark_scale_factor
= 10;
254 static unsigned long __meminitdata nr_kernel_pages
;
255 static unsigned long __meminitdata nr_all_pages
;
256 static unsigned long __meminitdata dma_reserve
;
258 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
259 static unsigned long __meminitdata arch_zone_lowest_possible_pfn
[MAX_NR_ZONES
];
260 static unsigned long __meminitdata arch_zone_highest_possible_pfn
[MAX_NR_ZONES
];
261 static unsigned long __initdata required_kernelcore
;
262 static unsigned long __initdata required_movablecore
;
263 static unsigned long __meminitdata zone_movable_pfn
[MAX_NUMNODES
];
264 static bool mirrored_kernelcore
;
266 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
268 EXPORT_SYMBOL(movable_zone
);
269 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
272 int nr_node_ids __read_mostly
= MAX_NUMNODES
;
273 int nr_online_nodes __read_mostly
= 1;
274 EXPORT_SYMBOL(nr_node_ids
);
275 EXPORT_SYMBOL(nr_online_nodes
);
278 int page_group_by_mobility_disabled __read_mostly
;
280 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
281 static inline void reset_deferred_meminit(pg_data_t
*pgdat
)
283 pgdat
->first_deferred_pfn
= ULONG_MAX
;
286 /* Returns true if the struct page for the pfn is uninitialised */
287 static inline bool __meminit
early_page_uninitialised(unsigned long pfn
)
289 if (pfn
>= NODE_DATA(early_pfn_to_nid(pfn
))->first_deferred_pfn
)
295 static inline bool early_page_nid_uninitialised(unsigned long pfn
, int nid
)
297 if (pfn
>= NODE_DATA(nid
)->first_deferred_pfn
)
304 * Returns false when the remaining initialisation should be deferred until
305 * later in the boot cycle when it can be parallelised.
307 static inline bool update_defer_init(pg_data_t
*pgdat
,
308 unsigned long pfn
, unsigned long zone_end
,
309 unsigned long *nr_initialised
)
311 unsigned long max_initialise
;
313 /* Always populate low zones for address-contrained allocations */
314 if (zone_end
< pgdat_end_pfn(pgdat
))
317 * Initialise at least 2G of a node but also take into account that
318 * two large system hashes that can take up 1GB for 0.25TB/node.
320 max_initialise
= max(2UL << (30 - PAGE_SHIFT
),
321 (pgdat
->node_spanned_pages
>> 8));
324 if ((*nr_initialised
> max_initialise
) &&
325 (pfn
& (PAGES_PER_SECTION
- 1)) == 0) {
326 pgdat
->first_deferred_pfn
= pfn
;
333 static inline void reset_deferred_meminit(pg_data_t
*pgdat
)
337 static inline bool early_page_uninitialised(unsigned long pfn
)
342 static inline bool early_page_nid_uninitialised(unsigned long pfn
, int nid
)
347 static inline bool update_defer_init(pg_data_t
*pgdat
,
348 unsigned long pfn
, unsigned long zone_end
,
349 unsigned long *nr_initialised
)
356 void set_pageblock_migratetype(struct page
*page
, int migratetype
)
358 if (unlikely(page_group_by_mobility_disabled
&&
359 migratetype
< MIGRATE_PCPTYPES
))
360 migratetype
= MIGRATE_UNMOVABLE
;
362 set_pageblock_flags_group(page
, (unsigned long)migratetype
,
363 PB_migrate
, PB_migrate_end
);
366 #ifdef CONFIG_DEBUG_VM
367 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
371 unsigned long pfn
= page_to_pfn(page
);
372 unsigned long sp
, start_pfn
;
375 seq
= zone_span_seqbegin(zone
);
376 start_pfn
= zone
->zone_start_pfn
;
377 sp
= zone
->spanned_pages
;
378 if (!zone_spans_pfn(zone
, pfn
))
380 } while (zone_span_seqretry(zone
, seq
));
383 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
384 pfn
, zone_to_nid(zone
), zone
->name
,
385 start_pfn
, start_pfn
+ sp
);
390 static int page_is_consistent(struct zone
*zone
, struct page
*page
)
392 if (!pfn_valid_within(page_to_pfn(page
)))
394 if (zone
!= page_zone(page
))
400 * Temporary debugging check for pages not lying within a given zone.
402 static int bad_range(struct zone
*zone
, struct page
*page
)
404 if (page_outside_zone_boundaries(zone
, page
))
406 if (!page_is_consistent(zone
, page
))
412 static inline int bad_range(struct zone
*zone
, struct page
*page
)
418 static void bad_page(struct page
*page
, const char *reason
,
419 unsigned long bad_flags
)
421 static unsigned long resume
;
422 static unsigned long nr_shown
;
423 static unsigned long nr_unshown
;
425 /* Don't complain about poisoned pages */
426 if (PageHWPoison(page
)) {
427 page_mapcount_reset(page
); /* remove PageBuddy */
432 * Allow a burst of 60 reports, then keep quiet for that minute;
433 * or allow a steady drip of one report per second.
435 if (nr_shown
== 60) {
436 if (time_before(jiffies
, resume
)) {
442 "BUG: Bad page state: %lu messages suppressed\n",
449 resume
= jiffies
+ 60 * HZ
;
451 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
452 current
->comm
, page_to_pfn(page
));
453 __dump_page(page
, reason
);
454 bad_flags
&= page
->flags
;
456 pr_alert("bad because of flags: %#lx(%pGp)\n",
457 bad_flags
, &bad_flags
);
458 dump_page_owner(page
);
463 /* Leave bad fields for debug, except PageBuddy could make trouble */
464 page_mapcount_reset(page
); /* remove PageBuddy */
465 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
469 * Higher-order pages are called "compound pages". They are structured thusly:
471 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
473 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
474 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
476 * The first tail page's ->compound_dtor holds the offset in array of compound
477 * page destructors. See compound_page_dtors.
479 * The first tail page's ->compound_order holds the order of allocation.
480 * This usage means that zero-order pages may not be compound.
483 void free_compound_page(struct page
*page
)
485 __free_pages_ok(page
, compound_order(page
));
488 void prep_compound_page(struct page
*page
, unsigned int order
)
491 int nr_pages
= 1 << order
;
493 set_compound_page_dtor(page
, COMPOUND_PAGE_DTOR
);
494 set_compound_order(page
, order
);
496 for (i
= 1; i
< nr_pages
; i
++) {
497 struct page
*p
= page
+ i
;
498 set_page_count(p
, 0);
499 p
->mapping
= TAIL_MAPPING
;
500 set_compound_head(p
, page
);
502 atomic_set(compound_mapcount_ptr(page
), -1);
505 #ifdef CONFIG_DEBUG_PAGEALLOC
506 unsigned int _debug_guardpage_minorder
;
507 bool _debug_pagealloc_enabled __read_mostly
508 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT
);
509 EXPORT_SYMBOL(_debug_pagealloc_enabled
);
510 bool _debug_guardpage_enabled __read_mostly
;
512 static int __init
early_debug_pagealloc(char *buf
)
517 if (strcmp(buf
, "on") == 0)
518 _debug_pagealloc_enabled
= true;
520 if (strcmp(buf
, "off") == 0)
521 _debug_pagealloc_enabled
= false;
525 early_param("debug_pagealloc", early_debug_pagealloc
);
527 static bool need_debug_guardpage(void)
529 /* If we don't use debug_pagealloc, we don't need guard page */
530 if (!debug_pagealloc_enabled())
536 static void init_debug_guardpage(void)
538 if (!debug_pagealloc_enabled())
541 _debug_guardpage_enabled
= true;
544 struct page_ext_operations debug_guardpage_ops
= {
545 .need
= need_debug_guardpage
,
546 .init
= init_debug_guardpage
,
549 static int __init
debug_guardpage_minorder_setup(char *buf
)
553 if (kstrtoul(buf
, 10, &res
) < 0 || res
> MAX_ORDER
/ 2) {
554 pr_err("Bad debug_guardpage_minorder value\n");
557 _debug_guardpage_minorder
= res
;
558 pr_info("Setting debug_guardpage_minorder to %lu\n", res
);
561 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup
);
563 static inline void set_page_guard(struct zone
*zone
, struct page
*page
,
564 unsigned int order
, int migratetype
)
566 struct page_ext
*page_ext
;
568 if (!debug_guardpage_enabled())
571 page_ext
= lookup_page_ext(page
);
572 __set_bit(PAGE_EXT_DEBUG_GUARD
, &page_ext
->flags
);
574 INIT_LIST_HEAD(&page
->lru
);
575 set_page_private(page
, order
);
576 /* Guard pages are not available for any usage */
577 __mod_zone_freepage_state(zone
, -(1 << order
), migratetype
);
580 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
581 unsigned int order
, int migratetype
)
583 struct page_ext
*page_ext
;
585 if (!debug_guardpage_enabled())
588 page_ext
= lookup_page_ext(page
);
589 __clear_bit(PAGE_EXT_DEBUG_GUARD
, &page_ext
->flags
);
591 set_page_private(page
, 0);
592 if (!is_migrate_isolate(migratetype
))
593 __mod_zone_freepage_state(zone
, (1 << order
), migratetype
);
596 struct page_ext_operations debug_guardpage_ops
= { NULL
, };
597 static inline void set_page_guard(struct zone
*zone
, struct page
*page
,
598 unsigned int order
, int migratetype
) {}
599 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
600 unsigned int order
, int migratetype
) {}
603 static inline void set_page_order(struct page
*page
, unsigned int order
)
605 set_page_private(page
, order
);
606 __SetPageBuddy(page
);
609 static inline void rmv_page_order(struct page
*page
)
611 __ClearPageBuddy(page
);
612 set_page_private(page
, 0);
616 * This function checks whether a page is free && is the buddy
617 * we can do coalesce a page and its buddy if
618 * (a) the buddy is not in a hole &&
619 * (b) the buddy is in the buddy system &&
620 * (c) a page and its buddy have the same order &&
621 * (d) a page and its buddy are in the same zone.
623 * For recording whether a page is in the buddy system, we set ->_mapcount
624 * PAGE_BUDDY_MAPCOUNT_VALUE.
625 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
626 * serialized by zone->lock.
628 * For recording page's order, we use page_private(page).
630 static inline int page_is_buddy(struct page
*page
, struct page
*buddy
,
633 if (!pfn_valid_within(page_to_pfn(buddy
)))
636 if (page_is_guard(buddy
) && page_order(buddy
) == order
) {
637 if (page_zone_id(page
) != page_zone_id(buddy
))
640 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
645 if (PageBuddy(buddy
) && page_order(buddy
) == order
) {
647 * zone check is done late to avoid uselessly
648 * calculating zone/node ids for pages that could
651 if (page_zone_id(page
) != page_zone_id(buddy
))
654 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
662 * Freeing function for a buddy system allocator.
664 * The concept of a buddy system is to maintain direct-mapped table
665 * (containing bit values) for memory blocks of various "orders".
666 * The bottom level table contains the map for the smallest allocatable
667 * units of memory (here, pages), and each level above it describes
668 * pairs of units from the levels below, hence, "buddies".
669 * At a high level, all that happens here is marking the table entry
670 * at the bottom level available, and propagating the changes upward
671 * as necessary, plus some accounting needed to play nicely with other
672 * parts of the VM system.
673 * At each level, we keep a list of pages, which are heads of continuous
674 * free pages of length of (1 << order) and marked with _mapcount
675 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
677 * So when we are allocating or freeing one, we can derive the state of the
678 * other. That is, if we allocate a small block, and both were
679 * free, the remainder of the region must be split into blocks.
680 * If a block is freed, and its buddy is also free, then this
681 * triggers coalescing into a block of larger size.
686 static inline void __free_one_page(struct page
*page
,
688 struct zone
*zone
, unsigned int order
,
691 unsigned long page_idx
;
692 unsigned long combined_idx
;
693 unsigned long uninitialized_var(buddy_idx
);
695 unsigned int max_order
;
697 max_order
= min_t(unsigned int, MAX_ORDER
, pageblock_order
+ 1);
699 VM_BUG_ON(!zone_is_initialized(zone
));
700 VM_BUG_ON_PAGE(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
, page
);
702 VM_BUG_ON(migratetype
== -1);
703 if (likely(!is_migrate_isolate(migratetype
)))
704 __mod_zone_freepage_state(zone
, 1 << order
, migratetype
);
706 page_idx
= pfn
& ((1 << MAX_ORDER
) - 1);
708 VM_BUG_ON_PAGE(page_idx
& ((1 << order
) - 1), page
);
709 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
712 while (order
< max_order
- 1) {
713 buddy_idx
= __find_buddy_index(page_idx
, order
);
714 buddy
= page
+ (buddy_idx
- page_idx
);
715 if (!page_is_buddy(page
, buddy
, order
))
718 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
719 * merge with it and move up one order.
721 if (page_is_guard(buddy
)) {
722 clear_page_guard(zone
, buddy
, order
, migratetype
);
724 list_del(&buddy
->lru
);
725 zone
->free_area
[order
].nr_free
--;
726 rmv_page_order(buddy
);
728 combined_idx
= buddy_idx
& page_idx
;
729 page
= page
+ (combined_idx
- page_idx
);
730 page_idx
= combined_idx
;
733 if (max_order
< MAX_ORDER
) {
734 /* If we are here, it means order is >= pageblock_order.
735 * We want to prevent merge between freepages on isolate
736 * pageblock and normal pageblock. Without this, pageblock
737 * isolation could cause incorrect freepage or CMA accounting.
739 * We don't want to hit this code for the more frequent
742 if (unlikely(has_isolate_pageblock(zone
))) {
745 buddy_idx
= __find_buddy_index(page_idx
, order
);
746 buddy
= page
+ (buddy_idx
- page_idx
);
747 buddy_mt
= get_pageblock_migratetype(buddy
);
749 if (migratetype
!= buddy_mt
750 && (is_migrate_isolate(migratetype
) ||
751 is_migrate_isolate(buddy_mt
)))
755 goto continue_merging
;
759 set_page_order(page
, order
);
762 * If this is not the largest possible page, check if the buddy
763 * of the next-highest order is free. If it is, it's possible
764 * that pages are being freed that will coalesce soon. In case,
765 * that is happening, add the free page to the tail of the list
766 * so it's less likely to be used soon and more likely to be merged
767 * as a higher order page
769 if ((order
< MAX_ORDER
-2) && pfn_valid_within(page_to_pfn(buddy
))) {
770 struct page
*higher_page
, *higher_buddy
;
771 combined_idx
= buddy_idx
& page_idx
;
772 higher_page
= page
+ (combined_idx
- page_idx
);
773 buddy_idx
= __find_buddy_index(combined_idx
, order
+ 1);
774 higher_buddy
= higher_page
+ (buddy_idx
- combined_idx
);
775 if (page_is_buddy(higher_page
, higher_buddy
, order
+ 1)) {
776 list_add_tail(&page
->lru
,
777 &zone
->free_area
[order
].free_list
[migratetype
]);
782 list_add(&page
->lru
, &zone
->free_area
[order
].free_list
[migratetype
]);
784 zone
->free_area
[order
].nr_free
++;
788 * A bad page could be due to a number of fields. Instead of multiple branches,
789 * try and check multiple fields with one check. The caller must do a detailed
790 * check if necessary.
792 static inline bool page_expected_state(struct page
*page
,
793 unsigned long check_flags
)
795 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
798 if (unlikely((unsigned long)page
->mapping
|
799 page_ref_count(page
) |
801 (unsigned long)page
->mem_cgroup
|
803 (page
->flags
& check_flags
)))
809 static inline int free_pages_check(struct page
*page
)
811 const char *bad_reason
;
812 unsigned long bad_flags
;
814 if (page_expected_state(page
, PAGE_FLAGS_CHECK_AT_FREE
)) {
815 page_cpupid_reset_last(page
);
816 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
820 /* Something has gone sideways, find it */
824 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
825 bad_reason
= "nonzero mapcount";
826 if (unlikely(page
->mapping
!= NULL
))
827 bad_reason
= "non-NULL mapping";
828 if (unlikely(page_ref_count(page
) != 0))
829 bad_reason
= "nonzero _refcount";
830 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_FREE
)) {
831 bad_reason
= "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
832 bad_flags
= PAGE_FLAGS_CHECK_AT_FREE
;
835 if (unlikely(page
->mem_cgroup
))
836 bad_reason
= "page still charged to cgroup";
838 bad_page(page
, bad_reason
, bad_flags
);
843 * Frees a number of pages from the PCP lists
844 * Assumes all pages on list are in same zone, and of same order.
845 * count is the number of pages to free.
847 * If the zone was previously in an "all pages pinned" state then look to
848 * see if this freeing clears that state.
850 * And clear the zone's pages_scanned counter, to hold off the "all pages are
851 * pinned" detection logic.
853 static void free_pcppages_bulk(struct zone
*zone
, int count
,
854 struct per_cpu_pages
*pcp
)
859 unsigned long nr_scanned
;
860 bool isolated_pageblocks
;
862 spin_lock(&zone
->lock
);
863 isolated_pageblocks
= has_isolate_pageblock(zone
);
864 nr_scanned
= zone_page_state(zone
, NR_PAGES_SCANNED
);
866 __mod_zone_page_state(zone
, NR_PAGES_SCANNED
, -nr_scanned
);
870 struct list_head
*list
;
873 * Remove pages from lists in a round-robin fashion. A
874 * batch_free count is maintained that is incremented when an
875 * empty list is encountered. This is so more pages are freed
876 * off fuller lists instead of spinning excessively around empty
881 if (++migratetype
== MIGRATE_PCPTYPES
)
883 list
= &pcp
->lists
[migratetype
];
884 } while (list_empty(list
));
886 /* This is the only non-empty list. Free them all. */
887 if (batch_free
== MIGRATE_PCPTYPES
)
888 batch_free
= to_free
;
891 int mt
; /* migratetype of the to-be-freed page */
893 page
= list_last_entry(list
, struct page
, lru
);
894 /* must delete as __free_one_page list manipulates */
895 list_del(&page
->lru
);
897 mt
= get_pcppage_migratetype(page
);
898 /* MIGRATE_ISOLATE page should not go to pcplists */
899 VM_BUG_ON_PAGE(is_migrate_isolate(mt
), page
);
900 /* Pageblock could have been isolated meanwhile */
901 if (unlikely(isolated_pageblocks
))
902 mt
= get_pageblock_migratetype(page
);
904 __free_one_page(page
, page_to_pfn(page
), zone
, 0, mt
);
905 trace_mm_page_pcpu_drain(page
, 0, mt
);
906 } while (--to_free
&& --batch_free
&& !list_empty(list
));
908 spin_unlock(&zone
->lock
);
911 static void free_one_page(struct zone
*zone
,
912 struct page
*page
, unsigned long pfn
,
916 unsigned long nr_scanned
;
917 spin_lock(&zone
->lock
);
918 nr_scanned
= zone_page_state(zone
, NR_PAGES_SCANNED
);
920 __mod_zone_page_state(zone
, NR_PAGES_SCANNED
, -nr_scanned
);
922 if (unlikely(has_isolate_pageblock(zone
) ||
923 is_migrate_isolate(migratetype
))) {
924 migratetype
= get_pfnblock_migratetype(page
, pfn
);
926 __free_one_page(page
, pfn
, zone
, order
, migratetype
);
927 spin_unlock(&zone
->lock
);
930 static int free_tail_pages_check(struct page
*head_page
, struct page
*page
)
935 * We rely page->lru.next never has bit 0 set, unless the page
936 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
938 BUILD_BUG_ON((unsigned long)LIST_POISON1
& 1);
940 if (!IS_ENABLED(CONFIG_DEBUG_VM
)) {
944 switch (page
- head_page
) {
946 /* the first tail page: ->mapping is compound_mapcount() */
947 if (unlikely(compound_mapcount(page
))) {
948 bad_page(page
, "nonzero compound_mapcount", 0);
954 * the second tail page: ->mapping is
955 * page_deferred_list().next -- ignore value.
959 if (page
->mapping
!= TAIL_MAPPING
) {
960 bad_page(page
, "corrupted mapping in tail page", 0);
965 if (unlikely(!PageTail(page
))) {
966 bad_page(page
, "PageTail not set", 0);
969 if (unlikely(compound_head(page
) != head_page
)) {
970 bad_page(page
, "compound_head not consistent", 0);
975 page
->mapping
= NULL
;
976 clear_compound_head(page
);
980 static void __meminit
__init_single_page(struct page
*page
, unsigned long pfn
,
981 unsigned long zone
, int nid
)
983 set_page_links(page
, zone
, nid
, pfn
);
984 init_page_count(page
);
985 page_mapcount_reset(page
);
986 page_cpupid_reset_last(page
);
988 INIT_LIST_HEAD(&page
->lru
);
989 #ifdef WANT_PAGE_VIRTUAL
990 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
991 if (!is_highmem_idx(zone
))
992 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
996 static void __meminit
__init_single_pfn(unsigned long pfn
, unsigned long zone
,
999 return __init_single_page(pfn_to_page(pfn
), pfn
, zone
, nid
);
1002 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1003 static void init_reserved_page(unsigned long pfn
)
1008 if (!early_page_uninitialised(pfn
))
1011 nid
= early_pfn_to_nid(pfn
);
1012 pgdat
= NODE_DATA(nid
);
1014 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1015 struct zone
*zone
= &pgdat
->node_zones
[zid
];
1017 if (pfn
>= zone
->zone_start_pfn
&& pfn
< zone_end_pfn(zone
))
1020 __init_single_pfn(pfn
, zid
, nid
);
1023 static inline void init_reserved_page(unsigned long pfn
)
1026 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1029 * Initialised pages do not have PageReserved set. This function is
1030 * called for each range allocated by the bootmem allocator and
1031 * marks the pages PageReserved. The remaining valid pages are later
1032 * sent to the buddy page allocator.
1034 void __meminit
reserve_bootmem_region(unsigned long start
, unsigned long end
)
1036 unsigned long start_pfn
= PFN_DOWN(start
);
1037 unsigned long end_pfn
= PFN_UP(end
);
1039 for (; start_pfn
< end_pfn
; start_pfn
++) {
1040 if (pfn_valid(start_pfn
)) {
1041 struct page
*page
= pfn_to_page(start_pfn
);
1043 init_reserved_page(start_pfn
);
1045 /* Avoid false-positive PageTail() */
1046 INIT_LIST_HEAD(&page
->lru
);
1048 SetPageReserved(page
);
1053 static bool free_pages_prepare(struct page
*page
, unsigned int order
)
1057 VM_BUG_ON_PAGE(PageTail(page
), page
);
1059 trace_mm_page_free(page
, order
);
1060 kmemcheck_free_shadow(page
, order
);
1061 kasan_free_pages(page
, order
);
1064 * Check tail pages before head page information is cleared to
1065 * avoid checking PageCompound for order-0 pages.
1067 if (unlikely(order
)) {
1068 bool compound
= PageCompound(page
);
1071 VM_BUG_ON_PAGE(compound
&& compound_order(page
) != order
, page
);
1073 for (i
= 1; i
< (1 << order
); i
++) {
1075 bad
+= free_tail_pages_check(page
, page
+ i
);
1076 bad
+= free_pages_check(page
+ i
);
1079 if (PageAnonHead(page
))
1080 page
->mapping
= NULL
;
1081 bad
+= free_pages_check(page
);
1085 reset_page_owner(page
, order
);
1087 if (!PageHighMem(page
)) {
1088 debug_check_no_locks_freed(page_address(page
),
1089 PAGE_SIZE
<< order
);
1090 debug_check_no_obj_freed(page_address(page
),
1091 PAGE_SIZE
<< order
);
1093 arch_free_page(page
, order
);
1094 kernel_poison_pages(page
, 1 << order
, 0);
1095 kernel_map_pages(page
, 1 << order
, 0);
1100 static void __free_pages_ok(struct page
*page
, unsigned int order
)
1102 unsigned long flags
;
1104 unsigned long pfn
= page_to_pfn(page
);
1106 if (!free_pages_prepare(page
, order
))
1109 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1110 local_irq_save(flags
);
1111 __count_vm_events(PGFREE
, 1 << order
);
1112 free_one_page(page_zone(page
), page
, pfn
, order
, migratetype
);
1113 local_irq_restore(flags
);
1116 static void __init
__free_pages_boot_core(struct page
*page
, unsigned int order
)
1118 unsigned int nr_pages
= 1 << order
;
1119 struct page
*p
= page
;
1123 for (loop
= 0; loop
< (nr_pages
- 1); loop
++, p
++) {
1125 __ClearPageReserved(p
);
1126 set_page_count(p
, 0);
1128 __ClearPageReserved(p
);
1129 set_page_count(p
, 0);
1131 page_zone(page
)->managed_pages
+= nr_pages
;
1132 set_page_refcounted(page
);
1133 __free_pages(page
, order
);
1136 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1137 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1139 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata
;
1141 int __meminit
early_pfn_to_nid(unsigned long pfn
)
1143 static DEFINE_SPINLOCK(early_pfn_lock
);
1146 spin_lock(&early_pfn_lock
);
1147 nid
= __early_pfn_to_nid(pfn
, &early_pfnnid_cache
);
1150 spin_unlock(&early_pfn_lock
);
1156 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1157 static inline bool __meminit
meminit_pfn_in_nid(unsigned long pfn
, int node
,
1158 struct mminit_pfnnid_cache
*state
)
1162 nid
= __early_pfn_to_nid(pfn
, state
);
1163 if (nid
>= 0 && nid
!= node
)
1168 /* Only safe to use early in boot when initialisation is single-threaded */
1169 static inline bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
1171 return meminit_pfn_in_nid(pfn
, node
, &early_pfnnid_cache
);
1176 static inline bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
1180 static inline bool __meminit
meminit_pfn_in_nid(unsigned long pfn
, int node
,
1181 struct mminit_pfnnid_cache
*state
)
1188 void __init
__free_pages_bootmem(struct page
*page
, unsigned long pfn
,
1191 if (early_page_uninitialised(pfn
))
1193 return __free_pages_boot_core(page
, order
);
1197 * Check that the whole (or subset of) a pageblock given by the interval of
1198 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1199 * with the migration of free compaction scanner. The scanners then need to
1200 * use only pfn_valid_within() check for arches that allow holes within
1203 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1205 * It's possible on some configurations to have a setup like node0 node1 node0
1206 * i.e. it's possible that all pages within a zones range of pages do not
1207 * belong to a single zone. We assume that a border between node0 and node1
1208 * can occur within a single pageblock, but not a node0 node1 node0
1209 * interleaving within a single pageblock. It is therefore sufficient to check
1210 * the first and last page of a pageblock and avoid checking each individual
1211 * page in a pageblock.
1213 struct page
*__pageblock_pfn_to_page(unsigned long start_pfn
,
1214 unsigned long end_pfn
, struct zone
*zone
)
1216 struct page
*start_page
;
1217 struct page
*end_page
;
1219 /* end_pfn is one past the range we are checking */
1222 if (!pfn_valid(start_pfn
) || !pfn_valid(end_pfn
))
1225 start_page
= pfn_to_page(start_pfn
);
1227 if (page_zone(start_page
) != zone
)
1230 end_page
= pfn_to_page(end_pfn
);
1232 /* This gives a shorter code than deriving page_zone(end_page) */
1233 if (page_zone_id(start_page
) != page_zone_id(end_page
))
1239 void set_zone_contiguous(struct zone
*zone
)
1241 unsigned long block_start_pfn
= zone
->zone_start_pfn
;
1242 unsigned long block_end_pfn
;
1244 block_end_pfn
= ALIGN(block_start_pfn
+ 1, pageblock_nr_pages
);
1245 for (; block_start_pfn
< zone_end_pfn(zone
);
1246 block_start_pfn
= block_end_pfn
,
1247 block_end_pfn
+= pageblock_nr_pages
) {
1249 block_end_pfn
= min(block_end_pfn
, zone_end_pfn(zone
));
1251 if (!__pageblock_pfn_to_page(block_start_pfn
,
1252 block_end_pfn
, zone
))
1256 /* We confirm that there is no hole */
1257 zone
->contiguous
= true;
1260 void clear_zone_contiguous(struct zone
*zone
)
1262 zone
->contiguous
= false;
1265 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1266 static void __init
deferred_free_range(struct page
*page
,
1267 unsigned long pfn
, int nr_pages
)
1274 /* Free a large naturally-aligned chunk if possible */
1275 if (nr_pages
== MAX_ORDER_NR_PAGES
&&
1276 (pfn
& (MAX_ORDER_NR_PAGES
-1)) == 0) {
1277 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
1278 __free_pages_boot_core(page
, MAX_ORDER
-1);
1282 for (i
= 0; i
< nr_pages
; i
++, page
++)
1283 __free_pages_boot_core(page
, 0);
1286 /* Completion tracking for deferred_init_memmap() threads */
1287 static atomic_t pgdat_init_n_undone __initdata
;
1288 static __initdata
DECLARE_COMPLETION(pgdat_init_all_done_comp
);
1290 static inline void __init
pgdat_init_report_one_done(void)
1292 if (atomic_dec_and_test(&pgdat_init_n_undone
))
1293 complete(&pgdat_init_all_done_comp
);
1296 /* Initialise remaining memory on a node */
1297 static int __init
deferred_init_memmap(void *data
)
1299 pg_data_t
*pgdat
= data
;
1300 int nid
= pgdat
->node_id
;
1301 struct mminit_pfnnid_cache nid_init_state
= { };
1302 unsigned long start
= jiffies
;
1303 unsigned long nr_pages
= 0;
1304 unsigned long walk_start
, walk_end
;
1307 unsigned long first_init_pfn
= pgdat
->first_deferred_pfn
;
1308 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
1310 if (first_init_pfn
== ULONG_MAX
) {
1311 pgdat_init_report_one_done();
1315 /* Bind memory initialisation thread to a local node if possible */
1316 if (!cpumask_empty(cpumask
))
1317 set_cpus_allowed_ptr(current
, cpumask
);
1319 /* Sanity check boundaries */
1320 BUG_ON(pgdat
->first_deferred_pfn
< pgdat
->node_start_pfn
);
1321 BUG_ON(pgdat
->first_deferred_pfn
> pgdat_end_pfn(pgdat
));
1322 pgdat
->first_deferred_pfn
= ULONG_MAX
;
1324 /* Only the highest zone is deferred so find it */
1325 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1326 zone
= pgdat
->node_zones
+ zid
;
1327 if (first_init_pfn
< zone_end_pfn(zone
))
1331 for_each_mem_pfn_range(i
, nid
, &walk_start
, &walk_end
, NULL
) {
1332 unsigned long pfn
, end_pfn
;
1333 struct page
*page
= NULL
;
1334 struct page
*free_base_page
= NULL
;
1335 unsigned long free_base_pfn
= 0;
1338 end_pfn
= min(walk_end
, zone_end_pfn(zone
));
1339 pfn
= first_init_pfn
;
1340 if (pfn
< walk_start
)
1342 if (pfn
< zone
->zone_start_pfn
)
1343 pfn
= zone
->zone_start_pfn
;
1345 for (; pfn
< end_pfn
; pfn
++) {
1346 if (!pfn_valid_within(pfn
))
1350 * Ensure pfn_valid is checked every
1351 * MAX_ORDER_NR_PAGES for memory holes
1353 if ((pfn
& (MAX_ORDER_NR_PAGES
- 1)) == 0) {
1354 if (!pfn_valid(pfn
)) {
1360 if (!meminit_pfn_in_nid(pfn
, nid
, &nid_init_state
)) {
1365 /* Minimise pfn page lookups and scheduler checks */
1366 if (page
&& (pfn
& (MAX_ORDER_NR_PAGES
- 1)) != 0) {
1369 nr_pages
+= nr_to_free
;
1370 deferred_free_range(free_base_page
,
1371 free_base_pfn
, nr_to_free
);
1372 free_base_page
= NULL
;
1373 free_base_pfn
= nr_to_free
= 0;
1375 page
= pfn_to_page(pfn
);
1380 VM_BUG_ON(page_zone(page
) != zone
);
1384 __init_single_page(page
, pfn
, zid
, nid
);
1385 if (!free_base_page
) {
1386 free_base_page
= page
;
1387 free_base_pfn
= pfn
;
1392 /* Where possible, batch up pages for a single free */
1395 /* Free the current block of pages to allocator */
1396 nr_pages
+= nr_to_free
;
1397 deferred_free_range(free_base_page
, free_base_pfn
,
1399 free_base_page
= NULL
;
1400 free_base_pfn
= nr_to_free
= 0;
1403 first_init_pfn
= max(end_pfn
, first_init_pfn
);
1406 /* Sanity check that the next zone really is unpopulated */
1407 WARN_ON(++zid
< MAX_NR_ZONES
&& populated_zone(++zone
));
1409 pr_info("node %d initialised, %lu pages in %ums\n", nid
, nr_pages
,
1410 jiffies_to_msecs(jiffies
- start
));
1412 pgdat_init_report_one_done();
1415 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1417 void __init
page_alloc_init_late(void)
1421 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1424 /* There will be num_node_state(N_MEMORY) threads */
1425 atomic_set(&pgdat_init_n_undone
, num_node_state(N_MEMORY
));
1426 for_each_node_state(nid
, N_MEMORY
) {
1427 kthread_run(deferred_init_memmap
, NODE_DATA(nid
), "pgdatinit%d", nid
);
1430 /* Block until all are initialised */
1431 wait_for_completion(&pgdat_init_all_done_comp
);
1433 /* Reinit limits that are based on free pages after the kernel is up */
1434 files_maxfiles_init();
1437 for_each_populated_zone(zone
)
1438 set_zone_contiguous(zone
);
1442 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1443 void __init
init_cma_reserved_pageblock(struct page
*page
)
1445 unsigned i
= pageblock_nr_pages
;
1446 struct page
*p
= page
;
1449 __ClearPageReserved(p
);
1450 set_page_count(p
, 0);
1453 set_pageblock_migratetype(page
, MIGRATE_CMA
);
1455 if (pageblock_order
>= MAX_ORDER
) {
1456 i
= pageblock_nr_pages
;
1459 set_page_refcounted(p
);
1460 __free_pages(p
, MAX_ORDER
- 1);
1461 p
+= MAX_ORDER_NR_PAGES
;
1462 } while (i
-= MAX_ORDER_NR_PAGES
);
1464 set_page_refcounted(page
);
1465 __free_pages(page
, pageblock_order
);
1468 adjust_managed_page_count(page
, pageblock_nr_pages
);
1473 * The order of subdivision here is critical for the IO subsystem.
1474 * Please do not alter this order without good reasons and regression
1475 * testing. Specifically, as large blocks of memory are subdivided,
1476 * the order in which smaller blocks are delivered depends on the order
1477 * they're subdivided in this function. This is the primary factor
1478 * influencing the order in which pages are delivered to the IO
1479 * subsystem according to empirical testing, and this is also justified
1480 * by considering the behavior of a buddy system containing a single
1481 * large block of memory acted on by a series of small allocations.
1482 * This behavior is a critical factor in sglist merging's success.
1486 static inline void expand(struct zone
*zone
, struct page
*page
,
1487 int low
, int high
, struct free_area
*area
,
1490 unsigned long size
= 1 << high
;
1492 while (high
> low
) {
1496 VM_BUG_ON_PAGE(bad_range(zone
, &page
[size
]), &page
[size
]);
1498 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC
) &&
1499 debug_guardpage_enabled() &&
1500 high
< debug_guardpage_minorder()) {
1502 * Mark as guard pages (or page), that will allow to
1503 * merge back to allocator when buddy will be freed.
1504 * Corresponding page table entries will not be touched,
1505 * pages will stay not present in virtual address space
1507 set_page_guard(zone
, &page
[size
], high
, migratetype
);
1510 list_add(&page
[size
].lru
, &area
->free_list
[migratetype
]);
1512 set_page_order(&page
[size
], high
);
1517 * This page is about to be returned from the page allocator
1519 static inline int check_new_page(struct page
*page
)
1521 const char *bad_reason
;
1522 unsigned long bad_flags
;
1524 if (page_expected_state(page
, PAGE_FLAGS_CHECK_AT_PREP
|__PG_HWPOISON
))
1529 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
1530 bad_reason
= "nonzero mapcount";
1531 if (unlikely(page
->mapping
!= NULL
))
1532 bad_reason
= "non-NULL mapping";
1533 if (unlikely(page_ref_count(page
) != 0))
1534 bad_reason
= "nonzero _count";
1535 if (unlikely(page
->flags
& __PG_HWPOISON
)) {
1536 bad_reason
= "HWPoisoned (hardware-corrupted)";
1537 bad_flags
= __PG_HWPOISON
;
1539 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)) {
1540 bad_reason
= "PAGE_FLAGS_CHECK_AT_PREP flag set";
1541 bad_flags
= PAGE_FLAGS_CHECK_AT_PREP
;
1544 if (unlikely(page
->mem_cgroup
))
1545 bad_reason
= "page still charged to cgroup";
1547 if (unlikely(bad_reason
)) {
1548 bad_page(page
, bad_reason
, bad_flags
);
1554 static inline bool free_pages_prezeroed(bool poisoned
)
1556 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO
) &&
1557 page_poisoning_enabled() && poisoned
;
1560 static int prep_new_page(struct page
*page
, unsigned int order
, gfp_t gfp_flags
,
1561 unsigned int alloc_flags
)
1564 bool poisoned
= true;
1566 for (i
= 0; i
< (1 << order
); i
++) {
1567 struct page
*p
= page
+ i
;
1568 if (unlikely(check_new_page(p
)))
1571 poisoned
&= page_is_poisoned(p
);
1574 set_page_private(page
, 0);
1575 set_page_refcounted(page
);
1577 arch_alloc_page(page
, order
);
1578 kernel_map_pages(page
, 1 << order
, 1);
1579 kernel_poison_pages(page
, 1 << order
, 1);
1580 kasan_alloc_pages(page
, order
);
1582 if (!free_pages_prezeroed(poisoned
) && (gfp_flags
& __GFP_ZERO
))
1583 for (i
= 0; i
< (1 << order
); i
++)
1584 clear_highpage(page
+ i
);
1586 if (order
&& (gfp_flags
& __GFP_COMP
))
1587 prep_compound_page(page
, order
);
1589 set_page_owner(page
, order
, gfp_flags
);
1592 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1593 * allocate the page. The expectation is that the caller is taking
1594 * steps that will free more memory. The caller should avoid the page
1595 * being used for !PFMEMALLOC purposes.
1597 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
1598 set_page_pfmemalloc(page
);
1600 clear_page_pfmemalloc(page
);
1606 * Go through the free lists for the given migratetype and remove
1607 * the smallest available page from the freelists
1610 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
1613 unsigned int current_order
;
1614 struct free_area
*area
;
1617 /* Find a page of the appropriate size in the preferred list */
1618 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
1619 area
= &(zone
->free_area
[current_order
]);
1620 page
= list_first_entry_or_null(&area
->free_list
[migratetype
],
1624 list_del(&page
->lru
);
1625 rmv_page_order(page
);
1627 expand(zone
, page
, order
, current_order
, area
, migratetype
);
1628 set_pcppage_migratetype(page
, migratetype
);
1637 * This array describes the order lists are fallen back to when
1638 * the free lists for the desirable migrate type are depleted
1640 static int fallbacks
[MIGRATE_TYPES
][4] = {
1641 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_TYPES
},
1642 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_TYPES
},
1643 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_TYPES
},
1645 [MIGRATE_CMA
] = { MIGRATE_TYPES
}, /* Never used */
1647 #ifdef CONFIG_MEMORY_ISOLATION
1648 [MIGRATE_ISOLATE
] = { MIGRATE_TYPES
}, /* Never used */
1653 static struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
1656 return __rmqueue_smallest(zone
, order
, MIGRATE_CMA
);
1659 static inline struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
1660 unsigned int order
) { return NULL
; }
1664 * Move the free pages in a range to the free lists of the requested type.
1665 * Note that start_page and end_pages are not aligned on a pageblock
1666 * boundary. If alignment is required, use move_freepages_block()
1668 int move_freepages(struct zone
*zone
,
1669 struct page
*start_page
, struct page
*end_page
,
1674 int pages_moved
= 0;
1676 #ifndef CONFIG_HOLES_IN_ZONE
1678 * page_zone is not safe to call in this context when
1679 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1680 * anyway as we check zone boundaries in move_freepages_block().
1681 * Remove at a later date when no bug reports exist related to
1682 * grouping pages by mobility
1684 VM_BUG_ON(page_zone(start_page
) != page_zone(end_page
));
1687 for (page
= start_page
; page
<= end_page
;) {
1688 /* Make sure we are not inadvertently changing nodes */
1689 VM_BUG_ON_PAGE(page_to_nid(page
) != zone_to_nid(zone
), page
);
1691 if (!pfn_valid_within(page_to_pfn(page
))) {
1696 if (!PageBuddy(page
)) {
1701 order
= page_order(page
);
1702 list_move(&page
->lru
,
1703 &zone
->free_area
[order
].free_list
[migratetype
]);
1705 pages_moved
+= 1 << order
;
1711 int move_freepages_block(struct zone
*zone
, struct page
*page
,
1714 unsigned long start_pfn
, end_pfn
;
1715 struct page
*start_page
, *end_page
;
1717 start_pfn
= page_to_pfn(page
);
1718 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
1719 start_page
= pfn_to_page(start_pfn
);
1720 end_page
= start_page
+ pageblock_nr_pages
- 1;
1721 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
1723 /* Do not cross zone boundaries */
1724 if (!zone_spans_pfn(zone
, start_pfn
))
1726 if (!zone_spans_pfn(zone
, end_pfn
))
1729 return move_freepages(zone
, start_page
, end_page
, migratetype
);
1732 static void change_pageblock_range(struct page
*pageblock_page
,
1733 int start_order
, int migratetype
)
1735 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
1737 while (nr_pageblocks
--) {
1738 set_pageblock_migratetype(pageblock_page
, migratetype
);
1739 pageblock_page
+= pageblock_nr_pages
;
1744 * When we are falling back to another migratetype during allocation, try to
1745 * steal extra free pages from the same pageblocks to satisfy further
1746 * allocations, instead of polluting multiple pageblocks.
1748 * If we are stealing a relatively large buddy page, it is likely there will
1749 * be more free pages in the pageblock, so try to steal them all. For
1750 * reclaimable and unmovable allocations, we steal regardless of page size,
1751 * as fragmentation caused by those allocations polluting movable pageblocks
1752 * is worse than movable allocations stealing from unmovable and reclaimable
1755 static bool can_steal_fallback(unsigned int order
, int start_mt
)
1758 * Leaving this order check is intended, although there is
1759 * relaxed order check in next check. The reason is that
1760 * we can actually steal whole pageblock if this condition met,
1761 * but, below check doesn't guarantee it and that is just heuristic
1762 * so could be changed anytime.
1764 if (order
>= pageblock_order
)
1767 if (order
>= pageblock_order
/ 2 ||
1768 start_mt
== MIGRATE_RECLAIMABLE
||
1769 start_mt
== MIGRATE_UNMOVABLE
||
1770 page_group_by_mobility_disabled
)
1777 * This function implements actual steal behaviour. If order is large enough,
1778 * we can steal whole pageblock. If not, we first move freepages in this
1779 * pageblock and check whether half of pages are moved or not. If half of
1780 * pages are moved, we can change migratetype of pageblock and permanently
1781 * use it's pages as requested migratetype in the future.
1783 static void steal_suitable_fallback(struct zone
*zone
, struct page
*page
,
1786 unsigned int current_order
= page_order(page
);
1789 /* Take ownership for orders >= pageblock_order */
1790 if (current_order
>= pageblock_order
) {
1791 change_pageblock_range(page
, current_order
, start_type
);
1795 pages
= move_freepages_block(zone
, page
, start_type
);
1797 /* Claim the whole block if over half of it is free */
1798 if (pages
>= (1 << (pageblock_order
-1)) ||
1799 page_group_by_mobility_disabled
)
1800 set_pageblock_migratetype(page
, start_type
);
1804 * Check whether there is a suitable fallback freepage with requested order.
1805 * If only_stealable is true, this function returns fallback_mt only if
1806 * we can steal other freepages all together. This would help to reduce
1807 * fragmentation due to mixed migratetype pages in one pageblock.
1809 int find_suitable_fallback(struct free_area
*area
, unsigned int order
,
1810 int migratetype
, bool only_stealable
, bool *can_steal
)
1815 if (area
->nr_free
== 0)
1820 fallback_mt
= fallbacks
[migratetype
][i
];
1821 if (fallback_mt
== MIGRATE_TYPES
)
1824 if (list_empty(&area
->free_list
[fallback_mt
]))
1827 if (can_steal_fallback(order
, migratetype
))
1830 if (!only_stealable
)
1841 * Reserve a pageblock for exclusive use of high-order atomic allocations if
1842 * there are no empty page blocks that contain a page with a suitable order
1844 static void reserve_highatomic_pageblock(struct page
*page
, struct zone
*zone
,
1845 unsigned int alloc_order
)
1848 unsigned long max_managed
, flags
;
1851 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
1852 * Check is race-prone but harmless.
1854 max_managed
= (zone
->managed_pages
/ 100) + pageblock_nr_pages
;
1855 if (zone
->nr_reserved_highatomic
>= max_managed
)
1858 spin_lock_irqsave(&zone
->lock
, flags
);
1860 /* Recheck the nr_reserved_highatomic limit under the lock */
1861 if (zone
->nr_reserved_highatomic
>= max_managed
)
1865 mt
= get_pageblock_migratetype(page
);
1866 if (mt
!= MIGRATE_HIGHATOMIC
&&
1867 !is_migrate_isolate(mt
) && !is_migrate_cma(mt
)) {
1868 zone
->nr_reserved_highatomic
+= pageblock_nr_pages
;
1869 set_pageblock_migratetype(page
, MIGRATE_HIGHATOMIC
);
1870 move_freepages_block(zone
, page
, MIGRATE_HIGHATOMIC
);
1874 spin_unlock_irqrestore(&zone
->lock
, flags
);
1878 * Used when an allocation is about to fail under memory pressure. This
1879 * potentially hurts the reliability of high-order allocations when under
1880 * intense memory pressure but failed atomic allocations should be easier
1881 * to recover from than an OOM.
1883 static void unreserve_highatomic_pageblock(const struct alloc_context
*ac
)
1885 struct zonelist
*zonelist
= ac
->zonelist
;
1886 unsigned long flags
;
1892 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, ac
->high_zoneidx
,
1894 /* Preserve at least one pageblock */
1895 if (zone
->nr_reserved_highatomic
<= pageblock_nr_pages
)
1898 spin_lock_irqsave(&zone
->lock
, flags
);
1899 for (order
= 0; order
< MAX_ORDER
; order
++) {
1900 struct free_area
*area
= &(zone
->free_area
[order
]);
1902 page
= list_first_entry_or_null(
1903 &area
->free_list
[MIGRATE_HIGHATOMIC
],
1909 * It should never happen but changes to locking could
1910 * inadvertently allow a per-cpu drain to add pages
1911 * to MIGRATE_HIGHATOMIC while unreserving so be safe
1912 * and watch for underflows.
1914 zone
->nr_reserved_highatomic
-= min(pageblock_nr_pages
,
1915 zone
->nr_reserved_highatomic
);
1918 * Convert to ac->migratetype and avoid the normal
1919 * pageblock stealing heuristics. Minimally, the caller
1920 * is doing the work and needs the pages. More
1921 * importantly, if the block was always converted to
1922 * MIGRATE_UNMOVABLE or another type then the number
1923 * of pageblocks that cannot be completely freed
1926 set_pageblock_migratetype(page
, ac
->migratetype
);
1927 move_freepages_block(zone
, page
, ac
->migratetype
);
1928 spin_unlock_irqrestore(&zone
->lock
, flags
);
1931 spin_unlock_irqrestore(&zone
->lock
, flags
);
1935 /* Remove an element from the buddy allocator from the fallback list */
1936 static inline struct page
*
1937 __rmqueue_fallback(struct zone
*zone
, unsigned int order
, int start_migratetype
)
1939 struct free_area
*area
;
1940 unsigned int current_order
;
1945 /* Find the largest possible block of pages in the other list */
1946 for (current_order
= MAX_ORDER
-1;
1947 current_order
>= order
&& current_order
<= MAX_ORDER
-1;
1949 area
= &(zone
->free_area
[current_order
]);
1950 fallback_mt
= find_suitable_fallback(area
, current_order
,
1951 start_migratetype
, false, &can_steal
);
1952 if (fallback_mt
== -1)
1955 page
= list_first_entry(&area
->free_list
[fallback_mt
],
1958 steal_suitable_fallback(zone
, page
, start_migratetype
);
1960 /* Remove the page from the freelists */
1962 list_del(&page
->lru
);
1963 rmv_page_order(page
);
1965 expand(zone
, page
, order
, current_order
, area
,
1968 * The pcppage_migratetype may differ from pageblock's
1969 * migratetype depending on the decisions in
1970 * find_suitable_fallback(). This is OK as long as it does not
1971 * differ for MIGRATE_CMA pageblocks. Those can be used as
1972 * fallback only via special __rmqueue_cma_fallback() function
1974 set_pcppage_migratetype(page
, start_migratetype
);
1976 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
1977 start_migratetype
, fallback_mt
);
1986 * Do the hard work of removing an element from the buddy allocator.
1987 * Call me with the zone->lock already held.
1989 static struct page
*__rmqueue(struct zone
*zone
, unsigned int order
,
1994 page
= __rmqueue_smallest(zone
, order
, migratetype
);
1995 if (unlikely(!page
)) {
1996 if (migratetype
== MIGRATE_MOVABLE
)
1997 page
= __rmqueue_cma_fallback(zone
, order
);
2000 page
= __rmqueue_fallback(zone
, order
, migratetype
);
2003 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
2008 * Obtain a specified number of elements from the buddy allocator, all under
2009 * a single hold of the lock, for efficiency. Add them to the supplied list.
2010 * Returns the number of new pages which were placed at *list.
2012 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
2013 unsigned long count
, struct list_head
*list
,
2014 int migratetype
, bool cold
)
2018 spin_lock(&zone
->lock
);
2019 for (i
= 0; i
< count
; ++i
) {
2020 struct page
*page
= __rmqueue(zone
, order
, migratetype
);
2021 if (unlikely(page
== NULL
))
2025 * Split buddy pages returned by expand() are received here
2026 * in physical page order. The page is added to the callers and
2027 * list and the list head then moves forward. From the callers
2028 * perspective, the linked list is ordered by page number in
2029 * some conditions. This is useful for IO devices that can
2030 * merge IO requests if the physical pages are ordered
2034 list_add(&page
->lru
, list
);
2036 list_add_tail(&page
->lru
, list
);
2038 if (is_migrate_cma(get_pcppage_migratetype(page
)))
2039 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
,
2042 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
2043 spin_unlock(&zone
->lock
);
2049 * Called from the vmstat counter updater to drain pagesets of this
2050 * currently executing processor on remote nodes after they have
2053 * Note that this function must be called with the thread pinned to
2054 * a single processor.
2056 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
2058 unsigned long flags
;
2059 int to_drain
, batch
;
2061 local_irq_save(flags
);
2062 batch
= READ_ONCE(pcp
->batch
);
2063 to_drain
= min(pcp
->count
, batch
);
2065 free_pcppages_bulk(zone
, to_drain
, pcp
);
2066 pcp
->count
-= to_drain
;
2068 local_irq_restore(flags
);
2073 * Drain pcplists of the indicated processor and zone.
2075 * The processor must either be the current processor and the
2076 * thread pinned to the current processor or a processor that
2079 static void drain_pages_zone(unsigned int cpu
, struct zone
*zone
)
2081 unsigned long flags
;
2082 struct per_cpu_pageset
*pset
;
2083 struct per_cpu_pages
*pcp
;
2085 local_irq_save(flags
);
2086 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
2090 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
2093 local_irq_restore(flags
);
2097 * Drain pcplists of all zones on the indicated processor.
2099 * The processor must either be the current processor and the
2100 * thread pinned to the current processor or a processor that
2103 static void drain_pages(unsigned int cpu
)
2107 for_each_populated_zone(zone
) {
2108 drain_pages_zone(cpu
, zone
);
2113 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2115 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2116 * the single zone's pages.
2118 void drain_local_pages(struct zone
*zone
)
2120 int cpu
= smp_processor_id();
2123 drain_pages_zone(cpu
, zone
);
2129 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2131 * When zone parameter is non-NULL, spill just the single zone's pages.
2133 * Note that this code is protected against sending an IPI to an offline
2134 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
2135 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
2136 * nothing keeps CPUs from showing up after we populated the cpumask and
2137 * before the call to on_each_cpu_mask().
2139 void drain_all_pages(struct zone
*zone
)
2144 * Allocate in the BSS so we wont require allocation in
2145 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2147 static cpumask_t cpus_with_pcps
;
2150 * We don't care about racing with CPU hotplug event
2151 * as offline notification will cause the notified
2152 * cpu to drain that CPU pcps and on_each_cpu_mask
2153 * disables preemption as part of its processing
2155 for_each_online_cpu(cpu
) {
2156 struct per_cpu_pageset
*pcp
;
2158 bool has_pcps
= false;
2161 pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
2165 for_each_populated_zone(z
) {
2166 pcp
= per_cpu_ptr(z
->pageset
, cpu
);
2167 if (pcp
->pcp
.count
) {
2175 cpumask_set_cpu(cpu
, &cpus_with_pcps
);
2177 cpumask_clear_cpu(cpu
, &cpus_with_pcps
);
2179 on_each_cpu_mask(&cpus_with_pcps
, (smp_call_func_t
) drain_local_pages
,
2183 #ifdef CONFIG_HIBERNATION
2185 void mark_free_pages(struct zone
*zone
)
2187 unsigned long pfn
, max_zone_pfn
;
2188 unsigned long flags
;
2189 unsigned int order
, t
;
2192 if (zone_is_empty(zone
))
2195 spin_lock_irqsave(&zone
->lock
, flags
);
2197 max_zone_pfn
= zone_end_pfn(zone
);
2198 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
2199 if (pfn_valid(pfn
)) {
2200 page
= pfn_to_page(pfn
);
2202 if (page_zone(page
) != zone
)
2205 if (!swsusp_page_is_forbidden(page
))
2206 swsusp_unset_page_free(page
);
2209 for_each_migratetype_order(order
, t
) {
2210 list_for_each_entry(page
,
2211 &zone
->free_area
[order
].free_list
[t
], lru
) {
2214 pfn
= page_to_pfn(page
);
2215 for (i
= 0; i
< (1UL << order
); i
++)
2216 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
2219 spin_unlock_irqrestore(&zone
->lock
, flags
);
2221 #endif /* CONFIG_PM */
2224 * Free a 0-order page
2225 * cold == true ? free a cold page : free a hot page
2227 void free_hot_cold_page(struct page
*page
, bool cold
)
2229 struct zone
*zone
= page_zone(page
);
2230 struct per_cpu_pages
*pcp
;
2231 unsigned long flags
;
2232 unsigned long pfn
= page_to_pfn(page
);
2235 if (!free_pages_prepare(page
, 0))
2238 migratetype
= get_pfnblock_migratetype(page
, pfn
);
2239 set_pcppage_migratetype(page
, migratetype
);
2240 local_irq_save(flags
);
2241 __count_vm_event(PGFREE
);
2244 * We only track unmovable, reclaimable and movable on pcp lists.
2245 * Free ISOLATE pages back to the allocator because they are being
2246 * offlined but treat RESERVE as movable pages so we can get those
2247 * areas back if necessary. Otherwise, we may have to free
2248 * excessively into the page allocator
2250 if (migratetype
>= MIGRATE_PCPTYPES
) {
2251 if (unlikely(is_migrate_isolate(migratetype
))) {
2252 free_one_page(zone
, page
, pfn
, 0, migratetype
);
2255 migratetype
= MIGRATE_MOVABLE
;
2258 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
2260 list_add(&page
->lru
, &pcp
->lists
[migratetype
]);
2262 list_add_tail(&page
->lru
, &pcp
->lists
[migratetype
]);
2264 if (pcp
->count
>= pcp
->high
) {
2265 unsigned long batch
= READ_ONCE(pcp
->batch
);
2266 free_pcppages_bulk(zone
, batch
, pcp
);
2267 pcp
->count
-= batch
;
2271 local_irq_restore(flags
);
2275 * Free a list of 0-order pages
2277 void free_hot_cold_page_list(struct list_head
*list
, bool cold
)
2279 struct page
*page
, *next
;
2281 list_for_each_entry_safe(page
, next
, list
, lru
) {
2282 trace_mm_page_free_batched(page
, cold
);
2283 free_hot_cold_page(page
, cold
);
2288 * split_page takes a non-compound higher-order page, and splits it into
2289 * n (1<<order) sub-pages: page[0..n]
2290 * Each sub-page must be freed individually.
2292 * Note: this is probably too low level an operation for use in drivers.
2293 * Please consult with lkml before using this in your driver.
2295 void split_page(struct page
*page
, unsigned int order
)
2300 VM_BUG_ON_PAGE(PageCompound(page
), page
);
2301 VM_BUG_ON_PAGE(!page_count(page
), page
);
2303 #ifdef CONFIG_KMEMCHECK
2305 * Split shadow pages too, because free(page[0]) would
2306 * otherwise free the whole shadow.
2308 if (kmemcheck_page_is_tracked(page
))
2309 split_page(virt_to_page(page
[0].shadow
), order
);
2312 gfp_mask
= get_page_owner_gfp(page
);
2313 set_page_owner(page
, 0, gfp_mask
);
2314 for (i
= 1; i
< (1 << order
); i
++) {
2315 set_page_refcounted(page
+ i
);
2316 set_page_owner(page
+ i
, 0, gfp_mask
);
2319 EXPORT_SYMBOL_GPL(split_page
);
2321 int __isolate_free_page(struct page
*page
, unsigned int order
)
2323 unsigned long watermark
;
2327 BUG_ON(!PageBuddy(page
));
2329 zone
= page_zone(page
);
2330 mt
= get_pageblock_migratetype(page
);
2332 if (!is_migrate_isolate(mt
)) {
2333 /* Obey watermarks as if the page was being allocated */
2334 watermark
= low_wmark_pages(zone
) + (1 << order
);
2335 if (!zone_watermark_ok(zone
, 0, watermark
, 0, 0))
2338 __mod_zone_freepage_state(zone
, -(1UL << order
), mt
);
2341 /* Remove page from free list */
2342 list_del(&page
->lru
);
2343 zone
->free_area
[order
].nr_free
--;
2344 rmv_page_order(page
);
2346 set_page_owner(page
, order
, __GFP_MOVABLE
);
2348 /* Set the pageblock if the isolated page is at least a pageblock */
2349 if (order
>= pageblock_order
- 1) {
2350 struct page
*endpage
= page
+ (1 << order
) - 1;
2351 for (; page
< endpage
; page
+= pageblock_nr_pages
) {
2352 int mt
= get_pageblock_migratetype(page
);
2353 if (!is_migrate_isolate(mt
) && !is_migrate_cma(mt
))
2354 set_pageblock_migratetype(page
,
2360 return 1UL << order
;
2364 * Similar to split_page except the page is already free. As this is only
2365 * being used for migration, the migratetype of the block also changes.
2366 * As this is called with interrupts disabled, the caller is responsible
2367 * for calling arch_alloc_page() and kernel_map_page() after interrupts
2370 * Note: this is probably too low level an operation for use in drivers.
2371 * Please consult with lkml before using this in your driver.
2373 int split_free_page(struct page
*page
)
2378 order
= page_order(page
);
2380 nr_pages
= __isolate_free_page(page
, order
);
2384 /* Split into individual pages */
2385 set_page_refcounted(page
);
2386 split_page(page
, order
);
2391 * Update NUMA hit/miss statistics
2393 * Must be called with interrupts disabled.
2395 * When __GFP_OTHER_NODE is set assume the node of the preferred
2396 * zone is the local node. This is useful for daemons who allocate
2397 * memory on behalf of other processes.
2399 static inline void zone_statistics(struct zone
*preferred_zone
, struct zone
*z
,
2403 int local_nid
= numa_node_id();
2404 enum zone_stat_item local_stat
= NUMA_LOCAL
;
2406 if (unlikely(flags
& __GFP_OTHER_NODE
)) {
2407 local_stat
= NUMA_OTHER
;
2408 local_nid
= preferred_zone
->node
;
2411 if (z
->node
== local_nid
) {
2412 __inc_zone_state(z
, NUMA_HIT
);
2413 __inc_zone_state(z
, local_stat
);
2415 __inc_zone_state(z
, NUMA_MISS
);
2416 __inc_zone_state(preferred_zone
, NUMA_FOREIGN
);
2422 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2425 struct page
*buffered_rmqueue(struct zone
*preferred_zone
,
2426 struct zone
*zone
, unsigned int order
,
2427 gfp_t gfp_flags
, unsigned int alloc_flags
,
2430 unsigned long flags
;
2432 bool cold
= ((gfp_flags
& __GFP_COLD
) != 0);
2434 if (likely(order
== 0)) {
2435 struct per_cpu_pages
*pcp
;
2436 struct list_head
*list
;
2438 local_irq_save(flags
);
2439 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
2440 list
= &pcp
->lists
[migratetype
];
2441 if (list_empty(list
)) {
2442 pcp
->count
+= rmqueue_bulk(zone
, 0,
2445 if (unlikely(list_empty(list
)))
2450 page
= list_last_entry(list
, struct page
, lru
);
2452 page
= list_first_entry(list
, struct page
, lru
);
2454 __dec_zone_state(zone
, NR_ALLOC_BATCH
);
2455 list_del(&page
->lru
);
2459 * We most definitely don't want callers attempting to
2460 * allocate greater than order-1 page units with __GFP_NOFAIL.
2462 WARN_ON_ONCE((gfp_flags
& __GFP_NOFAIL
) && (order
> 1));
2463 spin_lock_irqsave(&zone
->lock
, flags
);
2466 if (alloc_flags
& ALLOC_HARDER
) {
2467 page
= __rmqueue_smallest(zone
, order
, MIGRATE_HIGHATOMIC
);
2469 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
2472 page
= __rmqueue(zone
, order
, migratetype
);
2473 spin_unlock(&zone
->lock
);
2476 __mod_zone_page_state(zone
, NR_ALLOC_BATCH
, -(1 << order
));
2477 __mod_zone_freepage_state(zone
, -(1 << order
),
2478 get_pcppage_migratetype(page
));
2481 if (atomic_long_read(&zone
->vm_stat
[NR_ALLOC_BATCH
]) <= 0 &&
2482 !test_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
))
2483 set_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
);
2485 __count_zone_vm_events(PGALLOC
, zone
, 1 << order
);
2486 zone_statistics(preferred_zone
, zone
, gfp_flags
);
2487 local_irq_restore(flags
);
2489 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
2493 local_irq_restore(flags
);
2497 #ifdef CONFIG_FAIL_PAGE_ALLOC
2500 struct fault_attr attr
;
2502 bool ignore_gfp_highmem
;
2503 bool ignore_gfp_reclaim
;
2505 } fail_page_alloc
= {
2506 .attr
= FAULT_ATTR_INITIALIZER
,
2507 .ignore_gfp_reclaim
= true,
2508 .ignore_gfp_highmem
= true,
2512 static int __init
setup_fail_page_alloc(char *str
)
2514 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
2516 __setup("fail_page_alloc=", setup_fail_page_alloc
);
2518 static bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
2520 if (order
< fail_page_alloc
.min_order
)
2522 if (gfp_mask
& __GFP_NOFAIL
)
2524 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
2526 if (fail_page_alloc
.ignore_gfp_reclaim
&&
2527 (gfp_mask
& __GFP_DIRECT_RECLAIM
))
2530 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
2533 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2535 static int __init
fail_page_alloc_debugfs(void)
2537 umode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
2540 dir
= fault_create_debugfs_attr("fail_page_alloc", NULL
,
2541 &fail_page_alloc
.attr
);
2543 return PTR_ERR(dir
);
2545 if (!debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
2546 &fail_page_alloc
.ignore_gfp_reclaim
))
2548 if (!debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
2549 &fail_page_alloc
.ignore_gfp_highmem
))
2551 if (!debugfs_create_u32("min-order", mode
, dir
,
2552 &fail_page_alloc
.min_order
))
2557 debugfs_remove_recursive(dir
);
2562 late_initcall(fail_page_alloc_debugfs
);
2564 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2566 #else /* CONFIG_FAIL_PAGE_ALLOC */
2568 static inline bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
2573 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2576 * Return true if free base pages are above 'mark'. For high-order checks it
2577 * will return true of the order-0 watermark is reached and there is at least
2578 * one free page of a suitable size. Checking now avoids taking the zone lock
2579 * to check in the allocation paths if no pages are free.
2581 static bool __zone_watermark_ok(struct zone
*z
, unsigned int order
,
2582 unsigned long mark
, int classzone_idx
,
2583 unsigned int alloc_flags
,
2588 const bool alloc_harder
= (alloc_flags
& ALLOC_HARDER
);
2590 /* free_pages may go negative - that's OK */
2591 free_pages
-= (1 << order
) - 1;
2593 if (alloc_flags
& ALLOC_HIGH
)
2597 * If the caller does not have rights to ALLOC_HARDER then subtract
2598 * the high-atomic reserves. This will over-estimate the size of the
2599 * atomic reserve but it avoids a search.
2601 if (likely(!alloc_harder
))
2602 free_pages
-= z
->nr_reserved_highatomic
;
2607 /* If allocation can't use CMA areas don't use free CMA pages */
2608 if (!(alloc_flags
& ALLOC_CMA
))
2609 free_pages
-= zone_page_state(z
, NR_FREE_CMA_PAGES
);
2613 * Check watermarks for an order-0 allocation request. If these
2614 * are not met, then a high-order request also cannot go ahead
2615 * even if a suitable page happened to be free.
2617 if (free_pages
<= min
+ z
->lowmem_reserve
[classzone_idx
])
2620 /* If this is an order-0 request then the watermark is fine */
2624 /* For a high-order request, check at least one suitable page is free */
2625 for (o
= order
; o
< MAX_ORDER
; o
++) {
2626 struct free_area
*area
= &z
->free_area
[o
];
2635 for (mt
= 0; mt
< MIGRATE_PCPTYPES
; mt
++) {
2636 if (!list_empty(&area
->free_list
[mt
]))
2641 if ((alloc_flags
& ALLOC_CMA
) &&
2642 !list_empty(&area
->free_list
[MIGRATE_CMA
])) {
2650 bool zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
2651 int classzone_idx
, unsigned int alloc_flags
)
2653 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
2654 zone_page_state(z
, NR_FREE_PAGES
));
2657 static inline bool zone_watermark_fast(struct zone
*z
, unsigned int order
,
2658 unsigned long mark
, int classzone_idx
, unsigned int alloc_flags
)
2660 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
2664 /* If allocation can't use CMA areas don't use free CMA pages */
2665 if (!(alloc_flags
& ALLOC_CMA
))
2666 cma_pages
= zone_page_state(z
, NR_FREE_CMA_PAGES
);
2670 * Fast check for order-0 only. If this fails then the reserves
2671 * need to be calculated. There is a corner case where the check
2672 * passes but only the high-order atomic reserve are free. If
2673 * the caller is !atomic then it'll uselessly search the free
2674 * list. That corner case is then slower but it is harmless.
2676 if (!order
&& (free_pages
- cma_pages
) > mark
+ z
->lowmem_reserve
[classzone_idx
])
2679 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
2683 bool zone_watermark_ok_safe(struct zone
*z
, unsigned int order
,
2684 unsigned long mark
, int classzone_idx
)
2686 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
2688 if (z
->percpu_drift_mark
&& free_pages
< z
->percpu_drift_mark
)
2689 free_pages
= zone_page_state_snapshot(z
, NR_FREE_PAGES
);
2691 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, 0,
2696 static bool zone_local(struct zone
*local_zone
, struct zone
*zone
)
2698 return local_zone
->node
== zone
->node
;
2701 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
2703 return node_distance(zone_to_nid(local_zone
), zone_to_nid(zone
)) <
2706 #else /* CONFIG_NUMA */
2707 static bool zone_local(struct zone
*local_zone
, struct zone
*zone
)
2712 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
2716 #endif /* CONFIG_NUMA */
2718 static void reset_alloc_batches(struct zone
*preferred_zone
)
2720 struct zone
*zone
= preferred_zone
->zone_pgdat
->node_zones
;
2723 mod_zone_page_state(zone
, NR_ALLOC_BATCH
,
2724 high_wmark_pages(zone
) - low_wmark_pages(zone
) -
2725 atomic_long_read(&zone
->vm_stat
[NR_ALLOC_BATCH
]));
2726 clear_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
);
2727 } while (zone
++ != preferred_zone
);
2731 * get_page_from_freelist goes through the zonelist trying to allocate
2734 static struct page
*
2735 get_page_from_freelist(gfp_t gfp_mask
, unsigned int order
, int alloc_flags
,
2736 const struct alloc_context
*ac
)
2738 struct zoneref
*z
= ac
->preferred_zoneref
;
2740 bool fair_skipped
= false;
2741 bool apply_fair
= (alloc_flags
& ALLOC_FAIR
);
2745 * Scan zonelist, looking for a zone with enough free.
2746 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2748 for_next_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
2753 if (cpusets_enabled() &&
2754 (alloc_flags
& ALLOC_CPUSET
) &&
2755 !cpuset_zone_allowed(zone
, gfp_mask
))
2758 * Distribute pages in proportion to the individual
2759 * zone size to ensure fair page aging. The zone a
2760 * page was allocated in should have no effect on the
2761 * time the page has in memory before being reclaimed.
2764 if (test_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
)) {
2765 fair_skipped
= true;
2768 if (!zone_local(ac
->preferred_zoneref
->zone
, zone
)) {
2775 * When allocating a page cache page for writing, we
2776 * want to get it from a zone that is within its dirty
2777 * limit, such that no single zone holds more than its
2778 * proportional share of globally allowed dirty pages.
2779 * The dirty limits take into account the zone's
2780 * lowmem reserves and high watermark so that kswapd
2781 * should be able to balance it without having to
2782 * write pages from its LRU list.
2784 * This may look like it could increase pressure on
2785 * lower zones by failing allocations in higher zones
2786 * before they are full. But the pages that do spill
2787 * over are limited as the lower zones are protected
2788 * by this very same mechanism. It should not become
2789 * a practical burden to them.
2791 * XXX: For now, allow allocations to potentially
2792 * exceed the per-zone dirty limit in the slowpath
2793 * (spread_dirty_pages unset) before going into reclaim,
2794 * which is important when on a NUMA setup the allowed
2795 * zones are together not big enough to reach the
2796 * global limit. The proper fix for these situations
2797 * will require awareness of zones in the
2798 * dirty-throttling and the flusher threads.
2800 if (ac
->spread_dirty_pages
&& !zone_dirty_ok(zone
))
2803 mark
= zone
->watermark
[alloc_flags
& ALLOC_WMARK_MASK
];
2804 if (!zone_watermark_fast(zone
, order
, mark
,
2805 ac_classzone_idx(ac
), alloc_flags
)) {
2808 /* Checked here to keep the fast path fast */
2809 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
2810 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
2813 if (zone_reclaim_mode
== 0 ||
2814 !zone_allows_reclaim(ac
->preferred_zoneref
->zone
, zone
))
2817 ret
= zone_reclaim(zone
, gfp_mask
, order
);
2819 case ZONE_RECLAIM_NOSCAN
:
2822 case ZONE_RECLAIM_FULL
:
2823 /* scanned but unreclaimable */
2826 /* did we reclaim enough */
2827 if (zone_watermark_ok(zone
, order
, mark
,
2828 ac_classzone_idx(ac
), alloc_flags
))
2836 page
= buffered_rmqueue(ac
->preferred_zoneref
->zone
, zone
, order
,
2837 gfp_mask
, alloc_flags
, ac
->migratetype
);
2839 if (prep_new_page(page
, order
, gfp_mask
, alloc_flags
))
2843 * If this is a high-order atomic allocation then check
2844 * if the pageblock should be reserved for the future
2846 if (unlikely(order
&& (alloc_flags
& ALLOC_HARDER
)))
2847 reserve_highatomic_pageblock(page
, zone
, order
);
2854 * The first pass makes sure allocations are spread fairly within the
2855 * local node. However, the local node might have free pages left
2856 * after the fairness batches are exhausted, and remote zones haven't
2857 * even been considered yet. Try once more without fairness, and
2858 * include remote zones now, before entering the slowpath and waking
2859 * kswapd: prefer spilling to a remote zone over swapping locally.
2864 fair_skipped
= false;
2865 reset_alloc_batches(ac
->preferred_zoneref
->zone
);
2873 * Large machines with many possible nodes should not always dump per-node
2874 * meminfo in irq context.
2876 static inline bool should_suppress_show_mem(void)
2881 ret
= in_interrupt();
2886 static DEFINE_RATELIMIT_STATE(nopage_rs
,
2887 DEFAULT_RATELIMIT_INTERVAL
,
2888 DEFAULT_RATELIMIT_BURST
);
2890 void warn_alloc_failed(gfp_t gfp_mask
, unsigned int order
, const char *fmt
, ...)
2892 unsigned int filter
= SHOW_MEM_FILTER_NODES
;
2894 if ((gfp_mask
& __GFP_NOWARN
) || !__ratelimit(&nopage_rs
) ||
2895 debug_guardpage_minorder() > 0)
2899 * This documents exceptions given to allocations in certain
2900 * contexts that are allowed to allocate outside current's set
2903 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
2904 if (test_thread_flag(TIF_MEMDIE
) ||
2905 (current
->flags
& (PF_MEMALLOC
| PF_EXITING
)))
2906 filter
&= ~SHOW_MEM_FILTER_NODES
;
2907 if (in_interrupt() || !(gfp_mask
& __GFP_DIRECT_RECLAIM
))
2908 filter
&= ~SHOW_MEM_FILTER_NODES
;
2911 struct va_format vaf
;
2914 va_start(args
, fmt
);
2919 pr_warn("%pV", &vaf
);
2924 pr_warn("%s: page allocation failure: order:%u, mode:%#x(%pGg)\n",
2925 current
->comm
, order
, gfp_mask
, &gfp_mask
);
2927 if (!should_suppress_show_mem())
2931 static inline struct page
*
2932 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
2933 const struct alloc_context
*ac
, unsigned long *did_some_progress
)
2935 struct oom_control oc
= {
2936 .zonelist
= ac
->zonelist
,
2937 .nodemask
= ac
->nodemask
,
2938 .gfp_mask
= gfp_mask
,
2943 *did_some_progress
= 0;
2946 * Acquire the oom lock. If that fails, somebody else is
2947 * making progress for us.
2949 if (!mutex_trylock(&oom_lock
)) {
2950 *did_some_progress
= 1;
2951 schedule_timeout_uninterruptible(1);
2956 * Go through the zonelist yet one more time, keep very high watermark
2957 * here, this is only to catch a parallel oom killing, we must fail if
2958 * we're still under heavy pressure.
2960 page
= get_page_from_freelist(gfp_mask
| __GFP_HARDWALL
, order
,
2961 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
, ac
);
2965 if (!(gfp_mask
& __GFP_NOFAIL
)) {
2966 /* Coredumps can quickly deplete all memory reserves */
2967 if (current
->flags
& PF_DUMPCORE
)
2969 /* The OOM killer will not help higher order allocs */
2970 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
2972 /* The OOM killer does not needlessly kill tasks for lowmem */
2973 if (ac
->high_zoneidx
< ZONE_NORMAL
)
2975 if (pm_suspended_storage())
2978 * XXX: GFP_NOFS allocations should rather fail than rely on
2979 * other request to make a forward progress.
2980 * We are in an unfortunate situation where out_of_memory cannot
2981 * do much for this context but let's try it to at least get
2982 * access to memory reserved if the current task is killed (see
2983 * out_of_memory). Once filesystems are ready to handle allocation
2984 * failures more gracefully we should just bail out here.
2987 /* The OOM killer may not free memory on a specific node */
2988 if (gfp_mask
& __GFP_THISNODE
)
2991 /* Exhausted what can be done so it's blamo time */
2992 if (out_of_memory(&oc
) || WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
)) {
2993 *did_some_progress
= 1;
2995 if (gfp_mask
& __GFP_NOFAIL
) {
2996 page
= get_page_from_freelist(gfp_mask
, order
,
2997 ALLOC_NO_WATERMARKS
|ALLOC_CPUSET
, ac
);
2999 * fallback to ignore cpuset restriction if our nodes
3003 page
= get_page_from_freelist(gfp_mask
, order
,
3004 ALLOC_NO_WATERMARKS
, ac
);
3008 mutex_unlock(&oom_lock
);
3012 #ifdef CONFIG_COMPACTION
3013 /* Try memory compaction for high-order allocations before reclaim */
3014 static struct page
*
3015 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
3016 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3017 enum migrate_mode mode
, int *contended_compaction
,
3018 bool *deferred_compaction
)
3020 unsigned long compact_result
;
3026 current
->flags
|= PF_MEMALLOC
;
3027 compact_result
= try_to_compact_pages(gfp_mask
, order
, alloc_flags
, ac
,
3028 mode
, contended_compaction
);
3029 current
->flags
&= ~PF_MEMALLOC
;
3031 switch (compact_result
) {
3032 case COMPACT_DEFERRED
:
3033 *deferred_compaction
= true;
3035 case COMPACT_SKIPPED
:
3042 * At least in one zone compaction wasn't deferred or skipped, so let's
3043 * count a compaction stall
3045 count_vm_event(COMPACTSTALL
);
3047 page
= get_page_from_freelist(gfp_mask
, order
,
3048 alloc_flags
& ~ALLOC_NO_WATERMARKS
, ac
);
3051 struct zone
*zone
= page_zone(page
);
3053 zone
->compact_blockskip_flush
= false;
3054 compaction_defer_reset(zone
, order
, true);
3055 count_vm_event(COMPACTSUCCESS
);
3060 * It's bad if compaction run occurs and fails. The most likely reason
3061 * is that pages exist, but not enough to satisfy watermarks.
3063 count_vm_event(COMPACTFAIL
);
3070 static inline struct page
*
3071 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
3072 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3073 enum migrate_mode mode
, int *contended_compaction
,
3074 bool *deferred_compaction
)
3078 #endif /* CONFIG_COMPACTION */
3080 /* Perform direct synchronous page reclaim */
3082 __perform_reclaim(gfp_t gfp_mask
, unsigned int order
,
3083 const struct alloc_context
*ac
)
3085 struct reclaim_state reclaim_state
;
3090 /* We now go into synchronous reclaim */
3091 cpuset_memory_pressure_bump();
3092 current
->flags
|= PF_MEMALLOC
;
3093 lockdep_set_current_reclaim_state(gfp_mask
);
3094 reclaim_state
.reclaimed_slab
= 0;
3095 current
->reclaim_state
= &reclaim_state
;
3097 progress
= try_to_free_pages(ac
->zonelist
, order
, gfp_mask
,
3100 current
->reclaim_state
= NULL
;
3101 lockdep_clear_current_reclaim_state();
3102 current
->flags
&= ~PF_MEMALLOC
;
3109 /* The really slow allocator path where we enter direct reclaim */
3110 static inline struct page
*
3111 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
3112 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3113 unsigned long *did_some_progress
)
3115 struct page
*page
= NULL
;
3116 bool drained
= false;
3118 *did_some_progress
= __perform_reclaim(gfp_mask
, order
, ac
);
3119 if (unlikely(!(*did_some_progress
)))
3123 page
= get_page_from_freelist(gfp_mask
, order
,
3124 alloc_flags
& ~ALLOC_NO_WATERMARKS
, ac
);
3127 * If an allocation failed after direct reclaim, it could be because
3128 * pages are pinned on the per-cpu lists or in high alloc reserves.
3129 * Shrink them them and try again
3131 if (!page
&& !drained
) {
3132 unreserve_highatomic_pageblock(ac
);
3133 drain_all_pages(NULL
);
3141 static void wake_all_kswapds(unsigned int order
, const struct alloc_context
*ac
)
3146 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
,
3147 ac
->high_zoneidx
, ac
->nodemask
)
3148 wakeup_kswapd(zone
, order
, ac_classzone_idx(ac
));
3151 static inline unsigned int
3152 gfp_to_alloc_flags(gfp_t gfp_mask
)
3154 unsigned int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
3156 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3157 BUILD_BUG_ON(__GFP_HIGH
!= (__force gfp_t
) ALLOC_HIGH
);
3160 * The caller may dip into page reserves a bit more if the caller
3161 * cannot run direct reclaim, or if the caller has realtime scheduling
3162 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3163 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3165 alloc_flags
|= (__force
int) (gfp_mask
& __GFP_HIGH
);
3167 if (gfp_mask
& __GFP_ATOMIC
) {
3169 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3170 * if it can't schedule.
3172 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
3173 alloc_flags
|= ALLOC_HARDER
;
3175 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3176 * comment for __cpuset_node_allowed().
3178 alloc_flags
&= ~ALLOC_CPUSET
;
3179 } else if (unlikely(rt_task(current
)) && !in_interrupt())
3180 alloc_flags
|= ALLOC_HARDER
;
3182 if (likely(!(gfp_mask
& __GFP_NOMEMALLOC
))) {
3183 if (gfp_mask
& __GFP_MEMALLOC
)
3184 alloc_flags
|= ALLOC_NO_WATERMARKS
;
3185 else if (in_serving_softirq() && (current
->flags
& PF_MEMALLOC
))
3186 alloc_flags
|= ALLOC_NO_WATERMARKS
;
3187 else if (!in_interrupt() &&
3188 ((current
->flags
& PF_MEMALLOC
) ||
3189 unlikely(test_thread_flag(TIF_MEMDIE
))))
3190 alloc_flags
|= ALLOC_NO_WATERMARKS
;
3193 if (gfpflags_to_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
3194 alloc_flags
|= ALLOC_CMA
;
3199 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask
)
3201 return !!(gfp_to_alloc_flags(gfp_mask
) & ALLOC_NO_WATERMARKS
);
3204 static inline bool is_thp_gfp_mask(gfp_t gfp_mask
)
3206 return (gfp_mask
& (GFP_TRANSHUGE
| __GFP_KSWAPD_RECLAIM
)) == GFP_TRANSHUGE
;
3209 static inline struct page
*
3210 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
3211 struct alloc_context
*ac
)
3213 bool can_direct_reclaim
= gfp_mask
& __GFP_DIRECT_RECLAIM
;
3214 struct page
*page
= NULL
;
3215 unsigned int alloc_flags
;
3216 unsigned long pages_reclaimed
= 0;
3217 unsigned long did_some_progress
;
3218 enum migrate_mode migration_mode
= MIGRATE_ASYNC
;
3219 bool deferred_compaction
= false;
3220 int contended_compaction
= COMPACT_CONTENDED_NONE
;
3223 * In the slowpath, we sanity check order to avoid ever trying to
3224 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3225 * be using allocators in order of preference for an area that is
3228 if (order
>= MAX_ORDER
) {
3229 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
3234 * We also sanity check to catch abuse of atomic reserves being used by
3235 * callers that are not in atomic context.
3237 if (WARN_ON_ONCE((gfp_mask
& (__GFP_ATOMIC
|__GFP_DIRECT_RECLAIM
)) ==
3238 (__GFP_ATOMIC
|__GFP_DIRECT_RECLAIM
)))
3239 gfp_mask
&= ~__GFP_ATOMIC
;
3242 if (gfp_mask
& __GFP_KSWAPD_RECLAIM
)
3243 wake_all_kswapds(order
, ac
);
3246 * OK, we're below the kswapd watermark and have kicked background
3247 * reclaim. Now things get more complex, so set up alloc_flags according
3248 * to how we want to proceed.
3250 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
3252 /* This is the last chance, in general, before the goto nopage. */
3253 page
= get_page_from_freelist(gfp_mask
, order
,
3254 alloc_flags
& ~ALLOC_NO_WATERMARKS
, ac
);
3258 /* Allocate without watermarks if the context allows */
3259 if (alloc_flags
& ALLOC_NO_WATERMARKS
) {
3261 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
3262 * the allocation is high priority and these type of
3263 * allocations are system rather than user orientated
3265 ac
->zonelist
= node_zonelist(numa_node_id(), gfp_mask
);
3266 page
= get_page_from_freelist(gfp_mask
, order
,
3267 ALLOC_NO_WATERMARKS
, ac
);
3272 /* Caller is not willing to reclaim, we can't balance anything */
3273 if (!can_direct_reclaim
) {
3275 * All existing users of the __GFP_NOFAIL are blockable, so warn
3276 * of any new users that actually allow this type of allocation
3279 WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
);
3283 /* Avoid recursion of direct reclaim */
3284 if (current
->flags
& PF_MEMALLOC
) {
3286 * __GFP_NOFAIL request from this context is rather bizarre
3287 * because we cannot reclaim anything and only can loop waiting
3288 * for somebody to do a work for us.
3290 if (WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
)) {
3297 /* Avoid allocations with no watermarks from looping endlessly */
3298 if (test_thread_flag(TIF_MEMDIE
) && !(gfp_mask
& __GFP_NOFAIL
))
3302 * Try direct compaction. The first pass is asynchronous. Subsequent
3303 * attempts after direct reclaim are synchronous
3305 page
= __alloc_pages_direct_compact(gfp_mask
, order
, alloc_flags
, ac
,
3307 &contended_compaction
,
3308 &deferred_compaction
);
3312 /* Checks for THP-specific high-order allocations */
3313 if (is_thp_gfp_mask(gfp_mask
)) {
3315 * If compaction is deferred for high-order allocations, it is
3316 * because sync compaction recently failed. If this is the case
3317 * and the caller requested a THP allocation, we do not want
3318 * to heavily disrupt the system, so we fail the allocation
3319 * instead of entering direct reclaim.
3321 if (deferred_compaction
)
3325 * In all zones where compaction was attempted (and not
3326 * deferred or skipped), lock contention has been detected.
3327 * For THP allocation we do not want to disrupt the others
3328 * so we fallback to base pages instead.
3330 if (contended_compaction
== COMPACT_CONTENDED_LOCK
)
3334 * If compaction was aborted due to need_resched(), we do not
3335 * want to further increase allocation latency, unless it is
3336 * khugepaged trying to collapse.
3338 if (contended_compaction
== COMPACT_CONTENDED_SCHED
3339 && !(current
->flags
& PF_KTHREAD
))
3344 * It can become very expensive to allocate transparent hugepages at
3345 * fault, so use asynchronous memory compaction for THP unless it is
3346 * khugepaged trying to collapse.
3348 if (!is_thp_gfp_mask(gfp_mask
) || (current
->flags
& PF_KTHREAD
))
3349 migration_mode
= MIGRATE_SYNC_LIGHT
;
3351 /* Try direct reclaim and then allocating */
3352 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
, alloc_flags
, ac
,
3353 &did_some_progress
);
3357 /* Do not loop if specifically requested */
3358 if (gfp_mask
& __GFP_NORETRY
)
3361 /* Keep reclaiming pages as long as there is reasonable progress */
3362 pages_reclaimed
+= did_some_progress
;
3363 if ((did_some_progress
&& order
<= PAGE_ALLOC_COSTLY_ORDER
) ||
3364 ((gfp_mask
& __GFP_REPEAT
) && pages_reclaimed
< (1 << order
))) {
3365 /* Wait for some write requests to complete then retry */
3366 wait_iff_congested(ac
->preferred_zoneref
->zone
, BLK_RW_ASYNC
, HZ
/50);
3370 /* Reclaim has failed us, start killing things */
3371 page
= __alloc_pages_may_oom(gfp_mask
, order
, ac
, &did_some_progress
);
3375 /* Retry as long as the OOM killer is making progress */
3376 if (did_some_progress
)
3381 * High-order allocations do not necessarily loop after
3382 * direct reclaim and reclaim/compaction depends on compaction
3383 * being called after reclaim so call directly if necessary
3385 page
= __alloc_pages_direct_compact(gfp_mask
, order
, alloc_flags
,
3387 &contended_compaction
,
3388 &deferred_compaction
);
3392 warn_alloc_failed(gfp_mask
, order
, NULL
);
3398 * This is the 'heart' of the zoned buddy allocator.
3401 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
,
3402 struct zonelist
*zonelist
, nodemask_t
*nodemask
)
3405 unsigned int cpuset_mems_cookie
;
3406 unsigned int alloc_flags
= ALLOC_WMARK_LOW
|ALLOC_FAIR
;
3407 gfp_t alloc_mask
= gfp_mask
; /* The gfp_t that was actually used for allocation */
3408 struct alloc_context ac
= {
3409 .high_zoneidx
= gfp_zone(gfp_mask
),
3410 .zonelist
= zonelist
,
3411 .nodemask
= nodemask
,
3412 .migratetype
= gfpflags_to_migratetype(gfp_mask
),
3415 if (cpusets_enabled()) {
3416 alloc_mask
|= __GFP_HARDWALL
;
3417 alloc_flags
|= ALLOC_CPUSET
;
3419 ac
.nodemask
= &cpuset_current_mems_allowed
;
3422 gfp_mask
&= gfp_allowed_mask
;
3424 lockdep_trace_alloc(gfp_mask
);
3426 might_sleep_if(gfp_mask
& __GFP_DIRECT_RECLAIM
);
3428 if (should_fail_alloc_page(gfp_mask
, order
))
3432 * Check the zones suitable for the gfp_mask contain at least one
3433 * valid zone. It's possible to have an empty zonelist as a result
3434 * of __GFP_THISNODE and a memoryless node
3436 if (unlikely(!zonelist
->_zonerefs
->zone
))
3439 if (IS_ENABLED(CONFIG_CMA
) && ac
.migratetype
== MIGRATE_MOVABLE
)
3440 alloc_flags
|= ALLOC_CMA
;
3443 cpuset_mems_cookie
= read_mems_allowed_begin();
3445 /* Dirty zone balancing only done in the fast path */
3446 ac
.spread_dirty_pages
= (gfp_mask
& __GFP_WRITE
);
3448 /* The preferred zone is used for statistics later */
3449 ac
.preferred_zoneref
= first_zones_zonelist(ac
.zonelist
,
3450 ac
.high_zoneidx
, ac
.nodemask
);
3451 if (!ac
.preferred_zoneref
) {
3456 /* First allocation attempt */
3457 page
= get_page_from_freelist(alloc_mask
, order
, alloc_flags
, &ac
);
3462 * Runtime PM, block IO and its error handling path can deadlock
3463 * because I/O on the device might not complete.
3465 alloc_mask
= memalloc_noio_flags(gfp_mask
);
3466 ac
.spread_dirty_pages
= false;
3468 page
= __alloc_pages_slowpath(alloc_mask
, order
, &ac
);
3472 * When updating a task's mems_allowed, it is possible to race with
3473 * parallel threads in such a way that an allocation can fail while
3474 * the mask is being updated. If a page allocation is about to fail,
3475 * check if the cpuset changed during allocation and if so, retry.
3477 if (unlikely(!page
&& read_mems_allowed_retry(cpuset_mems_cookie
))) {
3478 alloc_mask
= gfp_mask
;
3483 if (kmemcheck_enabled
&& page
)
3484 kmemcheck_pagealloc_alloc(page
, order
, gfp_mask
);
3486 trace_mm_page_alloc(page
, order
, alloc_mask
, ac
.migratetype
);
3490 EXPORT_SYMBOL(__alloc_pages_nodemask
);
3493 * Common helper functions.
3495 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
3500 * __get_free_pages() returns a 32-bit address, which cannot represent
3503 VM_BUG_ON((gfp_mask
& __GFP_HIGHMEM
) != 0);
3505 page
= alloc_pages(gfp_mask
, order
);
3508 return (unsigned long) page_address(page
);
3510 EXPORT_SYMBOL(__get_free_pages
);
3512 unsigned long get_zeroed_page(gfp_t gfp_mask
)
3514 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
3516 EXPORT_SYMBOL(get_zeroed_page
);
3518 void __free_pages(struct page
*page
, unsigned int order
)
3520 if (put_page_testzero(page
)) {
3522 free_hot_cold_page(page
, false);
3524 __free_pages_ok(page
, order
);
3528 EXPORT_SYMBOL(__free_pages
);
3530 void free_pages(unsigned long addr
, unsigned int order
)
3533 VM_BUG_ON(!virt_addr_valid((void *)addr
));
3534 __free_pages(virt_to_page((void *)addr
), order
);
3538 EXPORT_SYMBOL(free_pages
);
3542 * An arbitrary-length arbitrary-offset area of memory which resides
3543 * within a 0 or higher order page. Multiple fragments within that page
3544 * are individually refcounted, in the page's reference counter.
3546 * The page_frag functions below provide a simple allocation framework for
3547 * page fragments. This is used by the network stack and network device
3548 * drivers to provide a backing region of memory for use as either an
3549 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3551 static struct page
*__page_frag_refill(struct page_frag_cache
*nc
,
3554 struct page
*page
= NULL
;
3555 gfp_t gfp
= gfp_mask
;
3557 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3558 gfp_mask
|= __GFP_COMP
| __GFP_NOWARN
| __GFP_NORETRY
|
3560 page
= alloc_pages_node(NUMA_NO_NODE
, gfp_mask
,
3561 PAGE_FRAG_CACHE_MAX_ORDER
);
3562 nc
->size
= page
? PAGE_FRAG_CACHE_MAX_SIZE
: PAGE_SIZE
;
3564 if (unlikely(!page
))
3565 page
= alloc_pages_node(NUMA_NO_NODE
, gfp
, 0);
3567 nc
->va
= page
? page_address(page
) : NULL
;
3572 void *__alloc_page_frag(struct page_frag_cache
*nc
,
3573 unsigned int fragsz
, gfp_t gfp_mask
)
3575 unsigned int size
= PAGE_SIZE
;
3579 if (unlikely(!nc
->va
)) {
3581 page
= __page_frag_refill(nc
, gfp_mask
);
3585 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3586 /* if size can vary use size else just use PAGE_SIZE */
3589 /* Even if we own the page, we do not use atomic_set().
3590 * This would break get_page_unless_zero() users.
3592 page_ref_add(page
, size
- 1);
3594 /* reset page count bias and offset to start of new frag */
3595 nc
->pfmemalloc
= page_is_pfmemalloc(page
);
3596 nc
->pagecnt_bias
= size
;
3600 offset
= nc
->offset
- fragsz
;
3601 if (unlikely(offset
< 0)) {
3602 page
= virt_to_page(nc
->va
);
3604 if (!page_ref_sub_and_test(page
, nc
->pagecnt_bias
))
3607 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3608 /* if size can vary use size else just use PAGE_SIZE */
3611 /* OK, page count is 0, we can safely set it */
3612 set_page_count(page
, size
);
3614 /* reset page count bias and offset to start of new frag */
3615 nc
->pagecnt_bias
= size
;
3616 offset
= size
- fragsz
;
3620 nc
->offset
= offset
;
3622 return nc
->va
+ offset
;
3624 EXPORT_SYMBOL(__alloc_page_frag
);
3627 * Frees a page fragment allocated out of either a compound or order 0 page.
3629 void __free_page_frag(void *addr
)
3631 struct page
*page
= virt_to_head_page(addr
);
3633 if (unlikely(put_page_testzero(page
)))
3634 __free_pages_ok(page
, compound_order(page
));
3636 EXPORT_SYMBOL(__free_page_frag
);
3639 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
3640 * of the current memory cgroup if __GFP_ACCOUNT is set, other than that it is
3641 * equivalent to alloc_pages.
3643 * It should be used when the caller would like to use kmalloc, but since the
3644 * allocation is large, it has to fall back to the page allocator.
3646 struct page
*alloc_kmem_pages(gfp_t gfp_mask
, unsigned int order
)
3650 page
= alloc_pages(gfp_mask
, order
);
3651 if (page
&& memcg_kmem_charge(page
, gfp_mask
, order
) != 0) {
3652 __free_pages(page
, order
);
3658 struct page
*alloc_kmem_pages_node(int nid
, gfp_t gfp_mask
, unsigned int order
)
3662 page
= alloc_pages_node(nid
, gfp_mask
, order
);
3663 if (page
&& memcg_kmem_charge(page
, gfp_mask
, order
) != 0) {
3664 __free_pages(page
, order
);
3671 * __free_kmem_pages and free_kmem_pages will free pages allocated with
3674 void __free_kmem_pages(struct page
*page
, unsigned int order
)
3676 memcg_kmem_uncharge(page
, order
);
3677 __free_pages(page
, order
);
3680 void free_kmem_pages(unsigned long addr
, unsigned int order
)
3683 VM_BUG_ON(!virt_addr_valid((void *)addr
));
3684 __free_kmem_pages(virt_to_page((void *)addr
), order
);
3688 static void *make_alloc_exact(unsigned long addr
, unsigned int order
,
3692 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
3693 unsigned long used
= addr
+ PAGE_ALIGN(size
);
3695 split_page(virt_to_page((void *)addr
), order
);
3696 while (used
< alloc_end
) {
3701 return (void *)addr
;
3705 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
3706 * @size: the number of bytes to allocate
3707 * @gfp_mask: GFP flags for the allocation
3709 * This function is similar to alloc_pages(), except that it allocates the
3710 * minimum number of pages to satisfy the request. alloc_pages() can only
3711 * allocate memory in power-of-two pages.
3713 * This function is also limited by MAX_ORDER.
3715 * Memory allocated by this function must be released by free_pages_exact().
3717 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
3719 unsigned int order
= get_order(size
);
3722 addr
= __get_free_pages(gfp_mask
, order
);
3723 return make_alloc_exact(addr
, order
, size
);
3725 EXPORT_SYMBOL(alloc_pages_exact
);
3728 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3730 * @nid: the preferred node ID where memory should be allocated
3731 * @size: the number of bytes to allocate
3732 * @gfp_mask: GFP flags for the allocation
3734 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3737 void * __meminit
alloc_pages_exact_nid(int nid
, size_t size
, gfp_t gfp_mask
)
3739 unsigned int order
= get_order(size
);
3740 struct page
*p
= alloc_pages_node(nid
, gfp_mask
, order
);
3743 return make_alloc_exact((unsigned long)page_address(p
), order
, size
);
3747 * free_pages_exact - release memory allocated via alloc_pages_exact()
3748 * @virt: the value returned by alloc_pages_exact.
3749 * @size: size of allocation, same value as passed to alloc_pages_exact().
3751 * Release the memory allocated by a previous call to alloc_pages_exact.
3753 void free_pages_exact(void *virt
, size_t size
)
3755 unsigned long addr
= (unsigned long)virt
;
3756 unsigned long end
= addr
+ PAGE_ALIGN(size
);
3758 while (addr
< end
) {
3763 EXPORT_SYMBOL(free_pages_exact
);
3766 * nr_free_zone_pages - count number of pages beyond high watermark
3767 * @offset: The zone index of the highest zone
3769 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3770 * high watermark within all zones at or below a given zone index. For each
3771 * zone, the number of pages is calculated as:
3772 * managed_pages - high_pages
3774 static unsigned long nr_free_zone_pages(int offset
)
3779 /* Just pick one node, since fallback list is circular */
3780 unsigned long sum
= 0;
3782 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
3784 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
3785 unsigned long size
= zone
->managed_pages
;
3786 unsigned long high
= high_wmark_pages(zone
);
3795 * nr_free_buffer_pages - count number of pages beyond high watermark
3797 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3798 * watermark within ZONE_DMA and ZONE_NORMAL.
3800 unsigned long nr_free_buffer_pages(void)
3802 return nr_free_zone_pages(gfp_zone(GFP_USER
));
3804 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
3807 * nr_free_pagecache_pages - count number of pages beyond high watermark
3809 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3810 * high watermark within all zones.
3812 unsigned long nr_free_pagecache_pages(void)
3814 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
3817 static inline void show_node(struct zone
*zone
)
3819 if (IS_ENABLED(CONFIG_NUMA
))
3820 printk("Node %d ", zone_to_nid(zone
));
3823 long si_mem_available(void)
3826 unsigned long pagecache
;
3827 unsigned long wmark_low
= 0;
3828 unsigned long pages
[NR_LRU_LISTS
];
3832 for (lru
= LRU_BASE
; lru
< NR_LRU_LISTS
; lru
++)
3833 pages
[lru
] = global_page_state(NR_LRU_BASE
+ lru
);
3836 wmark_low
+= zone
->watermark
[WMARK_LOW
];
3839 * Estimate the amount of memory available for userspace allocations,
3840 * without causing swapping.
3842 available
= global_page_state(NR_FREE_PAGES
) - totalreserve_pages
;
3845 * Not all the page cache can be freed, otherwise the system will
3846 * start swapping. Assume at least half of the page cache, or the
3847 * low watermark worth of cache, needs to stay.
3849 pagecache
= pages
[LRU_ACTIVE_FILE
] + pages
[LRU_INACTIVE_FILE
];
3850 pagecache
-= min(pagecache
/ 2, wmark_low
);
3851 available
+= pagecache
;
3854 * Part of the reclaimable slab consists of items that are in use,
3855 * and cannot be freed. Cap this estimate at the low watermark.
3857 available
+= global_page_state(NR_SLAB_RECLAIMABLE
) -
3858 min(global_page_state(NR_SLAB_RECLAIMABLE
) / 2, wmark_low
);
3864 EXPORT_SYMBOL_GPL(si_mem_available
);
3866 void si_meminfo(struct sysinfo
*val
)
3868 val
->totalram
= totalram_pages
;
3869 val
->sharedram
= global_page_state(NR_SHMEM
);
3870 val
->freeram
= global_page_state(NR_FREE_PAGES
);
3871 val
->bufferram
= nr_blockdev_pages();
3872 val
->totalhigh
= totalhigh_pages
;
3873 val
->freehigh
= nr_free_highpages();
3874 val
->mem_unit
= PAGE_SIZE
;
3877 EXPORT_SYMBOL(si_meminfo
);
3880 void si_meminfo_node(struct sysinfo
*val
, int nid
)
3882 int zone_type
; /* needs to be signed */
3883 unsigned long managed_pages
= 0;
3884 unsigned long managed_highpages
= 0;
3885 unsigned long free_highpages
= 0;
3886 pg_data_t
*pgdat
= NODE_DATA(nid
);
3888 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++)
3889 managed_pages
+= pgdat
->node_zones
[zone_type
].managed_pages
;
3890 val
->totalram
= managed_pages
;
3891 val
->sharedram
= node_page_state(nid
, NR_SHMEM
);
3892 val
->freeram
= node_page_state(nid
, NR_FREE_PAGES
);
3893 #ifdef CONFIG_HIGHMEM
3894 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
3895 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
3897 if (is_highmem(zone
)) {
3898 managed_highpages
+= zone
->managed_pages
;
3899 free_highpages
+= zone_page_state(zone
, NR_FREE_PAGES
);
3902 val
->totalhigh
= managed_highpages
;
3903 val
->freehigh
= free_highpages
;
3905 val
->totalhigh
= managed_highpages
;
3906 val
->freehigh
= free_highpages
;
3908 val
->mem_unit
= PAGE_SIZE
;
3913 * Determine whether the node should be displayed or not, depending on whether
3914 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3916 bool skip_free_areas_node(unsigned int flags
, int nid
)
3919 unsigned int cpuset_mems_cookie
;
3921 if (!(flags
& SHOW_MEM_FILTER_NODES
))
3925 cpuset_mems_cookie
= read_mems_allowed_begin();
3926 ret
= !node_isset(nid
, cpuset_current_mems_allowed
);
3927 } while (read_mems_allowed_retry(cpuset_mems_cookie
));
3932 #define K(x) ((x) << (PAGE_SHIFT-10))
3934 static void show_migration_types(unsigned char type
)
3936 static const char types
[MIGRATE_TYPES
] = {
3937 [MIGRATE_UNMOVABLE
] = 'U',
3938 [MIGRATE_MOVABLE
] = 'M',
3939 [MIGRATE_RECLAIMABLE
] = 'E',
3940 [MIGRATE_HIGHATOMIC
] = 'H',
3942 [MIGRATE_CMA
] = 'C',
3944 #ifdef CONFIG_MEMORY_ISOLATION
3945 [MIGRATE_ISOLATE
] = 'I',
3948 char tmp
[MIGRATE_TYPES
+ 1];
3952 for (i
= 0; i
< MIGRATE_TYPES
; i
++) {
3953 if (type
& (1 << i
))
3958 printk("(%s) ", tmp
);
3962 * Show free area list (used inside shift_scroll-lock stuff)
3963 * We also calculate the percentage fragmentation. We do this by counting the
3964 * memory on each free list with the exception of the first item on the list.
3967 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
3970 void show_free_areas(unsigned int filter
)
3972 unsigned long free_pcp
= 0;
3976 for_each_populated_zone(zone
) {
3977 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3980 for_each_online_cpu(cpu
)
3981 free_pcp
+= per_cpu_ptr(zone
->pageset
, cpu
)->pcp
.count
;
3984 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3985 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3986 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
3987 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3988 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3989 " free:%lu free_pcp:%lu free_cma:%lu\n",
3990 global_page_state(NR_ACTIVE_ANON
),
3991 global_page_state(NR_INACTIVE_ANON
),
3992 global_page_state(NR_ISOLATED_ANON
),
3993 global_page_state(NR_ACTIVE_FILE
),
3994 global_page_state(NR_INACTIVE_FILE
),
3995 global_page_state(NR_ISOLATED_FILE
),
3996 global_page_state(NR_UNEVICTABLE
),
3997 global_page_state(NR_FILE_DIRTY
),
3998 global_page_state(NR_WRITEBACK
),
3999 global_page_state(NR_UNSTABLE_NFS
),
4000 global_page_state(NR_SLAB_RECLAIMABLE
),
4001 global_page_state(NR_SLAB_UNRECLAIMABLE
),
4002 global_page_state(NR_FILE_MAPPED
),
4003 global_page_state(NR_SHMEM
),
4004 global_page_state(NR_PAGETABLE
),
4005 global_page_state(NR_BOUNCE
),
4006 global_page_state(NR_FREE_PAGES
),
4008 global_page_state(NR_FREE_CMA_PAGES
));
4010 for_each_populated_zone(zone
) {
4013 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
4017 for_each_online_cpu(cpu
)
4018 free_pcp
+= per_cpu_ptr(zone
->pageset
, cpu
)->pcp
.count
;
4026 " active_anon:%lukB"
4027 " inactive_anon:%lukB"
4028 " active_file:%lukB"
4029 " inactive_file:%lukB"
4030 " unevictable:%lukB"
4031 " isolated(anon):%lukB"
4032 " isolated(file):%lukB"
4040 " slab_reclaimable:%lukB"
4041 " slab_unreclaimable:%lukB"
4042 " kernel_stack:%lukB"
4049 " writeback_tmp:%lukB"
4050 " pages_scanned:%lu"
4051 " all_unreclaimable? %s"
4054 K(zone_page_state(zone
, NR_FREE_PAGES
)),
4055 K(min_wmark_pages(zone
)),
4056 K(low_wmark_pages(zone
)),
4057 K(high_wmark_pages(zone
)),
4058 K(zone_page_state(zone
, NR_ACTIVE_ANON
)),
4059 K(zone_page_state(zone
, NR_INACTIVE_ANON
)),
4060 K(zone_page_state(zone
, NR_ACTIVE_FILE
)),
4061 K(zone_page_state(zone
, NR_INACTIVE_FILE
)),
4062 K(zone_page_state(zone
, NR_UNEVICTABLE
)),
4063 K(zone_page_state(zone
, NR_ISOLATED_ANON
)),
4064 K(zone_page_state(zone
, NR_ISOLATED_FILE
)),
4065 K(zone
->present_pages
),
4066 K(zone
->managed_pages
),
4067 K(zone_page_state(zone
, NR_MLOCK
)),
4068 K(zone_page_state(zone
, NR_FILE_DIRTY
)),
4069 K(zone_page_state(zone
, NR_WRITEBACK
)),
4070 K(zone_page_state(zone
, NR_FILE_MAPPED
)),
4071 K(zone_page_state(zone
, NR_SHMEM
)),
4072 K(zone_page_state(zone
, NR_SLAB_RECLAIMABLE
)),
4073 K(zone_page_state(zone
, NR_SLAB_UNRECLAIMABLE
)),
4074 zone_page_state(zone
, NR_KERNEL_STACK
) *
4076 K(zone_page_state(zone
, NR_PAGETABLE
)),
4077 K(zone_page_state(zone
, NR_UNSTABLE_NFS
)),
4078 K(zone_page_state(zone
, NR_BOUNCE
)),
4080 K(this_cpu_read(zone
->pageset
->pcp
.count
)),
4081 K(zone_page_state(zone
, NR_FREE_CMA_PAGES
)),
4082 K(zone_page_state(zone
, NR_WRITEBACK_TEMP
)),
4083 K(zone_page_state(zone
, NR_PAGES_SCANNED
)),
4084 (!zone_reclaimable(zone
) ? "yes" : "no")
4086 printk("lowmem_reserve[]:");
4087 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4088 printk(" %ld", zone
->lowmem_reserve
[i
]);
4092 for_each_populated_zone(zone
) {
4094 unsigned long nr
[MAX_ORDER
], flags
, total
= 0;
4095 unsigned char types
[MAX_ORDER
];
4097 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
4100 printk("%s: ", zone
->name
);
4102 spin_lock_irqsave(&zone
->lock
, flags
);
4103 for (order
= 0; order
< MAX_ORDER
; order
++) {
4104 struct free_area
*area
= &zone
->free_area
[order
];
4107 nr
[order
] = area
->nr_free
;
4108 total
+= nr
[order
] << order
;
4111 for (type
= 0; type
< MIGRATE_TYPES
; type
++) {
4112 if (!list_empty(&area
->free_list
[type
]))
4113 types
[order
] |= 1 << type
;
4116 spin_unlock_irqrestore(&zone
->lock
, flags
);
4117 for (order
= 0; order
< MAX_ORDER
; order
++) {
4118 printk("%lu*%lukB ", nr
[order
], K(1UL) << order
);
4120 show_migration_types(types
[order
]);
4122 printk("= %lukB\n", K(total
));
4125 hugetlb_show_meminfo();
4127 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES
));
4129 show_swap_cache_info();
4132 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
4134 zoneref
->zone
= zone
;
4135 zoneref
->zone_idx
= zone_idx(zone
);
4139 * Builds allocation fallback zone lists.
4141 * Add all populated zones of a node to the zonelist.
4143 static int build_zonelists_node(pg_data_t
*pgdat
, struct zonelist
*zonelist
,
4147 enum zone_type zone_type
= MAX_NR_ZONES
;
4151 zone
= pgdat
->node_zones
+ zone_type
;
4152 if (populated_zone(zone
)) {
4153 zoneref_set_zone(zone
,
4154 &zonelist
->_zonerefs
[nr_zones
++]);
4155 check_highest_zone(zone_type
);
4157 } while (zone_type
);
4165 * 0 = automatic detection of better ordering.
4166 * 1 = order by ([node] distance, -zonetype)
4167 * 2 = order by (-zonetype, [node] distance)
4169 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
4170 * the same zonelist. So only NUMA can configure this param.
4172 #define ZONELIST_ORDER_DEFAULT 0
4173 #define ZONELIST_ORDER_NODE 1
4174 #define ZONELIST_ORDER_ZONE 2
4176 /* zonelist order in the kernel.
4177 * set_zonelist_order() will set this to NODE or ZONE.
4179 static int current_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
4180 static char zonelist_order_name
[3][8] = {"Default", "Node", "Zone"};
4184 /* The value user specified ....changed by config */
4185 static int user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
4186 /* string for sysctl */
4187 #define NUMA_ZONELIST_ORDER_LEN 16
4188 char numa_zonelist_order
[16] = "default";
4191 * interface for configure zonelist ordering.
4192 * command line option "numa_zonelist_order"
4193 * = "[dD]efault - default, automatic configuration.
4194 * = "[nN]ode - order by node locality, then by zone within node
4195 * = "[zZ]one - order by zone, then by locality within zone
4198 static int __parse_numa_zonelist_order(char *s
)
4200 if (*s
== 'd' || *s
== 'D') {
4201 user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
4202 } else if (*s
== 'n' || *s
== 'N') {
4203 user_zonelist_order
= ZONELIST_ORDER_NODE
;
4204 } else if (*s
== 'z' || *s
== 'Z') {
4205 user_zonelist_order
= ZONELIST_ORDER_ZONE
;
4207 pr_warn("Ignoring invalid numa_zonelist_order value: %s\n", s
);
4213 static __init
int setup_numa_zonelist_order(char *s
)
4220 ret
= __parse_numa_zonelist_order(s
);
4222 strlcpy(numa_zonelist_order
, s
, NUMA_ZONELIST_ORDER_LEN
);
4226 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
4229 * sysctl handler for numa_zonelist_order
4231 int numa_zonelist_order_handler(struct ctl_table
*table
, int write
,
4232 void __user
*buffer
, size_t *length
,
4235 char saved_string
[NUMA_ZONELIST_ORDER_LEN
];
4237 static DEFINE_MUTEX(zl_order_mutex
);
4239 mutex_lock(&zl_order_mutex
);
4241 if (strlen((char *)table
->data
) >= NUMA_ZONELIST_ORDER_LEN
) {
4245 strcpy(saved_string
, (char *)table
->data
);
4247 ret
= proc_dostring(table
, write
, buffer
, length
, ppos
);
4251 int oldval
= user_zonelist_order
;
4253 ret
= __parse_numa_zonelist_order((char *)table
->data
);
4256 * bogus value. restore saved string
4258 strncpy((char *)table
->data
, saved_string
,
4259 NUMA_ZONELIST_ORDER_LEN
);
4260 user_zonelist_order
= oldval
;
4261 } else if (oldval
!= user_zonelist_order
) {
4262 mutex_lock(&zonelists_mutex
);
4263 build_all_zonelists(NULL
, NULL
);
4264 mutex_unlock(&zonelists_mutex
);
4268 mutex_unlock(&zl_order_mutex
);
4273 #define MAX_NODE_LOAD (nr_online_nodes)
4274 static int node_load
[MAX_NUMNODES
];
4277 * find_next_best_node - find the next node that should appear in a given node's fallback list
4278 * @node: node whose fallback list we're appending
4279 * @used_node_mask: nodemask_t of already used nodes
4281 * We use a number of factors to determine which is the next node that should
4282 * appear on a given node's fallback list. The node should not have appeared
4283 * already in @node's fallback list, and it should be the next closest node
4284 * according to the distance array (which contains arbitrary distance values
4285 * from each node to each node in the system), and should also prefer nodes
4286 * with no CPUs, since presumably they'll have very little allocation pressure
4287 * on them otherwise.
4288 * It returns -1 if no node is found.
4290 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
4293 int min_val
= INT_MAX
;
4294 int best_node
= NUMA_NO_NODE
;
4295 const struct cpumask
*tmp
= cpumask_of_node(0);
4297 /* Use the local node if we haven't already */
4298 if (!node_isset(node
, *used_node_mask
)) {
4299 node_set(node
, *used_node_mask
);
4303 for_each_node_state(n
, N_MEMORY
) {
4305 /* Don't want a node to appear more than once */
4306 if (node_isset(n
, *used_node_mask
))
4309 /* Use the distance array to find the distance */
4310 val
= node_distance(node
, n
);
4312 /* Penalize nodes under us ("prefer the next node") */
4315 /* Give preference to headless and unused nodes */
4316 tmp
= cpumask_of_node(n
);
4317 if (!cpumask_empty(tmp
))
4318 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
4320 /* Slight preference for less loaded node */
4321 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
4322 val
+= node_load
[n
];
4324 if (val
< min_val
) {
4331 node_set(best_node
, *used_node_mask
);
4338 * Build zonelists ordered by node and zones within node.
4339 * This results in maximum locality--normal zone overflows into local
4340 * DMA zone, if any--but risks exhausting DMA zone.
4342 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int node
)
4345 struct zonelist
*zonelist
;
4347 zonelist
= &pgdat
->node_zonelists
[0];
4348 for (j
= 0; zonelist
->_zonerefs
[j
].zone
!= NULL
; j
++)
4350 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
4351 zonelist
->_zonerefs
[j
].zone
= NULL
;
4352 zonelist
->_zonerefs
[j
].zone_idx
= 0;
4356 * Build gfp_thisnode zonelists
4358 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
4361 struct zonelist
*zonelist
;
4363 zonelist
= &pgdat
->node_zonelists
[1];
4364 j
= build_zonelists_node(pgdat
, zonelist
, 0);
4365 zonelist
->_zonerefs
[j
].zone
= NULL
;
4366 zonelist
->_zonerefs
[j
].zone_idx
= 0;
4370 * Build zonelists ordered by zone and nodes within zones.
4371 * This results in conserving DMA zone[s] until all Normal memory is
4372 * exhausted, but results in overflowing to remote node while memory
4373 * may still exist in local DMA zone.
4375 static int node_order
[MAX_NUMNODES
];
4377 static void build_zonelists_in_zone_order(pg_data_t
*pgdat
, int nr_nodes
)
4380 int zone_type
; /* needs to be signed */
4382 struct zonelist
*zonelist
;
4384 zonelist
= &pgdat
->node_zonelists
[0];
4386 for (zone_type
= MAX_NR_ZONES
- 1; zone_type
>= 0; zone_type
--) {
4387 for (j
= 0; j
< nr_nodes
; j
++) {
4388 node
= node_order
[j
];
4389 z
= &NODE_DATA(node
)->node_zones
[zone_type
];
4390 if (populated_zone(z
)) {
4392 &zonelist
->_zonerefs
[pos
++]);
4393 check_highest_zone(zone_type
);
4397 zonelist
->_zonerefs
[pos
].zone
= NULL
;
4398 zonelist
->_zonerefs
[pos
].zone_idx
= 0;
4401 #if defined(CONFIG_64BIT)
4403 * Devices that require DMA32/DMA are relatively rare and do not justify a
4404 * penalty to every machine in case the specialised case applies. Default
4405 * to Node-ordering on 64-bit NUMA machines
4407 static int default_zonelist_order(void)
4409 return ZONELIST_ORDER_NODE
;
4413 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4414 * by the kernel. If processes running on node 0 deplete the low memory zone
4415 * then reclaim will occur more frequency increasing stalls and potentially
4416 * be easier to OOM if a large percentage of the zone is under writeback or
4417 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4418 * Hence, default to zone ordering on 32-bit.
4420 static int default_zonelist_order(void)
4422 return ZONELIST_ORDER_ZONE
;
4424 #endif /* CONFIG_64BIT */
4426 static void set_zonelist_order(void)
4428 if (user_zonelist_order
== ZONELIST_ORDER_DEFAULT
)
4429 current_zonelist_order
= default_zonelist_order();
4431 current_zonelist_order
= user_zonelist_order
;
4434 static void build_zonelists(pg_data_t
*pgdat
)
4437 nodemask_t used_mask
;
4438 int local_node
, prev_node
;
4439 struct zonelist
*zonelist
;
4440 unsigned int order
= current_zonelist_order
;
4442 /* initialize zonelists */
4443 for (i
= 0; i
< MAX_ZONELISTS
; i
++) {
4444 zonelist
= pgdat
->node_zonelists
+ i
;
4445 zonelist
->_zonerefs
[0].zone
= NULL
;
4446 zonelist
->_zonerefs
[0].zone_idx
= 0;
4449 /* NUMA-aware ordering of nodes */
4450 local_node
= pgdat
->node_id
;
4451 load
= nr_online_nodes
;
4452 prev_node
= local_node
;
4453 nodes_clear(used_mask
);
4455 memset(node_order
, 0, sizeof(node_order
));
4458 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
4460 * We don't want to pressure a particular node.
4461 * So adding penalty to the first node in same
4462 * distance group to make it round-robin.
4464 if (node_distance(local_node
, node
) !=
4465 node_distance(local_node
, prev_node
))
4466 node_load
[node
] = load
;
4470 if (order
== ZONELIST_ORDER_NODE
)
4471 build_zonelists_in_node_order(pgdat
, node
);
4473 node_order
[i
++] = node
; /* remember order */
4476 if (order
== ZONELIST_ORDER_ZONE
) {
4477 /* calculate node order -- i.e., DMA last! */
4478 build_zonelists_in_zone_order(pgdat
, i
);
4481 build_thisnode_zonelists(pgdat
);
4484 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4486 * Return node id of node used for "local" allocations.
4487 * I.e., first node id of first zone in arg node's generic zonelist.
4488 * Used for initializing percpu 'numa_mem', which is used primarily
4489 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4491 int local_memory_node(int node
)
4495 z
= first_zones_zonelist(node_zonelist(node
, GFP_KERNEL
),
4496 gfp_zone(GFP_KERNEL
),
4498 return z
->zone
->node
;
4502 #else /* CONFIG_NUMA */
4504 static void set_zonelist_order(void)
4506 current_zonelist_order
= ZONELIST_ORDER_ZONE
;
4509 static void build_zonelists(pg_data_t
*pgdat
)
4511 int node
, local_node
;
4513 struct zonelist
*zonelist
;
4515 local_node
= pgdat
->node_id
;
4517 zonelist
= &pgdat
->node_zonelists
[0];
4518 j
= build_zonelists_node(pgdat
, zonelist
, 0);
4521 * Now we build the zonelist so that it contains the zones
4522 * of all the other nodes.
4523 * We don't want to pressure a particular node, so when
4524 * building the zones for node N, we make sure that the
4525 * zones coming right after the local ones are those from
4526 * node N+1 (modulo N)
4528 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
4529 if (!node_online(node
))
4531 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
4533 for (node
= 0; node
< local_node
; node
++) {
4534 if (!node_online(node
))
4536 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
4539 zonelist
->_zonerefs
[j
].zone
= NULL
;
4540 zonelist
->_zonerefs
[j
].zone_idx
= 0;
4543 #endif /* CONFIG_NUMA */
4546 * Boot pageset table. One per cpu which is going to be used for all
4547 * zones and all nodes. The parameters will be set in such a way
4548 * that an item put on a list will immediately be handed over to
4549 * the buddy list. This is safe since pageset manipulation is done
4550 * with interrupts disabled.
4552 * The boot_pagesets must be kept even after bootup is complete for
4553 * unused processors and/or zones. They do play a role for bootstrapping
4554 * hotplugged processors.
4556 * zoneinfo_show() and maybe other functions do
4557 * not check if the processor is online before following the pageset pointer.
4558 * Other parts of the kernel may not check if the zone is available.
4560 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
);
4561 static DEFINE_PER_CPU(struct per_cpu_pageset
, boot_pageset
);
4562 static void setup_zone_pageset(struct zone
*zone
);
4565 * Global mutex to protect against size modification of zonelists
4566 * as well as to serialize pageset setup for the new populated zone.
4568 DEFINE_MUTEX(zonelists_mutex
);
4570 /* return values int ....just for stop_machine() */
4571 static int __build_all_zonelists(void *data
)
4575 pg_data_t
*self
= data
;
4578 memset(node_load
, 0, sizeof(node_load
));
4581 if (self
&& !node_online(self
->node_id
)) {
4582 build_zonelists(self
);
4585 for_each_online_node(nid
) {
4586 pg_data_t
*pgdat
= NODE_DATA(nid
);
4588 build_zonelists(pgdat
);
4592 * Initialize the boot_pagesets that are going to be used
4593 * for bootstrapping processors. The real pagesets for
4594 * each zone will be allocated later when the per cpu
4595 * allocator is available.
4597 * boot_pagesets are used also for bootstrapping offline
4598 * cpus if the system is already booted because the pagesets
4599 * are needed to initialize allocators on a specific cpu too.
4600 * F.e. the percpu allocator needs the page allocator which
4601 * needs the percpu allocator in order to allocate its pagesets
4602 * (a chicken-egg dilemma).
4604 for_each_possible_cpu(cpu
) {
4605 setup_pageset(&per_cpu(boot_pageset
, cpu
), 0);
4607 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4609 * We now know the "local memory node" for each node--
4610 * i.e., the node of the first zone in the generic zonelist.
4611 * Set up numa_mem percpu variable for on-line cpus. During
4612 * boot, only the boot cpu should be on-line; we'll init the
4613 * secondary cpus' numa_mem as they come on-line. During
4614 * node/memory hotplug, we'll fixup all on-line cpus.
4616 if (cpu_online(cpu
))
4617 set_cpu_numa_mem(cpu
, local_memory_node(cpu_to_node(cpu
)));
4624 static noinline
void __init
4625 build_all_zonelists_init(void)
4627 __build_all_zonelists(NULL
);
4628 mminit_verify_zonelist();
4629 cpuset_init_current_mems_allowed();
4633 * Called with zonelists_mutex held always
4634 * unless system_state == SYSTEM_BOOTING.
4636 * __ref due to (1) call of __meminit annotated setup_zone_pageset
4637 * [we're only called with non-NULL zone through __meminit paths] and
4638 * (2) call of __init annotated helper build_all_zonelists_init
4639 * [protected by SYSTEM_BOOTING].
4641 void __ref
build_all_zonelists(pg_data_t
*pgdat
, struct zone
*zone
)
4643 set_zonelist_order();
4645 if (system_state
== SYSTEM_BOOTING
) {
4646 build_all_zonelists_init();
4648 #ifdef CONFIG_MEMORY_HOTPLUG
4650 setup_zone_pageset(zone
);
4652 /* we have to stop all cpus to guarantee there is no user
4654 stop_machine(__build_all_zonelists
, pgdat
, NULL
);
4655 /* cpuset refresh routine should be here */
4657 vm_total_pages
= nr_free_pagecache_pages();
4659 * Disable grouping by mobility if the number of pages in the
4660 * system is too low to allow the mechanism to work. It would be
4661 * more accurate, but expensive to check per-zone. This check is
4662 * made on memory-hotadd so a system can start with mobility
4663 * disabled and enable it later
4665 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
4666 page_group_by_mobility_disabled
= 1;
4668 page_group_by_mobility_disabled
= 0;
4670 pr_info("Built %i zonelists in %s order, mobility grouping %s. Total pages: %ld\n",
4672 zonelist_order_name
[current_zonelist_order
],
4673 page_group_by_mobility_disabled
? "off" : "on",
4676 pr_info("Policy zone: %s\n", zone_names
[policy_zone
]);
4681 * Helper functions to size the waitqueue hash table.
4682 * Essentially these want to choose hash table sizes sufficiently
4683 * large so that collisions trying to wait on pages are rare.
4684 * But in fact, the number of active page waitqueues on typical
4685 * systems is ridiculously low, less than 200. So this is even
4686 * conservative, even though it seems large.
4688 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
4689 * waitqueues, i.e. the size of the waitq table given the number of pages.
4691 #define PAGES_PER_WAITQUEUE 256
4693 #ifndef CONFIG_MEMORY_HOTPLUG
4694 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
4696 unsigned long size
= 1;
4698 pages
/= PAGES_PER_WAITQUEUE
;
4700 while (size
< pages
)
4704 * Once we have dozens or even hundreds of threads sleeping
4705 * on IO we've got bigger problems than wait queue collision.
4706 * Limit the size of the wait table to a reasonable size.
4708 size
= min(size
, 4096UL);
4710 return max(size
, 4UL);
4714 * A zone's size might be changed by hot-add, so it is not possible to determine
4715 * a suitable size for its wait_table. So we use the maximum size now.
4717 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
4719 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
4720 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
4721 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
4723 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
4724 * or more by the traditional way. (See above). It equals:
4726 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
4727 * ia64(16K page size) : = ( 8G + 4M)byte.
4728 * powerpc (64K page size) : = (32G +16M)byte.
4730 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
4737 * This is an integer logarithm so that shifts can be used later
4738 * to extract the more random high bits from the multiplicative
4739 * hash function before the remainder is taken.
4741 static inline unsigned long wait_table_bits(unsigned long size
)
4747 * Initially all pages are reserved - free ones are freed
4748 * up by free_all_bootmem() once the early boot process is
4749 * done. Non-atomic initialization, single-pass.
4751 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
4752 unsigned long start_pfn
, enum memmap_context context
)
4754 struct vmem_altmap
*altmap
= to_vmem_altmap(__pfn_to_phys(start_pfn
));
4755 unsigned long end_pfn
= start_pfn
+ size
;
4756 pg_data_t
*pgdat
= NODE_DATA(nid
);
4758 unsigned long nr_initialised
= 0;
4759 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4760 struct memblock_region
*r
= NULL
, *tmp
;
4763 if (highest_memmap_pfn
< end_pfn
- 1)
4764 highest_memmap_pfn
= end_pfn
- 1;
4767 * Honor reservation requested by the driver for this ZONE_DEVICE
4770 if (altmap
&& start_pfn
== altmap
->base_pfn
)
4771 start_pfn
+= altmap
->reserve
;
4773 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
4775 * There can be holes in boot-time mem_map[]s handed to this
4776 * function. They do not exist on hotplugged memory.
4778 if (context
!= MEMMAP_EARLY
)
4781 if (!early_pfn_valid(pfn
))
4783 if (!early_pfn_in_nid(pfn
, nid
))
4785 if (!update_defer_init(pgdat
, pfn
, end_pfn
, &nr_initialised
))
4788 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4790 * If not mirrored_kernelcore and ZONE_MOVABLE exists, range
4791 * from zone_movable_pfn[nid] to end of each node should be
4792 * ZONE_MOVABLE not ZONE_NORMAL. skip it.
4794 if (!mirrored_kernelcore
&& zone_movable_pfn
[nid
])
4795 if (zone
== ZONE_NORMAL
&& pfn
>= zone_movable_pfn
[nid
])
4799 * Check given memblock attribute by firmware which can affect
4800 * kernel memory layout. If zone==ZONE_MOVABLE but memory is
4801 * mirrored, it's an overlapped memmap init. skip it.
4803 if (mirrored_kernelcore
&& zone
== ZONE_MOVABLE
) {
4804 if (!r
|| pfn
>= memblock_region_memory_end_pfn(r
)) {
4805 for_each_memblock(memory
, tmp
)
4806 if (pfn
< memblock_region_memory_end_pfn(tmp
))
4810 if (pfn
>= memblock_region_memory_base_pfn(r
) &&
4811 memblock_is_mirror(r
)) {
4812 /* already initialized as NORMAL */
4813 pfn
= memblock_region_memory_end_pfn(r
);
4821 * Mark the block movable so that blocks are reserved for
4822 * movable at startup. This will force kernel allocations
4823 * to reserve their blocks rather than leaking throughout
4824 * the address space during boot when many long-lived
4825 * kernel allocations are made.
4827 * bitmap is created for zone's valid pfn range. but memmap
4828 * can be created for invalid pages (for alignment)
4829 * check here not to call set_pageblock_migratetype() against
4832 if (!(pfn
& (pageblock_nr_pages
- 1))) {
4833 struct page
*page
= pfn_to_page(pfn
);
4835 __init_single_page(page
, pfn
, zone
, nid
);
4836 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
4838 __init_single_pfn(pfn
, zone
, nid
);
4843 static void __meminit
zone_init_free_lists(struct zone
*zone
)
4845 unsigned int order
, t
;
4846 for_each_migratetype_order(order
, t
) {
4847 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
4848 zone
->free_area
[order
].nr_free
= 0;
4852 #ifndef __HAVE_ARCH_MEMMAP_INIT
4853 #define memmap_init(size, nid, zone, start_pfn) \
4854 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4857 static int zone_batchsize(struct zone
*zone
)
4863 * The per-cpu-pages pools are set to around 1000th of the
4864 * size of the zone. But no more than 1/2 of a meg.
4866 * OK, so we don't know how big the cache is. So guess.
4868 batch
= zone
->managed_pages
/ 1024;
4869 if (batch
* PAGE_SIZE
> 512 * 1024)
4870 batch
= (512 * 1024) / PAGE_SIZE
;
4871 batch
/= 4; /* We effectively *= 4 below */
4876 * Clamp the batch to a 2^n - 1 value. Having a power
4877 * of 2 value was found to be more likely to have
4878 * suboptimal cache aliasing properties in some cases.
4880 * For example if 2 tasks are alternately allocating
4881 * batches of pages, one task can end up with a lot
4882 * of pages of one half of the possible page colors
4883 * and the other with pages of the other colors.
4885 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
4890 /* The deferral and batching of frees should be suppressed under NOMMU
4893 * The problem is that NOMMU needs to be able to allocate large chunks
4894 * of contiguous memory as there's no hardware page translation to
4895 * assemble apparent contiguous memory from discontiguous pages.
4897 * Queueing large contiguous runs of pages for batching, however,
4898 * causes the pages to actually be freed in smaller chunks. As there
4899 * can be a significant delay between the individual batches being
4900 * recycled, this leads to the once large chunks of space being
4901 * fragmented and becoming unavailable for high-order allocations.
4908 * pcp->high and pcp->batch values are related and dependent on one another:
4909 * ->batch must never be higher then ->high.
4910 * The following function updates them in a safe manner without read side
4913 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4914 * those fields changing asynchronously (acording the the above rule).
4916 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4917 * outside of boot time (or some other assurance that no concurrent updaters
4920 static void pageset_update(struct per_cpu_pages
*pcp
, unsigned long high
,
4921 unsigned long batch
)
4923 /* start with a fail safe value for batch */
4927 /* Update high, then batch, in order */
4934 /* a companion to pageset_set_high() */
4935 static void pageset_set_batch(struct per_cpu_pageset
*p
, unsigned long batch
)
4937 pageset_update(&p
->pcp
, 6 * batch
, max(1UL, 1 * batch
));
4940 static void pageset_init(struct per_cpu_pageset
*p
)
4942 struct per_cpu_pages
*pcp
;
4945 memset(p
, 0, sizeof(*p
));
4949 for (migratetype
= 0; migratetype
< MIGRATE_PCPTYPES
; migratetype
++)
4950 INIT_LIST_HEAD(&pcp
->lists
[migratetype
]);
4953 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
4956 pageset_set_batch(p
, batch
);
4960 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4961 * to the value high for the pageset p.
4963 static void pageset_set_high(struct per_cpu_pageset
*p
,
4966 unsigned long batch
= max(1UL, high
/ 4);
4967 if ((high
/ 4) > (PAGE_SHIFT
* 8))
4968 batch
= PAGE_SHIFT
* 8;
4970 pageset_update(&p
->pcp
, high
, batch
);
4973 static void pageset_set_high_and_batch(struct zone
*zone
,
4974 struct per_cpu_pageset
*pcp
)
4976 if (percpu_pagelist_fraction
)
4977 pageset_set_high(pcp
,
4978 (zone
->managed_pages
/
4979 percpu_pagelist_fraction
));
4981 pageset_set_batch(pcp
, zone_batchsize(zone
));
4984 static void __meminit
zone_pageset_init(struct zone
*zone
, int cpu
)
4986 struct per_cpu_pageset
*pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
4989 pageset_set_high_and_batch(zone
, pcp
);
4992 static void __meminit
setup_zone_pageset(struct zone
*zone
)
4995 zone
->pageset
= alloc_percpu(struct per_cpu_pageset
);
4996 for_each_possible_cpu(cpu
)
4997 zone_pageset_init(zone
, cpu
);
5001 * Allocate per cpu pagesets and initialize them.
5002 * Before this call only boot pagesets were available.
5004 void __init
setup_per_cpu_pageset(void)
5008 for_each_populated_zone(zone
)
5009 setup_zone_pageset(zone
);
5012 static noinline __init_refok
5013 int zone_wait_table_init(struct zone
*zone
, unsigned long zone_size_pages
)
5019 * The per-page waitqueue mechanism uses hashed waitqueues
5022 zone
->wait_table_hash_nr_entries
=
5023 wait_table_hash_nr_entries(zone_size_pages
);
5024 zone
->wait_table_bits
=
5025 wait_table_bits(zone
->wait_table_hash_nr_entries
);
5026 alloc_size
= zone
->wait_table_hash_nr_entries
5027 * sizeof(wait_queue_head_t
);
5029 if (!slab_is_available()) {
5030 zone
->wait_table
= (wait_queue_head_t
*)
5031 memblock_virt_alloc_node_nopanic(
5032 alloc_size
, zone
->zone_pgdat
->node_id
);
5035 * This case means that a zone whose size was 0 gets new memory
5036 * via memory hot-add.
5037 * But it may be the case that a new node was hot-added. In
5038 * this case vmalloc() will not be able to use this new node's
5039 * memory - this wait_table must be initialized to use this new
5040 * node itself as well.
5041 * To use this new node's memory, further consideration will be
5044 zone
->wait_table
= vmalloc(alloc_size
);
5046 if (!zone
->wait_table
)
5049 for (i
= 0; i
< zone
->wait_table_hash_nr_entries
; ++i
)
5050 init_waitqueue_head(zone
->wait_table
+ i
);
5055 static __meminit
void zone_pcp_init(struct zone
*zone
)
5058 * per cpu subsystem is not up at this point. The following code
5059 * relies on the ability of the linker to provide the
5060 * offset of a (static) per cpu variable into the per cpu area.
5062 zone
->pageset
= &boot_pageset
;
5064 if (populated_zone(zone
))
5065 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%u\n",
5066 zone
->name
, zone
->present_pages
,
5067 zone_batchsize(zone
));
5070 int __meminit
init_currently_empty_zone(struct zone
*zone
,
5071 unsigned long zone_start_pfn
,
5074 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
5076 ret
= zone_wait_table_init(zone
, size
);
5079 pgdat
->nr_zones
= zone_idx(zone
) + 1;
5081 zone
->zone_start_pfn
= zone_start_pfn
;
5083 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
5084 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5086 (unsigned long)zone_idx(zone
),
5087 zone_start_pfn
, (zone_start_pfn
+ size
));
5089 zone_init_free_lists(zone
);
5094 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5095 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5098 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5100 int __meminit
__early_pfn_to_nid(unsigned long pfn
,
5101 struct mminit_pfnnid_cache
*state
)
5103 unsigned long start_pfn
, end_pfn
;
5106 if (state
->last_start
<= pfn
&& pfn
< state
->last_end
)
5107 return state
->last_nid
;
5109 nid
= memblock_search_pfn_nid(pfn
, &start_pfn
, &end_pfn
);
5111 state
->last_start
= start_pfn
;
5112 state
->last_end
= end_pfn
;
5113 state
->last_nid
= nid
;
5118 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5121 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5122 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5123 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5125 * If an architecture guarantees that all ranges registered contain no holes
5126 * and may be freed, this this function may be used instead of calling
5127 * memblock_free_early_nid() manually.
5129 void __init
free_bootmem_with_active_regions(int nid
, unsigned long max_low_pfn
)
5131 unsigned long start_pfn
, end_pfn
;
5134 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
) {
5135 start_pfn
= min(start_pfn
, max_low_pfn
);
5136 end_pfn
= min(end_pfn
, max_low_pfn
);
5138 if (start_pfn
< end_pfn
)
5139 memblock_free_early_nid(PFN_PHYS(start_pfn
),
5140 (end_pfn
- start_pfn
) << PAGE_SHIFT
,
5146 * sparse_memory_present_with_active_regions - Call memory_present for each active range
5147 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
5149 * If an architecture guarantees that all ranges registered contain no holes and may
5150 * be freed, this function may be used instead of calling memory_present() manually.
5152 void __init
sparse_memory_present_with_active_regions(int nid
)
5154 unsigned long start_pfn
, end_pfn
;
5157 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
)
5158 memory_present(this_nid
, start_pfn
, end_pfn
);
5162 * get_pfn_range_for_nid - Return the start and end page frames for a node
5163 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5164 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5165 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5167 * It returns the start and end page frame of a node based on information
5168 * provided by memblock_set_node(). If called for a node
5169 * with no available memory, a warning is printed and the start and end
5172 void __meminit
get_pfn_range_for_nid(unsigned int nid
,
5173 unsigned long *start_pfn
, unsigned long *end_pfn
)
5175 unsigned long this_start_pfn
, this_end_pfn
;
5181 for_each_mem_pfn_range(i
, nid
, &this_start_pfn
, &this_end_pfn
, NULL
) {
5182 *start_pfn
= min(*start_pfn
, this_start_pfn
);
5183 *end_pfn
= max(*end_pfn
, this_end_pfn
);
5186 if (*start_pfn
== -1UL)
5191 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5192 * assumption is made that zones within a node are ordered in monotonic
5193 * increasing memory addresses so that the "highest" populated zone is used
5195 static void __init
find_usable_zone_for_movable(void)
5198 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
5199 if (zone_index
== ZONE_MOVABLE
)
5202 if (arch_zone_highest_possible_pfn
[zone_index
] >
5203 arch_zone_lowest_possible_pfn
[zone_index
])
5207 VM_BUG_ON(zone_index
== -1);
5208 movable_zone
= zone_index
;
5212 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5213 * because it is sized independent of architecture. Unlike the other zones,
5214 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5215 * in each node depending on the size of each node and how evenly kernelcore
5216 * is distributed. This helper function adjusts the zone ranges
5217 * provided by the architecture for a given node by using the end of the
5218 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5219 * zones within a node are in order of monotonic increases memory addresses
5221 static void __meminit
adjust_zone_range_for_zone_movable(int nid
,
5222 unsigned long zone_type
,
5223 unsigned long node_start_pfn
,
5224 unsigned long node_end_pfn
,
5225 unsigned long *zone_start_pfn
,
5226 unsigned long *zone_end_pfn
)
5228 /* Only adjust if ZONE_MOVABLE is on this node */
5229 if (zone_movable_pfn
[nid
]) {
5230 /* Size ZONE_MOVABLE */
5231 if (zone_type
== ZONE_MOVABLE
) {
5232 *zone_start_pfn
= zone_movable_pfn
[nid
];
5233 *zone_end_pfn
= min(node_end_pfn
,
5234 arch_zone_highest_possible_pfn
[movable_zone
]);
5236 /* Check if this whole range is within ZONE_MOVABLE */
5237 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
5238 *zone_start_pfn
= *zone_end_pfn
;
5243 * Return the number of pages a zone spans in a node, including holes
5244 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5246 static unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
5247 unsigned long zone_type
,
5248 unsigned long node_start_pfn
,
5249 unsigned long node_end_pfn
,
5250 unsigned long *zone_start_pfn
,
5251 unsigned long *zone_end_pfn
,
5252 unsigned long *ignored
)
5254 /* When hotadd a new node from cpu_up(), the node should be empty */
5255 if (!node_start_pfn
&& !node_end_pfn
)
5258 /* Get the start and end of the zone */
5259 *zone_start_pfn
= arch_zone_lowest_possible_pfn
[zone_type
];
5260 *zone_end_pfn
= arch_zone_highest_possible_pfn
[zone_type
];
5261 adjust_zone_range_for_zone_movable(nid
, zone_type
,
5262 node_start_pfn
, node_end_pfn
,
5263 zone_start_pfn
, zone_end_pfn
);
5265 /* Check that this node has pages within the zone's required range */
5266 if (*zone_end_pfn
< node_start_pfn
|| *zone_start_pfn
> node_end_pfn
)
5269 /* Move the zone boundaries inside the node if necessary */
5270 *zone_end_pfn
= min(*zone_end_pfn
, node_end_pfn
);
5271 *zone_start_pfn
= max(*zone_start_pfn
, node_start_pfn
);
5273 /* Return the spanned pages */
5274 return *zone_end_pfn
- *zone_start_pfn
;
5278 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5279 * then all holes in the requested range will be accounted for.
5281 unsigned long __meminit
__absent_pages_in_range(int nid
,
5282 unsigned long range_start_pfn
,
5283 unsigned long range_end_pfn
)
5285 unsigned long nr_absent
= range_end_pfn
- range_start_pfn
;
5286 unsigned long start_pfn
, end_pfn
;
5289 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
5290 start_pfn
= clamp(start_pfn
, range_start_pfn
, range_end_pfn
);
5291 end_pfn
= clamp(end_pfn
, range_start_pfn
, range_end_pfn
);
5292 nr_absent
-= end_pfn
- start_pfn
;
5298 * absent_pages_in_range - Return number of page frames in holes within a range
5299 * @start_pfn: The start PFN to start searching for holes
5300 * @end_pfn: The end PFN to stop searching for holes
5302 * It returns the number of pages frames in memory holes within a range.
5304 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
5305 unsigned long end_pfn
)
5307 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
5310 /* Return the number of page frames in holes in a zone on a node */
5311 static unsigned long __meminit
zone_absent_pages_in_node(int nid
,
5312 unsigned long zone_type
,
5313 unsigned long node_start_pfn
,
5314 unsigned long node_end_pfn
,
5315 unsigned long *ignored
)
5317 unsigned long zone_low
= arch_zone_lowest_possible_pfn
[zone_type
];
5318 unsigned long zone_high
= arch_zone_highest_possible_pfn
[zone_type
];
5319 unsigned long zone_start_pfn
, zone_end_pfn
;
5320 unsigned long nr_absent
;
5322 /* When hotadd a new node from cpu_up(), the node should be empty */
5323 if (!node_start_pfn
&& !node_end_pfn
)
5326 zone_start_pfn
= clamp(node_start_pfn
, zone_low
, zone_high
);
5327 zone_end_pfn
= clamp(node_end_pfn
, zone_low
, zone_high
);
5329 adjust_zone_range_for_zone_movable(nid
, zone_type
,
5330 node_start_pfn
, node_end_pfn
,
5331 &zone_start_pfn
, &zone_end_pfn
);
5332 nr_absent
= __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
5335 * ZONE_MOVABLE handling.
5336 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
5339 if (zone_movable_pfn
[nid
]) {
5340 if (mirrored_kernelcore
) {
5341 unsigned long start_pfn
, end_pfn
;
5342 struct memblock_region
*r
;
5344 for_each_memblock(memory
, r
) {
5345 start_pfn
= clamp(memblock_region_memory_base_pfn(r
),
5346 zone_start_pfn
, zone_end_pfn
);
5347 end_pfn
= clamp(memblock_region_memory_end_pfn(r
),
5348 zone_start_pfn
, zone_end_pfn
);
5350 if (zone_type
== ZONE_MOVABLE
&&
5351 memblock_is_mirror(r
))
5352 nr_absent
+= end_pfn
- start_pfn
;
5354 if (zone_type
== ZONE_NORMAL
&&
5355 !memblock_is_mirror(r
))
5356 nr_absent
+= end_pfn
- start_pfn
;
5359 if (zone_type
== ZONE_NORMAL
)
5360 nr_absent
+= node_end_pfn
- zone_movable_pfn
[nid
];
5367 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5368 static inline unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
5369 unsigned long zone_type
,
5370 unsigned long node_start_pfn
,
5371 unsigned long node_end_pfn
,
5372 unsigned long *zone_start_pfn
,
5373 unsigned long *zone_end_pfn
,
5374 unsigned long *zones_size
)
5378 *zone_start_pfn
= node_start_pfn
;
5379 for (zone
= 0; zone
< zone_type
; zone
++)
5380 *zone_start_pfn
+= zones_size
[zone
];
5382 *zone_end_pfn
= *zone_start_pfn
+ zones_size
[zone_type
];
5384 return zones_size
[zone_type
];
5387 static inline unsigned long __meminit
zone_absent_pages_in_node(int nid
,
5388 unsigned long zone_type
,
5389 unsigned long node_start_pfn
,
5390 unsigned long node_end_pfn
,
5391 unsigned long *zholes_size
)
5396 return zholes_size
[zone_type
];
5399 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5401 static void __meminit
calculate_node_totalpages(struct pglist_data
*pgdat
,
5402 unsigned long node_start_pfn
,
5403 unsigned long node_end_pfn
,
5404 unsigned long *zones_size
,
5405 unsigned long *zholes_size
)
5407 unsigned long realtotalpages
= 0, totalpages
= 0;
5410 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5411 struct zone
*zone
= pgdat
->node_zones
+ i
;
5412 unsigned long zone_start_pfn
, zone_end_pfn
;
5413 unsigned long size
, real_size
;
5415 size
= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
5421 real_size
= size
- zone_absent_pages_in_node(pgdat
->node_id
, i
,
5422 node_start_pfn
, node_end_pfn
,
5425 zone
->zone_start_pfn
= zone_start_pfn
;
5427 zone
->zone_start_pfn
= 0;
5428 zone
->spanned_pages
= size
;
5429 zone
->present_pages
= real_size
;
5432 realtotalpages
+= real_size
;
5435 pgdat
->node_spanned_pages
= totalpages
;
5436 pgdat
->node_present_pages
= realtotalpages
;
5437 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
5441 #ifndef CONFIG_SPARSEMEM
5443 * Calculate the size of the zone->blockflags rounded to an unsigned long
5444 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5445 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5446 * round what is now in bits to nearest long in bits, then return it in
5449 static unsigned long __init
usemap_size(unsigned long zone_start_pfn
, unsigned long zonesize
)
5451 unsigned long usemapsize
;
5453 zonesize
+= zone_start_pfn
& (pageblock_nr_pages
-1);
5454 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
5455 usemapsize
= usemapsize
>> pageblock_order
;
5456 usemapsize
*= NR_PAGEBLOCK_BITS
;
5457 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
5459 return usemapsize
/ 8;
5462 static void __init
setup_usemap(struct pglist_data
*pgdat
,
5464 unsigned long zone_start_pfn
,
5465 unsigned long zonesize
)
5467 unsigned long usemapsize
= usemap_size(zone_start_pfn
, zonesize
);
5468 zone
->pageblock_flags
= NULL
;
5470 zone
->pageblock_flags
=
5471 memblock_virt_alloc_node_nopanic(usemapsize
,
5475 static inline void setup_usemap(struct pglist_data
*pgdat
, struct zone
*zone
,
5476 unsigned long zone_start_pfn
, unsigned long zonesize
) {}
5477 #endif /* CONFIG_SPARSEMEM */
5479 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5481 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5482 void __paginginit
set_pageblock_order(void)
5486 /* Check that pageblock_nr_pages has not already been setup */
5487 if (pageblock_order
)
5490 if (HPAGE_SHIFT
> PAGE_SHIFT
)
5491 order
= HUGETLB_PAGE_ORDER
;
5493 order
= MAX_ORDER
- 1;
5496 * Assume the largest contiguous order of interest is a huge page.
5497 * This value may be variable depending on boot parameters on IA64 and
5500 pageblock_order
= order
;
5502 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5505 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5506 * is unused as pageblock_order is set at compile-time. See
5507 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5510 void __paginginit
set_pageblock_order(void)
5514 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5516 static unsigned long __paginginit
calc_memmap_size(unsigned long spanned_pages
,
5517 unsigned long present_pages
)
5519 unsigned long pages
= spanned_pages
;
5522 * Provide a more accurate estimation if there are holes within
5523 * the zone and SPARSEMEM is in use. If there are holes within the
5524 * zone, each populated memory region may cost us one or two extra
5525 * memmap pages due to alignment because memmap pages for each
5526 * populated regions may not naturally algined on page boundary.
5527 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5529 if (spanned_pages
> present_pages
+ (present_pages
>> 4) &&
5530 IS_ENABLED(CONFIG_SPARSEMEM
))
5531 pages
= present_pages
;
5533 return PAGE_ALIGN(pages
* sizeof(struct page
)) >> PAGE_SHIFT
;
5537 * Set up the zone data structures:
5538 * - mark all pages reserved
5539 * - mark all memory queues empty
5540 * - clear the memory bitmaps
5542 * NOTE: pgdat should get zeroed by caller.
5544 static void __paginginit
free_area_init_core(struct pglist_data
*pgdat
)
5547 int nid
= pgdat
->node_id
;
5550 pgdat_resize_init(pgdat
);
5551 #ifdef CONFIG_NUMA_BALANCING
5552 spin_lock_init(&pgdat
->numabalancing_migrate_lock
);
5553 pgdat
->numabalancing_migrate_nr_pages
= 0;
5554 pgdat
->numabalancing_migrate_next_window
= jiffies
;
5556 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5557 spin_lock_init(&pgdat
->split_queue_lock
);
5558 INIT_LIST_HEAD(&pgdat
->split_queue
);
5559 pgdat
->split_queue_len
= 0;
5561 init_waitqueue_head(&pgdat
->kswapd_wait
);
5562 init_waitqueue_head(&pgdat
->pfmemalloc_wait
);
5563 #ifdef CONFIG_COMPACTION
5564 init_waitqueue_head(&pgdat
->kcompactd_wait
);
5566 pgdat_page_ext_init(pgdat
);
5568 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
5569 struct zone
*zone
= pgdat
->node_zones
+ j
;
5570 unsigned long size
, realsize
, freesize
, memmap_pages
;
5571 unsigned long zone_start_pfn
= zone
->zone_start_pfn
;
5573 size
= zone
->spanned_pages
;
5574 realsize
= freesize
= zone
->present_pages
;
5577 * Adjust freesize so that it accounts for how much memory
5578 * is used by this zone for memmap. This affects the watermark
5579 * and per-cpu initialisations
5581 memmap_pages
= calc_memmap_size(size
, realsize
);
5582 if (!is_highmem_idx(j
)) {
5583 if (freesize
>= memmap_pages
) {
5584 freesize
-= memmap_pages
;
5587 " %s zone: %lu pages used for memmap\n",
5588 zone_names
[j
], memmap_pages
);
5590 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
5591 zone_names
[j
], memmap_pages
, freesize
);
5594 /* Account for reserved pages */
5595 if (j
== 0 && freesize
> dma_reserve
) {
5596 freesize
-= dma_reserve
;
5597 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
5598 zone_names
[0], dma_reserve
);
5601 if (!is_highmem_idx(j
))
5602 nr_kernel_pages
+= freesize
;
5603 /* Charge for highmem memmap if there are enough kernel pages */
5604 else if (nr_kernel_pages
> memmap_pages
* 2)
5605 nr_kernel_pages
-= memmap_pages
;
5606 nr_all_pages
+= freesize
;
5609 * Set an approximate value for lowmem here, it will be adjusted
5610 * when the bootmem allocator frees pages into the buddy system.
5611 * And all highmem pages will be managed by the buddy system.
5613 zone
->managed_pages
= is_highmem_idx(j
) ? realsize
: freesize
;
5616 zone
->min_unmapped_pages
= (freesize
*sysctl_min_unmapped_ratio
)
5618 zone
->min_slab_pages
= (freesize
* sysctl_min_slab_ratio
) / 100;
5620 zone
->name
= zone_names
[j
];
5621 spin_lock_init(&zone
->lock
);
5622 spin_lock_init(&zone
->lru_lock
);
5623 zone_seqlock_init(zone
);
5624 zone
->zone_pgdat
= pgdat
;
5625 zone_pcp_init(zone
);
5627 /* For bootup, initialized properly in watermark setup */
5628 mod_zone_page_state(zone
, NR_ALLOC_BATCH
, zone
->managed_pages
);
5630 lruvec_init(&zone
->lruvec
);
5634 set_pageblock_order();
5635 setup_usemap(pgdat
, zone
, zone_start_pfn
, size
);
5636 ret
= init_currently_empty_zone(zone
, zone_start_pfn
, size
);
5638 memmap_init(size
, nid
, j
, zone_start_pfn
);
5642 static void __init_refok
alloc_node_mem_map(struct pglist_data
*pgdat
)
5644 unsigned long __maybe_unused start
= 0;
5645 unsigned long __maybe_unused offset
= 0;
5647 /* Skip empty nodes */
5648 if (!pgdat
->node_spanned_pages
)
5651 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5652 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
5653 offset
= pgdat
->node_start_pfn
- start
;
5654 /* ia64 gets its own node_mem_map, before this, without bootmem */
5655 if (!pgdat
->node_mem_map
) {
5656 unsigned long size
, end
;
5660 * The zone's endpoints aren't required to be MAX_ORDER
5661 * aligned but the node_mem_map endpoints must be in order
5662 * for the buddy allocator to function correctly.
5664 end
= pgdat_end_pfn(pgdat
);
5665 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
5666 size
= (end
- start
) * sizeof(struct page
);
5667 map
= alloc_remap(pgdat
->node_id
, size
);
5669 map
= memblock_virt_alloc_node_nopanic(size
,
5671 pgdat
->node_mem_map
= map
+ offset
;
5673 #ifndef CONFIG_NEED_MULTIPLE_NODES
5675 * With no DISCONTIG, the global mem_map is just set as node 0's
5677 if (pgdat
== NODE_DATA(0)) {
5678 mem_map
= NODE_DATA(0)->node_mem_map
;
5679 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
5680 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
5682 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5685 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
5688 void __paginginit
free_area_init_node(int nid
, unsigned long *zones_size
,
5689 unsigned long node_start_pfn
, unsigned long *zholes_size
)
5691 pg_data_t
*pgdat
= NODE_DATA(nid
);
5692 unsigned long start_pfn
= 0;
5693 unsigned long end_pfn
= 0;
5695 /* pg_data_t should be reset to zero when it's allocated */
5696 WARN_ON(pgdat
->nr_zones
|| pgdat
->classzone_idx
);
5698 reset_deferred_meminit(pgdat
);
5699 pgdat
->node_id
= nid
;
5700 pgdat
->node_start_pfn
= node_start_pfn
;
5701 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5702 get_pfn_range_for_nid(nid
, &start_pfn
, &end_pfn
);
5703 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid
,
5704 (u64
)start_pfn
<< PAGE_SHIFT
,
5705 end_pfn
? ((u64
)end_pfn
<< PAGE_SHIFT
) - 1 : 0);
5707 start_pfn
= node_start_pfn
;
5709 calculate_node_totalpages(pgdat
, start_pfn
, end_pfn
,
5710 zones_size
, zholes_size
);
5712 alloc_node_mem_map(pgdat
);
5713 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5714 printk(KERN_DEBUG
"free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5715 nid
, (unsigned long)pgdat
,
5716 (unsigned long)pgdat
->node_mem_map
);
5719 free_area_init_core(pgdat
);
5722 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5724 #if MAX_NUMNODES > 1
5726 * Figure out the number of possible node ids.
5728 void __init
setup_nr_node_ids(void)
5730 unsigned int highest
;
5732 highest
= find_last_bit(node_possible_map
.bits
, MAX_NUMNODES
);
5733 nr_node_ids
= highest
+ 1;
5738 * node_map_pfn_alignment - determine the maximum internode alignment
5740 * This function should be called after node map is populated and sorted.
5741 * It calculates the maximum power of two alignment which can distinguish
5744 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5745 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5746 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5747 * shifted, 1GiB is enough and this function will indicate so.
5749 * This is used to test whether pfn -> nid mapping of the chosen memory
5750 * model has fine enough granularity to avoid incorrect mapping for the
5751 * populated node map.
5753 * Returns the determined alignment in pfn's. 0 if there is no alignment
5754 * requirement (single node).
5756 unsigned long __init
node_map_pfn_alignment(void)
5758 unsigned long accl_mask
= 0, last_end
= 0;
5759 unsigned long start
, end
, mask
;
5763 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start
, &end
, &nid
) {
5764 if (!start
|| last_nid
< 0 || last_nid
== nid
) {
5771 * Start with a mask granular enough to pin-point to the
5772 * start pfn and tick off bits one-by-one until it becomes
5773 * too coarse to separate the current node from the last.
5775 mask
= ~((1 << __ffs(start
)) - 1);
5776 while (mask
&& last_end
<= (start
& (mask
<< 1)))
5779 /* accumulate all internode masks */
5783 /* convert mask to number of pages */
5784 return ~accl_mask
+ 1;
5787 /* Find the lowest pfn for a node */
5788 static unsigned long __init
find_min_pfn_for_node(int nid
)
5790 unsigned long min_pfn
= ULONG_MAX
;
5791 unsigned long start_pfn
;
5794 for_each_mem_pfn_range(i
, nid
, &start_pfn
, NULL
, NULL
)
5795 min_pfn
= min(min_pfn
, start_pfn
);
5797 if (min_pfn
== ULONG_MAX
) {
5798 pr_warn("Could not find start_pfn for node %d\n", nid
);
5806 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5808 * It returns the minimum PFN based on information provided via
5809 * memblock_set_node().
5811 unsigned long __init
find_min_pfn_with_active_regions(void)
5813 return find_min_pfn_for_node(MAX_NUMNODES
);
5817 * early_calculate_totalpages()
5818 * Sum pages in active regions for movable zone.
5819 * Populate N_MEMORY for calculating usable_nodes.
5821 static unsigned long __init
early_calculate_totalpages(void)
5823 unsigned long totalpages
= 0;
5824 unsigned long start_pfn
, end_pfn
;
5827 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
5828 unsigned long pages
= end_pfn
- start_pfn
;
5830 totalpages
+= pages
;
5832 node_set_state(nid
, N_MEMORY
);
5838 * Find the PFN the Movable zone begins in each node. Kernel memory
5839 * is spread evenly between nodes as long as the nodes have enough
5840 * memory. When they don't, some nodes will have more kernelcore than
5843 static void __init
find_zone_movable_pfns_for_nodes(void)
5846 unsigned long usable_startpfn
;
5847 unsigned long kernelcore_node
, kernelcore_remaining
;
5848 /* save the state before borrow the nodemask */
5849 nodemask_t saved_node_state
= node_states
[N_MEMORY
];
5850 unsigned long totalpages
= early_calculate_totalpages();
5851 int usable_nodes
= nodes_weight(node_states
[N_MEMORY
]);
5852 struct memblock_region
*r
;
5854 /* Need to find movable_zone earlier when movable_node is specified. */
5855 find_usable_zone_for_movable();
5858 * If movable_node is specified, ignore kernelcore and movablecore
5861 if (movable_node_is_enabled()) {
5862 for_each_memblock(memory
, r
) {
5863 if (!memblock_is_hotpluggable(r
))
5868 usable_startpfn
= PFN_DOWN(r
->base
);
5869 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
5870 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
5878 * If kernelcore=mirror is specified, ignore movablecore option
5880 if (mirrored_kernelcore
) {
5881 bool mem_below_4gb_not_mirrored
= false;
5883 for_each_memblock(memory
, r
) {
5884 if (memblock_is_mirror(r
))
5889 usable_startpfn
= memblock_region_memory_base_pfn(r
);
5891 if (usable_startpfn
< 0x100000) {
5892 mem_below_4gb_not_mirrored
= true;
5896 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
5897 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
5901 if (mem_below_4gb_not_mirrored
)
5902 pr_warn("This configuration results in unmirrored kernel memory.");
5908 * If movablecore=nn[KMG] was specified, calculate what size of
5909 * kernelcore that corresponds so that memory usable for
5910 * any allocation type is evenly spread. If both kernelcore
5911 * and movablecore are specified, then the value of kernelcore
5912 * will be used for required_kernelcore if it's greater than
5913 * what movablecore would have allowed.
5915 if (required_movablecore
) {
5916 unsigned long corepages
;
5919 * Round-up so that ZONE_MOVABLE is at least as large as what
5920 * was requested by the user
5922 required_movablecore
=
5923 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
5924 required_movablecore
= min(totalpages
, required_movablecore
);
5925 corepages
= totalpages
- required_movablecore
;
5927 required_kernelcore
= max(required_kernelcore
, corepages
);
5931 * If kernelcore was not specified or kernelcore size is larger
5932 * than totalpages, there is no ZONE_MOVABLE.
5934 if (!required_kernelcore
|| required_kernelcore
>= totalpages
)
5937 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5938 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
5941 /* Spread kernelcore memory as evenly as possible throughout nodes */
5942 kernelcore_node
= required_kernelcore
/ usable_nodes
;
5943 for_each_node_state(nid
, N_MEMORY
) {
5944 unsigned long start_pfn
, end_pfn
;
5947 * Recalculate kernelcore_node if the division per node
5948 * now exceeds what is necessary to satisfy the requested
5949 * amount of memory for the kernel
5951 if (required_kernelcore
< kernelcore_node
)
5952 kernelcore_node
= required_kernelcore
/ usable_nodes
;
5955 * As the map is walked, we track how much memory is usable
5956 * by the kernel using kernelcore_remaining. When it is
5957 * 0, the rest of the node is usable by ZONE_MOVABLE
5959 kernelcore_remaining
= kernelcore_node
;
5961 /* Go through each range of PFNs within this node */
5962 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
5963 unsigned long size_pages
;
5965 start_pfn
= max(start_pfn
, zone_movable_pfn
[nid
]);
5966 if (start_pfn
>= end_pfn
)
5969 /* Account for what is only usable for kernelcore */
5970 if (start_pfn
< usable_startpfn
) {
5971 unsigned long kernel_pages
;
5972 kernel_pages
= min(end_pfn
, usable_startpfn
)
5975 kernelcore_remaining
-= min(kernel_pages
,
5976 kernelcore_remaining
);
5977 required_kernelcore
-= min(kernel_pages
,
5978 required_kernelcore
);
5980 /* Continue if range is now fully accounted */
5981 if (end_pfn
<= usable_startpfn
) {
5984 * Push zone_movable_pfn to the end so
5985 * that if we have to rebalance
5986 * kernelcore across nodes, we will
5987 * not double account here
5989 zone_movable_pfn
[nid
] = end_pfn
;
5992 start_pfn
= usable_startpfn
;
5996 * The usable PFN range for ZONE_MOVABLE is from
5997 * start_pfn->end_pfn. Calculate size_pages as the
5998 * number of pages used as kernelcore
6000 size_pages
= end_pfn
- start_pfn
;
6001 if (size_pages
> kernelcore_remaining
)
6002 size_pages
= kernelcore_remaining
;
6003 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
6006 * Some kernelcore has been met, update counts and
6007 * break if the kernelcore for this node has been
6010 required_kernelcore
-= min(required_kernelcore
,
6012 kernelcore_remaining
-= size_pages
;
6013 if (!kernelcore_remaining
)
6019 * If there is still required_kernelcore, we do another pass with one
6020 * less node in the count. This will push zone_movable_pfn[nid] further
6021 * along on the nodes that still have memory until kernelcore is
6025 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
6029 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
6030 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
6031 zone_movable_pfn
[nid
] =
6032 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
6035 /* restore the node_state */
6036 node_states
[N_MEMORY
] = saved_node_state
;
6039 /* Any regular or high memory on that node ? */
6040 static void check_for_memory(pg_data_t
*pgdat
, int nid
)
6042 enum zone_type zone_type
;
6044 if (N_MEMORY
== N_NORMAL_MEMORY
)
6047 for (zone_type
= 0; zone_type
<= ZONE_MOVABLE
- 1; zone_type
++) {
6048 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
6049 if (populated_zone(zone
)) {
6050 node_set_state(nid
, N_HIGH_MEMORY
);
6051 if (N_NORMAL_MEMORY
!= N_HIGH_MEMORY
&&
6052 zone_type
<= ZONE_NORMAL
)
6053 node_set_state(nid
, N_NORMAL_MEMORY
);
6060 * free_area_init_nodes - Initialise all pg_data_t and zone data
6061 * @max_zone_pfn: an array of max PFNs for each zone
6063 * This will call free_area_init_node() for each active node in the system.
6064 * Using the page ranges provided by memblock_set_node(), the size of each
6065 * zone in each node and their holes is calculated. If the maximum PFN
6066 * between two adjacent zones match, it is assumed that the zone is empty.
6067 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
6068 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
6069 * starts where the previous one ended. For example, ZONE_DMA32 starts
6070 * at arch_max_dma_pfn.
6072 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
6074 unsigned long start_pfn
, end_pfn
;
6077 /* Record where the zone boundaries are */
6078 memset(arch_zone_lowest_possible_pfn
, 0,
6079 sizeof(arch_zone_lowest_possible_pfn
));
6080 memset(arch_zone_highest_possible_pfn
, 0,
6081 sizeof(arch_zone_highest_possible_pfn
));
6082 arch_zone_lowest_possible_pfn
[0] = find_min_pfn_with_active_regions();
6083 arch_zone_highest_possible_pfn
[0] = max_zone_pfn
[0];
6084 for (i
= 1; i
< MAX_NR_ZONES
; i
++) {
6085 if (i
== ZONE_MOVABLE
)
6087 arch_zone_lowest_possible_pfn
[i
] =
6088 arch_zone_highest_possible_pfn
[i
-1];
6089 arch_zone_highest_possible_pfn
[i
] =
6090 max(max_zone_pfn
[i
], arch_zone_lowest_possible_pfn
[i
]);
6092 arch_zone_lowest_possible_pfn
[ZONE_MOVABLE
] = 0;
6093 arch_zone_highest_possible_pfn
[ZONE_MOVABLE
] = 0;
6095 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
6096 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
6097 find_zone_movable_pfns_for_nodes();
6099 /* Print out the zone ranges */
6100 pr_info("Zone ranges:\n");
6101 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
6102 if (i
== ZONE_MOVABLE
)
6104 pr_info(" %-8s ", zone_names
[i
]);
6105 if (arch_zone_lowest_possible_pfn
[i
] ==
6106 arch_zone_highest_possible_pfn
[i
])
6109 pr_cont("[mem %#018Lx-%#018Lx]\n",
6110 (u64
)arch_zone_lowest_possible_pfn
[i
]
6112 ((u64
)arch_zone_highest_possible_pfn
[i
]
6113 << PAGE_SHIFT
) - 1);
6116 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
6117 pr_info("Movable zone start for each node\n");
6118 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6119 if (zone_movable_pfn
[i
])
6120 pr_info(" Node %d: %#018Lx\n", i
,
6121 (u64
)zone_movable_pfn
[i
] << PAGE_SHIFT
);
6124 /* Print out the early node map */
6125 pr_info("Early memory node ranges\n");
6126 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
)
6127 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid
,
6128 (u64
)start_pfn
<< PAGE_SHIFT
,
6129 ((u64
)end_pfn
<< PAGE_SHIFT
) - 1);
6131 /* Initialise every node */
6132 mminit_verify_pageflags_layout();
6133 setup_nr_node_ids();
6134 for_each_online_node(nid
) {
6135 pg_data_t
*pgdat
= NODE_DATA(nid
);
6136 free_area_init_node(nid
, NULL
,
6137 find_min_pfn_for_node(nid
), NULL
);
6139 /* Any memory on that node */
6140 if (pgdat
->node_present_pages
)
6141 node_set_state(nid
, N_MEMORY
);
6142 check_for_memory(pgdat
, nid
);
6146 static int __init
cmdline_parse_core(char *p
, unsigned long *core
)
6148 unsigned long long coremem
;
6152 coremem
= memparse(p
, &p
);
6153 *core
= coremem
>> PAGE_SHIFT
;
6155 /* Paranoid check that UL is enough for the coremem value */
6156 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
6162 * kernelcore=size sets the amount of memory for use for allocations that
6163 * cannot be reclaimed or migrated.
6165 static int __init
cmdline_parse_kernelcore(char *p
)
6167 /* parse kernelcore=mirror */
6168 if (parse_option_str(p
, "mirror")) {
6169 mirrored_kernelcore
= true;
6173 return cmdline_parse_core(p
, &required_kernelcore
);
6177 * movablecore=size sets the amount of memory for use for allocations that
6178 * can be reclaimed or migrated.
6180 static int __init
cmdline_parse_movablecore(char *p
)
6182 return cmdline_parse_core(p
, &required_movablecore
);
6185 early_param("kernelcore", cmdline_parse_kernelcore
);
6186 early_param("movablecore", cmdline_parse_movablecore
);
6188 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6190 void adjust_managed_page_count(struct page
*page
, long count
)
6192 spin_lock(&managed_page_count_lock
);
6193 page_zone(page
)->managed_pages
+= count
;
6194 totalram_pages
+= count
;
6195 #ifdef CONFIG_HIGHMEM
6196 if (PageHighMem(page
))
6197 totalhigh_pages
+= count
;
6199 spin_unlock(&managed_page_count_lock
);
6201 EXPORT_SYMBOL(adjust_managed_page_count
);
6203 unsigned long free_reserved_area(void *start
, void *end
, int poison
, char *s
)
6206 unsigned long pages
= 0;
6208 start
= (void *)PAGE_ALIGN((unsigned long)start
);
6209 end
= (void *)((unsigned long)end
& PAGE_MASK
);
6210 for (pos
= start
; pos
< end
; pos
+= PAGE_SIZE
, pages
++) {
6211 if ((unsigned int)poison
<= 0xFF)
6212 memset(pos
, poison
, PAGE_SIZE
);
6213 free_reserved_page(virt_to_page(pos
));
6217 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
6218 s
, pages
<< (PAGE_SHIFT
- 10), start
, end
);
6222 EXPORT_SYMBOL(free_reserved_area
);
6224 #ifdef CONFIG_HIGHMEM
6225 void free_highmem_page(struct page
*page
)
6227 __free_reserved_page(page
);
6229 page_zone(page
)->managed_pages
++;
6235 void __init
mem_init_print_info(const char *str
)
6237 unsigned long physpages
, codesize
, datasize
, rosize
, bss_size
;
6238 unsigned long init_code_size
, init_data_size
;
6240 physpages
= get_num_physpages();
6241 codesize
= _etext
- _stext
;
6242 datasize
= _edata
- _sdata
;
6243 rosize
= __end_rodata
- __start_rodata
;
6244 bss_size
= __bss_stop
- __bss_start
;
6245 init_data_size
= __init_end
- __init_begin
;
6246 init_code_size
= _einittext
- _sinittext
;
6249 * Detect special cases and adjust section sizes accordingly:
6250 * 1) .init.* may be embedded into .data sections
6251 * 2) .init.text.* may be out of [__init_begin, __init_end],
6252 * please refer to arch/tile/kernel/vmlinux.lds.S.
6253 * 3) .rodata.* may be embedded into .text or .data sections.
6255 #define adj_init_size(start, end, size, pos, adj) \
6257 if (start <= pos && pos < end && size > adj) \
6261 adj_init_size(__init_begin
, __init_end
, init_data_size
,
6262 _sinittext
, init_code_size
);
6263 adj_init_size(_stext
, _etext
, codesize
, _sinittext
, init_code_size
);
6264 adj_init_size(_sdata
, _edata
, datasize
, __init_begin
, init_data_size
);
6265 adj_init_size(_stext
, _etext
, codesize
, __start_rodata
, rosize
);
6266 adj_init_size(_sdata
, _edata
, datasize
, __start_rodata
, rosize
);
6268 #undef adj_init_size
6270 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
6271 #ifdef CONFIG_HIGHMEM
6275 nr_free_pages() << (PAGE_SHIFT
- 10),
6276 physpages
<< (PAGE_SHIFT
- 10),
6277 codesize
>> 10, datasize
>> 10, rosize
>> 10,
6278 (init_data_size
+ init_code_size
) >> 10, bss_size
>> 10,
6279 (physpages
- totalram_pages
- totalcma_pages
) << (PAGE_SHIFT
- 10),
6280 totalcma_pages
<< (PAGE_SHIFT
- 10),
6281 #ifdef CONFIG_HIGHMEM
6282 totalhigh_pages
<< (PAGE_SHIFT
- 10),
6284 str
? ", " : "", str
? str
: "");
6288 * set_dma_reserve - set the specified number of pages reserved in the first zone
6289 * @new_dma_reserve: The number of pages to mark reserved
6291 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
6292 * In the DMA zone, a significant percentage may be consumed by kernel image
6293 * and other unfreeable allocations which can skew the watermarks badly. This
6294 * function may optionally be used to account for unfreeable pages in the
6295 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
6296 * smaller per-cpu batchsize.
6298 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
6300 dma_reserve
= new_dma_reserve
;
6303 void __init
free_area_init(unsigned long *zones_size
)
6305 free_area_init_node(0, zones_size
,
6306 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
6309 static int page_alloc_cpu_notify(struct notifier_block
*self
,
6310 unsigned long action
, void *hcpu
)
6312 int cpu
= (unsigned long)hcpu
;
6314 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
) {
6315 lru_add_drain_cpu(cpu
);
6319 * Spill the event counters of the dead processor
6320 * into the current processors event counters.
6321 * This artificially elevates the count of the current
6324 vm_events_fold_cpu(cpu
);
6327 * Zero the differential counters of the dead processor
6328 * so that the vm statistics are consistent.
6330 * This is only okay since the processor is dead and cannot
6331 * race with what we are doing.
6333 cpu_vm_stats_fold(cpu
);
6338 void __init
page_alloc_init(void)
6340 hotcpu_notifier(page_alloc_cpu_notify
, 0);
6344 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6345 * or min_free_kbytes changes.
6347 static void calculate_totalreserve_pages(void)
6349 struct pglist_data
*pgdat
;
6350 unsigned long reserve_pages
= 0;
6351 enum zone_type i
, j
;
6353 for_each_online_pgdat(pgdat
) {
6354 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
6355 struct zone
*zone
= pgdat
->node_zones
+ i
;
6358 /* Find valid and maximum lowmem_reserve in the zone */
6359 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
6360 if (zone
->lowmem_reserve
[j
] > max
)
6361 max
= zone
->lowmem_reserve
[j
];
6364 /* we treat the high watermark as reserved pages. */
6365 max
+= high_wmark_pages(zone
);
6367 if (max
> zone
->managed_pages
)
6368 max
= zone
->managed_pages
;
6370 zone
->totalreserve_pages
= max
;
6372 reserve_pages
+= max
;
6375 totalreserve_pages
= reserve_pages
;
6379 * setup_per_zone_lowmem_reserve - called whenever
6380 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6381 * has a correct pages reserved value, so an adequate number of
6382 * pages are left in the zone after a successful __alloc_pages().
6384 static void setup_per_zone_lowmem_reserve(void)
6386 struct pglist_data
*pgdat
;
6387 enum zone_type j
, idx
;
6389 for_each_online_pgdat(pgdat
) {
6390 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
6391 struct zone
*zone
= pgdat
->node_zones
+ j
;
6392 unsigned long managed_pages
= zone
->managed_pages
;
6394 zone
->lowmem_reserve
[j
] = 0;
6398 struct zone
*lower_zone
;
6402 if (sysctl_lowmem_reserve_ratio
[idx
] < 1)
6403 sysctl_lowmem_reserve_ratio
[idx
] = 1;
6405 lower_zone
= pgdat
->node_zones
+ idx
;
6406 lower_zone
->lowmem_reserve
[j
] = managed_pages
/
6407 sysctl_lowmem_reserve_ratio
[idx
];
6408 managed_pages
+= lower_zone
->managed_pages
;
6413 /* update totalreserve_pages */
6414 calculate_totalreserve_pages();
6417 static void __setup_per_zone_wmarks(void)
6419 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
6420 unsigned long lowmem_pages
= 0;
6422 unsigned long flags
;
6424 /* Calculate total number of !ZONE_HIGHMEM pages */
6425 for_each_zone(zone
) {
6426 if (!is_highmem(zone
))
6427 lowmem_pages
+= zone
->managed_pages
;
6430 for_each_zone(zone
) {
6433 spin_lock_irqsave(&zone
->lock
, flags
);
6434 tmp
= (u64
)pages_min
* zone
->managed_pages
;
6435 do_div(tmp
, lowmem_pages
);
6436 if (is_highmem(zone
)) {
6438 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6439 * need highmem pages, so cap pages_min to a small
6442 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6443 * deltas control asynch page reclaim, and so should
6444 * not be capped for highmem.
6446 unsigned long min_pages
;
6448 min_pages
= zone
->managed_pages
/ 1024;
6449 min_pages
= clamp(min_pages
, SWAP_CLUSTER_MAX
, 128UL);
6450 zone
->watermark
[WMARK_MIN
] = min_pages
;
6453 * If it's a lowmem zone, reserve a number of pages
6454 * proportionate to the zone's size.
6456 zone
->watermark
[WMARK_MIN
] = tmp
;
6460 * Set the kswapd watermarks distance according to the
6461 * scale factor in proportion to available memory, but
6462 * ensure a minimum size on small systems.
6464 tmp
= max_t(u64
, tmp
>> 2,
6465 mult_frac(zone
->managed_pages
,
6466 watermark_scale_factor
, 10000));
6468 zone
->watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + tmp
;
6469 zone
->watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + tmp
* 2;
6471 __mod_zone_page_state(zone
, NR_ALLOC_BATCH
,
6472 high_wmark_pages(zone
) - low_wmark_pages(zone
) -
6473 atomic_long_read(&zone
->vm_stat
[NR_ALLOC_BATCH
]));
6475 spin_unlock_irqrestore(&zone
->lock
, flags
);
6478 /* update totalreserve_pages */
6479 calculate_totalreserve_pages();
6483 * setup_per_zone_wmarks - called when min_free_kbytes changes
6484 * or when memory is hot-{added|removed}
6486 * Ensures that the watermark[min,low,high] values for each zone are set
6487 * correctly with respect to min_free_kbytes.
6489 void setup_per_zone_wmarks(void)
6491 mutex_lock(&zonelists_mutex
);
6492 __setup_per_zone_wmarks();
6493 mutex_unlock(&zonelists_mutex
);
6497 * The inactive anon list should be small enough that the VM never has to
6498 * do too much work, but large enough that each inactive page has a chance
6499 * to be referenced again before it is swapped out.
6501 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
6502 * INACTIVE_ANON pages on this zone's LRU, maintained by the
6503 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
6504 * the anonymous pages are kept on the inactive list.
6507 * memory ratio inactive anon
6508 * -------------------------------------
6517 static void __meminit
calculate_zone_inactive_ratio(struct zone
*zone
)
6519 unsigned int gb
, ratio
;
6521 /* Zone size in gigabytes */
6522 gb
= zone
->managed_pages
>> (30 - PAGE_SHIFT
);
6524 ratio
= int_sqrt(10 * gb
);
6528 zone
->inactive_ratio
= ratio
;
6531 static void __meminit
setup_per_zone_inactive_ratio(void)
6536 calculate_zone_inactive_ratio(zone
);
6540 * Initialise min_free_kbytes.
6542 * For small machines we want it small (128k min). For large machines
6543 * we want it large (64MB max). But it is not linear, because network
6544 * bandwidth does not increase linearly with machine size. We use
6546 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6547 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6563 int __meminit
init_per_zone_wmark_min(void)
6565 unsigned long lowmem_kbytes
;
6566 int new_min_free_kbytes
;
6568 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
6569 new_min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
6571 if (new_min_free_kbytes
> user_min_free_kbytes
) {
6572 min_free_kbytes
= new_min_free_kbytes
;
6573 if (min_free_kbytes
< 128)
6574 min_free_kbytes
= 128;
6575 if (min_free_kbytes
> 65536)
6576 min_free_kbytes
= 65536;
6578 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6579 new_min_free_kbytes
, user_min_free_kbytes
);
6581 setup_per_zone_wmarks();
6582 refresh_zone_stat_thresholds();
6583 setup_per_zone_lowmem_reserve();
6584 setup_per_zone_inactive_ratio();
6587 core_initcall(init_per_zone_wmark_min
)
6590 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6591 * that we can call two helper functions whenever min_free_kbytes
6594 int min_free_kbytes_sysctl_handler(struct ctl_table
*table
, int write
,
6595 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6599 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6604 user_min_free_kbytes
= min_free_kbytes
;
6605 setup_per_zone_wmarks();
6610 int watermark_scale_factor_sysctl_handler(struct ctl_table
*table
, int write
,
6611 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6615 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6620 setup_per_zone_wmarks();
6626 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
6627 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6632 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6637 zone
->min_unmapped_pages
= (zone
->managed_pages
*
6638 sysctl_min_unmapped_ratio
) / 100;
6642 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
6643 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6648 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6653 zone
->min_slab_pages
= (zone
->managed_pages
*
6654 sysctl_min_slab_ratio
) / 100;
6660 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6661 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6662 * whenever sysctl_lowmem_reserve_ratio changes.
6664 * The reserve ratio obviously has absolutely no relation with the
6665 * minimum watermarks. The lowmem reserve ratio can only make sense
6666 * if in function of the boot time zone sizes.
6668 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
6669 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6671 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6672 setup_per_zone_lowmem_reserve();
6677 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6678 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6679 * pagelist can have before it gets flushed back to buddy allocator.
6681 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table
*table
, int write
,
6682 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6685 int old_percpu_pagelist_fraction
;
6688 mutex_lock(&pcp_batch_high_lock
);
6689 old_percpu_pagelist_fraction
= percpu_pagelist_fraction
;
6691 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6692 if (!write
|| ret
< 0)
6695 /* Sanity checking to avoid pcp imbalance */
6696 if (percpu_pagelist_fraction
&&
6697 percpu_pagelist_fraction
< MIN_PERCPU_PAGELIST_FRACTION
) {
6698 percpu_pagelist_fraction
= old_percpu_pagelist_fraction
;
6704 if (percpu_pagelist_fraction
== old_percpu_pagelist_fraction
)
6707 for_each_populated_zone(zone
) {
6710 for_each_possible_cpu(cpu
)
6711 pageset_set_high_and_batch(zone
,
6712 per_cpu_ptr(zone
->pageset
, cpu
));
6715 mutex_unlock(&pcp_batch_high_lock
);
6720 int hashdist
= HASHDIST_DEFAULT
;
6722 static int __init
set_hashdist(char *str
)
6726 hashdist
= simple_strtoul(str
, &str
, 0);
6729 __setup("hashdist=", set_hashdist
);
6733 * allocate a large system hash table from bootmem
6734 * - it is assumed that the hash table must contain an exact power-of-2
6735 * quantity of entries
6736 * - limit is the number of hash buckets, not the total allocation size
6738 void *__init
alloc_large_system_hash(const char *tablename
,
6739 unsigned long bucketsize
,
6740 unsigned long numentries
,
6743 unsigned int *_hash_shift
,
6744 unsigned int *_hash_mask
,
6745 unsigned long low_limit
,
6746 unsigned long high_limit
)
6748 unsigned long long max
= high_limit
;
6749 unsigned long log2qty
, size
;
6752 /* allow the kernel cmdline to have a say */
6754 /* round applicable memory size up to nearest megabyte */
6755 numentries
= nr_kernel_pages
;
6757 /* It isn't necessary when PAGE_SIZE >= 1MB */
6758 if (PAGE_SHIFT
< 20)
6759 numentries
= round_up(numentries
, (1<<20)/PAGE_SIZE
);
6761 /* limit to 1 bucket per 2^scale bytes of low memory */
6762 if (scale
> PAGE_SHIFT
)
6763 numentries
>>= (scale
- PAGE_SHIFT
);
6765 numentries
<<= (PAGE_SHIFT
- scale
);
6767 /* Make sure we've got at least a 0-order allocation.. */
6768 if (unlikely(flags
& HASH_SMALL
)) {
6769 /* Makes no sense without HASH_EARLY */
6770 WARN_ON(!(flags
& HASH_EARLY
));
6771 if (!(numentries
>> *_hash_shift
)) {
6772 numentries
= 1UL << *_hash_shift
;
6773 BUG_ON(!numentries
);
6775 } else if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
6776 numentries
= PAGE_SIZE
/ bucketsize
;
6778 numentries
= roundup_pow_of_two(numentries
);
6780 /* limit allocation size to 1/16 total memory by default */
6782 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
6783 do_div(max
, bucketsize
);
6785 max
= min(max
, 0x80000000ULL
);
6787 if (numentries
< low_limit
)
6788 numentries
= low_limit
;
6789 if (numentries
> max
)
6792 log2qty
= ilog2(numentries
);
6795 size
= bucketsize
<< log2qty
;
6796 if (flags
& HASH_EARLY
)
6797 table
= memblock_virt_alloc_nopanic(size
, 0);
6799 table
= __vmalloc(size
, GFP_ATOMIC
, PAGE_KERNEL
);
6802 * If bucketsize is not a power-of-two, we may free
6803 * some pages at the end of hash table which
6804 * alloc_pages_exact() automatically does
6806 if (get_order(size
) < MAX_ORDER
) {
6807 table
= alloc_pages_exact(size
, GFP_ATOMIC
);
6808 kmemleak_alloc(table
, size
, 1, GFP_ATOMIC
);
6811 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
6814 panic("Failed to allocate %s hash table\n", tablename
);
6816 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
6817 tablename
, 1UL << log2qty
, ilog2(size
) - PAGE_SHIFT
, size
);
6820 *_hash_shift
= log2qty
;
6822 *_hash_mask
= (1 << log2qty
) - 1;
6827 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6828 static inline unsigned long *get_pageblock_bitmap(struct page
*page
,
6831 #ifdef CONFIG_SPARSEMEM
6832 return __pfn_to_section(pfn
)->pageblock_flags
;
6834 return page_zone(page
)->pageblock_flags
;
6835 #endif /* CONFIG_SPARSEMEM */
6838 static inline int pfn_to_bitidx(struct page
*page
, unsigned long pfn
)
6840 #ifdef CONFIG_SPARSEMEM
6841 pfn
&= (PAGES_PER_SECTION
-1);
6842 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
6844 pfn
= pfn
- round_down(page_zone(page
)->zone_start_pfn
, pageblock_nr_pages
);
6845 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
6846 #endif /* CONFIG_SPARSEMEM */
6850 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6851 * @page: The page within the block of interest
6852 * @pfn: The target page frame number
6853 * @end_bitidx: The last bit of interest to retrieve
6854 * @mask: mask of bits that the caller is interested in
6856 * Return: pageblock_bits flags
6858 unsigned long get_pfnblock_flags_mask(struct page
*page
, unsigned long pfn
,
6859 unsigned long end_bitidx
,
6862 unsigned long *bitmap
;
6863 unsigned long bitidx
, word_bitidx
;
6866 bitmap
= get_pageblock_bitmap(page
, pfn
);
6867 bitidx
= pfn_to_bitidx(page
, pfn
);
6868 word_bitidx
= bitidx
/ BITS_PER_LONG
;
6869 bitidx
&= (BITS_PER_LONG
-1);
6871 word
= bitmap
[word_bitidx
];
6872 bitidx
+= end_bitidx
;
6873 return (word
>> (BITS_PER_LONG
- bitidx
- 1)) & mask
;
6877 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6878 * @page: The page within the block of interest
6879 * @flags: The flags to set
6880 * @pfn: The target page frame number
6881 * @end_bitidx: The last bit of interest
6882 * @mask: mask of bits that the caller is interested in
6884 void set_pfnblock_flags_mask(struct page
*page
, unsigned long flags
,
6886 unsigned long end_bitidx
,
6889 unsigned long *bitmap
;
6890 unsigned long bitidx
, word_bitidx
;
6891 unsigned long old_word
, word
;
6893 BUILD_BUG_ON(NR_PAGEBLOCK_BITS
!= 4);
6895 bitmap
= get_pageblock_bitmap(page
, pfn
);
6896 bitidx
= pfn_to_bitidx(page
, pfn
);
6897 word_bitidx
= bitidx
/ BITS_PER_LONG
;
6898 bitidx
&= (BITS_PER_LONG
-1);
6900 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page
), pfn
), page
);
6902 bitidx
+= end_bitidx
;
6903 mask
<<= (BITS_PER_LONG
- bitidx
- 1);
6904 flags
<<= (BITS_PER_LONG
- bitidx
- 1);
6906 word
= READ_ONCE(bitmap
[word_bitidx
]);
6908 old_word
= cmpxchg(&bitmap
[word_bitidx
], word
, (word
& ~mask
) | flags
);
6909 if (word
== old_word
)
6916 * This function checks whether pageblock includes unmovable pages or not.
6917 * If @count is not zero, it is okay to include less @count unmovable pages
6919 * PageLRU check without isolation or lru_lock could race so that
6920 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6921 * expect this function should be exact.
6923 bool has_unmovable_pages(struct zone
*zone
, struct page
*page
, int count
,
6924 bool skip_hwpoisoned_pages
)
6926 unsigned long pfn
, iter
, found
;
6930 * For avoiding noise data, lru_add_drain_all() should be called
6931 * If ZONE_MOVABLE, the zone never contains unmovable pages
6933 if (zone_idx(zone
) == ZONE_MOVABLE
)
6935 mt
= get_pageblock_migratetype(page
);
6936 if (mt
== MIGRATE_MOVABLE
|| is_migrate_cma(mt
))
6939 pfn
= page_to_pfn(page
);
6940 for (found
= 0, iter
= 0; iter
< pageblock_nr_pages
; iter
++) {
6941 unsigned long check
= pfn
+ iter
;
6943 if (!pfn_valid_within(check
))
6946 page
= pfn_to_page(check
);
6949 * Hugepages are not in LRU lists, but they're movable.
6950 * We need not scan over tail pages bacause we don't
6951 * handle each tail page individually in migration.
6953 if (PageHuge(page
)) {
6954 iter
= round_up(iter
+ 1, 1<<compound_order(page
)) - 1;
6959 * We can't use page_count without pin a page
6960 * because another CPU can free compound page.
6961 * This check already skips compound tails of THP
6962 * because their page->_refcount is zero at all time.
6964 if (!page_ref_count(page
)) {
6965 if (PageBuddy(page
))
6966 iter
+= (1 << page_order(page
)) - 1;
6971 * The HWPoisoned page may be not in buddy system, and
6972 * page_count() is not 0.
6974 if (skip_hwpoisoned_pages
&& PageHWPoison(page
))
6980 * If there are RECLAIMABLE pages, we need to check
6981 * it. But now, memory offline itself doesn't call
6982 * shrink_node_slabs() and it still to be fixed.
6985 * If the page is not RAM, page_count()should be 0.
6986 * we don't need more check. This is an _used_ not-movable page.
6988 * The problematic thing here is PG_reserved pages. PG_reserved
6989 * is set to both of a memory hole page and a _used_ kernel
6998 bool is_pageblock_removable_nolock(struct page
*page
)
7004 * We have to be careful here because we are iterating over memory
7005 * sections which are not zone aware so we might end up outside of
7006 * the zone but still within the section.
7007 * We have to take care about the node as well. If the node is offline
7008 * its NODE_DATA will be NULL - see page_zone.
7010 if (!node_online(page_to_nid(page
)))
7013 zone
= page_zone(page
);
7014 pfn
= page_to_pfn(page
);
7015 if (!zone_spans_pfn(zone
, pfn
))
7018 return !has_unmovable_pages(zone
, page
, 0, true);
7021 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
7023 static unsigned long pfn_max_align_down(unsigned long pfn
)
7025 return pfn
& ~(max_t(unsigned long, MAX_ORDER_NR_PAGES
,
7026 pageblock_nr_pages
) - 1);
7029 static unsigned long pfn_max_align_up(unsigned long pfn
)
7031 return ALIGN(pfn
, max_t(unsigned long, MAX_ORDER_NR_PAGES
,
7032 pageblock_nr_pages
));
7035 /* [start, end) must belong to a single zone. */
7036 static int __alloc_contig_migrate_range(struct compact_control
*cc
,
7037 unsigned long start
, unsigned long end
)
7039 /* This function is based on compact_zone() from compaction.c. */
7040 unsigned long nr_reclaimed
;
7041 unsigned long pfn
= start
;
7042 unsigned int tries
= 0;
7047 while (pfn
< end
|| !list_empty(&cc
->migratepages
)) {
7048 if (fatal_signal_pending(current
)) {
7053 if (list_empty(&cc
->migratepages
)) {
7054 cc
->nr_migratepages
= 0;
7055 pfn
= isolate_migratepages_range(cc
, pfn
, end
);
7061 } else if (++tries
== 5) {
7062 ret
= ret
< 0 ? ret
: -EBUSY
;
7066 nr_reclaimed
= reclaim_clean_pages_from_list(cc
->zone
,
7068 cc
->nr_migratepages
-= nr_reclaimed
;
7070 ret
= migrate_pages(&cc
->migratepages
, alloc_migrate_target
,
7071 NULL
, 0, cc
->mode
, MR_CMA
);
7074 putback_movable_pages(&cc
->migratepages
);
7081 * alloc_contig_range() -- tries to allocate given range of pages
7082 * @start: start PFN to allocate
7083 * @end: one-past-the-last PFN to allocate
7084 * @migratetype: migratetype of the underlaying pageblocks (either
7085 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
7086 * in range must have the same migratetype and it must
7087 * be either of the two.
7089 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
7090 * aligned, however it's the caller's responsibility to guarantee that
7091 * we are the only thread that changes migrate type of pageblocks the
7094 * The PFN range must belong to a single zone.
7096 * Returns zero on success or negative error code. On success all
7097 * pages which PFN is in [start, end) are allocated for the caller and
7098 * need to be freed with free_contig_range().
7100 int alloc_contig_range(unsigned long start
, unsigned long end
,
7101 unsigned migratetype
)
7103 unsigned long outer_start
, outer_end
;
7107 struct compact_control cc
= {
7108 .nr_migratepages
= 0,
7110 .zone
= page_zone(pfn_to_page(start
)),
7111 .mode
= MIGRATE_SYNC
,
7112 .ignore_skip_hint
= true,
7114 INIT_LIST_HEAD(&cc
.migratepages
);
7117 * What we do here is we mark all pageblocks in range as
7118 * MIGRATE_ISOLATE. Because pageblock and max order pages may
7119 * have different sizes, and due to the way page allocator
7120 * work, we align the range to biggest of the two pages so
7121 * that page allocator won't try to merge buddies from
7122 * different pageblocks and change MIGRATE_ISOLATE to some
7123 * other migration type.
7125 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
7126 * migrate the pages from an unaligned range (ie. pages that
7127 * we are interested in). This will put all the pages in
7128 * range back to page allocator as MIGRATE_ISOLATE.
7130 * When this is done, we take the pages in range from page
7131 * allocator removing them from the buddy system. This way
7132 * page allocator will never consider using them.
7134 * This lets us mark the pageblocks back as
7135 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
7136 * aligned range but not in the unaligned, original range are
7137 * put back to page allocator so that buddy can use them.
7140 ret
= start_isolate_page_range(pfn_max_align_down(start
),
7141 pfn_max_align_up(end
), migratetype
,
7147 * In case of -EBUSY, we'd like to know which page causes problem.
7148 * So, just fall through. We will check it in test_pages_isolated().
7150 ret
= __alloc_contig_migrate_range(&cc
, start
, end
);
7151 if (ret
&& ret
!= -EBUSY
)
7155 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
7156 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
7157 * more, all pages in [start, end) are free in page allocator.
7158 * What we are going to do is to allocate all pages from
7159 * [start, end) (that is remove them from page allocator).
7161 * The only problem is that pages at the beginning and at the
7162 * end of interesting range may be not aligned with pages that
7163 * page allocator holds, ie. they can be part of higher order
7164 * pages. Because of this, we reserve the bigger range and
7165 * once this is done free the pages we are not interested in.
7167 * We don't have to hold zone->lock here because the pages are
7168 * isolated thus they won't get removed from buddy.
7171 lru_add_drain_all();
7172 drain_all_pages(cc
.zone
);
7175 outer_start
= start
;
7176 while (!PageBuddy(pfn_to_page(outer_start
))) {
7177 if (++order
>= MAX_ORDER
) {
7178 outer_start
= start
;
7181 outer_start
&= ~0UL << order
;
7184 if (outer_start
!= start
) {
7185 order
= page_order(pfn_to_page(outer_start
));
7188 * outer_start page could be small order buddy page and
7189 * it doesn't include start page. Adjust outer_start
7190 * in this case to report failed page properly
7191 * on tracepoint in test_pages_isolated()
7193 if (outer_start
+ (1UL << order
) <= start
)
7194 outer_start
= start
;
7197 /* Make sure the range is really isolated. */
7198 if (test_pages_isolated(outer_start
, end
, false)) {
7199 pr_info("%s: [%lx, %lx) PFNs busy\n",
7200 __func__
, outer_start
, end
);
7205 /* Grab isolated pages from freelists. */
7206 outer_end
= isolate_freepages_range(&cc
, outer_start
, end
);
7212 /* Free head and tail (if any) */
7213 if (start
!= outer_start
)
7214 free_contig_range(outer_start
, start
- outer_start
);
7215 if (end
!= outer_end
)
7216 free_contig_range(end
, outer_end
- end
);
7219 undo_isolate_page_range(pfn_max_align_down(start
),
7220 pfn_max_align_up(end
), migratetype
);
7224 void free_contig_range(unsigned long pfn
, unsigned nr_pages
)
7226 unsigned int count
= 0;
7228 for (; nr_pages
--; pfn
++) {
7229 struct page
*page
= pfn_to_page(pfn
);
7231 count
+= page_count(page
) != 1;
7234 WARN(count
!= 0, "%d pages are still in use!\n", count
);
7238 #ifdef CONFIG_MEMORY_HOTPLUG
7240 * The zone indicated has a new number of managed_pages; batch sizes and percpu
7241 * page high values need to be recalulated.
7243 void __meminit
zone_pcp_update(struct zone
*zone
)
7246 mutex_lock(&pcp_batch_high_lock
);
7247 for_each_possible_cpu(cpu
)
7248 pageset_set_high_and_batch(zone
,
7249 per_cpu_ptr(zone
->pageset
, cpu
));
7250 mutex_unlock(&pcp_batch_high_lock
);
7254 void zone_pcp_reset(struct zone
*zone
)
7256 unsigned long flags
;
7258 struct per_cpu_pageset
*pset
;
7260 /* avoid races with drain_pages() */
7261 local_irq_save(flags
);
7262 if (zone
->pageset
!= &boot_pageset
) {
7263 for_each_online_cpu(cpu
) {
7264 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
7265 drain_zonestat(zone
, pset
);
7267 free_percpu(zone
->pageset
);
7268 zone
->pageset
= &boot_pageset
;
7270 local_irq_restore(flags
);
7273 #ifdef CONFIG_MEMORY_HOTREMOVE
7275 * All pages in the range must be in a single zone and isolated
7276 * before calling this.
7279 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
7283 unsigned int order
, i
;
7285 unsigned long flags
;
7286 /* find the first valid pfn */
7287 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
7292 zone
= page_zone(pfn_to_page(pfn
));
7293 spin_lock_irqsave(&zone
->lock
, flags
);
7295 while (pfn
< end_pfn
) {
7296 if (!pfn_valid(pfn
)) {
7300 page
= pfn_to_page(pfn
);
7302 * The HWPoisoned page may be not in buddy system, and
7303 * page_count() is not 0.
7305 if (unlikely(!PageBuddy(page
) && PageHWPoison(page
))) {
7307 SetPageReserved(page
);
7311 BUG_ON(page_count(page
));
7312 BUG_ON(!PageBuddy(page
));
7313 order
= page_order(page
);
7314 #ifdef CONFIG_DEBUG_VM
7315 pr_info("remove from free list %lx %d %lx\n",
7316 pfn
, 1 << order
, end_pfn
);
7318 list_del(&page
->lru
);
7319 rmv_page_order(page
);
7320 zone
->free_area
[order
].nr_free
--;
7321 for (i
= 0; i
< (1 << order
); i
++)
7322 SetPageReserved((page
+i
));
7323 pfn
+= (1 << order
);
7325 spin_unlock_irqrestore(&zone
->lock
, flags
);
7329 bool is_free_buddy_page(struct page
*page
)
7331 struct zone
*zone
= page_zone(page
);
7332 unsigned long pfn
= page_to_pfn(page
);
7333 unsigned long flags
;
7336 spin_lock_irqsave(&zone
->lock
, flags
);
7337 for (order
= 0; order
< MAX_ORDER
; order
++) {
7338 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
7340 if (PageBuddy(page_head
) && page_order(page_head
) >= order
)
7343 spin_unlock_irqrestore(&zone
->lock
, flags
);
7345 return order
< MAX_ORDER
;