2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kasan.h>
28 #include <linux/module.h>
29 #include <linux/suspend.h>
30 #include <linux/pagevec.h>
31 #include <linux/blkdev.h>
32 #include <linux/slab.h>
33 #include <linux/ratelimit.h>
34 #include <linux/oom.h>
35 #include <linux/notifier.h>
36 #include <linux/topology.h>
37 #include <linux/sysctl.h>
38 #include <linux/cpu.h>
39 #include <linux/cpuset.h>
40 #include <linux/memory_hotplug.h>
41 #include <linux/nodemask.h>
42 #include <linux/vmalloc.h>
43 #include <linux/vmstat.h>
44 #include <linux/mempolicy.h>
45 #include <linux/memremap.h>
46 #include <linux/stop_machine.h>
47 #include <linux/sort.h>
48 #include <linux/pfn.h>
49 #include <linux/backing-dev.h>
50 #include <linux/fault-inject.h>
51 #include <linux/page-isolation.h>
52 #include <linux/page_ext.h>
53 #include <linux/debugobjects.h>
54 #include <linux/kmemleak.h>
55 #include <linux/compaction.h>
56 #include <trace/events/kmem.h>
57 #include <trace/events/oom.h>
58 #include <linux/prefetch.h>
59 #include <linux/mm_inline.h>
60 #include <linux/migrate.h>
61 #include <linux/hugetlb.h>
62 #include <linux/sched/rt.h>
63 #include <linux/sched/mm.h>
64 #include <linux/page_owner.h>
65 #include <linux/kthread.h>
66 #include <linux/memcontrol.h>
67 #include <linux/ftrace.h>
68 #include <linux/lockdep.h>
69 #include <linux/nmi.h>
71 #include <asm/sections.h>
72 #include <asm/tlbflush.h>
73 #include <asm/div64.h>
76 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
77 static DEFINE_MUTEX(pcp_batch_high_lock
);
78 #define MIN_PERCPU_PAGELIST_FRACTION (8)
80 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
81 DEFINE_PER_CPU(int, numa_node
);
82 EXPORT_PER_CPU_SYMBOL(numa_node
);
85 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
87 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
88 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
89 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
90 * defined in <linux/topology.h>.
92 DEFINE_PER_CPU(int, _numa_mem_
); /* Kernel "local memory" node */
93 EXPORT_PER_CPU_SYMBOL(_numa_mem_
);
94 int _node_numa_mem_
[MAX_NUMNODES
];
97 /* work_structs for global per-cpu drains */
98 DEFINE_MUTEX(pcpu_drain_mutex
);
99 DEFINE_PER_CPU(struct work_struct
, pcpu_drain
);
101 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
102 volatile unsigned long latent_entropy __latent_entropy
;
103 EXPORT_SYMBOL(latent_entropy
);
107 * Array of node states.
109 nodemask_t node_states
[NR_NODE_STATES
] __read_mostly
= {
110 [N_POSSIBLE
] = NODE_MASK_ALL
,
111 [N_ONLINE
] = { { [0] = 1UL } },
113 [N_NORMAL_MEMORY
] = { { [0] = 1UL } },
114 #ifdef CONFIG_HIGHMEM
115 [N_HIGH_MEMORY
] = { { [0] = 1UL } },
117 [N_MEMORY
] = { { [0] = 1UL } },
118 [N_CPU
] = { { [0] = 1UL } },
121 EXPORT_SYMBOL(node_states
);
123 /* Protect totalram_pages and zone->managed_pages */
124 static DEFINE_SPINLOCK(managed_page_count_lock
);
126 unsigned long totalram_pages __read_mostly
;
127 unsigned long totalreserve_pages __read_mostly
;
128 unsigned long totalcma_pages __read_mostly
;
130 int percpu_pagelist_fraction
;
131 gfp_t gfp_allowed_mask __read_mostly
= GFP_BOOT_MASK
;
133 #ifdef CONFIG_PM_SLEEP
135 * The following functions are used by the suspend/hibernate code to temporarily
136 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
137 * while devices are suspended. To avoid races with the suspend/hibernate code,
138 * they should always be called with pm_mutex held (gfp_allowed_mask also should
139 * only be modified with pm_mutex held, unless the suspend/hibernate code is
140 * guaranteed not to run in parallel with that modification).
143 static gfp_t saved_gfp_mask
;
145 void pm_restore_gfp_mask(void)
147 WARN_ON(!mutex_is_locked(&pm_mutex
));
148 if (saved_gfp_mask
) {
149 gfp_allowed_mask
= saved_gfp_mask
;
154 void pm_restrict_gfp_mask(void)
156 WARN_ON(!mutex_is_locked(&pm_mutex
));
157 WARN_ON(saved_gfp_mask
);
158 saved_gfp_mask
= gfp_allowed_mask
;
159 gfp_allowed_mask
&= ~(__GFP_IO
| __GFP_FS
);
162 bool pm_suspended_storage(void)
164 if ((gfp_allowed_mask
& (__GFP_IO
| __GFP_FS
)) == (__GFP_IO
| __GFP_FS
))
168 #endif /* CONFIG_PM_SLEEP */
170 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
171 unsigned int pageblock_order __read_mostly
;
174 static void __free_pages_ok(struct page
*page
, unsigned int order
);
177 * results with 256, 32 in the lowmem_reserve sysctl:
178 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
179 * 1G machine -> (16M dma, 784M normal, 224M high)
180 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
181 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
182 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
184 * TBD: should special case ZONE_DMA32 machines here - in those we normally
185 * don't need any ZONE_NORMAL reservation
187 int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
-1] = {
188 #ifdef CONFIG_ZONE_DMA
191 #ifdef CONFIG_ZONE_DMA32
194 #ifdef CONFIG_HIGHMEM
200 EXPORT_SYMBOL(totalram_pages
);
202 static char * const zone_names
[MAX_NR_ZONES
] = {
203 #ifdef CONFIG_ZONE_DMA
206 #ifdef CONFIG_ZONE_DMA32
210 #ifdef CONFIG_HIGHMEM
214 #ifdef CONFIG_ZONE_DEVICE
219 char * const migratetype_names
[MIGRATE_TYPES
] = {
227 #ifdef CONFIG_MEMORY_ISOLATION
232 compound_page_dtor
* const compound_page_dtors
[] = {
235 #ifdef CONFIG_HUGETLB_PAGE
238 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
244 * Try to keep at least this much lowmem free. Do not allow normal
245 * allocations below this point, only high priority ones. Automatically
246 * tuned according to the amount of memory in the system.
248 int min_free_kbytes
= 1024;
249 int user_min_free_kbytes
= -1;
250 int watermark_scale_factor
= 10;
253 * Extra memory for the system to try freeing. Used to temporarily
254 * free memory, to make space for new workloads. Anyone can allocate
255 * down to the min watermarks controlled by min_free_kbytes above.
257 int extra_free_kbytes
= 0;
259 static unsigned long __meminitdata nr_kernel_pages
;
260 static unsigned long __meminitdata nr_all_pages
;
261 static unsigned long __meminitdata dma_reserve
;
263 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
264 static unsigned long __meminitdata arch_zone_lowest_possible_pfn
[MAX_NR_ZONES
];
265 static unsigned long __meminitdata arch_zone_highest_possible_pfn
[MAX_NR_ZONES
];
266 static unsigned long __initdata required_kernelcore
;
267 static unsigned long __initdata required_movablecore
;
268 static unsigned long __meminitdata zone_movable_pfn
[MAX_NUMNODES
];
269 static bool mirrored_kernelcore
;
271 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
273 EXPORT_SYMBOL(movable_zone
);
274 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
277 int nr_node_ids __read_mostly
= MAX_NUMNODES
;
278 int nr_online_nodes __read_mostly
= 1;
279 EXPORT_SYMBOL(nr_node_ids
);
280 EXPORT_SYMBOL(nr_online_nodes
);
283 int page_group_by_mobility_disabled __read_mostly
;
285 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
288 * Determine how many pages need to be initialized durig early boot
289 * (non-deferred initialization).
290 * The value of first_deferred_pfn will be set later, once non-deferred pages
291 * are initialized, but for now set it ULONG_MAX.
293 static inline void reset_deferred_meminit(pg_data_t
*pgdat
)
295 phys_addr_t start_addr
, end_addr
;
296 unsigned long max_pgcnt
;
297 unsigned long reserved
;
300 * Initialise at least 2G of a node but also take into account that
301 * two large system hashes that can take up 1GB for 0.25TB/node.
303 max_pgcnt
= max(2UL << (30 - PAGE_SHIFT
),
304 (pgdat
->node_spanned_pages
>> 8));
307 * Compensate the all the memblock reservations (e.g. crash kernel)
308 * from the initial estimation to make sure we will initialize enough
311 start_addr
= PFN_PHYS(pgdat
->node_start_pfn
);
312 end_addr
= PFN_PHYS(pgdat
->node_start_pfn
+ max_pgcnt
);
313 reserved
= memblock_reserved_memory_within(start_addr
, end_addr
);
314 max_pgcnt
+= PHYS_PFN(reserved
);
316 pgdat
->static_init_pgcnt
= min(max_pgcnt
, pgdat
->node_spanned_pages
);
317 pgdat
->first_deferred_pfn
= ULONG_MAX
;
320 /* Returns true if the struct page for the pfn is uninitialised */
321 static inline bool __meminit
early_page_uninitialised(unsigned long pfn
)
323 int nid
= early_pfn_to_nid(pfn
);
325 if (node_online(nid
) && pfn
>= NODE_DATA(nid
)->first_deferred_pfn
)
332 * Returns false when the remaining initialisation should be deferred until
333 * later in the boot cycle when it can be parallelised.
335 static inline bool update_defer_init(pg_data_t
*pgdat
,
336 unsigned long pfn
, unsigned long zone_end
,
337 unsigned long *nr_initialised
)
339 /* Always populate low zones for address-contrained allocations */
340 if (zone_end
< pgdat_end_pfn(pgdat
))
343 if ((*nr_initialised
> pgdat
->static_init_pgcnt
) &&
344 (pfn
& (PAGES_PER_SECTION
- 1)) == 0) {
345 pgdat
->first_deferred_pfn
= pfn
;
352 static inline void reset_deferred_meminit(pg_data_t
*pgdat
)
356 static inline bool early_page_uninitialised(unsigned long pfn
)
361 static inline bool update_defer_init(pg_data_t
*pgdat
,
362 unsigned long pfn
, unsigned long zone_end
,
363 unsigned long *nr_initialised
)
369 /* Return a pointer to the bitmap storing bits affecting a block of pages */
370 static inline unsigned long *get_pageblock_bitmap(struct page
*page
,
373 #ifdef CONFIG_SPARSEMEM
374 return __pfn_to_section(pfn
)->pageblock_flags
;
376 return page_zone(page
)->pageblock_flags
;
377 #endif /* CONFIG_SPARSEMEM */
380 static inline int pfn_to_bitidx(struct page
*page
, unsigned long pfn
)
382 #ifdef CONFIG_SPARSEMEM
383 pfn
&= (PAGES_PER_SECTION
-1);
384 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
386 pfn
= pfn
- round_down(page_zone(page
)->zone_start_pfn
, pageblock_nr_pages
);
387 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
388 #endif /* CONFIG_SPARSEMEM */
392 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
393 * @page: The page within the block of interest
394 * @pfn: The target page frame number
395 * @end_bitidx: The last bit of interest to retrieve
396 * @mask: mask of bits that the caller is interested in
398 * Return: pageblock_bits flags
400 static __always_inline
unsigned long __get_pfnblock_flags_mask(struct page
*page
,
402 unsigned long end_bitidx
,
405 unsigned long *bitmap
;
406 unsigned long bitidx
, word_bitidx
;
409 bitmap
= get_pageblock_bitmap(page
, pfn
);
410 bitidx
= pfn_to_bitidx(page
, pfn
);
411 word_bitidx
= bitidx
/ BITS_PER_LONG
;
412 bitidx
&= (BITS_PER_LONG
-1);
414 word
= bitmap
[word_bitidx
];
415 bitidx
+= end_bitidx
;
416 return (word
>> (BITS_PER_LONG
- bitidx
- 1)) & mask
;
419 unsigned long get_pfnblock_flags_mask(struct page
*page
, unsigned long pfn
,
420 unsigned long end_bitidx
,
423 return __get_pfnblock_flags_mask(page
, pfn
, end_bitidx
, mask
);
426 static __always_inline
int get_pfnblock_migratetype(struct page
*page
, unsigned long pfn
)
428 return __get_pfnblock_flags_mask(page
, pfn
, PB_migrate_end
, MIGRATETYPE_MASK
);
432 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
433 * @page: The page within the block of interest
434 * @flags: The flags to set
435 * @pfn: The target page frame number
436 * @end_bitidx: The last bit of interest
437 * @mask: mask of bits that the caller is interested in
439 void set_pfnblock_flags_mask(struct page
*page
, unsigned long flags
,
441 unsigned long end_bitidx
,
444 unsigned long *bitmap
;
445 unsigned long bitidx
, word_bitidx
;
446 unsigned long old_word
, word
;
448 BUILD_BUG_ON(NR_PAGEBLOCK_BITS
!= 4);
450 bitmap
= get_pageblock_bitmap(page
, pfn
);
451 bitidx
= pfn_to_bitidx(page
, pfn
);
452 word_bitidx
= bitidx
/ BITS_PER_LONG
;
453 bitidx
&= (BITS_PER_LONG
-1);
455 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page
), pfn
), page
);
457 bitidx
+= end_bitidx
;
458 mask
<<= (BITS_PER_LONG
- bitidx
- 1);
459 flags
<<= (BITS_PER_LONG
- bitidx
- 1);
461 word
= READ_ONCE(bitmap
[word_bitidx
]);
463 old_word
= cmpxchg(&bitmap
[word_bitidx
], word
, (word
& ~mask
) | flags
);
464 if (word
== old_word
)
470 void set_pageblock_migratetype(struct page
*page
, int migratetype
)
472 if (unlikely(page_group_by_mobility_disabled
&&
473 migratetype
< MIGRATE_PCPTYPES
))
474 migratetype
= MIGRATE_UNMOVABLE
;
476 set_pageblock_flags_group(page
, (unsigned long)migratetype
,
477 PB_migrate
, PB_migrate_end
);
480 #ifdef CONFIG_DEBUG_VM
481 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
485 unsigned long pfn
= page_to_pfn(page
);
486 unsigned long sp
, start_pfn
;
489 seq
= zone_span_seqbegin(zone
);
490 start_pfn
= zone
->zone_start_pfn
;
491 sp
= zone
->spanned_pages
;
492 if (!zone_spans_pfn(zone
, pfn
))
494 } while (zone_span_seqretry(zone
, seq
));
497 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
498 pfn
, zone_to_nid(zone
), zone
->name
,
499 start_pfn
, start_pfn
+ sp
);
504 static int page_is_consistent(struct zone
*zone
, struct page
*page
)
506 if (!pfn_valid_within(page_to_pfn(page
)))
508 if (zone
!= page_zone(page
))
514 * Temporary debugging check for pages not lying within a given zone.
516 static int __maybe_unused
bad_range(struct zone
*zone
, struct page
*page
)
518 if (page_outside_zone_boundaries(zone
, page
))
520 if (!page_is_consistent(zone
, page
))
526 static inline int __maybe_unused
bad_range(struct zone
*zone
, struct page
*page
)
532 static void bad_page(struct page
*page
, const char *reason
,
533 unsigned long bad_flags
)
535 static unsigned long resume
;
536 static unsigned long nr_shown
;
537 static unsigned long nr_unshown
;
540 * Allow a burst of 60 reports, then keep quiet for that minute;
541 * or allow a steady drip of one report per second.
543 if (nr_shown
== 60) {
544 if (time_before(jiffies
, resume
)) {
550 "BUG: Bad page state: %lu messages suppressed\n",
557 resume
= jiffies
+ 60 * HZ
;
559 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
560 current
->comm
, page_to_pfn(page
));
561 __dump_page(page
, reason
);
562 bad_flags
&= page
->flags
;
564 pr_alert("bad because of flags: %#lx(%pGp)\n",
565 bad_flags
, &bad_flags
);
566 dump_page_owner(page
);
571 /* Leave bad fields for debug, except PageBuddy could make trouble */
572 page_mapcount_reset(page
); /* remove PageBuddy */
573 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
577 * Higher-order pages are called "compound pages". They are structured thusly:
579 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
581 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
582 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
584 * The first tail page's ->compound_dtor holds the offset in array of compound
585 * page destructors. See compound_page_dtors.
587 * The first tail page's ->compound_order holds the order of allocation.
588 * This usage means that zero-order pages may not be compound.
591 void free_compound_page(struct page
*page
)
593 __free_pages_ok(page
, compound_order(page
));
596 void prep_compound_page(struct page
*page
, unsigned int order
)
599 int nr_pages
= 1 << order
;
601 set_compound_page_dtor(page
, COMPOUND_PAGE_DTOR
);
602 set_compound_order(page
, order
);
604 for (i
= 1; i
< nr_pages
; i
++) {
605 struct page
*p
= page
+ i
;
606 set_page_count(p
, 0);
607 p
->mapping
= TAIL_MAPPING
;
608 set_compound_head(p
, page
);
610 atomic_set(compound_mapcount_ptr(page
), -1);
613 #ifdef CONFIG_DEBUG_PAGEALLOC
614 unsigned int _debug_guardpage_minorder
;
615 bool _debug_pagealloc_enabled __read_mostly
616 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT
);
617 EXPORT_SYMBOL(_debug_pagealloc_enabled
);
618 bool _debug_guardpage_enabled __read_mostly
;
620 static int __init
early_debug_pagealloc(char *buf
)
624 return kstrtobool(buf
, &_debug_pagealloc_enabled
);
626 early_param("debug_pagealloc", early_debug_pagealloc
);
628 static bool need_debug_guardpage(void)
630 /* If we don't use debug_pagealloc, we don't need guard page */
631 if (!debug_pagealloc_enabled())
634 if (!debug_guardpage_minorder())
640 static void init_debug_guardpage(void)
642 if (!debug_pagealloc_enabled())
645 if (!debug_guardpage_minorder())
648 _debug_guardpage_enabled
= true;
651 struct page_ext_operations debug_guardpage_ops
= {
652 .need
= need_debug_guardpage
,
653 .init
= init_debug_guardpage
,
656 static int __init
debug_guardpage_minorder_setup(char *buf
)
660 if (kstrtoul(buf
, 10, &res
) < 0 || res
> MAX_ORDER
/ 2) {
661 pr_err("Bad debug_guardpage_minorder value\n");
664 _debug_guardpage_minorder
= res
;
665 pr_info("Setting debug_guardpage_minorder to %lu\n", res
);
668 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup
);
670 static inline bool set_page_guard(struct zone
*zone
, struct page
*page
,
671 unsigned int order
, int migratetype
)
673 struct page_ext
*page_ext
;
675 if (!debug_guardpage_enabled())
678 if (order
>= debug_guardpage_minorder())
681 page_ext
= lookup_page_ext(page
);
682 if (unlikely(!page_ext
))
685 __set_bit(PAGE_EXT_DEBUG_GUARD
, &page_ext
->flags
);
687 INIT_LIST_HEAD(&page
->lru
);
688 set_page_private(page
, order
);
689 /* Guard pages are not available for any usage */
690 __mod_zone_freepage_state(zone
, -(1 << order
), migratetype
);
695 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
696 unsigned int order
, int migratetype
)
698 struct page_ext
*page_ext
;
700 if (!debug_guardpage_enabled())
703 page_ext
= lookup_page_ext(page
);
704 if (unlikely(!page_ext
))
707 __clear_bit(PAGE_EXT_DEBUG_GUARD
, &page_ext
->flags
);
709 set_page_private(page
, 0);
710 if (!is_migrate_isolate(migratetype
))
711 __mod_zone_freepage_state(zone
, (1 << order
), migratetype
);
714 struct page_ext_operations debug_guardpage_ops
;
715 static inline bool set_page_guard(struct zone
*zone
, struct page
*page
,
716 unsigned int order
, int migratetype
) { return false; }
717 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
718 unsigned int order
, int migratetype
) {}
721 static inline void set_page_order(struct page
*page
, unsigned int order
)
723 set_page_private(page
, order
);
724 __SetPageBuddy(page
);
727 static inline void rmv_page_order(struct page
*page
)
729 __ClearPageBuddy(page
);
730 set_page_private(page
, 0);
734 * This function checks whether a page is free && is the buddy
735 * we can do coalesce a page and its buddy if
736 * (a) the buddy is not in a hole (check before calling!) &&
737 * (b) the buddy is in the buddy system &&
738 * (c) a page and its buddy have the same order &&
739 * (d) a page and its buddy are in the same zone.
741 * For recording whether a page is in the buddy system, we set ->_mapcount
742 * PAGE_BUDDY_MAPCOUNT_VALUE.
743 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
744 * serialized by zone->lock.
746 * For recording page's order, we use page_private(page).
748 static inline int page_is_buddy(struct page
*page
, struct page
*buddy
,
751 if (page_is_guard(buddy
) && page_order(buddy
) == order
) {
752 if (page_zone_id(page
) != page_zone_id(buddy
))
755 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
760 if (PageBuddy(buddy
) && page_order(buddy
) == order
) {
762 * zone check is done late to avoid uselessly
763 * calculating zone/node ids for pages that could
766 if (page_zone_id(page
) != page_zone_id(buddy
))
769 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
777 * Freeing function for a buddy system allocator.
779 * The concept of a buddy system is to maintain direct-mapped table
780 * (containing bit values) for memory blocks of various "orders".
781 * The bottom level table contains the map for the smallest allocatable
782 * units of memory (here, pages), and each level above it describes
783 * pairs of units from the levels below, hence, "buddies".
784 * At a high level, all that happens here is marking the table entry
785 * at the bottom level available, and propagating the changes upward
786 * as necessary, plus some accounting needed to play nicely with other
787 * parts of the VM system.
788 * At each level, we keep a list of pages, which are heads of continuous
789 * free pages of length of (1 << order) and marked with _mapcount
790 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
792 * So when we are allocating or freeing one, we can derive the state of the
793 * other. That is, if we allocate a small block, and both were
794 * free, the remainder of the region must be split into blocks.
795 * If a block is freed, and its buddy is also free, then this
796 * triggers coalescing into a block of larger size.
801 static inline void __free_one_page(struct page
*page
,
803 struct zone
*zone
, unsigned int order
,
806 unsigned long combined_pfn
;
807 unsigned long uninitialized_var(buddy_pfn
);
809 unsigned int max_order
;
811 max_order
= min_t(unsigned int, MAX_ORDER
, pageblock_order
+ 1);
813 VM_BUG_ON(!zone_is_initialized(zone
));
814 VM_BUG_ON_PAGE(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
, page
);
816 VM_BUG_ON(migratetype
== -1);
817 if (likely(!is_migrate_isolate(migratetype
)))
818 __mod_zone_freepage_state(zone
, 1 << order
, migratetype
);
820 VM_BUG_ON_PAGE(pfn
& ((1 << order
) - 1), page
);
821 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
824 while (order
< max_order
- 1) {
825 buddy_pfn
= __find_buddy_pfn(pfn
, order
);
826 buddy
= page
+ (buddy_pfn
- pfn
);
828 if (!pfn_valid_within(buddy_pfn
))
830 if (!page_is_buddy(page
, buddy
, order
))
833 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
834 * merge with it and move up one order.
836 if (page_is_guard(buddy
)) {
837 clear_page_guard(zone
, buddy
, order
, migratetype
);
839 list_del(&buddy
->lru
);
840 zone
->free_area
[order
].nr_free
--;
841 rmv_page_order(buddy
);
843 combined_pfn
= buddy_pfn
& pfn
;
844 page
= page
+ (combined_pfn
- pfn
);
848 if (max_order
< MAX_ORDER
) {
849 /* If we are here, it means order is >= pageblock_order.
850 * We want to prevent merge between freepages on isolate
851 * pageblock and normal pageblock. Without this, pageblock
852 * isolation could cause incorrect freepage or CMA accounting.
854 * We don't want to hit this code for the more frequent
857 if (unlikely(has_isolate_pageblock(zone
))) {
860 buddy_pfn
= __find_buddy_pfn(pfn
, order
);
861 buddy
= page
+ (buddy_pfn
- pfn
);
862 buddy_mt
= get_pageblock_migratetype(buddy
);
864 if (migratetype
!= buddy_mt
865 && (is_migrate_isolate(migratetype
) ||
866 is_migrate_isolate(buddy_mt
)))
870 goto continue_merging
;
874 set_page_order(page
, order
);
877 * If this is not the largest possible page, check if the buddy
878 * of the next-highest order is free. If it is, it's possible
879 * that pages are being freed that will coalesce soon. In case,
880 * that is happening, add the free page to the tail of the list
881 * so it's less likely to be used soon and more likely to be merged
882 * as a higher order page
884 if ((order
< MAX_ORDER
-2) && pfn_valid_within(buddy_pfn
)) {
885 struct page
*higher_page
, *higher_buddy
;
886 combined_pfn
= buddy_pfn
& pfn
;
887 higher_page
= page
+ (combined_pfn
- pfn
);
888 buddy_pfn
= __find_buddy_pfn(combined_pfn
, order
+ 1);
889 higher_buddy
= higher_page
+ (buddy_pfn
- combined_pfn
);
890 if (pfn_valid_within(buddy_pfn
) &&
891 page_is_buddy(higher_page
, higher_buddy
, order
+ 1)) {
892 list_add_tail(&page
->lru
,
893 &zone
->free_area
[order
].free_list
[migratetype
]);
898 list_add(&page
->lru
, &zone
->free_area
[order
].free_list
[migratetype
]);
900 zone
->free_area
[order
].nr_free
++;
904 * A bad page could be due to a number of fields. Instead of multiple branches,
905 * try and check multiple fields with one check. The caller must do a detailed
906 * check if necessary.
908 static inline bool page_expected_state(struct page
*page
,
909 unsigned long check_flags
)
911 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
914 if (unlikely((unsigned long)page
->mapping
|
915 page_ref_count(page
) |
917 (unsigned long)page
->mem_cgroup
|
919 (page
->flags
& check_flags
)))
925 static void free_pages_check_bad(struct page
*page
)
927 const char *bad_reason
;
928 unsigned long bad_flags
;
933 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
934 bad_reason
= "nonzero mapcount";
935 if (unlikely(page
->mapping
!= NULL
))
936 bad_reason
= "non-NULL mapping";
937 if (unlikely(page_ref_count(page
) != 0))
938 bad_reason
= "nonzero _refcount";
939 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_FREE
)) {
940 bad_reason
= "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
941 bad_flags
= PAGE_FLAGS_CHECK_AT_FREE
;
944 if (unlikely(page
->mem_cgroup
))
945 bad_reason
= "page still charged to cgroup";
947 bad_page(page
, bad_reason
, bad_flags
);
950 static inline int free_pages_check(struct page
*page
)
952 if (likely(page_expected_state(page
, PAGE_FLAGS_CHECK_AT_FREE
)))
955 /* Something has gone sideways, find it */
956 free_pages_check_bad(page
);
960 static int free_tail_pages_check(struct page
*head_page
, struct page
*page
)
965 * We rely page->lru.next never has bit 0 set, unless the page
966 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
968 BUILD_BUG_ON((unsigned long)LIST_POISON1
& 1);
970 if (!IS_ENABLED(CONFIG_DEBUG_VM
)) {
974 switch (page
- head_page
) {
976 /* the first tail page: ->mapping is compound_mapcount() */
977 if (unlikely(compound_mapcount(page
))) {
978 bad_page(page
, "nonzero compound_mapcount", 0);
984 * the second tail page: ->mapping is
985 * page_deferred_list().next -- ignore value.
989 if (page
->mapping
!= TAIL_MAPPING
) {
990 bad_page(page
, "corrupted mapping in tail page", 0);
995 if (unlikely(!PageTail(page
))) {
996 bad_page(page
, "PageTail not set", 0);
999 if (unlikely(compound_head(page
) != head_page
)) {
1000 bad_page(page
, "compound_head not consistent", 0);
1005 page
->mapping
= NULL
;
1006 clear_compound_head(page
);
1010 static __always_inline
bool free_pages_prepare(struct page
*page
,
1011 unsigned int order
, bool check_free
)
1015 VM_BUG_ON_PAGE(PageTail(page
), page
);
1017 trace_mm_page_free(page
, order
);
1020 * Check tail pages before head page information is cleared to
1021 * avoid checking PageCompound for order-0 pages.
1023 if (unlikely(order
)) {
1024 bool compound
= PageCompound(page
);
1027 VM_BUG_ON_PAGE(compound
&& compound_order(page
) != order
, page
);
1030 ClearPageDoubleMap(page
);
1031 for (i
= 1; i
< (1 << order
); i
++) {
1033 bad
+= free_tail_pages_check(page
, page
+ i
);
1034 if (unlikely(free_pages_check(page
+ i
))) {
1038 (page
+ i
)->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
1041 if (PageMappingFlags(page
))
1042 page
->mapping
= NULL
;
1043 if (memcg_kmem_enabled() && PageKmemcg(page
))
1044 memcg_kmem_uncharge(page
, order
);
1046 bad
+= free_pages_check(page
);
1050 page_cpupid_reset_last(page
);
1051 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
1052 reset_page_owner(page
, order
);
1054 if (!PageHighMem(page
)) {
1055 debug_check_no_locks_freed(page_address(page
),
1056 PAGE_SIZE
<< order
);
1057 debug_check_no_obj_freed(page_address(page
),
1058 PAGE_SIZE
<< order
);
1060 arch_free_page(page
, order
);
1061 kernel_poison_pages(page
, 1 << order
, 0);
1062 kernel_map_pages(page
, 1 << order
, 0);
1063 kasan_free_pages(page
, order
);
1068 #ifdef CONFIG_DEBUG_VM
1069 static inline bool free_pcp_prepare(struct page
*page
)
1071 return free_pages_prepare(page
, 0, true);
1074 static inline bool bulkfree_pcp_prepare(struct page
*page
)
1079 static bool free_pcp_prepare(struct page
*page
)
1081 return free_pages_prepare(page
, 0, false);
1084 static bool bulkfree_pcp_prepare(struct page
*page
)
1086 return free_pages_check(page
);
1088 #endif /* CONFIG_DEBUG_VM */
1091 * Frees a number of pages from the PCP lists
1092 * Assumes all pages on list are in same zone, and of same order.
1093 * count is the number of pages to free.
1095 * If the zone was previously in an "all pages pinned" state then look to
1096 * see if this freeing clears that state.
1098 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1099 * pinned" detection logic.
1101 static void free_pcppages_bulk(struct zone
*zone
, int count
,
1102 struct per_cpu_pages
*pcp
)
1104 int migratetype
= 0;
1106 bool isolated_pageblocks
;
1108 spin_lock(&zone
->lock
);
1109 isolated_pageblocks
= has_isolate_pageblock(zone
);
1113 struct list_head
*list
;
1116 * Remove pages from lists in a round-robin fashion. A
1117 * batch_free count is maintained that is incremented when an
1118 * empty list is encountered. This is so more pages are freed
1119 * off fuller lists instead of spinning excessively around empty
1124 if (++migratetype
== MIGRATE_PCPTYPES
)
1126 list
= &pcp
->lists
[migratetype
];
1127 } while (list_empty(list
));
1129 /* This is the only non-empty list. Free them all. */
1130 if (batch_free
== MIGRATE_PCPTYPES
)
1134 int mt
; /* migratetype of the to-be-freed page */
1136 page
= list_last_entry(list
, struct page
, lru
);
1137 /* must delete as __free_one_page list manipulates */
1138 list_del(&page
->lru
);
1140 mt
= get_pcppage_migratetype(page
);
1141 /* MIGRATE_ISOLATE page should not go to pcplists */
1142 VM_BUG_ON_PAGE(is_migrate_isolate(mt
), page
);
1143 /* Pageblock could have been isolated meanwhile */
1144 if (unlikely(isolated_pageblocks
))
1145 mt
= get_pageblock_migratetype(page
);
1147 if (bulkfree_pcp_prepare(page
))
1150 __free_one_page(page
, page_to_pfn(page
), zone
, 0, mt
);
1151 trace_mm_page_pcpu_drain(page
, 0, mt
);
1152 } while (--count
&& --batch_free
&& !list_empty(list
));
1154 spin_unlock(&zone
->lock
);
1157 static void free_one_page(struct zone
*zone
,
1158 struct page
*page
, unsigned long pfn
,
1162 spin_lock(&zone
->lock
);
1163 if (unlikely(has_isolate_pageblock(zone
) ||
1164 is_migrate_isolate(migratetype
))) {
1165 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1167 __free_one_page(page
, pfn
, zone
, order
, migratetype
);
1168 spin_unlock(&zone
->lock
);
1171 static void __meminit
__init_single_page(struct page
*page
, unsigned long pfn
,
1172 unsigned long zone
, int nid
)
1174 set_page_links(page
, zone
, nid
, pfn
);
1175 init_page_count(page
);
1176 page_mapcount_reset(page
);
1177 page_cpupid_reset_last(page
);
1179 INIT_LIST_HEAD(&page
->lru
);
1180 #ifdef WANT_PAGE_VIRTUAL
1181 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1182 if (!is_highmem_idx(zone
))
1183 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
1187 static void __meminit
__init_single_pfn(unsigned long pfn
, unsigned long zone
,
1190 return __init_single_page(pfn_to_page(pfn
), pfn
, zone
, nid
);
1193 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1194 static void __meminit
init_reserved_page(unsigned long pfn
)
1199 if (!early_page_uninitialised(pfn
))
1202 nid
= early_pfn_to_nid(pfn
);
1203 pgdat
= NODE_DATA(nid
);
1205 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1206 struct zone
*zone
= &pgdat
->node_zones
[zid
];
1208 if (pfn
>= zone
->zone_start_pfn
&& pfn
< zone_end_pfn(zone
))
1211 __init_single_pfn(pfn
, zid
, nid
);
1214 static inline void init_reserved_page(unsigned long pfn
)
1217 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1220 * Initialised pages do not have PageReserved set. This function is
1221 * called for each range allocated by the bootmem allocator and
1222 * marks the pages PageReserved. The remaining valid pages are later
1223 * sent to the buddy page allocator.
1225 void __meminit
reserve_bootmem_region(phys_addr_t start
, phys_addr_t end
)
1227 unsigned long start_pfn
= PFN_DOWN(start
);
1228 unsigned long end_pfn
= PFN_UP(end
);
1230 for (; start_pfn
< end_pfn
; start_pfn
++) {
1231 if (pfn_valid(start_pfn
)) {
1232 struct page
*page
= pfn_to_page(start_pfn
);
1234 init_reserved_page(start_pfn
);
1236 /* Avoid false-positive PageTail() */
1237 INIT_LIST_HEAD(&page
->lru
);
1239 SetPageReserved(page
);
1244 static void __free_pages_ok(struct page
*page
, unsigned int order
)
1246 unsigned long flags
;
1248 unsigned long pfn
= page_to_pfn(page
);
1250 if (!free_pages_prepare(page
, order
, true))
1253 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1254 local_irq_save(flags
);
1255 __count_vm_events(PGFREE
, 1 << order
);
1256 free_one_page(page_zone(page
), page
, pfn
, order
, migratetype
);
1257 local_irq_restore(flags
);
1260 static void __init
__free_pages_boot_core(struct page
*page
, unsigned int order
)
1262 unsigned int nr_pages
= 1 << order
;
1263 struct page
*p
= page
;
1267 for (loop
= 0; loop
< (nr_pages
- 1); loop
++, p
++) {
1269 __ClearPageReserved(p
);
1270 set_page_count(p
, 0);
1272 __ClearPageReserved(p
);
1273 set_page_count(p
, 0);
1275 page_zone(page
)->managed_pages
+= nr_pages
;
1276 set_page_refcounted(page
);
1277 __free_pages(page
, order
);
1280 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1281 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1283 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata
;
1285 int __meminit
early_pfn_to_nid(unsigned long pfn
)
1287 static DEFINE_SPINLOCK(early_pfn_lock
);
1290 spin_lock(&early_pfn_lock
);
1291 nid
= __early_pfn_to_nid(pfn
, &early_pfnnid_cache
);
1293 nid
= first_online_node
;
1294 spin_unlock(&early_pfn_lock
);
1300 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1301 static inline bool __meminit __maybe_unused
1302 meminit_pfn_in_nid(unsigned long pfn
, int node
,
1303 struct mminit_pfnnid_cache
*state
)
1307 nid
= __early_pfn_to_nid(pfn
, state
);
1308 if (nid
>= 0 && nid
!= node
)
1313 /* Only safe to use early in boot when initialisation is single-threaded */
1314 static inline bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
1316 return meminit_pfn_in_nid(pfn
, node
, &early_pfnnid_cache
);
1321 static inline bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
1325 static inline bool __meminit __maybe_unused
1326 meminit_pfn_in_nid(unsigned long pfn
, int node
,
1327 struct mminit_pfnnid_cache
*state
)
1334 void __init
__free_pages_bootmem(struct page
*page
, unsigned long pfn
,
1337 if (early_page_uninitialised(pfn
))
1339 return __free_pages_boot_core(page
, order
);
1343 * Check that the whole (or subset of) a pageblock given by the interval of
1344 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1345 * with the migration of free compaction scanner. The scanners then need to
1346 * use only pfn_valid_within() check for arches that allow holes within
1349 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1351 * It's possible on some configurations to have a setup like node0 node1 node0
1352 * i.e. it's possible that all pages within a zones range of pages do not
1353 * belong to a single zone. We assume that a border between node0 and node1
1354 * can occur within a single pageblock, but not a node0 node1 node0
1355 * interleaving within a single pageblock. It is therefore sufficient to check
1356 * the first and last page of a pageblock and avoid checking each individual
1357 * page in a pageblock.
1359 struct page
*__pageblock_pfn_to_page(unsigned long start_pfn
,
1360 unsigned long end_pfn
, struct zone
*zone
)
1362 struct page
*start_page
;
1363 struct page
*end_page
;
1365 /* end_pfn is one past the range we are checking */
1368 if (!pfn_valid(start_pfn
) || !pfn_valid(end_pfn
))
1371 start_page
= pfn_to_online_page(start_pfn
);
1375 if (page_zone(start_page
) != zone
)
1378 end_page
= pfn_to_page(end_pfn
);
1380 /* This gives a shorter code than deriving page_zone(end_page) */
1381 if (page_zone_id(start_page
) != page_zone_id(end_page
))
1387 void set_zone_contiguous(struct zone
*zone
)
1389 unsigned long block_start_pfn
= zone
->zone_start_pfn
;
1390 unsigned long block_end_pfn
;
1392 block_end_pfn
= ALIGN(block_start_pfn
+ 1, pageblock_nr_pages
);
1393 for (; block_start_pfn
< zone_end_pfn(zone
);
1394 block_start_pfn
= block_end_pfn
,
1395 block_end_pfn
+= pageblock_nr_pages
) {
1397 block_end_pfn
= min(block_end_pfn
, zone_end_pfn(zone
));
1399 if (!__pageblock_pfn_to_page(block_start_pfn
,
1400 block_end_pfn
, zone
))
1404 /* We confirm that there is no hole */
1405 zone
->contiguous
= true;
1408 void clear_zone_contiguous(struct zone
*zone
)
1410 zone
->contiguous
= false;
1413 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1414 static void __init
deferred_free_range(struct page
*page
,
1415 unsigned long pfn
, int nr_pages
)
1422 /* Free a large naturally-aligned chunk if possible */
1423 if (nr_pages
== pageblock_nr_pages
&&
1424 (pfn
& (pageblock_nr_pages
- 1)) == 0) {
1425 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
1426 __free_pages_boot_core(page
, pageblock_order
);
1430 for (i
= 0; i
< nr_pages
; i
++, page
++, pfn
++) {
1431 if ((pfn
& (pageblock_nr_pages
- 1)) == 0)
1432 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
1433 __free_pages_boot_core(page
, 0);
1437 /* Completion tracking for deferred_init_memmap() threads */
1438 static atomic_t pgdat_init_n_undone __initdata
;
1439 static __initdata
DECLARE_COMPLETION(pgdat_init_all_done_comp
);
1441 static inline void __init
pgdat_init_report_one_done(void)
1443 if (atomic_dec_and_test(&pgdat_init_n_undone
))
1444 complete(&pgdat_init_all_done_comp
);
1447 /* Initialise remaining memory on a node */
1448 static int __init
deferred_init_memmap(void *data
)
1450 pg_data_t
*pgdat
= data
;
1451 int nid
= pgdat
->node_id
;
1452 struct mminit_pfnnid_cache nid_init_state
= { };
1453 unsigned long start
= jiffies
;
1454 unsigned long nr_pages
= 0;
1455 unsigned long walk_start
, walk_end
;
1458 unsigned long first_init_pfn
= pgdat
->first_deferred_pfn
;
1459 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
1461 if (first_init_pfn
== ULONG_MAX
) {
1462 pgdat_init_report_one_done();
1466 /* Bind memory initialisation thread to a local node if possible */
1467 if (!cpumask_empty(cpumask
))
1468 set_cpus_allowed_ptr(current
, cpumask
);
1470 /* Sanity check boundaries */
1471 BUG_ON(pgdat
->first_deferred_pfn
< pgdat
->node_start_pfn
);
1472 BUG_ON(pgdat
->first_deferred_pfn
> pgdat_end_pfn(pgdat
));
1473 pgdat
->first_deferred_pfn
= ULONG_MAX
;
1475 /* Only the highest zone is deferred so find it */
1476 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1477 zone
= pgdat
->node_zones
+ zid
;
1478 if (first_init_pfn
< zone_end_pfn(zone
))
1482 for_each_mem_pfn_range(i
, nid
, &walk_start
, &walk_end
, NULL
) {
1483 unsigned long pfn
, end_pfn
;
1484 struct page
*page
= NULL
;
1485 struct page
*free_base_page
= NULL
;
1486 unsigned long free_base_pfn
= 0;
1489 end_pfn
= min(walk_end
, zone_end_pfn(zone
));
1490 pfn
= first_init_pfn
;
1491 if (pfn
< walk_start
)
1493 if (pfn
< zone
->zone_start_pfn
)
1494 pfn
= zone
->zone_start_pfn
;
1496 for (; pfn
< end_pfn
; pfn
++) {
1497 if (!pfn_valid_within(pfn
))
1501 * Ensure pfn_valid is checked every
1502 * pageblock_nr_pages for memory holes
1504 if ((pfn
& (pageblock_nr_pages
- 1)) == 0) {
1505 if (!pfn_valid(pfn
)) {
1511 if (!meminit_pfn_in_nid(pfn
, nid
, &nid_init_state
)) {
1516 /* Minimise pfn page lookups and scheduler checks */
1517 if (page
&& (pfn
& (pageblock_nr_pages
- 1)) != 0) {
1520 nr_pages
+= nr_to_free
;
1521 deferred_free_range(free_base_page
,
1522 free_base_pfn
, nr_to_free
);
1523 free_base_page
= NULL
;
1524 free_base_pfn
= nr_to_free
= 0;
1526 page
= pfn_to_page(pfn
);
1531 VM_BUG_ON(page_zone(page
) != zone
);
1535 __init_single_page(page
, pfn
, zid
, nid
);
1536 if (!free_base_page
) {
1537 free_base_page
= page
;
1538 free_base_pfn
= pfn
;
1543 /* Where possible, batch up pages for a single free */
1546 /* Free the current block of pages to allocator */
1547 nr_pages
+= nr_to_free
;
1548 deferred_free_range(free_base_page
, free_base_pfn
,
1550 free_base_page
= NULL
;
1551 free_base_pfn
= nr_to_free
= 0;
1553 /* Free the last block of pages to allocator */
1554 nr_pages
+= nr_to_free
;
1555 deferred_free_range(free_base_page
, free_base_pfn
, nr_to_free
);
1557 first_init_pfn
= max(end_pfn
, first_init_pfn
);
1560 /* Sanity check that the next zone really is unpopulated */
1561 WARN_ON(++zid
< MAX_NR_ZONES
&& populated_zone(++zone
));
1563 pr_info("node %d initialised, %lu pages in %ums\n", nid
, nr_pages
,
1564 jiffies_to_msecs(jiffies
- start
));
1566 pgdat_init_report_one_done();
1569 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1571 void __init
page_alloc_init_late(void)
1575 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1578 /* There will be num_node_state(N_MEMORY) threads */
1579 atomic_set(&pgdat_init_n_undone
, num_node_state(N_MEMORY
));
1580 for_each_node_state(nid
, N_MEMORY
) {
1581 kthread_run(deferred_init_memmap
, NODE_DATA(nid
), "pgdatinit%d", nid
);
1584 /* Block until all are initialised */
1585 wait_for_completion(&pgdat_init_all_done_comp
);
1587 /* Reinit limits that are based on free pages after the kernel is up */
1588 files_maxfiles_init();
1590 #ifdef CONFIG_ARCH_DISCARD_MEMBLOCK
1591 /* Discard memblock private memory */
1595 for_each_populated_zone(zone
)
1596 set_zone_contiguous(zone
);
1600 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1601 void __init
init_cma_reserved_pageblock(struct page
*page
)
1603 unsigned i
= pageblock_nr_pages
;
1604 struct page
*p
= page
;
1607 __ClearPageReserved(p
);
1608 set_page_count(p
, 0);
1611 set_pageblock_migratetype(page
, MIGRATE_CMA
);
1613 if (pageblock_order
>= MAX_ORDER
) {
1614 i
= pageblock_nr_pages
;
1617 set_page_refcounted(p
);
1618 __free_pages(p
, MAX_ORDER
- 1);
1619 p
+= MAX_ORDER_NR_PAGES
;
1620 } while (i
-= MAX_ORDER_NR_PAGES
);
1622 set_page_refcounted(page
);
1623 __free_pages(page
, pageblock_order
);
1626 adjust_managed_page_count(page
, pageblock_nr_pages
);
1631 * The order of subdivision here is critical for the IO subsystem.
1632 * Please do not alter this order without good reasons and regression
1633 * testing. Specifically, as large blocks of memory are subdivided,
1634 * the order in which smaller blocks are delivered depends on the order
1635 * they're subdivided in this function. This is the primary factor
1636 * influencing the order in which pages are delivered to the IO
1637 * subsystem according to empirical testing, and this is also justified
1638 * by considering the behavior of a buddy system containing a single
1639 * large block of memory acted on by a series of small allocations.
1640 * This behavior is a critical factor in sglist merging's success.
1644 static inline void expand(struct zone
*zone
, struct page
*page
,
1645 int low
, int high
, struct free_area
*area
,
1648 unsigned long size
= 1 << high
;
1650 while (high
> low
) {
1654 VM_BUG_ON_PAGE(bad_range(zone
, &page
[size
]), &page
[size
]);
1657 * Mark as guard pages (or page), that will allow to
1658 * merge back to allocator when buddy will be freed.
1659 * Corresponding page table entries will not be touched,
1660 * pages will stay not present in virtual address space
1662 if (set_page_guard(zone
, &page
[size
], high
, migratetype
))
1665 list_add(&page
[size
].lru
, &area
->free_list
[migratetype
]);
1667 set_page_order(&page
[size
], high
);
1671 static void check_new_page_bad(struct page
*page
)
1673 const char *bad_reason
= NULL
;
1674 unsigned long bad_flags
= 0;
1676 if (unlikely(atomic_read(&page
->_mapcount
) != -1))
1677 bad_reason
= "nonzero mapcount";
1678 if (unlikely(page
->mapping
!= NULL
))
1679 bad_reason
= "non-NULL mapping";
1680 if (unlikely(page_ref_count(page
) != 0))
1681 bad_reason
= "nonzero _count";
1682 if (unlikely(page
->flags
& __PG_HWPOISON
)) {
1683 bad_reason
= "HWPoisoned (hardware-corrupted)";
1684 bad_flags
= __PG_HWPOISON
;
1685 /* Don't complain about hwpoisoned pages */
1686 page_mapcount_reset(page
); /* remove PageBuddy */
1689 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)) {
1690 bad_reason
= "PAGE_FLAGS_CHECK_AT_PREP flag set";
1691 bad_flags
= PAGE_FLAGS_CHECK_AT_PREP
;
1694 if (unlikely(page
->mem_cgroup
))
1695 bad_reason
= "page still charged to cgroup";
1697 bad_page(page
, bad_reason
, bad_flags
);
1701 * This page is about to be returned from the page allocator
1703 static inline int check_new_page(struct page
*page
)
1705 if (likely(page_expected_state(page
,
1706 PAGE_FLAGS_CHECK_AT_PREP
|__PG_HWPOISON
)))
1709 check_new_page_bad(page
);
1713 static inline bool free_pages_prezeroed(void)
1715 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO
) &&
1716 page_poisoning_enabled();
1719 #ifdef CONFIG_DEBUG_VM
1720 static bool check_pcp_refill(struct page
*page
)
1725 static bool check_new_pcp(struct page
*page
)
1727 return check_new_page(page
);
1730 static bool check_pcp_refill(struct page
*page
)
1732 return check_new_page(page
);
1734 static bool check_new_pcp(struct page
*page
)
1738 #endif /* CONFIG_DEBUG_VM */
1740 static bool check_new_pages(struct page
*page
, unsigned int order
)
1743 for (i
= 0; i
< (1 << order
); i
++) {
1744 struct page
*p
= page
+ i
;
1746 if (unlikely(check_new_page(p
)))
1753 inline void post_alloc_hook(struct page
*page
, unsigned int order
,
1756 set_page_private(page
, 0);
1757 set_page_refcounted(page
);
1759 arch_alloc_page(page
, order
);
1760 kernel_map_pages(page
, 1 << order
, 1);
1761 kernel_poison_pages(page
, 1 << order
, 1);
1762 kasan_alloc_pages(page
, order
);
1763 set_page_owner(page
, order
, gfp_flags
);
1766 static void prep_new_page(struct page
*page
, unsigned int order
, gfp_t gfp_flags
,
1767 unsigned int alloc_flags
)
1771 post_alloc_hook(page
, order
, gfp_flags
);
1773 if (!free_pages_prezeroed() && (gfp_flags
& __GFP_ZERO
))
1774 for (i
= 0; i
< (1 << order
); i
++)
1775 clear_highpage(page
+ i
);
1777 if (order
&& (gfp_flags
& __GFP_COMP
))
1778 prep_compound_page(page
, order
);
1781 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1782 * allocate the page. The expectation is that the caller is taking
1783 * steps that will free more memory. The caller should avoid the page
1784 * being used for !PFMEMALLOC purposes.
1786 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
1787 set_page_pfmemalloc(page
);
1789 clear_page_pfmemalloc(page
);
1793 * Go through the free lists for the given migratetype and remove
1794 * the smallest available page from the freelists
1797 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
1800 unsigned int current_order
;
1801 struct free_area
*area
;
1804 /* Find a page of the appropriate size in the preferred list */
1805 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
1806 area
= &(zone
->free_area
[current_order
]);
1807 page
= list_first_entry_or_null(&area
->free_list
[migratetype
],
1811 list_del(&page
->lru
);
1812 rmv_page_order(page
);
1814 expand(zone
, page
, order
, current_order
, area
, migratetype
);
1815 set_pcppage_migratetype(page
, migratetype
);
1824 * This array describes the order lists are fallen back to when
1825 * the free lists for the desirable migrate type are depleted
1827 static int fallbacks
[MIGRATE_TYPES
][4] = {
1828 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_TYPES
},
1829 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_TYPES
},
1830 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_TYPES
},
1832 [MIGRATE_CMA
] = { MIGRATE_TYPES
}, /* Never used */
1834 #ifdef CONFIG_MEMORY_ISOLATION
1835 [MIGRATE_ISOLATE
] = { MIGRATE_TYPES
}, /* Never used */
1840 static struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
1843 return __rmqueue_smallest(zone
, order
, MIGRATE_CMA
);
1846 static inline struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
1847 unsigned int order
) { return NULL
; }
1851 * Move the free pages in a range to the free lists of the requested type.
1852 * Note that start_page and end_pages are not aligned on a pageblock
1853 * boundary. If alignment is required, use move_freepages_block()
1855 static int move_freepages(struct zone
*zone
,
1856 struct page
*start_page
, struct page
*end_page
,
1857 int migratetype
, int *num_movable
)
1861 int pages_moved
= 0;
1863 #ifndef CONFIG_HOLES_IN_ZONE
1865 * page_zone is not safe to call in this context when
1866 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1867 * anyway as we check zone boundaries in move_freepages_block().
1868 * Remove at a later date when no bug reports exist related to
1869 * grouping pages by mobility
1871 VM_BUG_ON(page_zone(start_page
) != page_zone(end_page
));
1877 for (page
= start_page
; page
<= end_page
;) {
1878 if (!pfn_valid_within(page_to_pfn(page
))) {
1883 /* Make sure we are not inadvertently changing nodes */
1884 VM_BUG_ON_PAGE(page_to_nid(page
) != zone_to_nid(zone
), page
);
1886 if (!PageBuddy(page
)) {
1888 * We assume that pages that could be isolated for
1889 * migration are movable. But we don't actually try
1890 * isolating, as that would be expensive.
1893 (PageLRU(page
) || __PageMovable(page
)))
1900 order
= page_order(page
);
1901 list_move(&page
->lru
,
1902 &zone
->free_area
[order
].free_list
[migratetype
]);
1904 pages_moved
+= 1 << order
;
1910 int move_freepages_block(struct zone
*zone
, struct page
*page
,
1911 int migratetype
, int *num_movable
)
1913 unsigned long start_pfn
, end_pfn
;
1914 struct page
*start_page
, *end_page
;
1916 start_pfn
= page_to_pfn(page
);
1917 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
1918 start_page
= pfn_to_page(start_pfn
);
1919 end_page
= start_page
+ pageblock_nr_pages
- 1;
1920 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
1922 /* Do not cross zone boundaries */
1923 if (!zone_spans_pfn(zone
, start_pfn
))
1925 if (!zone_spans_pfn(zone
, end_pfn
))
1928 return move_freepages(zone
, start_page
, end_page
, migratetype
,
1932 static void change_pageblock_range(struct page
*pageblock_page
,
1933 int start_order
, int migratetype
)
1935 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
1937 while (nr_pageblocks
--) {
1938 set_pageblock_migratetype(pageblock_page
, migratetype
);
1939 pageblock_page
+= pageblock_nr_pages
;
1944 * When we are falling back to another migratetype during allocation, try to
1945 * steal extra free pages from the same pageblocks to satisfy further
1946 * allocations, instead of polluting multiple pageblocks.
1948 * If we are stealing a relatively large buddy page, it is likely there will
1949 * be more free pages in the pageblock, so try to steal them all. For
1950 * reclaimable and unmovable allocations, we steal regardless of page size,
1951 * as fragmentation caused by those allocations polluting movable pageblocks
1952 * is worse than movable allocations stealing from unmovable and reclaimable
1955 static bool can_steal_fallback(unsigned int order
, int start_mt
)
1958 * Leaving this order check is intended, although there is
1959 * relaxed order check in next check. The reason is that
1960 * we can actually steal whole pageblock if this condition met,
1961 * but, below check doesn't guarantee it and that is just heuristic
1962 * so could be changed anytime.
1964 if (order
>= pageblock_order
)
1967 if (order
>= pageblock_order
/ 2 ||
1968 start_mt
== MIGRATE_RECLAIMABLE
||
1969 start_mt
== MIGRATE_UNMOVABLE
||
1970 page_group_by_mobility_disabled
)
1977 * This function implements actual steal behaviour. If order is large enough,
1978 * we can steal whole pageblock. If not, we first move freepages in this
1979 * pageblock to our migratetype and determine how many already-allocated pages
1980 * are there in the pageblock with a compatible migratetype. If at least half
1981 * of pages are free or compatible, we can change migratetype of the pageblock
1982 * itself, so pages freed in the future will be put on the correct free list.
1984 static void steal_suitable_fallback(struct zone
*zone
, struct page
*page
,
1985 int start_type
, bool whole_block
)
1987 unsigned int current_order
= page_order(page
);
1988 struct free_area
*area
;
1989 int free_pages
, movable_pages
, alike_pages
;
1992 old_block_type
= get_pageblock_migratetype(page
);
1995 * This can happen due to races and we want to prevent broken
1996 * highatomic accounting.
1998 if (is_migrate_highatomic(old_block_type
))
2001 /* Take ownership for orders >= pageblock_order */
2002 if (current_order
>= pageblock_order
) {
2003 change_pageblock_range(page
, current_order
, start_type
);
2007 /* We are not allowed to try stealing from the whole block */
2011 free_pages
= move_freepages_block(zone
, page
, start_type
,
2014 * Determine how many pages are compatible with our allocation.
2015 * For movable allocation, it's the number of movable pages which
2016 * we just obtained. For other types it's a bit more tricky.
2018 if (start_type
== MIGRATE_MOVABLE
) {
2019 alike_pages
= movable_pages
;
2022 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2023 * to MOVABLE pageblock, consider all non-movable pages as
2024 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2025 * vice versa, be conservative since we can't distinguish the
2026 * exact migratetype of non-movable pages.
2028 if (old_block_type
== MIGRATE_MOVABLE
)
2029 alike_pages
= pageblock_nr_pages
2030 - (free_pages
+ movable_pages
);
2035 /* moving whole block can fail due to zone boundary conditions */
2040 * If a sufficient number of pages in the block are either free or of
2041 * comparable migratability as our allocation, claim the whole block.
2043 if (free_pages
+ alike_pages
>= (1 << (pageblock_order
-1)) ||
2044 page_group_by_mobility_disabled
)
2045 set_pageblock_migratetype(page
, start_type
);
2050 area
= &zone
->free_area
[current_order
];
2051 list_move(&page
->lru
, &area
->free_list
[start_type
]);
2055 * Check whether there is a suitable fallback freepage with requested order.
2056 * If only_stealable is true, this function returns fallback_mt only if
2057 * we can steal other freepages all together. This would help to reduce
2058 * fragmentation due to mixed migratetype pages in one pageblock.
2060 int find_suitable_fallback(struct free_area
*area
, unsigned int order
,
2061 int migratetype
, bool only_stealable
, bool *can_steal
)
2066 if (area
->nr_free
== 0)
2071 fallback_mt
= fallbacks
[migratetype
][i
];
2072 if (fallback_mt
== MIGRATE_TYPES
)
2075 if (list_empty(&area
->free_list
[fallback_mt
]))
2078 if (can_steal_fallback(order
, migratetype
))
2081 if (!only_stealable
)
2092 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2093 * there are no empty page blocks that contain a page with a suitable order
2095 static void reserve_highatomic_pageblock(struct page
*page
, struct zone
*zone
,
2096 unsigned int alloc_order
)
2099 unsigned long max_managed
, flags
;
2102 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2103 * Check is race-prone but harmless.
2105 max_managed
= (zone
->managed_pages
/ 100) + pageblock_nr_pages
;
2106 if (zone
->nr_reserved_highatomic
>= max_managed
)
2109 spin_lock_irqsave(&zone
->lock
, flags
);
2111 /* Recheck the nr_reserved_highatomic limit under the lock */
2112 if (zone
->nr_reserved_highatomic
>= max_managed
)
2116 mt
= get_pageblock_migratetype(page
);
2117 if (!is_migrate_highatomic(mt
) && !is_migrate_isolate(mt
)
2118 && !is_migrate_cma(mt
)) {
2119 zone
->nr_reserved_highatomic
+= pageblock_nr_pages
;
2120 set_pageblock_migratetype(page
, MIGRATE_HIGHATOMIC
);
2121 move_freepages_block(zone
, page
, MIGRATE_HIGHATOMIC
, NULL
);
2125 spin_unlock_irqrestore(&zone
->lock
, flags
);
2129 * Used when an allocation is about to fail under memory pressure. This
2130 * potentially hurts the reliability of high-order allocations when under
2131 * intense memory pressure but failed atomic allocations should be easier
2132 * to recover from than an OOM.
2134 * If @force is true, try to unreserve a pageblock even though highatomic
2135 * pageblock is exhausted.
2137 static bool unreserve_highatomic_pageblock(const struct alloc_context
*ac
,
2140 struct zonelist
*zonelist
= ac
->zonelist
;
2141 unsigned long flags
;
2148 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, ac
->high_zoneidx
,
2151 * Preserve at least one pageblock unless memory pressure
2154 if (!force
&& zone
->nr_reserved_highatomic
<=
2158 spin_lock_irqsave(&zone
->lock
, flags
);
2159 for (order
= 0; order
< MAX_ORDER
; order
++) {
2160 struct free_area
*area
= &(zone
->free_area
[order
]);
2162 page
= list_first_entry_or_null(
2163 &area
->free_list
[MIGRATE_HIGHATOMIC
],
2169 * In page freeing path, migratetype change is racy so
2170 * we can counter several free pages in a pageblock
2171 * in this loop althoug we changed the pageblock type
2172 * from highatomic to ac->migratetype. So we should
2173 * adjust the count once.
2175 if (is_migrate_highatomic_page(page
)) {
2177 * It should never happen but changes to
2178 * locking could inadvertently allow a per-cpu
2179 * drain to add pages to MIGRATE_HIGHATOMIC
2180 * while unreserving so be safe and watch for
2183 zone
->nr_reserved_highatomic
-= min(
2185 zone
->nr_reserved_highatomic
);
2189 * Convert to ac->migratetype and avoid the normal
2190 * pageblock stealing heuristics. Minimally, the caller
2191 * is doing the work and needs the pages. More
2192 * importantly, if the block was always converted to
2193 * MIGRATE_UNMOVABLE or another type then the number
2194 * of pageblocks that cannot be completely freed
2197 set_pageblock_migratetype(page
, ac
->migratetype
);
2198 ret
= move_freepages_block(zone
, page
, ac
->migratetype
,
2201 spin_unlock_irqrestore(&zone
->lock
, flags
);
2205 spin_unlock_irqrestore(&zone
->lock
, flags
);
2212 * Try finding a free buddy page on the fallback list and put it on the free
2213 * list of requested migratetype, possibly along with other pages from the same
2214 * block, depending on fragmentation avoidance heuristics. Returns true if
2215 * fallback was found so that __rmqueue_smallest() can grab it.
2217 * The use of signed ints for order and current_order is a deliberate
2218 * deviation from the rest of this file, to make the for loop
2219 * condition simpler.
2222 __rmqueue_fallback(struct zone
*zone
, int order
, int start_migratetype
)
2224 struct free_area
*area
;
2231 * Find the largest available free page in the other list. This roughly
2232 * approximates finding the pageblock with the most free pages, which
2233 * would be too costly to do exactly.
2235 for (current_order
= MAX_ORDER
- 1; current_order
>= order
;
2237 area
= &(zone
->free_area
[current_order
]);
2238 fallback_mt
= find_suitable_fallback(area
, current_order
,
2239 start_migratetype
, false, &can_steal
);
2240 if (fallback_mt
== -1)
2244 * We cannot steal all free pages from the pageblock and the
2245 * requested migratetype is movable. In that case it's better to
2246 * steal and split the smallest available page instead of the
2247 * largest available page, because even if the next movable
2248 * allocation falls back into a different pageblock than this
2249 * one, it won't cause permanent fragmentation.
2251 if (!can_steal
&& start_migratetype
== MIGRATE_MOVABLE
2252 && current_order
> order
)
2261 for (current_order
= order
; current_order
< MAX_ORDER
;
2263 area
= &(zone
->free_area
[current_order
]);
2264 fallback_mt
= find_suitable_fallback(area
, current_order
,
2265 start_migratetype
, false, &can_steal
);
2266 if (fallback_mt
!= -1)
2271 * This should not happen - we already found a suitable fallback
2272 * when looking for the largest page.
2274 VM_BUG_ON(current_order
== MAX_ORDER
);
2277 page
= list_first_entry(&area
->free_list
[fallback_mt
],
2280 steal_suitable_fallback(zone
, page
, start_migratetype
, can_steal
);
2282 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
2283 start_migratetype
, fallback_mt
);
2290 * Do the hard work of removing an element from the buddy allocator.
2291 * Call me with the zone->lock already held.
2292 * If gfp mask of the page allocation has GFP_HIGHUSER_MOVABLE, @migratetype
2293 * is changed from MIGRATE_MOVABLE to MIGRATE_CMA in rmqueue() to select the
2294 * free list of MIGRATE_CMA. It helps depleting CMA free pages so that
2295 * evaluation of watermark for unmovable page allocations is not too different
2296 * from movable page allocations.
2297 * If @migratetype is MIGRATE_CMA, it should be corrected to MIGRATE_MOVABLE
2298 * after the free list of MIGRATE_CMA is searched to have a chance to search the
2299 * free list of MIGRATE_MOVABLE. It also records correct migrate type in the
2300 * trace as intended by the page allocation.
2302 static struct page
*__rmqueue(struct zone
*zone
, unsigned int order
,
2305 struct page
*page
= NULL
;
2308 if (migratetype
== MIGRATE_CMA
) {
2310 if (migratetype
== MIGRATE_MOVABLE
) {
2312 page
= __rmqueue_cma_fallback(zone
, order
);
2313 migratetype
= MIGRATE_MOVABLE
;
2317 page
= __rmqueue_smallest(zone
, order
, migratetype
);
2318 if (unlikely(!page
) &&
2319 !__rmqueue_fallback(zone
, order
, migratetype
))
2323 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
2328 * Obtain a specified number of elements from the buddy allocator, all under
2329 * a single hold of the lock, for efficiency. Add them to the supplied list.
2330 * Returns the number of new pages which were placed at *list.
2332 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
2333 unsigned long count
, struct list_head
*list
,
2334 int migratetype
, bool cold
)
2338 spin_lock(&zone
->lock
);
2339 for (i
= 0; i
< count
; ++i
) {
2340 struct page
*page
= __rmqueue(zone
, order
, migratetype
);
2341 if (unlikely(page
== NULL
))
2344 if (unlikely(check_pcp_refill(page
)))
2348 * Split buddy pages returned by expand() are received here
2349 * in physical page order. The page is added to the callers and
2350 * list and the list head then moves forward. From the callers
2351 * perspective, the linked list is ordered by page number in
2352 * some conditions. This is useful for IO devices that can
2353 * merge IO requests if the physical pages are ordered
2357 list_add(&page
->lru
, list
);
2359 list_add_tail(&page
->lru
, list
);
2362 if (is_migrate_cma(get_pcppage_migratetype(page
)))
2363 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
,
2368 * i pages were removed from the buddy list even if some leak due
2369 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2370 * on i. Do not confuse with 'alloced' which is the number of
2371 * pages added to the pcp list.
2373 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
2374 spin_unlock(&zone
->lock
);
2380 * Called from the vmstat counter updater to drain pagesets of this
2381 * currently executing processor on remote nodes after they have
2384 * Note that this function must be called with the thread pinned to
2385 * a single processor.
2387 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
2389 unsigned long flags
;
2390 int to_drain
, batch
;
2392 local_irq_save(flags
);
2393 batch
= READ_ONCE(pcp
->batch
);
2394 to_drain
= min(pcp
->count
, batch
);
2396 free_pcppages_bulk(zone
, to_drain
, pcp
);
2397 pcp
->count
-= to_drain
;
2399 local_irq_restore(flags
);
2404 * Drain pcplists of the indicated processor and zone.
2406 * The processor must either be the current processor and the
2407 * thread pinned to the current processor or a processor that
2410 static void drain_pages_zone(unsigned int cpu
, struct zone
*zone
)
2412 unsigned long flags
;
2413 struct per_cpu_pageset
*pset
;
2414 struct per_cpu_pages
*pcp
;
2416 local_irq_save(flags
);
2417 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
2421 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
2424 local_irq_restore(flags
);
2428 * Drain pcplists of all zones on the indicated processor.
2430 * The processor must either be the current processor and the
2431 * thread pinned to the current processor or a processor that
2434 static void drain_pages(unsigned int cpu
)
2438 for_each_populated_zone(zone
) {
2439 drain_pages_zone(cpu
, zone
);
2444 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2446 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2447 * the single zone's pages.
2449 void drain_local_pages(struct zone
*zone
)
2451 int cpu
= smp_processor_id();
2454 drain_pages_zone(cpu
, zone
);
2459 static void drain_local_pages_wq(struct work_struct
*work
)
2462 * drain_all_pages doesn't use proper cpu hotplug protection so
2463 * we can race with cpu offline when the WQ can move this from
2464 * a cpu pinned worker to an unbound one. We can operate on a different
2465 * cpu which is allright but we also have to make sure to not move to
2469 drain_local_pages(NULL
);
2474 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2476 * When zone parameter is non-NULL, spill just the single zone's pages.
2478 * Note that this can be extremely slow as the draining happens in a workqueue.
2480 void drain_all_pages(struct zone
*zone
)
2485 * Allocate in the BSS so we wont require allocation in
2486 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2488 static cpumask_t cpus_with_pcps
;
2491 * Make sure nobody triggers this path before mm_percpu_wq is fully
2494 if (WARN_ON_ONCE(!mm_percpu_wq
))
2498 * Do not drain if one is already in progress unless it's specific to
2499 * a zone. Such callers are primarily CMA and memory hotplug and need
2500 * the drain to be complete when the call returns.
2502 if (unlikely(!mutex_trylock(&pcpu_drain_mutex
))) {
2505 mutex_lock(&pcpu_drain_mutex
);
2509 * We don't care about racing with CPU hotplug event
2510 * as offline notification will cause the notified
2511 * cpu to drain that CPU pcps and on_each_cpu_mask
2512 * disables preemption as part of its processing
2514 for_each_online_cpu(cpu
) {
2515 struct per_cpu_pageset
*pcp
;
2517 bool has_pcps
= false;
2520 pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
2524 for_each_populated_zone(z
) {
2525 pcp
= per_cpu_ptr(z
->pageset
, cpu
);
2526 if (pcp
->pcp
.count
) {
2534 cpumask_set_cpu(cpu
, &cpus_with_pcps
);
2536 cpumask_clear_cpu(cpu
, &cpus_with_pcps
);
2539 for_each_cpu(cpu
, &cpus_with_pcps
) {
2540 struct work_struct
*work
= per_cpu_ptr(&pcpu_drain
, cpu
);
2541 INIT_WORK(work
, drain_local_pages_wq
);
2542 queue_work_on(cpu
, mm_percpu_wq
, work
);
2544 for_each_cpu(cpu
, &cpus_with_pcps
)
2545 flush_work(per_cpu_ptr(&pcpu_drain
, cpu
));
2547 mutex_unlock(&pcpu_drain_mutex
);
2550 #ifdef CONFIG_HIBERNATION
2553 * Touch the watchdog for every WD_PAGE_COUNT pages.
2555 #define WD_PAGE_COUNT (128*1024)
2557 void mark_free_pages(struct zone
*zone
)
2559 unsigned long pfn
, max_zone_pfn
, page_count
= WD_PAGE_COUNT
;
2560 unsigned long flags
;
2561 unsigned int order
, t
;
2564 if (zone_is_empty(zone
))
2567 spin_lock_irqsave(&zone
->lock
, flags
);
2569 max_zone_pfn
= zone_end_pfn(zone
);
2570 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
2571 if (pfn_valid(pfn
)) {
2572 page
= pfn_to_page(pfn
);
2574 if (!--page_count
) {
2575 touch_nmi_watchdog();
2576 page_count
= WD_PAGE_COUNT
;
2579 if (page_zone(page
) != zone
)
2582 if (!swsusp_page_is_forbidden(page
))
2583 swsusp_unset_page_free(page
);
2586 for_each_migratetype_order(order
, t
) {
2587 list_for_each_entry(page
,
2588 &zone
->free_area
[order
].free_list
[t
], lru
) {
2591 pfn
= page_to_pfn(page
);
2592 for (i
= 0; i
< (1UL << order
); i
++) {
2593 if (!--page_count
) {
2594 touch_nmi_watchdog();
2595 page_count
= WD_PAGE_COUNT
;
2597 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
2601 spin_unlock_irqrestore(&zone
->lock
, flags
);
2603 #endif /* CONFIG_PM */
2606 * Free a 0-order page
2607 * cold == true ? free a cold page : free a hot page
2609 void free_hot_cold_page(struct page
*page
, bool cold
)
2611 struct zone
*zone
= page_zone(page
);
2612 struct per_cpu_pages
*pcp
;
2613 unsigned long flags
;
2614 unsigned long pfn
= page_to_pfn(page
);
2617 if (!free_pcp_prepare(page
))
2620 migratetype
= get_pfnblock_migratetype(page
, pfn
);
2621 set_pcppage_migratetype(page
, migratetype
);
2622 local_irq_save(flags
);
2623 __count_vm_event(PGFREE
);
2626 * We only track unmovable, reclaimable and movable on pcp lists.
2627 * Free ISOLATE pages back to the allocator because they are being
2628 * offlined but treat HIGHATOMIC as movable pages so we can get those
2629 * areas back if necessary. Otherwise, we may have to free
2630 * excessively into the page allocator
2632 if (migratetype
>= MIGRATE_PCPTYPES
) {
2633 if (unlikely(is_migrate_isolate(migratetype
))) {
2634 free_one_page(zone
, page
, pfn
, 0, migratetype
);
2637 migratetype
= MIGRATE_MOVABLE
;
2640 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
2642 list_add(&page
->lru
, &pcp
->lists
[migratetype
]);
2644 list_add_tail(&page
->lru
, &pcp
->lists
[migratetype
]);
2646 if (pcp
->count
>= pcp
->high
) {
2647 unsigned long batch
= READ_ONCE(pcp
->batch
);
2648 free_pcppages_bulk(zone
, batch
, pcp
);
2649 pcp
->count
-= batch
;
2653 local_irq_restore(flags
);
2657 * Free a list of 0-order pages
2659 void free_hot_cold_page_list(struct list_head
*list
, bool cold
)
2661 struct page
*page
, *next
;
2663 list_for_each_entry_safe(page
, next
, list
, lru
) {
2664 trace_mm_page_free_batched(page
, cold
);
2665 free_hot_cold_page(page
, cold
);
2670 * split_page takes a non-compound higher-order page, and splits it into
2671 * n (1<<order) sub-pages: page[0..n]
2672 * Each sub-page must be freed individually.
2674 * Note: this is probably too low level an operation for use in drivers.
2675 * Please consult with lkml before using this in your driver.
2677 void split_page(struct page
*page
, unsigned int order
)
2681 VM_BUG_ON_PAGE(PageCompound(page
), page
);
2682 VM_BUG_ON_PAGE(!page_count(page
), page
);
2684 for (i
= 1; i
< (1 << order
); i
++)
2685 set_page_refcounted(page
+ i
);
2686 split_page_owner(page
, order
);
2688 EXPORT_SYMBOL_GPL(split_page
);
2690 int __isolate_free_page(struct page
*page
, unsigned int order
)
2692 unsigned long watermark
;
2696 BUG_ON(!PageBuddy(page
));
2698 zone
= page_zone(page
);
2699 mt
= get_pageblock_migratetype(page
);
2701 if (!is_migrate_isolate(mt
)) {
2703 * Obey watermarks as if the page was being allocated. We can
2704 * emulate a high-order watermark check with a raised order-0
2705 * watermark, because we already know our high-order page
2708 watermark
= min_wmark_pages(zone
) + (1UL << order
);
2709 if (!zone_watermark_ok(zone
, 0, watermark
, 0, ALLOC_CMA
))
2712 __mod_zone_freepage_state(zone
, -(1UL << order
), mt
);
2715 /* Remove page from free list */
2716 list_del(&page
->lru
);
2717 zone
->free_area
[order
].nr_free
--;
2718 rmv_page_order(page
);
2721 * Set the pageblock if the isolated page is at least half of a
2724 if (order
>= pageblock_order
- 1) {
2725 struct page
*endpage
= page
+ (1 << order
) - 1;
2726 for (; page
< endpage
; page
+= pageblock_nr_pages
) {
2727 int mt
= get_pageblock_migratetype(page
);
2728 if (!is_migrate_isolate(mt
) && !is_migrate_cma(mt
)
2729 && !is_migrate_highatomic(mt
))
2730 set_pageblock_migratetype(page
,
2736 return 1UL << order
;
2740 * Update NUMA hit/miss statistics
2742 * Must be called with interrupts disabled.
2744 static inline void zone_statistics(struct zone
*preferred_zone
, struct zone
*z
)
2747 enum numa_stat_item local_stat
= NUMA_LOCAL
;
2749 if (z
->node
!= numa_node_id())
2750 local_stat
= NUMA_OTHER
;
2752 if (z
->node
== preferred_zone
->node
)
2753 __inc_numa_state(z
, NUMA_HIT
);
2755 __inc_numa_state(z
, NUMA_MISS
);
2756 __inc_numa_state(preferred_zone
, NUMA_FOREIGN
);
2758 __inc_numa_state(z
, local_stat
);
2762 /* Remove page from the per-cpu list, caller must protect the list */
2763 static struct page
*__rmqueue_pcplist(struct zone
*zone
, int migratetype
,
2764 bool cold
, struct per_cpu_pages
*pcp
,
2765 struct list_head
*list
)
2770 if (list_empty(list
)) {
2771 pcp
->count
+= rmqueue_bulk(zone
, 0,
2774 if (unlikely(list_empty(list
)))
2779 page
= list_last_entry(list
, struct page
, lru
);
2781 page
= list_first_entry(list
, struct page
, lru
);
2784 * If the head or the tail page in the pcp list is CMA page and
2785 * the gfp flags is not GFP_HIGHUSER_MOVABLE, do not allocate a
2786 * page from the pcp list. The free list of MIGRATE_CMA is a
2787 * special case of the free list of MIGRATE_MOVABLE and the
2788 * pages from the free list of MIGRATE_CMA are pushed to the pcp
2789 * list of MIGRATE_MOVABLE. Since the pcp list of
2790 * MIGRATE_MOVABLE is selected if the gfp flags has GFP_MOVABLE,
2791 * we should avoid the case that a cma page in the pcp list of
2792 * MIGRATE_MOVABLE is allocated to a movable allocation without
2793 * GFP_HIGHUSER_MOVABLE.
2794 * If this is the case, allocate a movable page from the free
2795 * list of MIGRATE_MOVABLE instead of pcp list of
2799 if (is_migrate_cma_page(page
) && (migratetype
!= MIGRATE_CMA
))
2802 list_del(&page
->lru
);
2804 } while (check_new_pcp(page
));
2809 /* Lock and remove page from the per-cpu list */
2810 static struct page
*rmqueue_pcplist(struct zone
*preferred_zone
,
2811 struct zone
*zone
, unsigned int order
,
2812 gfp_t gfp_flags
, int migratetype
,
2813 int migratetype_rmqueue
)
2815 struct per_cpu_pages
*pcp
;
2816 struct list_head
*list
;
2817 bool cold
= ((gfp_flags
& __GFP_COLD
) != 0);
2819 unsigned long flags
;
2821 local_irq_save(flags
);
2822 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
2823 list
= &pcp
->lists
[migratetype
];
2824 page
= __rmqueue_pcplist(zone
, migratetype_rmqueue
, cold
, pcp
, list
);
2826 __count_zid_vm_events(PGALLOC
, page_zonenum(page
), 1 << order
);
2827 zone_statistics(preferred_zone
, zone
);
2829 local_irq_restore(flags
);
2834 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2837 struct page
*rmqueue(struct zone
*preferred_zone
,
2838 struct zone
*zone
, unsigned int order
,
2839 gfp_t gfp_flags
, unsigned int alloc_flags
,
2842 unsigned long flags
;
2844 int migratetype_rmqueue
= migratetype
;
2847 if ((migratetype_rmqueue
== MIGRATE_MOVABLE
) &&
2848 ((gfp_flags
& GFP_HIGHUSER_MOVABLE
) == GFP_HIGHUSER_MOVABLE
))
2849 migratetype_rmqueue
= MIGRATE_CMA
;
2851 if (likely(order
== 0)) {
2852 page
= rmqueue_pcplist(preferred_zone
, zone
, order
,
2853 gfp_flags
, migratetype
, migratetype_rmqueue
);
2855 * Allocation with GFP_MOVABLE and !GFP_HIGHMEM will have
2856 * another chance of page allocation from the free list.
2857 * See the comment in __rmqueue_pcplist().
2860 if (likely(page
) || (migratetype_rmqueue
!= MIGRATE_MOVABLE
))
2866 * We most definitely don't want callers attempting to
2867 * allocate greater than order-1 page units with __GFP_NOFAIL.
2869 WARN_ON_ONCE((gfp_flags
& __GFP_NOFAIL
) && (order
> 1));
2870 spin_lock_irqsave(&zone
->lock
, flags
);
2874 if (alloc_flags
& ALLOC_HARDER
) {
2875 page
= __rmqueue_smallest(zone
, order
, MIGRATE_HIGHATOMIC
);
2877 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
2880 page
= __rmqueue(zone
, order
, migratetype_rmqueue
);
2881 } while (page
&& check_new_pages(page
, order
));
2882 spin_unlock(&zone
->lock
);
2885 __mod_zone_freepage_state(zone
, -(1 << order
),
2886 get_pcppage_migratetype(page
));
2888 __count_zid_vm_events(PGALLOC
, page_zonenum(page
), 1 << order
);
2889 zone_statistics(preferred_zone
, zone
);
2890 local_irq_restore(flags
);
2893 VM_BUG_ON_PAGE(page
&& bad_range(zone
, page
), page
);
2897 local_irq_restore(flags
);
2901 #ifdef CONFIG_FAIL_PAGE_ALLOC
2904 struct fault_attr attr
;
2906 bool ignore_gfp_highmem
;
2907 bool ignore_gfp_reclaim
;
2909 } fail_page_alloc
= {
2910 .attr
= FAULT_ATTR_INITIALIZER
,
2911 .ignore_gfp_reclaim
= true,
2912 .ignore_gfp_highmem
= true,
2916 static int __init
setup_fail_page_alloc(char *str
)
2918 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
2920 __setup("fail_page_alloc=", setup_fail_page_alloc
);
2922 static bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
2924 if (order
< fail_page_alloc
.min_order
)
2926 if (gfp_mask
& __GFP_NOFAIL
)
2928 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
2930 if (fail_page_alloc
.ignore_gfp_reclaim
&&
2931 (gfp_mask
& __GFP_DIRECT_RECLAIM
))
2934 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
2937 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2939 static int __init
fail_page_alloc_debugfs(void)
2941 umode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
2944 dir
= fault_create_debugfs_attr("fail_page_alloc", NULL
,
2945 &fail_page_alloc
.attr
);
2947 return PTR_ERR(dir
);
2949 if (!debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
2950 &fail_page_alloc
.ignore_gfp_reclaim
))
2952 if (!debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
2953 &fail_page_alloc
.ignore_gfp_highmem
))
2955 if (!debugfs_create_u32("min-order", mode
, dir
,
2956 &fail_page_alloc
.min_order
))
2961 debugfs_remove_recursive(dir
);
2966 late_initcall(fail_page_alloc_debugfs
);
2968 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2970 #else /* CONFIG_FAIL_PAGE_ALLOC */
2972 static inline bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
2977 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2980 * Return true if free base pages are above 'mark'. For high-order checks it
2981 * will return true of the order-0 watermark is reached and there is at least
2982 * one free page of a suitable size. Checking now avoids taking the zone lock
2983 * to check in the allocation paths if no pages are free.
2985 bool __zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
2986 int classzone_idx
, unsigned int alloc_flags
,
2991 const bool alloc_harder
= (alloc_flags
& (ALLOC_HARDER
|ALLOC_OOM
));
2993 /* free_pages may go negative - that's OK */
2994 free_pages
-= (1 << order
) - 1;
2996 if (alloc_flags
& ALLOC_HIGH
)
3000 * If the caller does not have rights to ALLOC_HARDER then subtract
3001 * the high-atomic reserves. This will over-estimate the size of the
3002 * atomic reserve but it avoids a search.
3004 if (likely(!alloc_harder
)) {
3005 free_pages
-= z
->nr_reserved_highatomic
;
3008 * OOM victims can try even harder than normal ALLOC_HARDER
3009 * users on the grounds that it's definitely going to be in
3010 * the exit path shortly and free memory. Any allocation it
3011 * makes during the free path will be small and short-lived.
3013 if (alloc_flags
& ALLOC_OOM
)
3021 /* If allocation can't use CMA areas don't use free CMA pages */
3022 if (!(alloc_flags
& ALLOC_CMA
))
3023 free_pages
-= zone_page_state(z
, NR_FREE_CMA_PAGES
);
3027 * Check watermarks for an order-0 allocation request. If these
3028 * are not met, then a high-order request also cannot go ahead
3029 * even if a suitable page happened to be free.
3031 if (free_pages
<= min
+ z
->lowmem_reserve
[classzone_idx
])
3034 /* If this is an order-0 request then the watermark is fine */
3038 /* For a high-order request, check at least one suitable page is free */
3039 for (o
= order
; o
< MAX_ORDER
; o
++) {
3040 struct free_area
*area
= &z
->free_area
[o
];
3046 for (mt
= 0; mt
< MIGRATE_PCPTYPES
; mt
++) {
3047 if (!list_empty(&area
->free_list
[mt
]))
3052 if ((alloc_flags
& ALLOC_CMA
) &&
3053 !list_empty(&area
->free_list
[MIGRATE_CMA
])) {
3058 !list_empty(&area
->free_list
[MIGRATE_HIGHATOMIC
]))
3064 bool zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
3065 int classzone_idx
, unsigned int alloc_flags
)
3067 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
3068 zone_page_state(z
, NR_FREE_PAGES
));
3071 static inline bool zone_watermark_fast(struct zone
*z
, unsigned int order
,
3072 unsigned long mark
, int classzone_idx
, unsigned int alloc_flags
)
3074 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
3078 /* If allocation can't use CMA areas don't use free CMA pages */
3079 if (!(alloc_flags
& ALLOC_CMA
))
3080 cma_pages
= zone_page_state(z
, NR_FREE_CMA_PAGES
);
3084 * Fast check for order-0 only. If this fails then the reserves
3085 * need to be calculated. There is a corner case where the check
3086 * passes but only the high-order atomic reserve are free. If
3087 * the caller is !atomic then it'll uselessly search the free
3088 * list. That corner case is then slower but it is harmless.
3090 if (!order
&& (free_pages
- cma_pages
) > mark
+ z
->lowmem_reserve
[classzone_idx
])
3093 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
3097 bool zone_watermark_ok_safe(struct zone
*z
, unsigned int order
,
3098 unsigned long mark
, int classzone_idx
)
3100 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
3102 if (z
->percpu_drift_mark
&& free_pages
< z
->percpu_drift_mark
)
3103 free_pages
= zone_page_state_snapshot(z
, NR_FREE_PAGES
);
3105 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, 0,
3110 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
3112 return node_distance(zone_to_nid(local_zone
), zone_to_nid(zone
)) <=
3115 #else /* CONFIG_NUMA */
3116 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
3120 #endif /* CONFIG_NUMA */
3123 * get_page_from_freelist goes through the zonelist trying to allocate
3126 static struct page
*
3127 get_page_from_freelist(gfp_t gfp_mask
, unsigned int order
, int alloc_flags
,
3128 const struct alloc_context
*ac
)
3130 struct zoneref
*z
= ac
->preferred_zoneref
;
3132 struct pglist_data
*last_pgdat_dirty_limit
= NULL
;
3135 * Scan zonelist, looking for a zone with enough free.
3136 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3138 for_next_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
3143 if (cpusets_enabled() &&
3144 (alloc_flags
& ALLOC_CPUSET
) &&
3145 !__cpuset_zone_allowed(zone
, gfp_mask
))
3148 * When allocating a page cache page for writing, we
3149 * want to get it from a node that is within its dirty
3150 * limit, such that no single node holds more than its
3151 * proportional share of globally allowed dirty pages.
3152 * The dirty limits take into account the node's
3153 * lowmem reserves and high watermark so that kswapd
3154 * should be able to balance it without having to
3155 * write pages from its LRU list.
3157 * XXX: For now, allow allocations to potentially
3158 * exceed the per-node dirty limit in the slowpath
3159 * (spread_dirty_pages unset) before going into reclaim,
3160 * which is important when on a NUMA setup the allowed
3161 * nodes are together not big enough to reach the
3162 * global limit. The proper fix for these situations
3163 * will require awareness of nodes in the
3164 * dirty-throttling and the flusher threads.
3166 if (ac
->spread_dirty_pages
) {
3167 if (last_pgdat_dirty_limit
== zone
->zone_pgdat
)
3170 if (!node_dirty_ok(zone
->zone_pgdat
)) {
3171 last_pgdat_dirty_limit
= zone
->zone_pgdat
;
3176 mark
= zone
->watermark
[alloc_flags
& ALLOC_WMARK_MASK
];
3177 if (!zone_watermark_fast(zone
, order
, mark
,
3178 ac_classzone_idx(ac
), alloc_flags
)) {
3181 /* Checked here to keep the fast path fast */
3182 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
3183 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
3186 if (node_reclaim_mode
== 0 ||
3187 !zone_allows_reclaim(ac
->preferred_zoneref
->zone
, zone
))
3190 ret
= node_reclaim(zone
->zone_pgdat
, gfp_mask
, order
);
3192 case NODE_RECLAIM_NOSCAN
:
3195 case NODE_RECLAIM_FULL
:
3196 /* scanned but unreclaimable */
3199 /* did we reclaim enough */
3200 if (zone_watermark_ok(zone
, order
, mark
,
3201 ac_classzone_idx(ac
), alloc_flags
))
3209 page
= rmqueue(ac
->preferred_zoneref
->zone
, zone
, order
,
3210 gfp_mask
, alloc_flags
, ac
->migratetype
);
3212 prep_new_page(page
, order
, gfp_mask
, alloc_flags
);
3215 * If this is a high-order atomic allocation then check
3216 * if the pageblock should be reserved for the future
3218 if (unlikely(order
&& (alloc_flags
& ALLOC_HARDER
)))
3219 reserve_highatomic_pageblock(page
, zone
, order
);
3229 * Large machines with many possible nodes should not always dump per-node
3230 * meminfo in irq context.
3232 static inline bool should_suppress_show_mem(void)
3237 ret
= in_interrupt();
3242 static void warn_alloc_show_mem(gfp_t gfp_mask
, nodemask_t
*nodemask
)
3244 unsigned int filter
= SHOW_MEM_FILTER_NODES
;
3245 static DEFINE_RATELIMIT_STATE(show_mem_rs
, HZ
, 1);
3247 if (should_suppress_show_mem() || !__ratelimit(&show_mem_rs
))
3251 * This documents exceptions given to allocations in certain
3252 * contexts that are allowed to allocate outside current's set
3255 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
3256 if (tsk_is_oom_victim(current
) ||
3257 (current
->flags
& (PF_MEMALLOC
| PF_EXITING
)))
3258 filter
&= ~SHOW_MEM_FILTER_NODES
;
3259 if (in_interrupt() || !(gfp_mask
& __GFP_DIRECT_RECLAIM
))
3260 filter
&= ~SHOW_MEM_FILTER_NODES
;
3262 show_mem(filter
, nodemask
);
3265 void warn_alloc(gfp_t gfp_mask
, nodemask_t
*nodemask
, const char *fmt
, ...)
3267 struct va_format vaf
;
3269 static DEFINE_RATELIMIT_STATE(nopage_rs
, DEFAULT_RATELIMIT_INTERVAL
,
3270 DEFAULT_RATELIMIT_BURST
);
3272 if ((gfp_mask
& __GFP_NOWARN
) || !__ratelimit(&nopage_rs
))
3275 pr_warn("%s: ", current
->comm
);
3277 va_start(args
, fmt
);
3280 pr_cont("%pV", &vaf
);
3283 pr_cont(", mode:%#x(%pGg), nodemask=", gfp_mask
, &gfp_mask
);
3285 pr_cont("%*pbl\n", nodemask_pr_args(nodemask
));
3287 pr_cont("(null)\n");
3289 cpuset_print_current_mems_allowed();
3292 warn_alloc_show_mem(gfp_mask
, nodemask
);
3295 static inline struct page
*
3296 __alloc_pages_cpuset_fallback(gfp_t gfp_mask
, unsigned int order
,
3297 unsigned int alloc_flags
,
3298 const struct alloc_context
*ac
)
3302 page
= get_page_from_freelist(gfp_mask
, order
,
3303 alloc_flags
|ALLOC_CPUSET
, ac
);
3305 * fallback to ignore cpuset restriction if our nodes
3309 page
= get_page_from_freelist(gfp_mask
, order
,
3315 static inline struct page
*
3316 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
3317 const struct alloc_context
*ac
, unsigned long *did_some_progress
)
3319 struct oom_control oc
= {
3320 .zonelist
= ac
->zonelist
,
3321 .nodemask
= ac
->nodemask
,
3323 .gfp_mask
= gfp_mask
,
3328 *did_some_progress
= 0;
3331 * Acquire the oom lock. If that fails, somebody else is
3332 * making progress for us.
3334 if (!mutex_trylock(&oom_lock
)) {
3335 *did_some_progress
= 1;
3336 schedule_timeout_uninterruptible(1);
3341 * Go through the zonelist yet one more time, keep very high watermark
3342 * here, this is only to catch a parallel oom killing, we must fail if
3343 * we're still under heavy pressure. But make sure that this reclaim
3344 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3345 * allocation which will never fail due to oom_lock already held.
3347 page
= get_page_from_freelist((gfp_mask
| __GFP_HARDWALL
) &
3348 ~__GFP_DIRECT_RECLAIM
, order
,
3349 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
, ac
);
3353 /* Coredumps can quickly deplete all memory reserves */
3354 if (current
->flags
& PF_DUMPCORE
)
3356 /* The OOM killer will not help higher order allocs */
3357 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
3360 * We have already exhausted all our reclaim opportunities without any
3361 * success so it is time to admit defeat. We will skip the OOM killer
3362 * because it is very likely that the caller has a more reasonable
3363 * fallback than shooting a random task.
3365 if (gfp_mask
& __GFP_RETRY_MAYFAIL
)
3367 /* The OOM killer does not needlessly kill tasks for lowmem */
3368 if (ac
->high_zoneidx
< ZONE_NORMAL
)
3370 if (pm_suspended_storage())
3373 * XXX: GFP_NOFS allocations should rather fail than rely on
3374 * other request to make a forward progress.
3375 * We are in an unfortunate situation where out_of_memory cannot
3376 * do much for this context but let's try it to at least get
3377 * access to memory reserved if the current task is killed (see
3378 * out_of_memory). Once filesystems are ready to handle allocation
3379 * failures more gracefully we should just bail out here.
3382 /* The OOM killer may not free memory on a specific node */
3383 if (gfp_mask
& __GFP_THISNODE
)
3386 /* Exhausted what can be done so it's blamo time */
3387 if (out_of_memory(&oc
) || WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
)) {
3388 *did_some_progress
= 1;
3391 * Help non-failing allocations by giving them access to memory
3394 if (gfp_mask
& __GFP_NOFAIL
)
3395 page
= __alloc_pages_cpuset_fallback(gfp_mask
, order
,
3396 ALLOC_NO_WATERMARKS
, ac
);
3399 mutex_unlock(&oom_lock
);
3404 * Maximum number of compaction retries wit a progress before OOM
3405 * killer is consider as the only way to move forward.
3407 #define MAX_COMPACT_RETRIES 16
3409 #ifdef CONFIG_COMPACTION
3410 /* Try memory compaction for high-order allocations before reclaim */
3411 static struct page
*
3412 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
3413 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3414 enum compact_priority prio
, enum compact_result
*compact_result
)
3417 unsigned int noreclaim_flag
;
3422 noreclaim_flag
= memalloc_noreclaim_save();
3423 *compact_result
= try_to_compact_pages(gfp_mask
, order
, alloc_flags
, ac
,
3425 memalloc_noreclaim_restore(noreclaim_flag
);
3427 if (*compact_result
<= COMPACT_INACTIVE
)
3431 * At least in one zone compaction wasn't deferred or skipped, so let's
3432 * count a compaction stall
3434 count_vm_event(COMPACTSTALL
);
3436 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
3439 struct zone
*zone
= page_zone(page
);
3441 zone
->compact_blockskip_flush
= false;
3442 compaction_defer_reset(zone
, order
, true);
3443 count_vm_event(COMPACTSUCCESS
);
3448 * It's bad if compaction run occurs and fails. The most likely reason
3449 * is that pages exist, but not enough to satisfy watermarks.
3451 count_vm_event(COMPACTFAIL
);
3459 should_compact_retry(struct alloc_context
*ac
, int order
, int alloc_flags
,
3460 enum compact_result compact_result
,
3461 enum compact_priority
*compact_priority
,
3462 int *compaction_retries
)
3464 int max_retries
= MAX_COMPACT_RETRIES
;
3467 int retries
= *compaction_retries
;
3468 enum compact_priority priority
= *compact_priority
;
3473 if (compaction_made_progress(compact_result
))
3474 (*compaction_retries
)++;
3477 * compaction considers all the zone as desperately out of memory
3478 * so it doesn't really make much sense to retry except when the
3479 * failure could be caused by insufficient priority
3481 if (compaction_failed(compact_result
))
3482 goto check_priority
;
3485 * make sure the compaction wasn't deferred or didn't bail out early
3486 * due to locks contention before we declare that we should give up.
3487 * But do not retry if the given zonelist is not suitable for
3490 if (compaction_withdrawn(compact_result
)) {
3491 ret
= compaction_zonelist_suitable(ac
, order
, alloc_flags
);
3496 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
3497 * costly ones because they are de facto nofail and invoke OOM
3498 * killer to move on while costly can fail and users are ready
3499 * to cope with that. 1/4 retries is rather arbitrary but we
3500 * would need much more detailed feedback from compaction to
3501 * make a better decision.
3503 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
3505 if (*compaction_retries
<= max_retries
) {
3511 * Make sure there are attempts at the highest priority if we exhausted
3512 * all retries or failed at the lower priorities.
3515 min_priority
= (order
> PAGE_ALLOC_COSTLY_ORDER
) ?
3516 MIN_COMPACT_COSTLY_PRIORITY
: MIN_COMPACT_PRIORITY
;
3518 if (*compact_priority
> min_priority
) {
3519 (*compact_priority
)--;
3520 *compaction_retries
= 0;
3524 trace_compact_retry(order
, priority
, compact_result
, retries
, max_retries
, ret
);
3528 static inline struct page
*
3529 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
3530 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3531 enum compact_priority prio
, enum compact_result
*compact_result
)
3533 *compact_result
= COMPACT_SKIPPED
;
3538 should_compact_retry(struct alloc_context
*ac
, unsigned int order
, int alloc_flags
,
3539 enum compact_result compact_result
,
3540 enum compact_priority
*compact_priority
,
3541 int *compaction_retries
)
3546 if (!order
|| order
> PAGE_ALLOC_COSTLY_ORDER
)
3550 * There are setups with compaction disabled which would prefer to loop
3551 * inside the allocator rather than hit the oom killer prematurely.
3552 * Let's give them a good hope and keep retrying while the order-0
3553 * watermarks are OK.
3555 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
3557 if (zone_watermark_ok(zone
, 0, min_wmark_pages(zone
),
3558 ac_classzone_idx(ac
), alloc_flags
))
3563 #endif /* CONFIG_COMPACTION */
3565 #ifdef CONFIG_LOCKDEP
3566 struct lockdep_map __fs_reclaim_map
=
3567 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map
);
3569 static bool __need_fs_reclaim(gfp_t gfp_mask
)
3571 gfp_mask
= current_gfp_context(gfp_mask
);
3573 /* no reclaim without waiting on it */
3574 if (!(gfp_mask
& __GFP_DIRECT_RECLAIM
))
3577 /* this guy won't enter reclaim */
3578 if (current
->flags
& PF_MEMALLOC
)
3581 /* We're only interested __GFP_FS allocations for now */
3582 if (!(gfp_mask
& __GFP_FS
))
3585 if (gfp_mask
& __GFP_NOLOCKDEP
)
3591 void fs_reclaim_acquire(gfp_t gfp_mask
)
3593 if (__need_fs_reclaim(gfp_mask
))
3594 lock_map_acquire(&__fs_reclaim_map
);
3596 EXPORT_SYMBOL_GPL(fs_reclaim_acquire
);
3598 void fs_reclaim_release(gfp_t gfp_mask
)
3600 if (__need_fs_reclaim(gfp_mask
))
3601 lock_map_release(&__fs_reclaim_map
);
3603 EXPORT_SYMBOL_GPL(fs_reclaim_release
);
3606 /* Perform direct synchronous page reclaim */
3608 __perform_reclaim(gfp_t gfp_mask
, unsigned int order
,
3609 const struct alloc_context
*ac
)
3611 struct reclaim_state reclaim_state
;
3613 unsigned int noreclaim_flag
;
3617 /* We now go into synchronous reclaim */
3618 cpuset_memory_pressure_bump();
3619 noreclaim_flag
= memalloc_noreclaim_save();
3620 fs_reclaim_acquire(gfp_mask
);
3621 reclaim_state
.reclaimed_slab
= 0;
3622 current
->reclaim_state
= &reclaim_state
;
3624 progress
= try_to_free_pages(ac
->zonelist
, order
, gfp_mask
,
3627 current
->reclaim_state
= NULL
;
3628 fs_reclaim_release(gfp_mask
);
3629 memalloc_noreclaim_restore(noreclaim_flag
);
3636 /* The really slow allocator path where we enter direct reclaim */
3637 static inline struct page
*
3638 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
3639 unsigned int alloc_flags
, const struct alloc_context
*ac
,
3640 unsigned long *did_some_progress
)
3642 struct page
*page
= NULL
;
3643 bool drained
= false;
3645 *did_some_progress
= __perform_reclaim(gfp_mask
, order
, ac
);
3646 if (unlikely(!(*did_some_progress
)))
3650 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
3653 * If an allocation failed after direct reclaim, it could be because
3654 * pages are pinned on the per-cpu lists or in high alloc reserves.
3655 * Shrink them them and try again
3657 if (!page
&& !drained
) {
3658 unreserve_highatomic_pageblock(ac
, false);
3659 drain_all_pages(NULL
);
3667 static void wake_all_kswapds(unsigned int order
, const struct alloc_context
*ac
)
3671 pg_data_t
*last_pgdat
= NULL
;
3673 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
,
3674 ac
->high_zoneidx
, ac
->nodemask
) {
3675 if (last_pgdat
!= zone
->zone_pgdat
)
3676 wakeup_kswapd(zone
, order
, ac
->high_zoneidx
);
3677 last_pgdat
= zone
->zone_pgdat
;
3681 static inline unsigned int
3682 gfp_to_alloc_flags(gfp_t gfp_mask
)
3684 unsigned int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
3686 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3687 BUILD_BUG_ON(__GFP_HIGH
!= (__force gfp_t
) ALLOC_HIGH
);
3690 * The caller may dip into page reserves a bit more if the caller
3691 * cannot run direct reclaim, or if the caller has realtime scheduling
3692 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3693 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3695 alloc_flags
|= (__force
int) (gfp_mask
& __GFP_HIGH
);
3697 if (gfp_mask
& __GFP_ATOMIC
) {
3699 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3700 * if it can't schedule.
3702 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
3703 alloc_flags
|= ALLOC_HARDER
;
3705 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3706 * comment for __cpuset_node_allowed().
3708 alloc_flags
&= ~ALLOC_CPUSET
;
3709 } else if (unlikely(rt_task(current
)) && !in_interrupt())
3710 alloc_flags
|= ALLOC_HARDER
;
3713 if ((gfpflags_to_migratetype(gfp_mask
) == MIGRATE_MOVABLE
) ||
3714 ((gfp_mask
& GFP_HIGHUSER_MOVABLE
) == GFP_HIGHUSER_MOVABLE
))
3715 alloc_flags
|= ALLOC_CMA
;
3720 static bool oom_reserves_allowed(struct task_struct
*tsk
)
3722 if (!tsk_is_oom_victim(tsk
))
3726 * !MMU doesn't have oom reaper so give access to memory reserves
3727 * only to the thread with TIF_MEMDIE set
3729 if (!IS_ENABLED(CONFIG_MMU
) && !test_thread_flag(TIF_MEMDIE
))
3736 * Distinguish requests which really need access to full memory
3737 * reserves from oom victims which can live with a portion of it
3739 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask
)
3741 if (unlikely(gfp_mask
& __GFP_NOMEMALLOC
))
3743 if (gfp_mask
& __GFP_MEMALLOC
)
3744 return ALLOC_NO_WATERMARKS
;
3745 if (in_serving_softirq() && (current
->flags
& PF_MEMALLOC
))
3746 return ALLOC_NO_WATERMARKS
;
3747 if (!in_interrupt()) {
3748 if (current
->flags
& PF_MEMALLOC
)
3749 return ALLOC_NO_WATERMARKS
;
3750 else if (oom_reserves_allowed(current
))
3757 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask
)
3759 return !!__gfp_pfmemalloc_flags(gfp_mask
);
3763 * Checks whether it makes sense to retry the reclaim to make a forward progress
3764 * for the given allocation request.
3766 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
3767 * without success, or when we couldn't even meet the watermark if we
3768 * reclaimed all remaining pages on the LRU lists.
3770 * Returns true if a retry is viable or false to enter the oom path.
3773 should_reclaim_retry(gfp_t gfp_mask
, unsigned order
,
3774 struct alloc_context
*ac
, int alloc_flags
,
3775 bool did_some_progress
, int *no_progress_loops
)
3781 * Costly allocations might have made a progress but this doesn't mean
3782 * their order will become available due to high fragmentation so
3783 * always increment the no progress counter for them
3785 if (did_some_progress
&& order
<= PAGE_ALLOC_COSTLY_ORDER
)
3786 *no_progress_loops
= 0;
3788 (*no_progress_loops
)++;
3791 * Make sure we converge to OOM if we cannot make any progress
3792 * several times in the row.
3794 if (*no_progress_loops
> MAX_RECLAIM_RETRIES
) {
3795 /* Before OOM, exhaust highatomic_reserve */
3796 return unreserve_highatomic_pageblock(ac
, true);
3800 * Keep reclaiming pages while there is a chance this will lead
3801 * somewhere. If none of the target zones can satisfy our allocation
3802 * request even if all reclaimable pages are considered then we are
3803 * screwed and have to go OOM.
3805 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
3807 unsigned long available
;
3808 unsigned long reclaimable
;
3809 unsigned long min_wmark
= min_wmark_pages(zone
);
3812 available
= reclaimable
= zone_reclaimable_pages(zone
);
3813 available
+= zone_page_state_snapshot(zone
, NR_FREE_PAGES
);
3816 * Would the allocation succeed if we reclaimed all
3817 * reclaimable pages?
3819 wmark
= __zone_watermark_ok(zone
, order
, min_wmark
,
3820 ac_classzone_idx(ac
), alloc_flags
, available
);
3821 trace_reclaim_retry_zone(z
, order
, reclaimable
,
3822 available
, min_wmark
, *no_progress_loops
, wmark
);
3825 * If we didn't make any progress and have a lot of
3826 * dirty + writeback pages then we should wait for
3827 * an IO to complete to slow down the reclaim and
3828 * prevent from pre mature OOM
3830 if (!did_some_progress
) {
3831 unsigned long write_pending
;
3833 write_pending
= zone_page_state_snapshot(zone
,
3834 NR_ZONE_WRITE_PENDING
);
3836 if (2 * write_pending
> reclaimable
) {
3837 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
3843 * Memory allocation/reclaim might be called from a WQ
3844 * context and the current implementation of the WQ
3845 * concurrency control doesn't recognize that
3846 * a particular WQ is congested if the worker thread is
3847 * looping without ever sleeping. Therefore we have to
3848 * do a short sleep here rather than calling
3851 if (current
->flags
& PF_WQ_WORKER
)
3852 schedule_timeout_uninterruptible(1);
3864 check_retry_cpuset(int cpuset_mems_cookie
, struct alloc_context
*ac
)
3867 * It's possible that cpuset's mems_allowed and the nodemask from
3868 * mempolicy don't intersect. This should be normally dealt with by
3869 * policy_nodemask(), but it's possible to race with cpuset update in
3870 * such a way the check therein was true, and then it became false
3871 * before we got our cpuset_mems_cookie here.
3872 * This assumes that for all allocations, ac->nodemask can come only
3873 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
3874 * when it does not intersect with the cpuset restrictions) or the
3875 * caller can deal with a violated nodemask.
3877 if (cpusets_enabled() && ac
->nodemask
&&
3878 !cpuset_nodemask_valid_mems_allowed(ac
->nodemask
)) {
3879 ac
->nodemask
= NULL
;
3884 * When updating a task's mems_allowed or mempolicy nodemask, it is
3885 * possible to race with parallel threads in such a way that our
3886 * allocation can fail while the mask is being updated. If we are about
3887 * to fail, check if the cpuset changed during allocation and if so,
3890 if (read_mems_allowed_retry(cpuset_mems_cookie
))
3896 static inline struct page
*
3897 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
3898 struct alloc_context
*ac
)
3900 bool can_direct_reclaim
= gfp_mask
& __GFP_DIRECT_RECLAIM
;
3901 const bool costly_order
= order
> PAGE_ALLOC_COSTLY_ORDER
;
3902 struct page
*page
= NULL
;
3903 unsigned int alloc_flags
;
3904 unsigned long did_some_progress
;
3905 enum compact_priority compact_priority
;
3906 enum compact_result compact_result
;
3907 int compaction_retries
;
3908 int no_progress_loops
;
3909 unsigned long alloc_start
= jiffies
;
3910 unsigned int stall_timeout
= 10 * HZ
;
3911 unsigned int cpuset_mems_cookie
;
3915 * In the slowpath, we sanity check order to avoid ever trying to
3916 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3917 * be using allocators in order of preference for an area that is
3920 if (order
>= MAX_ORDER
) {
3921 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
3926 * We also sanity check to catch abuse of atomic reserves being used by
3927 * callers that are not in atomic context.
3929 if (WARN_ON_ONCE((gfp_mask
& (__GFP_ATOMIC
|__GFP_DIRECT_RECLAIM
)) ==
3930 (__GFP_ATOMIC
|__GFP_DIRECT_RECLAIM
)))
3931 gfp_mask
&= ~__GFP_ATOMIC
;
3934 compaction_retries
= 0;
3935 no_progress_loops
= 0;
3936 compact_priority
= DEF_COMPACT_PRIORITY
;
3937 cpuset_mems_cookie
= read_mems_allowed_begin();
3940 * The fast path uses conservative alloc_flags to succeed only until
3941 * kswapd needs to be woken up, and to avoid the cost of setting up
3942 * alloc_flags precisely. So we do that now.
3944 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
3947 * We need to recalculate the starting point for the zonelist iterator
3948 * because we might have used different nodemask in the fast path, or
3949 * there was a cpuset modification and we are retrying - otherwise we
3950 * could end up iterating over non-eligible zones endlessly.
3952 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
3953 ac
->high_zoneidx
, ac
->nodemask
);
3954 if (!ac
->preferred_zoneref
->zone
)
3957 if (gfp_mask
& __GFP_KSWAPD_RECLAIM
)
3958 wake_all_kswapds(order
, ac
);
3961 * The adjusted alloc_flags might result in immediate success, so try
3964 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
3969 * For costly allocations, try direct compaction first, as it's likely
3970 * that we have enough base pages and don't need to reclaim. For non-
3971 * movable high-order allocations, do that as well, as compaction will
3972 * try prevent permanent fragmentation by migrating from blocks of the
3974 * Don't try this for allocations that are allowed to ignore
3975 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
3977 if (can_direct_reclaim
&&
3979 (order
> 0 && ac
->migratetype
!= MIGRATE_MOVABLE
))
3980 && !gfp_pfmemalloc_allowed(gfp_mask
)) {
3981 page
= __alloc_pages_direct_compact(gfp_mask
, order
,
3983 INIT_COMPACT_PRIORITY
,
3989 * Checks for costly allocations with __GFP_NORETRY, which
3990 * includes THP page fault allocations
3992 if (costly_order
&& (gfp_mask
& __GFP_NORETRY
)) {
3994 * If compaction is deferred for high-order allocations,
3995 * it is because sync compaction recently failed. If
3996 * this is the case and the caller requested a THP
3997 * allocation, we do not want to heavily disrupt the
3998 * system, so we fail the allocation instead of entering
4001 if (compact_result
== COMPACT_DEFERRED
)
4005 * Looks like reclaim/compaction is worth trying, but
4006 * sync compaction could be very expensive, so keep
4007 * using async compaction.
4009 compact_priority
= INIT_COMPACT_PRIORITY
;
4014 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4015 if (gfp_mask
& __GFP_KSWAPD_RECLAIM
)
4016 wake_all_kswapds(order
, ac
);
4018 reserve_flags
= __gfp_pfmemalloc_flags(gfp_mask
);
4020 alloc_flags
= reserve_flags
;
4023 * Reset the zonelist iterators if memory policies can be ignored.
4024 * These allocations are high priority and system rather than user
4027 if (!(alloc_flags
& ALLOC_CPUSET
) || reserve_flags
) {
4028 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
4029 ac
->high_zoneidx
, ac
->nodemask
);
4032 /* Attempt with potentially adjusted zonelist and alloc_flags */
4033 page
= get_page_from_freelist(gfp_mask
, order
, alloc_flags
, ac
);
4037 /* Caller is not willing to reclaim, we can't balance anything */
4038 if (!can_direct_reclaim
)
4041 /* Make sure we know about allocations which stall for too long */
4042 if (time_after(jiffies
, alloc_start
+ stall_timeout
)) {
4043 warn_alloc(gfp_mask
& ~__GFP_NOWARN
, ac
->nodemask
,
4044 "page allocation stalls for %ums, order:%u",
4045 jiffies_to_msecs(jiffies
-alloc_start
), order
);
4046 stall_timeout
+= 10 * HZ
;
4049 /* Avoid recursion of direct reclaim */
4050 if (current
->flags
& PF_MEMALLOC
)
4053 /* Try direct reclaim and then allocating */
4054 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
, alloc_flags
, ac
,
4055 &did_some_progress
);
4059 /* Try direct compaction and then allocating */
4060 page
= __alloc_pages_direct_compact(gfp_mask
, order
, alloc_flags
, ac
,
4061 compact_priority
, &compact_result
);
4065 /* Do not loop if specifically requested */
4066 if (gfp_mask
& __GFP_NORETRY
)
4070 * Do not retry costly high order allocations unless they are
4071 * __GFP_RETRY_MAYFAIL
4073 if (costly_order
&& !(gfp_mask
& __GFP_RETRY_MAYFAIL
))
4076 if (should_reclaim_retry(gfp_mask
, order
, ac
, alloc_flags
,
4077 did_some_progress
> 0, &no_progress_loops
))
4081 * It doesn't make any sense to retry for the compaction if the order-0
4082 * reclaim is not able to make any progress because the current
4083 * implementation of the compaction depends on the sufficient amount
4084 * of free memory (see __compaction_suitable)
4086 if (did_some_progress
> 0 &&
4087 should_compact_retry(ac
, order
, alloc_flags
,
4088 compact_result
, &compact_priority
,
4089 &compaction_retries
))
4093 /* Deal with possible cpuset update races before we start OOM killing */
4094 if (check_retry_cpuset(cpuset_mems_cookie
, ac
))
4097 /* Reclaim has failed us, start killing things */
4098 page
= __alloc_pages_may_oom(gfp_mask
, order
, ac
, &did_some_progress
);
4102 /* Avoid allocations with no watermarks from looping endlessly */
4103 if (tsk_is_oom_victim(current
) &&
4104 (alloc_flags
== ALLOC_OOM
||
4105 (gfp_mask
& __GFP_NOMEMALLOC
)))
4108 /* Retry as long as the OOM killer is making progress */
4109 if (did_some_progress
) {
4110 no_progress_loops
= 0;
4115 /* Deal with possible cpuset update races before we fail */
4116 if (check_retry_cpuset(cpuset_mems_cookie
, ac
))
4120 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4123 if (gfp_mask
& __GFP_NOFAIL
) {
4125 * All existing users of the __GFP_NOFAIL are blockable, so warn
4126 * of any new users that actually require GFP_NOWAIT
4128 if (WARN_ON_ONCE(!can_direct_reclaim
))
4132 * PF_MEMALLOC request from this context is rather bizarre
4133 * because we cannot reclaim anything and only can loop waiting
4134 * for somebody to do a work for us
4136 WARN_ON_ONCE(current
->flags
& PF_MEMALLOC
);
4139 * non failing costly orders are a hard requirement which we
4140 * are not prepared for much so let's warn about these users
4141 * so that we can identify them and convert them to something
4144 WARN_ON_ONCE(order
> PAGE_ALLOC_COSTLY_ORDER
);
4147 * Help non-failing allocations by giving them access to memory
4148 * reserves but do not use ALLOC_NO_WATERMARKS because this
4149 * could deplete whole memory reserves which would just make
4150 * the situation worse
4152 page
= __alloc_pages_cpuset_fallback(gfp_mask
, order
, ALLOC_HARDER
, ac
);
4160 warn_alloc(gfp_mask
, ac
->nodemask
,
4161 "page allocation failure: order:%u", order
);
4166 static inline bool prepare_alloc_pages(gfp_t gfp_mask
, unsigned int order
,
4167 int preferred_nid
, nodemask_t
*nodemask
,
4168 struct alloc_context
*ac
, gfp_t
*alloc_mask
,
4169 unsigned int *alloc_flags
)
4171 ac
->high_zoneidx
= gfp_zone(gfp_mask
);
4172 ac
->zonelist
= node_zonelist(preferred_nid
, gfp_mask
);
4173 ac
->nodemask
= nodemask
;
4174 ac
->migratetype
= gfpflags_to_migratetype(gfp_mask
);
4176 if (cpusets_enabled()) {
4177 *alloc_mask
|= __GFP_HARDWALL
;
4179 ac
->nodemask
= &cpuset_current_mems_allowed
;
4181 *alloc_flags
|= ALLOC_CPUSET
;
4184 fs_reclaim_acquire(gfp_mask
);
4185 fs_reclaim_release(gfp_mask
);
4187 might_sleep_if(gfp_mask
& __GFP_DIRECT_RECLAIM
);
4189 if (should_fail_alloc_page(gfp_mask
, order
))
4192 if (IS_ENABLED(CONFIG_CMA
) && ac
->migratetype
== MIGRATE_MOVABLE
)
4193 *alloc_flags
|= ALLOC_CMA
;
4198 /* Determine whether to spread dirty pages and what the first usable zone */
4199 static inline void finalise_ac(gfp_t gfp_mask
,
4200 unsigned int order
, struct alloc_context
*ac
)
4202 /* Dirty zone balancing only done in the fast path */
4203 ac
->spread_dirty_pages
= (gfp_mask
& __GFP_WRITE
);
4206 * The preferred zone is used for statistics but crucially it is
4207 * also used as the starting point for the zonelist iterator. It
4208 * may get reset for allocations that ignore memory policies.
4210 ac
->preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
4211 ac
->high_zoneidx
, ac
->nodemask
);
4215 * This is the 'heart' of the zoned buddy allocator.
4218 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
, int preferred_nid
,
4219 nodemask_t
*nodemask
)
4222 unsigned int alloc_flags
= ALLOC_WMARK_LOW
;
4223 gfp_t alloc_mask
; /* The gfp_t that was actually used for allocation */
4224 struct alloc_context ac
= { };
4226 gfp_mask
&= gfp_allowed_mask
;
4227 alloc_mask
= gfp_mask
;
4228 if (!prepare_alloc_pages(gfp_mask
, order
, preferred_nid
, nodemask
, &ac
, &alloc_mask
, &alloc_flags
))
4231 finalise_ac(gfp_mask
, order
, &ac
);
4233 /* First allocation attempt */
4234 page
= get_page_from_freelist(alloc_mask
, order
, alloc_flags
, &ac
);
4239 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4240 * resp. GFP_NOIO which has to be inherited for all allocation requests
4241 * from a particular context which has been marked by
4242 * memalloc_no{fs,io}_{save,restore}.
4244 alloc_mask
= current_gfp_context(gfp_mask
);
4245 ac
.spread_dirty_pages
= false;
4248 * Restore the original nodemask if it was potentially replaced with
4249 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4251 if (unlikely(ac
.nodemask
!= nodemask
))
4252 ac
.nodemask
= nodemask
;
4254 page
= __alloc_pages_slowpath(alloc_mask
, order
, &ac
);
4257 if (memcg_kmem_enabled() && (gfp_mask
& __GFP_ACCOUNT
) && page
&&
4258 unlikely(memcg_kmem_charge(page
, gfp_mask
, order
) != 0)) {
4259 __free_pages(page
, order
);
4263 trace_mm_page_alloc(page
, order
, alloc_mask
, ac
.migratetype
);
4267 EXPORT_SYMBOL(__alloc_pages_nodemask
);
4270 * Common helper functions.
4272 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
4277 * __get_free_pages() returns a 32-bit address, which cannot represent
4280 VM_BUG_ON((gfp_mask
& __GFP_HIGHMEM
) != 0);
4282 page
= alloc_pages(gfp_mask
, order
);
4285 return (unsigned long) page_address(page
);
4287 EXPORT_SYMBOL(__get_free_pages
);
4289 unsigned long get_zeroed_page(gfp_t gfp_mask
)
4291 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
4293 EXPORT_SYMBOL(get_zeroed_page
);
4295 void __free_pages(struct page
*page
, unsigned int order
)
4297 if (put_page_testzero(page
)) {
4299 free_hot_cold_page(page
, false);
4301 __free_pages_ok(page
, order
);
4305 EXPORT_SYMBOL(__free_pages
);
4307 void free_pages(unsigned long addr
, unsigned int order
)
4310 VM_BUG_ON(!virt_addr_valid((void *)addr
));
4311 __free_pages(virt_to_page((void *)addr
), order
);
4315 EXPORT_SYMBOL(free_pages
);
4319 * An arbitrary-length arbitrary-offset area of memory which resides
4320 * within a 0 or higher order page. Multiple fragments within that page
4321 * are individually refcounted, in the page's reference counter.
4323 * The page_frag functions below provide a simple allocation framework for
4324 * page fragments. This is used by the network stack and network device
4325 * drivers to provide a backing region of memory for use as either an
4326 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4328 static struct page
*__page_frag_cache_refill(struct page_frag_cache
*nc
,
4331 struct page
*page
= NULL
;
4332 gfp_t gfp
= gfp_mask
;
4334 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4335 gfp_mask
|= __GFP_COMP
| __GFP_NOWARN
| __GFP_NORETRY
|
4337 page
= alloc_pages_node(NUMA_NO_NODE
, gfp_mask
,
4338 PAGE_FRAG_CACHE_MAX_ORDER
);
4339 nc
->size
= page
? PAGE_FRAG_CACHE_MAX_SIZE
: PAGE_SIZE
;
4341 if (unlikely(!page
))
4342 page
= alloc_pages_node(NUMA_NO_NODE
, gfp
, 0);
4344 nc
->va
= page
? page_address(page
) : NULL
;
4349 void __page_frag_cache_drain(struct page
*page
, unsigned int count
)
4351 VM_BUG_ON_PAGE(page_ref_count(page
) == 0, page
);
4353 if (page_ref_sub_and_test(page
, count
)) {
4354 unsigned int order
= compound_order(page
);
4357 free_hot_cold_page(page
, false);
4359 __free_pages_ok(page
, order
);
4362 EXPORT_SYMBOL(__page_frag_cache_drain
);
4364 void *page_frag_alloc(struct page_frag_cache
*nc
,
4365 unsigned int fragsz
, gfp_t gfp_mask
)
4367 unsigned int size
= PAGE_SIZE
;
4371 if (unlikely(!nc
->va
)) {
4373 page
= __page_frag_cache_refill(nc
, gfp_mask
);
4377 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4378 /* if size can vary use size else just use PAGE_SIZE */
4381 /* Even if we own the page, we do not use atomic_set().
4382 * This would break get_page_unless_zero() users.
4384 page_ref_add(page
, size
- 1);
4386 /* reset page count bias and offset to start of new frag */
4387 nc
->pfmemalloc
= page_is_pfmemalloc(page
);
4388 nc
->pagecnt_bias
= size
;
4392 offset
= nc
->offset
- fragsz
;
4393 if (unlikely(offset
< 0)) {
4394 page
= virt_to_page(nc
->va
);
4396 if (!page_ref_sub_and_test(page
, nc
->pagecnt_bias
))
4399 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4400 /* if size can vary use size else just use PAGE_SIZE */
4403 /* OK, page count is 0, we can safely set it */
4404 set_page_count(page
, size
);
4406 /* reset page count bias and offset to start of new frag */
4407 nc
->pagecnt_bias
= size
;
4408 offset
= size
- fragsz
;
4412 nc
->offset
= offset
;
4414 return nc
->va
+ offset
;
4416 EXPORT_SYMBOL(page_frag_alloc
);
4419 * Frees a page fragment allocated out of either a compound or order 0 page.
4421 void page_frag_free(void *addr
)
4423 struct page
*page
= virt_to_head_page(addr
);
4425 if (unlikely(put_page_testzero(page
)))
4426 __free_pages_ok(page
, compound_order(page
));
4428 EXPORT_SYMBOL(page_frag_free
);
4430 static void *make_alloc_exact(unsigned long addr
, unsigned int order
,
4434 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
4435 unsigned long used
= addr
+ PAGE_ALIGN(size
);
4437 split_page(virt_to_page((void *)addr
), order
);
4438 while (used
< alloc_end
) {
4443 return (void *)addr
;
4447 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4448 * @size: the number of bytes to allocate
4449 * @gfp_mask: GFP flags for the allocation
4451 * This function is similar to alloc_pages(), except that it allocates the
4452 * minimum number of pages to satisfy the request. alloc_pages() can only
4453 * allocate memory in power-of-two pages.
4455 * This function is also limited by MAX_ORDER.
4457 * Memory allocated by this function must be released by free_pages_exact().
4459 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
4461 unsigned int order
= get_order(size
);
4464 addr
= __get_free_pages(gfp_mask
, order
);
4465 return make_alloc_exact(addr
, order
, size
);
4467 EXPORT_SYMBOL(alloc_pages_exact
);
4470 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4472 * @nid: the preferred node ID where memory should be allocated
4473 * @size: the number of bytes to allocate
4474 * @gfp_mask: GFP flags for the allocation
4476 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4479 void * __meminit
alloc_pages_exact_nid(int nid
, size_t size
, gfp_t gfp_mask
)
4481 unsigned int order
= get_order(size
);
4482 struct page
*p
= alloc_pages_node(nid
, gfp_mask
, order
);
4485 return make_alloc_exact((unsigned long)page_address(p
), order
, size
);
4489 * free_pages_exact - release memory allocated via alloc_pages_exact()
4490 * @virt: the value returned by alloc_pages_exact.
4491 * @size: size of allocation, same value as passed to alloc_pages_exact().
4493 * Release the memory allocated by a previous call to alloc_pages_exact.
4495 void free_pages_exact(void *virt
, size_t size
)
4497 unsigned long addr
= (unsigned long)virt
;
4498 unsigned long end
= addr
+ PAGE_ALIGN(size
);
4500 while (addr
< end
) {
4505 EXPORT_SYMBOL(free_pages_exact
);
4508 * nr_free_zone_pages - count number of pages beyond high watermark
4509 * @offset: The zone index of the highest zone
4511 * nr_free_zone_pages() counts the number of counts pages which are beyond the
4512 * high watermark within all zones at or below a given zone index. For each
4513 * zone, the number of pages is calculated as:
4515 * nr_free_zone_pages = managed_pages - high_pages
4517 static unsigned long nr_free_zone_pages(int offset
)
4522 /* Just pick one node, since fallback list is circular */
4523 unsigned long sum
= 0;
4525 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
4527 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
4528 unsigned long size
= zone
->managed_pages
;
4529 unsigned long high
= high_wmark_pages(zone
);
4538 * nr_free_buffer_pages - count number of pages beyond high watermark
4540 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4541 * watermark within ZONE_DMA and ZONE_NORMAL.
4543 unsigned long nr_free_buffer_pages(void)
4545 return nr_free_zone_pages(gfp_zone(GFP_USER
));
4547 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
4550 * nr_free_pagecache_pages - count number of pages beyond high watermark
4552 * nr_free_pagecache_pages() counts the number of pages which are beyond the
4553 * high watermark within all zones.
4555 unsigned long nr_free_pagecache_pages(void)
4557 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
4560 static inline void show_node(struct zone
*zone
)
4562 if (IS_ENABLED(CONFIG_NUMA
))
4563 printk("Node %d ", zone_to_nid(zone
));
4566 long si_mem_available(void)
4569 unsigned long pagecache
;
4570 unsigned long wmark_low
= 0;
4571 unsigned long pages
[NR_LRU_LISTS
];
4575 for (lru
= LRU_BASE
; lru
< NR_LRU_LISTS
; lru
++)
4576 pages
[lru
] = global_node_page_state(NR_LRU_BASE
+ lru
);
4579 wmark_low
+= zone
->watermark
[WMARK_LOW
];
4582 * Estimate the amount of memory available for userspace allocations,
4583 * without causing swapping.
4585 available
= global_zone_page_state(NR_FREE_PAGES
) - totalreserve_pages
;
4588 * Not all the page cache can be freed, otherwise the system will
4589 * start swapping. Assume at least half of the page cache, or the
4590 * low watermark worth of cache, needs to stay.
4592 pagecache
= pages
[LRU_ACTIVE_FILE
] + pages
[LRU_INACTIVE_FILE
];
4593 pagecache
-= min(pagecache
/ 2, wmark_low
);
4594 available
+= pagecache
;
4597 * Part of the reclaimable slab consists of items that are in use,
4598 * and cannot be freed. Cap this estimate at the low watermark.
4600 available
+= global_node_page_state(NR_SLAB_RECLAIMABLE
) -
4601 min(global_node_page_state(NR_SLAB_RECLAIMABLE
) / 2,
4608 EXPORT_SYMBOL_GPL(si_mem_available
);
4610 void si_meminfo(struct sysinfo
*val
)
4612 val
->totalram
= totalram_pages
;
4613 val
->sharedram
= global_node_page_state(NR_SHMEM
);
4614 val
->freeram
= global_zone_page_state(NR_FREE_PAGES
);
4615 val
->bufferram
= nr_blockdev_pages();
4616 val
->totalhigh
= totalhigh_pages
;
4617 val
->freehigh
= nr_free_highpages();
4618 val
->mem_unit
= PAGE_SIZE
;
4621 EXPORT_SYMBOL(si_meminfo
);
4624 void si_meminfo_node(struct sysinfo
*val
, int nid
)
4626 int zone_type
; /* needs to be signed */
4627 unsigned long managed_pages
= 0;
4628 unsigned long managed_highpages
= 0;
4629 unsigned long free_highpages
= 0;
4630 pg_data_t
*pgdat
= NODE_DATA(nid
);
4632 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++)
4633 managed_pages
+= pgdat
->node_zones
[zone_type
].managed_pages
;
4634 val
->totalram
= managed_pages
;
4635 val
->sharedram
= node_page_state(pgdat
, NR_SHMEM
);
4636 val
->freeram
= sum_zone_node_page_state(nid
, NR_FREE_PAGES
);
4637 #ifdef CONFIG_HIGHMEM
4638 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
4639 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
4641 if (is_highmem(zone
)) {
4642 managed_highpages
+= zone
->managed_pages
;
4643 free_highpages
+= zone_page_state(zone
, NR_FREE_PAGES
);
4646 val
->totalhigh
= managed_highpages
;
4647 val
->freehigh
= free_highpages
;
4649 val
->totalhigh
= managed_highpages
;
4650 val
->freehigh
= free_highpages
;
4652 val
->mem_unit
= PAGE_SIZE
;
4657 * Determine whether the node should be displayed or not, depending on whether
4658 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
4660 static bool show_mem_node_skip(unsigned int flags
, int nid
, nodemask_t
*nodemask
)
4662 if (!(flags
& SHOW_MEM_FILTER_NODES
))
4666 * no node mask - aka implicit memory numa policy. Do not bother with
4667 * the synchronization - read_mems_allowed_begin - because we do not
4668 * have to be precise here.
4671 nodemask
= &cpuset_current_mems_allowed
;
4673 return !node_isset(nid
, *nodemask
);
4676 #define K(x) ((x) << (PAGE_SHIFT-10))
4678 static void show_migration_types(unsigned char type
)
4680 static const char types
[MIGRATE_TYPES
] = {
4681 [MIGRATE_UNMOVABLE
] = 'U',
4682 [MIGRATE_MOVABLE
] = 'M',
4683 [MIGRATE_RECLAIMABLE
] = 'E',
4684 [MIGRATE_HIGHATOMIC
] = 'H',
4686 [MIGRATE_CMA
] = 'C',
4688 #ifdef CONFIG_MEMORY_ISOLATION
4689 [MIGRATE_ISOLATE
] = 'I',
4692 char tmp
[MIGRATE_TYPES
+ 1];
4696 for (i
= 0; i
< MIGRATE_TYPES
; i
++) {
4697 if (type
& (1 << i
))
4702 printk(KERN_CONT
"(%s) ", tmp
);
4706 * Show free area list (used inside shift_scroll-lock stuff)
4707 * We also calculate the percentage fragmentation. We do this by counting the
4708 * memory on each free list with the exception of the first item on the list.
4711 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
4714 void show_free_areas(unsigned int filter
, nodemask_t
*nodemask
)
4716 unsigned long free_pcp
= 0;
4721 for_each_populated_zone(zone
) {
4722 if (show_mem_node_skip(filter
, zone_to_nid(zone
), nodemask
))
4725 for_each_online_cpu(cpu
)
4726 free_pcp
+= per_cpu_ptr(zone
->pageset
, cpu
)->pcp
.count
;
4729 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
4730 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
4731 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
4732 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
4733 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
4734 " free:%lu free_pcp:%lu free_cma:%lu\n",
4735 global_node_page_state(NR_ACTIVE_ANON
),
4736 global_node_page_state(NR_INACTIVE_ANON
),
4737 global_node_page_state(NR_ISOLATED_ANON
),
4738 global_node_page_state(NR_ACTIVE_FILE
),
4739 global_node_page_state(NR_INACTIVE_FILE
),
4740 global_node_page_state(NR_ISOLATED_FILE
),
4741 global_node_page_state(NR_UNEVICTABLE
),
4742 global_node_page_state(NR_FILE_DIRTY
),
4743 global_node_page_state(NR_WRITEBACK
),
4744 global_node_page_state(NR_UNSTABLE_NFS
),
4745 global_node_page_state(NR_SLAB_RECLAIMABLE
),
4746 global_node_page_state(NR_SLAB_UNRECLAIMABLE
),
4747 global_node_page_state(NR_FILE_MAPPED
),
4748 global_node_page_state(NR_SHMEM
),
4749 global_zone_page_state(NR_PAGETABLE
),
4750 global_zone_page_state(NR_BOUNCE
),
4751 global_zone_page_state(NR_FREE_PAGES
),
4753 global_zone_page_state(NR_FREE_CMA_PAGES
));
4755 for_each_online_pgdat(pgdat
) {
4756 if (show_mem_node_skip(filter
, pgdat
->node_id
, nodemask
))
4760 " active_anon:%lukB"
4761 " inactive_anon:%lukB"
4762 " active_file:%lukB"
4763 " inactive_file:%lukB"
4764 " unevictable:%lukB"
4765 " isolated(anon):%lukB"
4766 " isolated(file):%lukB"
4771 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4773 " shmem_pmdmapped: %lukB"
4776 " writeback_tmp:%lukB"
4778 " all_unreclaimable? %s"
4781 K(node_page_state(pgdat
, NR_ACTIVE_ANON
)),
4782 K(node_page_state(pgdat
, NR_INACTIVE_ANON
)),
4783 K(node_page_state(pgdat
, NR_ACTIVE_FILE
)),
4784 K(node_page_state(pgdat
, NR_INACTIVE_FILE
)),
4785 K(node_page_state(pgdat
, NR_UNEVICTABLE
)),
4786 K(node_page_state(pgdat
, NR_ISOLATED_ANON
)),
4787 K(node_page_state(pgdat
, NR_ISOLATED_FILE
)),
4788 K(node_page_state(pgdat
, NR_FILE_MAPPED
)),
4789 K(node_page_state(pgdat
, NR_FILE_DIRTY
)),
4790 K(node_page_state(pgdat
, NR_WRITEBACK
)),
4791 K(node_page_state(pgdat
, NR_SHMEM
)),
4792 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4793 K(node_page_state(pgdat
, NR_SHMEM_THPS
) * HPAGE_PMD_NR
),
4794 K(node_page_state(pgdat
, NR_SHMEM_PMDMAPPED
)
4796 K(node_page_state(pgdat
, NR_ANON_THPS
) * HPAGE_PMD_NR
),
4798 K(node_page_state(pgdat
, NR_WRITEBACK_TEMP
)),
4799 K(node_page_state(pgdat
, NR_UNSTABLE_NFS
)),
4800 pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
?
4804 for_each_populated_zone(zone
) {
4807 if (show_mem_node_skip(filter
, zone_to_nid(zone
), nodemask
))
4811 for_each_online_cpu(cpu
)
4812 free_pcp
+= per_cpu_ptr(zone
->pageset
, cpu
)->pcp
.count
;
4821 " active_anon:%lukB"
4822 " inactive_anon:%lukB"
4823 " active_file:%lukB"
4824 " inactive_file:%lukB"
4825 " unevictable:%lukB"
4826 " writepending:%lukB"
4830 " kernel_stack:%lukB"
4838 K(zone_page_state(zone
, NR_FREE_PAGES
)),
4839 K(min_wmark_pages(zone
)),
4840 K(low_wmark_pages(zone
)),
4841 K(high_wmark_pages(zone
)),
4842 K(zone_page_state(zone
, NR_ZONE_ACTIVE_ANON
)),
4843 K(zone_page_state(zone
, NR_ZONE_INACTIVE_ANON
)),
4844 K(zone_page_state(zone
, NR_ZONE_ACTIVE_FILE
)),
4845 K(zone_page_state(zone
, NR_ZONE_INACTIVE_FILE
)),
4846 K(zone_page_state(zone
, NR_ZONE_UNEVICTABLE
)),
4847 K(zone_page_state(zone
, NR_ZONE_WRITE_PENDING
)),
4848 K(zone
->present_pages
),
4849 K(zone
->managed_pages
),
4850 K(zone_page_state(zone
, NR_MLOCK
)),
4851 zone_page_state(zone
, NR_KERNEL_STACK_KB
),
4852 K(zone_page_state(zone
, NR_PAGETABLE
)),
4853 K(zone_page_state(zone
, NR_BOUNCE
)),
4855 K(this_cpu_read(zone
->pageset
->pcp
.count
)),
4856 K(zone_page_state(zone
, NR_FREE_CMA_PAGES
)));
4857 printk("lowmem_reserve[]:");
4858 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4859 printk(KERN_CONT
" %ld", zone
->lowmem_reserve
[i
]);
4860 printk(KERN_CONT
"\n");
4863 for_each_populated_zone(zone
) {
4865 unsigned long nr
[MAX_ORDER
], flags
, total
= 0;
4866 unsigned char types
[MAX_ORDER
];
4868 if (show_mem_node_skip(filter
, zone_to_nid(zone
), nodemask
))
4871 printk(KERN_CONT
"%s: ", zone
->name
);
4873 spin_lock_irqsave(&zone
->lock
, flags
);
4874 for (order
= 0; order
< MAX_ORDER
; order
++) {
4875 struct free_area
*area
= &zone
->free_area
[order
];
4878 nr
[order
] = area
->nr_free
;
4879 total
+= nr
[order
] << order
;
4882 for (type
= 0; type
< MIGRATE_TYPES
; type
++) {
4883 if (!list_empty(&area
->free_list
[type
]))
4884 types
[order
] |= 1 << type
;
4887 spin_unlock_irqrestore(&zone
->lock
, flags
);
4888 for (order
= 0; order
< MAX_ORDER
; order
++) {
4889 printk(KERN_CONT
"%lu*%lukB ",
4890 nr
[order
], K(1UL) << order
);
4892 show_migration_types(types
[order
]);
4894 printk(KERN_CONT
"= %lukB\n", K(total
));
4897 hugetlb_show_meminfo();
4899 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES
));
4901 show_swap_cache_info();
4904 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
4906 zoneref
->zone
= zone
;
4907 zoneref
->zone_idx
= zone_idx(zone
);
4911 * Builds allocation fallback zone lists.
4913 * Add all populated zones of a node to the zonelist.
4915 static int build_zonerefs_node(pg_data_t
*pgdat
, struct zoneref
*zonerefs
)
4918 enum zone_type zone_type
= MAX_NR_ZONES
;
4923 zone
= pgdat
->node_zones
+ zone_type
;
4924 if (managed_zone(zone
)) {
4925 zoneref_set_zone(zone
, &zonerefs
[nr_zones
++]);
4926 check_highest_zone(zone_type
);
4928 } while (zone_type
);
4935 static int __parse_numa_zonelist_order(char *s
)
4938 * We used to support different zonlists modes but they turned
4939 * out to be just not useful. Let's keep the warning in place
4940 * if somebody still use the cmd line parameter so that we do
4941 * not fail it silently
4943 if (!(*s
== 'd' || *s
== 'D' || *s
== 'n' || *s
== 'N')) {
4944 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s
);
4950 static __init
int setup_numa_zonelist_order(char *s
)
4955 return __parse_numa_zonelist_order(s
);
4957 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
4959 char numa_zonelist_order
[] = "Node";
4962 * sysctl handler for numa_zonelist_order
4964 int numa_zonelist_order_handler(struct ctl_table
*table
, int write
,
4965 void __user
*buffer
, size_t *length
,
4972 return proc_dostring(table
, write
, buffer
, length
, ppos
);
4973 str
= memdup_user_nul(buffer
, 16);
4975 return PTR_ERR(str
);
4977 ret
= __parse_numa_zonelist_order(str
);
4983 #define MAX_NODE_LOAD (nr_online_nodes)
4984 static int node_load
[MAX_NUMNODES
];
4987 * find_next_best_node - find the next node that should appear in a given node's fallback list
4988 * @node: node whose fallback list we're appending
4989 * @used_node_mask: nodemask_t of already used nodes
4991 * We use a number of factors to determine which is the next node that should
4992 * appear on a given node's fallback list. The node should not have appeared
4993 * already in @node's fallback list, and it should be the next closest node
4994 * according to the distance array (which contains arbitrary distance values
4995 * from each node to each node in the system), and should also prefer nodes
4996 * with no CPUs, since presumably they'll have very little allocation pressure
4997 * on them otherwise.
4998 * It returns -1 if no node is found.
5000 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
5003 int min_val
= INT_MAX
;
5004 int best_node
= NUMA_NO_NODE
;
5005 const struct cpumask
*tmp
= cpumask_of_node(0);
5007 /* Use the local node if we haven't already */
5008 if (!node_isset(node
, *used_node_mask
)) {
5009 node_set(node
, *used_node_mask
);
5013 for_each_node_state(n
, N_MEMORY
) {
5015 /* Don't want a node to appear more than once */
5016 if (node_isset(n
, *used_node_mask
))
5019 /* Use the distance array to find the distance */
5020 val
= node_distance(node
, n
);
5022 /* Penalize nodes under us ("prefer the next node") */
5025 /* Give preference to headless and unused nodes */
5026 tmp
= cpumask_of_node(n
);
5027 if (!cpumask_empty(tmp
))
5028 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
5030 /* Slight preference for less loaded node */
5031 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
5032 val
+= node_load
[n
];
5034 if (val
< min_val
) {
5041 node_set(best_node
, *used_node_mask
);
5048 * Build zonelists ordered by node and zones within node.
5049 * This results in maximum locality--normal zone overflows into local
5050 * DMA zone, if any--but risks exhausting DMA zone.
5052 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int *node_order
,
5055 struct zoneref
*zonerefs
;
5058 zonerefs
= pgdat
->node_zonelists
[ZONELIST_FALLBACK
]._zonerefs
;
5060 for (i
= 0; i
< nr_nodes
; i
++) {
5063 pg_data_t
*node
= NODE_DATA(node_order
[i
]);
5065 nr_zones
= build_zonerefs_node(node
, zonerefs
);
5066 zonerefs
+= nr_zones
;
5068 zonerefs
->zone
= NULL
;
5069 zonerefs
->zone_idx
= 0;
5073 * Build gfp_thisnode zonelists
5075 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
5077 struct zoneref
*zonerefs
;
5080 zonerefs
= pgdat
->node_zonelists
[ZONELIST_NOFALLBACK
]._zonerefs
;
5081 nr_zones
= build_zonerefs_node(pgdat
, zonerefs
);
5082 zonerefs
+= nr_zones
;
5083 zonerefs
->zone
= NULL
;
5084 zonerefs
->zone_idx
= 0;
5088 * Build zonelists ordered by zone and nodes within zones.
5089 * This results in conserving DMA zone[s] until all Normal memory is
5090 * exhausted, but results in overflowing to remote node while memory
5091 * may still exist in local DMA zone.
5094 static void build_zonelists(pg_data_t
*pgdat
)
5096 static int node_order
[MAX_NUMNODES
];
5097 int node
, load
, nr_nodes
= 0;
5098 nodemask_t used_mask
;
5099 int local_node
, prev_node
;
5101 /* NUMA-aware ordering of nodes */
5102 local_node
= pgdat
->node_id
;
5103 load
= nr_online_nodes
;
5104 prev_node
= local_node
;
5105 nodes_clear(used_mask
);
5107 memset(node_order
, 0, sizeof(node_order
));
5108 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
5110 * We don't want to pressure a particular node.
5111 * So adding penalty to the first node in same
5112 * distance group to make it round-robin.
5114 if (node_distance(local_node
, node
) !=
5115 node_distance(local_node
, prev_node
))
5116 node_load
[node
] = load
;
5118 node_order
[nr_nodes
++] = node
;
5123 build_zonelists_in_node_order(pgdat
, node_order
, nr_nodes
);
5124 build_thisnode_zonelists(pgdat
);
5127 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5129 * Return node id of node used for "local" allocations.
5130 * I.e., first node id of first zone in arg node's generic zonelist.
5131 * Used for initializing percpu 'numa_mem', which is used primarily
5132 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5134 int local_memory_node(int node
)
5138 z
= first_zones_zonelist(node_zonelist(node
, GFP_KERNEL
),
5139 gfp_zone(GFP_KERNEL
),
5141 return z
->zone
->node
;
5145 static void setup_min_unmapped_ratio(void);
5146 static void setup_min_slab_ratio(void);
5147 #else /* CONFIG_NUMA */
5149 static void build_zonelists(pg_data_t
*pgdat
)
5151 int node
, local_node
;
5152 struct zoneref
*zonerefs
;
5155 local_node
= pgdat
->node_id
;
5157 zonerefs
= pgdat
->node_zonelists
[ZONELIST_FALLBACK
]._zonerefs
;
5158 nr_zones
= build_zonerefs_node(pgdat
, zonerefs
);
5159 zonerefs
+= nr_zones
;
5162 * Now we build the zonelist so that it contains the zones
5163 * of all the other nodes.
5164 * We don't want to pressure a particular node, so when
5165 * building the zones for node N, we make sure that the
5166 * zones coming right after the local ones are those from
5167 * node N+1 (modulo N)
5169 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
5170 if (!node_online(node
))
5172 nr_zones
= build_zonerefs_node(NODE_DATA(node
), zonerefs
);
5173 zonerefs
+= nr_zones
;
5175 for (node
= 0; node
< local_node
; node
++) {
5176 if (!node_online(node
))
5178 nr_zones
= build_zonerefs_node(NODE_DATA(node
), zonerefs
);
5179 zonerefs
+= nr_zones
;
5182 zonerefs
->zone
= NULL
;
5183 zonerefs
->zone_idx
= 0;
5186 #endif /* CONFIG_NUMA */
5189 * Boot pageset table. One per cpu which is going to be used for all
5190 * zones and all nodes. The parameters will be set in such a way
5191 * that an item put on a list will immediately be handed over to
5192 * the buddy list. This is safe since pageset manipulation is done
5193 * with interrupts disabled.
5195 * The boot_pagesets must be kept even after bootup is complete for
5196 * unused processors and/or zones. They do play a role for bootstrapping
5197 * hotplugged processors.
5199 * zoneinfo_show() and maybe other functions do
5200 * not check if the processor is online before following the pageset pointer.
5201 * Other parts of the kernel may not check if the zone is available.
5203 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
);
5204 static DEFINE_PER_CPU(struct per_cpu_pageset
, boot_pageset
);
5205 static DEFINE_PER_CPU(struct per_cpu_nodestat
, boot_nodestats
);
5207 static void __build_all_zonelists(void *data
)
5210 int __maybe_unused cpu
;
5211 pg_data_t
*self
= data
;
5212 static DEFINE_SPINLOCK(lock
);
5217 memset(node_load
, 0, sizeof(node_load
));
5221 * This node is hotadded and no memory is yet present. So just
5222 * building zonelists is fine - no need to touch other nodes.
5224 if (self
&& !node_online(self
->node_id
)) {
5225 build_zonelists(self
);
5227 for_each_online_node(nid
) {
5228 pg_data_t
*pgdat
= NODE_DATA(nid
);
5230 build_zonelists(pgdat
);
5233 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5235 * We now know the "local memory node" for each node--
5236 * i.e., the node of the first zone in the generic zonelist.
5237 * Set up numa_mem percpu variable for on-line cpus. During
5238 * boot, only the boot cpu should be on-line; we'll init the
5239 * secondary cpus' numa_mem as they come on-line. During
5240 * node/memory hotplug, we'll fixup all on-line cpus.
5242 for_each_online_cpu(cpu
)
5243 set_cpu_numa_mem(cpu
, local_memory_node(cpu_to_node(cpu
)));
5250 static noinline
void __init
5251 build_all_zonelists_init(void)
5255 __build_all_zonelists(NULL
);
5258 * Initialize the boot_pagesets that are going to be used
5259 * for bootstrapping processors. The real pagesets for
5260 * each zone will be allocated later when the per cpu
5261 * allocator is available.
5263 * boot_pagesets are used also for bootstrapping offline
5264 * cpus if the system is already booted because the pagesets
5265 * are needed to initialize allocators on a specific cpu too.
5266 * F.e. the percpu allocator needs the page allocator which
5267 * needs the percpu allocator in order to allocate its pagesets
5268 * (a chicken-egg dilemma).
5270 for_each_possible_cpu(cpu
)
5271 setup_pageset(&per_cpu(boot_pageset
, cpu
), 0);
5273 mminit_verify_zonelist();
5274 cpuset_init_current_mems_allowed();
5278 * unless system_state == SYSTEM_BOOTING.
5280 * __ref due to call of __init annotated helper build_all_zonelists_init
5281 * [protected by SYSTEM_BOOTING].
5283 void __ref
build_all_zonelists(pg_data_t
*pgdat
)
5285 if (system_state
== SYSTEM_BOOTING
) {
5286 build_all_zonelists_init();
5288 __build_all_zonelists(pgdat
);
5289 /* cpuset refresh routine should be here */
5291 vm_total_pages
= nr_free_pagecache_pages();
5293 * Disable grouping by mobility if the number of pages in the
5294 * system is too low to allow the mechanism to work. It would be
5295 * more accurate, but expensive to check per-zone. This check is
5296 * made on memory-hotadd so a system can start with mobility
5297 * disabled and enable it later
5299 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
5300 page_group_by_mobility_disabled
= 1;
5302 page_group_by_mobility_disabled
= 0;
5304 pr_info("Built %i zonelists, mobility grouping %s. Total pages: %ld\n",
5306 page_group_by_mobility_disabled
? "off" : "on",
5309 pr_info("Policy zone: %s\n", zone_names
[policy_zone
]);
5314 * Initially all pages are reserved - free ones are freed
5315 * up by free_all_bootmem() once the early boot process is
5316 * done. Non-atomic initialization, single-pass.
5318 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
5319 unsigned long start_pfn
, enum memmap_context context
)
5321 struct vmem_altmap
*altmap
= to_vmem_altmap(__pfn_to_phys(start_pfn
));
5322 unsigned long end_pfn
= start_pfn
+ size
;
5323 pg_data_t
*pgdat
= NODE_DATA(nid
);
5325 unsigned long nr_initialised
= 0;
5326 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5327 struct memblock_region
*r
= NULL
, *tmp
;
5330 if (highest_memmap_pfn
< end_pfn
- 1)
5331 highest_memmap_pfn
= end_pfn
- 1;
5334 * Honor reservation requested by the driver for this ZONE_DEVICE
5337 if (altmap
&& start_pfn
== altmap
->base_pfn
)
5338 start_pfn
+= altmap
->reserve
;
5340 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
5342 * There can be holes in boot-time mem_map[]s handed to this
5343 * function. They do not exist on hotplugged memory.
5345 if (context
!= MEMMAP_EARLY
)
5348 if (!early_pfn_valid(pfn
))
5350 if (!early_pfn_in_nid(pfn
, nid
))
5352 if (!update_defer_init(pgdat
, pfn
, end_pfn
, &nr_initialised
))
5355 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5357 * Check given memblock attribute by firmware which can affect
5358 * kernel memory layout. If zone==ZONE_MOVABLE but memory is
5359 * mirrored, it's an overlapped memmap init. skip it.
5361 if (mirrored_kernelcore
&& zone
== ZONE_MOVABLE
) {
5362 if (!r
|| pfn
>= memblock_region_memory_end_pfn(r
)) {
5363 for_each_memblock(memory
, tmp
)
5364 if (pfn
< memblock_region_memory_end_pfn(tmp
))
5368 if (pfn
>= memblock_region_memory_base_pfn(r
) &&
5369 memblock_is_mirror(r
)) {
5370 /* already initialized as NORMAL */
5371 pfn
= memblock_region_memory_end_pfn(r
);
5379 * Mark the block movable so that blocks are reserved for
5380 * movable at startup. This will force kernel allocations
5381 * to reserve their blocks rather than leaking throughout
5382 * the address space during boot when many long-lived
5383 * kernel allocations are made.
5385 * bitmap is created for zone's valid pfn range. but memmap
5386 * can be created for invalid pages (for alignment)
5387 * check here not to call set_pageblock_migratetype() against
5390 if (!(pfn
& (pageblock_nr_pages
- 1))) {
5391 struct page
*page
= pfn_to_page(pfn
);
5393 __init_single_page(page
, pfn
, zone
, nid
);
5394 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
5397 __init_single_pfn(pfn
, zone
, nid
);
5402 static void __meminit
zone_init_free_lists(struct zone
*zone
)
5404 unsigned int order
, t
;
5405 for_each_migratetype_order(order
, t
) {
5406 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
5407 zone
->free_area
[order
].nr_free
= 0;
5411 #ifndef __HAVE_ARCH_MEMMAP_INIT
5412 #define memmap_init(size, nid, zone, start_pfn) \
5413 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
5416 static int zone_batchsize(struct zone
*zone
)
5422 * The per-cpu-pages pools are set to around 1000th of the
5423 * size of the zone. But no more than 1/2 of a meg.
5425 * OK, so we don't know how big the cache is. So guess.
5427 batch
= zone
->managed_pages
/ 1024;
5428 if (batch
* PAGE_SIZE
> 512 * 1024)
5429 batch
= (512 * 1024) / PAGE_SIZE
;
5430 batch
/= 4; /* We effectively *= 4 below */
5435 * Clamp the batch to a 2^n - 1 value. Having a power
5436 * of 2 value was found to be more likely to have
5437 * suboptimal cache aliasing properties in some cases.
5439 * For example if 2 tasks are alternately allocating
5440 * batches of pages, one task can end up with a lot
5441 * of pages of one half of the possible page colors
5442 * and the other with pages of the other colors.
5444 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
5449 /* The deferral and batching of frees should be suppressed under NOMMU
5452 * The problem is that NOMMU needs to be able to allocate large chunks
5453 * of contiguous memory as there's no hardware page translation to
5454 * assemble apparent contiguous memory from discontiguous pages.
5456 * Queueing large contiguous runs of pages for batching, however,
5457 * causes the pages to actually be freed in smaller chunks. As there
5458 * can be a significant delay between the individual batches being
5459 * recycled, this leads to the once large chunks of space being
5460 * fragmented and becoming unavailable for high-order allocations.
5467 * pcp->high and pcp->batch values are related and dependent on one another:
5468 * ->batch must never be higher then ->high.
5469 * The following function updates them in a safe manner without read side
5472 * Any new users of pcp->batch and pcp->high should ensure they can cope with
5473 * those fields changing asynchronously (acording the the above rule).
5475 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5476 * outside of boot time (or some other assurance that no concurrent updaters
5479 static void pageset_update(struct per_cpu_pages
*pcp
, unsigned long high
,
5480 unsigned long batch
)
5482 /* start with a fail safe value for batch */
5486 /* Update high, then batch, in order */
5493 /* a companion to pageset_set_high() */
5494 static void pageset_set_batch(struct per_cpu_pageset
*p
, unsigned long batch
)
5496 pageset_update(&p
->pcp
, 6 * batch
, max(1UL, 1 * batch
));
5499 static void pageset_init(struct per_cpu_pageset
*p
)
5501 struct per_cpu_pages
*pcp
;
5504 memset(p
, 0, sizeof(*p
));
5508 for (migratetype
= 0; migratetype
< MIGRATE_PCPTYPES
; migratetype
++)
5509 INIT_LIST_HEAD(&pcp
->lists
[migratetype
]);
5512 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
5515 pageset_set_batch(p
, batch
);
5519 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5520 * to the value high for the pageset p.
5522 static void pageset_set_high(struct per_cpu_pageset
*p
,
5525 unsigned long batch
= max(1UL, high
/ 4);
5526 if ((high
/ 4) > (PAGE_SHIFT
* 8))
5527 batch
= PAGE_SHIFT
* 8;
5529 pageset_update(&p
->pcp
, high
, batch
);
5532 static void pageset_set_high_and_batch(struct zone
*zone
,
5533 struct per_cpu_pageset
*pcp
)
5535 if (percpu_pagelist_fraction
)
5536 pageset_set_high(pcp
,
5537 (zone
->managed_pages
/
5538 percpu_pagelist_fraction
));
5540 pageset_set_batch(pcp
, zone_batchsize(zone
));
5543 static void __meminit
zone_pageset_init(struct zone
*zone
, int cpu
)
5545 struct per_cpu_pageset
*pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
5548 pageset_set_high_and_batch(zone
, pcp
);
5551 void __meminit
setup_zone_pageset(struct zone
*zone
)
5554 zone
->pageset
= alloc_percpu(struct per_cpu_pageset
);
5555 for_each_possible_cpu(cpu
)
5556 zone_pageset_init(zone
, cpu
);
5560 * Allocate per cpu pagesets and initialize them.
5561 * Before this call only boot pagesets were available.
5563 void __init
setup_per_cpu_pageset(void)
5565 struct pglist_data
*pgdat
;
5568 for_each_populated_zone(zone
)
5569 setup_zone_pageset(zone
);
5571 for_each_online_pgdat(pgdat
)
5572 pgdat
->per_cpu_nodestats
=
5573 alloc_percpu(struct per_cpu_nodestat
);
5576 static __meminit
void zone_pcp_init(struct zone
*zone
)
5579 * per cpu subsystem is not up at this point. The following code
5580 * relies on the ability of the linker to provide the
5581 * offset of a (static) per cpu variable into the per cpu area.
5583 zone
->pageset
= &boot_pageset
;
5585 if (populated_zone(zone
))
5586 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%u\n",
5587 zone
->name
, zone
->present_pages
,
5588 zone_batchsize(zone
));
5591 void __meminit
init_currently_empty_zone(struct zone
*zone
,
5592 unsigned long zone_start_pfn
,
5595 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
5597 pgdat
->nr_zones
= zone_idx(zone
) + 1;
5599 zone
->zone_start_pfn
= zone_start_pfn
;
5601 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
5602 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5604 (unsigned long)zone_idx(zone
),
5605 zone_start_pfn
, (zone_start_pfn
+ size
));
5607 zone_init_free_lists(zone
);
5608 zone
->initialized
= 1;
5611 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5612 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5615 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5617 int __meminit
__early_pfn_to_nid(unsigned long pfn
,
5618 struct mminit_pfnnid_cache
*state
)
5620 unsigned long start_pfn
, end_pfn
;
5623 if (state
->last_start
<= pfn
&& pfn
< state
->last_end
)
5624 return state
->last_nid
;
5626 nid
= memblock_search_pfn_nid(pfn
, &start_pfn
, &end_pfn
);
5628 state
->last_start
= start_pfn
;
5629 state
->last_end
= end_pfn
;
5630 state
->last_nid
= nid
;
5635 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5638 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5639 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5640 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5642 * If an architecture guarantees that all ranges registered contain no holes
5643 * and may be freed, this this function may be used instead of calling
5644 * memblock_free_early_nid() manually.
5646 void __init
free_bootmem_with_active_regions(int nid
, unsigned long max_low_pfn
)
5648 unsigned long start_pfn
, end_pfn
;
5651 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
) {
5652 start_pfn
= min(start_pfn
, max_low_pfn
);
5653 end_pfn
= min(end_pfn
, max_low_pfn
);
5655 if (start_pfn
< end_pfn
)
5656 memblock_free_early_nid(PFN_PHYS(start_pfn
),
5657 (end_pfn
- start_pfn
) << PAGE_SHIFT
,
5663 * sparse_memory_present_with_active_regions - Call memory_present for each active range
5664 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
5666 * If an architecture guarantees that all ranges registered contain no holes and may
5667 * be freed, this function may be used instead of calling memory_present() manually.
5669 void __init
sparse_memory_present_with_active_regions(int nid
)
5671 unsigned long start_pfn
, end_pfn
;
5674 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
)
5675 memory_present(this_nid
, start_pfn
, end_pfn
);
5679 * get_pfn_range_for_nid - Return the start and end page frames for a node
5680 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5681 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5682 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5684 * It returns the start and end page frame of a node based on information
5685 * provided by memblock_set_node(). If called for a node
5686 * with no available memory, a warning is printed and the start and end
5689 void __meminit
get_pfn_range_for_nid(unsigned int nid
,
5690 unsigned long *start_pfn
, unsigned long *end_pfn
)
5692 unsigned long this_start_pfn
, this_end_pfn
;
5698 for_each_mem_pfn_range(i
, nid
, &this_start_pfn
, &this_end_pfn
, NULL
) {
5699 *start_pfn
= min(*start_pfn
, this_start_pfn
);
5700 *end_pfn
= max(*end_pfn
, this_end_pfn
);
5703 if (*start_pfn
== -1UL)
5708 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5709 * assumption is made that zones within a node are ordered in monotonic
5710 * increasing memory addresses so that the "highest" populated zone is used
5712 static void __init
find_usable_zone_for_movable(void)
5715 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
5716 if (zone_index
== ZONE_MOVABLE
)
5719 if (arch_zone_highest_possible_pfn
[zone_index
] >
5720 arch_zone_lowest_possible_pfn
[zone_index
])
5724 VM_BUG_ON(zone_index
== -1);
5725 movable_zone
= zone_index
;
5729 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5730 * because it is sized independent of architecture. Unlike the other zones,
5731 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5732 * in each node depending on the size of each node and how evenly kernelcore
5733 * is distributed. This helper function adjusts the zone ranges
5734 * provided by the architecture for a given node by using the end of the
5735 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5736 * zones within a node are in order of monotonic increases memory addresses
5738 static void __meminit
adjust_zone_range_for_zone_movable(int nid
,
5739 unsigned long zone_type
,
5740 unsigned long node_start_pfn
,
5741 unsigned long node_end_pfn
,
5742 unsigned long *zone_start_pfn
,
5743 unsigned long *zone_end_pfn
)
5745 /* Only adjust if ZONE_MOVABLE is on this node */
5746 if (zone_movable_pfn
[nid
]) {
5747 /* Size ZONE_MOVABLE */
5748 if (zone_type
== ZONE_MOVABLE
) {
5749 *zone_start_pfn
= zone_movable_pfn
[nid
];
5750 *zone_end_pfn
= min(node_end_pfn
,
5751 arch_zone_highest_possible_pfn
[movable_zone
]);
5753 /* Adjust for ZONE_MOVABLE starting within this range */
5754 } else if (!mirrored_kernelcore
&&
5755 *zone_start_pfn
< zone_movable_pfn
[nid
] &&
5756 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
5757 *zone_end_pfn
= zone_movable_pfn
[nid
];
5759 /* Check if this whole range is within ZONE_MOVABLE */
5760 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
5761 *zone_start_pfn
= *zone_end_pfn
;
5766 * Return the number of pages a zone spans in a node, including holes
5767 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5769 static unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
5770 unsigned long zone_type
,
5771 unsigned long node_start_pfn
,
5772 unsigned long node_end_pfn
,
5773 unsigned long *zone_start_pfn
,
5774 unsigned long *zone_end_pfn
,
5775 unsigned long *ignored
)
5777 /* When hotadd a new node from cpu_up(), the node should be empty */
5778 if (!node_start_pfn
&& !node_end_pfn
)
5781 /* Get the start and end of the zone */
5782 *zone_start_pfn
= arch_zone_lowest_possible_pfn
[zone_type
];
5783 *zone_end_pfn
= arch_zone_highest_possible_pfn
[zone_type
];
5784 adjust_zone_range_for_zone_movable(nid
, zone_type
,
5785 node_start_pfn
, node_end_pfn
,
5786 zone_start_pfn
, zone_end_pfn
);
5788 /* Check that this node has pages within the zone's required range */
5789 if (*zone_end_pfn
< node_start_pfn
|| *zone_start_pfn
> node_end_pfn
)
5792 /* Move the zone boundaries inside the node if necessary */
5793 *zone_end_pfn
= min(*zone_end_pfn
, node_end_pfn
);
5794 *zone_start_pfn
= max(*zone_start_pfn
, node_start_pfn
);
5796 /* Return the spanned pages */
5797 return *zone_end_pfn
- *zone_start_pfn
;
5801 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5802 * then all holes in the requested range will be accounted for.
5804 unsigned long __meminit
__absent_pages_in_range(int nid
,
5805 unsigned long range_start_pfn
,
5806 unsigned long range_end_pfn
)
5808 unsigned long nr_absent
= range_end_pfn
- range_start_pfn
;
5809 unsigned long start_pfn
, end_pfn
;
5812 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
5813 start_pfn
= clamp(start_pfn
, range_start_pfn
, range_end_pfn
);
5814 end_pfn
= clamp(end_pfn
, range_start_pfn
, range_end_pfn
);
5815 nr_absent
-= end_pfn
- start_pfn
;
5821 * absent_pages_in_range - Return number of page frames in holes within a range
5822 * @start_pfn: The start PFN to start searching for holes
5823 * @end_pfn: The end PFN to stop searching for holes
5825 * It returns the number of pages frames in memory holes within a range.
5827 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
5828 unsigned long end_pfn
)
5830 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
5833 /* Return the number of page frames in holes in a zone on a node */
5834 static unsigned long __meminit
zone_absent_pages_in_node(int nid
,
5835 unsigned long zone_type
,
5836 unsigned long node_start_pfn
,
5837 unsigned long node_end_pfn
,
5838 unsigned long *ignored
)
5840 unsigned long zone_low
= arch_zone_lowest_possible_pfn
[zone_type
];
5841 unsigned long zone_high
= arch_zone_highest_possible_pfn
[zone_type
];
5842 unsigned long zone_start_pfn
, zone_end_pfn
;
5843 unsigned long nr_absent
;
5845 /* When hotadd a new node from cpu_up(), the node should be empty */
5846 if (!node_start_pfn
&& !node_end_pfn
)
5849 zone_start_pfn
= clamp(node_start_pfn
, zone_low
, zone_high
);
5850 zone_end_pfn
= clamp(node_end_pfn
, zone_low
, zone_high
);
5852 adjust_zone_range_for_zone_movable(nid
, zone_type
,
5853 node_start_pfn
, node_end_pfn
,
5854 &zone_start_pfn
, &zone_end_pfn
);
5855 nr_absent
= __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
5858 * ZONE_MOVABLE handling.
5859 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
5862 if (mirrored_kernelcore
&& zone_movable_pfn
[nid
]) {
5863 unsigned long start_pfn
, end_pfn
;
5864 struct memblock_region
*r
;
5866 for_each_memblock(memory
, r
) {
5867 start_pfn
= clamp(memblock_region_memory_base_pfn(r
),
5868 zone_start_pfn
, zone_end_pfn
);
5869 end_pfn
= clamp(memblock_region_memory_end_pfn(r
),
5870 zone_start_pfn
, zone_end_pfn
);
5872 if (zone_type
== ZONE_MOVABLE
&&
5873 memblock_is_mirror(r
))
5874 nr_absent
+= end_pfn
- start_pfn
;
5876 if (zone_type
== ZONE_NORMAL
&&
5877 !memblock_is_mirror(r
))
5878 nr_absent
+= end_pfn
- start_pfn
;
5885 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5886 static inline unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
5887 unsigned long zone_type
,
5888 unsigned long node_start_pfn
,
5889 unsigned long node_end_pfn
,
5890 unsigned long *zone_start_pfn
,
5891 unsigned long *zone_end_pfn
,
5892 unsigned long *zones_size
)
5896 *zone_start_pfn
= node_start_pfn
;
5897 for (zone
= 0; zone
< zone_type
; zone
++)
5898 *zone_start_pfn
+= zones_size
[zone
];
5900 *zone_end_pfn
= *zone_start_pfn
+ zones_size
[zone_type
];
5902 return zones_size
[zone_type
];
5905 static inline unsigned long __meminit
zone_absent_pages_in_node(int nid
,
5906 unsigned long zone_type
,
5907 unsigned long node_start_pfn
,
5908 unsigned long node_end_pfn
,
5909 unsigned long *zholes_size
)
5914 return zholes_size
[zone_type
];
5917 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5919 static void __meminit
calculate_node_totalpages(struct pglist_data
*pgdat
,
5920 unsigned long node_start_pfn
,
5921 unsigned long node_end_pfn
,
5922 unsigned long *zones_size
,
5923 unsigned long *zholes_size
)
5925 unsigned long realtotalpages
= 0, totalpages
= 0;
5928 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5929 struct zone
*zone
= pgdat
->node_zones
+ i
;
5930 unsigned long zone_start_pfn
, zone_end_pfn
;
5931 unsigned long size
, real_size
;
5933 size
= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
5939 real_size
= size
- zone_absent_pages_in_node(pgdat
->node_id
, i
,
5940 node_start_pfn
, node_end_pfn
,
5943 zone
->zone_start_pfn
= zone_start_pfn
;
5945 zone
->zone_start_pfn
= 0;
5946 zone
->spanned_pages
= size
;
5947 zone
->present_pages
= real_size
;
5950 realtotalpages
+= real_size
;
5953 pgdat
->node_spanned_pages
= totalpages
;
5954 pgdat
->node_present_pages
= realtotalpages
;
5955 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
5959 #ifndef CONFIG_SPARSEMEM
5961 * Calculate the size of the zone->blockflags rounded to an unsigned long
5962 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5963 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5964 * round what is now in bits to nearest long in bits, then return it in
5967 static unsigned long __init
usemap_size(unsigned long zone_start_pfn
, unsigned long zonesize
)
5969 unsigned long usemapsize
;
5971 zonesize
+= zone_start_pfn
& (pageblock_nr_pages
-1);
5972 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
5973 usemapsize
= usemapsize
>> pageblock_order
;
5974 usemapsize
*= NR_PAGEBLOCK_BITS
;
5975 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
5977 return usemapsize
/ 8;
5980 static void __init
setup_usemap(struct pglist_data
*pgdat
,
5982 unsigned long zone_start_pfn
,
5983 unsigned long zonesize
)
5985 unsigned long usemapsize
= usemap_size(zone_start_pfn
, zonesize
);
5986 zone
->pageblock_flags
= NULL
;
5988 zone
->pageblock_flags
=
5989 memblock_virt_alloc_node_nopanic(usemapsize
,
5993 static inline void setup_usemap(struct pglist_data
*pgdat
, struct zone
*zone
,
5994 unsigned long zone_start_pfn
, unsigned long zonesize
) {}
5995 #endif /* CONFIG_SPARSEMEM */
5997 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5999 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6000 void __paginginit
set_pageblock_order(void)
6004 /* Check that pageblock_nr_pages has not already been setup */
6005 if (pageblock_order
)
6008 if (HPAGE_SHIFT
> PAGE_SHIFT
)
6009 order
= HUGETLB_PAGE_ORDER
;
6011 order
= MAX_ORDER
- 1;
6014 * Assume the largest contiguous order of interest is a huge page.
6015 * This value may be variable depending on boot parameters on IA64 and
6018 pageblock_order
= order
;
6020 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6023 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6024 * is unused as pageblock_order is set at compile-time. See
6025 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6028 void __paginginit
set_pageblock_order(void)
6032 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6034 static unsigned long __paginginit
calc_memmap_size(unsigned long spanned_pages
,
6035 unsigned long present_pages
)
6037 unsigned long pages
= spanned_pages
;
6040 * Provide a more accurate estimation if there are holes within
6041 * the zone and SPARSEMEM is in use. If there are holes within the
6042 * zone, each populated memory region may cost us one or two extra
6043 * memmap pages due to alignment because memmap pages for each
6044 * populated regions may not be naturally aligned on page boundary.
6045 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6047 if (spanned_pages
> present_pages
+ (present_pages
>> 4) &&
6048 IS_ENABLED(CONFIG_SPARSEMEM
))
6049 pages
= present_pages
;
6051 return PAGE_ALIGN(pages
* sizeof(struct page
)) >> PAGE_SHIFT
;
6055 * Set up the zone data structures:
6056 * - mark all pages reserved
6057 * - mark all memory queues empty
6058 * - clear the memory bitmaps
6060 * NOTE: pgdat should get zeroed by caller.
6062 static void __paginginit
free_area_init_core(struct pglist_data
*pgdat
)
6065 int nid
= pgdat
->node_id
;
6067 pgdat_resize_init(pgdat
);
6068 #ifdef CONFIG_NUMA_BALANCING
6069 spin_lock_init(&pgdat
->numabalancing_migrate_lock
);
6070 pgdat
->numabalancing_migrate_nr_pages
= 0;
6071 pgdat
->numabalancing_migrate_next_window
= jiffies
;
6073 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6074 spin_lock_init(&pgdat
->split_queue_lock
);
6075 INIT_LIST_HEAD(&pgdat
->split_queue
);
6076 pgdat
->split_queue_len
= 0;
6078 init_waitqueue_head(&pgdat
->kswapd_wait
);
6079 init_waitqueue_head(&pgdat
->pfmemalloc_wait
);
6080 #ifdef CONFIG_COMPACTION
6081 init_waitqueue_head(&pgdat
->kcompactd_wait
);
6083 pgdat_page_ext_init(pgdat
);
6084 spin_lock_init(&pgdat
->lru_lock
);
6085 lruvec_init(node_lruvec(pgdat
));
6087 pgdat
->per_cpu_nodestats
= &boot_nodestats
;
6089 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
6090 struct zone
*zone
= pgdat
->node_zones
+ j
;
6091 unsigned long size
, realsize
, freesize
, memmap_pages
;
6092 unsigned long zone_start_pfn
= zone
->zone_start_pfn
;
6094 size
= zone
->spanned_pages
;
6095 realsize
= freesize
= zone
->present_pages
;
6098 * Adjust freesize so that it accounts for how much memory
6099 * is used by this zone for memmap. This affects the watermark
6100 * and per-cpu initialisations
6102 memmap_pages
= calc_memmap_size(size
, realsize
);
6103 if (!is_highmem_idx(j
)) {
6104 if (freesize
>= memmap_pages
) {
6105 freesize
-= memmap_pages
;
6108 " %s zone: %lu pages used for memmap\n",
6109 zone_names
[j
], memmap_pages
);
6111 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6112 zone_names
[j
], memmap_pages
, freesize
);
6115 /* Account for reserved pages */
6116 if (j
== 0 && freesize
> dma_reserve
) {
6117 freesize
-= dma_reserve
;
6118 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
6119 zone_names
[0], dma_reserve
);
6122 if (!is_highmem_idx(j
))
6123 nr_kernel_pages
+= freesize
;
6124 /* Charge for highmem memmap if there are enough kernel pages */
6125 else if (nr_kernel_pages
> memmap_pages
* 2)
6126 nr_kernel_pages
-= memmap_pages
;
6127 nr_all_pages
+= freesize
;
6130 * Set an approximate value for lowmem here, it will be adjusted
6131 * when the bootmem allocator frees pages into the buddy system.
6132 * And all highmem pages will be managed by the buddy system.
6134 zone
->managed_pages
= is_highmem_idx(j
) ? realsize
: freesize
;
6138 zone
->name
= zone_names
[j
];
6139 zone
->zone_pgdat
= pgdat
;
6140 spin_lock_init(&zone
->lock
);
6141 zone_seqlock_init(zone
);
6142 zone_pcp_init(zone
);
6147 set_pageblock_order();
6148 setup_usemap(pgdat
, zone
, zone_start_pfn
, size
);
6149 init_currently_empty_zone(zone
, zone_start_pfn
, size
);
6150 memmap_init(size
, nid
, j
, zone_start_pfn
);
6154 static void __ref
alloc_node_mem_map(struct pglist_data
*pgdat
)
6156 unsigned long __maybe_unused start
= 0;
6157 unsigned long __maybe_unused offset
= 0;
6159 /* Skip empty nodes */
6160 if (!pgdat
->node_spanned_pages
)
6163 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6164 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
6165 offset
= pgdat
->node_start_pfn
- start
;
6166 /* ia64 gets its own node_mem_map, before this, without bootmem */
6167 if (!pgdat
->node_mem_map
) {
6168 unsigned long size
, end
;
6172 * The zone's endpoints aren't required to be MAX_ORDER
6173 * aligned but the node_mem_map endpoints must be in order
6174 * for the buddy allocator to function correctly.
6176 end
= pgdat_end_pfn(pgdat
);
6177 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
6178 size
= (end
- start
) * sizeof(struct page
);
6179 map
= alloc_remap(pgdat
->node_id
, size
);
6181 map
= memblock_virt_alloc_node_nopanic(size
,
6183 pgdat
->node_mem_map
= map
+ offset
;
6185 #ifndef CONFIG_NEED_MULTIPLE_NODES
6187 * With no DISCONTIG, the global mem_map is just set as node 0's
6189 if (pgdat
== NODE_DATA(0)) {
6190 mem_map
= NODE_DATA(0)->node_mem_map
;
6191 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6192 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
6194 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6197 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6200 void __paginginit
free_area_init_node(int nid
, unsigned long *zones_size
,
6201 unsigned long node_start_pfn
, unsigned long *zholes_size
)
6203 pg_data_t
*pgdat
= NODE_DATA(nid
);
6204 unsigned long start_pfn
= 0;
6205 unsigned long end_pfn
= 0;
6207 /* pg_data_t should be reset to zero when it's allocated */
6208 WARN_ON(pgdat
->nr_zones
|| pgdat
->kswapd_classzone_idx
);
6210 pgdat
->node_id
= nid
;
6211 pgdat
->node_start_pfn
= node_start_pfn
;
6212 pgdat
->per_cpu_nodestats
= NULL
;
6213 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6214 get_pfn_range_for_nid(nid
, &start_pfn
, &end_pfn
);
6215 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid
,
6216 (u64
)start_pfn
<< PAGE_SHIFT
,
6217 end_pfn
? ((u64
)end_pfn
<< PAGE_SHIFT
) - 1 : 0);
6219 start_pfn
= node_start_pfn
;
6221 calculate_node_totalpages(pgdat
, start_pfn
, end_pfn
,
6222 zones_size
, zholes_size
);
6224 alloc_node_mem_map(pgdat
);
6225 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6226 printk(KERN_DEBUG
"free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
6227 nid
, (unsigned long)pgdat
,
6228 (unsigned long)pgdat
->node_mem_map
);
6231 reset_deferred_meminit(pgdat
);
6232 free_area_init_core(pgdat
);
6235 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6237 #if MAX_NUMNODES > 1
6239 * Figure out the number of possible node ids.
6241 void __init
setup_nr_node_ids(void)
6243 unsigned int highest
;
6245 highest
= find_last_bit(node_possible_map
.bits
, MAX_NUMNODES
);
6246 nr_node_ids
= highest
+ 1;
6251 * node_map_pfn_alignment - determine the maximum internode alignment
6253 * This function should be called after node map is populated and sorted.
6254 * It calculates the maximum power of two alignment which can distinguish
6257 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
6258 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
6259 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
6260 * shifted, 1GiB is enough and this function will indicate so.
6262 * This is used to test whether pfn -> nid mapping of the chosen memory
6263 * model has fine enough granularity to avoid incorrect mapping for the
6264 * populated node map.
6266 * Returns the determined alignment in pfn's. 0 if there is no alignment
6267 * requirement (single node).
6269 unsigned long __init
node_map_pfn_alignment(void)
6271 unsigned long accl_mask
= 0, last_end
= 0;
6272 unsigned long start
, end
, mask
;
6276 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start
, &end
, &nid
) {
6277 if (!start
|| last_nid
< 0 || last_nid
== nid
) {
6284 * Start with a mask granular enough to pin-point to the
6285 * start pfn and tick off bits one-by-one until it becomes
6286 * too coarse to separate the current node from the last.
6288 mask
= ~((1 << __ffs(start
)) - 1);
6289 while (mask
&& last_end
<= (start
& (mask
<< 1)))
6292 /* accumulate all internode masks */
6296 /* convert mask to number of pages */
6297 return ~accl_mask
+ 1;
6300 /* Find the lowest pfn for a node */
6301 static unsigned long __init
find_min_pfn_for_node(int nid
)
6303 unsigned long min_pfn
= ULONG_MAX
;
6304 unsigned long start_pfn
;
6307 for_each_mem_pfn_range(i
, nid
, &start_pfn
, NULL
, NULL
)
6308 min_pfn
= min(min_pfn
, start_pfn
);
6310 if (min_pfn
== ULONG_MAX
) {
6311 pr_warn("Could not find start_pfn for node %d\n", nid
);
6319 * find_min_pfn_with_active_regions - Find the minimum PFN registered
6321 * It returns the minimum PFN based on information provided via
6322 * memblock_set_node().
6324 unsigned long __init
find_min_pfn_with_active_regions(void)
6326 return find_min_pfn_for_node(MAX_NUMNODES
);
6330 * early_calculate_totalpages()
6331 * Sum pages in active regions for movable zone.
6332 * Populate N_MEMORY for calculating usable_nodes.
6334 static unsigned long __init
early_calculate_totalpages(void)
6336 unsigned long totalpages
= 0;
6337 unsigned long start_pfn
, end_pfn
;
6340 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
6341 unsigned long pages
= end_pfn
- start_pfn
;
6343 totalpages
+= pages
;
6345 node_set_state(nid
, N_MEMORY
);
6351 * Find the PFN the Movable zone begins in each node. Kernel memory
6352 * is spread evenly between nodes as long as the nodes have enough
6353 * memory. When they don't, some nodes will have more kernelcore than
6356 static void __init
find_zone_movable_pfns_for_nodes(void)
6359 unsigned long usable_startpfn
;
6360 unsigned long kernelcore_node
, kernelcore_remaining
;
6361 /* save the state before borrow the nodemask */
6362 nodemask_t saved_node_state
= node_states
[N_MEMORY
];
6363 unsigned long totalpages
= early_calculate_totalpages();
6364 int usable_nodes
= nodes_weight(node_states
[N_MEMORY
]);
6365 struct memblock_region
*r
;
6367 /* Need to find movable_zone earlier when movable_node is specified. */
6368 find_usable_zone_for_movable();
6371 * If movable_node is specified, ignore kernelcore and movablecore
6374 if (movable_node_is_enabled()) {
6375 for_each_memblock(memory
, r
) {
6376 if (!memblock_is_hotpluggable(r
))
6381 usable_startpfn
= PFN_DOWN(r
->base
);
6382 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
6383 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
6391 * If kernelcore=mirror is specified, ignore movablecore option
6393 if (mirrored_kernelcore
) {
6394 bool mem_below_4gb_not_mirrored
= false;
6396 for_each_memblock(memory
, r
) {
6397 if (memblock_is_mirror(r
))
6402 usable_startpfn
= memblock_region_memory_base_pfn(r
);
6404 if (usable_startpfn
< 0x100000) {
6405 mem_below_4gb_not_mirrored
= true;
6409 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
6410 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
6414 if (mem_below_4gb_not_mirrored
)
6415 pr_warn("This configuration results in unmirrored kernel memory.");
6421 * If movablecore=nn[KMG] was specified, calculate what size of
6422 * kernelcore that corresponds so that memory usable for
6423 * any allocation type is evenly spread. If both kernelcore
6424 * and movablecore are specified, then the value of kernelcore
6425 * will be used for required_kernelcore if it's greater than
6426 * what movablecore would have allowed.
6428 if (required_movablecore
) {
6429 unsigned long corepages
;
6432 * Round-up so that ZONE_MOVABLE is at least as large as what
6433 * was requested by the user
6435 required_movablecore
=
6436 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
6437 required_movablecore
= min(totalpages
, required_movablecore
);
6438 corepages
= totalpages
- required_movablecore
;
6440 required_kernelcore
= max(required_kernelcore
, corepages
);
6444 * If kernelcore was not specified or kernelcore size is larger
6445 * than totalpages, there is no ZONE_MOVABLE.
6447 if (!required_kernelcore
|| required_kernelcore
>= totalpages
)
6450 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
6451 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
6454 /* Spread kernelcore memory as evenly as possible throughout nodes */
6455 kernelcore_node
= required_kernelcore
/ usable_nodes
;
6456 for_each_node_state(nid
, N_MEMORY
) {
6457 unsigned long start_pfn
, end_pfn
;
6460 * Recalculate kernelcore_node if the division per node
6461 * now exceeds what is necessary to satisfy the requested
6462 * amount of memory for the kernel
6464 if (required_kernelcore
< kernelcore_node
)
6465 kernelcore_node
= required_kernelcore
/ usable_nodes
;
6468 * As the map is walked, we track how much memory is usable
6469 * by the kernel using kernelcore_remaining. When it is
6470 * 0, the rest of the node is usable by ZONE_MOVABLE
6472 kernelcore_remaining
= kernelcore_node
;
6474 /* Go through each range of PFNs within this node */
6475 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
6476 unsigned long size_pages
;
6478 start_pfn
= max(start_pfn
, zone_movable_pfn
[nid
]);
6479 if (start_pfn
>= end_pfn
)
6482 /* Account for what is only usable for kernelcore */
6483 if (start_pfn
< usable_startpfn
) {
6484 unsigned long kernel_pages
;
6485 kernel_pages
= min(end_pfn
, usable_startpfn
)
6488 kernelcore_remaining
-= min(kernel_pages
,
6489 kernelcore_remaining
);
6490 required_kernelcore
-= min(kernel_pages
,
6491 required_kernelcore
);
6493 /* Continue if range is now fully accounted */
6494 if (end_pfn
<= usable_startpfn
) {
6497 * Push zone_movable_pfn to the end so
6498 * that if we have to rebalance
6499 * kernelcore across nodes, we will
6500 * not double account here
6502 zone_movable_pfn
[nid
] = end_pfn
;
6505 start_pfn
= usable_startpfn
;
6509 * The usable PFN range for ZONE_MOVABLE is from
6510 * start_pfn->end_pfn. Calculate size_pages as the
6511 * number of pages used as kernelcore
6513 size_pages
= end_pfn
- start_pfn
;
6514 if (size_pages
> kernelcore_remaining
)
6515 size_pages
= kernelcore_remaining
;
6516 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
6519 * Some kernelcore has been met, update counts and
6520 * break if the kernelcore for this node has been
6523 required_kernelcore
-= min(required_kernelcore
,
6525 kernelcore_remaining
-= size_pages
;
6526 if (!kernelcore_remaining
)
6532 * If there is still required_kernelcore, we do another pass with one
6533 * less node in the count. This will push zone_movable_pfn[nid] further
6534 * along on the nodes that still have memory until kernelcore is
6538 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
6542 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
6543 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
6544 zone_movable_pfn
[nid
] =
6545 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
6548 /* restore the node_state */
6549 node_states
[N_MEMORY
] = saved_node_state
;
6552 /* Any regular or high memory on that node ? */
6553 static void check_for_memory(pg_data_t
*pgdat
, int nid
)
6555 enum zone_type zone_type
;
6557 if (N_MEMORY
== N_NORMAL_MEMORY
)
6560 for (zone_type
= 0; zone_type
<= ZONE_MOVABLE
- 1; zone_type
++) {
6561 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
6562 if (populated_zone(zone
)) {
6563 node_set_state(nid
, N_HIGH_MEMORY
);
6564 if (N_NORMAL_MEMORY
!= N_HIGH_MEMORY
&&
6565 zone_type
<= ZONE_NORMAL
)
6566 node_set_state(nid
, N_NORMAL_MEMORY
);
6573 * free_area_init_nodes - Initialise all pg_data_t and zone data
6574 * @max_zone_pfn: an array of max PFNs for each zone
6576 * This will call free_area_init_node() for each active node in the system.
6577 * Using the page ranges provided by memblock_set_node(), the size of each
6578 * zone in each node and their holes is calculated. If the maximum PFN
6579 * between two adjacent zones match, it is assumed that the zone is empty.
6580 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
6581 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
6582 * starts where the previous one ended. For example, ZONE_DMA32 starts
6583 * at arch_max_dma_pfn.
6585 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
6587 unsigned long start_pfn
, end_pfn
;
6590 /* Record where the zone boundaries are */
6591 memset(arch_zone_lowest_possible_pfn
, 0,
6592 sizeof(arch_zone_lowest_possible_pfn
));
6593 memset(arch_zone_highest_possible_pfn
, 0,
6594 sizeof(arch_zone_highest_possible_pfn
));
6596 start_pfn
= find_min_pfn_with_active_regions();
6598 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
6599 if (i
== ZONE_MOVABLE
)
6602 end_pfn
= max(max_zone_pfn
[i
], start_pfn
);
6603 arch_zone_lowest_possible_pfn
[i
] = start_pfn
;
6604 arch_zone_highest_possible_pfn
[i
] = end_pfn
;
6606 start_pfn
= end_pfn
;
6609 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
6610 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
6611 find_zone_movable_pfns_for_nodes();
6613 /* Print out the zone ranges */
6614 pr_info("Zone ranges:\n");
6615 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
6616 if (i
== ZONE_MOVABLE
)
6618 pr_info(" %-8s ", zone_names
[i
]);
6619 if (arch_zone_lowest_possible_pfn
[i
] ==
6620 arch_zone_highest_possible_pfn
[i
])
6623 pr_cont("[mem %#018Lx-%#018Lx]\n",
6624 (u64
)arch_zone_lowest_possible_pfn
[i
]
6626 ((u64
)arch_zone_highest_possible_pfn
[i
]
6627 << PAGE_SHIFT
) - 1);
6630 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
6631 pr_info("Movable zone start for each node\n");
6632 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6633 if (zone_movable_pfn
[i
])
6634 pr_info(" Node %d: %#018Lx\n", i
,
6635 (u64
)zone_movable_pfn
[i
] << PAGE_SHIFT
);
6638 /* Print out the early node map */
6639 pr_info("Early memory node ranges\n");
6640 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
)
6641 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid
,
6642 (u64
)start_pfn
<< PAGE_SHIFT
,
6643 ((u64
)end_pfn
<< PAGE_SHIFT
) - 1);
6645 /* Initialise every node */
6646 mminit_verify_pageflags_layout();
6647 setup_nr_node_ids();
6648 for_each_online_node(nid
) {
6649 pg_data_t
*pgdat
= NODE_DATA(nid
);
6650 free_area_init_node(nid
, NULL
,
6651 find_min_pfn_for_node(nid
), NULL
);
6653 /* Any memory on that node */
6654 if (pgdat
->node_present_pages
)
6655 node_set_state(nid
, N_MEMORY
);
6656 check_for_memory(pgdat
, nid
);
6660 static int __init
cmdline_parse_core(char *p
, unsigned long *core
)
6662 unsigned long long coremem
;
6666 coremem
= memparse(p
, &p
);
6667 *core
= coremem
>> PAGE_SHIFT
;
6669 /* Paranoid check that UL is enough for the coremem value */
6670 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
6676 * kernelcore=size sets the amount of memory for use for allocations that
6677 * cannot be reclaimed or migrated.
6679 static int __init
cmdline_parse_kernelcore(char *p
)
6681 /* parse kernelcore=mirror */
6682 if (parse_option_str(p
, "mirror")) {
6683 mirrored_kernelcore
= true;
6687 return cmdline_parse_core(p
, &required_kernelcore
);
6691 * movablecore=size sets the amount of memory for use for allocations that
6692 * can be reclaimed or migrated.
6694 static int __init
cmdline_parse_movablecore(char *p
)
6696 return cmdline_parse_core(p
, &required_movablecore
);
6699 early_param("kernelcore", cmdline_parse_kernelcore
);
6700 early_param("movablecore", cmdline_parse_movablecore
);
6702 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6704 void adjust_managed_page_count(struct page
*page
, long count
)
6706 spin_lock(&managed_page_count_lock
);
6707 page_zone(page
)->managed_pages
+= count
;
6708 totalram_pages
+= count
;
6709 #ifdef CONFIG_HIGHMEM
6710 if (PageHighMem(page
))
6711 totalhigh_pages
+= count
;
6713 spin_unlock(&managed_page_count_lock
);
6715 EXPORT_SYMBOL(adjust_managed_page_count
);
6717 unsigned long free_reserved_area(void *start
, void *end
, int poison
, char *s
)
6720 unsigned long pages
= 0;
6722 start
= (void *)PAGE_ALIGN((unsigned long)start
);
6723 end
= (void *)((unsigned long)end
& PAGE_MASK
);
6724 for (pos
= start
; pos
< end
; pos
+= PAGE_SIZE
, pages
++) {
6725 if ((unsigned int)poison
<= 0xFF)
6726 memset(pos
, poison
, PAGE_SIZE
);
6727 free_reserved_page(virt_to_page(pos
));
6731 pr_info("Freeing %s memory: %ldK\n",
6732 s
, pages
<< (PAGE_SHIFT
- 10));
6736 EXPORT_SYMBOL(free_reserved_area
);
6738 #ifdef CONFIG_HIGHMEM
6739 void free_highmem_page(struct page
*page
)
6741 __free_reserved_page(page
);
6743 page_zone(page
)->managed_pages
++;
6749 void __init
mem_init_print_info(const char *str
)
6751 unsigned long physpages
, codesize
, datasize
, rosize
, bss_size
;
6752 unsigned long init_code_size
, init_data_size
;
6754 physpages
= get_num_physpages();
6755 codesize
= _etext
- _stext
;
6756 datasize
= _edata
- _sdata
;
6757 rosize
= __end_rodata
- __start_rodata
;
6758 bss_size
= __bss_stop
- __bss_start
;
6759 init_data_size
= __init_end
- __init_begin
;
6760 init_code_size
= _einittext
- _sinittext
;
6763 * Detect special cases and adjust section sizes accordingly:
6764 * 1) .init.* may be embedded into .data sections
6765 * 2) .init.text.* may be out of [__init_begin, __init_end],
6766 * please refer to arch/tile/kernel/vmlinux.lds.S.
6767 * 3) .rodata.* may be embedded into .text or .data sections.
6769 #define adj_init_size(start, end, size, pos, adj) \
6771 if (start <= pos && pos < end && size > adj) \
6775 adj_init_size(__init_begin
, __init_end
, init_data_size
,
6776 _sinittext
, init_code_size
);
6777 adj_init_size(_stext
, _etext
, codesize
, _sinittext
, init_code_size
);
6778 adj_init_size(_sdata
, _edata
, datasize
, __init_begin
, init_data_size
);
6779 adj_init_size(_stext
, _etext
, codesize
, __start_rodata
, rosize
);
6780 adj_init_size(_sdata
, _edata
, datasize
, __start_rodata
, rosize
);
6782 #undef adj_init_size
6784 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
6785 #ifdef CONFIG_HIGHMEM
6789 nr_free_pages() << (PAGE_SHIFT
- 10),
6790 physpages
<< (PAGE_SHIFT
- 10),
6791 codesize
>> 10, datasize
>> 10, rosize
>> 10,
6792 (init_data_size
+ init_code_size
) >> 10, bss_size
>> 10,
6793 (physpages
- totalram_pages
- totalcma_pages
) << (PAGE_SHIFT
- 10),
6794 totalcma_pages
<< (PAGE_SHIFT
- 10),
6795 #ifdef CONFIG_HIGHMEM
6796 totalhigh_pages
<< (PAGE_SHIFT
- 10),
6798 str
? ", " : "", str
? str
: "");
6802 * set_dma_reserve - set the specified number of pages reserved in the first zone
6803 * @new_dma_reserve: The number of pages to mark reserved
6805 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
6806 * In the DMA zone, a significant percentage may be consumed by kernel image
6807 * and other unfreeable allocations which can skew the watermarks badly. This
6808 * function may optionally be used to account for unfreeable pages in the
6809 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
6810 * smaller per-cpu batchsize.
6812 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
6814 dma_reserve
= new_dma_reserve
;
6817 void __init
free_area_init(unsigned long *zones_size
)
6819 free_area_init_node(0, zones_size
,
6820 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
6823 static int page_alloc_cpu_dead(unsigned int cpu
)
6826 lru_add_drain_cpu(cpu
);
6830 * Spill the event counters of the dead processor
6831 * into the current processors event counters.
6832 * This artificially elevates the count of the current
6835 vm_events_fold_cpu(cpu
);
6838 * Zero the differential counters of the dead processor
6839 * so that the vm statistics are consistent.
6841 * This is only okay since the processor is dead and cannot
6842 * race with what we are doing.
6844 cpu_vm_stats_fold(cpu
);
6848 void __init
page_alloc_init(void)
6852 ret
= cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD
,
6853 "mm/page_alloc:dead", NULL
,
6854 page_alloc_cpu_dead
);
6859 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6860 * or min_free_kbytes changes.
6862 static void calculate_totalreserve_pages(void)
6864 struct pglist_data
*pgdat
;
6865 unsigned long reserve_pages
= 0;
6866 enum zone_type i
, j
;
6868 for_each_online_pgdat(pgdat
) {
6870 pgdat
->totalreserve_pages
= 0;
6872 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
6873 struct zone
*zone
= pgdat
->node_zones
+ i
;
6876 /* Find valid and maximum lowmem_reserve in the zone */
6877 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
6878 if (zone
->lowmem_reserve
[j
] > max
)
6879 max
= zone
->lowmem_reserve
[j
];
6882 /* we treat the high watermark as reserved pages. */
6883 max
+= high_wmark_pages(zone
);
6885 if (max
> zone
->managed_pages
)
6886 max
= zone
->managed_pages
;
6888 pgdat
->totalreserve_pages
+= max
;
6890 reserve_pages
+= max
;
6893 totalreserve_pages
= reserve_pages
;
6897 * setup_per_zone_lowmem_reserve - called whenever
6898 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6899 * has a correct pages reserved value, so an adequate number of
6900 * pages are left in the zone after a successful __alloc_pages().
6902 static void setup_per_zone_lowmem_reserve(void)
6904 struct pglist_data
*pgdat
;
6905 enum zone_type j
, idx
;
6907 for_each_online_pgdat(pgdat
) {
6908 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
6909 struct zone
*zone
= pgdat
->node_zones
+ j
;
6910 unsigned long managed_pages
= zone
->managed_pages
;
6912 zone
->lowmem_reserve
[j
] = 0;
6916 struct zone
*lower_zone
;
6920 if (sysctl_lowmem_reserve_ratio
[idx
] < 1)
6921 sysctl_lowmem_reserve_ratio
[idx
] = 1;
6923 lower_zone
= pgdat
->node_zones
+ idx
;
6924 lower_zone
->lowmem_reserve
[j
] = managed_pages
/
6925 sysctl_lowmem_reserve_ratio
[idx
];
6926 managed_pages
+= lower_zone
->managed_pages
;
6931 /* update totalreserve_pages */
6932 calculate_totalreserve_pages();
6935 static void __setup_per_zone_wmarks(void)
6937 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
6938 unsigned long pages_low
= extra_free_kbytes
>> (PAGE_SHIFT
- 10);
6939 unsigned long lowmem_pages
= 0;
6941 unsigned long flags
;
6943 /* Calculate total number of !ZONE_HIGHMEM pages */
6944 for_each_zone(zone
) {
6945 if (!is_highmem(zone
))
6946 lowmem_pages
+= zone
->managed_pages
;
6949 for_each_zone(zone
) {
6952 spin_lock_irqsave(&zone
->lock
, flags
);
6953 min
= (u64
)pages_min
* zone
->managed_pages
;
6954 do_div(min
, lowmem_pages
);
6955 low
= (u64
)pages_low
* zone
->managed_pages
;
6956 do_div(low
, vm_total_pages
);
6958 if (is_highmem(zone
)) {
6960 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6961 * need highmem pages, so cap pages_min to a small
6964 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6965 * deltas control asynch page reclaim, and so should
6966 * not be capped for highmem.
6968 unsigned long min_pages
;
6970 min_pages
= zone
->managed_pages
/ 1024;
6971 min_pages
= clamp(min_pages
, SWAP_CLUSTER_MAX
, 128UL);
6972 zone
->watermark
[WMARK_MIN
] = min_pages
;
6975 * If it's a lowmem zone, reserve a number of pages
6976 * proportionate to the zone's size.
6978 zone
->watermark
[WMARK_MIN
] = min
;
6982 * Set the kswapd watermarks distance according to the
6983 * scale factor in proportion to available memory, but
6984 * ensure a minimum size on small systems.
6986 min
= max_t(u64
, min
>> 2,
6987 mult_frac(zone
->managed_pages
,
6988 watermark_scale_factor
, 10000));
6990 zone
->watermark
[WMARK_LOW
] = min_wmark_pages(zone
) +
6992 zone
->watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) +
6995 spin_unlock_irqrestore(&zone
->lock
, flags
);
6998 /* update totalreserve_pages */
6999 calculate_totalreserve_pages();
7003 * setup_per_zone_wmarks - called when min_free_kbytes changes
7004 * or when memory is hot-{added|removed}
7006 * Ensures that the watermark[min,low,high] values for each zone are set
7007 * correctly with respect to min_free_kbytes.
7009 void setup_per_zone_wmarks(void)
7011 static DEFINE_SPINLOCK(lock
);
7014 __setup_per_zone_wmarks();
7019 * Initialise min_free_kbytes.
7021 * For small machines we want it small (128k min). For large machines
7022 * we want it large (64MB max). But it is not linear, because network
7023 * bandwidth does not increase linearly with machine size. We use
7025 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7026 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7042 int __meminit
init_per_zone_wmark_min(void)
7044 unsigned long lowmem_kbytes
;
7045 int new_min_free_kbytes
;
7047 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
7048 new_min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
7050 if (new_min_free_kbytes
> user_min_free_kbytes
) {
7051 min_free_kbytes
= new_min_free_kbytes
;
7052 if (min_free_kbytes
< 128)
7053 min_free_kbytes
= 128;
7054 if (min_free_kbytes
> 65536)
7055 min_free_kbytes
= 65536;
7057 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7058 new_min_free_kbytes
, user_min_free_kbytes
);
7060 setup_per_zone_wmarks();
7061 refresh_zone_stat_thresholds();
7062 setup_per_zone_lowmem_reserve();
7065 setup_min_unmapped_ratio();
7066 setup_min_slab_ratio();
7071 core_initcall(init_per_zone_wmark_min
)
7074 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7075 * that we can call two helper functions whenever min_free_kbytes
7076 * or extra_free_kbytes changes.
7078 int min_free_kbytes_sysctl_handler(struct ctl_table
*table
, int write
,
7079 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7083 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7088 user_min_free_kbytes
= min_free_kbytes
;
7089 setup_per_zone_wmarks();
7094 int watermark_scale_factor_sysctl_handler(struct ctl_table
*table
, int write
,
7095 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7099 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7104 setup_per_zone_wmarks();
7110 static void setup_min_unmapped_ratio(void)
7115 for_each_online_pgdat(pgdat
)
7116 pgdat
->min_unmapped_pages
= 0;
7119 zone
->zone_pgdat
->min_unmapped_pages
+= (zone
->managed_pages
*
7120 sysctl_min_unmapped_ratio
) / 100;
7124 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
7125 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7129 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7133 setup_min_unmapped_ratio();
7138 static void setup_min_slab_ratio(void)
7143 for_each_online_pgdat(pgdat
)
7144 pgdat
->min_slab_pages
= 0;
7147 zone
->zone_pgdat
->min_slab_pages
+= (zone
->managed_pages
*
7148 sysctl_min_slab_ratio
) / 100;
7151 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
7152 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7156 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7160 setup_min_slab_ratio();
7167 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
7168 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
7169 * whenever sysctl_lowmem_reserve_ratio changes.
7171 * The reserve ratio obviously has absolutely no relation with the
7172 * minimum watermarks. The lowmem reserve ratio can only make sense
7173 * if in function of the boot time zone sizes.
7175 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
7176 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7178 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7179 setup_per_zone_lowmem_reserve();
7184 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
7185 * cpu. It is the fraction of total pages in each zone that a hot per cpu
7186 * pagelist can have before it gets flushed back to buddy allocator.
7188 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table
*table
, int write
,
7189 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
7192 int old_percpu_pagelist_fraction
;
7195 mutex_lock(&pcp_batch_high_lock
);
7196 old_percpu_pagelist_fraction
= percpu_pagelist_fraction
;
7198 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
7199 if (!write
|| ret
< 0)
7202 /* Sanity checking to avoid pcp imbalance */
7203 if (percpu_pagelist_fraction
&&
7204 percpu_pagelist_fraction
< MIN_PERCPU_PAGELIST_FRACTION
) {
7205 percpu_pagelist_fraction
= old_percpu_pagelist_fraction
;
7211 if (percpu_pagelist_fraction
== old_percpu_pagelist_fraction
)
7214 for_each_populated_zone(zone
) {
7217 for_each_possible_cpu(cpu
)
7218 pageset_set_high_and_batch(zone
,
7219 per_cpu_ptr(zone
->pageset
, cpu
));
7222 mutex_unlock(&pcp_batch_high_lock
);
7227 int hashdist
= HASHDIST_DEFAULT
;
7229 static int __init
set_hashdist(char *str
)
7233 hashdist
= simple_strtoul(str
, &str
, 0);
7236 __setup("hashdist=", set_hashdist
);
7239 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
7241 * Returns the number of pages that arch has reserved but
7242 * is not known to alloc_large_system_hash().
7244 static unsigned long __init
arch_reserved_kernel_pages(void)
7251 * Adaptive scale is meant to reduce sizes of hash tables on large memory
7252 * machines. As memory size is increased the scale is also increased but at
7253 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
7254 * quadruples the scale is increased by one, which means the size of hash table
7255 * only doubles, instead of quadrupling as well.
7256 * Because 32-bit systems cannot have large physical memory, where this scaling
7257 * makes sense, it is disabled on such platforms.
7259 #if __BITS_PER_LONG > 32
7260 #define ADAPT_SCALE_BASE (64ul << 30)
7261 #define ADAPT_SCALE_SHIFT 2
7262 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
7266 * allocate a large system hash table from bootmem
7267 * - it is assumed that the hash table must contain an exact power-of-2
7268 * quantity of entries
7269 * - limit is the number of hash buckets, not the total allocation size
7271 void *__init
alloc_large_system_hash(const char *tablename
,
7272 unsigned long bucketsize
,
7273 unsigned long numentries
,
7276 unsigned int *_hash_shift
,
7277 unsigned int *_hash_mask
,
7278 unsigned long low_limit
,
7279 unsigned long high_limit
)
7281 unsigned long long max
= high_limit
;
7282 unsigned long log2qty
, size
;
7286 /* allow the kernel cmdline to have a say */
7288 /* round applicable memory size up to nearest megabyte */
7289 numentries
= nr_kernel_pages
;
7290 numentries
-= arch_reserved_kernel_pages();
7292 /* It isn't necessary when PAGE_SIZE >= 1MB */
7293 if (PAGE_SHIFT
< 20)
7294 numentries
= round_up(numentries
, (1<<20)/PAGE_SIZE
);
7296 #if __BITS_PER_LONG > 32
7298 unsigned long adapt
;
7300 for (adapt
= ADAPT_SCALE_NPAGES
; adapt
< numentries
;
7301 adapt
<<= ADAPT_SCALE_SHIFT
)
7306 /* limit to 1 bucket per 2^scale bytes of low memory */
7307 if (scale
> PAGE_SHIFT
)
7308 numentries
>>= (scale
- PAGE_SHIFT
);
7310 numentries
<<= (PAGE_SHIFT
- scale
);
7312 /* Make sure we've got at least a 0-order allocation.. */
7313 if (unlikely(flags
& HASH_SMALL
)) {
7314 /* Makes no sense without HASH_EARLY */
7315 WARN_ON(!(flags
& HASH_EARLY
));
7316 if (!(numentries
>> *_hash_shift
)) {
7317 numentries
= 1UL << *_hash_shift
;
7318 BUG_ON(!numentries
);
7320 } else if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
7321 numentries
= PAGE_SIZE
/ bucketsize
;
7323 numentries
= roundup_pow_of_two(numentries
);
7325 /* limit allocation size to 1/16 total memory by default */
7327 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
7328 do_div(max
, bucketsize
);
7330 max
= min(max
, 0x80000000ULL
);
7332 if (numentries
< low_limit
)
7333 numentries
= low_limit
;
7334 if (numentries
> max
)
7337 log2qty
= ilog2(numentries
);
7340 * memblock allocator returns zeroed memory already, so HASH_ZERO is
7341 * currently not used when HASH_EARLY is specified.
7343 gfp_flags
= (flags
& HASH_ZERO
) ? GFP_ATOMIC
| __GFP_ZERO
: GFP_ATOMIC
;
7345 size
= bucketsize
<< log2qty
;
7346 if (flags
& HASH_EARLY
)
7347 table
= memblock_virt_alloc_nopanic(size
, 0);
7349 table
= __vmalloc(size
, gfp_flags
, PAGE_KERNEL
);
7352 * If bucketsize is not a power-of-two, we may free
7353 * some pages at the end of hash table which
7354 * alloc_pages_exact() automatically does
7356 if (get_order(size
) < MAX_ORDER
) {
7357 table
= alloc_pages_exact(size
, gfp_flags
);
7358 kmemleak_alloc(table
, size
, 1, gfp_flags
);
7361 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
7364 panic("Failed to allocate %s hash table\n", tablename
);
7366 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
7367 tablename
, 1UL << log2qty
, ilog2(size
) - PAGE_SHIFT
, size
);
7370 *_hash_shift
= log2qty
;
7372 *_hash_mask
= (1 << log2qty
) - 1;
7378 * This function checks whether pageblock includes unmovable pages or not.
7379 * If @count is not zero, it is okay to include less @count unmovable pages
7381 * PageLRU check without isolation or lru_lock could race so that
7382 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
7383 * check without lock_page also may miss some movable non-lru pages at
7384 * race condition. So you can't expect this function should be exact.
7386 bool has_unmovable_pages(struct zone
*zone
, struct page
*page
, int count
,
7387 bool skip_hwpoisoned_pages
)
7389 unsigned long pfn
, iter
, found
;
7393 * For avoiding noise data, lru_add_drain_all() should be called
7394 * If ZONE_MOVABLE, the zone never contains unmovable pages
7396 if (zone_idx(zone
) == ZONE_MOVABLE
)
7398 mt
= get_pageblock_migratetype(page
);
7399 if (mt
== MIGRATE_MOVABLE
|| is_migrate_cma(mt
))
7402 pfn
= page_to_pfn(page
);
7403 for (found
= 0, iter
= 0; iter
< pageblock_nr_pages
; iter
++) {
7404 unsigned long check
= pfn
+ iter
;
7406 if (!pfn_valid_within(check
))
7409 page
= pfn_to_page(check
);
7412 * Hugepages are not in LRU lists, but they're movable.
7413 * We need not scan over tail pages bacause we don't
7414 * handle each tail page individually in migration.
7416 if (PageHuge(page
)) {
7417 iter
= round_up(iter
+ 1, 1<<compound_order(page
)) - 1;
7422 * We can't use page_count without pin a page
7423 * because another CPU can free compound page.
7424 * This check already skips compound tails of THP
7425 * because their page->_refcount is zero at all time.
7427 if (!page_ref_count(page
)) {
7428 if (PageBuddy(page
))
7429 iter
+= (1 << page_order(page
)) - 1;
7434 * The HWPoisoned page may be not in buddy system, and
7435 * page_count() is not 0.
7437 if (skip_hwpoisoned_pages
&& PageHWPoison(page
))
7440 if (__PageMovable(page
))
7446 * If there are RECLAIMABLE pages, we need to check
7447 * it. But now, memory offline itself doesn't call
7448 * shrink_node_slabs() and it still to be fixed.
7451 * If the page is not RAM, page_count()should be 0.
7452 * we don't need more check. This is an _used_ not-movable page.
7454 * The problematic thing here is PG_reserved pages. PG_reserved
7455 * is set to both of a memory hole page and a _used_ kernel
7464 bool is_pageblock_removable_nolock(struct page
*page
)
7470 * We have to be careful here because we are iterating over memory
7471 * sections which are not zone aware so we might end up outside of
7472 * the zone but still within the section.
7473 * We have to take care about the node as well. If the node is offline
7474 * its NODE_DATA will be NULL - see page_zone.
7476 if (!node_online(page_to_nid(page
)))
7479 zone
= page_zone(page
);
7480 pfn
= page_to_pfn(page
);
7481 if (!zone_spans_pfn(zone
, pfn
))
7484 return !has_unmovable_pages(zone
, page
, 0, true);
7487 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
7489 static unsigned long pfn_max_align_down(unsigned long pfn
)
7491 return pfn
& ~(max_t(unsigned long, MAX_ORDER_NR_PAGES
,
7492 pageblock_nr_pages
) - 1);
7495 static unsigned long pfn_max_align_up(unsigned long pfn
)
7497 return ALIGN(pfn
, max_t(unsigned long, MAX_ORDER_NR_PAGES
,
7498 pageblock_nr_pages
));
7501 /* [start, end) must belong to a single zone. */
7502 static int __alloc_contig_migrate_range(struct compact_control
*cc
,
7503 unsigned long start
, unsigned long end
,
7506 /* This function is based on compact_zone() from compaction.c. */
7507 unsigned long nr_reclaimed
;
7508 unsigned long pfn
= start
;
7509 unsigned int tries
= 0;
7515 while (pfn
< end
|| !list_empty(&cc
->migratepages
)) {
7516 if (fatal_signal_pending(current
)) {
7521 if (list_empty(&cc
->migratepages
)) {
7522 cc
->nr_migratepages
= 0;
7523 pfn
= isolate_migratepages_range(cc
, pfn
, end
);
7529 } else if (++tries
== 5) {
7530 ret
= ret
< 0 ? ret
: -EBUSY
;
7534 nr_reclaimed
= reclaim_clean_pages_from_list(cc
->zone
,
7536 cc
->nr_migratepages
-= nr_reclaimed
;
7538 ret
= migrate_pages(&cc
->migratepages
, alloc_migrate_target
,
7539 NULL
, 0, cc
->mode
, drain
? MR_CMA
: MR_HPA
);
7542 putback_movable_pages(&cc
->migratepages
);
7549 * alloc_contig_range() -- tries to allocate given range of pages
7550 * @start: start PFN to allocate
7551 * @end: one-past-the-last PFN to allocate
7552 * @migratetype: migratetype of the underlaying pageblocks (either
7553 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
7554 * in range must have the same migratetype and it must
7555 * be either of the two.
7556 * @gfp_mask: GFP mask to use during compaction
7558 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
7559 * aligned, however it's the caller's responsibility to guarantee that
7560 * we are the only thread that changes migrate type of pageblocks the
7563 * The PFN range must belong to a single zone.
7565 * Returns zero on success or negative error code. On success all
7566 * pages which PFN is in [start, end) are allocated for the caller and
7567 * need to be freed with free_contig_range().
7569 int __alloc_contig_range(unsigned long start
, unsigned long end
,
7570 unsigned migratetype
, gfp_t gfp_mask
, bool drain
)
7572 unsigned long outer_start
, outer_end
;
7576 struct compact_control cc
= {
7577 .nr_migratepages
= 0,
7579 .zone
= page_zone(pfn_to_page(start
)),
7580 .mode
= MIGRATE_SYNC
,
7581 .ignore_skip_hint
= true,
7582 .gfp_mask
= current_gfp_context(gfp_mask
),
7584 INIT_LIST_HEAD(&cc
.migratepages
);
7587 * What we do here is we mark all pageblocks in range as
7588 * MIGRATE_ISOLATE. Because pageblock and max order pages may
7589 * have different sizes, and due to the way page allocator
7590 * work, we align the range to biggest of the two pages so
7591 * that page allocator won't try to merge buddies from
7592 * different pageblocks and change MIGRATE_ISOLATE to some
7593 * other migration type.
7595 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
7596 * migrate the pages from an unaligned range (ie. pages that
7597 * we are interested in). This will put all the pages in
7598 * range back to page allocator as MIGRATE_ISOLATE.
7600 * When this is done, we take the pages in range from page
7601 * allocator removing them from the buddy system. This way
7602 * page allocator will never consider using them.
7604 * This lets us mark the pageblocks back as
7605 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
7606 * aligned range but not in the unaligned, original range are
7607 * put back to page allocator so that buddy can use them.
7610 ret
= start_isolate_page_range(pfn_max_align_down(start
),
7611 pfn_max_align_up(end
), migratetype
,
7617 * In case of -EBUSY, we'd like to know which page causes problem.
7618 * So, just fall through. test_pages_isolated() has a tracepoint
7619 * which will report the busy page.
7621 * It is possible that busy pages could become available before
7622 * the call to test_pages_isolated, and the range will actually be
7623 * allocated. So, if we fall through be sure to clear ret so that
7624 * -EBUSY is not accidentally used or returned to caller.
7626 ret
= __alloc_contig_migrate_range(&cc
, start
, end
, drain
);
7627 if (ret
&& ret
!= -EBUSY
)
7632 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
7633 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
7634 * more, all pages in [start, end) are free in page allocator.
7635 * What we are going to do is to allocate all pages from
7636 * [start, end) (that is remove them from page allocator).
7638 * The only problem is that pages at the beginning and at the
7639 * end of interesting range may be not aligned with pages that
7640 * page allocator holds, ie. they can be part of higher order
7641 * pages. Because of this, we reserve the bigger range and
7642 * once this is done free the pages we are not interested in.
7644 * We don't have to hold zone->lock here because the pages are
7645 * isolated thus they won't get removed from buddy.
7649 outer_start
= start
;
7652 lru_add_drain_all();
7653 drain_all_pages(cc
.zone
);
7655 while (!PageBuddy(pfn_to_page(outer_start
))) {
7656 if (++order
>= MAX_ORDER
) {
7657 outer_start
= start
;
7660 outer_start
&= ~0UL << order
;
7663 if (outer_start
!= start
) {
7664 order
= page_order(pfn_to_page(outer_start
));
7667 * outer_start page could be small order buddy page and
7668 * it doesn't include start page. Adjust outer_start
7669 * in this case to report failed page properly
7670 * on tracepoint in test_pages_isolated()
7672 if (outer_start
+ (1UL << order
) <= start
)
7673 outer_start
= start
;
7676 /* Make sure the range is really isolated. */
7677 if (test_pages_isolated(outer_start
, end
, false)) {
7678 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
7679 __func__
, outer_start
, end
);
7685 /* Grab isolated pages from freelists. */
7686 outer_end
= isolate_freepages_range(&cc
, outer_start
, end
);
7692 /* Free head and tail (if any) */
7693 if (start
!= outer_start
)
7694 free_contig_range(outer_start
, start
- outer_start
);
7695 if (end
!= outer_end
)
7696 free_contig_range(end
, outer_end
- end
);
7699 undo_isolate_page_range(pfn_max_align_down(start
),
7700 pfn_max_align_up(end
), migratetype
);
7704 int alloc_contig_range(unsigned long start
, unsigned long end
,
7705 unsigned migratetype
, gfp_t gfp_mask
)
7707 return __alloc_contig_range(start
, end
, migratetype
, gfp_mask
, true);
7710 int alloc_contig_range_fast(unsigned long start
, unsigned long end
,
7711 unsigned migratetype
)
7713 return __alloc_contig_range(start
, end
, migratetype
, GFP_KERNEL
, false);
7716 void free_contig_range(unsigned long pfn
, unsigned nr_pages
)
7718 unsigned int count
= 0;
7720 for (; nr_pages
--; pfn
++) {
7721 struct page
*page
= pfn_to_page(pfn
);
7723 count
+= page_count(page
) != 1;
7726 WARN(count
!= 0, "%d pages are still in use!\n", count
);
7730 #ifdef CONFIG_MEMORY_HOTPLUG
7732 * The zone indicated has a new number of managed_pages; batch sizes and percpu
7733 * page high values need to be recalulated.
7735 void __meminit
zone_pcp_update(struct zone
*zone
)
7738 mutex_lock(&pcp_batch_high_lock
);
7739 for_each_possible_cpu(cpu
)
7740 pageset_set_high_and_batch(zone
,
7741 per_cpu_ptr(zone
->pageset
, cpu
));
7742 mutex_unlock(&pcp_batch_high_lock
);
7746 void zone_pcp_reset(struct zone
*zone
)
7748 unsigned long flags
;
7750 struct per_cpu_pageset
*pset
;
7752 /* avoid races with drain_pages() */
7753 local_irq_save(flags
);
7754 if (zone
->pageset
!= &boot_pageset
) {
7755 for_each_online_cpu(cpu
) {
7756 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
7757 drain_zonestat(zone
, pset
);
7759 free_percpu(zone
->pageset
);
7760 zone
->pageset
= &boot_pageset
;
7762 local_irq_restore(flags
);
7765 #ifdef CONFIG_MEMORY_HOTREMOVE
7767 * All pages in the range must be in a single zone and isolated
7768 * before calling this.
7771 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
7775 unsigned int order
, i
;
7777 unsigned long flags
;
7778 /* find the first valid pfn */
7779 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
7784 offline_mem_sections(pfn
, end_pfn
);
7785 zone
= page_zone(pfn_to_page(pfn
));
7786 spin_lock_irqsave(&zone
->lock
, flags
);
7788 while (pfn
< end_pfn
) {
7789 if (!pfn_valid(pfn
)) {
7793 page
= pfn_to_page(pfn
);
7795 * The HWPoisoned page may be not in buddy system, and
7796 * page_count() is not 0.
7798 if (unlikely(!PageBuddy(page
) && PageHWPoison(page
))) {
7800 SetPageReserved(page
);
7804 BUG_ON(page_count(page
));
7805 BUG_ON(!PageBuddy(page
));
7806 order
= page_order(page
);
7807 #ifdef CONFIG_DEBUG_VM
7808 pr_info("remove from free list %lx %d %lx\n",
7809 pfn
, 1 << order
, end_pfn
);
7811 list_del(&page
->lru
);
7812 rmv_page_order(page
);
7813 zone
->free_area
[order
].nr_free
--;
7814 for (i
= 0; i
< (1 << order
); i
++)
7815 SetPageReserved((page
+i
));
7816 pfn
+= (1 << order
);
7818 spin_unlock_irqrestore(&zone
->lock
, flags
);
7822 bool is_free_buddy_page(struct page
*page
)
7824 struct zone
*zone
= page_zone(page
);
7825 unsigned long pfn
= page_to_pfn(page
);
7826 unsigned long flags
;
7829 spin_lock_irqsave(&zone
->lock
, flags
);
7830 for (order
= 0; order
< MAX_ORDER
; order
++) {
7831 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
7833 if (PageBuddy(page_head
) && page_order(page_head
) >= order
)
7836 spin_unlock_irqrestore(&zone
->lock
, flags
);
7838 return order
< MAX_ORDER
;