2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kmemcheck.h>
28 #include <linux/kasan.h>
29 #include <linux/module.h>
30 #include <linux/suspend.h>
31 #include <linux/pagevec.h>
32 #include <linux/blkdev.h>
33 #include <linux/slab.h>
34 #include <linux/ratelimit.h>
35 #include <linux/oom.h>
36 #include <linux/notifier.h>
37 #include <linux/topology.h>
38 #include <linux/sysctl.h>
39 #include <linux/cpu.h>
40 #include <linux/cpuset.h>
41 #include <linux/memory_hotplug.h>
42 #include <linux/nodemask.h>
43 #include <linux/vmalloc.h>
44 #include <linux/vmstat.h>
45 #include <linux/mempolicy.h>
46 #include <linux/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 <linux/prefetch.h>
58 #include <linux/mm_inline.h>
59 #include <linux/migrate.h>
60 #include <linux/page_ext.h>
61 #include <linux/hugetlb.h>
62 #include <linux/sched/rt.h>
63 #include <linux/page_owner.h>
65 #include <asm/sections.h>
66 #include <asm/tlbflush.h>
67 #include <asm/div64.h>
70 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
71 static DEFINE_MUTEX(pcp_batch_high_lock
);
72 #define MIN_PERCPU_PAGELIST_FRACTION (8)
74 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
75 DEFINE_PER_CPU(int, numa_node
);
76 EXPORT_PER_CPU_SYMBOL(numa_node
);
79 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
81 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
82 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
83 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
84 * defined in <linux/topology.h>.
86 DEFINE_PER_CPU(int, _numa_mem_
); /* Kernel "local memory" node */
87 EXPORT_PER_CPU_SYMBOL(_numa_mem_
);
88 int _node_numa_mem_
[MAX_NUMNODES
];
92 * Array of node states.
94 nodemask_t node_states
[NR_NODE_STATES
] __read_mostly
= {
95 [N_POSSIBLE
] = NODE_MASK_ALL
,
96 [N_ONLINE
] = { { [0] = 1UL } },
98 [N_NORMAL_MEMORY
] = { { [0] = 1UL } },
100 [N_HIGH_MEMORY
] = { { [0] = 1UL } },
102 #ifdef CONFIG_MOVABLE_NODE
103 [N_MEMORY
] = { { [0] = 1UL } },
105 [N_CPU
] = { { [0] = 1UL } },
108 EXPORT_SYMBOL(node_states
);
110 /* Protect totalram_pages and zone->managed_pages */
111 static DEFINE_SPINLOCK(managed_page_count_lock
);
113 unsigned long totalram_pages __read_mostly
;
114 unsigned long totalreserve_pages __read_mostly
;
115 unsigned long totalcma_pages __read_mostly
;
117 * When calculating the number of globally allowed dirty pages, there
118 * is a certain number of per-zone reserves that should not be
119 * considered dirtyable memory. This is the sum of those reserves
120 * over all existing zones that contribute dirtyable memory.
122 unsigned long dirty_balance_reserve __read_mostly
;
124 int percpu_pagelist_fraction
;
125 gfp_t gfp_allowed_mask __read_mostly
= GFP_BOOT_MASK
;
127 #ifdef CONFIG_PM_SLEEP
129 * The following functions are used by the suspend/hibernate code to temporarily
130 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
131 * while devices are suspended. To avoid races with the suspend/hibernate code,
132 * they should always be called with pm_mutex held (gfp_allowed_mask also should
133 * only be modified with pm_mutex held, unless the suspend/hibernate code is
134 * guaranteed not to run in parallel with that modification).
137 static gfp_t saved_gfp_mask
;
139 void pm_restore_gfp_mask(void)
141 WARN_ON(!mutex_is_locked(&pm_mutex
));
142 if (saved_gfp_mask
) {
143 gfp_allowed_mask
= saved_gfp_mask
;
148 void pm_restrict_gfp_mask(void)
150 WARN_ON(!mutex_is_locked(&pm_mutex
));
151 WARN_ON(saved_gfp_mask
);
152 saved_gfp_mask
= gfp_allowed_mask
;
153 gfp_allowed_mask
&= ~GFP_IOFS
;
156 bool pm_suspended_storage(void)
158 if ((gfp_allowed_mask
& GFP_IOFS
) == GFP_IOFS
)
162 #endif /* CONFIG_PM_SLEEP */
164 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
165 int pageblock_order __read_mostly
;
168 static void __free_pages_ok(struct page
*page
, unsigned int order
);
171 * results with 256, 32 in the lowmem_reserve sysctl:
172 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
173 * 1G machine -> (16M dma, 784M normal, 224M high)
174 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
175 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
176 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
178 * TBD: should special case ZONE_DMA32 machines here - in those we normally
179 * don't need any ZONE_NORMAL reservation
181 int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
-1] = {
182 #ifdef CONFIG_ZONE_DMA
185 #ifdef CONFIG_ZONE_DMA32
188 #ifdef CONFIG_HIGHMEM
194 EXPORT_SYMBOL(totalram_pages
);
196 static char * const zone_names
[MAX_NR_ZONES
] = {
197 #ifdef CONFIG_ZONE_DMA
200 #ifdef CONFIG_ZONE_DMA32
204 #ifdef CONFIG_HIGHMEM
210 int min_free_kbytes
= 1024;
211 int user_min_free_kbytes
= -1;
213 static unsigned long __meminitdata nr_kernel_pages
;
214 static unsigned long __meminitdata nr_all_pages
;
215 static unsigned long __meminitdata dma_reserve
;
217 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
218 static unsigned long __meminitdata arch_zone_lowest_possible_pfn
[MAX_NR_ZONES
];
219 static unsigned long __meminitdata arch_zone_highest_possible_pfn
[MAX_NR_ZONES
];
220 static unsigned long __initdata required_kernelcore
;
221 static unsigned long __initdata required_movablecore
;
222 static unsigned long __meminitdata zone_movable_pfn
[MAX_NUMNODES
];
224 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
226 EXPORT_SYMBOL(movable_zone
);
227 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
230 int nr_node_ids __read_mostly
= MAX_NUMNODES
;
231 int nr_online_nodes __read_mostly
= 1;
232 EXPORT_SYMBOL(nr_node_ids
);
233 EXPORT_SYMBOL(nr_online_nodes
);
236 int page_group_by_mobility_disabled __read_mostly
;
238 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
239 static inline void reset_deferred_meminit(pg_data_t
*pgdat
)
241 pgdat
->first_deferred_pfn
= ULONG_MAX
;
244 /* Returns true if the struct page for the pfn is uninitialised */
245 static inline bool __defermem_init
early_page_uninitialised(unsigned long pfn
)
247 int nid
= early_pfn_to_nid(pfn
);
249 if (pfn
>= NODE_DATA(nid
)->first_deferred_pfn
)
255 static inline bool early_page_nid_uninitialised(unsigned long pfn
, int nid
)
257 if (pfn
>= NODE_DATA(nid
)->first_deferred_pfn
)
264 * Returns false when the remaining initialisation should be deferred until
265 * later in the boot cycle when it can be parallelised.
267 static inline bool update_defer_init(pg_data_t
*pgdat
,
268 unsigned long pfn
, unsigned long zone_end
,
269 unsigned long *nr_initialised
)
271 /* Always populate low zones for address-contrained allocations */
272 if (zone_end
< pgdat_end_pfn(pgdat
))
275 /* Initialise at least 2G of the highest zone */
277 if (*nr_initialised
> (2UL << (30 - PAGE_SHIFT
)) &&
278 (pfn
& (PAGES_PER_SECTION
- 1)) == 0) {
279 pgdat
->first_deferred_pfn
= pfn
;
286 static inline void reset_deferred_meminit(pg_data_t
*pgdat
)
290 static inline bool early_page_uninitialised(unsigned long pfn
)
295 static inline bool early_page_nid_uninitialised(unsigned long pfn
, int nid
)
300 static inline bool update_defer_init(pg_data_t
*pgdat
,
301 unsigned long pfn
, unsigned long zone_end
,
302 unsigned long *nr_initialised
)
309 void set_pageblock_migratetype(struct page
*page
, int migratetype
)
311 if (unlikely(page_group_by_mobility_disabled
&&
312 migratetype
< MIGRATE_PCPTYPES
))
313 migratetype
= MIGRATE_UNMOVABLE
;
315 set_pageblock_flags_group(page
, (unsigned long)migratetype
,
316 PB_migrate
, PB_migrate_end
);
319 #ifdef CONFIG_DEBUG_VM
320 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
324 unsigned long pfn
= page_to_pfn(page
);
325 unsigned long sp
, start_pfn
;
328 seq
= zone_span_seqbegin(zone
);
329 start_pfn
= zone
->zone_start_pfn
;
330 sp
= zone
->spanned_pages
;
331 if (!zone_spans_pfn(zone
, pfn
))
333 } while (zone_span_seqretry(zone
, seq
));
336 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
337 pfn
, zone_to_nid(zone
), zone
->name
,
338 start_pfn
, start_pfn
+ sp
);
343 static int page_is_consistent(struct zone
*zone
, struct page
*page
)
345 if (!pfn_valid_within(page_to_pfn(page
)))
347 if (zone
!= page_zone(page
))
353 * Temporary debugging check for pages not lying within a given zone.
355 static int bad_range(struct zone
*zone
, struct page
*page
)
357 if (page_outside_zone_boundaries(zone
, page
))
359 if (!page_is_consistent(zone
, page
))
365 static inline int bad_range(struct zone
*zone
, struct page
*page
)
371 static void bad_page(struct page
*page
, const char *reason
,
372 unsigned long bad_flags
)
374 static unsigned long resume
;
375 static unsigned long nr_shown
;
376 static unsigned long nr_unshown
;
378 /* Don't complain about poisoned pages */
379 if (PageHWPoison(page
)) {
380 page_mapcount_reset(page
); /* remove PageBuddy */
385 * Allow a burst of 60 reports, then keep quiet for that minute;
386 * or allow a steady drip of one report per second.
388 if (nr_shown
== 60) {
389 if (time_before(jiffies
, resume
)) {
395 "BUG: Bad page state: %lu messages suppressed\n",
402 resume
= jiffies
+ 60 * HZ
;
404 printk(KERN_ALERT
"BUG: Bad page state in process %s pfn:%05lx\n",
405 current
->comm
, page_to_pfn(page
));
406 dump_page_badflags(page
, reason
, bad_flags
);
411 /* Leave bad fields for debug, except PageBuddy could make trouble */
412 page_mapcount_reset(page
); /* remove PageBuddy */
413 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
417 * Higher-order pages are called "compound pages". They are structured thusly:
419 * The first PAGE_SIZE page is called the "head page".
421 * The remaining PAGE_SIZE pages are called "tail pages".
423 * All pages have PG_compound set. All tail pages have their ->first_page
424 * pointing at the head page.
426 * The first tail page's ->lru.next holds the address of the compound page's
427 * put_page() function. Its ->lru.prev holds the order of allocation.
428 * This usage means that zero-order pages may not be compound.
431 static void free_compound_page(struct page
*page
)
433 __free_pages_ok(page
, compound_order(page
));
436 void prep_compound_page(struct page
*page
, unsigned long order
)
439 int nr_pages
= 1 << order
;
441 set_compound_page_dtor(page
, free_compound_page
);
442 set_compound_order(page
, order
);
444 for (i
= 1; i
< nr_pages
; i
++) {
445 struct page
*p
= page
+ i
;
446 set_page_count(p
, 0);
447 p
->first_page
= page
;
448 /* Make sure p->first_page is always valid for PageTail() */
454 #ifdef CONFIG_DEBUG_PAGEALLOC
455 unsigned int _debug_guardpage_minorder
;
456 bool _debug_pagealloc_enabled __read_mostly
;
457 bool _debug_guardpage_enabled __read_mostly
;
459 static int __init
early_debug_pagealloc(char *buf
)
464 if (strcmp(buf
, "on") == 0)
465 _debug_pagealloc_enabled
= true;
469 early_param("debug_pagealloc", early_debug_pagealloc
);
471 static bool need_debug_guardpage(void)
473 /* If we don't use debug_pagealloc, we don't need guard page */
474 if (!debug_pagealloc_enabled())
480 static void init_debug_guardpage(void)
482 if (!debug_pagealloc_enabled())
485 _debug_guardpage_enabled
= true;
488 struct page_ext_operations debug_guardpage_ops
= {
489 .need
= need_debug_guardpage
,
490 .init
= init_debug_guardpage
,
493 static int __init
debug_guardpage_minorder_setup(char *buf
)
497 if (kstrtoul(buf
, 10, &res
) < 0 || res
> MAX_ORDER
/ 2) {
498 printk(KERN_ERR
"Bad debug_guardpage_minorder value\n");
501 _debug_guardpage_minorder
= res
;
502 printk(KERN_INFO
"Setting debug_guardpage_minorder to %lu\n", res
);
505 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup
);
507 static inline void set_page_guard(struct zone
*zone
, struct page
*page
,
508 unsigned int order
, int migratetype
)
510 struct page_ext
*page_ext
;
512 if (!debug_guardpage_enabled())
515 page_ext
= lookup_page_ext(page
);
516 __set_bit(PAGE_EXT_DEBUG_GUARD
, &page_ext
->flags
);
518 INIT_LIST_HEAD(&page
->lru
);
519 set_page_private(page
, order
);
520 /* Guard pages are not available for any usage */
521 __mod_zone_freepage_state(zone
, -(1 << order
), migratetype
);
524 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
525 unsigned int order
, int migratetype
)
527 struct page_ext
*page_ext
;
529 if (!debug_guardpage_enabled())
532 page_ext
= lookup_page_ext(page
);
533 __clear_bit(PAGE_EXT_DEBUG_GUARD
, &page_ext
->flags
);
535 set_page_private(page
, 0);
536 if (!is_migrate_isolate(migratetype
))
537 __mod_zone_freepage_state(zone
, (1 << order
), migratetype
);
540 struct page_ext_operations debug_guardpage_ops
= { NULL
, };
541 static inline void set_page_guard(struct zone
*zone
, struct page
*page
,
542 unsigned int order
, int migratetype
) {}
543 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
544 unsigned int order
, int migratetype
) {}
547 static inline void set_page_order(struct page
*page
, unsigned int order
)
549 set_page_private(page
, order
);
550 __SetPageBuddy(page
);
553 static inline void rmv_page_order(struct page
*page
)
555 __ClearPageBuddy(page
);
556 set_page_private(page
, 0);
560 * This function checks whether a page is free && is the buddy
561 * we can do coalesce a page and its buddy if
562 * (a) the buddy is not in a hole &&
563 * (b) the buddy is in the buddy system &&
564 * (c) a page and its buddy have the same order &&
565 * (d) a page and its buddy are in the same zone.
567 * For recording whether a page is in the buddy system, we set ->_mapcount
568 * PAGE_BUDDY_MAPCOUNT_VALUE.
569 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
570 * serialized by zone->lock.
572 * For recording page's order, we use page_private(page).
574 static inline int page_is_buddy(struct page
*page
, struct page
*buddy
,
577 if (!pfn_valid_within(page_to_pfn(buddy
)))
580 if (page_is_guard(buddy
) && page_order(buddy
) == order
) {
581 if (page_zone_id(page
) != page_zone_id(buddy
))
584 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
589 if (PageBuddy(buddy
) && page_order(buddy
) == order
) {
591 * zone check is done late to avoid uselessly
592 * calculating zone/node ids for pages that could
595 if (page_zone_id(page
) != page_zone_id(buddy
))
598 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
606 * Freeing function for a buddy system allocator.
608 * The concept of a buddy system is to maintain direct-mapped table
609 * (containing bit values) for memory blocks of various "orders".
610 * The bottom level table contains the map for the smallest allocatable
611 * units of memory (here, pages), and each level above it describes
612 * pairs of units from the levels below, hence, "buddies".
613 * At a high level, all that happens here is marking the table entry
614 * at the bottom level available, and propagating the changes upward
615 * as necessary, plus some accounting needed to play nicely with other
616 * parts of the VM system.
617 * At each level, we keep a list of pages, which are heads of continuous
618 * free pages of length of (1 << order) and marked with _mapcount
619 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
621 * So when we are allocating or freeing one, we can derive the state of the
622 * other. That is, if we allocate a small block, and both were
623 * free, the remainder of the region must be split into blocks.
624 * If a block is freed, and its buddy is also free, then this
625 * triggers coalescing into a block of larger size.
630 static inline void __free_one_page(struct page
*page
,
632 struct zone
*zone
, unsigned int order
,
635 unsigned long page_idx
;
636 unsigned long combined_idx
;
637 unsigned long uninitialized_var(buddy_idx
);
639 int max_order
= MAX_ORDER
;
641 VM_BUG_ON(!zone_is_initialized(zone
));
642 VM_BUG_ON_PAGE(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
, page
);
644 VM_BUG_ON(migratetype
== -1);
645 if (is_migrate_isolate(migratetype
)) {
647 * We restrict max order of merging to prevent merge
648 * between freepages on isolate pageblock and normal
649 * pageblock. Without this, pageblock isolation
650 * could cause incorrect freepage accounting.
652 max_order
= min(MAX_ORDER
, pageblock_order
+ 1);
654 __mod_zone_freepage_state(zone
, 1 << order
, migratetype
);
657 page_idx
= pfn
& ((1 << max_order
) - 1);
659 VM_BUG_ON_PAGE(page_idx
& ((1 << order
) - 1), page
);
660 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
662 while (order
< max_order
- 1) {
663 buddy_idx
= __find_buddy_index(page_idx
, order
);
664 buddy
= page
+ (buddy_idx
- page_idx
);
665 if (!page_is_buddy(page
, buddy
, order
))
668 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
669 * merge with it and move up one order.
671 if (page_is_guard(buddy
)) {
672 clear_page_guard(zone
, buddy
, order
, migratetype
);
674 list_del(&buddy
->lru
);
675 zone
->free_area
[order
].nr_free
--;
676 rmv_page_order(buddy
);
678 combined_idx
= buddy_idx
& page_idx
;
679 page
= page
+ (combined_idx
- page_idx
);
680 page_idx
= combined_idx
;
683 set_page_order(page
, order
);
686 * If this is not the largest possible page, check if the buddy
687 * of the next-highest order is free. If it is, it's possible
688 * that pages are being freed that will coalesce soon. In case,
689 * that is happening, add the free page to the tail of the list
690 * so it's less likely to be used soon and more likely to be merged
691 * as a higher order page
693 if ((order
< MAX_ORDER
-2) && pfn_valid_within(page_to_pfn(buddy
))) {
694 struct page
*higher_page
, *higher_buddy
;
695 combined_idx
= buddy_idx
& page_idx
;
696 higher_page
= page
+ (combined_idx
- page_idx
);
697 buddy_idx
= __find_buddy_index(combined_idx
, order
+ 1);
698 higher_buddy
= higher_page
+ (buddy_idx
- combined_idx
);
699 if (page_is_buddy(higher_page
, higher_buddy
, order
+ 1)) {
700 list_add_tail(&page
->lru
,
701 &zone
->free_area
[order
].free_list
[migratetype
]);
706 list_add(&page
->lru
, &zone
->free_area
[order
].free_list
[migratetype
]);
708 zone
->free_area
[order
].nr_free
++;
711 static inline int free_pages_check(struct page
*page
)
713 const char *bad_reason
= NULL
;
714 unsigned long bad_flags
= 0;
716 if (unlikely(page_mapcount(page
)))
717 bad_reason
= "nonzero mapcount";
718 if (unlikely(page
->mapping
!= NULL
))
719 bad_reason
= "non-NULL mapping";
720 if (unlikely(atomic_read(&page
->_count
) != 0))
721 bad_reason
= "nonzero _count";
722 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_FREE
)) {
723 bad_reason
= "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
724 bad_flags
= PAGE_FLAGS_CHECK_AT_FREE
;
727 if (unlikely(page
->mem_cgroup
))
728 bad_reason
= "page still charged to cgroup";
730 if (unlikely(bad_reason
)) {
731 bad_page(page
, bad_reason
, bad_flags
);
734 page_cpupid_reset_last(page
);
735 if (page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)
736 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
741 * Frees a number of pages from the PCP lists
742 * Assumes all pages on list are in same zone, and of same order.
743 * count is the number of pages to free.
745 * If the zone was previously in an "all pages pinned" state then look to
746 * see if this freeing clears that state.
748 * And clear the zone's pages_scanned counter, to hold off the "all pages are
749 * pinned" detection logic.
751 static void free_pcppages_bulk(struct zone
*zone
, int count
,
752 struct per_cpu_pages
*pcp
)
757 unsigned long nr_scanned
;
759 spin_lock(&zone
->lock
);
760 nr_scanned
= zone_page_state(zone
, NR_PAGES_SCANNED
);
762 __mod_zone_page_state(zone
, NR_PAGES_SCANNED
, -nr_scanned
);
766 struct list_head
*list
;
769 * Remove pages from lists in a round-robin fashion. A
770 * batch_free count is maintained that is incremented when an
771 * empty list is encountered. This is so more pages are freed
772 * off fuller lists instead of spinning excessively around empty
777 if (++migratetype
== MIGRATE_PCPTYPES
)
779 list
= &pcp
->lists
[migratetype
];
780 } while (list_empty(list
));
782 /* This is the only non-empty list. Free them all. */
783 if (batch_free
== MIGRATE_PCPTYPES
)
784 batch_free
= to_free
;
787 int mt
; /* migratetype of the to-be-freed page */
789 page
= list_entry(list
->prev
, struct page
, lru
);
790 /* must delete as __free_one_page list manipulates */
791 list_del(&page
->lru
);
792 mt
= get_freepage_migratetype(page
);
793 if (unlikely(has_isolate_pageblock(zone
)))
794 mt
= get_pageblock_migratetype(page
);
796 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
797 __free_one_page(page
, page_to_pfn(page
), zone
, 0, mt
);
798 trace_mm_page_pcpu_drain(page
, 0, mt
);
799 } while (--to_free
&& --batch_free
&& !list_empty(list
));
801 spin_unlock(&zone
->lock
);
804 static void free_one_page(struct zone
*zone
,
805 struct page
*page
, unsigned long pfn
,
809 unsigned long nr_scanned
;
810 spin_lock(&zone
->lock
);
811 nr_scanned
= zone_page_state(zone
, NR_PAGES_SCANNED
);
813 __mod_zone_page_state(zone
, NR_PAGES_SCANNED
, -nr_scanned
);
815 if (unlikely(has_isolate_pageblock(zone
) ||
816 is_migrate_isolate(migratetype
))) {
817 migratetype
= get_pfnblock_migratetype(page
, pfn
);
819 __free_one_page(page
, pfn
, zone
, order
, migratetype
);
820 spin_unlock(&zone
->lock
);
823 static int free_tail_pages_check(struct page
*head_page
, struct page
*page
)
825 if (!IS_ENABLED(CONFIG_DEBUG_VM
))
827 if (unlikely(!PageTail(page
))) {
828 bad_page(page
, "PageTail not set", 0);
831 if (unlikely(page
->first_page
!= head_page
)) {
832 bad_page(page
, "first_page not consistent", 0);
838 static void __meminit
__init_single_page(struct page
*page
, unsigned long pfn
,
839 unsigned long zone
, int nid
)
841 set_page_links(page
, zone
, nid
, pfn
);
842 mminit_verify_page_links(page
, zone
, nid
, pfn
);
843 init_page_count(page
);
844 page_mapcount_reset(page
);
845 page_cpupid_reset_last(page
);
847 INIT_LIST_HEAD(&page
->lru
);
848 #ifdef WANT_PAGE_VIRTUAL
849 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
850 if (!is_highmem_idx(zone
))
851 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
855 static void __meminit
__init_single_pfn(unsigned long pfn
, unsigned long zone
,
858 return __init_single_page(pfn_to_page(pfn
), pfn
, zone
, nid
);
861 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
862 static void init_reserved_page(unsigned long pfn
)
867 if (!early_page_uninitialised(pfn
))
870 nid
= early_pfn_to_nid(pfn
);
871 pgdat
= NODE_DATA(nid
);
873 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
874 struct zone
*zone
= &pgdat
->node_zones
[zid
];
876 if (pfn
>= zone
->zone_start_pfn
&& pfn
< zone_end_pfn(zone
))
879 __init_single_pfn(pfn
, zid
, nid
);
882 static inline void init_reserved_page(unsigned long pfn
)
885 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
888 * Initialised pages do not have PageReserved set. This function is
889 * called for each range allocated by the bootmem allocator and
890 * marks the pages PageReserved. The remaining valid pages are later
891 * sent to the buddy page allocator.
893 void __meminit
reserve_bootmem_region(unsigned long start
, unsigned long end
)
895 unsigned long start_pfn
= PFN_DOWN(start
);
896 unsigned long end_pfn
= PFN_UP(end
);
898 for (; start_pfn
< end_pfn
; start_pfn
++) {
899 if (pfn_valid(start_pfn
)) {
900 struct page
*page
= pfn_to_page(start_pfn
);
902 init_reserved_page(start_pfn
);
903 SetPageReserved(page
);
908 static bool free_pages_prepare(struct page
*page
, unsigned int order
)
910 bool compound
= PageCompound(page
);
913 VM_BUG_ON_PAGE(PageTail(page
), page
);
914 VM_BUG_ON_PAGE(compound
&& compound_order(page
) != order
, page
);
916 trace_mm_page_free(page
, order
);
917 kmemcheck_free_shadow(page
, order
);
918 kasan_free_pages(page
, order
);
921 page
->mapping
= NULL
;
922 bad
+= free_pages_check(page
);
923 for (i
= 1; i
< (1 << order
); i
++) {
925 bad
+= free_tail_pages_check(page
, page
+ i
);
926 bad
+= free_pages_check(page
+ i
);
931 reset_page_owner(page
, order
);
933 if (!PageHighMem(page
)) {
934 debug_check_no_locks_freed(page_address(page
),
936 debug_check_no_obj_freed(page_address(page
),
939 arch_free_page(page
, order
);
940 kernel_map_pages(page
, 1 << order
, 0);
945 static void __free_pages_ok(struct page
*page
, unsigned int order
)
949 unsigned long pfn
= page_to_pfn(page
);
951 if (!free_pages_prepare(page
, order
))
954 migratetype
= get_pfnblock_migratetype(page
, pfn
);
955 local_irq_save(flags
);
956 __count_vm_events(PGFREE
, 1 << order
);
957 set_freepage_migratetype(page
, migratetype
);
958 free_one_page(page_zone(page
), page
, pfn
, order
, migratetype
);
959 local_irq_restore(flags
);
962 static void __defer_init
__free_pages_boot_core(struct page
*page
,
963 unsigned long pfn
, unsigned int order
)
965 unsigned int nr_pages
= 1 << order
;
966 struct page
*p
= page
;
970 for (loop
= 0; loop
< (nr_pages
- 1); loop
++, p
++) {
972 __ClearPageReserved(p
);
973 set_page_count(p
, 0);
975 __ClearPageReserved(p
);
976 set_page_count(p
, 0);
978 page_zone(page
)->managed_pages
+= nr_pages
;
979 set_page_refcounted(page
);
980 __free_pages(page
, order
);
983 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
984 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
985 /* Only safe to use early in boot when initialisation is single-threaded */
986 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata
;
988 int __meminit
early_pfn_to_nid(unsigned long pfn
)
992 /* The system will behave unpredictably otherwise */
993 BUG_ON(system_state
!= SYSTEM_BOOTING
);
995 nid
= __early_pfn_to_nid(pfn
, &early_pfnnid_cache
);
1003 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1004 static inline bool __meminit
meminit_pfn_in_nid(unsigned long pfn
, int node
,
1005 struct mminit_pfnnid_cache
*state
)
1009 nid
= __early_pfn_to_nid(pfn
, state
);
1010 if (nid
>= 0 && nid
!= node
)
1015 /* Only safe to use early in boot when initialisation is single-threaded */
1016 static inline bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
1018 return meminit_pfn_in_nid(pfn
, node
, &early_pfnnid_cache
);
1023 static inline bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
1027 static inline bool __meminit
meminit_pfn_in_nid(unsigned long pfn
, int node
,
1028 struct mminit_pfnnid_cache
*state
)
1035 void __defer_init
__free_pages_bootmem(struct page
*page
, unsigned long pfn
,
1038 if (early_page_uninitialised(pfn
))
1040 return __free_pages_boot_core(page
, pfn
, order
);
1043 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1044 static void __defermem_init
deferred_free_range(struct page
*page
,
1045 unsigned long pfn
, int nr_pages
)
1052 /* Free a large naturally-aligned chunk if possible */
1053 if (nr_pages
== MAX_ORDER_NR_PAGES
&&
1054 (pfn
& (MAX_ORDER_NR_PAGES
-1)) == 0) {
1055 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
1056 __free_pages_boot_core(page
, pfn
, MAX_ORDER
-1);
1060 for (i
= 0; i
< nr_pages
; i
++, page
++, pfn
++)
1061 __free_pages_boot_core(page
, pfn
, 0);
1064 /* Initialise remaining memory on a node */
1065 void __defermem_init
deferred_init_memmap(int nid
)
1067 struct mminit_pfnnid_cache nid_init_state
= { };
1068 unsigned long start
= jiffies
;
1069 unsigned long nr_pages
= 0;
1070 unsigned long walk_start
, walk_end
;
1073 pg_data_t
*pgdat
= NODE_DATA(nid
);
1074 unsigned long first_init_pfn
= pgdat
->first_deferred_pfn
;
1076 if (first_init_pfn
== ULONG_MAX
)
1079 /* Sanity check boundaries */
1080 BUG_ON(pgdat
->first_deferred_pfn
< pgdat
->node_start_pfn
);
1081 BUG_ON(pgdat
->first_deferred_pfn
> pgdat_end_pfn(pgdat
));
1082 pgdat
->first_deferred_pfn
= ULONG_MAX
;
1084 /* Only the highest zone is deferred so find it */
1085 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1086 zone
= pgdat
->node_zones
+ zid
;
1087 if (first_init_pfn
< zone_end_pfn(zone
))
1091 for_each_mem_pfn_range(i
, nid
, &walk_start
, &walk_end
, NULL
) {
1092 unsigned long pfn
, end_pfn
;
1093 struct page
*page
= NULL
;
1094 struct page
*free_base_page
= NULL
;
1095 unsigned long free_base_pfn
= 0;
1098 end_pfn
= min(walk_end
, zone_end_pfn(zone
));
1099 pfn
= first_init_pfn
;
1100 if (pfn
< walk_start
)
1102 if (pfn
< zone
->zone_start_pfn
)
1103 pfn
= zone
->zone_start_pfn
;
1105 for (; pfn
< end_pfn
; pfn
++) {
1106 if (!pfn_valid_within(pfn
))
1110 * Ensure pfn_valid is checked every
1111 * MAX_ORDER_NR_PAGES for memory holes
1113 if ((pfn
& (MAX_ORDER_NR_PAGES
- 1)) == 0) {
1114 if (!pfn_valid(pfn
)) {
1120 if (!meminit_pfn_in_nid(pfn
, nid
, &nid_init_state
)) {
1125 /* Minimise pfn page lookups and scheduler checks */
1126 if (page
&& (pfn
& (MAX_ORDER_NR_PAGES
- 1)) != 0) {
1129 nr_pages
+= nr_to_free
;
1130 deferred_free_range(free_base_page
,
1131 free_base_pfn
, nr_to_free
);
1132 free_base_page
= NULL
;
1133 free_base_pfn
= nr_to_free
= 0;
1135 page
= pfn_to_page(pfn
);
1140 VM_BUG_ON(page_zone(page
) != zone
);
1144 __init_single_page(page
, pfn
, zid
, nid
);
1145 if (!free_base_page
) {
1146 free_base_page
= page
;
1147 free_base_pfn
= pfn
;
1152 /* Where possible, batch up pages for a single free */
1155 /* Free the current block of pages to allocator */
1156 nr_pages
+= nr_to_free
;
1157 deferred_free_range(free_base_page
, free_base_pfn
,
1159 free_base_page
= NULL
;
1160 free_base_pfn
= nr_to_free
= 0;
1163 first_init_pfn
= max(end_pfn
, first_init_pfn
);
1166 /* Sanity check that the next zone really is unpopulated */
1167 WARN_ON(++zid
< MAX_NR_ZONES
&& populated_zone(++zone
));
1169 pr_info("kswapd %d initialised %lu pages in %ums\n", nid
, nr_pages
,
1170 jiffies_to_msecs(jiffies
- start
));
1172 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1175 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1176 void __init
init_cma_reserved_pageblock(struct page
*page
)
1178 unsigned i
= pageblock_nr_pages
;
1179 struct page
*p
= page
;
1182 __ClearPageReserved(p
);
1183 set_page_count(p
, 0);
1186 set_pageblock_migratetype(page
, MIGRATE_CMA
);
1188 if (pageblock_order
>= MAX_ORDER
) {
1189 i
= pageblock_nr_pages
;
1192 set_page_refcounted(p
);
1193 __free_pages(p
, MAX_ORDER
- 1);
1194 p
+= MAX_ORDER_NR_PAGES
;
1195 } while (i
-= MAX_ORDER_NR_PAGES
);
1197 set_page_refcounted(page
);
1198 __free_pages(page
, pageblock_order
);
1201 adjust_managed_page_count(page
, pageblock_nr_pages
);
1206 * The order of subdivision here is critical for the IO subsystem.
1207 * Please do not alter this order without good reasons and regression
1208 * testing. Specifically, as large blocks of memory are subdivided,
1209 * the order in which smaller blocks are delivered depends on the order
1210 * they're subdivided in this function. This is the primary factor
1211 * influencing the order in which pages are delivered to the IO
1212 * subsystem according to empirical testing, and this is also justified
1213 * by considering the behavior of a buddy system containing a single
1214 * large block of memory acted on by a series of small allocations.
1215 * This behavior is a critical factor in sglist merging's success.
1219 static inline void expand(struct zone
*zone
, struct page
*page
,
1220 int low
, int high
, struct free_area
*area
,
1223 unsigned long size
= 1 << high
;
1225 while (high
> low
) {
1229 VM_BUG_ON_PAGE(bad_range(zone
, &page
[size
]), &page
[size
]);
1231 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC
) &&
1232 debug_guardpage_enabled() &&
1233 high
< debug_guardpage_minorder()) {
1235 * Mark as guard pages (or page), that will allow to
1236 * merge back to allocator when buddy will be freed.
1237 * Corresponding page table entries will not be touched,
1238 * pages will stay not present in virtual address space
1240 set_page_guard(zone
, &page
[size
], high
, migratetype
);
1243 list_add(&page
[size
].lru
, &area
->free_list
[migratetype
]);
1245 set_page_order(&page
[size
], high
);
1250 * This page is about to be returned from the page allocator
1252 static inline int check_new_page(struct page
*page
)
1254 const char *bad_reason
= NULL
;
1255 unsigned long bad_flags
= 0;
1257 if (unlikely(page_mapcount(page
)))
1258 bad_reason
= "nonzero mapcount";
1259 if (unlikely(page
->mapping
!= NULL
))
1260 bad_reason
= "non-NULL mapping";
1261 if (unlikely(atomic_read(&page
->_count
) != 0))
1262 bad_reason
= "nonzero _count";
1263 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)) {
1264 bad_reason
= "PAGE_FLAGS_CHECK_AT_PREP flag set";
1265 bad_flags
= PAGE_FLAGS_CHECK_AT_PREP
;
1268 if (unlikely(page
->mem_cgroup
))
1269 bad_reason
= "page still charged to cgroup";
1271 if (unlikely(bad_reason
)) {
1272 bad_page(page
, bad_reason
, bad_flags
);
1278 static int prep_new_page(struct page
*page
, unsigned int order
, gfp_t gfp_flags
,
1283 for (i
= 0; i
< (1 << order
); i
++) {
1284 struct page
*p
= page
+ i
;
1285 if (unlikely(check_new_page(p
)))
1289 set_page_private(page
, 0);
1290 set_page_refcounted(page
);
1292 arch_alloc_page(page
, order
);
1293 kernel_map_pages(page
, 1 << order
, 1);
1294 kasan_alloc_pages(page
, order
);
1296 if (gfp_flags
& __GFP_ZERO
)
1297 for (i
= 0; i
< (1 << order
); i
++)
1298 clear_highpage(page
+ i
);
1300 if (order
&& (gfp_flags
& __GFP_COMP
))
1301 prep_compound_page(page
, order
);
1303 set_page_owner(page
, order
, gfp_flags
);
1306 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was necessary to
1307 * allocate the page. The expectation is that the caller is taking
1308 * steps that will free more memory. The caller should avoid the page
1309 * being used for !PFMEMALLOC purposes.
1311 page
->pfmemalloc
= !!(alloc_flags
& ALLOC_NO_WATERMARKS
);
1317 * Go through the free lists for the given migratetype and remove
1318 * the smallest available page from the freelists
1321 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
1324 unsigned int current_order
;
1325 struct free_area
*area
;
1328 /* Find a page of the appropriate size in the preferred list */
1329 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
1330 area
= &(zone
->free_area
[current_order
]);
1331 if (list_empty(&area
->free_list
[migratetype
]))
1334 page
= list_entry(area
->free_list
[migratetype
].next
,
1336 list_del(&page
->lru
);
1337 rmv_page_order(page
);
1339 expand(zone
, page
, order
, current_order
, area
, migratetype
);
1340 set_freepage_migratetype(page
, migratetype
);
1349 * This array describes the order lists are fallen back to when
1350 * the free lists for the desirable migrate type are depleted
1352 static int fallbacks
[MIGRATE_TYPES
][4] = {
1353 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
1354 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
1355 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_RESERVE
},
1357 [MIGRATE_CMA
] = { MIGRATE_RESERVE
}, /* Never used */
1359 [MIGRATE_RESERVE
] = { MIGRATE_RESERVE
}, /* Never used */
1360 #ifdef CONFIG_MEMORY_ISOLATION
1361 [MIGRATE_ISOLATE
] = { MIGRATE_RESERVE
}, /* Never used */
1366 static struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
1369 return __rmqueue_smallest(zone
, order
, MIGRATE_CMA
);
1372 static inline struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
1373 unsigned int order
) { return NULL
; }
1377 * Move the free pages in a range to the free lists of the requested type.
1378 * Note that start_page and end_pages are not aligned on a pageblock
1379 * boundary. If alignment is required, use move_freepages_block()
1381 int move_freepages(struct zone
*zone
,
1382 struct page
*start_page
, struct page
*end_page
,
1386 unsigned long order
;
1387 int pages_moved
= 0;
1389 #ifndef CONFIG_HOLES_IN_ZONE
1391 * page_zone is not safe to call in this context when
1392 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1393 * anyway as we check zone boundaries in move_freepages_block().
1394 * Remove at a later date when no bug reports exist related to
1395 * grouping pages by mobility
1397 VM_BUG_ON(page_zone(start_page
) != page_zone(end_page
));
1400 for (page
= start_page
; page
<= end_page
;) {
1401 /* Make sure we are not inadvertently changing nodes */
1402 VM_BUG_ON_PAGE(page_to_nid(page
) != zone_to_nid(zone
), page
);
1404 if (!pfn_valid_within(page_to_pfn(page
))) {
1409 if (!PageBuddy(page
)) {
1414 order
= page_order(page
);
1415 list_move(&page
->lru
,
1416 &zone
->free_area
[order
].free_list
[migratetype
]);
1417 set_freepage_migratetype(page
, migratetype
);
1419 pages_moved
+= 1 << order
;
1425 int move_freepages_block(struct zone
*zone
, struct page
*page
,
1428 unsigned long start_pfn
, end_pfn
;
1429 struct page
*start_page
, *end_page
;
1431 start_pfn
= page_to_pfn(page
);
1432 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
1433 start_page
= pfn_to_page(start_pfn
);
1434 end_page
= start_page
+ pageblock_nr_pages
- 1;
1435 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
1437 /* Do not cross zone boundaries */
1438 if (!zone_spans_pfn(zone
, start_pfn
))
1440 if (!zone_spans_pfn(zone
, end_pfn
))
1443 return move_freepages(zone
, start_page
, end_page
, migratetype
);
1446 static void change_pageblock_range(struct page
*pageblock_page
,
1447 int start_order
, int migratetype
)
1449 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
1451 while (nr_pageblocks
--) {
1452 set_pageblock_migratetype(pageblock_page
, migratetype
);
1453 pageblock_page
+= pageblock_nr_pages
;
1458 * When we are falling back to another migratetype during allocation, try to
1459 * steal extra free pages from the same pageblocks to satisfy further
1460 * allocations, instead of polluting multiple pageblocks.
1462 * If we are stealing a relatively large buddy page, it is likely there will
1463 * be more free pages in the pageblock, so try to steal them all. For
1464 * reclaimable and unmovable allocations, we steal regardless of page size,
1465 * as fragmentation caused by those allocations polluting movable pageblocks
1466 * is worse than movable allocations stealing from unmovable and reclaimable
1469 static bool can_steal_fallback(unsigned int order
, int start_mt
)
1472 * Leaving this order check is intended, although there is
1473 * relaxed order check in next check. The reason is that
1474 * we can actually steal whole pageblock if this condition met,
1475 * but, below check doesn't guarantee it and that is just heuristic
1476 * so could be changed anytime.
1478 if (order
>= pageblock_order
)
1481 if (order
>= pageblock_order
/ 2 ||
1482 start_mt
== MIGRATE_RECLAIMABLE
||
1483 start_mt
== MIGRATE_UNMOVABLE
||
1484 page_group_by_mobility_disabled
)
1491 * This function implements actual steal behaviour. If order is large enough,
1492 * we can steal whole pageblock. If not, we first move freepages in this
1493 * pageblock and check whether half of pages are moved or not. If half of
1494 * pages are moved, we can change migratetype of pageblock and permanently
1495 * use it's pages as requested migratetype in the future.
1497 static void steal_suitable_fallback(struct zone
*zone
, struct page
*page
,
1500 int current_order
= page_order(page
);
1503 /* Take ownership for orders >= pageblock_order */
1504 if (current_order
>= pageblock_order
) {
1505 change_pageblock_range(page
, current_order
, start_type
);
1509 pages
= move_freepages_block(zone
, page
, start_type
);
1511 /* Claim the whole block if over half of it is free */
1512 if (pages
>= (1 << (pageblock_order
-1)) ||
1513 page_group_by_mobility_disabled
)
1514 set_pageblock_migratetype(page
, start_type
);
1518 * Check whether there is a suitable fallback freepage with requested order.
1519 * If only_stealable is true, this function returns fallback_mt only if
1520 * we can steal other freepages all together. This would help to reduce
1521 * fragmentation due to mixed migratetype pages in one pageblock.
1523 int find_suitable_fallback(struct free_area
*area
, unsigned int order
,
1524 int migratetype
, bool only_stealable
, bool *can_steal
)
1529 if (area
->nr_free
== 0)
1534 fallback_mt
= fallbacks
[migratetype
][i
];
1535 if (fallback_mt
== MIGRATE_RESERVE
)
1538 if (list_empty(&area
->free_list
[fallback_mt
]))
1541 if (can_steal_fallback(order
, migratetype
))
1544 if (!only_stealable
)
1554 /* Remove an element from the buddy allocator from the fallback list */
1555 static inline struct page
*
1556 __rmqueue_fallback(struct zone
*zone
, unsigned int order
, int start_migratetype
)
1558 struct free_area
*area
;
1559 unsigned int current_order
;
1564 /* Find the largest possible block of pages in the other list */
1565 for (current_order
= MAX_ORDER
-1;
1566 current_order
>= order
&& current_order
<= MAX_ORDER
-1;
1568 area
= &(zone
->free_area
[current_order
]);
1569 fallback_mt
= find_suitable_fallback(area
, current_order
,
1570 start_migratetype
, false, &can_steal
);
1571 if (fallback_mt
== -1)
1574 page
= list_entry(area
->free_list
[fallback_mt
].next
,
1577 steal_suitable_fallback(zone
, page
, start_migratetype
);
1579 /* Remove the page from the freelists */
1581 list_del(&page
->lru
);
1582 rmv_page_order(page
);
1584 expand(zone
, page
, order
, current_order
, area
,
1587 * The freepage_migratetype may differ from pageblock's
1588 * migratetype depending on the decisions in
1589 * try_to_steal_freepages(). This is OK as long as it
1590 * does not differ for MIGRATE_CMA pageblocks. For CMA
1591 * we need to make sure unallocated pages flushed from
1592 * pcp lists are returned to the correct freelist.
1594 set_freepage_migratetype(page
, start_migratetype
);
1596 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
1597 start_migratetype
, fallback_mt
);
1606 * Do the hard work of removing an element from the buddy allocator.
1607 * Call me with the zone->lock already held.
1609 static struct page
*__rmqueue(struct zone
*zone
, unsigned int order
,
1615 page
= __rmqueue_smallest(zone
, order
, migratetype
);
1617 if (unlikely(!page
) && migratetype
!= MIGRATE_RESERVE
) {
1618 if (migratetype
== MIGRATE_MOVABLE
)
1619 page
= __rmqueue_cma_fallback(zone
, order
);
1622 page
= __rmqueue_fallback(zone
, order
, migratetype
);
1625 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1626 * is used because __rmqueue_smallest is an inline function
1627 * and we want just one call site
1630 migratetype
= MIGRATE_RESERVE
;
1635 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
1640 * Obtain a specified number of elements from the buddy allocator, all under
1641 * a single hold of the lock, for efficiency. Add them to the supplied list.
1642 * Returns the number of new pages which were placed at *list.
1644 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
1645 unsigned long count
, struct list_head
*list
,
1646 int migratetype
, bool cold
)
1650 spin_lock(&zone
->lock
);
1651 for (i
= 0; i
< count
; ++i
) {
1652 struct page
*page
= __rmqueue(zone
, order
, migratetype
);
1653 if (unlikely(page
== NULL
))
1657 * Split buddy pages returned by expand() are received here
1658 * in physical page order. The page is added to the callers and
1659 * list and the list head then moves forward. From the callers
1660 * perspective, the linked list is ordered by page number in
1661 * some conditions. This is useful for IO devices that can
1662 * merge IO requests if the physical pages are ordered
1666 list_add(&page
->lru
, list
);
1668 list_add_tail(&page
->lru
, list
);
1670 if (is_migrate_cma(get_freepage_migratetype(page
)))
1671 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
,
1674 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
1675 spin_unlock(&zone
->lock
);
1681 * Called from the vmstat counter updater to drain pagesets of this
1682 * currently executing processor on remote nodes after they have
1685 * Note that this function must be called with the thread pinned to
1686 * a single processor.
1688 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
1690 unsigned long flags
;
1691 int to_drain
, batch
;
1693 local_irq_save(flags
);
1694 batch
= READ_ONCE(pcp
->batch
);
1695 to_drain
= min(pcp
->count
, batch
);
1697 free_pcppages_bulk(zone
, to_drain
, pcp
);
1698 pcp
->count
-= to_drain
;
1700 local_irq_restore(flags
);
1705 * Drain pcplists of the indicated processor and zone.
1707 * The processor must either be the current processor and the
1708 * thread pinned to the current processor or a processor that
1711 static void drain_pages_zone(unsigned int cpu
, struct zone
*zone
)
1713 unsigned long flags
;
1714 struct per_cpu_pageset
*pset
;
1715 struct per_cpu_pages
*pcp
;
1717 local_irq_save(flags
);
1718 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
1722 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
1725 local_irq_restore(flags
);
1729 * Drain pcplists of all zones on the indicated processor.
1731 * The processor must either be the current processor and the
1732 * thread pinned to the current processor or a processor that
1735 static void drain_pages(unsigned int cpu
)
1739 for_each_populated_zone(zone
) {
1740 drain_pages_zone(cpu
, zone
);
1745 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1747 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
1748 * the single zone's pages.
1750 void drain_local_pages(struct zone
*zone
)
1752 int cpu
= smp_processor_id();
1755 drain_pages_zone(cpu
, zone
);
1761 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1763 * When zone parameter is non-NULL, spill just the single zone's pages.
1765 * Note that this code is protected against sending an IPI to an offline
1766 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1767 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1768 * nothing keeps CPUs from showing up after we populated the cpumask and
1769 * before the call to on_each_cpu_mask().
1771 void drain_all_pages(struct zone
*zone
)
1776 * Allocate in the BSS so we wont require allocation in
1777 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1779 static cpumask_t cpus_with_pcps
;
1782 * We don't care about racing with CPU hotplug event
1783 * as offline notification will cause the notified
1784 * cpu to drain that CPU pcps and on_each_cpu_mask
1785 * disables preemption as part of its processing
1787 for_each_online_cpu(cpu
) {
1788 struct per_cpu_pageset
*pcp
;
1790 bool has_pcps
= false;
1793 pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
1797 for_each_populated_zone(z
) {
1798 pcp
= per_cpu_ptr(z
->pageset
, cpu
);
1799 if (pcp
->pcp
.count
) {
1807 cpumask_set_cpu(cpu
, &cpus_with_pcps
);
1809 cpumask_clear_cpu(cpu
, &cpus_with_pcps
);
1811 on_each_cpu_mask(&cpus_with_pcps
, (smp_call_func_t
) drain_local_pages
,
1815 #ifdef CONFIG_HIBERNATION
1817 void mark_free_pages(struct zone
*zone
)
1819 unsigned long pfn
, max_zone_pfn
;
1820 unsigned long flags
;
1821 unsigned int order
, t
;
1822 struct list_head
*curr
;
1824 if (zone_is_empty(zone
))
1827 spin_lock_irqsave(&zone
->lock
, flags
);
1829 max_zone_pfn
= zone_end_pfn(zone
);
1830 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
1831 if (pfn_valid(pfn
)) {
1832 struct page
*page
= pfn_to_page(pfn
);
1834 if (!swsusp_page_is_forbidden(page
))
1835 swsusp_unset_page_free(page
);
1838 for_each_migratetype_order(order
, t
) {
1839 list_for_each(curr
, &zone
->free_area
[order
].free_list
[t
]) {
1842 pfn
= page_to_pfn(list_entry(curr
, struct page
, lru
));
1843 for (i
= 0; i
< (1UL << order
); i
++)
1844 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
1847 spin_unlock_irqrestore(&zone
->lock
, flags
);
1849 #endif /* CONFIG_PM */
1852 * Free a 0-order page
1853 * cold == true ? free a cold page : free a hot page
1855 void free_hot_cold_page(struct page
*page
, bool cold
)
1857 struct zone
*zone
= page_zone(page
);
1858 struct per_cpu_pages
*pcp
;
1859 unsigned long flags
;
1860 unsigned long pfn
= page_to_pfn(page
);
1863 if (!free_pages_prepare(page
, 0))
1866 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1867 set_freepage_migratetype(page
, migratetype
);
1868 local_irq_save(flags
);
1869 __count_vm_event(PGFREE
);
1872 * We only track unmovable, reclaimable and movable on pcp lists.
1873 * Free ISOLATE pages back to the allocator because they are being
1874 * offlined but treat RESERVE as movable pages so we can get those
1875 * areas back if necessary. Otherwise, we may have to free
1876 * excessively into the page allocator
1878 if (migratetype
>= MIGRATE_PCPTYPES
) {
1879 if (unlikely(is_migrate_isolate(migratetype
))) {
1880 free_one_page(zone
, page
, pfn
, 0, migratetype
);
1883 migratetype
= MIGRATE_MOVABLE
;
1886 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
1888 list_add(&page
->lru
, &pcp
->lists
[migratetype
]);
1890 list_add_tail(&page
->lru
, &pcp
->lists
[migratetype
]);
1892 if (pcp
->count
>= pcp
->high
) {
1893 unsigned long batch
= READ_ONCE(pcp
->batch
);
1894 free_pcppages_bulk(zone
, batch
, pcp
);
1895 pcp
->count
-= batch
;
1899 local_irq_restore(flags
);
1903 * Free a list of 0-order pages
1905 void free_hot_cold_page_list(struct list_head
*list
, bool cold
)
1907 struct page
*page
, *next
;
1909 list_for_each_entry_safe(page
, next
, list
, lru
) {
1910 trace_mm_page_free_batched(page
, cold
);
1911 free_hot_cold_page(page
, cold
);
1916 * split_page takes a non-compound higher-order page, and splits it into
1917 * n (1<<order) sub-pages: page[0..n]
1918 * Each sub-page must be freed individually.
1920 * Note: this is probably too low level an operation for use in drivers.
1921 * Please consult with lkml before using this in your driver.
1923 void split_page(struct page
*page
, unsigned int order
)
1927 VM_BUG_ON_PAGE(PageCompound(page
), page
);
1928 VM_BUG_ON_PAGE(!page_count(page
), page
);
1930 #ifdef CONFIG_KMEMCHECK
1932 * Split shadow pages too, because free(page[0]) would
1933 * otherwise free the whole shadow.
1935 if (kmemcheck_page_is_tracked(page
))
1936 split_page(virt_to_page(page
[0].shadow
), order
);
1939 set_page_owner(page
, 0, 0);
1940 for (i
= 1; i
< (1 << order
); i
++) {
1941 set_page_refcounted(page
+ i
);
1942 set_page_owner(page
+ i
, 0, 0);
1945 EXPORT_SYMBOL_GPL(split_page
);
1947 int __isolate_free_page(struct page
*page
, unsigned int order
)
1949 unsigned long watermark
;
1953 BUG_ON(!PageBuddy(page
));
1955 zone
= page_zone(page
);
1956 mt
= get_pageblock_migratetype(page
);
1958 if (!is_migrate_isolate(mt
)) {
1959 /* Obey watermarks as if the page was being allocated */
1960 watermark
= low_wmark_pages(zone
) + (1 << order
);
1961 if (!zone_watermark_ok(zone
, 0, watermark
, 0, 0))
1964 __mod_zone_freepage_state(zone
, -(1UL << order
), mt
);
1967 /* Remove page from free list */
1968 list_del(&page
->lru
);
1969 zone
->free_area
[order
].nr_free
--;
1970 rmv_page_order(page
);
1972 /* Set the pageblock if the isolated page is at least a pageblock */
1973 if (order
>= pageblock_order
- 1) {
1974 struct page
*endpage
= page
+ (1 << order
) - 1;
1975 for (; page
< endpage
; page
+= pageblock_nr_pages
) {
1976 int mt
= get_pageblock_migratetype(page
);
1977 if (!is_migrate_isolate(mt
) && !is_migrate_cma(mt
))
1978 set_pageblock_migratetype(page
,
1983 set_page_owner(page
, order
, 0);
1984 return 1UL << order
;
1988 * Similar to split_page except the page is already free. As this is only
1989 * being used for migration, the migratetype of the block also changes.
1990 * As this is called with interrupts disabled, the caller is responsible
1991 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1994 * Note: this is probably too low level an operation for use in drivers.
1995 * Please consult with lkml before using this in your driver.
1997 int split_free_page(struct page
*page
)
2002 order
= page_order(page
);
2004 nr_pages
= __isolate_free_page(page
, order
);
2008 /* Split into individual pages */
2009 set_page_refcounted(page
);
2010 split_page(page
, order
);
2015 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2018 struct page
*buffered_rmqueue(struct zone
*preferred_zone
,
2019 struct zone
*zone
, unsigned int order
,
2020 gfp_t gfp_flags
, int migratetype
)
2022 unsigned long flags
;
2024 bool cold
= ((gfp_flags
& __GFP_COLD
) != 0);
2026 if (likely(order
== 0)) {
2027 struct per_cpu_pages
*pcp
;
2028 struct list_head
*list
;
2030 local_irq_save(flags
);
2031 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
2032 list
= &pcp
->lists
[migratetype
];
2033 if (list_empty(list
)) {
2034 pcp
->count
+= rmqueue_bulk(zone
, 0,
2037 if (unlikely(list_empty(list
)))
2042 page
= list_entry(list
->prev
, struct page
, lru
);
2044 page
= list_entry(list
->next
, struct page
, lru
);
2046 list_del(&page
->lru
);
2049 if (unlikely(gfp_flags
& __GFP_NOFAIL
)) {
2051 * __GFP_NOFAIL is not to be used in new code.
2053 * All __GFP_NOFAIL callers should be fixed so that they
2054 * properly detect and handle allocation failures.
2056 * We most definitely don't want callers attempting to
2057 * allocate greater than order-1 page units with
2060 WARN_ON_ONCE(order
> 1);
2062 spin_lock_irqsave(&zone
->lock
, flags
);
2063 page
= __rmqueue(zone
, order
, migratetype
);
2064 spin_unlock(&zone
->lock
);
2067 __mod_zone_freepage_state(zone
, -(1 << order
),
2068 get_freepage_migratetype(page
));
2071 __mod_zone_page_state(zone
, NR_ALLOC_BATCH
, -(1 << order
));
2072 if (atomic_long_read(&zone
->vm_stat
[NR_ALLOC_BATCH
]) <= 0 &&
2073 !test_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
))
2074 set_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
);
2076 __count_zone_vm_events(PGALLOC
, zone
, 1 << order
);
2077 zone_statistics(preferred_zone
, zone
, gfp_flags
);
2078 local_irq_restore(flags
);
2080 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
2084 local_irq_restore(flags
);
2088 #ifdef CONFIG_FAIL_PAGE_ALLOC
2091 struct fault_attr attr
;
2093 u32 ignore_gfp_highmem
;
2094 u32 ignore_gfp_wait
;
2096 } fail_page_alloc
= {
2097 .attr
= FAULT_ATTR_INITIALIZER
,
2098 .ignore_gfp_wait
= 1,
2099 .ignore_gfp_highmem
= 1,
2103 static int __init
setup_fail_page_alloc(char *str
)
2105 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
2107 __setup("fail_page_alloc=", setup_fail_page_alloc
);
2109 static bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
2111 if (order
< fail_page_alloc
.min_order
)
2113 if (gfp_mask
& __GFP_NOFAIL
)
2115 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
2117 if (fail_page_alloc
.ignore_gfp_wait
&& (gfp_mask
& __GFP_WAIT
))
2120 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
2123 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2125 static int __init
fail_page_alloc_debugfs(void)
2127 umode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
2130 dir
= fault_create_debugfs_attr("fail_page_alloc", NULL
,
2131 &fail_page_alloc
.attr
);
2133 return PTR_ERR(dir
);
2135 if (!debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
2136 &fail_page_alloc
.ignore_gfp_wait
))
2138 if (!debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
2139 &fail_page_alloc
.ignore_gfp_highmem
))
2141 if (!debugfs_create_u32("min-order", mode
, dir
,
2142 &fail_page_alloc
.min_order
))
2147 debugfs_remove_recursive(dir
);
2152 late_initcall(fail_page_alloc_debugfs
);
2154 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2156 #else /* CONFIG_FAIL_PAGE_ALLOC */
2158 static inline bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
2163 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2166 * Return true if free pages are above 'mark'. This takes into account the order
2167 * of the allocation.
2169 static bool __zone_watermark_ok(struct zone
*z
, unsigned int order
,
2170 unsigned long mark
, int classzone_idx
, int alloc_flags
,
2173 /* free_pages may go negative - that's OK */
2178 free_pages
-= (1 << order
) - 1;
2179 if (alloc_flags
& ALLOC_HIGH
)
2181 if (alloc_flags
& ALLOC_HARDER
)
2184 /* If allocation can't use CMA areas don't use free CMA pages */
2185 if (!(alloc_flags
& ALLOC_CMA
))
2186 free_cma
= zone_page_state(z
, NR_FREE_CMA_PAGES
);
2189 if (free_pages
- free_cma
<= min
+ z
->lowmem_reserve
[classzone_idx
])
2191 for (o
= 0; o
< order
; o
++) {
2192 /* At the next order, this order's pages become unavailable */
2193 free_pages
-= z
->free_area
[o
].nr_free
<< o
;
2195 /* Require fewer higher order pages to be free */
2198 if (free_pages
<= min
)
2204 bool zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
2205 int classzone_idx
, int alloc_flags
)
2207 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
2208 zone_page_state(z
, NR_FREE_PAGES
));
2211 bool zone_watermark_ok_safe(struct zone
*z
, unsigned int order
,
2212 unsigned long mark
, int classzone_idx
, int alloc_flags
)
2214 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
2216 if (z
->percpu_drift_mark
&& free_pages
< z
->percpu_drift_mark
)
2217 free_pages
= zone_page_state_snapshot(z
, NR_FREE_PAGES
);
2219 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
2225 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
2226 * skip over zones that are not allowed by the cpuset, or that have
2227 * been recently (in last second) found to be nearly full. See further
2228 * comments in mmzone.h. Reduces cache footprint of zonelist scans
2229 * that have to skip over a lot of full or unallowed zones.
2231 * If the zonelist cache is present in the passed zonelist, then
2232 * returns a pointer to the allowed node mask (either the current
2233 * tasks mems_allowed, or node_states[N_MEMORY].)
2235 * If the zonelist cache is not available for this zonelist, does
2236 * nothing and returns NULL.
2238 * If the fullzones BITMAP in the zonelist cache is stale (more than
2239 * a second since last zap'd) then we zap it out (clear its bits.)
2241 * We hold off even calling zlc_setup, until after we've checked the
2242 * first zone in the zonelist, on the theory that most allocations will
2243 * be satisfied from that first zone, so best to examine that zone as
2244 * quickly as we can.
2246 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
2248 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
2249 nodemask_t
*allowednodes
; /* zonelist_cache approximation */
2251 zlc
= zonelist
->zlcache_ptr
;
2255 if (time_after(jiffies
, zlc
->last_full_zap
+ HZ
)) {
2256 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
2257 zlc
->last_full_zap
= jiffies
;
2260 allowednodes
= !in_interrupt() && (alloc_flags
& ALLOC_CPUSET
) ?
2261 &cpuset_current_mems_allowed
:
2262 &node_states
[N_MEMORY
];
2263 return allowednodes
;
2267 * Given 'z' scanning a zonelist, run a couple of quick checks to see
2268 * if it is worth looking at further for free memory:
2269 * 1) Check that the zone isn't thought to be full (doesn't have its
2270 * bit set in the zonelist_cache fullzones BITMAP).
2271 * 2) Check that the zones node (obtained from the zonelist_cache
2272 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
2273 * Return true (non-zero) if zone is worth looking at further, or
2274 * else return false (zero) if it is not.
2276 * This check -ignores- the distinction between various watermarks,
2277 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
2278 * found to be full for any variation of these watermarks, it will
2279 * be considered full for up to one second by all requests, unless
2280 * we are so low on memory on all allowed nodes that we are forced
2281 * into the second scan of the zonelist.
2283 * In the second scan we ignore this zonelist cache and exactly
2284 * apply the watermarks to all zones, even it is slower to do so.
2285 * We are low on memory in the second scan, and should leave no stone
2286 * unturned looking for a free page.
2288 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
2289 nodemask_t
*allowednodes
)
2291 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
2292 int i
; /* index of *z in zonelist zones */
2293 int n
; /* node that zone *z is on */
2295 zlc
= zonelist
->zlcache_ptr
;
2299 i
= z
- zonelist
->_zonerefs
;
2302 /* This zone is worth trying if it is allowed but not full */
2303 return node_isset(n
, *allowednodes
) && !test_bit(i
, zlc
->fullzones
);
2307 * Given 'z' scanning a zonelist, set the corresponding bit in
2308 * zlc->fullzones, so that subsequent attempts to allocate a page
2309 * from that zone don't waste time re-examining it.
2311 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
2313 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
2314 int i
; /* index of *z in zonelist zones */
2316 zlc
= zonelist
->zlcache_ptr
;
2320 i
= z
- zonelist
->_zonerefs
;
2322 set_bit(i
, zlc
->fullzones
);
2326 * clear all zones full, called after direct reclaim makes progress so that
2327 * a zone that was recently full is not skipped over for up to a second
2329 static void zlc_clear_zones_full(struct zonelist
*zonelist
)
2331 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
2333 zlc
= zonelist
->zlcache_ptr
;
2337 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
2340 static bool zone_local(struct zone
*local_zone
, struct zone
*zone
)
2342 return local_zone
->node
== zone
->node
;
2345 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
2347 return node_distance(zone_to_nid(local_zone
), zone_to_nid(zone
)) <
2351 #else /* CONFIG_NUMA */
2353 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
2358 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
2359 nodemask_t
*allowednodes
)
2364 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
2368 static void zlc_clear_zones_full(struct zonelist
*zonelist
)
2372 static bool zone_local(struct zone
*local_zone
, struct zone
*zone
)
2377 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
2382 #endif /* CONFIG_NUMA */
2384 static void reset_alloc_batches(struct zone
*preferred_zone
)
2386 struct zone
*zone
= preferred_zone
->zone_pgdat
->node_zones
;
2389 mod_zone_page_state(zone
, NR_ALLOC_BATCH
,
2390 high_wmark_pages(zone
) - low_wmark_pages(zone
) -
2391 atomic_long_read(&zone
->vm_stat
[NR_ALLOC_BATCH
]));
2392 clear_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
);
2393 } while (zone
++ != preferred_zone
);
2397 * get_page_from_freelist goes through the zonelist trying to allocate
2400 static struct page
*
2401 get_page_from_freelist(gfp_t gfp_mask
, unsigned int order
, int alloc_flags
,
2402 const struct alloc_context
*ac
)
2404 struct zonelist
*zonelist
= ac
->zonelist
;
2406 struct page
*page
= NULL
;
2408 nodemask_t
*allowednodes
= NULL
;/* zonelist_cache approximation */
2409 int zlc_active
= 0; /* set if using zonelist_cache */
2410 int did_zlc_setup
= 0; /* just call zlc_setup() one time */
2411 bool consider_zone_dirty
= (alloc_flags
& ALLOC_WMARK_LOW
) &&
2412 (gfp_mask
& __GFP_WRITE
);
2413 int nr_fair_skipped
= 0;
2414 bool zonelist_rescan
;
2417 zonelist_rescan
= false;
2420 * Scan zonelist, looking for a zone with enough free.
2421 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2423 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, ac
->high_zoneidx
,
2427 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
&&
2428 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
2430 if (cpusets_enabled() &&
2431 (alloc_flags
& ALLOC_CPUSET
) &&
2432 !cpuset_zone_allowed(zone
, gfp_mask
))
2435 * Distribute pages in proportion to the individual
2436 * zone size to ensure fair page aging. The zone a
2437 * page was allocated in should have no effect on the
2438 * time the page has in memory before being reclaimed.
2440 if (alloc_flags
& ALLOC_FAIR
) {
2441 if (!zone_local(ac
->preferred_zone
, zone
))
2443 if (test_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
)) {
2449 * When allocating a page cache page for writing, we
2450 * want to get it from a zone that is within its dirty
2451 * limit, such that no single zone holds more than its
2452 * proportional share of globally allowed dirty pages.
2453 * The dirty limits take into account the zone's
2454 * lowmem reserves and high watermark so that kswapd
2455 * should be able to balance it without having to
2456 * write pages from its LRU list.
2458 * This may look like it could increase pressure on
2459 * lower zones by failing allocations in higher zones
2460 * before they are full. But the pages that do spill
2461 * over are limited as the lower zones are protected
2462 * by this very same mechanism. It should not become
2463 * a practical burden to them.
2465 * XXX: For now, allow allocations to potentially
2466 * exceed the per-zone dirty limit in the slowpath
2467 * (ALLOC_WMARK_LOW unset) before going into reclaim,
2468 * which is important when on a NUMA setup the allowed
2469 * zones are together not big enough to reach the
2470 * global limit. The proper fix for these situations
2471 * will require awareness of zones in the
2472 * dirty-throttling and the flusher threads.
2474 if (consider_zone_dirty
&& !zone_dirty_ok(zone
))
2477 mark
= zone
->watermark
[alloc_flags
& ALLOC_WMARK_MASK
];
2478 if (!zone_watermark_ok(zone
, order
, mark
,
2479 ac
->classzone_idx
, alloc_flags
)) {
2482 /* Checked here to keep the fast path fast */
2483 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
2484 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
2487 if (IS_ENABLED(CONFIG_NUMA
) &&
2488 !did_zlc_setup
&& nr_online_nodes
> 1) {
2490 * we do zlc_setup if there are multiple nodes
2491 * and before considering the first zone allowed
2494 allowednodes
= zlc_setup(zonelist
, alloc_flags
);
2499 if (zone_reclaim_mode
== 0 ||
2500 !zone_allows_reclaim(ac
->preferred_zone
, zone
))
2501 goto this_zone_full
;
2504 * As we may have just activated ZLC, check if the first
2505 * eligible zone has failed zone_reclaim recently.
2507 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
&&
2508 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
2511 ret
= zone_reclaim(zone
, gfp_mask
, order
);
2513 case ZONE_RECLAIM_NOSCAN
:
2516 case ZONE_RECLAIM_FULL
:
2517 /* scanned but unreclaimable */
2520 /* did we reclaim enough */
2521 if (zone_watermark_ok(zone
, order
, mark
,
2522 ac
->classzone_idx
, alloc_flags
))
2526 * Failed to reclaim enough to meet watermark.
2527 * Only mark the zone full if checking the min
2528 * watermark or if we failed to reclaim just
2529 * 1<<order pages or else the page allocator
2530 * fastpath will prematurely mark zones full
2531 * when the watermark is between the low and
2534 if (((alloc_flags
& ALLOC_WMARK_MASK
) == ALLOC_WMARK_MIN
) ||
2535 ret
== ZONE_RECLAIM_SOME
)
2536 goto this_zone_full
;
2543 page
= buffered_rmqueue(ac
->preferred_zone
, zone
, order
,
2544 gfp_mask
, ac
->migratetype
);
2546 if (prep_new_page(page
, order
, gfp_mask
, alloc_flags
))
2551 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
)
2552 zlc_mark_zone_full(zonelist
, z
);
2556 * The first pass makes sure allocations are spread fairly within the
2557 * local node. However, the local node might have free pages left
2558 * after the fairness batches are exhausted, and remote zones haven't
2559 * even been considered yet. Try once more without fairness, and
2560 * include remote zones now, before entering the slowpath and waking
2561 * kswapd: prefer spilling to a remote zone over swapping locally.
2563 if (alloc_flags
& ALLOC_FAIR
) {
2564 alloc_flags
&= ~ALLOC_FAIR
;
2565 if (nr_fair_skipped
) {
2566 zonelist_rescan
= true;
2567 reset_alloc_batches(ac
->preferred_zone
);
2569 if (nr_online_nodes
> 1)
2570 zonelist_rescan
= true;
2573 if (unlikely(IS_ENABLED(CONFIG_NUMA
) && zlc_active
)) {
2574 /* Disable zlc cache for second zonelist scan */
2576 zonelist_rescan
= true;
2579 if (zonelist_rescan
)
2586 * Large machines with many possible nodes should not always dump per-node
2587 * meminfo in irq context.
2589 static inline bool should_suppress_show_mem(void)
2594 ret
= in_interrupt();
2599 static DEFINE_RATELIMIT_STATE(nopage_rs
,
2600 DEFAULT_RATELIMIT_INTERVAL
,
2601 DEFAULT_RATELIMIT_BURST
);
2603 void warn_alloc_failed(gfp_t gfp_mask
, int order
, const char *fmt
, ...)
2605 unsigned int filter
= SHOW_MEM_FILTER_NODES
;
2607 if ((gfp_mask
& __GFP_NOWARN
) || !__ratelimit(&nopage_rs
) ||
2608 debug_guardpage_minorder() > 0)
2612 * This documents exceptions given to allocations in certain
2613 * contexts that are allowed to allocate outside current's set
2616 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
2617 if (test_thread_flag(TIF_MEMDIE
) ||
2618 (current
->flags
& (PF_MEMALLOC
| PF_EXITING
)))
2619 filter
&= ~SHOW_MEM_FILTER_NODES
;
2620 if (in_interrupt() || !(gfp_mask
& __GFP_WAIT
))
2621 filter
&= ~SHOW_MEM_FILTER_NODES
;
2624 struct va_format vaf
;
2627 va_start(args
, fmt
);
2632 pr_warn("%pV", &vaf
);
2637 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2638 current
->comm
, order
, gfp_mask
);
2641 if (!should_suppress_show_mem())
2645 static inline struct page
*
2646 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
2647 const struct alloc_context
*ac
, unsigned long *did_some_progress
)
2651 *did_some_progress
= 0;
2654 * Acquire the oom lock. If that fails, somebody else is
2655 * making progress for us.
2657 if (!mutex_trylock(&oom_lock
)) {
2658 *did_some_progress
= 1;
2659 schedule_timeout_uninterruptible(1);
2664 * Go through the zonelist yet one more time, keep very high watermark
2665 * here, this is only to catch a parallel oom killing, we must fail if
2666 * we're still under heavy pressure.
2668 page
= get_page_from_freelist(gfp_mask
| __GFP_HARDWALL
, order
,
2669 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
, ac
);
2673 if (!(gfp_mask
& __GFP_NOFAIL
)) {
2674 /* Coredumps can quickly deplete all memory reserves */
2675 if (current
->flags
& PF_DUMPCORE
)
2677 /* The OOM killer will not help higher order allocs */
2678 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
2680 /* The OOM killer does not needlessly kill tasks for lowmem */
2681 if (ac
->high_zoneidx
< ZONE_NORMAL
)
2683 /* The OOM killer does not compensate for IO-less reclaim */
2684 if (!(gfp_mask
& __GFP_FS
)) {
2686 * XXX: Page reclaim didn't yield anything,
2687 * and the OOM killer can't be invoked, but
2688 * keep looping as per tradition.
2690 *did_some_progress
= 1;
2693 if (pm_suspended_storage())
2695 /* The OOM killer may not free memory on a specific node */
2696 if (gfp_mask
& __GFP_THISNODE
)
2699 /* Exhausted what can be done so it's blamo time */
2700 if (out_of_memory(ac
->zonelist
, gfp_mask
, order
, ac
->nodemask
, false)
2701 || WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
))
2702 *did_some_progress
= 1;
2704 mutex_unlock(&oom_lock
);
2708 #ifdef CONFIG_COMPACTION
2709 /* Try memory compaction for high-order allocations before reclaim */
2710 static struct page
*
2711 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
2712 int alloc_flags
, const struct alloc_context
*ac
,
2713 enum migrate_mode mode
, int *contended_compaction
,
2714 bool *deferred_compaction
)
2716 unsigned long compact_result
;
2722 current
->flags
|= PF_MEMALLOC
;
2723 compact_result
= try_to_compact_pages(gfp_mask
, order
, alloc_flags
, ac
,
2724 mode
, contended_compaction
);
2725 current
->flags
&= ~PF_MEMALLOC
;
2727 switch (compact_result
) {
2728 case COMPACT_DEFERRED
:
2729 *deferred_compaction
= true;
2731 case COMPACT_SKIPPED
:
2738 * At least in one zone compaction wasn't deferred or skipped, so let's
2739 * count a compaction stall
2741 count_vm_event(COMPACTSTALL
);
2743 page
= get_page_from_freelist(gfp_mask
, order
,
2744 alloc_flags
& ~ALLOC_NO_WATERMARKS
, ac
);
2747 struct zone
*zone
= page_zone(page
);
2749 zone
->compact_blockskip_flush
= false;
2750 compaction_defer_reset(zone
, order
, true);
2751 count_vm_event(COMPACTSUCCESS
);
2756 * It's bad if compaction run occurs and fails. The most likely reason
2757 * is that pages exist, but not enough to satisfy watermarks.
2759 count_vm_event(COMPACTFAIL
);
2766 static inline struct page
*
2767 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
2768 int alloc_flags
, const struct alloc_context
*ac
,
2769 enum migrate_mode mode
, int *contended_compaction
,
2770 bool *deferred_compaction
)
2774 #endif /* CONFIG_COMPACTION */
2776 /* Perform direct synchronous page reclaim */
2778 __perform_reclaim(gfp_t gfp_mask
, unsigned int order
,
2779 const struct alloc_context
*ac
)
2781 struct reclaim_state reclaim_state
;
2786 /* We now go into synchronous reclaim */
2787 cpuset_memory_pressure_bump();
2788 current
->flags
|= PF_MEMALLOC
;
2789 lockdep_set_current_reclaim_state(gfp_mask
);
2790 reclaim_state
.reclaimed_slab
= 0;
2791 current
->reclaim_state
= &reclaim_state
;
2793 progress
= try_to_free_pages(ac
->zonelist
, order
, gfp_mask
,
2796 current
->reclaim_state
= NULL
;
2797 lockdep_clear_current_reclaim_state();
2798 current
->flags
&= ~PF_MEMALLOC
;
2805 /* The really slow allocator path where we enter direct reclaim */
2806 static inline struct page
*
2807 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
2808 int alloc_flags
, const struct alloc_context
*ac
,
2809 unsigned long *did_some_progress
)
2811 struct page
*page
= NULL
;
2812 bool drained
= false;
2814 *did_some_progress
= __perform_reclaim(gfp_mask
, order
, ac
);
2815 if (unlikely(!(*did_some_progress
)))
2818 /* After successful reclaim, reconsider all zones for allocation */
2819 if (IS_ENABLED(CONFIG_NUMA
))
2820 zlc_clear_zones_full(ac
->zonelist
);
2823 page
= get_page_from_freelist(gfp_mask
, order
,
2824 alloc_flags
& ~ALLOC_NO_WATERMARKS
, ac
);
2827 * If an allocation failed after direct reclaim, it could be because
2828 * pages are pinned on the per-cpu lists. Drain them and try again
2830 if (!page
&& !drained
) {
2831 drain_all_pages(NULL
);
2840 * This is called in the allocator slow-path if the allocation request is of
2841 * sufficient urgency to ignore watermarks and take other desperate measures
2843 static inline struct page
*
2844 __alloc_pages_high_priority(gfp_t gfp_mask
, unsigned int order
,
2845 const struct alloc_context
*ac
)
2850 page
= get_page_from_freelist(gfp_mask
, order
,
2851 ALLOC_NO_WATERMARKS
, ac
);
2853 if (!page
&& gfp_mask
& __GFP_NOFAIL
)
2854 wait_iff_congested(ac
->preferred_zone
, BLK_RW_ASYNC
,
2856 } while (!page
&& (gfp_mask
& __GFP_NOFAIL
));
2861 static void wake_all_kswapds(unsigned int order
, const struct alloc_context
*ac
)
2866 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
,
2867 ac
->high_zoneidx
, ac
->nodemask
)
2868 wakeup_kswapd(zone
, order
, zone_idx(ac
->preferred_zone
));
2872 gfp_to_alloc_flags(gfp_t gfp_mask
)
2874 int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
2875 const bool atomic
= !(gfp_mask
& (__GFP_WAIT
| __GFP_NO_KSWAPD
));
2877 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2878 BUILD_BUG_ON(__GFP_HIGH
!= (__force gfp_t
) ALLOC_HIGH
);
2881 * The caller may dip into page reserves a bit more if the caller
2882 * cannot run direct reclaim, or if the caller has realtime scheduling
2883 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2884 * set both ALLOC_HARDER (atomic == true) and ALLOC_HIGH (__GFP_HIGH).
2886 alloc_flags
|= (__force
int) (gfp_mask
& __GFP_HIGH
);
2890 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
2891 * if it can't schedule.
2893 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
2894 alloc_flags
|= ALLOC_HARDER
;
2896 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
2897 * comment for __cpuset_node_allowed().
2899 alloc_flags
&= ~ALLOC_CPUSET
;
2900 } else if (unlikely(rt_task(current
)) && !in_interrupt())
2901 alloc_flags
|= ALLOC_HARDER
;
2903 if (likely(!(gfp_mask
& __GFP_NOMEMALLOC
))) {
2904 if (gfp_mask
& __GFP_MEMALLOC
)
2905 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2906 else if (in_serving_softirq() && (current
->flags
& PF_MEMALLOC
))
2907 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2908 else if (!in_interrupt() &&
2909 ((current
->flags
& PF_MEMALLOC
) ||
2910 unlikely(test_thread_flag(TIF_MEMDIE
))))
2911 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2914 if (gfpflags_to_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
2915 alloc_flags
|= ALLOC_CMA
;
2920 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask
)
2922 return !!(gfp_to_alloc_flags(gfp_mask
) & ALLOC_NO_WATERMARKS
);
2925 static inline struct page
*
2926 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
2927 struct alloc_context
*ac
)
2929 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
2930 struct page
*page
= NULL
;
2932 unsigned long pages_reclaimed
= 0;
2933 unsigned long did_some_progress
;
2934 enum migrate_mode migration_mode
= MIGRATE_ASYNC
;
2935 bool deferred_compaction
= false;
2936 int contended_compaction
= COMPACT_CONTENDED_NONE
;
2939 * In the slowpath, we sanity check order to avoid ever trying to
2940 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2941 * be using allocators in order of preference for an area that is
2944 if (order
>= MAX_ORDER
) {
2945 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
2950 * If this allocation cannot block and it is for a specific node, then
2951 * fail early. There's no need to wakeup kswapd or retry for a
2952 * speculative node-specific allocation.
2954 if (IS_ENABLED(CONFIG_NUMA
) && (gfp_mask
& __GFP_THISNODE
) && !wait
)
2958 if (!(gfp_mask
& __GFP_NO_KSWAPD
))
2959 wake_all_kswapds(order
, ac
);
2962 * OK, we're below the kswapd watermark and have kicked background
2963 * reclaim. Now things get more complex, so set up alloc_flags according
2964 * to how we want to proceed.
2966 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
2969 * Find the true preferred zone if the allocation is unconstrained by
2972 if (!(alloc_flags
& ALLOC_CPUSET
) && !ac
->nodemask
) {
2973 struct zoneref
*preferred_zoneref
;
2974 preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
2975 ac
->high_zoneidx
, NULL
, &ac
->preferred_zone
);
2976 ac
->classzone_idx
= zonelist_zone_idx(preferred_zoneref
);
2979 /* This is the last chance, in general, before the goto nopage. */
2980 page
= get_page_from_freelist(gfp_mask
, order
,
2981 alloc_flags
& ~ALLOC_NO_WATERMARKS
, ac
);
2985 /* Allocate without watermarks if the context allows */
2986 if (alloc_flags
& ALLOC_NO_WATERMARKS
) {
2988 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2989 * the allocation is high priority and these type of
2990 * allocations are system rather than user orientated
2992 ac
->zonelist
= node_zonelist(numa_node_id(), gfp_mask
);
2994 page
= __alloc_pages_high_priority(gfp_mask
, order
, ac
);
3001 /* Atomic allocations - we can't balance anything */
3004 * All existing users of the deprecated __GFP_NOFAIL are
3005 * blockable, so warn of any new users that actually allow this
3006 * type of allocation to fail.
3008 WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
);
3012 /* Avoid recursion of direct reclaim */
3013 if (current
->flags
& PF_MEMALLOC
)
3016 /* Avoid allocations with no watermarks from looping endlessly */
3017 if (test_thread_flag(TIF_MEMDIE
) && !(gfp_mask
& __GFP_NOFAIL
))
3021 * Try direct compaction. The first pass is asynchronous. Subsequent
3022 * attempts after direct reclaim are synchronous
3024 page
= __alloc_pages_direct_compact(gfp_mask
, order
, alloc_flags
, ac
,
3026 &contended_compaction
,
3027 &deferred_compaction
);
3031 /* Checks for THP-specific high-order allocations */
3032 if ((gfp_mask
& GFP_TRANSHUGE
) == GFP_TRANSHUGE
) {
3034 * If compaction is deferred for high-order allocations, it is
3035 * because sync compaction recently failed. If this is the case
3036 * and the caller requested a THP allocation, we do not want
3037 * to heavily disrupt the system, so we fail the allocation
3038 * instead of entering direct reclaim.
3040 if (deferred_compaction
)
3044 * In all zones where compaction was attempted (and not
3045 * deferred or skipped), lock contention has been detected.
3046 * For THP allocation we do not want to disrupt the others
3047 * so we fallback to base pages instead.
3049 if (contended_compaction
== COMPACT_CONTENDED_LOCK
)
3053 * If compaction was aborted due to need_resched(), we do not
3054 * want to further increase allocation latency, unless it is
3055 * khugepaged trying to collapse.
3057 if (contended_compaction
== COMPACT_CONTENDED_SCHED
3058 && !(current
->flags
& PF_KTHREAD
))
3063 * It can become very expensive to allocate transparent hugepages at
3064 * fault, so use asynchronous memory compaction for THP unless it is
3065 * khugepaged trying to collapse.
3067 if ((gfp_mask
& GFP_TRANSHUGE
) != GFP_TRANSHUGE
||
3068 (current
->flags
& PF_KTHREAD
))
3069 migration_mode
= MIGRATE_SYNC_LIGHT
;
3071 /* Try direct reclaim and then allocating */
3072 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
, alloc_flags
, ac
,
3073 &did_some_progress
);
3077 /* Do not loop if specifically requested */
3078 if (gfp_mask
& __GFP_NORETRY
)
3081 /* Keep reclaiming pages as long as there is reasonable progress */
3082 pages_reclaimed
+= did_some_progress
;
3083 if ((did_some_progress
&& order
<= PAGE_ALLOC_COSTLY_ORDER
) ||
3084 ((gfp_mask
& __GFP_REPEAT
) && pages_reclaimed
< (1 << order
))) {
3085 /* Wait for some write requests to complete then retry */
3086 wait_iff_congested(ac
->preferred_zone
, BLK_RW_ASYNC
, HZ
/50);
3090 /* Reclaim has failed us, start killing things */
3091 page
= __alloc_pages_may_oom(gfp_mask
, order
, ac
, &did_some_progress
);
3095 /* Retry as long as the OOM killer is making progress */
3096 if (did_some_progress
)
3101 * High-order allocations do not necessarily loop after
3102 * direct reclaim and reclaim/compaction depends on compaction
3103 * being called after reclaim so call directly if necessary
3105 page
= __alloc_pages_direct_compact(gfp_mask
, order
, alloc_flags
,
3107 &contended_compaction
,
3108 &deferred_compaction
);
3112 warn_alloc_failed(gfp_mask
, order
, NULL
);
3118 * This is the 'heart' of the zoned buddy allocator.
3121 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
,
3122 struct zonelist
*zonelist
, nodemask_t
*nodemask
)
3124 struct zoneref
*preferred_zoneref
;
3125 struct page
*page
= NULL
;
3126 unsigned int cpuset_mems_cookie
;
3127 int alloc_flags
= ALLOC_WMARK_LOW
|ALLOC_CPUSET
|ALLOC_FAIR
;
3128 gfp_t alloc_mask
; /* The gfp_t that was actually used for allocation */
3129 struct alloc_context ac
= {
3130 .high_zoneidx
= gfp_zone(gfp_mask
),
3131 .nodemask
= nodemask
,
3132 .migratetype
= gfpflags_to_migratetype(gfp_mask
),
3135 gfp_mask
&= gfp_allowed_mask
;
3137 lockdep_trace_alloc(gfp_mask
);
3139 might_sleep_if(gfp_mask
& __GFP_WAIT
);
3141 if (should_fail_alloc_page(gfp_mask
, order
))
3145 * Check the zones suitable for the gfp_mask contain at least one
3146 * valid zone. It's possible to have an empty zonelist as a result
3147 * of __GFP_THISNODE and a memoryless node
3149 if (unlikely(!zonelist
->_zonerefs
->zone
))
3152 if (IS_ENABLED(CONFIG_CMA
) && ac
.migratetype
== MIGRATE_MOVABLE
)
3153 alloc_flags
|= ALLOC_CMA
;
3156 cpuset_mems_cookie
= read_mems_allowed_begin();
3158 /* We set it here, as __alloc_pages_slowpath might have changed it */
3159 ac
.zonelist
= zonelist
;
3160 /* The preferred zone is used for statistics later */
3161 preferred_zoneref
= first_zones_zonelist(ac
.zonelist
, ac
.high_zoneidx
,
3162 ac
.nodemask
? : &cpuset_current_mems_allowed
,
3163 &ac
.preferred_zone
);
3164 if (!ac
.preferred_zone
)
3166 ac
.classzone_idx
= zonelist_zone_idx(preferred_zoneref
);
3168 /* First allocation attempt */
3169 alloc_mask
= gfp_mask
|__GFP_HARDWALL
;
3170 page
= get_page_from_freelist(alloc_mask
, order
, alloc_flags
, &ac
);
3171 if (unlikely(!page
)) {
3173 * Runtime PM, block IO and its error handling path
3174 * can deadlock because I/O on the device might not
3177 alloc_mask
= memalloc_noio_flags(gfp_mask
);
3179 page
= __alloc_pages_slowpath(alloc_mask
, order
, &ac
);
3182 if (kmemcheck_enabled
&& page
)
3183 kmemcheck_pagealloc_alloc(page
, order
, gfp_mask
);
3185 trace_mm_page_alloc(page
, order
, alloc_mask
, ac
.migratetype
);
3189 * When updating a task's mems_allowed, it is possible to race with
3190 * parallel threads in such a way that an allocation can fail while
3191 * the mask is being updated. If a page allocation is about to fail,
3192 * check if the cpuset changed during allocation and if so, retry.
3194 if (unlikely(!page
&& read_mems_allowed_retry(cpuset_mems_cookie
)))
3199 EXPORT_SYMBOL(__alloc_pages_nodemask
);
3202 * Common helper functions.
3204 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
3209 * __get_free_pages() returns a 32-bit address, which cannot represent
3212 VM_BUG_ON((gfp_mask
& __GFP_HIGHMEM
) != 0);
3214 page
= alloc_pages(gfp_mask
, order
);
3217 return (unsigned long) page_address(page
);
3219 EXPORT_SYMBOL(__get_free_pages
);
3221 unsigned long get_zeroed_page(gfp_t gfp_mask
)
3223 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
3225 EXPORT_SYMBOL(get_zeroed_page
);
3227 void __free_pages(struct page
*page
, unsigned int order
)
3229 if (put_page_testzero(page
)) {
3231 free_hot_cold_page(page
, false);
3233 __free_pages_ok(page
, order
);
3237 EXPORT_SYMBOL(__free_pages
);
3239 void free_pages(unsigned long addr
, unsigned int order
)
3242 VM_BUG_ON(!virt_addr_valid((void *)addr
));
3243 __free_pages(virt_to_page((void *)addr
), order
);
3247 EXPORT_SYMBOL(free_pages
);
3251 * An arbitrary-length arbitrary-offset area of memory which resides
3252 * within a 0 or higher order page. Multiple fragments within that page
3253 * are individually refcounted, in the page's reference counter.
3255 * The page_frag functions below provide a simple allocation framework for
3256 * page fragments. This is used by the network stack and network device
3257 * drivers to provide a backing region of memory for use as either an
3258 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3260 static struct page
*__page_frag_refill(struct page_frag_cache
*nc
,
3263 struct page
*page
= NULL
;
3264 gfp_t gfp
= gfp_mask
;
3266 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3267 gfp_mask
|= __GFP_COMP
| __GFP_NOWARN
| __GFP_NORETRY
|
3269 page
= alloc_pages_node(NUMA_NO_NODE
, gfp_mask
,
3270 PAGE_FRAG_CACHE_MAX_ORDER
);
3271 nc
->size
= page
? PAGE_FRAG_CACHE_MAX_SIZE
: PAGE_SIZE
;
3273 if (unlikely(!page
))
3274 page
= alloc_pages_node(NUMA_NO_NODE
, gfp
, 0);
3276 nc
->va
= page
? page_address(page
) : NULL
;
3281 void *__alloc_page_frag(struct page_frag_cache
*nc
,
3282 unsigned int fragsz
, gfp_t gfp_mask
)
3284 unsigned int size
= PAGE_SIZE
;
3288 if (unlikely(!nc
->va
)) {
3290 page
= __page_frag_refill(nc
, gfp_mask
);
3294 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3295 /* if size can vary use size else just use PAGE_SIZE */
3298 /* Even if we own the page, we do not use atomic_set().
3299 * This would break get_page_unless_zero() users.
3301 atomic_add(size
- 1, &page
->_count
);
3303 /* reset page count bias and offset to start of new frag */
3304 nc
->pfmemalloc
= page
->pfmemalloc
;
3305 nc
->pagecnt_bias
= size
;
3309 offset
= nc
->offset
- fragsz
;
3310 if (unlikely(offset
< 0)) {
3311 page
= virt_to_page(nc
->va
);
3313 if (!atomic_sub_and_test(nc
->pagecnt_bias
, &page
->_count
))
3316 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3317 /* if size can vary use size else just use PAGE_SIZE */
3320 /* OK, page count is 0, we can safely set it */
3321 atomic_set(&page
->_count
, size
);
3323 /* reset page count bias and offset to start of new frag */
3324 nc
->pagecnt_bias
= size
;
3325 offset
= size
- fragsz
;
3329 nc
->offset
= offset
;
3331 return nc
->va
+ offset
;
3333 EXPORT_SYMBOL(__alloc_page_frag
);
3336 * Frees a page fragment allocated out of either a compound or order 0 page.
3338 void __free_page_frag(void *addr
)
3340 struct page
*page
= virt_to_head_page(addr
);
3342 if (unlikely(put_page_testzero(page
)))
3343 __free_pages_ok(page
, compound_order(page
));
3345 EXPORT_SYMBOL(__free_page_frag
);
3348 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
3349 * of the current memory cgroup.
3351 * It should be used when the caller would like to use kmalloc, but since the
3352 * allocation is large, it has to fall back to the page allocator.
3354 struct page
*alloc_kmem_pages(gfp_t gfp_mask
, unsigned int order
)
3357 struct mem_cgroup
*memcg
= NULL
;
3359 if (!memcg_kmem_newpage_charge(gfp_mask
, &memcg
, order
))
3361 page
= alloc_pages(gfp_mask
, order
);
3362 memcg_kmem_commit_charge(page
, memcg
, order
);
3366 struct page
*alloc_kmem_pages_node(int nid
, gfp_t gfp_mask
, unsigned int order
)
3369 struct mem_cgroup
*memcg
= NULL
;
3371 if (!memcg_kmem_newpage_charge(gfp_mask
, &memcg
, order
))
3373 page
= alloc_pages_node(nid
, gfp_mask
, order
);
3374 memcg_kmem_commit_charge(page
, memcg
, order
);
3379 * __free_kmem_pages and free_kmem_pages will free pages allocated with
3382 void __free_kmem_pages(struct page
*page
, unsigned int order
)
3384 memcg_kmem_uncharge_pages(page
, order
);
3385 __free_pages(page
, order
);
3388 void free_kmem_pages(unsigned long addr
, unsigned int order
)
3391 VM_BUG_ON(!virt_addr_valid((void *)addr
));
3392 __free_kmem_pages(virt_to_page((void *)addr
), order
);
3396 static void *make_alloc_exact(unsigned long addr
, unsigned order
, size_t size
)
3399 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
3400 unsigned long used
= addr
+ PAGE_ALIGN(size
);
3402 split_page(virt_to_page((void *)addr
), order
);
3403 while (used
< alloc_end
) {
3408 return (void *)addr
;
3412 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
3413 * @size: the number of bytes to allocate
3414 * @gfp_mask: GFP flags for the allocation
3416 * This function is similar to alloc_pages(), except that it allocates the
3417 * minimum number of pages to satisfy the request. alloc_pages() can only
3418 * allocate memory in power-of-two pages.
3420 * This function is also limited by MAX_ORDER.
3422 * Memory allocated by this function must be released by free_pages_exact().
3424 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
3426 unsigned int order
= get_order(size
);
3429 addr
= __get_free_pages(gfp_mask
, order
);
3430 return make_alloc_exact(addr
, order
, size
);
3432 EXPORT_SYMBOL(alloc_pages_exact
);
3435 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3437 * @nid: the preferred node ID where memory should be allocated
3438 * @size: the number of bytes to allocate
3439 * @gfp_mask: GFP flags for the allocation
3441 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3443 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
3446 void * __meminit
alloc_pages_exact_nid(int nid
, size_t size
, gfp_t gfp_mask
)
3448 unsigned order
= get_order(size
);
3449 struct page
*p
= alloc_pages_node(nid
, gfp_mask
, order
);
3452 return make_alloc_exact((unsigned long)page_address(p
), order
, size
);
3456 * free_pages_exact - release memory allocated via alloc_pages_exact()
3457 * @virt: the value returned by alloc_pages_exact.
3458 * @size: size of allocation, same value as passed to alloc_pages_exact().
3460 * Release the memory allocated by a previous call to alloc_pages_exact.
3462 void free_pages_exact(void *virt
, size_t size
)
3464 unsigned long addr
= (unsigned long)virt
;
3465 unsigned long end
= addr
+ PAGE_ALIGN(size
);
3467 while (addr
< end
) {
3472 EXPORT_SYMBOL(free_pages_exact
);
3475 * nr_free_zone_pages - count number of pages beyond high watermark
3476 * @offset: The zone index of the highest zone
3478 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3479 * high watermark within all zones at or below a given zone index. For each
3480 * zone, the number of pages is calculated as:
3481 * managed_pages - high_pages
3483 static unsigned long nr_free_zone_pages(int offset
)
3488 /* Just pick one node, since fallback list is circular */
3489 unsigned long sum
= 0;
3491 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
3493 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
3494 unsigned long size
= zone
->managed_pages
;
3495 unsigned long high
= high_wmark_pages(zone
);
3504 * nr_free_buffer_pages - count number of pages beyond high watermark
3506 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3507 * watermark within ZONE_DMA and ZONE_NORMAL.
3509 unsigned long nr_free_buffer_pages(void)
3511 return nr_free_zone_pages(gfp_zone(GFP_USER
));
3513 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
3516 * nr_free_pagecache_pages - count number of pages beyond high watermark
3518 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3519 * high watermark within all zones.
3521 unsigned long nr_free_pagecache_pages(void)
3523 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
3526 static inline void show_node(struct zone
*zone
)
3528 if (IS_ENABLED(CONFIG_NUMA
))
3529 printk("Node %d ", zone_to_nid(zone
));
3532 void si_meminfo(struct sysinfo
*val
)
3534 val
->totalram
= totalram_pages
;
3535 val
->sharedram
= global_page_state(NR_SHMEM
);
3536 val
->freeram
= global_page_state(NR_FREE_PAGES
);
3537 val
->bufferram
= nr_blockdev_pages();
3538 val
->totalhigh
= totalhigh_pages
;
3539 val
->freehigh
= nr_free_highpages();
3540 val
->mem_unit
= PAGE_SIZE
;
3543 EXPORT_SYMBOL(si_meminfo
);
3546 void si_meminfo_node(struct sysinfo
*val
, int nid
)
3548 int zone_type
; /* needs to be signed */
3549 unsigned long managed_pages
= 0;
3550 pg_data_t
*pgdat
= NODE_DATA(nid
);
3552 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++)
3553 managed_pages
+= pgdat
->node_zones
[zone_type
].managed_pages
;
3554 val
->totalram
= managed_pages
;
3555 val
->sharedram
= node_page_state(nid
, NR_SHMEM
);
3556 val
->freeram
= node_page_state(nid
, NR_FREE_PAGES
);
3557 #ifdef CONFIG_HIGHMEM
3558 val
->totalhigh
= pgdat
->node_zones
[ZONE_HIGHMEM
].managed_pages
;
3559 val
->freehigh
= zone_page_state(&pgdat
->node_zones
[ZONE_HIGHMEM
],
3565 val
->mem_unit
= PAGE_SIZE
;
3570 * Determine whether the node should be displayed or not, depending on whether
3571 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3573 bool skip_free_areas_node(unsigned int flags
, int nid
)
3576 unsigned int cpuset_mems_cookie
;
3578 if (!(flags
& SHOW_MEM_FILTER_NODES
))
3582 cpuset_mems_cookie
= read_mems_allowed_begin();
3583 ret
= !node_isset(nid
, cpuset_current_mems_allowed
);
3584 } while (read_mems_allowed_retry(cpuset_mems_cookie
));
3589 #define K(x) ((x) << (PAGE_SHIFT-10))
3591 static void show_migration_types(unsigned char type
)
3593 static const char types
[MIGRATE_TYPES
] = {
3594 [MIGRATE_UNMOVABLE
] = 'U',
3595 [MIGRATE_RECLAIMABLE
] = 'E',
3596 [MIGRATE_MOVABLE
] = 'M',
3597 [MIGRATE_RESERVE
] = 'R',
3599 [MIGRATE_CMA
] = 'C',
3601 #ifdef CONFIG_MEMORY_ISOLATION
3602 [MIGRATE_ISOLATE
] = 'I',
3605 char tmp
[MIGRATE_TYPES
+ 1];
3609 for (i
= 0; i
< MIGRATE_TYPES
; i
++) {
3610 if (type
& (1 << i
))
3615 printk("(%s) ", tmp
);
3619 * Show free area list (used inside shift_scroll-lock stuff)
3620 * We also calculate the percentage fragmentation. We do this by counting the
3621 * memory on each free list with the exception of the first item on the list.
3624 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
3627 void show_free_areas(unsigned int filter
)
3629 unsigned long free_pcp
= 0;
3633 for_each_populated_zone(zone
) {
3634 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3637 for_each_online_cpu(cpu
)
3638 free_pcp
+= per_cpu_ptr(zone
->pageset
, cpu
)->pcp
.count
;
3641 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3642 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3643 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
3644 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3645 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3646 " free:%lu free_pcp:%lu free_cma:%lu\n",
3647 global_page_state(NR_ACTIVE_ANON
),
3648 global_page_state(NR_INACTIVE_ANON
),
3649 global_page_state(NR_ISOLATED_ANON
),
3650 global_page_state(NR_ACTIVE_FILE
),
3651 global_page_state(NR_INACTIVE_FILE
),
3652 global_page_state(NR_ISOLATED_FILE
),
3653 global_page_state(NR_UNEVICTABLE
),
3654 global_page_state(NR_FILE_DIRTY
),
3655 global_page_state(NR_WRITEBACK
),
3656 global_page_state(NR_UNSTABLE_NFS
),
3657 global_page_state(NR_SLAB_RECLAIMABLE
),
3658 global_page_state(NR_SLAB_UNRECLAIMABLE
),
3659 global_page_state(NR_FILE_MAPPED
),
3660 global_page_state(NR_SHMEM
),
3661 global_page_state(NR_PAGETABLE
),
3662 global_page_state(NR_BOUNCE
),
3663 global_page_state(NR_FREE_PAGES
),
3665 global_page_state(NR_FREE_CMA_PAGES
));
3667 for_each_populated_zone(zone
) {
3670 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3674 for_each_online_cpu(cpu
)
3675 free_pcp
+= per_cpu_ptr(zone
->pageset
, cpu
)->pcp
.count
;
3683 " active_anon:%lukB"
3684 " inactive_anon:%lukB"
3685 " active_file:%lukB"
3686 " inactive_file:%lukB"
3687 " unevictable:%lukB"
3688 " isolated(anon):%lukB"
3689 " isolated(file):%lukB"
3697 " slab_reclaimable:%lukB"
3698 " slab_unreclaimable:%lukB"
3699 " kernel_stack:%lukB"
3706 " writeback_tmp:%lukB"
3707 " pages_scanned:%lu"
3708 " all_unreclaimable? %s"
3711 K(zone_page_state(zone
, NR_FREE_PAGES
)),
3712 K(min_wmark_pages(zone
)),
3713 K(low_wmark_pages(zone
)),
3714 K(high_wmark_pages(zone
)),
3715 K(zone_page_state(zone
, NR_ACTIVE_ANON
)),
3716 K(zone_page_state(zone
, NR_INACTIVE_ANON
)),
3717 K(zone_page_state(zone
, NR_ACTIVE_FILE
)),
3718 K(zone_page_state(zone
, NR_INACTIVE_FILE
)),
3719 K(zone_page_state(zone
, NR_UNEVICTABLE
)),
3720 K(zone_page_state(zone
, NR_ISOLATED_ANON
)),
3721 K(zone_page_state(zone
, NR_ISOLATED_FILE
)),
3722 K(zone
->present_pages
),
3723 K(zone
->managed_pages
),
3724 K(zone_page_state(zone
, NR_MLOCK
)),
3725 K(zone_page_state(zone
, NR_FILE_DIRTY
)),
3726 K(zone_page_state(zone
, NR_WRITEBACK
)),
3727 K(zone_page_state(zone
, NR_FILE_MAPPED
)),
3728 K(zone_page_state(zone
, NR_SHMEM
)),
3729 K(zone_page_state(zone
, NR_SLAB_RECLAIMABLE
)),
3730 K(zone_page_state(zone
, NR_SLAB_UNRECLAIMABLE
)),
3731 zone_page_state(zone
, NR_KERNEL_STACK
) *
3733 K(zone_page_state(zone
, NR_PAGETABLE
)),
3734 K(zone_page_state(zone
, NR_UNSTABLE_NFS
)),
3735 K(zone_page_state(zone
, NR_BOUNCE
)),
3737 K(this_cpu_read(zone
->pageset
->pcp
.count
)),
3738 K(zone_page_state(zone
, NR_FREE_CMA_PAGES
)),
3739 K(zone_page_state(zone
, NR_WRITEBACK_TEMP
)),
3740 K(zone_page_state(zone
, NR_PAGES_SCANNED
)),
3741 (!zone_reclaimable(zone
) ? "yes" : "no")
3743 printk("lowmem_reserve[]:");
3744 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
3745 printk(" %ld", zone
->lowmem_reserve
[i
]);
3749 for_each_populated_zone(zone
) {
3750 unsigned long nr
[MAX_ORDER
], flags
, order
, total
= 0;
3751 unsigned char types
[MAX_ORDER
];
3753 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3756 printk("%s: ", zone
->name
);
3758 spin_lock_irqsave(&zone
->lock
, flags
);
3759 for (order
= 0; order
< MAX_ORDER
; order
++) {
3760 struct free_area
*area
= &zone
->free_area
[order
];
3763 nr
[order
] = area
->nr_free
;
3764 total
+= nr
[order
] << order
;
3767 for (type
= 0; type
< MIGRATE_TYPES
; type
++) {
3768 if (!list_empty(&area
->free_list
[type
]))
3769 types
[order
] |= 1 << type
;
3772 spin_unlock_irqrestore(&zone
->lock
, flags
);
3773 for (order
= 0; order
< MAX_ORDER
; order
++) {
3774 printk("%lu*%lukB ", nr
[order
], K(1UL) << order
);
3776 show_migration_types(types
[order
]);
3778 printk("= %lukB\n", K(total
));
3781 hugetlb_show_meminfo();
3783 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES
));
3785 show_swap_cache_info();
3788 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
3790 zoneref
->zone
= zone
;
3791 zoneref
->zone_idx
= zone_idx(zone
);
3795 * Builds allocation fallback zone lists.
3797 * Add all populated zones of a node to the zonelist.
3799 static int build_zonelists_node(pg_data_t
*pgdat
, struct zonelist
*zonelist
,
3803 enum zone_type zone_type
= MAX_NR_ZONES
;
3807 zone
= pgdat
->node_zones
+ zone_type
;
3808 if (populated_zone(zone
)) {
3809 zoneref_set_zone(zone
,
3810 &zonelist
->_zonerefs
[nr_zones
++]);
3811 check_highest_zone(zone_type
);
3813 } while (zone_type
);
3821 * 0 = automatic detection of better ordering.
3822 * 1 = order by ([node] distance, -zonetype)
3823 * 2 = order by (-zonetype, [node] distance)
3825 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3826 * the same zonelist. So only NUMA can configure this param.
3828 #define ZONELIST_ORDER_DEFAULT 0
3829 #define ZONELIST_ORDER_NODE 1
3830 #define ZONELIST_ORDER_ZONE 2
3832 /* zonelist order in the kernel.
3833 * set_zonelist_order() will set this to NODE or ZONE.
3835 static int current_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3836 static char zonelist_order_name
[3][8] = {"Default", "Node", "Zone"};
3840 /* The value user specified ....changed by config */
3841 static int user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3842 /* string for sysctl */
3843 #define NUMA_ZONELIST_ORDER_LEN 16
3844 char numa_zonelist_order
[16] = "default";
3847 * interface for configure zonelist ordering.
3848 * command line option "numa_zonelist_order"
3849 * = "[dD]efault - default, automatic configuration.
3850 * = "[nN]ode - order by node locality, then by zone within node
3851 * = "[zZ]one - order by zone, then by locality within zone
3854 static int __parse_numa_zonelist_order(char *s
)
3856 if (*s
== 'd' || *s
== 'D') {
3857 user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3858 } else if (*s
== 'n' || *s
== 'N') {
3859 user_zonelist_order
= ZONELIST_ORDER_NODE
;
3860 } else if (*s
== 'z' || *s
== 'Z') {
3861 user_zonelist_order
= ZONELIST_ORDER_ZONE
;
3864 "Ignoring invalid numa_zonelist_order value: "
3871 static __init
int setup_numa_zonelist_order(char *s
)
3878 ret
= __parse_numa_zonelist_order(s
);
3880 strlcpy(numa_zonelist_order
, s
, NUMA_ZONELIST_ORDER_LEN
);
3884 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
3887 * sysctl handler for numa_zonelist_order
3889 int numa_zonelist_order_handler(struct ctl_table
*table
, int write
,
3890 void __user
*buffer
, size_t *length
,
3893 char saved_string
[NUMA_ZONELIST_ORDER_LEN
];
3895 static DEFINE_MUTEX(zl_order_mutex
);
3897 mutex_lock(&zl_order_mutex
);
3899 if (strlen((char *)table
->data
) >= NUMA_ZONELIST_ORDER_LEN
) {
3903 strcpy(saved_string
, (char *)table
->data
);
3905 ret
= proc_dostring(table
, write
, buffer
, length
, ppos
);
3909 int oldval
= user_zonelist_order
;
3911 ret
= __parse_numa_zonelist_order((char *)table
->data
);
3914 * bogus value. restore saved string
3916 strncpy((char *)table
->data
, saved_string
,
3917 NUMA_ZONELIST_ORDER_LEN
);
3918 user_zonelist_order
= oldval
;
3919 } else if (oldval
!= user_zonelist_order
) {
3920 mutex_lock(&zonelists_mutex
);
3921 build_all_zonelists(NULL
, NULL
);
3922 mutex_unlock(&zonelists_mutex
);
3926 mutex_unlock(&zl_order_mutex
);
3931 #define MAX_NODE_LOAD (nr_online_nodes)
3932 static int node_load
[MAX_NUMNODES
];
3935 * find_next_best_node - find the next node that should appear in a given node's fallback list
3936 * @node: node whose fallback list we're appending
3937 * @used_node_mask: nodemask_t of already used nodes
3939 * We use a number of factors to determine which is the next node that should
3940 * appear on a given node's fallback list. The node should not have appeared
3941 * already in @node's fallback list, and it should be the next closest node
3942 * according to the distance array (which contains arbitrary distance values
3943 * from each node to each node in the system), and should also prefer nodes
3944 * with no CPUs, since presumably they'll have very little allocation pressure
3945 * on them otherwise.
3946 * It returns -1 if no node is found.
3948 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
3951 int min_val
= INT_MAX
;
3952 int best_node
= NUMA_NO_NODE
;
3953 const struct cpumask
*tmp
= cpumask_of_node(0);
3955 /* Use the local node if we haven't already */
3956 if (!node_isset(node
, *used_node_mask
)) {
3957 node_set(node
, *used_node_mask
);
3961 for_each_node_state(n
, N_MEMORY
) {
3963 /* Don't want a node to appear more than once */
3964 if (node_isset(n
, *used_node_mask
))
3967 /* Use the distance array to find the distance */
3968 val
= node_distance(node
, n
);
3970 /* Penalize nodes under us ("prefer the next node") */
3973 /* Give preference to headless and unused nodes */
3974 tmp
= cpumask_of_node(n
);
3975 if (!cpumask_empty(tmp
))
3976 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
3978 /* Slight preference for less loaded node */
3979 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
3980 val
+= node_load
[n
];
3982 if (val
< min_val
) {
3989 node_set(best_node
, *used_node_mask
);
3996 * Build zonelists ordered by node and zones within node.
3997 * This results in maximum locality--normal zone overflows into local
3998 * DMA zone, if any--but risks exhausting DMA zone.
4000 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int node
)
4003 struct zonelist
*zonelist
;
4005 zonelist
= &pgdat
->node_zonelists
[0];
4006 for (j
= 0; zonelist
->_zonerefs
[j
].zone
!= NULL
; j
++)
4008 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
4009 zonelist
->_zonerefs
[j
].zone
= NULL
;
4010 zonelist
->_zonerefs
[j
].zone_idx
= 0;
4014 * Build gfp_thisnode zonelists
4016 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
4019 struct zonelist
*zonelist
;
4021 zonelist
= &pgdat
->node_zonelists
[1];
4022 j
= build_zonelists_node(pgdat
, zonelist
, 0);
4023 zonelist
->_zonerefs
[j
].zone
= NULL
;
4024 zonelist
->_zonerefs
[j
].zone_idx
= 0;
4028 * Build zonelists ordered by zone and nodes within zones.
4029 * This results in conserving DMA zone[s] until all Normal memory is
4030 * exhausted, but results in overflowing to remote node while memory
4031 * may still exist in local DMA zone.
4033 static int node_order
[MAX_NUMNODES
];
4035 static void build_zonelists_in_zone_order(pg_data_t
*pgdat
, int nr_nodes
)
4038 int zone_type
; /* needs to be signed */
4040 struct zonelist
*zonelist
;
4042 zonelist
= &pgdat
->node_zonelists
[0];
4044 for (zone_type
= MAX_NR_ZONES
- 1; zone_type
>= 0; zone_type
--) {
4045 for (j
= 0; j
< nr_nodes
; j
++) {
4046 node
= node_order
[j
];
4047 z
= &NODE_DATA(node
)->node_zones
[zone_type
];
4048 if (populated_zone(z
)) {
4050 &zonelist
->_zonerefs
[pos
++]);
4051 check_highest_zone(zone_type
);
4055 zonelist
->_zonerefs
[pos
].zone
= NULL
;
4056 zonelist
->_zonerefs
[pos
].zone_idx
= 0;
4059 #if defined(CONFIG_64BIT)
4061 * Devices that require DMA32/DMA are relatively rare and do not justify a
4062 * penalty to every machine in case the specialised case applies. Default
4063 * to Node-ordering on 64-bit NUMA machines
4065 static int default_zonelist_order(void)
4067 return ZONELIST_ORDER_NODE
;
4071 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4072 * by the kernel. If processes running on node 0 deplete the low memory zone
4073 * then reclaim will occur more frequency increasing stalls and potentially
4074 * be easier to OOM if a large percentage of the zone is under writeback or
4075 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4076 * Hence, default to zone ordering on 32-bit.
4078 static int default_zonelist_order(void)
4080 return ZONELIST_ORDER_ZONE
;
4082 #endif /* CONFIG_64BIT */
4084 static void set_zonelist_order(void)
4086 if (user_zonelist_order
== ZONELIST_ORDER_DEFAULT
)
4087 current_zonelist_order
= default_zonelist_order();
4089 current_zonelist_order
= user_zonelist_order
;
4092 static void build_zonelists(pg_data_t
*pgdat
)
4096 nodemask_t used_mask
;
4097 int local_node
, prev_node
;
4098 struct zonelist
*zonelist
;
4099 int order
= current_zonelist_order
;
4101 /* initialize zonelists */
4102 for (i
= 0; i
< MAX_ZONELISTS
; i
++) {
4103 zonelist
= pgdat
->node_zonelists
+ i
;
4104 zonelist
->_zonerefs
[0].zone
= NULL
;
4105 zonelist
->_zonerefs
[0].zone_idx
= 0;
4108 /* NUMA-aware ordering of nodes */
4109 local_node
= pgdat
->node_id
;
4110 load
= nr_online_nodes
;
4111 prev_node
= local_node
;
4112 nodes_clear(used_mask
);
4114 memset(node_order
, 0, sizeof(node_order
));
4117 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
4119 * We don't want to pressure a particular node.
4120 * So adding penalty to the first node in same
4121 * distance group to make it round-robin.
4123 if (node_distance(local_node
, node
) !=
4124 node_distance(local_node
, prev_node
))
4125 node_load
[node
] = load
;
4129 if (order
== ZONELIST_ORDER_NODE
)
4130 build_zonelists_in_node_order(pgdat
, node
);
4132 node_order
[j
++] = node
; /* remember order */
4135 if (order
== ZONELIST_ORDER_ZONE
) {
4136 /* calculate node order -- i.e., DMA last! */
4137 build_zonelists_in_zone_order(pgdat
, j
);
4140 build_thisnode_zonelists(pgdat
);
4143 /* Construct the zonelist performance cache - see further mmzone.h */
4144 static void build_zonelist_cache(pg_data_t
*pgdat
)
4146 struct zonelist
*zonelist
;
4147 struct zonelist_cache
*zlc
;
4150 zonelist
= &pgdat
->node_zonelists
[0];
4151 zonelist
->zlcache_ptr
= zlc
= &zonelist
->zlcache
;
4152 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
4153 for (z
= zonelist
->_zonerefs
; z
->zone
; z
++)
4154 zlc
->z_to_n
[z
- zonelist
->_zonerefs
] = zonelist_node_idx(z
);
4157 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4159 * Return node id of node used for "local" allocations.
4160 * I.e., first node id of first zone in arg node's generic zonelist.
4161 * Used for initializing percpu 'numa_mem', which is used primarily
4162 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4164 int local_memory_node(int node
)
4168 (void)first_zones_zonelist(node_zonelist(node
, GFP_KERNEL
),
4169 gfp_zone(GFP_KERNEL
),
4176 #else /* CONFIG_NUMA */
4178 static void set_zonelist_order(void)
4180 current_zonelist_order
= ZONELIST_ORDER_ZONE
;
4183 static void build_zonelists(pg_data_t
*pgdat
)
4185 int node
, local_node
;
4187 struct zonelist
*zonelist
;
4189 local_node
= pgdat
->node_id
;
4191 zonelist
= &pgdat
->node_zonelists
[0];
4192 j
= build_zonelists_node(pgdat
, zonelist
, 0);
4195 * Now we build the zonelist so that it contains the zones
4196 * of all the other nodes.
4197 * We don't want to pressure a particular node, so when
4198 * building the zones for node N, we make sure that the
4199 * zones coming right after the local ones are those from
4200 * node N+1 (modulo N)
4202 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
4203 if (!node_online(node
))
4205 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
4207 for (node
= 0; node
< local_node
; node
++) {
4208 if (!node_online(node
))
4210 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
4213 zonelist
->_zonerefs
[j
].zone
= NULL
;
4214 zonelist
->_zonerefs
[j
].zone_idx
= 0;
4217 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
4218 static void build_zonelist_cache(pg_data_t
*pgdat
)
4220 pgdat
->node_zonelists
[0].zlcache_ptr
= NULL
;
4223 #endif /* CONFIG_NUMA */
4226 * Boot pageset table. One per cpu which is going to be used for all
4227 * zones and all nodes. The parameters will be set in such a way
4228 * that an item put on a list will immediately be handed over to
4229 * the buddy list. This is safe since pageset manipulation is done
4230 * with interrupts disabled.
4232 * The boot_pagesets must be kept even after bootup is complete for
4233 * unused processors and/or zones. They do play a role for bootstrapping
4234 * hotplugged processors.
4236 * zoneinfo_show() and maybe other functions do
4237 * not check if the processor is online before following the pageset pointer.
4238 * Other parts of the kernel may not check if the zone is available.
4240 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
);
4241 static DEFINE_PER_CPU(struct per_cpu_pageset
, boot_pageset
);
4242 static void setup_zone_pageset(struct zone
*zone
);
4245 * Global mutex to protect against size modification of zonelists
4246 * as well as to serialize pageset setup for the new populated zone.
4248 DEFINE_MUTEX(zonelists_mutex
);
4250 /* return values int ....just for stop_machine() */
4251 static int __build_all_zonelists(void *data
)
4255 pg_data_t
*self
= data
;
4258 memset(node_load
, 0, sizeof(node_load
));
4261 if (self
&& !node_online(self
->node_id
)) {
4262 build_zonelists(self
);
4263 build_zonelist_cache(self
);
4266 for_each_online_node(nid
) {
4267 pg_data_t
*pgdat
= NODE_DATA(nid
);
4269 build_zonelists(pgdat
);
4270 build_zonelist_cache(pgdat
);
4274 * Initialize the boot_pagesets that are going to be used
4275 * for bootstrapping processors. The real pagesets for
4276 * each zone will be allocated later when the per cpu
4277 * allocator is available.
4279 * boot_pagesets are used also for bootstrapping offline
4280 * cpus if the system is already booted because the pagesets
4281 * are needed to initialize allocators on a specific cpu too.
4282 * F.e. the percpu allocator needs the page allocator which
4283 * needs the percpu allocator in order to allocate its pagesets
4284 * (a chicken-egg dilemma).
4286 for_each_possible_cpu(cpu
) {
4287 setup_pageset(&per_cpu(boot_pageset
, cpu
), 0);
4289 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4291 * We now know the "local memory node" for each node--
4292 * i.e., the node of the first zone in the generic zonelist.
4293 * Set up numa_mem percpu variable for on-line cpus. During
4294 * boot, only the boot cpu should be on-line; we'll init the
4295 * secondary cpus' numa_mem as they come on-line. During
4296 * node/memory hotplug, we'll fixup all on-line cpus.
4298 if (cpu_online(cpu
))
4299 set_cpu_numa_mem(cpu
, local_memory_node(cpu_to_node(cpu
)));
4306 static noinline
void __init
4307 build_all_zonelists_init(void)
4309 __build_all_zonelists(NULL
);
4310 mminit_verify_zonelist();
4311 cpuset_init_current_mems_allowed();
4315 * Called with zonelists_mutex held always
4316 * unless system_state == SYSTEM_BOOTING.
4318 * __ref due to (1) call of __meminit annotated setup_zone_pageset
4319 * [we're only called with non-NULL zone through __meminit paths] and
4320 * (2) call of __init annotated helper build_all_zonelists_init
4321 * [protected by SYSTEM_BOOTING].
4323 void __ref
build_all_zonelists(pg_data_t
*pgdat
, struct zone
*zone
)
4325 set_zonelist_order();
4327 if (system_state
== SYSTEM_BOOTING
) {
4328 build_all_zonelists_init();
4330 #ifdef CONFIG_MEMORY_HOTPLUG
4332 setup_zone_pageset(zone
);
4334 /* we have to stop all cpus to guarantee there is no user
4336 stop_machine(__build_all_zonelists
, pgdat
, NULL
);
4337 /* cpuset refresh routine should be here */
4339 vm_total_pages
= nr_free_pagecache_pages();
4341 * Disable grouping by mobility if the number of pages in the
4342 * system is too low to allow the mechanism to work. It would be
4343 * more accurate, but expensive to check per-zone. This check is
4344 * made on memory-hotadd so a system can start with mobility
4345 * disabled and enable it later
4347 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
4348 page_group_by_mobility_disabled
= 1;
4350 page_group_by_mobility_disabled
= 0;
4352 pr_info("Built %i zonelists in %s order, mobility grouping %s. "
4353 "Total pages: %ld\n",
4355 zonelist_order_name
[current_zonelist_order
],
4356 page_group_by_mobility_disabled
? "off" : "on",
4359 pr_info("Policy zone: %s\n", zone_names
[policy_zone
]);
4364 * Helper functions to size the waitqueue hash table.
4365 * Essentially these want to choose hash table sizes sufficiently
4366 * large so that collisions trying to wait on pages are rare.
4367 * But in fact, the number of active page waitqueues on typical
4368 * systems is ridiculously low, less than 200. So this is even
4369 * conservative, even though it seems large.
4371 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
4372 * waitqueues, i.e. the size of the waitq table given the number of pages.
4374 #define PAGES_PER_WAITQUEUE 256
4376 #ifndef CONFIG_MEMORY_HOTPLUG
4377 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
4379 unsigned long size
= 1;
4381 pages
/= PAGES_PER_WAITQUEUE
;
4383 while (size
< pages
)
4387 * Once we have dozens or even hundreds of threads sleeping
4388 * on IO we've got bigger problems than wait queue collision.
4389 * Limit the size of the wait table to a reasonable size.
4391 size
= min(size
, 4096UL);
4393 return max(size
, 4UL);
4397 * A zone's size might be changed by hot-add, so it is not possible to determine
4398 * a suitable size for its wait_table. So we use the maximum size now.
4400 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
4402 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
4403 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
4404 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
4406 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
4407 * or more by the traditional way. (See above). It equals:
4409 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
4410 * ia64(16K page size) : = ( 8G + 4M)byte.
4411 * powerpc (64K page size) : = (32G +16M)byte.
4413 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
4420 * This is an integer logarithm so that shifts can be used later
4421 * to extract the more random high bits from the multiplicative
4422 * hash function before the remainder is taken.
4424 static inline unsigned long wait_table_bits(unsigned long size
)
4430 * Check if a pageblock contains reserved pages
4432 static int pageblock_is_reserved(unsigned long start_pfn
, unsigned long end_pfn
)
4436 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
4437 if (!pfn_valid_within(pfn
) || PageReserved(pfn_to_page(pfn
)))
4444 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
4445 * of blocks reserved is based on min_wmark_pages(zone). The memory within
4446 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
4447 * higher will lead to a bigger reserve which will get freed as contiguous
4448 * blocks as reclaim kicks in
4450 static void setup_zone_migrate_reserve(struct zone
*zone
)
4452 unsigned long start_pfn
, pfn
, end_pfn
, block_end_pfn
;
4454 unsigned long block_migratetype
;
4459 * Get the start pfn, end pfn and the number of blocks to reserve
4460 * We have to be careful to be aligned to pageblock_nr_pages to
4461 * make sure that we always check pfn_valid for the first page in
4464 start_pfn
= zone
->zone_start_pfn
;
4465 end_pfn
= zone_end_pfn(zone
);
4466 start_pfn
= roundup(start_pfn
, pageblock_nr_pages
);
4467 reserve
= roundup(min_wmark_pages(zone
), pageblock_nr_pages
) >>
4471 * Reserve blocks are generally in place to help high-order atomic
4472 * allocations that are short-lived. A min_free_kbytes value that
4473 * would result in more than 2 reserve blocks for atomic allocations
4474 * is assumed to be in place to help anti-fragmentation for the
4475 * future allocation of hugepages at runtime.
4477 reserve
= min(2, reserve
);
4478 old_reserve
= zone
->nr_migrate_reserve_block
;
4480 /* When memory hot-add, we almost always need to do nothing */
4481 if (reserve
== old_reserve
)
4483 zone
->nr_migrate_reserve_block
= reserve
;
4485 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= pageblock_nr_pages
) {
4486 if (!early_page_nid_uninitialised(pfn
, zone_to_nid(zone
)))
4489 if (!pfn_valid(pfn
))
4491 page
= pfn_to_page(pfn
);
4493 /* Watch out for overlapping nodes */
4494 if (page_to_nid(page
) != zone_to_nid(zone
))
4497 block_migratetype
= get_pageblock_migratetype(page
);
4499 /* Only test what is necessary when the reserves are not met */
4502 * Blocks with reserved pages will never free, skip
4505 block_end_pfn
= min(pfn
+ pageblock_nr_pages
, end_pfn
);
4506 if (pageblock_is_reserved(pfn
, block_end_pfn
))
4509 /* If this block is reserved, account for it */
4510 if (block_migratetype
== MIGRATE_RESERVE
) {
4515 /* Suitable for reserving if this block is movable */
4516 if (block_migratetype
== MIGRATE_MOVABLE
) {
4517 set_pageblock_migratetype(page
,
4519 move_freepages_block(zone
, page
,
4524 } else if (!old_reserve
) {
4526 * At boot time we don't need to scan the whole zone
4527 * for turning off MIGRATE_RESERVE.
4533 * If the reserve is met and this is a previous reserved block,
4536 if (block_migratetype
== MIGRATE_RESERVE
) {
4537 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
4538 move_freepages_block(zone
, page
, MIGRATE_MOVABLE
);
4544 * Initially all pages are reserved - free ones are freed
4545 * up by free_all_bootmem() once the early boot process is
4546 * done. Non-atomic initialization, single-pass.
4548 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
4549 unsigned long start_pfn
, enum memmap_context context
)
4551 pg_data_t
*pgdat
= NODE_DATA(nid
);
4552 unsigned long end_pfn
= start_pfn
+ size
;
4555 unsigned long nr_initialised
= 0;
4557 if (highest_memmap_pfn
< end_pfn
- 1)
4558 highest_memmap_pfn
= end_pfn
- 1;
4560 z
= &pgdat
->node_zones
[zone
];
4561 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
4563 * There can be holes in boot-time mem_map[]s
4564 * handed to this function. They do not
4565 * exist on hotplugged memory.
4567 if (context
== MEMMAP_EARLY
) {
4568 if (!early_pfn_valid(pfn
))
4570 if (!early_pfn_in_nid(pfn
, nid
))
4572 if (!update_defer_init(pgdat
, pfn
, end_pfn
,
4578 * Mark the block movable so that blocks are reserved for
4579 * movable at startup. This will force kernel allocations
4580 * to reserve their blocks rather than leaking throughout
4581 * the address space during boot when many long-lived
4582 * kernel allocations are made. Later some blocks near
4583 * the start are marked MIGRATE_RESERVE by
4584 * setup_zone_migrate_reserve()
4586 * bitmap is created for zone's valid pfn range. but memmap
4587 * can be created for invalid pages (for alignment)
4588 * check here not to call set_pageblock_migratetype() against
4591 if (!(pfn
& (pageblock_nr_pages
- 1))) {
4592 struct page
*page
= pfn_to_page(pfn
);
4594 __init_single_page(page
, pfn
, zone
, nid
);
4595 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
4597 __init_single_pfn(pfn
, zone
, nid
);
4602 static void __meminit
zone_init_free_lists(struct zone
*zone
)
4604 unsigned int order
, t
;
4605 for_each_migratetype_order(order
, t
) {
4606 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
4607 zone
->free_area
[order
].nr_free
= 0;
4611 #ifndef __HAVE_ARCH_MEMMAP_INIT
4612 #define memmap_init(size, nid, zone, start_pfn) \
4613 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4616 static int zone_batchsize(struct zone
*zone
)
4622 * The per-cpu-pages pools are set to around 1000th of the
4623 * size of the zone. But no more than 1/2 of a meg.
4625 * OK, so we don't know how big the cache is. So guess.
4627 batch
= zone
->managed_pages
/ 1024;
4628 if (batch
* PAGE_SIZE
> 512 * 1024)
4629 batch
= (512 * 1024) / PAGE_SIZE
;
4630 batch
/= 4; /* We effectively *= 4 below */
4635 * Clamp the batch to a 2^n - 1 value. Having a power
4636 * of 2 value was found to be more likely to have
4637 * suboptimal cache aliasing properties in some cases.
4639 * For example if 2 tasks are alternately allocating
4640 * batches of pages, one task can end up with a lot
4641 * of pages of one half of the possible page colors
4642 * and the other with pages of the other colors.
4644 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
4649 /* The deferral and batching of frees should be suppressed under NOMMU
4652 * The problem is that NOMMU needs to be able to allocate large chunks
4653 * of contiguous memory as there's no hardware page translation to
4654 * assemble apparent contiguous memory from discontiguous pages.
4656 * Queueing large contiguous runs of pages for batching, however,
4657 * causes the pages to actually be freed in smaller chunks. As there
4658 * can be a significant delay between the individual batches being
4659 * recycled, this leads to the once large chunks of space being
4660 * fragmented and becoming unavailable for high-order allocations.
4667 * pcp->high and pcp->batch values are related and dependent on one another:
4668 * ->batch must never be higher then ->high.
4669 * The following function updates them in a safe manner without read side
4672 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4673 * those fields changing asynchronously (acording the the above rule).
4675 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4676 * outside of boot time (or some other assurance that no concurrent updaters
4679 static void pageset_update(struct per_cpu_pages
*pcp
, unsigned long high
,
4680 unsigned long batch
)
4682 /* start with a fail safe value for batch */
4686 /* Update high, then batch, in order */
4693 /* a companion to pageset_set_high() */
4694 static void pageset_set_batch(struct per_cpu_pageset
*p
, unsigned long batch
)
4696 pageset_update(&p
->pcp
, 6 * batch
, max(1UL, 1 * batch
));
4699 static void pageset_init(struct per_cpu_pageset
*p
)
4701 struct per_cpu_pages
*pcp
;
4704 memset(p
, 0, sizeof(*p
));
4708 for (migratetype
= 0; migratetype
< MIGRATE_PCPTYPES
; migratetype
++)
4709 INIT_LIST_HEAD(&pcp
->lists
[migratetype
]);
4712 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
4715 pageset_set_batch(p
, batch
);
4719 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4720 * to the value high for the pageset p.
4722 static void pageset_set_high(struct per_cpu_pageset
*p
,
4725 unsigned long batch
= max(1UL, high
/ 4);
4726 if ((high
/ 4) > (PAGE_SHIFT
* 8))
4727 batch
= PAGE_SHIFT
* 8;
4729 pageset_update(&p
->pcp
, high
, batch
);
4732 static void pageset_set_high_and_batch(struct zone
*zone
,
4733 struct per_cpu_pageset
*pcp
)
4735 if (percpu_pagelist_fraction
)
4736 pageset_set_high(pcp
,
4737 (zone
->managed_pages
/
4738 percpu_pagelist_fraction
));
4740 pageset_set_batch(pcp
, zone_batchsize(zone
));
4743 static void __meminit
zone_pageset_init(struct zone
*zone
, int cpu
)
4745 struct per_cpu_pageset
*pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
4748 pageset_set_high_and_batch(zone
, pcp
);
4751 static void __meminit
setup_zone_pageset(struct zone
*zone
)
4754 zone
->pageset
= alloc_percpu(struct per_cpu_pageset
);
4755 for_each_possible_cpu(cpu
)
4756 zone_pageset_init(zone
, cpu
);
4760 * Allocate per cpu pagesets and initialize them.
4761 * Before this call only boot pagesets were available.
4763 void __init
setup_per_cpu_pageset(void)
4767 for_each_populated_zone(zone
)
4768 setup_zone_pageset(zone
);
4771 static noinline __init_refok
4772 int zone_wait_table_init(struct zone
*zone
, unsigned long zone_size_pages
)
4778 * The per-page waitqueue mechanism uses hashed waitqueues
4781 zone
->wait_table_hash_nr_entries
=
4782 wait_table_hash_nr_entries(zone_size_pages
);
4783 zone
->wait_table_bits
=
4784 wait_table_bits(zone
->wait_table_hash_nr_entries
);
4785 alloc_size
= zone
->wait_table_hash_nr_entries
4786 * sizeof(wait_queue_head_t
);
4788 if (!slab_is_available()) {
4789 zone
->wait_table
= (wait_queue_head_t
*)
4790 memblock_virt_alloc_node_nopanic(
4791 alloc_size
, zone
->zone_pgdat
->node_id
);
4794 * This case means that a zone whose size was 0 gets new memory
4795 * via memory hot-add.
4796 * But it may be the case that a new node was hot-added. In
4797 * this case vmalloc() will not be able to use this new node's
4798 * memory - this wait_table must be initialized to use this new
4799 * node itself as well.
4800 * To use this new node's memory, further consideration will be
4803 zone
->wait_table
= vmalloc(alloc_size
);
4805 if (!zone
->wait_table
)
4808 for (i
= 0; i
< zone
->wait_table_hash_nr_entries
; ++i
)
4809 init_waitqueue_head(zone
->wait_table
+ i
);
4814 static __meminit
void zone_pcp_init(struct zone
*zone
)
4817 * per cpu subsystem is not up at this point. The following code
4818 * relies on the ability of the linker to provide the
4819 * offset of a (static) per cpu variable into the per cpu area.
4821 zone
->pageset
= &boot_pageset
;
4823 if (populated_zone(zone
))
4824 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%u\n",
4825 zone
->name
, zone
->present_pages
,
4826 zone_batchsize(zone
));
4829 int __meminit
init_currently_empty_zone(struct zone
*zone
,
4830 unsigned long zone_start_pfn
,
4832 enum memmap_context context
)
4834 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
4836 ret
= zone_wait_table_init(zone
, size
);
4839 pgdat
->nr_zones
= zone_idx(zone
) + 1;
4841 zone
->zone_start_pfn
= zone_start_pfn
;
4843 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
4844 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4846 (unsigned long)zone_idx(zone
),
4847 zone_start_pfn
, (zone_start_pfn
+ size
));
4849 zone_init_free_lists(zone
);
4854 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4855 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4858 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4860 int __meminit
__early_pfn_to_nid(unsigned long pfn
,
4861 struct mminit_pfnnid_cache
*state
)
4863 unsigned long start_pfn
, end_pfn
;
4866 if (state
->last_start
<= pfn
&& pfn
< state
->last_end
)
4867 return state
->last_nid
;
4869 nid
= memblock_search_pfn_nid(pfn
, &start_pfn
, &end_pfn
);
4871 state
->last_start
= start_pfn
;
4872 state
->last_end
= end_pfn
;
4873 state
->last_nid
= nid
;
4878 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4881 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4882 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4883 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4885 * If an architecture guarantees that all ranges registered contain no holes
4886 * and may be freed, this this function may be used instead of calling
4887 * memblock_free_early_nid() manually.
4889 void __init
free_bootmem_with_active_regions(int nid
, unsigned long max_low_pfn
)
4891 unsigned long start_pfn
, end_pfn
;
4894 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
) {
4895 start_pfn
= min(start_pfn
, max_low_pfn
);
4896 end_pfn
= min(end_pfn
, max_low_pfn
);
4898 if (start_pfn
< end_pfn
)
4899 memblock_free_early_nid(PFN_PHYS(start_pfn
),
4900 (end_pfn
- start_pfn
) << PAGE_SHIFT
,
4906 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4907 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4909 * If an architecture guarantees that all ranges registered contain no holes and may
4910 * be freed, this function may be used instead of calling memory_present() manually.
4912 void __init
sparse_memory_present_with_active_regions(int nid
)
4914 unsigned long start_pfn
, end_pfn
;
4917 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
)
4918 memory_present(this_nid
, start_pfn
, end_pfn
);
4922 * get_pfn_range_for_nid - Return the start and end page frames for a node
4923 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4924 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4925 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4927 * It returns the start and end page frame of a node based on information
4928 * provided by memblock_set_node(). If called for a node
4929 * with no available memory, a warning is printed and the start and end
4932 void __meminit
get_pfn_range_for_nid(unsigned int nid
,
4933 unsigned long *start_pfn
, unsigned long *end_pfn
)
4935 unsigned long this_start_pfn
, this_end_pfn
;
4941 for_each_mem_pfn_range(i
, nid
, &this_start_pfn
, &this_end_pfn
, NULL
) {
4942 *start_pfn
= min(*start_pfn
, this_start_pfn
);
4943 *end_pfn
= max(*end_pfn
, this_end_pfn
);
4946 if (*start_pfn
== -1UL)
4951 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4952 * assumption is made that zones within a node are ordered in monotonic
4953 * increasing memory addresses so that the "highest" populated zone is used
4955 static void __init
find_usable_zone_for_movable(void)
4958 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
4959 if (zone_index
== ZONE_MOVABLE
)
4962 if (arch_zone_highest_possible_pfn
[zone_index
] >
4963 arch_zone_lowest_possible_pfn
[zone_index
])
4967 VM_BUG_ON(zone_index
== -1);
4968 movable_zone
= zone_index
;
4972 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4973 * because it is sized independent of architecture. Unlike the other zones,
4974 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4975 * in each node depending on the size of each node and how evenly kernelcore
4976 * is distributed. This helper function adjusts the zone ranges
4977 * provided by the architecture for a given node by using the end of the
4978 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4979 * zones within a node are in order of monotonic increases memory addresses
4981 static void __meminit
adjust_zone_range_for_zone_movable(int nid
,
4982 unsigned long zone_type
,
4983 unsigned long node_start_pfn
,
4984 unsigned long node_end_pfn
,
4985 unsigned long *zone_start_pfn
,
4986 unsigned long *zone_end_pfn
)
4988 /* Only adjust if ZONE_MOVABLE is on this node */
4989 if (zone_movable_pfn
[nid
]) {
4990 /* Size ZONE_MOVABLE */
4991 if (zone_type
== ZONE_MOVABLE
) {
4992 *zone_start_pfn
= zone_movable_pfn
[nid
];
4993 *zone_end_pfn
= min(node_end_pfn
,
4994 arch_zone_highest_possible_pfn
[movable_zone
]);
4996 /* Adjust for ZONE_MOVABLE starting within this range */
4997 } else if (*zone_start_pfn
< zone_movable_pfn
[nid
] &&
4998 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
4999 *zone_end_pfn
= zone_movable_pfn
[nid
];
5001 /* Check if this whole range is within ZONE_MOVABLE */
5002 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
5003 *zone_start_pfn
= *zone_end_pfn
;
5008 * Return the number of pages a zone spans in a node, including holes
5009 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5011 static unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
5012 unsigned long zone_type
,
5013 unsigned long node_start_pfn
,
5014 unsigned long node_end_pfn
,
5015 unsigned long *ignored
)
5017 unsigned long zone_start_pfn
, zone_end_pfn
;
5019 /* Get the start and end of the zone */
5020 zone_start_pfn
= arch_zone_lowest_possible_pfn
[zone_type
];
5021 zone_end_pfn
= arch_zone_highest_possible_pfn
[zone_type
];
5022 adjust_zone_range_for_zone_movable(nid
, zone_type
,
5023 node_start_pfn
, node_end_pfn
,
5024 &zone_start_pfn
, &zone_end_pfn
);
5026 /* Check that this node has pages within the zone's required range */
5027 if (zone_end_pfn
< node_start_pfn
|| zone_start_pfn
> node_end_pfn
)
5030 /* Move the zone boundaries inside the node if necessary */
5031 zone_end_pfn
= min(zone_end_pfn
, node_end_pfn
);
5032 zone_start_pfn
= max(zone_start_pfn
, node_start_pfn
);
5034 /* Return the spanned pages */
5035 return zone_end_pfn
- zone_start_pfn
;
5039 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5040 * then all holes in the requested range will be accounted for.
5042 unsigned long __meminit
__absent_pages_in_range(int nid
,
5043 unsigned long range_start_pfn
,
5044 unsigned long range_end_pfn
)
5046 unsigned long nr_absent
= range_end_pfn
- range_start_pfn
;
5047 unsigned long start_pfn
, end_pfn
;
5050 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
5051 start_pfn
= clamp(start_pfn
, range_start_pfn
, range_end_pfn
);
5052 end_pfn
= clamp(end_pfn
, range_start_pfn
, range_end_pfn
);
5053 nr_absent
-= end_pfn
- start_pfn
;
5059 * absent_pages_in_range - Return number of page frames in holes within a range
5060 * @start_pfn: The start PFN to start searching for holes
5061 * @end_pfn: The end PFN to stop searching for holes
5063 * It returns the number of pages frames in memory holes within a range.
5065 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
5066 unsigned long end_pfn
)
5068 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
5071 /* Return the number of page frames in holes in a zone on a node */
5072 static unsigned long __meminit
zone_absent_pages_in_node(int nid
,
5073 unsigned long zone_type
,
5074 unsigned long node_start_pfn
,
5075 unsigned long node_end_pfn
,
5076 unsigned long *ignored
)
5078 unsigned long zone_low
= arch_zone_lowest_possible_pfn
[zone_type
];
5079 unsigned long zone_high
= arch_zone_highest_possible_pfn
[zone_type
];
5080 unsigned long zone_start_pfn
, zone_end_pfn
;
5082 zone_start_pfn
= clamp(node_start_pfn
, zone_low
, zone_high
);
5083 zone_end_pfn
= clamp(node_end_pfn
, zone_low
, zone_high
);
5085 adjust_zone_range_for_zone_movable(nid
, zone_type
,
5086 node_start_pfn
, node_end_pfn
,
5087 &zone_start_pfn
, &zone_end_pfn
);
5088 return __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
5091 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5092 static inline unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
5093 unsigned long zone_type
,
5094 unsigned long node_start_pfn
,
5095 unsigned long node_end_pfn
,
5096 unsigned long *zones_size
)
5098 return zones_size
[zone_type
];
5101 static inline unsigned long __meminit
zone_absent_pages_in_node(int nid
,
5102 unsigned long zone_type
,
5103 unsigned long node_start_pfn
,
5104 unsigned long node_end_pfn
,
5105 unsigned long *zholes_size
)
5110 return zholes_size
[zone_type
];
5113 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5115 static void __meminit
calculate_node_totalpages(struct pglist_data
*pgdat
,
5116 unsigned long node_start_pfn
,
5117 unsigned long node_end_pfn
,
5118 unsigned long *zones_size
,
5119 unsigned long *zholes_size
)
5121 unsigned long realtotalpages
= 0, totalpages
= 0;
5124 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5125 struct zone
*zone
= pgdat
->node_zones
+ i
;
5126 unsigned long size
, real_size
;
5128 size
= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
5132 real_size
= size
- zone_absent_pages_in_node(pgdat
->node_id
, i
,
5133 node_start_pfn
, node_end_pfn
,
5135 zone
->spanned_pages
= size
;
5136 zone
->present_pages
= real_size
;
5139 realtotalpages
+= real_size
;
5142 pgdat
->node_spanned_pages
= totalpages
;
5143 pgdat
->node_present_pages
= realtotalpages
;
5144 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
5148 #ifndef CONFIG_SPARSEMEM
5150 * Calculate the size of the zone->blockflags rounded to an unsigned long
5151 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5152 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5153 * round what is now in bits to nearest long in bits, then return it in
5156 static unsigned long __init
usemap_size(unsigned long zone_start_pfn
, unsigned long zonesize
)
5158 unsigned long usemapsize
;
5160 zonesize
+= zone_start_pfn
& (pageblock_nr_pages
-1);
5161 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
5162 usemapsize
= usemapsize
>> pageblock_order
;
5163 usemapsize
*= NR_PAGEBLOCK_BITS
;
5164 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
5166 return usemapsize
/ 8;
5169 static void __init
setup_usemap(struct pglist_data
*pgdat
,
5171 unsigned long zone_start_pfn
,
5172 unsigned long zonesize
)
5174 unsigned long usemapsize
= usemap_size(zone_start_pfn
, zonesize
);
5175 zone
->pageblock_flags
= NULL
;
5177 zone
->pageblock_flags
=
5178 memblock_virt_alloc_node_nopanic(usemapsize
,
5182 static inline void setup_usemap(struct pglist_data
*pgdat
, struct zone
*zone
,
5183 unsigned long zone_start_pfn
, unsigned long zonesize
) {}
5184 #endif /* CONFIG_SPARSEMEM */
5186 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5188 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5189 void __paginginit
set_pageblock_order(void)
5193 /* Check that pageblock_nr_pages has not already been setup */
5194 if (pageblock_order
)
5197 if (HPAGE_SHIFT
> PAGE_SHIFT
)
5198 order
= HUGETLB_PAGE_ORDER
;
5200 order
= MAX_ORDER
- 1;
5203 * Assume the largest contiguous order of interest is a huge page.
5204 * This value may be variable depending on boot parameters on IA64 and
5207 pageblock_order
= order
;
5209 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5212 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5213 * is unused as pageblock_order is set at compile-time. See
5214 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5217 void __paginginit
set_pageblock_order(void)
5221 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5223 static unsigned long __paginginit
calc_memmap_size(unsigned long spanned_pages
,
5224 unsigned long present_pages
)
5226 unsigned long pages
= spanned_pages
;
5229 * Provide a more accurate estimation if there are holes within
5230 * the zone and SPARSEMEM is in use. If there are holes within the
5231 * zone, each populated memory region may cost us one or two extra
5232 * memmap pages due to alignment because memmap pages for each
5233 * populated regions may not naturally algined on page boundary.
5234 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5236 if (spanned_pages
> present_pages
+ (present_pages
>> 4) &&
5237 IS_ENABLED(CONFIG_SPARSEMEM
))
5238 pages
= present_pages
;
5240 return PAGE_ALIGN(pages
* sizeof(struct page
)) >> PAGE_SHIFT
;
5244 * Set up the zone data structures:
5245 * - mark all pages reserved
5246 * - mark all memory queues empty
5247 * - clear the memory bitmaps
5249 * NOTE: pgdat should get zeroed by caller.
5251 static void __paginginit
free_area_init_core(struct pglist_data
*pgdat
,
5252 unsigned long node_start_pfn
, unsigned long node_end_pfn
)
5255 int nid
= pgdat
->node_id
;
5256 unsigned long zone_start_pfn
= pgdat
->node_start_pfn
;
5259 pgdat_resize_init(pgdat
);
5260 #ifdef CONFIG_NUMA_BALANCING
5261 spin_lock_init(&pgdat
->numabalancing_migrate_lock
);
5262 pgdat
->numabalancing_migrate_nr_pages
= 0;
5263 pgdat
->numabalancing_migrate_next_window
= jiffies
;
5265 init_waitqueue_head(&pgdat
->kswapd_wait
);
5266 init_waitqueue_head(&pgdat
->pfmemalloc_wait
);
5267 pgdat_page_ext_init(pgdat
);
5269 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
5270 struct zone
*zone
= pgdat
->node_zones
+ j
;
5271 unsigned long size
, realsize
, freesize
, memmap_pages
;
5273 size
= zone
->spanned_pages
;
5274 realsize
= freesize
= zone
->present_pages
;
5277 * Adjust freesize so that it accounts for how much memory
5278 * is used by this zone for memmap. This affects the watermark
5279 * and per-cpu initialisations
5281 memmap_pages
= calc_memmap_size(size
, realsize
);
5282 if (!is_highmem_idx(j
)) {
5283 if (freesize
>= memmap_pages
) {
5284 freesize
-= memmap_pages
;
5287 " %s zone: %lu pages used for memmap\n",
5288 zone_names
[j
], memmap_pages
);
5291 " %s zone: %lu pages exceeds freesize %lu\n",
5292 zone_names
[j
], memmap_pages
, freesize
);
5295 /* Account for reserved pages */
5296 if (j
== 0 && freesize
> dma_reserve
) {
5297 freesize
-= dma_reserve
;
5298 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
5299 zone_names
[0], dma_reserve
);
5302 if (!is_highmem_idx(j
))
5303 nr_kernel_pages
+= freesize
;
5304 /* Charge for highmem memmap if there are enough kernel pages */
5305 else if (nr_kernel_pages
> memmap_pages
* 2)
5306 nr_kernel_pages
-= memmap_pages
;
5307 nr_all_pages
+= freesize
;
5310 * Set an approximate value for lowmem here, it will be adjusted
5311 * when the bootmem allocator frees pages into the buddy system.
5312 * And all highmem pages will be managed by the buddy system.
5314 zone
->managed_pages
= is_highmem_idx(j
) ? realsize
: freesize
;
5317 zone
->min_unmapped_pages
= (freesize
*sysctl_min_unmapped_ratio
)
5319 zone
->min_slab_pages
= (freesize
* sysctl_min_slab_ratio
) / 100;
5321 zone
->name
= zone_names
[j
];
5322 spin_lock_init(&zone
->lock
);
5323 spin_lock_init(&zone
->lru_lock
);
5324 zone_seqlock_init(zone
);
5325 zone
->zone_pgdat
= pgdat
;
5326 zone_pcp_init(zone
);
5328 /* For bootup, initialized properly in watermark setup */
5329 mod_zone_page_state(zone
, NR_ALLOC_BATCH
, zone
->managed_pages
);
5331 lruvec_init(&zone
->lruvec
);
5335 set_pageblock_order();
5336 setup_usemap(pgdat
, zone
, zone_start_pfn
, size
);
5337 ret
= init_currently_empty_zone(zone
, zone_start_pfn
,
5338 size
, MEMMAP_EARLY
);
5340 memmap_init(size
, nid
, j
, zone_start_pfn
);
5341 zone_start_pfn
+= size
;
5345 static void __init_refok
alloc_node_mem_map(struct pglist_data
*pgdat
)
5347 /* Skip empty nodes */
5348 if (!pgdat
->node_spanned_pages
)
5351 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5352 /* ia64 gets its own node_mem_map, before this, without bootmem */
5353 if (!pgdat
->node_mem_map
) {
5354 unsigned long size
, start
, end
;
5358 * The zone's endpoints aren't required to be MAX_ORDER
5359 * aligned but the node_mem_map endpoints must be in order
5360 * for the buddy allocator to function correctly.
5362 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
5363 end
= pgdat_end_pfn(pgdat
);
5364 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
5365 size
= (end
- start
) * sizeof(struct page
);
5366 map
= alloc_remap(pgdat
->node_id
, size
);
5368 map
= memblock_virt_alloc_node_nopanic(size
,
5370 pgdat
->node_mem_map
= map
+ (pgdat
->node_start_pfn
- start
);
5372 #ifndef CONFIG_NEED_MULTIPLE_NODES
5374 * With no DISCONTIG, the global mem_map is just set as node 0's
5376 if (pgdat
== NODE_DATA(0)) {
5377 mem_map
= NODE_DATA(0)->node_mem_map
;
5378 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5379 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
5380 mem_map
-= (pgdat
->node_start_pfn
- ARCH_PFN_OFFSET
);
5381 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5384 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
5387 void __paginginit
free_area_init_node(int nid
, unsigned long *zones_size
,
5388 unsigned long node_start_pfn
, unsigned long *zholes_size
)
5390 pg_data_t
*pgdat
= NODE_DATA(nid
);
5391 unsigned long start_pfn
= 0;
5392 unsigned long end_pfn
= 0;
5394 /* pg_data_t should be reset to zero when it's allocated */
5395 WARN_ON(pgdat
->nr_zones
|| pgdat
->classzone_idx
);
5397 reset_deferred_meminit(pgdat
);
5398 pgdat
->node_id
= nid
;
5399 pgdat
->node_start_pfn
= node_start_pfn
;
5400 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5401 get_pfn_range_for_nid(nid
, &start_pfn
, &end_pfn
);
5402 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid
,
5403 (u64
)start_pfn
<< PAGE_SHIFT
, ((u64
)end_pfn
<< PAGE_SHIFT
) - 1);
5405 calculate_node_totalpages(pgdat
, start_pfn
, end_pfn
,
5406 zones_size
, zholes_size
);
5408 alloc_node_mem_map(pgdat
);
5409 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5410 printk(KERN_DEBUG
"free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5411 nid
, (unsigned long)pgdat
,
5412 (unsigned long)pgdat
->node_mem_map
);
5415 free_area_init_core(pgdat
, start_pfn
, end_pfn
);
5418 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5420 #if MAX_NUMNODES > 1
5422 * Figure out the number of possible node ids.
5424 void __init
setup_nr_node_ids(void)
5427 unsigned int highest
= 0;
5429 for_each_node_mask(node
, node_possible_map
)
5431 nr_node_ids
= highest
+ 1;
5436 * node_map_pfn_alignment - determine the maximum internode alignment
5438 * This function should be called after node map is populated and sorted.
5439 * It calculates the maximum power of two alignment which can distinguish
5442 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5443 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5444 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5445 * shifted, 1GiB is enough and this function will indicate so.
5447 * This is used to test whether pfn -> nid mapping of the chosen memory
5448 * model has fine enough granularity to avoid incorrect mapping for the
5449 * populated node map.
5451 * Returns the determined alignment in pfn's. 0 if there is no alignment
5452 * requirement (single node).
5454 unsigned long __init
node_map_pfn_alignment(void)
5456 unsigned long accl_mask
= 0, last_end
= 0;
5457 unsigned long start
, end
, mask
;
5461 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start
, &end
, &nid
) {
5462 if (!start
|| last_nid
< 0 || last_nid
== nid
) {
5469 * Start with a mask granular enough to pin-point to the
5470 * start pfn and tick off bits one-by-one until it becomes
5471 * too coarse to separate the current node from the last.
5473 mask
= ~((1 << __ffs(start
)) - 1);
5474 while (mask
&& last_end
<= (start
& (mask
<< 1)))
5477 /* accumulate all internode masks */
5481 /* convert mask to number of pages */
5482 return ~accl_mask
+ 1;
5485 /* Find the lowest pfn for a node */
5486 static unsigned long __init
find_min_pfn_for_node(int nid
)
5488 unsigned long min_pfn
= ULONG_MAX
;
5489 unsigned long start_pfn
;
5492 for_each_mem_pfn_range(i
, nid
, &start_pfn
, NULL
, NULL
)
5493 min_pfn
= min(min_pfn
, start_pfn
);
5495 if (min_pfn
== ULONG_MAX
) {
5497 "Could not find start_pfn for node %d\n", nid
);
5505 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5507 * It returns the minimum PFN based on information provided via
5508 * memblock_set_node().
5510 unsigned long __init
find_min_pfn_with_active_regions(void)
5512 return find_min_pfn_for_node(MAX_NUMNODES
);
5516 * early_calculate_totalpages()
5517 * Sum pages in active regions for movable zone.
5518 * Populate N_MEMORY for calculating usable_nodes.
5520 static unsigned long __init
early_calculate_totalpages(void)
5522 unsigned long totalpages
= 0;
5523 unsigned long start_pfn
, end_pfn
;
5526 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
5527 unsigned long pages
= end_pfn
- start_pfn
;
5529 totalpages
+= pages
;
5531 node_set_state(nid
, N_MEMORY
);
5537 * Find the PFN the Movable zone begins in each node. Kernel memory
5538 * is spread evenly between nodes as long as the nodes have enough
5539 * memory. When they don't, some nodes will have more kernelcore than
5542 static void __init
find_zone_movable_pfns_for_nodes(void)
5545 unsigned long usable_startpfn
;
5546 unsigned long kernelcore_node
, kernelcore_remaining
;
5547 /* save the state before borrow the nodemask */
5548 nodemask_t saved_node_state
= node_states
[N_MEMORY
];
5549 unsigned long totalpages
= early_calculate_totalpages();
5550 int usable_nodes
= nodes_weight(node_states
[N_MEMORY
]);
5551 struct memblock_region
*r
;
5553 /* Need to find movable_zone earlier when movable_node is specified. */
5554 find_usable_zone_for_movable();
5557 * If movable_node is specified, ignore kernelcore and movablecore
5560 if (movable_node_is_enabled()) {
5561 for_each_memblock(memory
, r
) {
5562 if (!memblock_is_hotpluggable(r
))
5567 usable_startpfn
= PFN_DOWN(r
->base
);
5568 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
5569 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
5577 * If movablecore=nn[KMG] was specified, calculate what size of
5578 * kernelcore that corresponds so that memory usable for
5579 * any allocation type is evenly spread. If both kernelcore
5580 * and movablecore are specified, then the value of kernelcore
5581 * will be used for required_kernelcore if it's greater than
5582 * what movablecore would have allowed.
5584 if (required_movablecore
) {
5585 unsigned long corepages
;
5588 * Round-up so that ZONE_MOVABLE is at least as large as what
5589 * was requested by the user
5591 required_movablecore
=
5592 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
5593 corepages
= totalpages
- required_movablecore
;
5595 required_kernelcore
= max(required_kernelcore
, corepages
);
5598 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
5599 if (!required_kernelcore
)
5602 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5603 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
5606 /* Spread kernelcore memory as evenly as possible throughout nodes */
5607 kernelcore_node
= required_kernelcore
/ usable_nodes
;
5608 for_each_node_state(nid
, N_MEMORY
) {
5609 unsigned long start_pfn
, end_pfn
;
5612 * Recalculate kernelcore_node if the division per node
5613 * now exceeds what is necessary to satisfy the requested
5614 * amount of memory for the kernel
5616 if (required_kernelcore
< kernelcore_node
)
5617 kernelcore_node
= required_kernelcore
/ usable_nodes
;
5620 * As the map is walked, we track how much memory is usable
5621 * by the kernel using kernelcore_remaining. When it is
5622 * 0, the rest of the node is usable by ZONE_MOVABLE
5624 kernelcore_remaining
= kernelcore_node
;
5626 /* Go through each range of PFNs within this node */
5627 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
5628 unsigned long size_pages
;
5630 start_pfn
= max(start_pfn
, zone_movable_pfn
[nid
]);
5631 if (start_pfn
>= end_pfn
)
5634 /* Account for what is only usable for kernelcore */
5635 if (start_pfn
< usable_startpfn
) {
5636 unsigned long kernel_pages
;
5637 kernel_pages
= min(end_pfn
, usable_startpfn
)
5640 kernelcore_remaining
-= min(kernel_pages
,
5641 kernelcore_remaining
);
5642 required_kernelcore
-= min(kernel_pages
,
5643 required_kernelcore
);
5645 /* Continue if range is now fully accounted */
5646 if (end_pfn
<= usable_startpfn
) {
5649 * Push zone_movable_pfn to the end so
5650 * that if we have to rebalance
5651 * kernelcore across nodes, we will
5652 * not double account here
5654 zone_movable_pfn
[nid
] = end_pfn
;
5657 start_pfn
= usable_startpfn
;
5661 * The usable PFN range for ZONE_MOVABLE is from
5662 * start_pfn->end_pfn. Calculate size_pages as the
5663 * number of pages used as kernelcore
5665 size_pages
= end_pfn
- start_pfn
;
5666 if (size_pages
> kernelcore_remaining
)
5667 size_pages
= kernelcore_remaining
;
5668 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
5671 * Some kernelcore has been met, update counts and
5672 * break if the kernelcore for this node has been
5675 required_kernelcore
-= min(required_kernelcore
,
5677 kernelcore_remaining
-= size_pages
;
5678 if (!kernelcore_remaining
)
5684 * If there is still required_kernelcore, we do another pass with one
5685 * less node in the count. This will push zone_movable_pfn[nid] further
5686 * along on the nodes that still have memory until kernelcore is
5690 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
5694 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5695 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
5696 zone_movable_pfn
[nid
] =
5697 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
5700 /* restore the node_state */
5701 node_states
[N_MEMORY
] = saved_node_state
;
5704 /* Any regular or high memory on that node ? */
5705 static void check_for_memory(pg_data_t
*pgdat
, int nid
)
5707 enum zone_type zone_type
;
5709 if (N_MEMORY
== N_NORMAL_MEMORY
)
5712 for (zone_type
= 0; zone_type
<= ZONE_MOVABLE
- 1; zone_type
++) {
5713 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
5714 if (populated_zone(zone
)) {
5715 node_set_state(nid
, N_HIGH_MEMORY
);
5716 if (N_NORMAL_MEMORY
!= N_HIGH_MEMORY
&&
5717 zone_type
<= ZONE_NORMAL
)
5718 node_set_state(nid
, N_NORMAL_MEMORY
);
5725 * free_area_init_nodes - Initialise all pg_data_t and zone data
5726 * @max_zone_pfn: an array of max PFNs for each zone
5728 * This will call free_area_init_node() for each active node in the system.
5729 * Using the page ranges provided by memblock_set_node(), the size of each
5730 * zone in each node and their holes is calculated. If the maximum PFN
5731 * between two adjacent zones match, it is assumed that the zone is empty.
5732 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5733 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5734 * starts where the previous one ended. For example, ZONE_DMA32 starts
5735 * at arch_max_dma_pfn.
5737 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
5739 unsigned long start_pfn
, end_pfn
;
5742 /* Record where the zone boundaries are */
5743 memset(arch_zone_lowest_possible_pfn
, 0,
5744 sizeof(arch_zone_lowest_possible_pfn
));
5745 memset(arch_zone_highest_possible_pfn
, 0,
5746 sizeof(arch_zone_highest_possible_pfn
));
5747 arch_zone_lowest_possible_pfn
[0] = find_min_pfn_with_active_regions();
5748 arch_zone_highest_possible_pfn
[0] = max_zone_pfn
[0];
5749 for (i
= 1; i
< MAX_NR_ZONES
; i
++) {
5750 if (i
== ZONE_MOVABLE
)
5752 arch_zone_lowest_possible_pfn
[i
] =
5753 arch_zone_highest_possible_pfn
[i
-1];
5754 arch_zone_highest_possible_pfn
[i
] =
5755 max(max_zone_pfn
[i
], arch_zone_lowest_possible_pfn
[i
]);
5757 arch_zone_lowest_possible_pfn
[ZONE_MOVABLE
] = 0;
5758 arch_zone_highest_possible_pfn
[ZONE_MOVABLE
] = 0;
5760 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5761 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
5762 find_zone_movable_pfns_for_nodes();
5764 /* Print out the zone ranges */
5765 pr_info("Zone ranges:\n");
5766 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5767 if (i
== ZONE_MOVABLE
)
5769 pr_info(" %-8s ", zone_names
[i
]);
5770 if (arch_zone_lowest_possible_pfn
[i
] ==
5771 arch_zone_highest_possible_pfn
[i
])
5774 pr_cont("[mem %#018Lx-%#018Lx]\n",
5775 (u64
)arch_zone_lowest_possible_pfn
[i
]
5777 ((u64
)arch_zone_highest_possible_pfn
[i
]
5778 << PAGE_SHIFT
) - 1);
5781 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5782 pr_info("Movable zone start for each node\n");
5783 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5784 if (zone_movable_pfn
[i
])
5785 pr_info(" Node %d: %#018Lx\n", i
,
5786 (u64
)zone_movable_pfn
[i
] << PAGE_SHIFT
);
5789 /* Print out the early node map */
5790 pr_info("Early memory node ranges\n");
5791 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
)
5792 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid
,
5793 (u64
)start_pfn
<< PAGE_SHIFT
,
5794 ((u64
)end_pfn
<< PAGE_SHIFT
) - 1);
5796 /* Initialise every node */
5797 mminit_verify_pageflags_layout();
5798 setup_nr_node_ids();
5799 for_each_online_node(nid
) {
5800 pg_data_t
*pgdat
= NODE_DATA(nid
);
5801 free_area_init_node(nid
, NULL
,
5802 find_min_pfn_for_node(nid
), NULL
);
5804 /* Any memory on that node */
5805 if (pgdat
->node_present_pages
)
5806 node_set_state(nid
, N_MEMORY
);
5807 check_for_memory(pgdat
, nid
);
5811 static int __init
cmdline_parse_core(char *p
, unsigned long *core
)
5813 unsigned long long coremem
;
5817 coremem
= memparse(p
, &p
);
5818 *core
= coremem
>> PAGE_SHIFT
;
5820 /* Paranoid check that UL is enough for the coremem value */
5821 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
5827 * kernelcore=size sets the amount of memory for use for allocations that
5828 * cannot be reclaimed or migrated.
5830 static int __init
cmdline_parse_kernelcore(char *p
)
5832 return cmdline_parse_core(p
, &required_kernelcore
);
5836 * movablecore=size sets the amount of memory for use for allocations that
5837 * can be reclaimed or migrated.
5839 static int __init
cmdline_parse_movablecore(char *p
)
5841 return cmdline_parse_core(p
, &required_movablecore
);
5844 early_param("kernelcore", cmdline_parse_kernelcore
);
5845 early_param("movablecore", cmdline_parse_movablecore
);
5847 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5849 void adjust_managed_page_count(struct page
*page
, long count
)
5851 spin_lock(&managed_page_count_lock
);
5852 page_zone(page
)->managed_pages
+= count
;
5853 totalram_pages
+= count
;
5854 #ifdef CONFIG_HIGHMEM
5855 if (PageHighMem(page
))
5856 totalhigh_pages
+= count
;
5858 spin_unlock(&managed_page_count_lock
);
5860 EXPORT_SYMBOL(adjust_managed_page_count
);
5862 unsigned long free_reserved_area(void *start
, void *end
, int poison
, char *s
)
5865 unsigned long pages
= 0;
5867 start
= (void *)PAGE_ALIGN((unsigned long)start
);
5868 end
= (void *)((unsigned long)end
& PAGE_MASK
);
5869 for (pos
= start
; pos
< end
; pos
+= PAGE_SIZE
, pages
++) {
5870 if ((unsigned int)poison
<= 0xFF)
5871 memset(pos
, poison
, PAGE_SIZE
);
5872 free_reserved_page(virt_to_page(pos
));
5876 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5877 s
, pages
<< (PAGE_SHIFT
- 10), start
, end
);
5881 EXPORT_SYMBOL(free_reserved_area
);
5883 #ifdef CONFIG_HIGHMEM
5884 void free_highmem_page(struct page
*page
)
5886 __free_reserved_page(page
);
5888 page_zone(page
)->managed_pages
++;
5894 void __init
mem_init_print_info(const char *str
)
5896 unsigned long physpages
, codesize
, datasize
, rosize
, bss_size
;
5897 unsigned long init_code_size
, init_data_size
;
5899 physpages
= get_num_physpages();
5900 codesize
= _etext
- _stext
;
5901 datasize
= _edata
- _sdata
;
5902 rosize
= __end_rodata
- __start_rodata
;
5903 bss_size
= __bss_stop
- __bss_start
;
5904 init_data_size
= __init_end
- __init_begin
;
5905 init_code_size
= _einittext
- _sinittext
;
5908 * Detect special cases and adjust section sizes accordingly:
5909 * 1) .init.* may be embedded into .data sections
5910 * 2) .init.text.* may be out of [__init_begin, __init_end],
5911 * please refer to arch/tile/kernel/vmlinux.lds.S.
5912 * 3) .rodata.* may be embedded into .text or .data sections.
5914 #define adj_init_size(start, end, size, pos, adj) \
5916 if (start <= pos && pos < end && size > adj) \
5920 adj_init_size(__init_begin
, __init_end
, init_data_size
,
5921 _sinittext
, init_code_size
);
5922 adj_init_size(_stext
, _etext
, codesize
, _sinittext
, init_code_size
);
5923 adj_init_size(_sdata
, _edata
, datasize
, __init_begin
, init_data_size
);
5924 adj_init_size(_stext
, _etext
, codesize
, __start_rodata
, rosize
);
5925 adj_init_size(_sdata
, _edata
, datasize
, __start_rodata
, rosize
);
5927 #undef adj_init_size
5929 pr_info("Memory: %luK/%luK available "
5930 "(%luK kernel code, %luK rwdata, %luK rodata, "
5931 "%luK init, %luK bss, %luK reserved, %luK cma-reserved"
5932 #ifdef CONFIG_HIGHMEM
5936 nr_free_pages() << (PAGE_SHIFT
-10), physpages
<< (PAGE_SHIFT
-10),
5937 codesize
>> 10, datasize
>> 10, rosize
>> 10,
5938 (init_data_size
+ init_code_size
) >> 10, bss_size
>> 10,
5939 (physpages
- totalram_pages
- totalcma_pages
) << (PAGE_SHIFT
-10),
5940 totalcma_pages
<< (PAGE_SHIFT
-10),
5941 #ifdef CONFIG_HIGHMEM
5942 totalhigh_pages
<< (PAGE_SHIFT
-10),
5944 str
? ", " : "", str
? str
: "");
5948 * set_dma_reserve - set the specified number of pages reserved in the first zone
5949 * @new_dma_reserve: The number of pages to mark reserved
5951 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5952 * In the DMA zone, a significant percentage may be consumed by kernel image
5953 * and other unfreeable allocations which can skew the watermarks badly. This
5954 * function may optionally be used to account for unfreeable pages in the
5955 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5956 * smaller per-cpu batchsize.
5958 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
5960 dma_reserve
= new_dma_reserve
;
5963 void __init
free_area_init(unsigned long *zones_size
)
5965 free_area_init_node(0, zones_size
,
5966 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
5969 static int page_alloc_cpu_notify(struct notifier_block
*self
,
5970 unsigned long action
, void *hcpu
)
5972 int cpu
= (unsigned long)hcpu
;
5974 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
) {
5975 lru_add_drain_cpu(cpu
);
5979 * Spill the event counters of the dead processor
5980 * into the current processors event counters.
5981 * This artificially elevates the count of the current
5984 vm_events_fold_cpu(cpu
);
5987 * Zero the differential counters of the dead processor
5988 * so that the vm statistics are consistent.
5990 * This is only okay since the processor is dead and cannot
5991 * race with what we are doing.
5993 cpu_vm_stats_fold(cpu
);
5998 void __init
page_alloc_init(void)
6000 hotcpu_notifier(page_alloc_cpu_notify
, 0);
6004 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
6005 * or min_free_kbytes changes.
6007 static void calculate_totalreserve_pages(void)
6009 struct pglist_data
*pgdat
;
6010 unsigned long reserve_pages
= 0;
6011 enum zone_type i
, j
;
6013 for_each_online_pgdat(pgdat
) {
6014 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
6015 struct zone
*zone
= pgdat
->node_zones
+ i
;
6018 /* Find valid and maximum lowmem_reserve in the zone */
6019 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
6020 if (zone
->lowmem_reserve
[j
] > max
)
6021 max
= zone
->lowmem_reserve
[j
];
6024 /* we treat the high watermark as reserved pages. */
6025 max
+= high_wmark_pages(zone
);
6027 if (max
> zone
->managed_pages
)
6028 max
= zone
->managed_pages
;
6029 reserve_pages
+= max
;
6031 * Lowmem reserves are not available to
6032 * GFP_HIGHUSER page cache allocations and
6033 * kswapd tries to balance zones to their high
6034 * watermark. As a result, neither should be
6035 * regarded as dirtyable memory, to prevent a
6036 * situation where reclaim has to clean pages
6037 * in order to balance the zones.
6039 zone
->dirty_balance_reserve
= max
;
6042 dirty_balance_reserve
= reserve_pages
;
6043 totalreserve_pages
= reserve_pages
;
6047 * setup_per_zone_lowmem_reserve - called whenever
6048 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
6049 * has a correct pages reserved value, so an adequate number of
6050 * pages are left in the zone after a successful __alloc_pages().
6052 static void setup_per_zone_lowmem_reserve(void)
6054 struct pglist_data
*pgdat
;
6055 enum zone_type j
, idx
;
6057 for_each_online_pgdat(pgdat
) {
6058 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
6059 struct zone
*zone
= pgdat
->node_zones
+ j
;
6060 unsigned long managed_pages
= zone
->managed_pages
;
6062 zone
->lowmem_reserve
[j
] = 0;
6066 struct zone
*lower_zone
;
6070 if (sysctl_lowmem_reserve_ratio
[idx
] < 1)
6071 sysctl_lowmem_reserve_ratio
[idx
] = 1;
6073 lower_zone
= pgdat
->node_zones
+ idx
;
6074 lower_zone
->lowmem_reserve
[j
] = managed_pages
/
6075 sysctl_lowmem_reserve_ratio
[idx
];
6076 managed_pages
+= lower_zone
->managed_pages
;
6081 /* update totalreserve_pages */
6082 calculate_totalreserve_pages();
6085 static void __setup_per_zone_wmarks(void)
6087 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
6088 unsigned long lowmem_pages
= 0;
6090 unsigned long flags
;
6092 /* Calculate total number of !ZONE_HIGHMEM pages */
6093 for_each_zone(zone
) {
6094 if (!is_highmem(zone
))
6095 lowmem_pages
+= zone
->managed_pages
;
6098 for_each_zone(zone
) {
6101 spin_lock_irqsave(&zone
->lock
, flags
);
6102 tmp
= (u64
)pages_min
* zone
->managed_pages
;
6103 do_div(tmp
, lowmem_pages
);
6104 if (is_highmem(zone
)) {
6106 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6107 * need highmem pages, so cap pages_min to a small
6110 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6111 * deltas control asynch page reclaim, and so should
6112 * not be capped for highmem.
6114 unsigned long min_pages
;
6116 min_pages
= zone
->managed_pages
/ 1024;
6117 min_pages
= clamp(min_pages
, SWAP_CLUSTER_MAX
, 128UL);
6118 zone
->watermark
[WMARK_MIN
] = min_pages
;
6121 * If it's a lowmem zone, reserve a number of pages
6122 * proportionate to the zone's size.
6124 zone
->watermark
[WMARK_MIN
] = tmp
;
6127 zone
->watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + (tmp
>> 2);
6128 zone
->watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + (tmp
>> 1);
6130 __mod_zone_page_state(zone
, NR_ALLOC_BATCH
,
6131 high_wmark_pages(zone
) - low_wmark_pages(zone
) -
6132 atomic_long_read(&zone
->vm_stat
[NR_ALLOC_BATCH
]));
6134 setup_zone_migrate_reserve(zone
);
6135 spin_unlock_irqrestore(&zone
->lock
, flags
);
6138 /* update totalreserve_pages */
6139 calculate_totalreserve_pages();
6143 * setup_per_zone_wmarks - called when min_free_kbytes changes
6144 * or when memory is hot-{added|removed}
6146 * Ensures that the watermark[min,low,high] values for each zone are set
6147 * correctly with respect to min_free_kbytes.
6149 void setup_per_zone_wmarks(void)
6151 mutex_lock(&zonelists_mutex
);
6152 __setup_per_zone_wmarks();
6153 mutex_unlock(&zonelists_mutex
);
6157 * The inactive anon list should be small enough that the VM never has to
6158 * do too much work, but large enough that each inactive page has a chance
6159 * to be referenced again before it is swapped out.
6161 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
6162 * INACTIVE_ANON pages on this zone's LRU, maintained by the
6163 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
6164 * the anonymous pages are kept on the inactive list.
6167 * memory ratio inactive anon
6168 * -------------------------------------
6177 static void __meminit
calculate_zone_inactive_ratio(struct zone
*zone
)
6179 unsigned int gb
, ratio
;
6181 /* Zone size in gigabytes */
6182 gb
= zone
->managed_pages
>> (30 - PAGE_SHIFT
);
6184 ratio
= int_sqrt(10 * gb
);
6188 zone
->inactive_ratio
= ratio
;
6191 static void __meminit
setup_per_zone_inactive_ratio(void)
6196 calculate_zone_inactive_ratio(zone
);
6200 * Initialise min_free_kbytes.
6202 * For small machines we want it small (128k min). For large machines
6203 * we want it large (64MB max). But it is not linear, because network
6204 * bandwidth does not increase linearly with machine size. We use
6206 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6207 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6223 int __meminit
init_per_zone_wmark_min(void)
6225 unsigned long lowmem_kbytes
;
6226 int new_min_free_kbytes
;
6228 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
6229 new_min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
6231 if (new_min_free_kbytes
> user_min_free_kbytes
) {
6232 min_free_kbytes
= new_min_free_kbytes
;
6233 if (min_free_kbytes
< 128)
6234 min_free_kbytes
= 128;
6235 if (min_free_kbytes
> 65536)
6236 min_free_kbytes
= 65536;
6238 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6239 new_min_free_kbytes
, user_min_free_kbytes
);
6241 setup_per_zone_wmarks();
6242 refresh_zone_stat_thresholds();
6243 setup_per_zone_lowmem_reserve();
6244 setup_per_zone_inactive_ratio();
6247 module_init(init_per_zone_wmark_min
)
6250 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6251 * that we can call two helper functions whenever min_free_kbytes
6254 int min_free_kbytes_sysctl_handler(struct ctl_table
*table
, int write
,
6255 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6259 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6264 user_min_free_kbytes
= min_free_kbytes
;
6265 setup_per_zone_wmarks();
6271 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
6272 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6277 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6282 zone
->min_unmapped_pages
= (zone
->managed_pages
*
6283 sysctl_min_unmapped_ratio
) / 100;
6287 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
6288 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6293 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6298 zone
->min_slab_pages
= (zone
->managed_pages
*
6299 sysctl_min_slab_ratio
) / 100;
6305 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6306 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6307 * whenever sysctl_lowmem_reserve_ratio changes.
6309 * The reserve ratio obviously has absolutely no relation with the
6310 * minimum watermarks. The lowmem reserve ratio can only make sense
6311 * if in function of the boot time zone sizes.
6313 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
6314 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6316 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6317 setup_per_zone_lowmem_reserve();
6322 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6323 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6324 * pagelist can have before it gets flushed back to buddy allocator.
6326 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table
*table
, int write
,
6327 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6330 int old_percpu_pagelist_fraction
;
6333 mutex_lock(&pcp_batch_high_lock
);
6334 old_percpu_pagelist_fraction
= percpu_pagelist_fraction
;
6336 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6337 if (!write
|| ret
< 0)
6340 /* Sanity checking to avoid pcp imbalance */
6341 if (percpu_pagelist_fraction
&&
6342 percpu_pagelist_fraction
< MIN_PERCPU_PAGELIST_FRACTION
) {
6343 percpu_pagelist_fraction
= old_percpu_pagelist_fraction
;
6349 if (percpu_pagelist_fraction
== old_percpu_pagelist_fraction
)
6352 for_each_populated_zone(zone
) {
6355 for_each_possible_cpu(cpu
)
6356 pageset_set_high_and_batch(zone
,
6357 per_cpu_ptr(zone
->pageset
, cpu
));
6360 mutex_unlock(&pcp_batch_high_lock
);
6365 int hashdist
= HASHDIST_DEFAULT
;
6367 static int __init
set_hashdist(char *str
)
6371 hashdist
= simple_strtoul(str
, &str
, 0);
6374 __setup("hashdist=", set_hashdist
);
6378 * allocate a large system hash table from bootmem
6379 * - it is assumed that the hash table must contain an exact power-of-2
6380 * quantity of entries
6381 * - limit is the number of hash buckets, not the total allocation size
6383 void *__init
alloc_large_system_hash(const char *tablename
,
6384 unsigned long bucketsize
,
6385 unsigned long numentries
,
6388 unsigned int *_hash_shift
,
6389 unsigned int *_hash_mask
,
6390 unsigned long low_limit
,
6391 unsigned long high_limit
)
6393 unsigned long long max
= high_limit
;
6394 unsigned long log2qty
, size
;
6397 /* allow the kernel cmdline to have a say */
6399 /* round applicable memory size up to nearest megabyte */
6400 numentries
= nr_kernel_pages
;
6402 /* It isn't necessary when PAGE_SIZE >= 1MB */
6403 if (PAGE_SHIFT
< 20)
6404 numentries
= round_up(numentries
, (1<<20)/PAGE_SIZE
);
6406 /* limit to 1 bucket per 2^scale bytes of low memory */
6407 if (scale
> PAGE_SHIFT
)
6408 numentries
>>= (scale
- PAGE_SHIFT
);
6410 numentries
<<= (PAGE_SHIFT
- scale
);
6412 /* Make sure we've got at least a 0-order allocation.. */
6413 if (unlikely(flags
& HASH_SMALL
)) {
6414 /* Makes no sense without HASH_EARLY */
6415 WARN_ON(!(flags
& HASH_EARLY
));
6416 if (!(numentries
>> *_hash_shift
)) {
6417 numentries
= 1UL << *_hash_shift
;
6418 BUG_ON(!numentries
);
6420 } else if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
6421 numentries
= PAGE_SIZE
/ bucketsize
;
6423 numentries
= roundup_pow_of_two(numentries
);
6425 /* limit allocation size to 1/16 total memory by default */
6427 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
6428 do_div(max
, bucketsize
);
6430 max
= min(max
, 0x80000000ULL
);
6432 if (numentries
< low_limit
)
6433 numentries
= low_limit
;
6434 if (numentries
> max
)
6437 log2qty
= ilog2(numentries
);
6440 size
= bucketsize
<< log2qty
;
6441 if (flags
& HASH_EARLY
)
6442 table
= memblock_virt_alloc_nopanic(size
, 0);
6444 table
= __vmalloc(size
, GFP_ATOMIC
, PAGE_KERNEL
);
6447 * If bucketsize is not a power-of-two, we may free
6448 * some pages at the end of hash table which
6449 * alloc_pages_exact() automatically does
6451 if (get_order(size
) < MAX_ORDER
) {
6452 table
= alloc_pages_exact(size
, GFP_ATOMIC
);
6453 kmemleak_alloc(table
, size
, 1, GFP_ATOMIC
);
6456 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
6459 panic("Failed to allocate %s hash table\n", tablename
);
6461 printk(KERN_INFO
"%s hash table entries: %ld (order: %d, %lu bytes)\n",
6464 ilog2(size
) - PAGE_SHIFT
,
6468 *_hash_shift
= log2qty
;
6470 *_hash_mask
= (1 << log2qty
) - 1;
6475 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6476 static inline unsigned long *get_pageblock_bitmap(struct zone
*zone
,
6479 #ifdef CONFIG_SPARSEMEM
6480 return __pfn_to_section(pfn
)->pageblock_flags
;
6482 return zone
->pageblock_flags
;
6483 #endif /* CONFIG_SPARSEMEM */
6486 static inline int pfn_to_bitidx(struct zone
*zone
, unsigned long pfn
)
6488 #ifdef CONFIG_SPARSEMEM
6489 pfn
&= (PAGES_PER_SECTION
-1);
6490 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
6492 pfn
= pfn
- round_down(zone
->zone_start_pfn
, pageblock_nr_pages
);
6493 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
6494 #endif /* CONFIG_SPARSEMEM */
6498 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6499 * @page: The page within the block of interest
6500 * @pfn: The target page frame number
6501 * @end_bitidx: The last bit of interest to retrieve
6502 * @mask: mask of bits that the caller is interested in
6504 * Return: pageblock_bits flags
6506 unsigned long get_pfnblock_flags_mask(struct page
*page
, unsigned long pfn
,
6507 unsigned long end_bitidx
,
6511 unsigned long *bitmap
;
6512 unsigned long bitidx
, word_bitidx
;
6515 zone
= page_zone(page
);
6516 bitmap
= get_pageblock_bitmap(zone
, pfn
);
6517 bitidx
= pfn_to_bitidx(zone
, pfn
);
6518 word_bitidx
= bitidx
/ BITS_PER_LONG
;
6519 bitidx
&= (BITS_PER_LONG
-1);
6521 word
= bitmap
[word_bitidx
];
6522 bitidx
+= end_bitidx
;
6523 return (word
>> (BITS_PER_LONG
- bitidx
- 1)) & mask
;
6527 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6528 * @page: The page within the block of interest
6529 * @flags: The flags to set
6530 * @pfn: The target page frame number
6531 * @end_bitidx: The last bit of interest
6532 * @mask: mask of bits that the caller is interested in
6534 void set_pfnblock_flags_mask(struct page
*page
, unsigned long flags
,
6536 unsigned long end_bitidx
,
6540 unsigned long *bitmap
;
6541 unsigned long bitidx
, word_bitidx
;
6542 unsigned long old_word
, word
;
6544 BUILD_BUG_ON(NR_PAGEBLOCK_BITS
!= 4);
6546 zone
= page_zone(page
);
6547 bitmap
= get_pageblock_bitmap(zone
, pfn
);
6548 bitidx
= pfn_to_bitidx(zone
, pfn
);
6549 word_bitidx
= bitidx
/ BITS_PER_LONG
;
6550 bitidx
&= (BITS_PER_LONG
-1);
6552 VM_BUG_ON_PAGE(!zone_spans_pfn(zone
, pfn
), page
);
6554 bitidx
+= end_bitidx
;
6555 mask
<<= (BITS_PER_LONG
- bitidx
- 1);
6556 flags
<<= (BITS_PER_LONG
- bitidx
- 1);
6558 word
= READ_ONCE(bitmap
[word_bitidx
]);
6560 old_word
= cmpxchg(&bitmap
[word_bitidx
], word
, (word
& ~mask
) | flags
);
6561 if (word
== old_word
)
6568 * This function checks whether pageblock includes unmovable pages or not.
6569 * If @count is not zero, it is okay to include less @count unmovable pages
6571 * PageLRU check without isolation or lru_lock could race so that
6572 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6573 * expect this function should be exact.
6575 bool has_unmovable_pages(struct zone
*zone
, struct page
*page
, int count
,
6576 bool skip_hwpoisoned_pages
)
6578 unsigned long pfn
, iter
, found
;
6582 * For avoiding noise data, lru_add_drain_all() should be called
6583 * If ZONE_MOVABLE, the zone never contains unmovable pages
6585 if (zone_idx(zone
) == ZONE_MOVABLE
)
6587 mt
= get_pageblock_migratetype(page
);
6588 if (mt
== MIGRATE_MOVABLE
|| is_migrate_cma(mt
))
6591 pfn
= page_to_pfn(page
);
6592 for (found
= 0, iter
= 0; iter
< pageblock_nr_pages
; iter
++) {
6593 unsigned long check
= pfn
+ iter
;
6595 if (!pfn_valid_within(check
))
6598 page
= pfn_to_page(check
);
6601 * Hugepages are not in LRU lists, but they're movable.
6602 * We need not scan over tail pages bacause we don't
6603 * handle each tail page individually in migration.
6605 if (PageHuge(page
)) {
6606 iter
= round_up(iter
+ 1, 1<<compound_order(page
)) - 1;
6611 * We can't use page_count without pin a page
6612 * because another CPU can free compound page.
6613 * This check already skips compound tails of THP
6614 * because their page->_count is zero at all time.
6616 if (!atomic_read(&page
->_count
)) {
6617 if (PageBuddy(page
))
6618 iter
+= (1 << page_order(page
)) - 1;
6623 * The HWPoisoned page may be not in buddy system, and
6624 * page_count() is not 0.
6626 if (skip_hwpoisoned_pages
&& PageHWPoison(page
))
6632 * If there are RECLAIMABLE pages, we need to check
6633 * it. But now, memory offline itself doesn't call
6634 * shrink_node_slabs() and it still to be fixed.
6637 * If the page is not RAM, page_count()should be 0.
6638 * we don't need more check. This is an _used_ not-movable page.
6640 * The problematic thing here is PG_reserved pages. PG_reserved
6641 * is set to both of a memory hole page and a _used_ kernel
6650 bool is_pageblock_removable_nolock(struct page
*page
)
6656 * We have to be careful here because we are iterating over memory
6657 * sections which are not zone aware so we might end up outside of
6658 * the zone but still within the section.
6659 * We have to take care about the node as well. If the node is offline
6660 * its NODE_DATA will be NULL - see page_zone.
6662 if (!node_online(page_to_nid(page
)))
6665 zone
= page_zone(page
);
6666 pfn
= page_to_pfn(page
);
6667 if (!zone_spans_pfn(zone
, pfn
))
6670 return !has_unmovable_pages(zone
, page
, 0, true);
6675 static unsigned long pfn_max_align_down(unsigned long pfn
)
6677 return pfn
& ~(max_t(unsigned long, MAX_ORDER_NR_PAGES
,
6678 pageblock_nr_pages
) - 1);
6681 static unsigned long pfn_max_align_up(unsigned long pfn
)
6683 return ALIGN(pfn
, max_t(unsigned long, MAX_ORDER_NR_PAGES
,
6684 pageblock_nr_pages
));
6687 /* [start, end) must belong to a single zone. */
6688 static int __alloc_contig_migrate_range(struct compact_control
*cc
,
6689 unsigned long start
, unsigned long end
)
6691 /* This function is based on compact_zone() from compaction.c. */
6692 unsigned long nr_reclaimed
;
6693 unsigned long pfn
= start
;
6694 unsigned int tries
= 0;
6699 while (pfn
< end
|| !list_empty(&cc
->migratepages
)) {
6700 if (fatal_signal_pending(current
)) {
6705 if (list_empty(&cc
->migratepages
)) {
6706 cc
->nr_migratepages
= 0;
6707 pfn
= isolate_migratepages_range(cc
, pfn
, end
);
6713 } else if (++tries
== 5) {
6714 ret
= ret
< 0 ? ret
: -EBUSY
;
6718 nr_reclaimed
= reclaim_clean_pages_from_list(cc
->zone
,
6720 cc
->nr_migratepages
-= nr_reclaimed
;
6722 ret
= migrate_pages(&cc
->migratepages
, alloc_migrate_target
,
6723 NULL
, 0, cc
->mode
, MR_CMA
);
6726 putback_movable_pages(&cc
->migratepages
);
6733 * alloc_contig_range() -- tries to allocate given range of pages
6734 * @start: start PFN to allocate
6735 * @end: one-past-the-last PFN to allocate
6736 * @migratetype: migratetype of the underlaying pageblocks (either
6737 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6738 * in range must have the same migratetype and it must
6739 * be either of the two.
6741 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6742 * aligned, however it's the caller's responsibility to guarantee that
6743 * we are the only thread that changes migrate type of pageblocks the
6746 * The PFN range must belong to a single zone.
6748 * Returns zero on success or negative error code. On success all
6749 * pages which PFN is in [start, end) are allocated for the caller and
6750 * need to be freed with free_contig_range().
6752 int alloc_contig_range(unsigned long start
, unsigned long end
,
6753 unsigned migratetype
)
6755 unsigned long outer_start
, outer_end
;
6758 struct compact_control cc
= {
6759 .nr_migratepages
= 0,
6761 .zone
= page_zone(pfn_to_page(start
)),
6762 .mode
= MIGRATE_SYNC
,
6763 .ignore_skip_hint
= true,
6765 INIT_LIST_HEAD(&cc
.migratepages
);
6768 * What we do here is we mark all pageblocks in range as
6769 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6770 * have different sizes, and due to the way page allocator
6771 * work, we align the range to biggest of the two pages so
6772 * that page allocator won't try to merge buddies from
6773 * different pageblocks and change MIGRATE_ISOLATE to some
6774 * other migration type.
6776 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6777 * migrate the pages from an unaligned range (ie. pages that
6778 * we are interested in). This will put all the pages in
6779 * range back to page allocator as MIGRATE_ISOLATE.
6781 * When this is done, we take the pages in range from page
6782 * allocator removing them from the buddy system. This way
6783 * page allocator will never consider using them.
6785 * This lets us mark the pageblocks back as
6786 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6787 * aligned range but not in the unaligned, original range are
6788 * put back to page allocator so that buddy can use them.
6791 ret
= start_isolate_page_range(pfn_max_align_down(start
),
6792 pfn_max_align_up(end
), migratetype
,
6797 ret
= __alloc_contig_migrate_range(&cc
, start
, end
);
6802 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6803 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6804 * more, all pages in [start, end) are free in page allocator.
6805 * What we are going to do is to allocate all pages from
6806 * [start, end) (that is remove them from page allocator).
6808 * The only problem is that pages at the beginning and at the
6809 * end of interesting range may be not aligned with pages that
6810 * page allocator holds, ie. they can be part of higher order
6811 * pages. Because of this, we reserve the bigger range and
6812 * once this is done free the pages we are not interested in.
6814 * We don't have to hold zone->lock here because the pages are
6815 * isolated thus they won't get removed from buddy.
6818 lru_add_drain_all();
6819 drain_all_pages(cc
.zone
);
6822 outer_start
= start
;
6823 while (!PageBuddy(pfn_to_page(outer_start
))) {
6824 if (++order
>= MAX_ORDER
) {
6828 outer_start
&= ~0UL << order
;
6831 /* Make sure the range is really isolated. */
6832 if (test_pages_isolated(outer_start
, end
, false)) {
6833 pr_info("%s: [%lx, %lx) PFNs busy\n",
6834 __func__
, outer_start
, end
);
6839 /* Grab isolated pages from freelists. */
6840 outer_end
= isolate_freepages_range(&cc
, outer_start
, end
);
6846 /* Free head and tail (if any) */
6847 if (start
!= outer_start
)
6848 free_contig_range(outer_start
, start
- outer_start
);
6849 if (end
!= outer_end
)
6850 free_contig_range(end
, outer_end
- end
);
6853 undo_isolate_page_range(pfn_max_align_down(start
),
6854 pfn_max_align_up(end
), migratetype
);
6858 void free_contig_range(unsigned long pfn
, unsigned nr_pages
)
6860 unsigned int count
= 0;
6862 for (; nr_pages
--; pfn
++) {
6863 struct page
*page
= pfn_to_page(pfn
);
6865 count
+= page_count(page
) != 1;
6868 WARN(count
!= 0, "%d pages are still in use!\n", count
);
6872 #ifdef CONFIG_MEMORY_HOTPLUG
6874 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6875 * page high values need to be recalulated.
6877 void __meminit
zone_pcp_update(struct zone
*zone
)
6880 mutex_lock(&pcp_batch_high_lock
);
6881 for_each_possible_cpu(cpu
)
6882 pageset_set_high_and_batch(zone
,
6883 per_cpu_ptr(zone
->pageset
, cpu
));
6884 mutex_unlock(&pcp_batch_high_lock
);
6888 void zone_pcp_reset(struct zone
*zone
)
6890 unsigned long flags
;
6892 struct per_cpu_pageset
*pset
;
6894 /* avoid races with drain_pages() */
6895 local_irq_save(flags
);
6896 if (zone
->pageset
!= &boot_pageset
) {
6897 for_each_online_cpu(cpu
) {
6898 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
6899 drain_zonestat(zone
, pset
);
6901 free_percpu(zone
->pageset
);
6902 zone
->pageset
= &boot_pageset
;
6904 local_irq_restore(flags
);
6907 #ifdef CONFIG_MEMORY_HOTREMOVE
6909 * All pages in the range must be isolated before calling this.
6912 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
6916 unsigned int order
, i
;
6918 unsigned long flags
;
6919 /* find the first valid pfn */
6920 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
6925 zone
= page_zone(pfn_to_page(pfn
));
6926 spin_lock_irqsave(&zone
->lock
, flags
);
6928 while (pfn
< end_pfn
) {
6929 if (!pfn_valid(pfn
)) {
6933 page
= pfn_to_page(pfn
);
6935 * The HWPoisoned page may be not in buddy system, and
6936 * page_count() is not 0.
6938 if (unlikely(!PageBuddy(page
) && PageHWPoison(page
))) {
6940 SetPageReserved(page
);
6944 BUG_ON(page_count(page
));
6945 BUG_ON(!PageBuddy(page
));
6946 order
= page_order(page
);
6947 #ifdef CONFIG_DEBUG_VM
6948 printk(KERN_INFO
"remove from free list %lx %d %lx\n",
6949 pfn
, 1 << order
, end_pfn
);
6951 list_del(&page
->lru
);
6952 rmv_page_order(page
);
6953 zone
->free_area
[order
].nr_free
--;
6954 for (i
= 0; i
< (1 << order
); i
++)
6955 SetPageReserved((page
+i
));
6956 pfn
+= (1 << order
);
6958 spin_unlock_irqrestore(&zone
->lock
, flags
);
6962 #ifdef CONFIG_MEMORY_FAILURE
6963 bool is_free_buddy_page(struct page
*page
)
6965 struct zone
*zone
= page_zone(page
);
6966 unsigned long pfn
= page_to_pfn(page
);
6967 unsigned long flags
;
6970 spin_lock_irqsave(&zone
->lock
, flags
);
6971 for (order
= 0; order
< MAX_ORDER
; order
++) {
6972 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
6974 if (PageBuddy(page_head
) && page_order(page_head
) >= order
)
6977 spin_unlock_irqrestore(&zone
->lock
, flags
);
6979 return order
< MAX_ORDER
;