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/compiler.h>
25 #include <linux/kernel.h>
26 #include <linux/kmemcheck.h>
27 #include <linux/module.h>
28 #include <linux/suspend.h>
29 #include <linux/pagevec.h>
30 #include <linux/blkdev.h>
31 #include <linux/slab.h>
32 #include <linux/oom.h>
33 #include <linux/notifier.h>
34 #include <linux/topology.h>
35 #include <linux/sysctl.h>
36 #include <linux/cpu.h>
37 #include <linux/cpuset.h>
38 #include <linux/memory_hotplug.h>
39 #include <linux/nodemask.h>
40 #include <linux/vmalloc.h>
41 #include <linux/mempolicy.h>
42 #include <linux/stop_machine.h>
43 #include <linux/sort.h>
44 #include <linux/pfn.h>
45 #include <linux/backing-dev.h>
46 #include <linux/fault-inject.h>
47 #include <linux/page-isolation.h>
48 #include <linux/page_cgroup.h>
49 #include <linux/debugobjects.h>
50 #include <linux/kmemleak.h>
51 #include <trace/events/kmem.h>
53 #include <asm/tlbflush.h>
54 #include <asm/div64.h>
58 * Array of node states.
60 nodemask_t node_states
[NR_NODE_STATES
] __read_mostly
= {
61 [N_POSSIBLE
] = NODE_MASK_ALL
,
62 [N_ONLINE
] = { { [0] = 1UL } },
64 [N_NORMAL_MEMORY
] = { { [0] = 1UL } },
66 [N_HIGH_MEMORY
] = { { [0] = 1UL } },
68 [N_CPU
] = { { [0] = 1UL } },
71 EXPORT_SYMBOL(node_states
);
73 unsigned long totalram_pages __read_mostly
;
74 unsigned long totalreserve_pages __read_mostly
;
75 unsigned long highest_memmap_pfn __read_mostly
;
76 int percpu_pagelist_fraction
;
77 gfp_t gfp_allowed_mask __read_mostly
= GFP_BOOT_MASK
;
79 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
80 int pageblock_order __read_mostly
;
83 static void __free_pages_ok(struct page
*page
, unsigned int order
);
86 * results with 256, 32 in the lowmem_reserve sysctl:
87 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
88 * 1G machine -> (16M dma, 784M normal, 224M high)
89 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
90 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
91 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
93 * TBD: should special case ZONE_DMA32 machines here - in those we normally
94 * don't need any ZONE_NORMAL reservation
96 int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
-1] = {
97 #ifdef CONFIG_ZONE_DMA
100 #ifdef CONFIG_ZONE_DMA32
103 #ifdef CONFIG_HIGHMEM
109 EXPORT_SYMBOL(totalram_pages
);
111 static char * const zone_names
[MAX_NR_ZONES
] = {
112 #ifdef CONFIG_ZONE_DMA
115 #ifdef CONFIG_ZONE_DMA32
119 #ifdef CONFIG_HIGHMEM
125 int min_free_kbytes
= 1024;
127 static unsigned long __meminitdata nr_kernel_pages
;
128 static unsigned long __meminitdata nr_all_pages
;
129 static unsigned long __meminitdata dma_reserve
;
131 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
133 * MAX_ACTIVE_REGIONS determines the maximum number of distinct
134 * ranges of memory (RAM) that may be registered with add_active_range().
135 * Ranges passed to add_active_range() will be merged if possible
136 * so the number of times add_active_range() can be called is
137 * related to the number of nodes and the number of holes
139 #ifdef CONFIG_MAX_ACTIVE_REGIONS
140 /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */
141 #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS
143 #if MAX_NUMNODES >= 32
144 /* If there can be many nodes, allow up to 50 holes per node */
145 #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50)
147 /* By default, allow up to 256 distinct regions */
148 #define MAX_ACTIVE_REGIONS 256
152 static struct node_active_region __meminitdata early_node_map
[MAX_ACTIVE_REGIONS
];
153 static int __meminitdata nr_nodemap_entries
;
154 static unsigned long __meminitdata arch_zone_lowest_possible_pfn
[MAX_NR_ZONES
];
155 static unsigned long __meminitdata arch_zone_highest_possible_pfn
[MAX_NR_ZONES
];
156 static unsigned long __initdata required_kernelcore
;
157 static unsigned long __initdata required_movablecore
;
158 static unsigned long __meminitdata zone_movable_pfn
[MAX_NUMNODES
];
160 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
162 EXPORT_SYMBOL(movable_zone
);
163 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
166 int nr_node_ids __read_mostly
= MAX_NUMNODES
;
167 int nr_online_nodes __read_mostly
= 1;
168 EXPORT_SYMBOL(nr_node_ids
);
169 EXPORT_SYMBOL(nr_online_nodes
);
172 int page_group_by_mobility_disabled __read_mostly
;
174 static void set_pageblock_migratetype(struct page
*page
, int migratetype
)
177 if (unlikely(page_group_by_mobility_disabled
))
178 migratetype
= MIGRATE_UNMOVABLE
;
180 set_pageblock_flags_group(page
, (unsigned long)migratetype
,
181 PB_migrate
, PB_migrate_end
);
184 bool oom_killer_disabled __read_mostly
;
186 #ifdef CONFIG_DEBUG_VM
187 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
191 unsigned long pfn
= page_to_pfn(page
);
194 seq
= zone_span_seqbegin(zone
);
195 if (pfn
>= zone
->zone_start_pfn
+ zone
->spanned_pages
)
197 else if (pfn
< zone
->zone_start_pfn
)
199 } while (zone_span_seqretry(zone
, seq
));
204 static int page_is_consistent(struct zone
*zone
, struct page
*page
)
206 if (!pfn_valid_within(page_to_pfn(page
)))
208 if (zone
!= page_zone(page
))
214 * Temporary debugging check for pages not lying within a given zone.
216 static int bad_range(struct zone
*zone
, struct page
*page
)
218 if (page_outside_zone_boundaries(zone
, page
))
220 if (!page_is_consistent(zone
, page
))
226 static inline int bad_range(struct zone
*zone
, struct page
*page
)
232 static void bad_page(struct page
*page
)
234 static unsigned long resume
;
235 static unsigned long nr_shown
;
236 static unsigned long nr_unshown
;
239 * Allow a burst of 60 reports, then keep quiet for that minute;
240 * or allow a steady drip of one report per second.
242 if (nr_shown
== 60) {
243 if (time_before(jiffies
, resume
)) {
249 "BUG: Bad page state: %lu messages suppressed\n",
256 resume
= jiffies
+ 60 * HZ
;
258 printk(KERN_ALERT
"BUG: Bad page state in process %s pfn:%05lx\n",
259 current
->comm
, page_to_pfn(page
));
261 "page:%p flags:%p count:%d mapcount:%d mapping:%p index:%lx\n",
262 page
, (void *)page
->flags
, page_count(page
),
263 page_mapcount(page
), page
->mapping
, page
->index
);
267 /* Leave bad fields for debug, except PageBuddy could make trouble */
268 __ClearPageBuddy(page
);
269 add_taint(TAINT_BAD_PAGE
);
273 * Higher-order pages are called "compound pages". They are structured thusly:
275 * The first PAGE_SIZE page is called the "head page".
277 * The remaining PAGE_SIZE pages are called "tail pages".
279 * All pages have PG_compound set. All pages have their ->private pointing at
280 * the head page (even the head page has this).
282 * The first tail page's ->lru.next holds the address of the compound page's
283 * put_page() function. Its ->lru.prev holds the order of allocation.
284 * This usage means that zero-order pages may not be compound.
287 static void free_compound_page(struct page
*page
)
289 __free_pages_ok(page
, compound_order(page
));
292 void prep_compound_page(struct page
*page
, unsigned long order
)
295 int nr_pages
= 1 << order
;
297 set_compound_page_dtor(page
, free_compound_page
);
298 set_compound_order(page
, order
);
300 for (i
= 1; i
< nr_pages
; i
++) {
301 struct page
*p
= page
+ i
;
304 p
->first_page
= page
;
308 static int destroy_compound_page(struct page
*page
, unsigned long order
)
311 int nr_pages
= 1 << order
;
314 if (unlikely(compound_order(page
) != order
) ||
315 unlikely(!PageHead(page
))) {
320 __ClearPageHead(page
);
322 for (i
= 1; i
< nr_pages
; i
++) {
323 struct page
*p
= page
+ i
;
325 if (unlikely(!PageTail(p
) || (p
->first_page
!= page
))) {
335 static inline void prep_zero_page(struct page
*page
, int order
, gfp_t gfp_flags
)
340 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
341 * and __GFP_HIGHMEM from hard or soft interrupt context.
343 VM_BUG_ON((gfp_flags
& __GFP_HIGHMEM
) && in_interrupt());
344 for (i
= 0; i
< (1 << order
); i
++)
345 clear_highpage(page
+ i
);
348 static inline void set_page_order(struct page
*page
, int order
)
350 set_page_private(page
, order
);
351 __SetPageBuddy(page
);
354 static inline void rmv_page_order(struct page
*page
)
356 __ClearPageBuddy(page
);
357 set_page_private(page
, 0);
361 * Locate the struct page for both the matching buddy in our
362 * pair (buddy1) and the combined O(n+1) page they form (page).
364 * 1) Any buddy B1 will have an order O twin B2 which satisfies
365 * the following equation:
367 * For example, if the starting buddy (buddy2) is #8 its order
369 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
371 * 2) Any buddy B will have an order O+1 parent P which
372 * satisfies the following equation:
375 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
377 static inline struct page
*
378 __page_find_buddy(struct page
*page
, unsigned long page_idx
, unsigned int order
)
380 unsigned long buddy_idx
= page_idx
^ (1 << order
);
382 return page
+ (buddy_idx
- page_idx
);
385 static inline unsigned long
386 __find_combined_index(unsigned long page_idx
, unsigned int order
)
388 return (page_idx
& ~(1 << order
));
392 * This function checks whether a page is free && is the buddy
393 * we can do coalesce a page and its buddy if
394 * (a) the buddy is not in a hole &&
395 * (b) the buddy is in the buddy system &&
396 * (c) a page and its buddy have the same order &&
397 * (d) a page and its buddy are in the same zone.
399 * For recording whether a page is in the buddy system, we use PG_buddy.
400 * Setting, clearing, and testing PG_buddy is serialized by zone->lock.
402 * For recording page's order, we use page_private(page).
404 static inline int page_is_buddy(struct page
*page
, struct page
*buddy
,
407 if (!pfn_valid_within(page_to_pfn(buddy
)))
410 if (page_zone_id(page
) != page_zone_id(buddy
))
413 if (PageBuddy(buddy
) && page_order(buddy
) == order
) {
414 VM_BUG_ON(page_count(buddy
) != 0);
421 * Freeing function for a buddy system allocator.
423 * The concept of a buddy system is to maintain direct-mapped table
424 * (containing bit values) for memory blocks of various "orders".
425 * The bottom level table contains the map for the smallest allocatable
426 * units of memory (here, pages), and each level above it describes
427 * pairs of units from the levels below, hence, "buddies".
428 * At a high level, all that happens here is marking the table entry
429 * at the bottom level available, and propagating the changes upward
430 * as necessary, plus some accounting needed to play nicely with other
431 * parts of the VM system.
432 * At each level, we keep a list of pages, which are heads of continuous
433 * free pages of length of (1 << order) and marked with PG_buddy. Page's
434 * order is recorded in page_private(page) field.
435 * So when we are allocating or freeing one, we can derive the state of the
436 * other. That is, if we allocate a small block, and both were
437 * free, the remainder of the region must be split into blocks.
438 * If a block is freed, and its buddy is also free, then this
439 * triggers coalescing into a block of larger size.
444 static inline void __free_one_page(struct page
*page
,
445 struct zone
*zone
, unsigned int order
,
448 unsigned long page_idx
;
450 if (unlikely(PageCompound(page
)))
451 if (unlikely(destroy_compound_page(page
, order
)))
454 VM_BUG_ON(migratetype
== -1);
456 page_idx
= page_to_pfn(page
) & ((1 << MAX_ORDER
) - 1);
458 VM_BUG_ON(page_idx
& ((1 << order
) - 1));
459 VM_BUG_ON(bad_range(zone
, page
));
461 while (order
< MAX_ORDER
-1) {
462 unsigned long combined_idx
;
465 buddy
= __page_find_buddy(page
, page_idx
, order
);
466 if (!page_is_buddy(page
, buddy
, order
))
469 /* Our buddy is free, merge with it and move up one order. */
470 list_del(&buddy
->lru
);
471 zone
->free_area
[order
].nr_free
--;
472 rmv_page_order(buddy
);
473 combined_idx
= __find_combined_index(page_idx
, order
);
474 page
= page
+ (combined_idx
- page_idx
);
475 page_idx
= combined_idx
;
478 set_page_order(page
, order
);
480 &zone
->free_area
[order
].free_list
[migratetype
]);
481 zone
->free_area
[order
].nr_free
++;
484 #ifdef CONFIG_HAVE_MLOCKED_PAGE_BIT
486 * free_page_mlock() -- clean up attempts to free and mlocked() page.
487 * Page should not be on lru, so no need to fix that up.
488 * free_pages_check() will verify...
490 static inline void free_page_mlock(struct page
*page
)
492 __dec_zone_page_state(page
, NR_MLOCK
);
493 __count_vm_event(UNEVICTABLE_MLOCKFREED
);
496 static void free_page_mlock(struct page
*page
) { }
499 static inline int free_pages_check(struct page
*page
)
501 if (unlikely(page_mapcount(page
) |
502 (page
->mapping
!= NULL
) |
503 (atomic_read(&page
->_count
) != 0) |
504 (page
->flags
& PAGE_FLAGS_CHECK_AT_FREE
))) {
508 if (page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)
509 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
514 * Frees a number of pages from the PCP lists
515 * Assumes all pages on list are in same zone, and of same order.
516 * count is the number of pages to free.
518 * If the zone was previously in an "all pages pinned" state then look to
519 * see if this freeing clears that state.
521 * And clear the zone's pages_scanned counter, to hold off the "all pages are
522 * pinned" detection logic.
524 static void free_pcppages_bulk(struct zone
*zone
, int count
,
525 struct per_cpu_pages
*pcp
)
529 spin_lock(&zone
->lock
);
530 zone_clear_flag(zone
, ZONE_ALL_UNRECLAIMABLE
);
531 zone
->pages_scanned
= 0;
533 __mod_zone_page_state(zone
, NR_FREE_PAGES
, count
);
536 struct list_head
*list
;
539 * Remove pages from lists in a round-robin fashion. This spinning
540 * around potentially empty lists is bloody awful, alternatives that
541 * don't suck are welcome
544 if (++migratetype
== MIGRATE_PCPTYPES
)
546 list
= &pcp
->lists
[migratetype
];
547 } while (list_empty(list
));
549 page
= list_entry(list
->prev
, struct page
, lru
);
550 /* have to delete it as __free_one_page list manipulates */
551 list_del(&page
->lru
);
552 trace_mm_page_pcpu_drain(page
, 0, migratetype
);
553 __free_one_page(page
, zone
, 0, migratetype
);
555 spin_unlock(&zone
->lock
);
558 static void free_one_page(struct zone
*zone
, struct page
*page
, int order
,
561 spin_lock(&zone
->lock
);
562 zone_clear_flag(zone
, ZONE_ALL_UNRECLAIMABLE
);
563 zone
->pages_scanned
= 0;
565 __mod_zone_page_state(zone
, NR_FREE_PAGES
, 1 << order
);
566 __free_one_page(page
, zone
, order
, migratetype
);
567 spin_unlock(&zone
->lock
);
570 static void __free_pages_ok(struct page
*page
, unsigned int order
)
575 int wasMlocked
= __TestClearPageMlocked(page
);
577 kmemcheck_free_shadow(page
, order
);
579 for (i
= 0 ; i
< (1 << order
) ; ++i
)
580 bad
+= free_pages_check(page
+ i
);
584 if (!PageHighMem(page
)) {
585 debug_check_no_locks_freed(page_address(page
),PAGE_SIZE
<<order
);
586 debug_check_no_obj_freed(page_address(page
),
589 arch_free_page(page
, order
);
590 kernel_map_pages(page
, 1 << order
, 0);
592 local_irq_save(flags
);
593 if (unlikely(wasMlocked
))
594 free_page_mlock(page
);
595 __count_vm_events(PGFREE
, 1 << order
);
596 free_one_page(page_zone(page
), page
, order
,
597 get_pageblock_migratetype(page
));
598 local_irq_restore(flags
);
602 * permit the bootmem allocator to evade page validation on high-order frees
604 void __meminit
__free_pages_bootmem(struct page
*page
, unsigned int order
)
607 __ClearPageReserved(page
);
608 set_page_count(page
, 0);
609 set_page_refcounted(page
);
615 for (loop
= 0; loop
< BITS_PER_LONG
; loop
++) {
616 struct page
*p
= &page
[loop
];
618 if (loop
+ 1 < BITS_PER_LONG
)
620 __ClearPageReserved(p
);
621 set_page_count(p
, 0);
624 set_page_refcounted(page
);
625 __free_pages(page
, order
);
631 * The order of subdivision here is critical for the IO subsystem.
632 * Please do not alter this order without good reasons and regression
633 * testing. Specifically, as large blocks of memory are subdivided,
634 * the order in which smaller blocks are delivered depends on the order
635 * they're subdivided in this function. This is the primary factor
636 * influencing the order in which pages are delivered to the IO
637 * subsystem according to empirical testing, and this is also justified
638 * by considering the behavior of a buddy system containing a single
639 * large block of memory acted on by a series of small allocations.
640 * This behavior is a critical factor in sglist merging's success.
644 static inline void expand(struct zone
*zone
, struct page
*page
,
645 int low
, int high
, struct free_area
*area
,
648 unsigned long size
= 1 << high
;
654 VM_BUG_ON(bad_range(zone
, &page
[size
]));
655 list_add(&page
[size
].lru
, &area
->free_list
[migratetype
]);
657 set_page_order(&page
[size
], high
);
662 * This page is about to be returned from the page allocator
664 static int prep_new_page(struct page
*page
, int order
, gfp_t gfp_flags
)
666 if (unlikely(page_mapcount(page
) |
667 (page
->mapping
!= NULL
) |
668 (atomic_read(&page
->_count
) != 0) |
669 (page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
))) {
674 set_page_private(page
, 0);
675 set_page_refcounted(page
);
677 arch_alloc_page(page
, order
);
678 kernel_map_pages(page
, 1 << order
, 1);
680 if (gfp_flags
& __GFP_ZERO
)
681 prep_zero_page(page
, order
, gfp_flags
);
683 if (order
&& (gfp_flags
& __GFP_COMP
))
684 prep_compound_page(page
, order
);
690 * Go through the free lists for the given migratetype and remove
691 * the smallest available page from the freelists
694 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
697 unsigned int current_order
;
698 struct free_area
* area
;
701 /* Find a page of the appropriate size in the preferred list */
702 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
703 area
= &(zone
->free_area
[current_order
]);
704 if (list_empty(&area
->free_list
[migratetype
]))
707 page
= list_entry(area
->free_list
[migratetype
].next
,
709 list_del(&page
->lru
);
710 rmv_page_order(page
);
712 expand(zone
, page
, order
, current_order
, area
, migratetype
);
721 * This array describes the order lists are fallen back to when
722 * the free lists for the desirable migrate type are depleted
724 static int fallbacks
[MIGRATE_TYPES
][MIGRATE_TYPES
-1] = {
725 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
726 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
727 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_RESERVE
},
728 [MIGRATE_RESERVE
] = { MIGRATE_RESERVE
, MIGRATE_RESERVE
, MIGRATE_RESERVE
}, /* Never used */
732 * Move the free pages in a range to the free lists of the requested type.
733 * Note that start_page and end_pages are not aligned on a pageblock
734 * boundary. If alignment is required, use move_freepages_block()
736 static int move_freepages(struct zone
*zone
,
737 struct page
*start_page
, struct page
*end_page
,
744 #ifndef CONFIG_HOLES_IN_ZONE
746 * page_zone is not safe to call in this context when
747 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
748 * anyway as we check zone boundaries in move_freepages_block().
749 * Remove at a later date when no bug reports exist related to
750 * grouping pages by mobility
752 BUG_ON(page_zone(start_page
) != page_zone(end_page
));
755 for (page
= start_page
; page
<= end_page
;) {
756 /* Make sure we are not inadvertently changing nodes */
757 VM_BUG_ON(page_to_nid(page
) != zone_to_nid(zone
));
759 if (!pfn_valid_within(page_to_pfn(page
))) {
764 if (!PageBuddy(page
)) {
769 order
= page_order(page
);
770 list_del(&page
->lru
);
772 &zone
->free_area
[order
].free_list
[migratetype
]);
774 pages_moved
+= 1 << order
;
780 static int move_freepages_block(struct zone
*zone
, struct page
*page
,
783 unsigned long start_pfn
, end_pfn
;
784 struct page
*start_page
, *end_page
;
786 start_pfn
= page_to_pfn(page
);
787 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
788 start_page
= pfn_to_page(start_pfn
);
789 end_page
= start_page
+ pageblock_nr_pages
- 1;
790 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
792 /* Do not cross zone boundaries */
793 if (start_pfn
< zone
->zone_start_pfn
)
795 if (end_pfn
>= zone
->zone_start_pfn
+ zone
->spanned_pages
)
798 return move_freepages(zone
, start_page
, end_page
, migratetype
);
801 static void change_pageblock_range(struct page
*pageblock_page
,
802 int start_order
, int migratetype
)
804 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
806 while (nr_pageblocks
--) {
807 set_pageblock_migratetype(pageblock_page
, migratetype
);
808 pageblock_page
+= pageblock_nr_pages
;
812 /* Remove an element from the buddy allocator from the fallback list */
813 static inline struct page
*
814 __rmqueue_fallback(struct zone
*zone
, int order
, int start_migratetype
)
816 struct free_area
* area
;
821 /* Find the largest possible block of pages in the other list */
822 for (current_order
= MAX_ORDER
-1; current_order
>= order
;
824 for (i
= 0; i
< MIGRATE_TYPES
- 1; i
++) {
825 migratetype
= fallbacks
[start_migratetype
][i
];
827 /* MIGRATE_RESERVE handled later if necessary */
828 if (migratetype
== MIGRATE_RESERVE
)
831 area
= &(zone
->free_area
[current_order
]);
832 if (list_empty(&area
->free_list
[migratetype
]))
835 page
= list_entry(area
->free_list
[migratetype
].next
,
840 * If breaking a large block of pages, move all free
841 * pages to the preferred allocation list. If falling
842 * back for a reclaimable kernel allocation, be more
843 * agressive about taking ownership of free pages
845 if (unlikely(current_order
>= (pageblock_order
>> 1)) ||
846 start_migratetype
== MIGRATE_RECLAIMABLE
||
847 page_group_by_mobility_disabled
) {
849 pages
= move_freepages_block(zone
, page
,
852 /* Claim the whole block if over half of it is free */
853 if (pages
>= (1 << (pageblock_order
-1)) ||
854 page_group_by_mobility_disabled
)
855 set_pageblock_migratetype(page
,
858 migratetype
= start_migratetype
;
861 /* Remove the page from the freelists */
862 list_del(&page
->lru
);
863 rmv_page_order(page
);
865 /* Take ownership for orders >= pageblock_order */
866 if (current_order
>= pageblock_order
)
867 change_pageblock_range(page
, current_order
,
870 expand(zone
, page
, order
, current_order
, area
, migratetype
);
872 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
873 start_migratetype
, migratetype
);
883 * Do the hard work of removing an element from the buddy allocator.
884 * Call me with the zone->lock already held.
886 static struct page
*__rmqueue(struct zone
*zone
, unsigned int order
,
892 page
= __rmqueue_smallest(zone
, order
, migratetype
);
894 if (unlikely(!page
) && migratetype
!= MIGRATE_RESERVE
) {
895 page
= __rmqueue_fallback(zone
, order
, migratetype
);
898 * Use MIGRATE_RESERVE rather than fail an allocation. goto
899 * is used because __rmqueue_smallest is an inline function
900 * and we want just one call site
903 migratetype
= MIGRATE_RESERVE
;
908 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
913 * Obtain a specified number of elements from the buddy allocator, all under
914 * a single hold of the lock, for efficiency. Add them to the supplied list.
915 * Returns the number of new pages which were placed at *list.
917 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
918 unsigned long count
, struct list_head
*list
,
919 int migratetype
, int cold
)
923 spin_lock(&zone
->lock
);
924 for (i
= 0; i
< count
; ++i
) {
925 struct page
*page
= __rmqueue(zone
, order
, migratetype
);
926 if (unlikely(page
== NULL
))
930 * Split buddy pages returned by expand() are received here
931 * in physical page order. The page is added to the callers and
932 * list and the list head then moves forward. From the callers
933 * perspective, the linked list is ordered by page number in
934 * some conditions. This is useful for IO devices that can
935 * merge IO requests if the physical pages are ordered
938 if (likely(cold
== 0))
939 list_add(&page
->lru
, list
);
941 list_add_tail(&page
->lru
, list
);
942 set_page_private(page
, migratetype
);
945 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
946 spin_unlock(&zone
->lock
);
952 * Called from the vmstat counter updater to drain pagesets of this
953 * currently executing processor on remote nodes after they have
956 * Note that this function must be called with the thread pinned to
957 * a single processor.
959 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
964 local_irq_save(flags
);
965 if (pcp
->count
>= pcp
->batch
)
966 to_drain
= pcp
->batch
;
968 to_drain
= pcp
->count
;
969 free_pcppages_bulk(zone
, to_drain
, pcp
);
970 pcp
->count
-= to_drain
;
971 local_irq_restore(flags
);
976 * Drain pages of the indicated processor.
978 * The processor must either be the current processor and the
979 * thread pinned to the current processor or a processor that
982 static void drain_pages(unsigned int cpu
)
987 for_each_populated_zone(zone
) {
988 struct per_cpu_pageset
*pset
;
989 struct per_cpu_pages
*pcp
;
991 pset
= zone_pcp(zone
, cpu
);
994 local_irq_save(flags
);
995 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
997 local_irq_restore(flags
);
1002 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1004 void drain_local_pages(void *arg
)
1006 drain_pages(smp_processor_id());
1010 * Spill all the per-cpu pages from all CPUs back into the buddy allocator
1012 void drain_all_pages(void)
1014 on_each_cpu(drain_local_pages
, NULL
, 1);
1017 #ifdef CONFIG_HIBERNATION
1019 void mark_free_pages(struct zone
*zone
)
1021 unsigned long pfn
, max_zone_pfn
;
1022 unsigned long flags
;
1024 struct list_head
*curr
;
1026 if (!zone
->spanned_pages
)
1029 spin_lock_irqsave(&zone
->lock
, flags
);
1031 max_zone_pfn
= zone
->zone_start_pfn
+ zone
->spanned_pages
;
1032 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
1033 if (pfn_valid(pfn
)) {
1034 struct page
*page
= pfn_to_page(pfn
);
1036 if (!swsusp_page_is_forbidden(page
))
1037 swsusp_unset_page_free(page
);
1040 for_each_migratetype_order(order
, t
) {
1041 list_for_each(curr
, &zone
->free_area
[order
].free_list
[t
]) {
1044 pfn
= page_to_pfn(list_entry(curr
, struct page
, lru
));
1045 for (i
= 0; i
< (1UL << order
); i
++)
1046 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
1049 spin_unlock_irqrestore(&zone
->lock
, flags
);
1051 #endif /* CONFIG_PM */
1054 * Free a 0-order page
1056 static void free_hot_cold_page(struct page
*page
, int cold
)
1058 struct zone
*zone
= page_zone(page
);
1059 struct per_cpu_pages
*pcp
;
1060 unsigned long flags
;
1062 int wasMlocked
= __TestClearPageMlocked(page
);
1064 kmemcheck_free_shadow(page
, 0);
1067 page
->mapping
= NULL
;
1068 if (free_pages_check(page
))
1071 if (!PageHighMem(page
)) {
1072 debug_check_no_locks_freed(page_address(page
), PAGE_SIZE
);
1073 debug_check_no_obj_freed(page_address(page
), PAGE_SIZE
);
1075 arch_free_page(page
, 0);
1076 kernel_map_pages(page
, 1, 0);
1078 pcp
= &zone_pcp(zone
, get_cpu())->pcp
;
1079 migratetype
= get_pageblock_migratetype(page
);
1080 set_page_private(page
, migratetype
);
1081 local_irq_save(flags
);
1082 if (unlikely(wasMlocked
))
1083 free_page_mlock(page
);
1084 __count_vm_event(PGFREE
);
1087 * We only track unmovable, reclaimable and movable on pcp lists.
1088 * Free ISOLATE pages back to the allocator because they are being
1089 * offlined but treat RESERVE as movable pages so we can get those
1090 * areas back if necessary. Otherwise, we may have to free
1091 * excessively into the page allocator
1093 if (migratetype
>= MIGRATE_PCPTYPES
) {
1094 if (unlikely(migratetype
== MIGRATE_ISOLATE
)) {
1095 free_one_page(zone
, page
, 0, migratetype
);
1098 migratetype
= MIGRATE_MOVABLE
;
1102 list_add_tail(&page
->lru
, &pcp
->lists
[migratetype
]);
1104 list_add(&page
->lru
, &pcp
->lists
[migratetype
]);
1106 if (pcp
->count
>= pcp
->high
) {
1107 free_pcppages_bulk(zone
, pcp
->batch
, pcp
);
1108 pcp
->count
-= pcp
->batch
;
1112 local_irq_restore(flags
);
1116 void free_hot_page(struct page
*page
)
1118 trace_mm_page_free_direct(page
, 0);
1119 free_hot_cold_page(page
, 0);
1123 * split_page takes a non-compound higher-order page, and splits it into
1124 * n (1<<order) sub-pages: page[0..n]
1125 * Each sub-page must be freed individually.
1127 * Note: this is probably too low level an operation for use in drivers.
1128 * Please consult with lkml before using this in your driver.
1130 void split_page(struct page
*page
, unsigned int order
)
1134 VM_BUG_ON(PageCompound(page
));
1135 VM_BUG_ON(!page_count(page
));
1137 #ifdef CONFIG_KMEMCHECK
1139 * Split shadow pages too, because free(page[0]) would
1140 * otherwise free the whole shadow.
1142 if (kmemcheck_page_is_tracked(page
))
1143 split_page(virt_to_page(page
[0].shadow
), order
);
1146 for (i
= 1; i
< (1 << order
); i
++)
1147 set_page_refcounted(page
+ i
);
1151 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1152 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1156 struct page
*buffered_rmqueue(struct zone
*preferred_zone
,
1157 struct zone
*zone
, int order
, gfp_t gfp_flags
,
1160 unsigned long flags
;
1162 int cold
= !!(gfp_flags
& __GFP_COLD
);
1167 if (likely(order
== 0)) {
1168 struct per_cpu_pages
*pcp
;
1169 struct list_head
*list
;
1171 pcp
= &zone_pcp(zone
, cpu
)->pcp
;
1172 list
= &pcp
->lists
[migratetype
];
1173 local_irq_save(flags
);
1174 if (list_empty(list
)) {
1175 pcp
->count
+= rmqueue_bulk(zone
, 0,
1178 if (unlikely(list_empty(list
)))
1183 page
= list_entry(list
->prev
, struct page
, lru
);
1185 page
= list_entry(list
->next
, struct page
, lru
);
1187 list_del(&page
->lru
);
1190 if (unlikely(gfp_flags
& __GFP_NOFAIL
)) {
1192 * __GFP_NOFAIL is not to be used in new code.
1194 * All __GFP_NOFAIL callers should be fixed so that they
1195 * properly detect and handle allocation failures.
1197 * We most definitely don't want callers attempting to
1198 * allocate greater than order-1 page units with
1201 WARN_ON_ONCE(order
> 1);
1203 spin_lock_irqsave(&zone
->lock
, flags
);
1204 page
= __rmqueue(zone
, order
, migratetype
);
1205 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(1 << order
));
1206 spin_unlock(&zone
->lock
);
1211 __count_zone_vm_events(PGALLOC
, zone
, 1 << order
);
1212 zone_statistics(preferred_zone
, zone
);
1213 local_irq_restore(flags
);
1216 VM_BUG_ON(bad_range(zone
, page
));
1217 if (prep_new_page(page
, order
, gfp_flags
))
1222 local_irq_restore(flags
);
1227 /* The ALLOC_WMARK bits are used as an index to zone->watermark */
1228 #define ALLOC_WMARK_MIN WMARK_MIN
1229 #define ALLOC_WMARK_LOW WMARK_LOW
1230 #define ALLOC_WMARK_HIGH WMARK_HIGH
1231 #define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */
1233 /* Mask to get the watermark bits */
1234 #define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1)
1236 #define ALLOC_HARDER 0x10 /* try to alloc harder */
1237 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
1238 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
1240 #ifdef CONFIG_FAIL_PAGE_ALLOC
1242 static struct fail_page_alloc_attr
{
1243 struct fault_attr attr
;
1245 u32 ignore_gfp_highmem
;
1246 u32 ignore_gfp_wait
;
1249 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1251 struct dentry
*ignore_gfp_highmem_file
;
1252 struct dentry
*ignore_gfp_wait_file
;
1253 struct dentry
*min_order_file
;
1255 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1257 } fail_page_alloc
= {
1258 .attr
= FAULT_ATTR_INITIALIZER
,
1259 .ignore_gfp_wait
= 1,
1260 .ignore_gfp_highmem
= 1,
1264 static int __init
setup_fail_page_alloc(char *str
)
1266 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
1268 __setup("fail_page_alloc=", setup_fail_page_alloc
);
1270 static int should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1272 if (order
< fail_page_alloc
.min_order
)
1274 if (gfp_mask
& __GFP_NOFAIL
)
1276 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
1278 if (fail_page_alloc
.ignore_gfp_wait
&& (gfp_mask
& __GFP_WAIT
))
1281 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
1284 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1286 static int __init
fail_page_alloc_debugfs(void)
1288 mode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
1292 err
= init_fault_attr_dentries(&fail_page_alloc
.attr
,
1296 dir
= fail_page_alloc
.attr
.dentries
.dir
;
1298 fail_page_alloc
.ignore_gfp_wait_file
=
1299 debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
1300 &fail_page_alloc
.ignore_gfp_wait
);
1302 fail_page_alloc
.ignore_gfp_highmem_file
=
1303 debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
1304 &fail_page_alloc
.ignore_gfp_highmem
);
1305 fail_page_alloc
.min_order_file
=
1306 debugfs_create_u32("min-order", mode
, dir
,
1307 &fail_page_alloc
.min_order
);
1309 if (!fail_page_alloc
.ignore_gfp_wait_file
||
1310 !fail_page_alloc
.ignore_gfp_highmem_file
||
1311 !fail_page_alloc
.min_order_file
) {
1313 debugfs_remove(fail_page_alloc
.ignore_gfp_wait_file
);
1314 debugfs_remove(fail_page_alloc
.ignore_gfp_highmem_file
);
1315 debugfs_remove(fail_page_alloc
.min_order_file
);
1316 cleanup_fault_attr_dentries(&fail_page_alloc
.attr
);
1322 late_initcall(fail_page_alloc_debugfs
);
1324 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1326 #else /* CONFIG_FAIL_PAGE_ALLOC */
1328 static inline int should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1333 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1336 * Return 1 if free pages are above 'mark'. This takes into account the order
1337 * of the allocation.
1339 int zone_watermark_ok(struct zone
*z
, int order
, unsigned long mark
,
1340 int classzone_idx
, int alloc_flags
)
1342 /* free_pages my go negative - that's OK */
1344 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
) - (1 << order
) + 1;
1347 if (alloc_flags
& ALLOC_HIGH
)
1349 if (alloc_flags
& ALLOC_HARDER
)
1352 if (free_pages
<= min
+ z
->lowmem_reserve
[classzone_idx
])
1354 for (o
= 0; o
< order
; o
++) {
1355 /* At the next order, this order's pages become unavailable */
1356 free_pages
-= z
->free_area
[o
].nr_free
<< o
;
1358 /* Require fewer higher order pages to be free */
1361 if (free_pages
<= min
)
1369 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1370 * skip over zones that are not allowed by the cpuset, or that have
1371 * been recently (in last second) found to be nearly full. See further
1372 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1373 * that have to skip over a lot of full or unallowed zones.
1375 * If the zonelist cache is present in the passed in zonelist, then
1376 * returns a pointer to the allowed node mask (either the current
1377 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1379 * If the zonelist cache is not available for this zonelist, does
1380 * nothing and returns NULL.
1382 * If the fullzones BITMAP in the zonelist cache is stale (more than
1383 * a second since last zap'd) then we zap it out (clear its bits.)
1385 * We hold off even calling zlc_setup, until after we've checked the
1386 * first zone in the zonelist, on the theory that most allocations will
1387 * be satisfied from that first zone, so best to examine that zone as
1388 * quickly as we can.
1390 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1392 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1393 nodemask_t
*allowednodes
; /* zonelist_cache approximation */
1395 zlc
= zonelist
->zlcache_ptr
;
1399 if (time_after(jiffies
, zlc
->last_full_zap
+ HZ
)) {
1400 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
1401 zlc
->last_full_zap
= jiffies
;
1404 allowednodes
= !in_interrupt() && (alloc_flags
& ALLOC_CPUSET
) ?
1405 &cpuset_current_mems_allowed
:
1406 &node_states
[N_HIGH_MEMORY
];
1407 return allowednodes
;
1411 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1412 * if it is worth looking at further for free memory:
1413 * 1) Check that the zone isn't thought to be full (doesn't have its
1414 * bit set in the zonelist_cache fullzones BITMAP).
1415 * 2) Check that the zones node (obtained from the zonelist_cache
1416 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1417 * Return true (non-zero) if zone is worth looking at further, or
1418 * else return false (zero) if it is not.
1420 * This check -ignores- the distinction between various watermarks,
1421 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1422 * found to be full for any variation of these watermarks, it will
1423 * be considered full for up to one second by all requests, unless
1424 * we are so low on memory on all allowed nodes that we are forced
1425 * into the second scan of the zonelist.
1427 * In the second scan we ignore this zonelist cache and exactly
1428 * apply the watermarks to all zones, even it is slower to do so.
1429 * We are low on memory in the second scan, and should leave no stone
1430 * unturned looking for a free page.
1432 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1433 nodemask_t
*allowednodes
)
1435 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1436 int i
; /* index of *z in zonelist zones */
1437 int n
; /* node that zone *z is on */
1439 zlc
= zonelist
->zlcache_ptr
;
1443 i
= z
- zonelist
->_zonerefs
;
1446 /* This zone is worth trying if it is allowed but not full */
1447 return node_isset(n
, *allowednodes
) && !test_bit(i
, zlc
->fullzones
);
1451 * Given 'z' scanning a zonelist, set the corresponding bit in
1452 * zlc->fullzones, so that subsequent attempts to allocate a page
1453 * from that zone don't waste time re-examining it.
1455 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1457 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1458 int i
; /* index of *z in zonelist zones */
1460 zlc
= zonelist
->zlcache_ptr
;
1464 i
= z
- zonelist
->_zonerefs
;
1466 set_bit(i
, zlc
->fullzones
);
1469 #else /* CONFIG_NUMA */
1471 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1476 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1477 nodemask_t
*allowednodes
)
1482 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1485 #endif /* CONFIG_NUMA */
1488 * get_page_from_freelist goes through the zonelist trying to allocate
1491 static struct page
*
1492 get_page_from_freelist(gfp_t gfp_mask
, nodemask_t
*nodemask
, unsigned int order
,
1493 struct zonelist
*zonelist
, int high_zoneidx
, int alloc_flags
,
1494 struct zone
*preferred_zone
, int migratetype
)
1497 struct page
*page
= NULL
;
1500 nodemask_t
*allowednodes
= NULL
;/* zonelist_cache approximation */
1501 int zlc_active
= 0; /* set if using zonelist_cache */
1502 int did_zlc_setup
= 0; /* just call zlc_setup() one time */
1504 classzone_idx
= zone_idx(preferred_zone
);
1507 * Scan zonelist, looking for a zone with enough free.
1508 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1510 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
1511 high_zoneidx
, nodemask
) {
1512 if (NUMA_BUILD
&& zlc_active
&&
1513 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
1515 if ((alloc_flags
& ALLOC_CPUSET
) &&
1516 !cpuset_zone_allowed_softwall(zone
, gfp_mask
))
1519 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
1520 if (!(alloc_flags
& ALLOC_NO_WATERMARKS
)) {
1524 mark
= zone
->watermark
[alloc_flags
& ALLOC_WMARK_MASK
];
1525 if (zone_watermark_ok(zone
, order
, mark
,
1526 classzone_idx
, alloc_flags
))
1529 if (zone_reclaim_mode
== 0)
1530 goto this_zone_full
;
1532 ret
= zone_reclaim(zone
, gfp_mask
, order
);
1534 case ZONE_RECLAIM_NOSCAN
:
1537 case ZONE_RECLAIM_FULL
:
1538 /* scanned but unreclaimable */
1539 goto this_zone_full
;
1541 /* did we reclaim enough */
1542 if (!zone_watermark_ok(zone
, order
, mark
,
1543 classzone_idx
, alloc_flags
))
1544 goto this_zone_full
;
1549 page
= buffered_rmqueue(preferred_zone
, zone
, order
,
1550 gfp_mask
, migratetype
);
1555 zlc_mark_zone_full(zonelist
, z
);
1557 if (NUMA_BUILD
&& !did_zlc_setup
&& nr_online_nodes
> 1) {
1559 * we do zlc_setup after the first zone is tried but only
1560 * if there are multiple nodes make it worthwhile
1562 allowednodes
= zlc_setup(zonelist
, alloc_flags
);
1568 if (unlikely(NUMA_BUILD
&& page
== NULL
&& zlc_active
)) {
1569 /* Disable zlc cache for second zonelist scan */
1577 should_alloc_retry(gfp_t gfp_mask
, unsigned int order
,
1578 unsigned long pages_reclaimed
)
1580 /* Do not loop if specifically requested */
1581 if (gfp_mask
& __GFP_NORETRY
)
1585 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
1586 * means __GFP_NOFAIL, but that may not be true in other
1589 if (order
<= PAGE_ALLOC_COSTLY_ORDER
)
1593 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
1594 * specified, then we retry until we no longer reclaim any pages
1595 * (above), or we've reclaimed an order of pages at least as
1596 * large as the allocation's order. In both cases, if the
1597 * allocation still fails, we stop retrying.
1599 if (gfp_mask
& __GFP_REPEAT
&& pages_reclaimed
< (1 << order
))
1603 * Don't let big-order allocations loop unless the caller
1604 * explicitly requests that.
1606 if (gfp_mask
& __GFP_NOFAIL
)
1612 static inline struct page
*
1613 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
1614 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1615 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
1620 /* Acquire the OOM killer lock for the zones in zonelist */
1621 if (!try_set_zone_oom(zonelist
, gfp_mask
)) {
1622 schedule_timeout_uninterruptible(1);
1627 * Go through the zonelist yet one more time, keep very high watermark
1628 * here, this is only to catch a parallel oom killing, we must fail if
1629 * we're still under heavy pressure.
1631 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
,
1632 order
, zonelist
, high_zoneidx
,
1633 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
,
1634 preferred_zone
, migratetype
);
1638 /* The OOM killer will not help higher order allocs */
1639 if (order
> PAGE_ALLOC_COSTLY_ORDER
&& !(gfp_mask
& __GFP_NOFAIL
))
1642 /* Exhausted what can be done so it's blamo time */
1643 out_of_memory(zonelist
, gfp_mask
, order
);
1646 clear_zonelist_oom(zonelist
, gfp_mask
);
1650 /* The really slow allocator path where we enter direct reclaim */
1651 static inline struct page
*
1652 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
1653 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1654 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
1655 int migratetype
, unsigned long *did_some_progress
)
1657 struct page
*page
= NULL
;
1658 struct reclaim_state reclaim_state
;
1659 struct task_struct
*p
= current
;
1663 /* We now go into synchronous reclaim */
1664 cpuset_memory_pressure_bump();
1665 p
->flags
|= PF_MEMALLOC
;
1666 lockdep_set_current_reclaim_state(gfp_mask
);
1667 reclaim_state
.reclaimed_slab
= 0;
1668 p
->reclaim_state
= &reclaim_state
;
1670 *did_some_progress
= try_to_free_pages(zonelist
, order
, gfp_mask
, nodemask
);
1672 p
->reclaim_state
= NULL
;
1673 lockdep_clear_current_reclaim_state();
1674 p
->flags
&= ~PF_MEMALLOC
;
1681 if (likely(*did_some_progress
))
1682 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
1683 zonelist
, high_zoneidx
,
1684 alloc_flags
, preferred_zone
,
1690 * This is called in the allocator slow-path if the allocation request is of
1691 * sufficient urgency to ignore watermarks and take other desperate measures
1693 static inline struct page
*
1694 __alloc_pages_high_priority(gfp_t gfp_mask
, unsigned int order
,
1695 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1696 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
1702 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
1703 zonelist
, high_zoneidx
, ALLOC_NO_WATERMARKS
,
1704 preferred_zone
, migratetype
);
1706 if (!page
&& gfp_mask
& __GFP_NOFAIL
)
1707 congestion_wait(BLK_RW_ASYNC
, HZ
/50);
1708 } while (!page
&& (gfp_mask
& __GFP_NOFAIL
));
1714 void wake_all_kswapd(unsigned int order
, struct zonelist
*zonelist
,
1715 enum zone_type high_zoneidx
)
1720 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
)
1721 wakeup_kswapd(zone
, order
);
1725 gfp_to_alloc_flags(gfp_t gfp_mask
)
1727 struct task_struct
*p
= current
;
1728 int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
1729 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
1731 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
1732 BUILD_BUG_ON(__GFP_HIGH
!= ALLOC_HIGH
);
1735 * The caller may dip into page reserves a bit more if the caller
1736 * cannot run direct reclaim, or if the caller has realtime scheduling
1737 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
1738 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
1740 alloc_flags
|= (gfp_mask
& __GFP_HIGH
);
1743 alloc_flags
|= ALLOC_HARDER
;
1745 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
1746 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1748 alloc_flags
&= ~ALLOC_CPUSET
;
1749 } else if (unlikely(rt_task(p
)))
1750 alloc_flags
|= ALLOC_HARDER
;
1752 if (likely(!(gfp_mask
& __GFP_NOMEMALLOC
))) {
1753 if (!in_interrupt() &&
1754 ((p
->flags
& PF_MEMALLOC
) ||
1755 unlikely(test_thread_flag(TIF_MEMDIE
))))
1756 alloc_flags
|= ALLOC_NO_WATERMARKS
;
1762 static inline struct page
*
1763 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
1764 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1765 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
1768 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
1769 struct page
*page
= NULL
;
1771 unsigned long pages_reclaimed
= 0;
1772 unsigned long did_some_progress
;
1773 struct task_struct
*p
= current
;
1776 * In the slowpath, we sanity check order to avoid ever trying to
1777 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
1778 * be using allocators in order of preference for an area that is
1781 if (order
>= MAX_ORDER
) {
1782 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
1787 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
1788 * __GFP_NOWARN set) should not cause reclaim since the subsystem
1789 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
1790 * using a larger set of nodes after it has established that the
1791 * allowed per node queues are empty and that nodes are
1794 if (NUMA_BUILD
&& (gfp_mask
& GFP_THISNODE
) == GFP_THISNODE
)
1797 wake_all_kswapd(order
, zonelist
, high_zoneidx
);
1801 * OK, we're below the kswapd watermark and have kicked background
1802 * reclaim. Now things get more complex, so set up alloc_flags according
1803 * to how we want to proceed.
1805 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
1807 /* This is the last chance, in general, before the goto nopage. */
1808 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
, zonelist
,
1809 high_zoneidx
, alloc_flags
& ~ALLOC_NO_WATERMARKS
,
1810 preferred_zone
, migratetype
);
1815 /* Allocate without watermarks if the context allows */
1816 if (alloc_flags
& ALLOC_NO_WATERMARKS
) {
1817 page
= __alloc_pages_high_priority(gfp_mask
, order
,
1818 zonelist
, high_zoneidx
, nodemask
,
1819 preferred_zone
, migratetype
);
1824 /* Atomic allocations - we can't balance anything */
1828 /* Avoid recursion of direct reclaim */
1829 if (p
->flags
& PF_MEMALLOC
)
1832 /* Avoid allocations with no watermarks from looping endlessly */
1833 if (test_thread_flag(TIF_MEMDIE
) && !(gfp_mask
& __GFP_NOFAIL
))
1836 /* Try direct reclaim and then allocating */
1837 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
,
1838 zonelist
, high_zoneidx
,
1840 alloc_flags
, preferred_zone
,
1841 migratetype
, &did_some_progress
);
1846 * If we failed to make any progress reclaiming, then we are
1847 * running out of options and have to consider going OOM
1849 if (!did_some_progress
) {
1850 if ((gfp_mask
& __GFP_FS
) && !(gfp_mask
& __GFP_NORETRY
)) {
1851 if (oom_killer_disabled
)
1853 page
= __alloc_pages_may_oom(gfp_mask
, order
,
1854 zonelist
, high_zoneidx
,
1855 nodemask
, preferred_zone
,
1861 * The OOM killer does not trigger for high-order
1862 * ~__GFP_NOFAIL allocations so if no progress is being
1863 * made, there are no other options and retrying is
1866 if (order
> PAGE_ALLOC_COSTLY_ORDER
&&
1867 !(gfp_mask
& __GFP_NOFAIL
))
1874 /* Check if we should retry the allocation */
1875 pages_reclaimed
+= did_some_progress
;
1876 if (should_alloc_retry(gfp_mask
, order
, pages_reclaimed
)) {
1877 /* Wait for some write requests to complete then retry */
1878 congestion_wait(BLK_RW_ASYNC
, HZ
/50);
1883 if (!(gfp_mask
& __GFP_NOWARN
) && printk_ratelimit()) {
1884 printk(KERN_WARNING
"%s: page allocation failure."
1885 " order:%d, mode:0x%x\n",
1886 p
->comm
, order
, gfp_mask
);
1892 if (kmemcheck_enabled
)
1893 kmemcheck_pagealloc_alloc(page
, order
, gfp_mask
);
1899 * This is the 'heart' of the zoned buddy allocator.
1902 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
,
1903 struct zonelist
*zonelist
, nodemask_t
*nodemask
)
1905 enum zone_type high_zoneidx
= gfp_zone(gfp_mask
);
1906 struct zone
*preferred_zone
;
1908 int migratetype
= allocflags_to_migratetype(gfp_mask
);
1910 gfp_mask
&= gfp_allowed_mask
;
1912 lockdep_trace_alloc(gfp_mask
);
1914 might_sleep_if(gfp_mask
& __GFP_WAIT
);
1916 if (should_fail_alloc_page(gfp_mask
, order
))
1920 * Check the zones suitable for the gfp_mask contain at least one
1921 * valid zone. It's possible to have an empty zonelist as a result
1922 * of GFP_THISNODE and a memoryless node
1924 if (unlikely(!zonelist
->_zonerefs
->zone
))
1927 /* The preferred zone is used for statistics later */
1928 first_zones_zonelist(zonelist
, high_zoneidx
, nodemask
, &preferred_zone
);
1929 if (!preferred_zone
)
1932 /* First allocation attempt */
1933 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
, order
,
1934 zonelist
, high_zoneidx
, ALLOC_WMARK_LOW
|ALLOC_CPUSET
,
1935 preferred_zone
, migratetype
);
1936 if (unlikely(!page
))
1937 page
= __alloc_pages_slowpath(gfp_mask
, order
,
1938 zonelist
, high_zoneidx
, nodemask
,
1939 preferred_zone
, migratetype
);
1941 trace_mm_page_alloc(page
, order
, gfp_mask
, migratetype
);
1944 EXPORT_SYMBOL(__alloc_pages_nodemask
);
1947 * Common helper functions.
1949 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
1954 * __get_free_pages() returns a 32-bit address, which cannot represent
1957 VM_BUG_ON((gfp_mask
& __GFP_HIGHMEM
) != 0);
1959 page
= alloc_pages(gfp_mask
, order
);
1962 return (unsigned long) page_address(page
);
1964 EXPORT_SYMBOL(__get_free_pages
);
1966 unsigned long get_zeroed_page(gfp_t gfp_mask
)
1968 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
1970 EXPORT_SYMBOL(get_zeroed_page
);
1972 void __pagevec_free(struct pagevec
*pvec
)
1974 int i
= pagevec_count(pvec
);
1977 trace_mm_pagevec_free(pvec
->pages
[i
], pvec
->cold
);
1978 free_hot_cold_page(pvec
->pages
[i
], pvec
->cold
);
1982 void __free_pages(struct page
*page
, unsigned int order
)
1984 if (put_page_testzero(page
)) {
1985 trace_mm_page_free_direct(page
, order
);
1987 free_hot_page(page
);
1989 __free_pages_ok(page
, order
);
1993 EXPORT_SYMBOL(__free_pages
);
1995 void free_pages(unsigned long addr
, unsigned int order
)
1998 VM_BUG_ON(!virt_addr_valid((void *)addr
));
1999 __free_pages(virt_to_page((void *)addr
), order
);
2003 EXPORT_SYMBOL(free_pages
);
2006 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2007 * @size: the number of bytes to allocate
2008 * @gfp_mask: GFP flags for the allocation
2010 * This function is similar to alloc_pages(), except that it allocates the
2011 * minimum number of pages to satisfy the request. alloc_pages() can only
2012 * allocate memory in power-of-two pages.
2014 * This function is also limited by MAX_ORDER.
2016 * Memory allocated by this function must be released by free_pages_exact().
2018 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
2020 unsigned int order
= get_order(size
);
2023 addr
= __get_free_pages(gfp_mask
, order
);
2025 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
2026 unsigned long used
= addr
+ PAGE_ALIGN(size
);
2028 split_page(virt_to_page((void *)addr
), order
);
2029 while (used
< alloc_end
) {
2035 return (void *)addr
;
2037 EXPORT_SYMBOL(alloc_pages_exact
);
2040 * free_pages_exact - release memory allocated via alloc_pages_exact()
2041 * @virt: the value returned by alloc_pages_exact.
2042 * @size: size of allocation, same value as passed to alloc_pages_exact().
2044 * Release the memory allocated by a previous call to alloc_pages_exact.
2046 void free_pages_exact(void *virt
, size_t size
)
2048 unsigned long addr
= (unsigned long)virt
;
2049 unsigned long end
= addr
+ PAGE_ALIGN(size
);
2051 while (addr
< end
) {
2056 EXPORT_SYMBOL(free_pages_exact
);
2058 static unsigned int nr_free_zone_pages(int offset
)
2063 /* Just pick one node, since fallback list is circular */
2064 unsigned int sum
= 0;
2066 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
2068 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
2069 unsigned long size
= zone
->present_pages
;
2070 unsigned long high
= high_wmark_pages(zone
);
2079 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2081 unsigned int nr_free_buffer_pages(void)
2083 return nr_free_zone_pages(gfp_zone(GFP_USER
));
2085 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
2088 * Amount of free RAM allocatable within all zones
2090 unsigned int nr_free_pagecache_pages(void)
2092 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
2095 static inline void show_node(struct zone
*zone
)
2098 printk("Node %d ", zone_to_nid(zone
));
2101 void si_meminfo(struct sysinfo
*val
)
2103 val
->totalram
= totalram_pages
;
2105 val
->freeram
= global_page_state(NR_FREE_PAGES
);
2106 val
->bufferram
= nr_blockdev_pages();
2107 val
->totalhigh
= totalhigh_pages
;
2108 val
->freehigh
= nr_free_highpages();
2109 val
->mem_unit
= PAGE_SIZE
;
2112 EXPORT_SYMBOL(si_meminfo
);
2115 void si_meminfo_node(struct sysinfo
*val
, int nid
)
2117 pg_data_t
*pgdat
= NODE_DATA(nid
);
2119 val
->totalram
= pgdat
->node_present_pages
;
2120 val
->freeram
= node_page_state(nid
, NR_FREE_PAGES
);
2121 #ifdef CONFIG_HIGHMEM
2122 val
->totalhigh
= pgdat
->node_zones
[ZONE_HIGHMEM
].present_pages
;
2123 val
->freehigh
= zone_page_state(&pgdat
->node_zones
[ZONE_HIGHMEM
],
2129 val
->mem_unit
= PAGE_SIZE
;
2133 #define K(x) ((x) << (PAGE_SHIFT-10))
2136 * Show free area list (used inside shift_scroll-lock stuff)
2137 * We also calculate the percentage fragmentation. We do this by counting the
2138 * memory on each free list with the exception of the first item on the list.
2140 void show_free_areas(void)
2145 for_each_populated_zone(zone
) {
2147 printk("%s per-cpu:\n", zone
->name
);
2149 for_each_online_cpu(cpu
) {
2150 struct per_cpu_pageset
*pageset
;
2152 pageset
= zone_pcp(zone
, cpu
);
2154 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2155 cpu
, pageset
->pcp
.high
,
2156 pageset
->pcp
.batch
, pageset
->pcp
.count
);
2160 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2161 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2163 " dirty:%lu writeback:%lu unstable:%lu buffer:%lu\n"
2164 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2165 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n",
2166 global_page_state(NR_ACTIVE_ANON
),
2167 global_page_state(NR_INACTIVE_ANON
),
2168 global_page_state(NR_ISOLATED_ANON
),
2169 global_page_state(NR_ACTIVE_FILE
),
2170 global_page_state(NR_INACTIVE_FILE
),
2171 global_page_state(NR_ISOLATED_FILE
),
2172 global_page_state(NR_UNEVICTABLE
),
2173 global_page_state(NR_FILE_DIRTY
),
2174 global_page_state(NR_WRITEBACK
),
2175 global_page_state(NR_UNSTABLE_NFS
),
2176 nr_blockdev_pages(),
2177 global_page_state(NR_FREE_PAGES
),
2178 global_page_state(NR_SLAB_RECLAIMABLE
),
2179 global_page_state(NR_SLAB_UNRECLAIMABLE
),
2180 global_page_state(NR_FILE_MAPPED
),
2181 global_page_state(NR_SHMEM
),
2182 global_page_state(NR_PAGETABLE
),
2183 global_page_state(NR_BOUNCE
));
2185 for_each_populated_zone(zone
) {
2194 " active_anon:%lukB"
2195 " inactive_anon:%lukB"
2196 " active_file:%lukB"
2197 " inactive_file:%lukB"
2198 " unevictable:%lukB"
2199 " isolated(anon):%lukB"
2200 " isolated(file):%lukB"
2207 " slab_reclaimable:%lukB"
2208 " slab_unreclaimable:%lukB"
2209 " kernel_stack:%lukB"
2213 " writeback_tmp:%lukB"
2214 " pages_scanned:%lu"
2215 " all_unreclaimable? %s"
2218 K(zone_page_state(zone
, NR_FREE_PAGES
)),
2219 K(min_wmark_pages(zone
)),
2220 K(low_wmark_pages(zone
)),
2221 K(high_wmark_pages(zone
)),
2222 K(zone_page_state(zone
, NR_ACTIVE_ANON
)),
2223 K(zone_page_state(zone
, NR_INACTIVE_ANON
)),
2224 K(zone_page_state(zone
, NR_ACTIVE_FILE
)),
2225 K(zone_page_state(zone
, NR_INACTIVE_FILE
)),
2226 K(zone_page_state(zone
, NR_UNEVICTABLE
)),
2227 K(zone_page_state(zone
, NR_ISOLATED_ANON
)),
2228 K(zone_page_state(zone
, NR_ISOLATED_FILE
)),
2229 K(zone
->present_pages
),
2230 K(zone_page_state(zone
, NR_MLOCK
)),
2231 K(zone_page_state(zone
, NR_FILE_DIRTY
)),
2232 K(zone_page_state(zone
, NR_WRITEBACK
)),
2233 K(zone_page_state(zone
, NR_FILE_MAPPED
)),
2234 K(zone_page_state(zone
, NR_SHMEM
)),
2235 K(zone_page_state(zone
, NR_SLAB_RECLAIMABLE
)),
2236 K(zone_page_state(zone
, NR_SLAB_UNRECLAIMABLE
)),
2237 zone_page_state(zone
, NR_KERNEL_STACK
) *
2239 K(zone_page_state(zone
, NR_PAGETABLE
)),
2240 K(zone_page_state(zone
, NR_UNSTABLE_NFS
)),
2241 K(zone_page_state(zone
, NR_BOUNCE
)),
2242 K(zone_page_state(zone
, NR_WRITEBACK_TEMP
)),
2243 zone
->pages_scanned
,
2244 (zone_is_all_unreclaimable(zone
) ? "yes" : "no")
2246 printk("lowmem_reserve[]:");
2247 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
2248 printk(" %lu", zone
->lowmem_reserve
[i
]);
2252 for_each_populated_zone(zone
) {
2253 unsigned long nr
[MAX_ORDER
], flags
, order
, total
= 0;
2256 printk("%s: ", zone
->name
);
2258 spin_lock_irqsave(&zone
->lock
, flags
);
2259 for (order
= 0; order
< MAX_ORDER
; order
++) {
2260 nr
[order
] = zone
->free_area
[order
].nr_free
;
2261 total
+= nr
[order
] << order
;
2263 spin_unlock_irqrestore(&zone
->lock
, flags
);
2264 for (order
= 0; order
< MAX_ORDER
; order
++)
2265 printk("%lu*%lukB ", nr
[order
], K(1UL) << order
);
2266 printk("= %lukB\n", K(total
));
2269 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES
));
2271 show_swap_cache_info();
2274 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
2276 zoneref
->zone
= zone
;
2277 zoneref
->zone_idx
= zone_idx(zone
);
2281 * Builds allocation fallback zone lists.
2283 * Add all populated zones of a node to the zonelist.
2285 static int build_zonelists_node(pg_data_t
*pgdat
, struct zonelist
*zonelist
,
2286 int nr_zones
, enum zone_type zone_type
)
2290 BUG_ON(zone_type
>= MAX_NR_ZONES
);
2295 zone
= pgdat
->node_zones
+ zone_type
;
2296 if (populated_zone(zone
)) {
2297 zoneref_set_zone(zone
,
2298 &zonelist
->_zonerefs
[nr_zones
++]);
2299 check_highest_zone(zone_type
);
2302 } while (zone_type
);
2309 * 0 = automatic detection of better ordering.
2310 * 1 = order by ([node] distance, -zonetype)
2311 * 2 = order by (-zonetype, [node] distance)
2313 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
2314 * the same zonelist. So only NUMA can configure this param.
2316 #define ZONELIST_ORDER_DEFAULT 0
2317 #define ZONELIST_ORDER_NODE 1
2318 #define ZONELIST_ORDER_ZONE 2
2320 /* zonelist order in the kernel.
2321 * set_zonelist_order() will set this to NODE or ZONE.
2323 static int current_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
2324 static char zonelist_order_name
[3][8] = {"Default", "Node", "Zone"};
2328 /* The value user specified ....changed by config */
2329 static int user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
2330 /* string for sysctl */
2331 #define NUMA_ZONELIST_ORDER_LEN 16
2332 char numa_zonelist_order
[16] = "default";
2335 * interface for configure zonelist ordering.
2336 * command line option "numa_zonelist_order"
2337 * = "[dD]efault - default, automatic configuration.
2338 * = "[nN]ode - order by node locality, then by zone within node
2339 * = "[zZ]one - order by zone, then by locality within zone
2342 static int __parse_numa_zonelist_order(char *s
)
2344 if (*s
== 'd' || *s
== 'D') {
2345 user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
2346 } else if (*s
== 'n' || *s
== 'N') {
2347 user_zonelist_order
= ZONELIST_ORDER_NODE
;
2348 } else if (*s
== 'z' || *s
== 'Z') {
2349 user_zonelist_order
= ZONELIST_ORDER_ZONE
;
2352 "Ignoring invalid numa_zonelist_order value: "
2359 static __init
int setup_numa_zonelist_order(char *s
)
2362 return __parse_numa_zonelist_order(s
);
2365 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
2368 * sysctl handler for numa_zonelist_order
2370 int numa_zonelist_order_handler(ctl_table
*table
, int write
,
2371 struct file
*file
, void __user
*buffer
, size_t *length
,
2374 char saved_string
[NUMA_ZONELIST_ORDER_LEN
];
2378 strncpy(saved_string
, (char*)table
->data
,
2379 NUMA_ZONELIST_ORDER_LEN
);
2380 ret
= proc_dostring(table
, write
, file
, buffer
, length
, ppos
);
2384 int oldval
= user_zonelist_order
;
2385 if (__parse_numa_zonelist_order((char*)table
->data
)) {
2387 * bogus value. restore saved string
2389 strncpy((char*)table
->data
, saved_string
,
2390 NUMA_ZONELIST_ORDER_LEN
);
2391 user_zonelist_order
= oldval
;
2392 } else if (oldval
!= user_zonelist_order
)
2393 build_all_zonelists();
2399 #define MAX_NODE_LOAD (nr_online_nodes)
2400 static int node_load
[MAX_NUMNODES
];
2403 * find_next_best_node - find the next node that should appear in a given node's fallback list
2404 * @node: node whose fallback list we're appending
2405 * @used_node_mask: nodemask_t of already used nodes
2407 * We use a number of factors to determine which is the next node that should
2408 * appear on a given node's fallback list. The node should not have appeared
2409 * already in @node's fallback list, and it should be the next closest node
2410 * according to the distance array (which contains arbitrary distance values
2411 * from each node to each node in the system), and should also prefer nodes
2412 * with no CPUs, since presumably they'll have very little allocation pressure
2413 * on them otherwise.
2414 * It returns -1 if no node is found.
2416 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
2419 int min_val
= INT_MAX
;
2421 const struct cpumask
*tmp
= cpumask_of_node(0);
2423 /* Use the local node if we haven't already */
2424 if (!node_isset(node
, *used_node_mask
)) {
2425 node_set(node
, *used_node_mask
);
2429 for_each_node_state(n
, N_HIGH_MEMORY
) {
2431 /* Don't want a node to appear more than once */
2432 if (node_isset(n
, *used_node_mask
))
2435 /* Use the distance array to find the distance */
2436 val
= node_distance(node
, n
);
2438 /* Penalize nodes under us ("prefer the next node") */
2441 /* Give preference to headless and unused nodes */
2442 tmp
= cpumask_of_node(n
);
2443 if (!cpumask_empty(tmp
))
2444 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
2446 /* Slight preference for less loaded node */
2447 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
2448 val
+= node_load
[n
];
2450 if (val
< min_val
) {
2457 node_set(best_node
, *used_node_mask
);
2464 * Build zonelists ordered by node and zones within node.
2465 * This results in maximum locality--normal zone overflows into local
2466 * DMA zone, if any--but risks exhausting DMA zone.
2468 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int node
)
2471 struct zonelist
*zonelist
;
2473 zonelist
= &pgdat
->node_zonelists
[0];
2474 for (j
= 0; zonelist
->_zonerefs
[j
].zone
!= NULL
; j
++)
2476 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
2478 zonelist
->_zonerefs
[j
].zone
= NULL
;
2479 zonelist
->_zonerefs
[j
].zone_idx
= 0;
2483 * Build gfp_thisnode zonelists
2485 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
2488 struct zonelist
*zonelist
;
2490 zonelist
= &pgdat
->node_zonelists
[1];
2491 j
= build_zonelists_node(pgdat
, zonelist
, 0, MAX_NR_ZONES
- 1);
2492 zonelist
->_zonerefs
[j
].zone
= NULL
;
2493 zonelist
->_zonerefs
[j
].zone_idx
= 0;
2497 * Build zonelists ordered by zone and nodes within zones.
2498 * This results in conserving DMA zone[s] until all Normal memory is
2499 * exhausted, but results in overflowing to remote node while memory
2500 * may still exist in local DMA zone.
2502 static int node_order
[MAX_NUMNODES
];
2504 static void build_zonelists_in_zone_order(pg_data_t
*pgdat
, int nr_nodes
)
2507 int zone_type
; /* needs to be signed */
2509 struct zonelist
*zonelist
;
2511 zonelist
= &pgdat
->node_zonelists
[0];
2513 for (zone_type
= MAX_NR_ZONES
- 1; zone_type
>= 0; zone_type
--) {
2514 for (j
= 0; j
< nr_nodes
; j
++) {
2515 node
= node_order
[j
];
2516 z
= &NODE_DATA(node
)->node_zones
[zone_type
];
2517 if (populated_zone(z
)) {
2519 &zonelist
->_zonerefs
[pos
++]);
2520 check_highest_zone(zone_type
);
2524 zonelist
->_zonerefs
[pos
].zone
= NULL
;
2525 zonelist
->_zonerefs
[pos
].zone_idx
= 0;
2528 static int default_zonelist_order(void)
2531 unsigned long low_kmem_size
,total_size
;
2535 * ZONE_DMA and ZONE_DMA32 can be very small area in the sytem.
2536 * If they are really small and used heavily, the system can fall
2537 * into OOM very easily.
2538 * This function detect ZONE_DMA/DMA32 size and confgigures zone order.
2540 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
2543 for_each_online_node(nid
) {
2544 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
2545 z
= &NODE_DATA(nid
)->node_zones
[zone_type
];
2546 if (populated_zone(z
)) {
2547 if (zone_type
< ZONE_NORMAL
)
2548 low_kmem_size
+= z
->present_pages
;
2549 total_size
+= z
->present_pages
;
2553 if (!low_kmem_size
|| /* there are no DMA area. */
2554 low_kmem_size
> total_size
/2) /* DMA/DMA32 is big. */
2555 return ZONELIST_ORDER_NODE
;
2557 * look into each node's config.
2558 * If there is a node whose DMA/DMA32 memory is very big area on
2559 * local memory, NODE_ORDER may be suitable.
2561 average_size
= total_size
/
2562 (nodes_weight(node_states
[N_HIGH_MEMORY
]) + 1);
2563 for_each_online_node(nid
) {
2566 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
2567 z
= &NODE_DATA(nid
)->node_zones
[zone_type
];
2568 if (populated_zone(z
)) {
2569 if (zone_type
< ZONE_NORMAL
)
2570 low_kmem_size
+= z
->present_pages
;
2571 total_size
+= z
->present_pages
;
2574 if (low_kmem_size
&&
2575 total_size
> average_size
&& /* ignore small node */
2576 low_kmem_size
> total_size
* 70/100)
2577 return ZONELIST_ORDER_NODE
;
2579 return ZONELIST_ORDER_ZONE
;
2582 static void set_zonelist_order(void)
2584 if (user_zonelist_order
== ZONELIST_ORDER_DEFAULT
)
2585 current_zonelist_order
= default_zonelist_order();
2587 current_zonelist_order
= user_zonelist_order
;
2590 static void build_zonelists(pg_data_t
*pgdat
)
2594 nodemask_t used_mask
;
2595 int local_node
, prev_node
;
2596 struct zonelist
*zonelist
;
2597 int order
= current_zonelist_order
;
2599 /* initialize zonelists */
2600 for (i
= 0; i
< MAX_ZONELISTS
; i
++) {
2601 zonelist
= pgdat
->node_zonelists
+ i
;
2602 zonelist
->_zonerefs
[0].zone
= NULL
;
2603 zonelist
->_zonerefs
[0].zone_idx
= 0;
2606 /* NUMA-aware ordering of nodes */
2607 local_node
= pgdat
->node_id
;
2608 load
= nr_online_nodes
;
2609 prev_node
= local_node
;
2610 nodes_clear(used_mask
);
2612 memset(node_order
, 0, sizeof(node_order
));
2615 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
2616 int distance
= node_distance(local_node
, node
);
2619 * If another node is sufficiently far away then it is better
2620 * to reclaim pages in a zone before going off node.
2622 if (distance
> RECLAIM_DISTANCE
)
2623 zone_reclaim_mode
= 1;
2626 * We don't want to pressure a particular node.
2627 * So adding penalty to the first node in same
2628 * distance group to make it round-robin.
2630 if (distance
!= node_distance(local_node
, prev_node
))
2631 node_load
[node
] = load
;
2635 if (order
== ZONELIST_ORDER_NODE
)
2636 build_zonelists_in_node_order(pgdat
, node
);
2638 node_order
[j
++] = node
; /* remember order */
2641 if (order
== ZONELIST_ORDER_ZONE
) {
2642 /* calculate node order -- i.e., DMA last! */
2643 build_zonelists_in_zone_order(pgdat
, j
);
2646 build_thisnode_zonelists(pgdat
);
2649 /* Construct the zonelist performance cache - see further mmzone.h */
2650 static void build_zonelist_cache(pg_data_t
*pgdat
)
2652 struct zonelist
*zonelist
;
2653 struct zonelist_cache
*zlc
;
2656 zonelist
= &pgdat
->node_zonelists
[0];
2657 zonelist
->zlcache_ptr
= zlc
= &zonelist
->zlcache
;
2658 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
2659 for (z
= zonelist
->_zonerefs
; z
->zone
; z
++)
2660 zlc
->z_to_n
[z
- zonelist
->_zonerefs
] = zonelist_node_idx(z
);
2664 #else /* CONFIG_NUMA */
2666 static void set_zonelist_order(void)
2668 current_zonelist_order
= ZONELIST_ORDER_ZONE
;
2671 static void build_zonelists(pg_data_t
*pgdat
)
2673 int node
, local_node
;
2675 struct zonelist
*zonelist
;
2677 local_node
= pgdat
->node_id
;
2679 zonelist
= &pgdat
->node_zonelists
[0];
2680 j
= build_zonelists_node(pgdat
, zonelist
, 0, MAX_NR_ZONES
- 1);
2683 * Now we build the zonelist so that it contains the zones
2684 * of all the other nodes.
2685 * We don't want to pressure a particular node, so when
2686 * building the zones for node N, we make sure that the
2687 * zones coming right after the local ones are those from
2688 * node N+1 (modulo N)
2690 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
2691 if (!node_online(node
))
2693 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
2696 for (node
= 0; node
< local_node
; node
++) {
2697 if (!node_online(node
))
2699 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
2703 zonelist
->_zonerefs
[j
].zone
= NULL
;
2704 zonelist
->_zonerefs
[j
].zone_idx
= 0;
2707 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
2708 static void build_zonelist_cache(pg_data_t
*pgdat
)
2710 pgdat
->node_zonelists
[0].zlcache_ptr
= NULL
;
2713 #endif /* CONFIG_NUMA */
2715 /* return values int ....just for stop_machine() */
2716 static int __build_all_zonelists(void *dummy
)
2721 memset(node_load
, 0, sizeof(node_load
));
2723 for_each_online_node(nid
) {
2724 pg_data_t
*pgdat
= NODE_DATA(nid
);
2726 build_zonelists(pgdat
);
2727 build_zonelist_cache(pgdat
);
2732 void build_all_zonelists(void)
2734 set_zonelist_order();
2736 if (system_state
== SYSTEM_BOOTING
) {
2737 __build_all_zonelists(NULL
);
2738 mminit_verify_zonelist();
2739 cpuset_init_current_mems_allowed();
2741 /* we have to stop all cpus to guarantee there is no user
2743 stop_machine(__build_all_zonelists
, NULL
, NULL
);
2744 /* cpuset refresh routine should be here */
2746 vm_total_pages
= nr_free_pagecache_pages();
2748 * Disable grouping by mobility if the number of pages in the
2749 * system is too low to allow the mechanism to work. It would be
2750 * more accurate, but expensive to check per-zone. This check is
2751 * made on memory-hotadd so a system can start with mobility
2752 * disabled and enable it later
2754 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
2755 page_group_by_mobility_disabled
= 1;
2757 page_group_by_mobility_disabled
= 0;
2759 printk("Built %i zonelists in %s order, mobility grouping %s. "
2760 "Total pages: %ld\n",
2762 zonelist_order_name
[current_zonelist_order
],
2763 page_group_by_mobility_disabled
? "off" : "on",
2766 printk("Policy zone: %s\n", zone_names
[policy_zone
]);
2771 * Helper functions to size the waitqueue hash table.
2772 * Essentially these want to choose hash table sizes sufficiently
2773 * large so that collisions trying to wait on pages are rare.
2774 * But in fact, the number of active page waitqueues on typical
2775 * systems is ridiculously low, less than 200. So this is even
2776 * conservative, even though it seems large.
2778 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
2779 * waitqueues, i.e. the size of the waitq table given the number of pages.
2781 #define PAGES_PER_WAITQUEUE 256
2783 #ifndef CONFIG_MEMORY_HOTPLUG
2784 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
2786 unsigned long size
= 1;
2788 pages
/= PAGES_PER_WAITQUEUE
;
2790 while (size
< pages
)
2794 * Once we have dozens or even hundreds of threads sleeping
2795 * on IO we've got bigger problems than wait queue collision.
2796 * Limit the size of the wait table to a reasonable size.
2798 size
= min(size
, 4096UL);
2800 return max(size
, 4UL);
2804 * A zone's size might be changed by hot-add, so it is not possible to determine
2805 * a suitable size for its wait_table. So we use the maximum size now.
2807 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
2809 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
2810 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
2811 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
2813 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
2814 * or more by the traditional way. (See above). It equals:
2816 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
2817 * ia64(16K page size) : = ( 8G + 4M)byte.
2818 * powerpc (64K page size) : = (32G +16M)byte.
2820 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
2827 * This is an integer logarithm so that shifts can be used later
2828 * to extract the more random high bits from the multiplicative
2829 * hash function before the remainder is taken.
2831 static inline unsigned long wait_table_bits(unsigned long size
)
2836 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
2839 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
2840 * of blocks reserved is based on min_wmark_pages(zone). The memory within
2841 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
2842 * higher will lead to a bigger reserve which will get freed as contiguous
2843 * blocks as reclaim kicks in
2845 static void setup_zone_migrate_reserve(struct zone
*zone
)
2847 unsigned long start_pfn
, pfn
, end_pfn
;
2849 unsigned long block_migratetype
;
2852 /* Get the start pfn, end pfn and the number of blocks to reserve */
2853 start_pfn
= zone
->zone_start_pfn
;
2854 end_pfn
= start_pfn
+ zone
->spanned_pages
;
2855 reserve
= roundup(min_wmark_pages(zone
), pageblock_nr_pages
) >>
2859 * Reserve blocks are generally in place to help high-order atomic
2860 * allocations that are short-lived. A min_free_kbytes value that
2861 * would result in more than 2 reserve blocks for atomic allocations
2862 * is assumed to be in place to help anti-fragmentation for the
2863 * future allocation of hugepages at runtime.
2865 reserve
= min(2, reserve
);
2867 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= pageblock_nr_pages
) {
2868 if (!pfn_valid(pfn
))
2870 page
= pfn_to_page(pfn
);
2872 /* Watch out for overlapping nodes */
2873 if (page_to_nid(page
) != zone_to_nid(zone
))
2876 /* Blocks with reserved pages will never free, skip them. */
2877 if (PageReserved(page
))
2880 block_migratetype
= get_pageblock_migratetype(page
);
2882 /* If this block is reserved, account for it */
2883 if (reserve
> 0 && block_migratetype
== MIGRATE_RESERVE
) {
2888 /* Suitable for reserving if this block is movable */
2889 if (reserve
> 0 && block_migratetype
== MIGRATE_MOVABLE
) {
2890 set_pageblock_migratetype(page
, MIGRATE_RESERVE
);
2891 move_freepages_block(zone
, page
, MIGRATE_RESERVE
);
2897 * If the reserve is met and this is a previous reserved block,
2900 if (block_migratetype
== MIGRATE_RESERVE
) {
2901 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
2902 move_freepages_block(zone
, page
, MIGRATE_MOVABLE
);
2908 * Initially all pages are reserved - free ones are freed
2909 * up by free_all_bootmem() once the early boot process is
2910 * done. Non-atomic initialization, single-pass.
2912 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
2913 unsigned long start_pfn
, enum memmap_context context
)
2916 unsigned long end_pfn
= start_pfn
+ size
;
2920 if (highest_memmap_pfn
< end_pfn
- 1)
2921 highest_memmap_pfn
= end_pfn
- 1;
2923 z
= &NODE_DATA(nid
)->node_zones
[zone
];
2924 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
2926 * There can be holes in boot-time mem_map[]s
2927 * handed to this function. They do not
2928 * exist on hotplugged memory.
2930 if (context
== MEMMAP_EARLY
) {
2931 if (!early_pfn_valid(pfn
))
2933 if (!early_pfn_in_nid(pfn
, nid
))
2936 page
= pfn_to_page(pfn
);
2937 set_page_links(page
, zone
, nid
, pfn
);
2938 mminit_verify_page_links(page
, zone
, nid
, pfn
);
2939 init_page_count(page
);
2940 reset_page_mapcount(page
);
2941 SetPageReserved(page
);
2943 * Mark the block movable so that blocks are reserved for
2944 * movable at startup. This will force kernel allocations
2945 * to reserve their blocks rather than leaking throughout
2946 * the address space during boot when many long-lived
2947 * kernel allocations are made. Later some blocks near
2948 * the start are marked MIGRATE_RESERVE by
2949 * setup_zone_migrate_reserve()
2951 * bitmap is created for zone's valid pfn range. but memmap
2952 * can be created for invalid pages (for alignment)
2953 * check here not to call set_pageblock_migratetype() against
2956 if ((z
->zone_start_pfn
<= pfn
)
2957 && (pfn
< z
->zone_start_pfn
+ z
->spanned_pages
)
2958 && !(pfn
& (pageblock_nr_pages
- 1)))
2959 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
2961 INIT_LIST_HEAD(&page
->lru
);
2962 #ifdef WANT_PAGE_VIRTUAL
2963 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
2964 if (!is_highmem_idx(zone
))
2965 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
2970 static void __meminit
zone_init_free_lists(struct zone
*zone
)
2973 for_each_migratetype_order(order
, t
) {
2974 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
2975 zone
->free_area
[order
].nr_free
= 0;
2979 #ifndef __HAVE_ARCH_MEMMAP_INIT
2980 #define memmap_init(size, nid, zone, start_pfn) \
2981 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
2984 static int zone_batchsize(struct zone
*zone
)
2990 * The per-cpu-pages pools are set to around 1000th of the
2991 * size of the zone. But no more than 1/2 of a meg.
2993 * OK, so we don't know how big the cache is. So guess.
2995 batch
= zone
->present_pages
/ 1024;
2996 if (batch
* PAGE_SIZE
> 512 * 1024)
2997 batch
= (512 * 1024) / PAGE_SIZE
;
2998 batch
/= 4; /* We effectively *= 4 below */
3003 * Clamp the batch to a 2^n - 1 value. Having a power
3004 * of 2 value was found to be more likely to have
3005 * suboptimal cache aliasing properties in some cases.
3007 * For example if 2 tasks are alternately allocating
3008 * batches of pages, one task can end up with a lot
3009 * of pages of one half of the possible page colors
3010 * and the other with pages of the other colors.
3012 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
3017 /* The deferral and batching of frees should be suppressed under NOMMU
3020 * The problem is that NOMMU needs to be able to allocate large chunks
3021 * of contiguous memory as there's no hardware page translation to
3022 * assemble apparent contiguous memory from discontiguous pages.
3024 * Queueing large contiguous runs of pages for batching, however,
3025 * causes the pages to actually be freed in smaller chunks. As there
3026 * can be a significant delay between the individual batches being
3027 * recycled, this leads to the once large chunks of space being
3028 * fragmented and becoming unavailable for high-order allocations.
3034 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
3036 struct per_cpu_pages
*pcp
;
3039 memset(p
, 0, sizeof(*p
));
3043 pcp
->high
= 6 * batch
;
3044 pcp
->batch
= max(1UL, 1 * batch
);
3045 for (migratetype
= 0; migratetype
< MIGRATE_PCPTYPES
; migratetype
++)
3046 INIT_LIST_HEAD(&pcp
->lists
[migratetype
]);
3050 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3051 * to the value high for the pageset p.
3054 static void setup_pagelist_highmark(struct per_cpu_pageset
*p
,
3057 struct per_cpu_pages
*pcp
;
3061 pcp
->batch
= max(1UL, high
/4);
3062 if ((high
/4) > (PAGE_SHIFT
* 8))
3063 pcp
->batch
= PAGE_SHIFT
* 8;
3069 * Boot pageset table. One per cpu which is going to be used for all
3070 * zones and all nodes. The parameters will be set in such a way
3071 * that an item put on a list will immediately be handed over to
3072 * the buddy list. This is safe since pageset manipulation is done
3073 * with interrupts disabled.
3075 * Some NUMA counter updates may also be caught by the boot pagesets.
3077 * The boot_pagesets must be kept even after bootup is complete for
3078 * unused processors and/or zones. They do play a role for bootstrapping
3079 * hotplugged processors.
3081 * zoneinfo_show() and maybe other functions do
3082 * not check if the processor is online before following the pageset pointer.
3083 * Other parts of the kernel may not check if the zone is available.
3085 static struct per_cpu_pageset boot_pageset
[NR_CPUS
];
3088 * Dynamically allocate memory for the
3089 * per cpu pageset array in struct zone.
3091 static int __cpuinit
process_zones(int cpu
)
3093 struct zone
*zone
, *dzone
;
3094 int node
= cpu_to_node(cpu
);
3096 node_set_state(node
, N_CPU
); /* this node has a cpu */
3098 for_each_populated_zone(zone
) {
3099 zone_pcp(zone
, cpu
) = kmalloc_node(sizeof(struct per_cpu_pageset
),
3101 if (!zone_pcp(zone
, cpu
))
3104 setup_pageset(zone_pcp(zone
, cpu
), zone_batchsize(zone
));
3106 if (percpu_pagelist_fraction
)
3107 setup_pagelist_highmark(zone_pcp(zone
, cpu
),
3108 (zone
->present_pages
/ percpu_pagelist_fraction
));
3113 for_each_zone(dzone
) {
3114 if (!populated_zone(dzone
))
3118 kfree(zone_pcp(dzone
, cpu
));
3119 zone_pcp(dzone
, cpu
) = &boot_pageset
[cpu
];
3124 static inline void free_zone_pagesets(int cpu
)
3128 for_each_zone(zone
) {
3129 struct per_cpu_pageset
*pset
= zone_pcp(zone
, cpu
);
3131 /* Free per_cpu_pageset if it is slab allocated */
3132 if (pset
!= &boot_pageset
[cpu
])
3134 zone_pcp(zone
, cpu
) = &boot_pageset
[cpu
];
3138 static int __cpuinit
pageset_cpuup_callback(struct notifier_block
*nfb
,
3139 unsigned long action
,
3142 int cpu
= (long)hcpu
;
3143 int ret
= NOTIFY_OK
;
3146 case CPU_UP_PREPARE
:
3147 case CPU_UP_PREPARE_FROZEN
:
3148 if (process_zones(cpu
))
3151 case CPU_UP_CANCELED
:
3152 case CPU_UP_CANCELED_FROZEN
:
3154 case CPU_DEAD_FROZEN
:
3155 free_zone_pagesets(cpu
);
3163 static struct notifier_block __cpuinitdata pageset_notifier
=
3164 { &pageset_cpuup_callback
, NULL
, 0 };
3166 void __init
setup_per_cpu_pageset(void)
3170 /* Initialize per_cpu_pageset for cpu 0.
3171 * A cpuup callback will do this for every cpu
3172 * as it comes online
3174 err
= process_zones(smp_processor_id());
3176 register_cpu_notifier(&pageset_notifier
);
3181 static noinline __init_refok
3182 int zone_wait_table_init(struct zone
*zone
, unsigned long zone_size_pages
)
3185 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
3189 * The per-page waitqueue mechanism uses hashed waitqueues
3192 zone
->wait_table_hash_nr_entries
=
3193 wait_table_hash_nr_entries(zone_size_pages
);
3194 zone
->wait_table_bits
=
3195 wait_table_bits(zone
->wait_table_hash_nr_entries
);
3196 alloc_size
= zone
->wait_table_hash_nr_entries
3197 * sizeof(wait_queue_head_t
);
3199 if (!slab_is_available()) {
3200 zone
->wait_table
= (wait_queue_head_t
*)
3201 alloc_bootmem_node(pgdat
, alloc_size
);
3204 * This case means that a zone whose size was 0 gets new memory
3205 * via memory hot-add.
3206 * But it may be the case that a new node was hot-added. In
3207 * this case vmalloc() will not be able to use this new node's
3208 * memory - this wait_table must be initialized to use this new
3209 * node itself as well.
3210 * To use this new node's memory, further consideration will be
3213 zone
->wait_table
= vmalloc(alloc_size
);
3215 if (!zone
->wait_table
)
3218 for(i
= 0; i
< zone
->wait_table_hash_nr_entries
; ++i
)
3219 init_waitqueue_head(zone
->wait_table
+ i
);
3224 static int __zone_pcp_update(void *data
)
3226 struct zone
*zone
= data
;
3228 unsigned long batch
= zone_batchsize(zone
), flags
;
3230 for (cpu
= 0; cpu
< NR_CPUS
; cpu
++) {
3231 struct per_cpu_pageset
*pset
;
3232 struct per_cpu_pages
*pcp
;
3234 pset
= zone_pcp(zone
, cpu
);
3237 local_irq_save(flags
);
3238 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
3239 setup_pageset(pset
, batch
);
3240 local_irq_restore(flags
);
3245 void zone_pcp_update(struct zone
*zone
)
3247 stop_machine(__zone_pcp_update
, zone
, NULL
);
3250 static __meminit
void zone_pcp_init(struct zone
*zone
)
3253 unsigned long batch
= zone_batchsize(zone
);
3255 for (cpu
= 0; cpu
< NR_CPUS
; cpu
++) {
3257 /* Early boot. Slab allocator not functional yet */
3258 zone_pcp(zone
, cpu
) = &boot_pageset
[cpu
];
3259 setup_pageset(&boot_pageset
[cpu
],0);
3261 setup_pageset(zone_pcp(zone
,cpu
), batch
);
3264 if (zone
->present_pages
)
3265 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%lu\n",
3266 zone
->name
, zone
->present_pages
, batch
);
3269 __meminit
int init_currently_empty_zone(struct zone
*zone
,
3270 unsigned long zone_start_pfn
,
3272 enum memmap_context context
)
3274 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
3276 ret
= zone_wait_table_init(zone
, size
);
3279 pgdat
->nr_zones
= zone_idx(zone
) + 1;
3281 zone
->zone_start_pfn
= zone_start_pfn
;
3283 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
3284 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
3286 (unsigned long)zone_idx(zone
),
3287 zone_start_pfn
, (zone_start_pfn
+ size
));
3289 zone_init_free_lists(zone
);
3294 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3296 * Basic iterator support. Return the first range of PFNs for a node
3297 * Note: nid == MAX_NUMNODES returns first region regardless of node
3299 static int __meminit
first_active_region_index_in_nid(int nid
)
3303 for (i
= 0; i
< nr_nodemap_entries
; i
++)
3304 if (nid
== MAX_NUMNODES
|| early_node_map
[i
].nid
== nid
)
3311 * Basic iterator support. Return the next active range of PFNs for a node
3312 * Note: nid == MAX_NUMNODES returns next region regardless of node
3314 static int __meminit
next_active_region_index_in_nid(int index
, int nid
)
3316 for (index
= index
+ 1; index
< nr_nodemap_entries
; index
++)
3317 if (nid
== MAX_NUMNODES
|| early_node_map
[index
].nid
== nid
)
3323 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
3325 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
3326 * Architectures may implement their own version but if add_active_range()
3327 * was used and there are no special requirements, this is a convenient
3330 int __meminit
__early_pfn_to_nid(unsigned long pfn
)
3334 for (i
= 0; i
< nr_nodemap_entries
; i
++) {
3335 unsigned long start_pfn
= early_node_map
[i
].start_pfn
;
3336 unsigned long end_pfn
= early_node_map
[i
].end_pfn
;
3338 if (start_pfn
<= pfn
&& pfn
< end_pfn
)
3339 return early_node_map
[i
].nid
;
3341 /* This is a memory hole */
3344 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
3346 int __meminit
early_pfn_to_nid(unsigned long pfn
)
3350 nid
= __early_pfn_to_nid(pfn
);
3353 /* just returns 0 */
3357 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
3358 bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
3362 nid
= __early_pfn_to_nid(pfn
);
3363 if (nid
>= 0 && nid
!= node
)
3369 /* Basic iterator support to walk early_node_map[] */
3370 #define for_each_active_range_index_in_nid(i, nid) \
3371 for (i = first_active_region_index_in_nid(nid); i != -1; \
3372 i = next_active_region_index_in_nid(i, nid))
3375 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
3376 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
3377 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
3379 * If an architecture guarantees that all ranges registered with
3380 * add_active_ranges() contain no holes and may be freed, this
3381 * this function may be used instead of calling free_bootmem() manually.
3383 void __init
free_bootmem_with_active_regions(int nid
,
3384 unsigned long max_low_pfn
)
3388 for_each_active_range_index_in_nid(i
, nid
) {
3389 unsigned long size_pages
= 0;
3390 unsigned long end_pfn
= early_node_map
[i
].end_pfn
;
3392 if (early_node_map
[i
].start_pfn
>= max_low_pfn
)
3395 if (end_pfn
> max_low_pfn
)
3396 end_pfn
= max_low_pfn
;
3398 size_pages
= end_pfn
- early_node_map
[i
].start_pfn
;
3399 free_bootmem_node(NODE_DATA(early_node_map
[i
].nid
),
3400 PFN_PHYS(early_node_map
[i
].start_pfn
),
3401 size_pages
<< PAGE_SHIFT
);
3405 void __init
work_with_active_regions(int nid
, work_fn_t work_fn
, void *data
)
3410 for_each_active_range_index_in_nid(i
, nid
) {
3411 ret
= work_fn(early_node_map
[i
].start_pfn
,
3412 early_node_map
[i
].end_pfn
, data
);
3418 * sparse_memory_present_with_active_regions - Call memory_present for each active range
3419 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
3421 * If an architecture guarantees that all ranges registered with
3422 * add_active_ranges() contain no holes and may be freed, this
3423 * function may be used instead of calling memory_present() manually.
3425 void __init
sparse_memory_present_with_active_regions(int nid
)
3429 for_each_active_range_index_in_nid(i
, nid
)
3430 memory_present(early_node_map
[i
].nid
,
3431 early_node_map
[i
].start_pfn
,
3432 early_node_map
[i
].end_pfn
);
3436 * get_pfn_range_for_nid - Return the start and end page frames for a node
3437 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
3438 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
3439 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
3441 * It returns the start and end page frame of a node based on information
3442 * provided by an arch calling add_active_range(). If called for a node
3443 * with no available memory, a warning is printed and the start and end
3446 void __meminit
get_pfn_range_for_nid(unsigned int nid
,
3447 unsigned long *start_pfn
, unsigned long *end_pfn
)
3453 for_each_active_range_index_in_nid(i
, nid
) {
3454 *start_pfn
= min(*start_pfn
, early_node_map
[i
].start_pfn
);
3455 *end_pfn
= max(*end_pfn
, early_node_map
[i
].end_pfn
);
3458 if (*start_pfn
== -1UL)
3463 * This finds a zone that can be used for ZONE_MOVABLE pages. The
3464 * assumption is made that zones within a node are ordered in monotonic
3465 * increasing memory addresses so that the "highest" populated zone is used
3467 static void __init
find_usable_zone_for_movable(void)
3470 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
3471 if (zone_index
== ZONE_MOVABLE
)
3474 if (arch_zone_highest_possible_pfn
[zone_index
] >
3475 arch_zone_lowest_possible_pfn
[zone_index
])
3479 VM_BUG_ON(zone_index
== -1);
3480 movable_zone
= zone_index
;
3484 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
3485 * because it is sized independant of architecture. Unlike the other zones,
3486 * the starting point for ZONE_MOVABLE is not fixed. It may be different
3487 * in each node depending on the size of each node and how evenly kernelcore
3488 * is distributed. This helper function adjusts the zone ranges
3489 * provided by the architecture for a given node by using the end of the
3490 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
3491 * zones within a node are in order of monotonic increases memory addresses
3493 static void __meminit
adjust_zone_range_for_zone_movable(int nid
,
3494 unsigned long zone_type
,
3495 unsigned long node_start_pfn
,
3496 unsigned long node_end_pfn
,
3497 unsigned long *zone_start_pfn
,
3498 unsigned long *zone_end_pfn
)
3500 /* Only adjust if ZONE_MOVABLE is on this node */
3501 if (zone_movable_pfn
[nid
]) {
3502 /* Size ZONE_MOVABLE */
3503 if (zone_type
== ZONE_MOVABLE
) {
3504 *zone_start_pfn
= zone_movable_pfn
[nid
];
3505 *zone_end_pfn
= min(node_end_pfn
,
3506 arch_zone_highest_possible_pfn
[movable_zone
]);
3508 /* Adjust for ZONE_MOVABLE starting within this range */
3509 } else if (*zone_start_pfn
< zone_movable_pfn
[nid
] &&
3510 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
3511 *zone_end_pfn
= zone_movable_pfn
[nid
];
3513 /* Check if this whole range is within ZONE_MOVABLE */
3514 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
3515 *zone_start_pfn
= *zone_end_pfn
;
3520 * Return the number of pages a zone spans in a node, including holes
3521 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
3523 static unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
3524 unsigned long zone_type
,
3525 unsigned long *ignored
)
3527 unsigned long node_start_pfn
, node_end_pfn
;
3528 unsigned long zone_start_pfn
, zone_end_pfn
;
3530 /* Get the start and end of the node and zone */
3531 get_pfn_range_for_nid(nid
, &node_start_pfn
, &node_end_pfn
);
3532 zone_start_pfn
= arch_zone_lowest_possible_pfn
[zone_type
];
3533 zone_end_pfn
= arch_zone_highest_possible_pfn
[zone_type
];
3534 adjust_zone_range_for_zone_movable(nid
, zone_type
,
3535 node_start_pfn
, node_end_pfn
,
3536 &zone_start_pfn
, &zone_end_pfn
);
3538 /* Check that this node has pages within the zone's required range */
3539 if (zone_end_pfn
< node_start_pfn
|| zone_start_pfn
> node_end_pfn
)
3542 /* Move the zone boundaries inside the node if necessary */
3543 zone_end_pfn
= min(zone_end_pfn
, node_end_pfn
);
3544 zone_start_pfn
= max(zone_start_pfn
, node_start_pfn
);
3546 /* Return the spanned pages */
3547 return zone_end_pfn
- zone_start_pfn
;
3551 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
3552 * then all holes in the requested range will be accounted for.
3554 static unsigned long __meminit
__absent_pages_in_range(int nid
,
3555 unsigned long range_start_pfn
,
3556 unsigned long range_end_pfn
)
3559 unsigned long prev_end_pfn
= 0, hole_pages
= 0;
3560 unsigned long start_pfn
;
3562 /* Find the end_pfn of the first active range of pfns in the node */
3563 i
= first_active_region_index_in_nid(nid
);
3567 prev_end_pfn
= min(early_node_map
[i
].start_pfn
, range_end_pfn
);
3569 /* Account for ranges before physical memory on this node */
3570 if (early_node_map
[i
].start_pfn
> range_start_pfn
)
3571 hole_pages
= prev_end_pfn
- range_start_pfn
;
3573 /* Find all holes for the zone within the node */
3574 for (; i
!= -1; i
= next_active_region_index_in_nid(i
, nid
)) {
3576 /* No need to continue if prev_end_pfn is outside the zone */
3577 if (prev_end_pfn
>= range_end_pfn
)
3580 /* Make sure the end of the zone is not within the hole */
3581 start_pfn
= min(early_node_map
[i
].start_pfn
, range_end_pfn
);
3582 prev_end_pfn
= max(prev_end_pfn
, range_start_pfn
);
3584 /* Update the hole size cound and move on */
3585 if (start_pfn
> range_start_pfn
) {
3586 BUG_ON(prev_end_pfn
> start_pfn
);
3587 hole_pages
+= start_pfn
- prev_end_pfn
;
3589 prev_end_pfn
= early_node_map
[i
].end_pfn
;
3592 /* Account for ranges past physical memory on this node */
3593 if (range_end_pfn
> prev_end_pfn
)
3594 hole_pages
+= range_end_pfn
-
3595 max(range_start_pfn
, prev_end_pfn
);
3601 * absent_pages_in_range - Return number of page frames in holes within a range
3602 * @start_pfn: The start PFN to start searching for holes
3603 * @end_pfn: The end PFN to stop searching for holes
3605 * It returns the number of pages frames in memory holes within a range.
3607 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
3608 unsigned long end_pfn
)
3610 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
3613 /* Return the number of page frames in holes in a zone on a node */
3614 static unsigned long __meminit
zone_absent_pages_in_node(int nid
,
3615 unsigned long zone_type
,
3616 unsigned long *ignored
)
3618 unsigned long node_start_pfn
, node_end_pfn
;
3619 unsigned long zone_start_pfn
, zone_end_pfn
;
3621 get_pfn_range_for_nid(nid
, &node_start_pfn
, &node_end_pfn
);
3622 zone_start_pfn
= max(arch_zone_lowest_possible_pfn
[zone_type
],
3624 zone_end_pfn
= min(arch_zone_highest_possible_pfn
[zone_type
],
3627 adjust_zone_range_for_zone_movable(nid
, zone_type
,
3628 node_start_pfn
, node_end_pfn
,
3629 &zone_start_pfn
, &zone_end_pfn
);
3630 return __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
3634 static inline unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
3635 unsigned long zone_type
,
3636 unsigned long *zones_size
)
3638 return zones_size
[zone_type
];
3641 static inline unsigned long __meminit
zone_absent_pages_in_node(int nid
,
3642 unsigned long zone_type
,
3643 unsigned long *zholes_size
)
3648 return zholes_size
[zone_type
];
3653 static void __meminit
calculate_node_totalpages(struct pglist_data
*pgdat
,
3654 unsigned long *zones_size
, unsigned long *zholes_size
)
3656 unsigned long realtotalpages
, totalpages
= 0;
3659 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
3660 totalpages
+= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
3662 pgdat
->node_spanned_pages
= totalpages
;
3664 realtotalpages
= totalpages
;
3665 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
3667 zone_absent_pages_in_node(pgdat
->node_id
, i
,
3669 pgdat
->node_present_pages
= realtotalpages
;
3670 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
3674 #ifndef CONFIG_SPARSEMEM
3676 * Calculate the size of the zone->blockflags rounded to an unsigned long
3677 * Start by making sure zonesize is a multiple of pageblock_order by rounding
3678 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
3679 * round what is now in bits to nearest long in bits, then return it in
3682 static unsigned long __init
usemap_size(unsigned long zonesize
)
3684 unsigned long usemapsize
;
3686 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
3687 usemapsize
= usemapsize
>> pageblock_order
;
3688 usemapsize
*= NR_PAGEBLOCK_BITS
;
3689 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
3691 return usemapsize
/ 8;
3694 static void __init
setup_usemap(struct pglist_data
*pgdat
,
3695 struct zone
*zone
, unsigned long zonesize
)
3697 unsigned long usemapsize
= usemap_size(zonesize
);
3698 zone
->pageblock_flags
= NULL
;
3700 zone
->pageblock_flags
= alloc_bootmem_node(pgdat
, usemapsize
);
3703 static void inline setup_usemap(struct pglist_data
*pgdat
,
3704 struct zone
*zone
, unsigned long zonesize
) {}
3705 #endif /* CONFIG_SPARSEMEM */
3707 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
3709 /* Return a sensible default order for the pageblock size. */
3710 static inline int pageblock_default_order(void)
3712 if (HPAGE_SHIFT
> PAGE_SHIFT
)
3713 return HUGETLB_PAGE_ORDER
;
3718 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
3719 static inline void __init
set_pageblock_order(unsigned int order
)
3721 /* Check that pageblock_nr_pages has not already been setup */
3722 if (pageblock_order
)
3726 * Assume the largest contiguous order of interest is a huge page.
3727 * This value may be variable depending on boot parameters on IA64
3729 pageblock_order
= order
;
3731 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
3734 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
3735 * and pageblock_default_order() are unused as pageblock_order is set
3736 * at compile-time. See include/linux/pageblock-flags.h for the values of
3737 * pageblock_order based on the kernel config
3739 static inline int pageblock_default_order(unsigned int order
)
3743 #define set_pageblock_order(x) do {} while (0)
3745 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
3748 * Set up the zone data structures:
3749 * - mark all pages reserved
3750 * - mark all memory queues empty
3751 * - clear the memory bitmaps
3753 static void __paginginit
free_area_init_core(struct pglist_data
*pgdat
,
3754 unsigned long *zones_size
, unsigned long *zholes_size
)
3757 int nid
= pgdat
->node_id
;
3758 unsigned long zone_start_pfn
= pgdat
->node_start_pfn
;
3761 pgdat_resize_init(pgdat
);
3762 pgdat
->nr_zones
= 0;
3763 init_waitqueue_head(&pgdat
->kswapd_wait
);
3764 pgdat
->kswapd_max_order
= 0;
3765 pgdat_page_cgroup_init(pgdat
);
3767 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
3768 struct zone
*zone
= pgdat
->node_zones
+ j
;
3769 unsigned long size
, realsize
, memmap_pages
;
3772 size
= zone_spanned_pages_in_node(nid
, j
, zones_size
);
3773 realsize
= size
- zone_absent_pages_in_node(nid
, j
,
3777 * Adjust realsize so that it accounts for how much memory
3778 * is used by this zone for memmap. This affects the watermark
3779 * and per-cpu initialisations
3782 PAGE_ALIGN(size
* sizeof(struct page
)) >> PAGE_SHIFT
;
3783 if (realsize
>= memmap_pages
) {
3784 realsize
-= memmap_pages
;
3787 " %s zone: %lu pages used for memmap\n",
3788 zone_names
[j
], memmap_pages
);
3791 " %s zone: %lu pages exceeds realsize %lu\n",
3792 zone_names
[j
], memmap_pages
, realsize
);
3794 /* Account for reserved pages */
3795 if (j
== 0 && realsize
> dma_reserve
) {
3796 realsize
-= dma_reserve
;
3797 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
3798 zone_names
[0], dma_reserve
);
3801 if (!is_highmem_idx(j
))
3802 nr_kernel_pages
+= realsize
;
3803 nr_all_pages
+= realsize
;
3805 zone
->spanned_pages
= size
;
3806 zone
->present_pages
= realsize
;
3809 zone
->min_unmapped_pages
= (realsize
*sysctl_min_unmapped_ratio
)
3811 zone
->min_slab_pages
= (realsize
* sysctl_min_slab_ratio
) / 100;
3813 zone
->name
= zone_names
[j
];
3814 spin_lock_init(&zone
->lock
);
3815 spin_lock_init(&zone
->lru_lock
);
3816 zone_seqlock_init(zone
);
3817 zone
->zone_pgdat
= pgdat
;
3819 zone
->prev_priority
= DEF_PRIORITY
;
3821 zone_pcp_init(zone
);
3823 INIT_LIST_HEAD(&zone
->lru
[l
].list
);
3824 zone
->reclaim_stat
.nr_saved_scan
[l
] = 0;
3826 zone
->reclaim_stat
.recent_rotated
[0] = 0;
3827 zone
->reclaim_stat
.recent_rotated
[1] = 0;
3828 zone
->reclaim_stat
.recent_scanned
[0] = 0;
3829 zone
->reclaim_stat
.recent_scanned
[1] = 0;
3830 zap_zone_vm_stats(zone
);
3835 set_pageblock_order(pageblock_default_order());
3836 setup_usemap(pgdat
, zone
, size
);
3837 ret
= init_currently_empty_zone(zone
, zone_start_pfn
,
3838 size
, MEMMAP_EARLY
);
3840 memmap_init(size
, nid
, j
, zone_start_pfn
);
3841 zone_start_pfn
+= size
;
3845 static void __init_refok
alloc_node_mem_map(struct pglist_data
*pgdat
)
3847 /* Skip empty nodes */
3848 if (!pgdat
->node_spanned_pages
)
3851 #ifdef CONFIG_FLAT_NODE_MEM_MAP
3852 /* ia64 gets its own node_mem_map, before this, without bootmem */
3853 if (!pgdat
->node_mem_map
) {
3854 unsigned long size
, start
, end
;
3858 * The zone's endpoints aren't required to be MAX_ORDER
3859 * aligned but the node_mem_map endpoints must be in order
3860 * for the buddy allocator to function correctly.
3862 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
3863 end
= pgdat
->node_start_pfn
+ pgdat
->node_spanned_pages
;
3864 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
3865 size
= (end
- start
) * sizeof(struct page
);
3866 map
= alloc_remap(pgdat
->node_id
, size
);
3868 map
= alloc_bootmem_node(pgdat
, size
);
3869 pgdat
->node_mem_map
= map
+ (pgdat
->node_start_pfn
- start
);
3871 #ifndef CONFIG_NEED_MULTIPLE_NODES
3873 * With no DISCONTIG, the global mem_map is just set as node 0's
3875 if (pgdat
== NODE_DATA(0)) {
3876 mem_map
= NODE_DATA(0)->node_mem_map
;
3877 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3878 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
3879 mem_map
-= (pgdat
->node_start_pfn
- ARCH_PFN_OFFSET
);
3880 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
3883 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
3886 void __paginginit
free_area_init_node(int nid
, unsigned long *zones_size
,
3887 unsigned long node_start_pfn
, unsigned long *zholes_size
)
3889 pg_data_t
*pgdat
= NODE_DATA(nid
);
3891 pgdat
->node_id
= nid
;
3892 pgdat
->node_start_pfn
= node_start_pfn
;
3893 calculate_node_totalpages(pgdat
, zones_size
, zholes_size
);
3895 alloc_node_mem_map(pgdat
);
3896 #ifdef CONFIG_FLAT_NODE_MEM_MAP
3897 printk(KERN_DEBUG
"free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
3898 nid
, (unsigned long)pgdat
,
3899 (unsigned long)pgdat
->node_mem_map
);
3902 free_area_init_core(pgdat
, zones_size
, zholes_size
);
3905 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3907 #if MAX_NUMNODES > 1
3909 * Figure out the number of possible node ids.
3911 static void __init
setup_nr_node_ids(void)
3914 unsigned int highest
= 0;
3916 for_each_node_mask(node
, node_possible_map
)
3918 nr_node_ids
= highest
+ 1;
3921 static inline void setup_nr_node_ids(void)
3927 * add_active_range - Register a range of PFNs backed by physical memory
3928 * @nid: The node ID the range resides on
3929 * @start_pfn: The start PFN of the available physical memory
3930 * @end_pfn: The end PFN of the available physical memory
3932 * These ranges are stored in an early_node_map[] and later used by
3933 * free_area_init_nodes() to calculate zone sizes and holes. If the
3934 * range spans a memory hole, it is up to the architecture to ensure
3935 * the memory is not freed by the bootmem allocator. If possible
3936 * the range being registered will be merged with existing ranges.
3938 void __init
add_active_range(unsigned int nid
, unsigned long start_pfn
,
3939 unsigned long end_pfn
)
3943 mminit_dprintk(MMINIT_TRACE
, "memory_register",
3944 "Entering add_active_range(%d, %#lx, %#lx) "
3945 "%d entries of %d used\n",
3946 nid
, start_pfn
, end_pfn
,
3947 nr_nodemap_entries
, MAX_ACTIVE_REGIONS
);
3949 mminit_validate_memmodel_limits(&start_pfn
, &end_pfn
);
3951 /* Merge with existing active regions if possible */
3952 for (i
= 0; i
< nr_nodemap_entries
; i
++) {
3953 if (early_node_map
[i
].nid
!= nid
)
3956 /* Skip if an existing region covers this new one */
3957 if (start_pfn
>= early_node_map
[i
].start_pfn
&&
3958 end_pfn
<= early_node_map
[i
].end_pfn
)
3961 /* Merge forward if suitable */
3962 if (start_pfn
<= early_node_map
[i
].end_pfn
&&
3963 end_pfn
> early_node_map
[i
].end_pfn
) {
3964 early_node_map
[i
].end_pfn
= end_pfn
;
3968 /* Merge backward if suitable */
3969 if (start_pfn
< early_node_map
[i
].end_pfn
&&
3970 end_pfn
>= early_node_map
[i
].start_pfn
) {
3971 early_node_map
[i
].start_pfn
= start_pfn
;
3976 /* Check that early_node_map is large enough */
3977 if (i
>= MAX_ACTIVE_REGIONS
) {
3978 printk(KERN_CRIT
"More than %d memory regions, truncating\n",
3979 MAX_ACTIVE_REGIONS
);
3983 early_node_map
[i
].nid
= nid
;
3984 early_node_map
[i
].start_pfn
= start_pfn
;
3985 early_node_map
[i
].end_pfn
= end_pfn
;
3986 nr_nodemap_entries
= i
+ 1;
3990 * remove_active_range - Shrink an existing registered range of PFNs
3991 * @nid: The node id the range is on that should be shrunk
3992 * @start_pfn: The new PFN of the range
3993 * @end_pfn: The new PFN of the range
3995 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
3996 * The map is kept near the end physical page range that has already been
3997 * registered. This function allows an arch to shrink an existing registered
4000 void __init
remove_active_range(unsigned int nid
, unsigned long start_pfn
,
4001 unsigned long end_pfn
)
4006 printk(KERN_DEBUG
"remove_active_range (%d, %lu, %lu)\n",
4007 nid
, start_pfn
, end_pfn
);
4009 /* Find the old active region end and shrink */
4010 for_each_active_range_index_in_nid(i
, nid
) {
4011 if (early_node_map
[i
].start_pfn
>= start_pfn
&&
4012 early_node_map
[i
].end_pfn
<= end_pfn
) {
4014 early_node_map
[i
].start_pfn
= 0;
4015 early_node_map
[i
].end_pfn
= 0;
4019 if (early_node_map
[i
].start_pfn
< start_pfn
&&
4020 early_node_map
[i
].end_pfn
> start_pfn
) {
4021 unsigned long temp_end_pfn
= early_node_map
[i
].end_pfn
;
4022 early_node_map
[i
].end_pfn
= start_pfn
;
4023 if (temp_end_pfn
> end_pfn
)
4024 add_active_range(nid
, end_pfn
, temp_end_pfn
);
4027 if (early_node_map
[i
].start_pfn
>= start_pfn
&&
4028 early_node_map
[i
].end_pfn
> end_pfn
&&
4029 early_node_map
[i
].start_pfn
< end_pfn
) {
4030 early_node_map
[i
].start_pfn
= end_pfn
;
4038 /* remove the blank ones */
4039 for (i
= nr_nodemap_entries
- 1; i
> 0; i
--) {
4040 if (early_node_map
[i
].nid
!= nid
)
4042 if (early_node_map
[i
].end_pfn
)
4044 /* we found it, get rid of it */
4045 for (j
= i
; j
< nr_nodemap_entries
- 1; j
++)
4046 memcpy(&early_node_map
[j
], &early_node_map
[j
+1],
4047 sizeof(early_node_map
[j
]));
4048 j
= nr_nodemap_entries
- 1;
4049 memset(&early_node_map
[j
], 0, sizeof(early_node_map
[j
]));
4050 nr_nodemap_entries
--;
4055 * remove_all_active_ranges - Remove all currently registered regions
4057 * During discovery, it may be found that a table like SRAT is invalid
4058 * and an alternative discovery method must be used. This function removes
4059 * all currently registered regions.
4061 void __init
remove_all_active_ranges(void)
4063 memset(early_node_map
, 0, sizeof(early_node_map
));
4064 nr_nodemap_entries
= 0;
4067 /* Compare two active node_active_regions */
4068 static int __init
cmp_node_active_region(const void *a
, const void *b
)
4070 struct node_active_region
*arange
= (struct node_active_region
*)a
;
4071 struct node_active_region
*brange
= (struct node_active_region
*)b
;
4073 /* Done this way to avoid overflows */
4074 if (arange
->start_pfn
> brange
->start_pfn
)
4076 if (arange
->start_pfn
< brange
->start_pfn
)
4082 /* sort the node_map by start_pfn */
4083 static void __init
sort_node_map(void)
4085 sort(early_node_map
, (size_t)nr_nodemap_entries
,
4086 sizeof(struct node_active_region
),
4087 cmp_node_active_region
, NULL
);
4090 /* Find the lowest pfn for a node */
4091 static unsigned long __init
find_min_pfn_for_node(int nid
)
4094 unsigned long min_pfn
= ULONG_MAX
;
4096 /* Assuming a sorted map, the first range found has the starting pfn */
4097 for_each_active_range_index_in_nid(i
, nid
)
4098 min_pfn
= min(min_pfn
, early_node_map
[i
].start_pfn
);
4100 if (min_pfn
== ULONG_MAX
) {
4102 "Could not find start_pfn for node %d\n", nid
);
4110 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4112 * It returns the minimum PFN based on information provided via
4113 * add_active_range().
4115 unsigned long __init
find_min_pfn_with_active_regions(void)
4117 return find_min_pfn_for_node(MAX_NUMNODES
);
4121 * early_calculate_totalpages()
4122 * Sum pages in active regions for movable zone.
4123 * Populate N_HIGH_MEMORY for calculating usable_nodes.
4125 static unsigned long __init
early_calculate_totalpages(void)
4128 unsigned long totalpages
= 0;
4130 for (i
= 0; i
< nr_nodemap_entries
; i
++) {
4131 unsigned long pages
= early_node_map
[i
].end_pfn
-
4132 early_node_map
[i
].start_pfn
;
4133 totalpages
+= pages
;
4135 node_set_state(early_node_map
[i
].nid
, N_HIGH_MEMORY
);
4141 * Find the PFN the Movable zone begins in each node. Kernel memory
4142 * is spread evenly between nodes as long as the nodes have enough
4143 * memory. When they don't, some nodes will have more kernelcore than
4146 static void __init
find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn
)
4149 unsigned long usable_startpfn
;
4150 unsigned long kernelcore_node
, kernelcore_remaining
;
4151 /* save the state before borrow the nodemask */
4152 nodemask_t saved_node_state
= node_states
[N_HIGH_MEMORY
];
4153 unsigned long totalpages
= early_calculate_totalpages();
4154 int usable_nodes
= nodes_weight(node_states
[N_HIGH_MEMORY
]);
4157 * If movablecore was specified, calculate what size of
4158 * kernelcore that corresponds so that memory usable for
4159 * any allocation type is evenly spread. If both kernelcore
4160 * and movablecore are specified, then the value of kernelcore
4161 * will be used for required_kernelcore if it's greater than
4162 * what movablecore would have allowed.
4164 if (required_movablecore
) {
4165 unsigned long corepages
;
4168 * Round-up so that ZONE_MOVABLE is at least as large as what
4169 * was requested by the user
4171 required_movablecore
=
4172 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
4173 corepages
= totalpages
- required_movablecore
;
4175 required_kernelcore
= max(required_kernelcore
, corepages
);
4178 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4179 if (!required_kernelcore
)
4182 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4183 find_usable_zone_for_movable();
4184 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
4187 /* Spread kernelcore memory as evenly as possible throughout nodes */
4188 kernelcore_node
= required_kernelcore
/ usable_nodes
;
4189 for_each_node_state(nid
, N_HIGH_MEMORY
) {
4191 * Recalculate kernelcore_node if the division per node
4192 * now exceeds what is necessary to satisfy the requested
4193 * amount of memory for the kernel
4195 if (required_kernelcore
< kernelcore_node
)
4196 kernelcore_node
= required_kernelcore
/ usable_nodes
;
4199 * As the map is walked, we track how much memory is usable
4200 * by the kernel using kernelcore_remaining. When it is
4201 * 0, the rest of the node is usable by ZONE_MOVABLE
4203 kernelcore_remaining
= kernelcore_node
;
4205 /* Go through each range of PFNs within this node */
4206 for_each_active_range_index_in_nid(i
, nid
) {
4207 unsigned long start_pfn
, end_pfn
;
4208 unsigned long size_pages
;
4210 start_pfn
= max(early_node_map
[i
].start_pfn
,
4211 zone_movable_pfn
[nid
]);
4212 end_pfn
= early_node_map
[i
].end_pfn
;
4213 if (start_pfn
>= end_pfn
)
4216 /* Account for what is only usable for kernelcore */
4217 if (start_pfn
< usable_startpfn
) {
4218 unsigned long kernel_pages
;
4219 kernel_pages
= min(end_pfn
, usable_startpfn
)
4222 kernelcore_remaining
-= min(kernel_pages
,
4223 kernelcore_remaining
);
4224 required_kernelcore
-= min(kernel_pages
,
4225 required_kernelcore
);
4227 /* Continue if range is now fully accounted */
4228 if (end_pfn
<= usable_startpfn
) {
4231 * Push zone_movable_pfn to the end so
4232 * that if we have to rebalance
4233 * kernelcore across nodes, we will
4234 * not double account here
4236 zone_movable_pfn
[nid
] = end_pfn
;
4239 start_pfn
= usable_startpfn
;
4243 * The usable PFN range for ZONE_MOVABLE is from
4244 * start_pfn->end_pfn. Calculate size_pages as the
4245 * number of pages used as kernelcore
4247 size_pages
= end_pfn
- start_pfn
;
4248 if (size_pages
> kernelcore_remaining
)
4249 size_pages
= kernelcore_remaining
;
4250 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
4253 * Some kernelcore has been met, update counts and
4254 * break if the kernelcore for this node has been
4257 required_kernelcore
-= min(required_kernelcore
,
4259 kernelcore_remaining
-= size_pages
;
4260 if (!kernelcore_remaining
)
4266 * If there is still required_kernelcore, we do another pass with one
4267 * less node in the count. This will push zone_movable_pfn[nid] further
4268 * along on the nodes that still have memory until kernelcore is
4272 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
4275 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4276 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
4277 zone_movable_pfn
[nid
] =
4278 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
4281 /* restore the node_state */
4282 node_states
[N_HIGH_MEMORY
] = saved_node_state
;
4285 /* Any regular memory on that node ? */
4286 static void check_for_regular_memory(pg_data_t
*pgdat
)
4288 #ifdef CONFIG_HIGHMEM
4289 enum zone_type zone_type
;
4291 for (zone_type
= 0; zone_type
<= ZONE_NORMAL
; zone_type
++) {
4292 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
4293 if (zone
->present_pages
)
4294 node_set_state(zone_to_nid(zone
), N_NORMAL_MEMORY
);
4300 * free_area_init_nodes - Initialise all pg_data_t and zone data
4301 * @max_zone_pfn: an array of max PFNs for each zone
4303 * This will call free_area_init_node() for each active node in the system.
4304 * Using the page ranges provided by add_active_range(), the size of each
4305 * zone in each node and their holes is calculated. If the maximum PFN
4306 * between two adjacent zones match, it is assumed that the zone is empty.
4307 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4308 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4309 * starts where the previous one ended. For example, ZONE_DMA32 starts
4310 * at arch_max_dma_pfn.
4312 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
4317 /* Sort early_node_map as initialisation assumes it is sorted */
4320 /* Record where the zone boundaries are */
4321 memset(arch_zone_lowest_possible_pfn
, 0,
4322 sizeof(arch_zone_lowest_possible_pfn
));
4323 memset(arch_zone_highest_possible_pfn
, 0,
4324 sizeof(arch_zone_highest_possible_pfn
));
4325 arch_zone_lowest_possible_pfn
[0] = find_min_pfn_with_active_regions();
4326 arch_zone_highest_possible_pfn
[0] = max_zone_pfn
[0];
4327 for (i
= 1; i
< MAX_NR_ZONES
; i
++) {
4328 if (i
== ZONE_MOVABLE
)
4330 arch_zone_lowest_possible_pfn
[i
] =
4331 arch_zone_highest_possible_pfn
[i
-1];
4332 arch_zone_highest_possible_pfn
[i
] =
4333 max(max_zone_pfn
[i
], arch_zone_lowest_possible_pfn
[i
]);
4335 arch_zone_lowest_possible_pfn
[ZONE_MOVABLE
] = 0;
4336 arch_zone_highest_possible_pfn
[ZONE_MOVABLE
] = 0;
4338 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4339 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
4340 find_zone_movable_pfns_for_nodes(zone_movable_pfn
);
4342 /* Print out the zone ranges */
4343 printk("Zone PFN ranges:\n");
4344 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
4345 if (i
== ZONE_MOVABLE
)
4347 printk(" %-8s %0#10lx -> %0#10lx\n",
4349 arch_zone_lowest_possible_pfn
[i
],
4350 arch_zone_highest_possible_pfn
[i
]);
4353 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4354 printk("Movable zone start PFN for each node\n");
4355 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
4356 if (zone_movable_pfn
[i
])
4357 printk(" Node %d: %lu\n", i
, zone_movable_pfn
[i
]);
4360 /* Print out the early_node_map[] */
4361 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries
);
4362 for (i
= 0; i
< nr_nodemap_entries
; i
++)
4363 printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map
[i
].nid
,
4364 early_node_map
[i
].start_pfn
,
4365 early_node_map
[i
].end_pfn
);
4367 /* Initialise every node */
4368 mminit_verify_pageflags_layout();
4369 setup_nr_node_ids();
4370 for_each_online_node(nid
) {
4371 pg_data_t
*pgdat
= NODE_DATA(nid
);
4372 free_area_init_node(nid
, NULL
,
4373 find_min_pfn_for_node(nid
), NULL
);
4375 /* Any memory on that node */
4376 if (pgdat
->node_present_pages
)
4377 node_set_state(nid
, N_HIGH_MEMORY
);
4378 check_for_regular_memory(pgdat
);
4382 static int __init
cmdline_parse_core(char *p
, unsigned long *core
)
4384 unsigned long long coremem
;
4388 coremem
= memparse(p
, &p
);
4389 *core
= coremem
>> PAGE_SHIFT
;
4391 /* Paranoid check that UL is enough for the coremem value */
4392 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
4398 * kernelcore=size sets the amount of memory for use for allocations that
4399 * cannot be reclaimed or migrated.
4401 static int __init
cmdline_parse_kernelcore(char *p
)
4403 return cmdline_parse_core(p
, &required_kernelcore
);
4407 * movablecore=size sets the amount of memory for use for allocations that
4408 * can be reclaimed or migrated.
4410 static int __init
cmdline_parse_movablecore(char *p
)
4412 return cmdline_parse_core(p
, &required_movablecore
);
4415 early_param("kernelcore", cmdline_parse_kernelcore
);
4416 early_param("movablecore", cmdline_parse_movablecore
);
4418 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4421 * set_dma_reserve - set the specified number of pages reserved in the first zone
4422 * @new_dma_reserve: The number of pages to mark reserved
4424 * The per-cpu batchsize and zone watermarks are determined by present_pages.
4425 * In the DMA zone, a significant percentage may be consumed by kernel image
4426 * and other unfreeable allocations which can skew the watermarks badly. This
4427 * function may optionally be used to account for unfreeable pages in the
4428 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
4429 * smaller per-cpu batchsize.
4431 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
4433 dma_reserve
= new_dma_reserve
;
4436 #ifndef CONFIG_NEED_MULTIPLE_NODES
4437 struct pglist_data __refdata contig_page_data
= { .bdata
= &bootmem_node_data
[0] };
4438 EXPORT_SYMBOL(contig_page_data
);
4441 void __init
free_area_init(unsigned long *zones_size
)
4443 free_area_init_node(0, zones_size
,
4444 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
4447 static int page_alloc_cpu_notify(struct notifier_block
*self
,
4448 unsigned long action
, void *hcpu
)
4450 int cpu
= (unsigned long)hcpu
;
4452 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
) {
4456 * Spill the event counters of the dead processor
4457 * into the current processors event counters.
4458 * This artificially elevates the count of the current
4461 vm_events_fold_cpu(cpu
);
4464 * Zero the differential counters of the dead processor
4465 * so that the vm statistics are consistent.
4467 * This is only okay since the processor is dead and cannot
4468 * race with what we are doing.
4470 refresh_cpu_vm_stats(cpu
);
4475 void __init
page_alloc_init(void)
4477 hotcpu_notifier(page_alloc_cpu_notify
, 0);
4481 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
4482 * or min_free_kbytes changes.
4484 static void calculate_totalreserve_pages(void)
4486 struct pglist_data
*pgdat
;
4487 unsigned long reserve_pages
= 0;
4488 enum zone_type i
, j
;
4490 for_each_online_pgdat(pgdat
) {
4491 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
4492 struct zone
*zone
= pgdat
->node_zones
+ i
;
4493 unsigned long max
= 0;
4495 /* Find valid and maximum lowmem_reserve in the zone */
4496 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
4497 if (zone
->lowmem_reserve
[j
] > max
)
4498 max
= zone
->lowmem_reserve
[j
];
4501 /* we treat the high watermark as reserved pages. */
4502 max
+= high_wmark_pages(zone
);
4504 if (max
> zone
->present_pages
)
4505 max
= zone
->present_pages
;
4506 reserve_pages
+= max
;
4509 totalreserve_pages
= reserve_pages
;
4513 * setup_per_zone_lowmem_reserve - called whenever
4514 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
4515 * has a correct pages reserved value, so an adequate number of
4516 * pages are left in the zone after a successful __alloc_pages().
4518 static void setup_per_zone_lowmem_reserve(void)
4520 struct pglist_data
*pgdat
;
4521 enum zone_type j
, idx
;
4523 for_each_online_pgdat(pgdat
) {
4524 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
4525 struct zone
*zone
= pgdat
->node_zones
+ j
;
4526 unsigned long present_pages
= zone
->present_pages
;
4528 zone
->lowmem_reserve
[j
] = 0;
4532 struct zone
*lower_zone
;
4536 if (sysctl_lowmem_reserve_ratio
[idx
] < 1)
4537 sysctl_lowmem_reserve_ratio
[idx
] = 1;
4539 lower_zone
= pgdat
->node_zones
+ idx
;
4540 lower_zone
->lowmem_reserve
[j
] = present_pages
/
4541 sysctl_lowmem_reserve_ratio
[idx
];
4542 present_pages
+= lower_zone
->present_pages
;
4547 /* update totalreserve_pages */
4548 calculate_totalreserve_pages();
4552 * setup_per_zone_wmarks - called when min_free_kbytes changes
4553 * or when memory is hot-{added|removed}
4555 * Ensures that the watermark[min,low,high] values for each zone are set
4556 * correctly with respect to min_free_kbytes.
4558 void setup_per_zone_wmarks(void)
4560 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
4561 unsigned long lowmem_pages
= 0;
4563 unsigned long flags
;
4565 /* Calculate total number of !ZONE_HIGHMEM pages */
4566 for_each_zone(zone
) {
4567 if (!is_highmem(zone
))
4568 lowmem_pages
+= zone
->present_pages
;
4571 for_each_zone(zone
) {
4574 spin_lock_irqsave(&zone
->lock
, flags
);
4575 tmp
= (u64
)pages_min
* zone
->present_pages
;
4576 do_div(tmp
, lowmem_pages
);
4577 if (is_highmem(zone
)) {
4579 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
4580 * need highmem pages, so cap pages_min to a small
4583 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
4584 * deltas controls asynch page reclaim, and so should
4585 * not be capped for highmem.
4589 min_pages
= zone
->present_pages
/ 1024;
4590 if (min_pages
< SWAP_CLUSTER_MAX
)
4591 min_pages
= SWAP_CLUSTER_MAX
;
4592 if (min_pages
> 128)
4594 zone
->watermark
[WMARK_MIN
] = min_pages
;
4597 * If it's a lowmem zone, reserve a number of pages
4598 * proportionate to the zone's size.
4600 zone
->watermark
[WMARK_MIN
] = tmp
;
4603 zone
->watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + (tmp
>> 2);
4604 zone
->watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + (tmp
>> 1);
4605 setup_zone_migrate_reserve(zone
);
4606 spin_unlock_irqrestore(&zone
->lock
, flags
);
4609 /* update totalreserve_pages */
4610 calculate_totalreserve_pages();
4614 * The inactive anon list should be small enough that the VM never has to
4615 * do too much work, but large enough that each inactive page has a chance
4616 * to be referenced again before it is swapped out.
4618 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
4619 * INACTIVE_ANON pages on this zone's LRU, maintained by the
4620 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
4621 * the anonymous pages are kept on the inactive list.
4624 * memory ratio inactive anon
4625 * -------------------------------------
4634 void calculate_zone_inactive_ratio(struct zone
*zone
)
4636 unsigned int gb
, ratio
;
4638 /* Zone size in gigabytes */
4639 gb
= zone
->present_pages
>> (30 - PAGE_SHIFT
);
4641 ratio
= int_sqrt(10 * gb
);
4645 zone
->inactive_ratio
= ratio
;
4648 static void __init
setup_per_zone_inactive_ratio(void)
4653 calculate_zone_inactive_ratio(zone
);
4657 * Initialise min_free_kbytes.
4659 * For small machines we want it small (128k min). For large machines
4660 * we want it large (64MB max). But it is not linear, because network
4661 * bandwidth does not increase linearly with machine size. We use
4663 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
4664 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
4680 static int __init
init_per_zone_wmark_min(void)
4682 unsigned long lowmem_kbytes
;
4684 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
4686 min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
4687 if (min_free_kbytes
< 128)
4688 min_free_kbytes
= 128;
4689 if (min_free_kbytes
> 65536)
4690 min_free_kbytes
= 65536;
4691 setup_per_zone_wmarks();
4692 setup_per_zone_lowmem_reserve();
4693 setup_per_zone_inactive_ratio();
4696 module_init(init_per_zone_wmark_min
)
4699 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
4700 * that we can call two helper functions whenever min_free_kbytes
4703 int min_free_kbytes_sysctl_handler(ctl_table
*table
, int write
,
4704 struct file
*file
, void __user
*buffer
, size_t *length
, loff_t
*ppos
)
4706 proc_dointvec(table
, write
, file
, buffer
, length
, ppos
);
4708 setup_per_zone_wmarks();
4713 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table
*table
, int write
,
4714 struct file
*file
, void __user
*buffer
, size_t *length
, loff_t
*ppos
)
4719 rc
= proc_dointvec_minmax(table
, write
, file
, buffer
, length
, ppos
);
4724 zone
->min_unmapped_pages
= (zone
->present_pages
*
4725 sysctl_min_unmapped_ratio
) / 100;
4729 int sysctl_min_slab_ratio_sysctl_handler(ctl_table
*table
, int write
,
4730 struct file
*file
, void __user
*buffer
, size_t *length
, loff_t
*ppos
)
4735 rc
= proc_dointvec_minmax(table
, write
, file
, buffer
, length
, ppos
);
4740 zone
->min_slab_pages
= (zone
->present_pages
*
4741 sysctl_min_slab_ratio
) / 100;
4747 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
4748 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
4749 * whenever sysctl_lowmem_reserve_ratio changes.
4751 * The reserve ratio obviously has absolutely no relation with the
4752 * minimum watermarks. The lowmem reserve ratio can only make sense
4753 * if in function of the boot time zone sizes.
4755 int lowmem_reserve_ratio_sysctl_handler(ctl_table
*table
, int write
,
4756 struct file
*file
, void __user
*buffer
, size_t *length
, loff_t
*ppos
)
4758 proc_dointvec_minmax(table
, write
, file
, buffer
, length
, ppos
);
4759 setup_per_zone_lowmem_reserve();
4764 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
4765 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
4766 * can have before it gets flushed back to buddy allocator.
4769 int percpu_pagelist_fraction_sysctl_handler(ctl_table
*table
, int write
,
4770 struct file
*file
, void __user
*buffer
, size_t *length
, loff_t
*ppos
)
4776 ret
= proc_dointvec_minmax(table
, write
, file
, buffer
, length
, ppos
);
4777 if (!write
|| (ret
== -EINVAL
))
4779 for_each_populated_zone(zone
) {
4780 for_each_online_cpu(cpu
) {
4782 high
= zone
->present_pages
/ percpu_pagelist_fraction
;
4783 setup_pagelist_highmark(zone_pcp(zone
, cpu
), high
);
4789 int hashdist
= HASHDIST_DEFAULT
;
4792 static int __init
set_hashdist(char *str
)
4796 hashdist
= simple_strtoul(str
, &str
, 0);
4799 __setup("hashdist=", set_hashdist
);
4803 * allocate a large system hash table from bootmem
4804 * - it is assumed that the hash table must contain an exact power-of-2
4805 * quantity of entries
4806 * - limit is the number of hash buckets, not the total allocation size
4808 void *__init
alloc_large_system_hash(const char *tablename
,
4809 unsigned long bucketsize
,
4810 unsigned long numentries
,
4813 unsigned int *_hash_shift
,
4814 unsigned int *_hash_mask
,
4815 unsigned long limit
)
4817 unsigned long long max
= limit
;
4818 unsigned long log2qty
, size
;
4821 /* allow the kernel cmdline to have a say */
4823 /* round applicable memory size up to nearest megabyte */
4824 numentries
= nr_kernel_pages
;
4825 numentries
+= (1UL << (20 - PAGE_SHIFT
)) - 1;
4826 numentries
>>= 20 - PAGE_SHIFT
;
4827 numentries
<<= 20 - PAGE_SHIFT
;
4829 /* limit to 1 bucket per 2^scale bytes of low memory */
4830 if (scale
> PAGE_SHIFT
)
4831 numentries
>>= (scale
- PAGE_SHIFT
);
4833 numentries
<<= (PAGE_SHIFT
- scale
);
4835 /* Make sure we've got at least a 0-order allocation.. */
4836 if (unlikely(flags
& HASH_SMALL
)) {
4837 /* Makes no sense without HASH_EARLY */
4838 WARN_ON(!(flags
& HASH_EARLY
));
4839 if (!(numentries
>> *_hash_shift
)) {
4840 numentries
= 1UL << *_hash_shift
;
4841 BUG_ON(!numentries
);
4843 } else if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
4844 numentries
= PAGE_SIZE
/ bucketsize
;
4846 numentries
= roundup_pow_of_two(numentries
);
4848 /* limit allocation size to 1/16 total memory by default */
4850 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
4851 do_div(max
, bucketsize
);
4854 if (numentries
> max
)
4857 log2qty
= ilog2(numentries
);
4860 size
= bucketsize
<< log2qty
;
4861 if (flags
& HASH_EARLY
)
4862 table
= alloc_bootmem_nopanic(size
);
4864 table
= __vmalloc(size
, GFP_ATOMIC
, PAGE_KERNEL
);
4867 * If bucketsize is not a power-of-two, we may free
4868 * some pages at the end of hash table which
4869 * alloc_pages_exact() automatically does
4871 if (get_order(size
) < MAX_ORDER
) {
4872 table
= alloc_pages_exact(size
, GFP_ATOMIC
);
4873 kmemleak_alloc(table
, size
, 1, GFP_ATOMIC
);
4876 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
4879 panic("Failed to allocate %s hash table\n", tablename
);
4881 printk(KERN_INFO
"%s hash table entries: %d (order: %d, %lu bytes)\n",
4884 ilog2(size
) - PAGE_SHIFT
,
4888 *_hash_shift
= log2qty
;
4890 *_hash_mask
= (1 << log2qty
) - 1;
4895 /* Return a pointer to the bitmap storing bits affecting a block of pages */
4896 static inline unsigned long *get_pageblock_bitmap(struct zone
*zone
,
4899 #ifdef CONFIG_SPARSEMEM
4900 return __pfn_to_section(pfn
)->pageblock_flags
;
4902 return zone
->pageblock_flags
;
4903 #endif /* CONFIG_SPARSEMEM */
4906 static inline int pfn_to_bitidx(struct zone
*zone
, unsigned long pfn
)
4908 #ifdef CONFIG_SPARSEMEM
4909 pfn
&= (PAGES_PER_SECTION
-1);
4910 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
4912 pfn
= pfn
- zone
->zone_start_pfn
;
4913 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
4914 #endif /* CONFIG_SPARSEMEM */
4918 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
4919 * @page: The page within the block of interest
4920 * @start_bitidx: The first bit of interest to retrieve
4921 * @end_bitidx: The last bit of interest
4922 * returns pageblock_bits flags
4924 unsigned long get_pageblock_flags_group(struct page
*page
,
4925 int start_bitidx
, int end_bitidx
)
4928 unsigned long *bitmap
;
4929 unsigned long pfn
, bitidx
;
4930 unsigned long flags
= 0;
4931 unsigned long value
= 1;
4933 zone
= page_zone(page
);
4934 pfn
= page_to_pfn(page
);
4935 bitmap
= get_pageblock_bitmap(zone
, pfn
);
4936 bitidx
= pfn_to_bitidx(zone
, pfn
);
4938 for (; start_bitidx
<= end_bitidx
; start_bitidx
++, value
<<= 1)
4939 if (test_bit(bitidx
+ start_bitidx
, bitmap
))
4946 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
4947 * @page: The page within the block of interest
4948 * @start_bitidx: The first bit of interest
4949 * @end_bitidx: The last bit of interest
4950 * @flags: The flags to set
4952 void set_pageblock_flags_group(struct page
*page
, unsigned long flags
,
4953 int start_bitidx
, int end_bitidx
)
4956 unsigned long *bitmap
;
4957 unsigned long pfn
, bitidx
;
4958 unsigned long value
= 1;
4960 zone
= page_zone(page
);
4961 pfn
= page_to_pfn(page
);
4962 bitmap
= get_pageblock_bitmap(zone
, pfn
);
4963 bitidx
= pfn_to_bitidx(zone
, pfn
);
4964 VM_BUG_ON(pfn
< zone
->zone_start_pfn
);
4965 VM_BUG_ON(pfn
>= zone
->zone_start_pfn
+ zone
->spanned_pages
);
4967 for (; start_bitidx
<= end_bitidx
; start_bitidx
++, value
<<= 1)
4969 __set_bit(bitidx
+ start_bitidx
, bitmap
);
4971 __clear_bit(bitidx
+ start_bitidx
, bitmap
);
4975 * This is designed as sub function...plz see page_isolation.c also.
4976 * set/clear page block's type to be ISOLATE.
4977 * page allocater never alloc memory from ISOLATE block.
4980 int set_migratetype_isolate(struct page
*page
)
4983 unsigned long flags
;
4987 zone
= page_zone(page
);
4988 zone_idx
= zone_idx(zone
);
4989 spin_lock_irqsave(&zone
->lock
, flags
);
4991 * In future, more migrate types will be able to be isolation target.
4993 if (get_pageblock_migratetype(page
) != MIGRATE_MOVABLE
&&
4994 zone_idx
!= ZONE_MOVABLE
)
4996 set_pageblock_migratetype(page
, MIGRATE_ISOLATE
);
4997 move_freepages_block(zone
, page
, MIGRATE_ISOLATE
);
5000 spin_unlock_irqrestore(&zone
->lock
, flags
);
5006 void unset_migratetype_isolate(struct page
*page
)
5009 unsigned long flags
;
5010 zone
= page_zone(page
);
5011 spin_lock_irqsave(&zone
->lock
, flags
);
5012 if (get_pageblock_migratetype(page
) != MIGRATE_ISOLATE
)
5014 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
5015 move_freepages_block(zone
, page
, MIGRATE_MOVABLE
);
5017 spin_unlock_irqrestore(&zone
->lock
, flags
);
5020 #ifdef CONFIG_MEMORY_HOTREMOVE
5022 * All pages in the range must be isolated before calling this.
5025 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
5031 unsigned long flags
;
5032 /* find the first valid pfn */
5033 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
5038 zone
= page_zone(pfn_to_page(pfn
));
5039 spin_lock_irqsave(&zone
->lock
, flags
);
5041 while (pfn
< end_pfn
) {
5042 if (!pfn_valid(pfn
)) {
5046 page
= pfn_to_page(pfn
);
5047 BUG_ON(page_count(page
));
5048 BUG_ON(!PageBuddy(page
));
5049 order
= page_order(page
);
5050 #ifdef CONFIG_DEBUG_VM
5051 printk(KERN_INFO
"remove from free list %lx %d %lx\n",
5052 pfn
, 1 << order
, end_pfn
);
5054 list_del(&page
->lru
);
5055 rmv_page_order(page
);
5056 zone
->free_area
[order
].nr_free
--;
5057 __mod_zone_page_state(zone
, NR_FREE_PAGES
,
5059 for (i
= 0; i
< (1 << order
); i
++)
5060 SetPageReserved((page
+i
));
5061 pfn
+= (1 << order
);
5063 spin_unlock_irqrestore(&zone
->lock
, flags
);