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/module.h>
27 #include <linux/suspend.h>
28 #include <linux/pagevec.h>
29 #include <linux/blkdev.h>
30 #include <linux/slab.h>
31 #include <linux/oom.h>
32 #include <linux/notifier.h>
33 #include <linux/topology.h>
34 #include <linux/sysctl.h>
35 #include <linux/cpu.h>
36 #include <linux/cpuset.h>
37 #include <linux/memory_hotplug.h>
38 #include <linux/nodemask.h>
39 #include <linux/vmalloc.h>
40 #include <linux/mempolicy.h>
41 #include <linux/stop_machine.h>
42 #include <linux/sort.h>
43 #include <linux/pfn.h>
44 #include <linux/backing-dev.h>
45 #include <linux/fault-inject.h>
46 #include <linux/page-isolation.h>
47 #include <linux/page_cgroup.h>
48 #include <linux/debugobjects.h>
49 #include <linux/kmemleak.h>
51 #include <asm/tlbflush.h>
52 #include <asm/div64.h>
56 * Array of node states.
58 nodemask_t node_states
[NR_NODE_STATES
] __read_mostly
= {
59 [N_POSSIBLE
] = NODE_MASK_ALL
,
60 [N_ONLINE
] = { { [0] = 1UL } },
62 [N_NORMAL_MEMORY
] = { { [0] = 1UL } },
64 [N_HIGH_MEMORY
] = { { [0] = 1UL } },
66 [N_CPU
] = { { [0] = 1UL } },
69 EXPORT_SYMBOL(node_states
);
71 unsigned long totalram_pages __read_mostly
;
72 unsigned long totalreserve_pages __read_mostly
;
73 unsigned long highest_memmap_pfn __read_mostly
;
74 int percpu_pagelist_fraction
;
76 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
77 int pageblock_order __read_mostly
;
80 static void __free_pages_ok(struct page
*page
, unsigned int order
);
83 * results with 256, 32 in the lowmem_reserve sysctl:
84 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
85 * 1G machine -> (16M dma, 784M normal, 224M high)
86 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
87 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
88 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
90 * TBD: should special case ZONE_DMA32 machines here - in those we normally
91 * don't need any ZONE_NORMAL reservation
93 int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
-1] = {
94 #ifdef CONFIG_ZONE_DMA
97 #ifdef CONFIG_ZONE_DMA32
100 #ifdef CONFIG_HIGHMEM
106 EXPORT_SYMBOL(totalram_pages
);
108 static char * const zone_names
[MAX_NR_ZONES
] = {
109 #ifdef CONFIG_ZONE_DMA
112 #ifdef CONFIG_ZONE_DMA32
116 #ifdef CONFIG_HIGHMEM
122 int min_free_kbytes
= 1024;
124 unsigned long __meminitdata nr_kernel_pages
;
125 unsigned long __meminitdata nr_all_pages
;
126 static unsigned long __meminitdata dma_reserve
;
128 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
130 * MAX_ACTIVE_REGIONS determines the maximum number of distinct
131 * ranges of memory (RAM) that may be registered with add_active_range().
132 * Ranges passed to add_active_range() will be merged if possible
133 * so the number of times add_active_range() can be called is
134 * related to the number of nodes and the number of holes
136 #ifdef CONFIG_MAX_ACTIVE_REGIONS
137 /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */
138 #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS
140 #if MAX_NUMNODES >= 32
141 /* If there can be many nodes, allow up to 50 holes per node */
142 #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50)
144 /* By default, allow up to 256 distinct regions */
145 #define MAX_ACTIVE_REGIONS 256
149 static struct node_active_region __meminitdata early_node_map
[MAX_ACTIVE_REGIONS
];
150 static int __meminitdata nr_nodemap_entries
;
151 static unsigned long __meminitdata arch_zone_lowest_possible_pfn
[MAX_NR_ZONES
];
152 static unsigned long __meminitdata arch_zone_highest_possible_pfn
[MAX_NR_ZONES
];
153 static unsigned long __initdata required_kernelcore
;
154 static unsigned long __initdata required_movablecore
;
155 static unsigned long __meminitdata zone_movable_pfn
[MAX_NUMNODES
];
157 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
159 EXPORT_SYMBOL(movable_zone
);
160 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
163 int nr_node_ids __read_mostly
= MAX_NUMNODES
;
164 int nr_online_nodes __read_mostly
= 1;
165 EXPORT_SYMBOL(nr_node_ids
);
166 EXPORT_SYMBOL(nr_online_nodes
);
169 int page_group_by_mobility_disabled __read_mostly
;
171 static void set_pageblock_migratetype(struct page
*page
, int migratetype
)
174 if (unlikely(page_group_by_mobility_disabled
))
175 migratetype
= MIGRATE_UNMOVABLE
;
177 set_pageblock_flags_group(page
, (unsigned long)migratetype
,
178 PB_migrate
, PB_migrate_end
);
181 #ifdef CONFIG_DEBUG_VM
182 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
186 unsigned long pfn
= page_to_pfn(page
);
189 seq
= zone_span_seqbegin(zone
);
190 if (pfn
>= zone
->zone_start_pfn
+ zone
->spanned_pages
)
192 else if (pfn
< zone
->zone_start_pfn
)
194 } while (zone_span_seqretry(zone
, seq
));
199 static int page_is_consistent(struct zone
*zone
, struct page
*page
)
201 if (!pfn_valid_within(page_to_pfn(page
)))
203 if (zone
!= page_zone(page
))
209 * Temporary debugging check for pages not lying within a given zone.
211 static int bad_range(struct zone
*zone
, struct page
*page
)
213 if (page_outside_zone_boundaries(zone
, page
))
215 if (!page_is_consistent(zone
, page
))
221 static inline int bad_range(struct zone
*zone
, struct page
*page
)
227 static void bad_page(struct page
*page
)
229 static unsigned long resume
;
230 static unsigned long nr_shown
;
231 static unsigned long nr_unshown
;
234 * Allow a burst of 60 reports, then keep quiet for that minute;
235 * or allow a steady drip of one report per second.
237 if (nr_shown
== 60) {
238 if (time_before(jiffies
, resume
)) {
244 "BUG: Bad page state: %lu messages suppressed\n",
251 resume
= jiffies
+ 60 * HZ
;
253 printk(KERN_ALERT
"BUG: Bad page state in process %s pfn:%05lx\n",
254 current
->comm
, page_to_pfn(page
));
256 "page:%p flags:%p count:%d mapcount:%d mapping:%p index:%lx\n",
257 page
, (void *)page
->flags
, page_count(page
),
258 page_mapcount(page
), page
->mapping
, page
->index
);
262 /* Leave bad fields for debug, except PageBuddy could make trouble */
263 __ClearPageBuddy(page
);
264 add_taint(TAINT_BAD_PAGE
);
268 * Higher-order pages are called "compound pages". They are structured thusly:
270 * The first PAGE_SIZE page is called the "head page".
272 * The remaining PAGE_SIZE pages are called "tail pages".
274 * All pages have PG_compound set. All pages have their ->private pointing at
275 * the head page (even the head page has this).
277 * The first tail page's ->lru.next holds the address of the compound page's
278 * put_page() function. Its ->lru.prev holds the order of allocation.
279 * This usage means that zero-order pages may not be compound.
282 static void free_compound_page(struct page
*page
)
284 __free_pages_ok(page
, compound_order(page
));
287 void prep_compound_page(struct page
*page
, unsigned long order
)
290 int nr_pages
= 1 << order
;
292 set_compound_page_dtor(page
, free_compound_page
);
293 set_compound_order(page
, order
);
295 for (i
= 1; i
< nr_pages
; i
++) {
296 struct page
*p
= page
+ i
;
299 p
->first_page
= page
;
303 #ifdef CONFIG_HUGETLBFS
304 void prep_compound_gigantic_page(struct page
*page
, unsigned long order
)
307 int nr_pages
= 1 << order
;
308 struct page
*p
= page
+ 1;
310 set_compound_page_dtor(page
, free_compound_page
);
311 set_compound_order(page
, order
);
313 for (i
= 1; i
< nr_pages
; i
++, p
= mem_map_next(p
, page
, i
)) {
315 p
->first_page
= page
;
320 static int destroy_compound_page(struct page
*page
, unsigned long order
)
323 int nr_pages
= 1 << order
;
326 if (unlikely(compound_order(page
) != order
) ||
327 unlikely(!PageHead(page
))) {
332 __ClearPageHead(page
);
334 for (i
= 1; i
< nr_pages
; i
++) {
335 struct page
*p
= page
+ i
;
337 if (unlikely(!PageTail(p
) || (p
->first_page
!= page
))) {
347 static inline void prep_zero_page(struct page
*page
, int order
, gfp_t gfp_flags
)
352 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
353 * and __GFP_HIGHMEM from hard or soft interrupt context.
355 VM_BUG_ON((gfp_flags
& __GFP_HIGHMEM
) && in_interrupt());
356 for (i
= 0; i
< (1 << order
); i
++)
357 clear_highpage(page
+ i
);
360 static inline void set_page_order(struct page
*page
, int order
)
362 set_page_private(page
, order
);
363 __SetPageBuddy(page
);
366 static inline void rmv_page_order(struct page
*page
)
368 __ClearPageBuddy(page
);
369 set_page_private(page
, 0);
373 * Locate the struct page for both the matching buddy in our
374 * pair (buddy1) and the combined O(n+1) page they form (page).
376 * 1) Any buddy B1 will have an order O twin B2 which satisfies
377 * the following equation:
379 * For example, if the starting buddy (buddy2) is #8 its order
381 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
383 * 2) Any buddy B will have an order O+1 parent P which
384 * satisfies the following equation:
387 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
389 static inline struct page
*
390 __page_find_buddy(struct page
*page
, unsigned long page_idx
, unsigned int order
)
392 unsigned long buddy_idx
= page_idx
^ (1 << order
);
394 return page
+ (buddy_idx
- page_idx
);
397 static inline unsigned long
398 __find_combined_index(unsigned long page_idx
, unsigned int order
)
400 return (page_idx
& ~(1 << order
));
404 * This function checks whether a page is free && is the buddy
405 * we can do coalesce a page and its buddy if
406 * (a) the buddy is not in a hole &&
407 * (b) the buddy is in the buddy system &&
408 * (c) a page and its buddy have the same order &&
409 * (d) a page and its buddy are in the same zone.
411 * For recording whether a page is in the buddy system, we use PG_buddy.
412 * Setting, clearing, and testing PG_buddy is serialized by zone->lock.
414 * For recording page's order, we use page_private(page).
416 static inline int page_is_buddy(struct page
*page
, struct page
*buddy
,
419 if (!pfn_valid_within(page_to_pfn(buddy
)))
422 if (page_zone_id(page
) != page_zone_id(buddy
))
425 if (PageBuddy(buddy
) && page_order(buddy
) == order
) {
426 VM_BUG_ON(page_count(buddy
) != 0);
433 * Freeing function for a buddy system allocator.
435 * The concept of a buddy system is to maintain direct-mapped table
436 * (containing bit values) for memory blocks of various "orders".
437 * The bottom level table contains the map for the smallest allocatable
438 * units of memory (here, pages), and each level above it describes
439 * pairs of units from the levels below, hence, "buddies".
440 * At a high level, all that happens here is marking the table entry
441 * at the bottom level available, and propagating the changes upward
442 * as necessary, plus some accounting needed to play nicely with other
443 * parts of the VM system.
444 * At each level, we keep a list of pages, which are heads of continuous
445 * free pages of length of (1 << order) and marked with PG_buddy. Page's
446 * order is recorded in page_private(page) field.
447 * So when we are allocating or freeing one, we can derive the state of the
448 * other. That is, if we allocate a small block, and both were
449 * free, the remainder of the region must be split into blocks.
450 * If a block is freed, and its buddy is also free, then this
451 * triggers coalescing into a block of larger size.
456 static inline void __free_one_page(struct page
*page
,
457 struct zone
*zone
, unsigned int order
,
460 unsigned long page_idx
;
462 if (unlikely(PageCompound(page
)))
463 if (unlikely(destroy_compound_page(page
, order
)))
466 VM_BUG_ON(migratetype
== -1);
468 page_idx
= page_to_pfn(page
) & ((1 << MAX_ORDER
) - 1);
470 VM_BUG_ON(page_idx
& ((1 << order
) - 1));
471 VM_BUG_ON(bad_range(zone
, page
));
473 while (order
< MAX_ORDER
-1) {
474 unsigned long combined_idx
;
477 buddy
= __page_find_buddy(page
, page_idx
, order
);
478 if (!page_is_buddy(page
, buddy
, order
))
481 /* Our buddy is free, merge with it and move up one order. */
482 list_del(&buddy
->lru
);
483 zone
->free_area
[order
].nr_free
--;
484 rmv_page_order(buddy
);
485 combined_idx
= __find_combined_index(page_idx
, order
);
486 page
= page
+ (combined_idx
- page_idx
);
487 page_idx
= combined_idx
;
490 set_page_order(page
, order
);
492 &zone
->free_area
[order
].free_list
[migratetype
]);
493 zone
->free_area
[order
].nr_free
++;
496 #ifdef CONFIG_HAVE_MLOCKED_PAGE_BIT
498 * free_page_mlock() -- clean up attempts to free and mlocked() page.
499 * Page should not be on lru, so no need to fix that up.
500 * free_pages_check() will verify...
502 static inline void free_page_mlock(struct page
*page
)
504 __ClearPageMlocked(page
);
505 __dec_zone_page_state(page
, NR_MLOCK
);
506 __count_vm_event(UNEVICTABLE_MLOCKFREED
);
509 static void free_page_mlock(struct page
*page
) { }
512 static inline int free_pages_check(struct page
*page
)
514 if (unlikely(page_mapcount(page
) |
515 (page
->mapping
!= NULL
) |
516 (atomic_read(&page
->_count
) != 0) |
517 (page
->flags
& PAGE_FLAGS_CHECK_AT_FREE
))) {
521 if (page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)
522 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
527 * Frees a list of pages.
528 * Assumes all pages on list are in same zone, and of same order.
529 * count is the number of pages to free.
531 * If the zone was previously in an "all pages pinned" state then look to
532 * see if this freeing clears that state.
534 * And clear the zone's pages_scanned counter, to hold off the "all pages are
535 * pinned" detection logic.
537 static void free_pages_bulk(struct zone
*zone
, int count
,
538 struct list_head
*list
, int order
)
540 spin_lock(&zone
->lock
);
541 zone_clear_flag(zone
, ZONE_ALL_UNRECLAIMABLE
);
542 zone
->pages_scanned
= 0;
544 __mod_zone_page_state(zone
, NR_FREE_PAGES
, count
<< order
);
548 VM_BUG_ON(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 __free_one_page(page
, zone
, order
, page_private(page
));
554 spin_unlock(&zone
->lock
);
557 static void free_one_page(struct zone
*zone
, struct page
*page
, int order
,
560 spin_lock(&zone
->lock
);
561 zone_clear_flag(zone
, ZONE_ALL_UNRECLAIMABLE
);
562 zone
->pages_scanned
= 0;
564 __mod_zone_page_state(zone
, NR_FREE_PAGES
, 1 << order
);
565 __free_one_page(page
, zone
, order
, migratetype
);
566 spin_unlock(&zone
->lock
);
569 static void __free_pages_ok(struct page
*page
, unsigned int order
)
574 int clearMlocked
= PageMlocked(page
);
576 for (i
= 0 ; i
< (1 << order
) ; ++i
)
577 bad
+= free_pages_check(page
+ i
);
581 if (!PageHighMem(page
)) {
582 debug_check_no_locks_freed(page_address(page
),PAGE_SIZE
<<order
);
583 debug_check_no_obj_freed(page_address(page
),
586 arch_free_page(page
, order
);
587 kernel_map_pages(page
, 1 << order
, 0);
589 local_irq_save(flags
);
590 if (unlikely(clearMlocked
))
591 free_page_mlock(page
);
592 __count_vm_events(PGFREE
, 1 << order
);
593 free_one_page(page_zone(page
), page
, order
,
594 get_pageblock_migratetype(page
));
595 local_irq_restore(flags
);
599 * permit the bootmem allocator to evade page validation on high-order frees
601 void __meminit
__free_pages_bootmem(struct page
*page
, unsigned int order
)
604 __ClearPageReserved(page
);
605 set_page_count(page
, 0);
606 set_page_refcounted(page
);
612 for (loop
= 0; loop
< BITS_PER_LONG
; loop
++) {
613 struct page
*p
= &page
[loop
];
615 if (loop
+ 1 < BITS_PER_LONG
)
617 __ClearPageReserved(p
);
618 set_page_count(p
, 0);
621 set_page_refcounted(page
);
622 __free_pages(page
, order
);
628 * The order of subdivision here is critical for the IO subsystem.
629 * Please do not alter this order without good reasons and regression
630 * testing. Specifically, as large blocks of memory are subdivided,
631 * the order in which smaller blocks are delivered depends on the order
632 * they're subdivided in this function. This is the primary factor
633 * influencing the order in which pages are delivered to the IO
634 * subsystem according to empirical testing, and this is also justified
635 * by considering the behavior of a buddy system containing a single
636 * large block of memory acted on by a series of small allocations.
637 * This behavior is a critical factor in sglist merging's success.
641 static inline void expand(struct zone
*zone
, struct page
*page
,
642 int low
, int high
, struct free_area
*area
,
645 unsigned long size
= 1 << high
;
651 VM_BUG_ON(bad_range(zone
, &page
[size
]));
652 list_add(&page
[size
].lru
, &area
->free_list
[migratetype
]);
654 set_page_order(&page
[size
], high
);
659 * This page is about to be returned from the page allocator
661 static int prep_new_page(struct page
*page
, int order
, gfp_t gfp_flags
)
663 if (unlikely(page_mapcount(page
) |
664 (page
->mapping
!= NULL
) |
665 (atomic_read(&page
->_count
) != 0) |
666 (page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
))) {
671 set_page_private(page
, 0);
672 set_page_refcounted(page
);
674 arch_alloc_page(page
, order
);
675 kernel_map_pages(page
, 1 << order
, 1);
677 if (gfp_flags
& __GFP_ZERO
)
678 prep_zero_page(page
, order
, gfp_flags
);
680 if (order
&& (gfp_flags
& __GFP_COMP
))
681 prep_compound_page(page
, order
);
687 * Go through the free lists for the given migratetype and remove
688 * the smallest available page from the freelists
691 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
694 unsigned int current_order
;
695 struct free_area
* area
;
698 /* Find a page of the appropriate size in the preferred list */
699 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
700 area
= &(zone
->free_area
[current_order
]);
701 if (list_empty(&area
->free_list
[migratetype
]))
704 page
= list_entry(area
->free_list
[migratetype
].next
,
706 list_del(&page
->lru
);
707 rmv_page_order(page
);
709 expand(zone
, page
, order
, current_order
, area
, migratetype
);
718 * This array describes the order lists are fallen back to when
719 * the free lists for the desirable migrate type are depleted
721 static int fallbacks
[MIGRATE_TYPES
][MIGRATE_TYPES
-1] = {
722 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
723 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
724 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_RESERVE
},
725 [MIGRATE_RESERVE
] = { MIGRATE_RESERVE
, MIGRATE_RESERVE
, MIGRATE_RESERVE
}, /* Never used */
729 * Move the free pages in a range to the free lists of the requested type.
730 * Note that start_page and end_pages are not aligned on a pageblock
731 * boundary. If alignment is required, use move_freepages_block()
733 static int move_freepages(struct zone
*zone
,
734 struct page
*start_page
, struct page
*end_page
,
741 #ifndef CONFIG_HOLES_IN_ZONE
743 * page_zone is not safe to call in this context when
744 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
745 * anyway as we check zone boundaries in move_freepages_block().
746 * Remove at a later date when no bug reports exist related to
747 * grouping pages by mobility
749 BUG_ON(page_zone(start_page
) != page_zone(end_page
));
752 for (page
= start_page
; page
<= end_page
;) {
753 /* Make sure we are not inadvertently changing nodes */
754 VM_BUG_ON(page_to_nid(page
) != zone_to_nid(zone
));
756 if (!pfn_valid_within(page_to_pfn(page
))) {
761 if (!PageBuddy(page
)) {
766 order
= page_order(page
);
767 list_del(&page
->lru
);
769 &zone
->free_area
[order
].free_list
[migratetype
]);
771 pages_moved
+= 1 << order
;
777 static int move_freepages_block(struct zone
*zone
, struct page
*page
,
780 unsigned long start_pfn
, end_pfn
;
781 struct page
*start_page
, *end_page
;
783 start_pfn
= page_to_pfn(page
);
784 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
785 start_page
= pfn_to_page(start_pfn
);
786 end_page
= start_page
+ pageblock_nr_pages
- 1;
787 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
789 /* Do not cross zone boundaries */
790 if (start_pfn
< zone
->zone_start_pfn
)
792 if (end_pfn
>= zone
->zone_start_pfn
+ zone
->spanned_pages
)
795 return move_freepages(zone
, start_page
, end_page
, migratetype
);
798 /* Remove an element from the buddy allocator from the fallback list */
799 static inline struct page
*
800 __rmqueue_fallback(struct zone
*zone
, int order
, int start_migratetype
)
802 struct free_area
* area
;
807 /* Find the largest possible block of pages in the other list */
808 for (current_order
= MAX_ORDER
-1; current_order
>= order
;
810 for (i
= 0; i
< MIGRATE_TYPES
- 1; i
++) {
811 migratetype
= fallbacks
[start_migratetype
][i
];
813 /* MIGRATE_RESERVE handled later if necessary */
814 if (migratetype
== MIGRATE_RESERVE
)
817 area
= &(zone
->free_area
[current_order
]);
818 if (list_empty(&area
->free_list
[migratetype
]))
821 page
= list_entry(area
->free_list
[migratetype
].next
,
826 * If breaking a large block of pages, move all free
827 * pages to the preferred allocation list. If falling
828 * back for a reclaimable kernel allocation, be more
829 * agressive about taking ownership of free pages
831 if (unlikely(current_order
>= (pageblock_order
>> 1)) ||
832 start_migratetype
== MIGRATE_RECLAIMABLE
) {
834 pages
= move_freepages_block(zone
, page
,
837 /* Claim the whole block if over half of it is free */
838 if (pages
>= (1 << (pageblock_order
-1)))
839 set_pageblock_migratetype(page
,
842 migratetype
= start_migratetype
;
845 /* Remove the page from the freelists */
846 list_del(&page
->lru
);
847 rmv_page_order(page
);
849 if (current_order
== pageblock_order
)
850 set_pageblock_migratetype(page
,
853 expand(zone
, page
, order
, current_order
, area
, migratetype
);
862 * Do the hard work of removing an element from the buddy allocator.
863 * Call me with the zone->lock already held.
865 static struct page
*__rmqueue(struct zone
*zone
, unsigned int order
,
871 page
= __rmqueue_smallest(zone
, order
, migratetype
);
873 if (unlikely(!page
) && migratetype
!= MIGRATE_RESERVE
) {
874 page
= __rmqueue_fallback(zone
, order
, migratetype
);
877 * Use MIGRATE_RESERVE rather than fail an allocation. goto
878 * is used because __rmqueue_smallest is an inline function
879 * and we want just one call site
882 migratetype
= MIGRATE_RESERVE
;
891 * Obtain a specified number of elements from the buddy allocator, all under
892 * a single hold of the lock, for efficiency. Add them to the supplied list.
893 * Returns the number of new pages which were placed at *list.
895 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
896 unsigned long count
, struct list_head
*list
,
901 spin_lock(&zone
->lock
);
902 for (i
= 0; i
< count
; ++i
) {
903 struct page
*page
= __rmqueue(zone
, order
, migratetype
);
904 if (unlikely(page
== NULL
))
908 * Split buddy pages returned by expand() are received here
909 * in physical page order. The page is added to the callers and
910 * list and the list head then moves forward. From the callers
911 * perspective, the linked list is ordered by page number in
912 * some conditions. This is useful for IO devices that can
913 * merge IO requests if the physical pages are ordered
916 list_add(&page
->lru
, list
);
917 set_page_private(page
, migratetype
);
920 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
921 spin_unlock(&zone
->lock
);
927 * Called from the vmstat counter updater to drain pagesets of this
928 * currently executing processor on remote nodes after they have
931 * Note that this function must be called with the thread pinned to
932 * a single processor.
934 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
939 local_irq_save(flags
);
940 if (pcp
->count
>= pcp
->batch
)
941 to_drain
= pcp
->batch
;
943 to_drain
= pcp
->count
;
944 free_pages_bulk(zone
, to_drain
, &pcp
->list
, 0);
945 pcp
->count
-= to_drain
;
946 local_irq_restore(flags
);
951 * Drain pages of the indicated processor.
953 * The processor must either be the current processor and the
954 * thread pinned to the current processor or a processor that
957 static void drain_pages(unsigned int cpu
)
962 for_each_populated_zone(zone
) {
963 struct per_cpu_pageset
*pset
;
964 struct per_cpu_pages
*pcp
;
966 pset
= zone_pcp(zone
, cpu
);
969 local_irq_save(flags
);
970 free_pages_bulk(zone
, pcp
->count
, &pcp
->list
, 0);
972 local_irq_restore(flags
);
977 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
979 void drain_local_pages(void *arg
)
981 drain_pages(smp_processor_id());
985 * Spill all the per-cpu pages from all CPUs back into the buddy allocator
987 void drain_all_pages(void)
989 on_each_cpu(drain_local_pages
, NULL
, 1);
992 #ifdef CONFIG_HIBERNATION
994 void mark_free_pages(struct zone
*zone
)
996 unsigned long pfn
, max_zone_pfn
;
999 struct list_head
*curr
;
1001 if (!zone
->spanned_pages
)
1004 spin_lock_irqsave(&zone
->lock
, flags
);
1006 max_zone_pfn
= zone
->zone_start_pfn
+ zone
->spanned_pages
;
1007 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
1008 if (pfn_valid(pfn
)) {
1009 struct page
*page
= pfn_to_page(pfn
);
1011 if (!swsusp_page_is_forbidden(page
))
1012 swsusp_unset_page_free(page
);
1015 for_each_migratetype_order(order
, t
) {
1016 list_for_each(curr
, &zone
->free_area
[order
].free_list
[t
]) {
1019 pfn
= page_to_pfn(list_entry(curr
, struct page
, lru
));
1020 for (i
= 0; i
< (1UL << order
); i
++)
1021 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
1024 spin_unlock_irqrestore(&zone
->lock
, flags
);
1026 #endif /* CONFIG_PM */
1029 * Free a 0-order page
1031 static void free_hot_cold_page(struct page
*page
, int cold
)
1033 struct zone
*zone
= page_zone(page
);
1034 struct per_cpu_pages
*pcp
;
1035 unsigned long flags
;
1036 int clearMlocked
= PageMlocked(page
);
1039 page
->mapping
= NULL
;
1040 if (free_pages_check(page
))
1043 if (!PageHighMem(page
)) {
1044 debug_check_no_locks_freed(page_address(page
), PAGE_SIZE
);
1045 debug_check_no_obj_freed(page_address(page
), PAGE_SIZE
);
1047 arch_free_page(page
, 0);
1048 kernel_map_pages(page
, 1, 0);
1050 pcp
= &zone_pcp(zone
, get_cpu())->pcp
;
1051 set_page_private(page
, get_pageblock_migratetype(page
));
1052 local_irq_save(flags
);
1053 if (unlikely(clearMlocked
))
1054 free_page_mlock(page
);
1055 __count_vm_event(PGFREE
);
1058 list_add_tail(&page
->lru
, &pcp
->list
);
1060 list_add(&page
->lru
, &pcp
->list
);
1062 if (pcp
->count
>= pcp
->high
) {
1063 free_pages_bulk(zone
, pcp
->batch
, &pcp
->list
, 0);
1064 pcp
->count
-= pcp
->batch
;
1066 local_irq_restore(flags
);
1070 void free_hot_page(struct page
*page
)
1072 free_hot_cold_page(page
, 0);
1075 void free_cold_page(struct page
*page
)
1077 free_hot_cold_page(page
, 1);
1081 * split_page takes a non-compound higher-order page, and splits it into
1082 * n (1<<order) sub-pages: page[0..n]
1083 * Each sub-page must be freed individually.
1085 * Note: this is probably too low level an operation for use in drivers.
1086 * Please consult with lkml before using this in your driver.
1088 void split_page(struct page
*page
, unsigned int order
)
1092 VM_BUG_ON(PageCompound(page
));
1093 VM_BUG_ON(!page_count(page
));
1094 for (i
= 1; i
< (1 << order
); i
++)
1095 set_page_refcounted(page
+ i
);
1099 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1100 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1104 struct page
*buffered_rmqueue(struct zone
*preferred_zone
,
1105 struct zone
*zone
, int order
, gfp_t gfp_flags
,
1108 unsigned long flags
;
1110 int cold
= !!(gfp_flags
& __GFP_COLD
);
1115 if (likely(order
== 0)) {
1116 struct per_cpu_pages
*pcp
;
1118 pcp
= &zone_pcp(zone
, cpu
)->pcp
;
1119 local_irq_save(flags
);
1121 pcp
->count
= rmqueue_bulk(zone
, 0,
1122 pcp
->batch
, &pcp
->list
, migratetype
);
1123 if (unlikely(!pcp
->count
))
1127 /* Find a page of the appropriate migrate type */
1129 list_for_each_entry_reverse(page
, &pcp
->list
, lru
)
1130 if (page_private(page
) == migratetype
)
1133 list_for_each_entry(page
, &pcp
->list
, lru
)
1134 if (page_private(page
) == migratetype
)
1138 /* Allocate more to the pcp list if necessary */
1139 if (unlikely(&page
->lru
== &pcp
->list
)) {
1140 pcp
->count
+= rmqueue_bulk(zone
, 0,
1141 pcp
->batch
, &pcp
->list
, migratetype
);
1142 page
= list_entry(pcp
->list
.next
, struct page
, lru
);
1145 list_del(&page
->lru
);
1148 spin_lock_irqsave(&zone
->lock
, flags
);
1149 page
= __rmqueue(zone
, order
, migratetype
);
1150 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(1 << order
));
1151 spin_unlock(&zone
->lock
);
1156 __count_zone_vm_events(PGALLOC
, zone
, 1 << order
);
1157 zone_statistics(preferred_zone
, zone
);
1158 local_irq_restore(flags
);
1161 VM_BUG_ON(bad_range(zone
, page
));
1162 if (prep_new_page(page
, order
, gfp_flags
))
1167 local_irq_restore(flags
);
1172 /* The ALLOC_WMARK bits are used as an index to zone->watermark */
1173 #define ALLOC_WMARK_MIN WMARK_MIN
1174 #define ALLOC_WMARK_LOW WMARK_LOW
1175 #define ALLOC_WMARK_HIGH WMARK_HIGH
1176 #define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */
1178 /* Mask to get the watermark bits */
1179 #define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1)
1181 #define ALLOC_HARDER 0x10 /* try to alloc harder */
1182 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
1183 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
1185 #ifdef CONFIG_FAIL_PAGE_ALLOC
1187 static struct fail_page_alloc_attr
{
1188 struct fault_attr attr
;
1190 u32 ignore_gfp_highmem
;
1191 u32 ignore_gfp_wait
;
1194 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1196 struct dentry
*ignore_gfp_highmem_file
;
1197 struct dentry
*ignore_gfp_wait_file
;
1198 struct dentry
*min_order_file
;
1200 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1202 } fail_page_alloc
= {
1203 .attr
= FAULT_ATTR_INITIALIZER
,
1204 .ignore_gfp_wait
= 1,
1205 .ignore_gfp_highmem
= 1,
1209 static int __init
setup_fail_page_alloc(char *str
)
1211 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
1213 __setup("fail_page_alloc=", setup_fail_page_alloc
);
1215 static int should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1217 if (order
< fail_page_alloc
.min_order
)
1219 if (gfp_mask
& __GFP_NOFAIL
)
1221 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
1223 if (fail_page_alloc
.ignore_gfp_wait
&& (gfp_mask
& __GFP_WAIT
))
1226 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
1229 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1231 static int __init
fail_page_alloc_debugfs(void)
1233 mode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
1237 err
= init_fault_attr_dentries(&fail_page_alloc
.attr
,
1241 dir
= fail_page_alloc
.attr
.dentries
.dir
;
1243 fail_page_alloc
.ignore_gfp_wait_file
=
1244 debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
1245 &fail_page_alloc
.ignore_gfp_wait
);
1247 fail_page_alloc
.ignore_gfp_highmem_file
=
1248 debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
1249 &fail_page_alloc
.ignore_gfp_highmem
);
1250 fail_page_alloc
.min_order_file
=
1251 debugfs_create_u32("min-order", mode
, dir
,
1252 &fail_page_alloc
.min_order
);
1254 if (!fail_page_alloc
.ignore_gfp_wait_file
||
1255 !fail_page_alloc
.ignore_gfp_highmem_file
||
1256 !fail_page_alloc
.min_order_file
) {
1258 debugfs_remove(fail_page_alloc
.ignore_gfp_wait_file
);
1259 debugfs_remove(fail_page_alloc
.ignore_gfp_highmem_file
);
1260 debugfs_remove(fail_page_alloc
.min_order_file
);
1261 cleanup_fault_attr_dentries(&fail_page_alloc
.attr
);
1267 late_initcall(fail_page_alloc_debugfs
);
1269 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1271 #else /* CONFIG_FAIL_PAGE_ALLOC */
1273 static inline int should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1278 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1281 * Return 1 if free pages are above 'mark'. This takes into account the order
1282 * of the allocation.
1284 int zone_watermark_ok(struct zone
*z
, int order
, unsigned long mark
,
1285 int classzone_idx
, int alloc_flags
)
1287 /* free_pages my go negative - that's OK */
1289 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
) - (1 << order
) + 1;
1292 if (alloc_flags
& ALLOC_HIGH
)
1294 if (alloc_flags
& ALLOC_HARDER
)
1297 if (free_pages
<= min
+ z
->lowmem_reserve
[classzone_idx
])
1299 for (o
= 0; o
< order
; o
++) {
1300 /* At the next order, this order's pages become unavailable */
1301 free_pages
-= z
->free_area
[o
].nr_free
<< o
;
1303 /* Require fewer higher order pages to be free */
1306 if (free_pages
<= min
)
1314 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1315 * skip over zones that are not allowed by the cpuset, or that have
1316 * been recently (in last second) found to be nearly full. See further
1317 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1318 * that have to skip over a lot of full or unallowed zones.
1320 * If the zonelist cache is present in the passed in zonelist, then
1321 * returns a pointer to the allowed node mask (either the current
1322 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1324 * If the zonelist cache is not available for this zonelist, does
1325 * nothing and returns NULL.
1327 * If the fullzones BITMAP in the zonelist cache is stale (more than
1328 * a second since last zap'd) then we zap it out (clear its bits.)
1330 * We hold off even calling zlc_setup, until after we've checked the
1331 * first zone in the zonelist, on the theory that most allocations will
1332 * be satisfied from that first zone, so best to examine that zone as
1333 * quickly as we can.
1335 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1337 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1338 nodemask_t
*allowednodes
; /* zonelist_cache approximation */
1340 zlc
= zonelist
->zlcache_ptr
;
1344 if (time_after(jiffies
, zlc
->last_full_zap
+ HZ
)) {
1345 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
1346 zlc
->last_full_zap
= jiffies
;
1349 allowednodes
= !in_interrupt() && (alloc_flags
& ALLOC_CPUSET
) ?
1350 &cpuset_current_mems_allowed
:
1351 &node_states
[N_HIGH_MEMORY
];
1352 return allowednodes
;
1356 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1357 * if it is worth looking at further for free memory:
1358 * 1) Check that the zone isn't thought to be full (doesn't have its
1359 * bit set in the zonelist_cache fullzones BITMAP).
1360 * 2) Check that the zones node (obtained from the zonelist_cache
1361 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1362 * Return true (non-zero) if zone is worth looking at further, or
1363 * else return false (zero) if it is not.
1365 * This check -ignores- the distinction between various watermarks,
1366 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1367 * found to be full for any variation of these watermarks, it will
1368 * be considered full for up to one second by all requests, unless
1369 * we are so low on memory on all allowed nodes that we are forced
1370 * into the second scan of the zonelist.
1372 * In the second scan we ignore this zonelist cache and exactly
1373 * apply the watermarks to all zones, even it is slower to do so.
1374 * We are low on memory in the second scan, and should leave no stone
1375 * unturned looking for a free page.
1377 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1378 nodemask_t
*allowednodes
)
1380 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1381 int i
; /* index of *z in zonelist zones */
1382 int n
; /* node that zone *z is on */
1384 zlc
= zonelist
->zlcache_ptr
;
1388 i
= z
- zonelist
->_zonerefs
;
1391 /* This zone is worth trying if it is allowed but not full */
1392 return node_isset(n
, *allowednodes
) && !test_bit(i
, zlc
->fullzones
);
1396 * Given 'z' scanning a zonelist, set the corresponding bit in
1397 * zlc->fullzones, so that subsequent attempts to allocate a page
1398 * from that zone don't waste time re-examining it.
1400 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1402 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1403 int i
; /* index of *z in zonelist zones */
1405 zlc
= zonelist
->zlcache_ptr
;
1409 i
= z
- zonelist
->_zonerefs
;
1411 set_bit(i
, zlc
->fullzones
);
1414 #else /* CONFIG_NUMA */
1416 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1421 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1422 nodemask_t
*allowednodes
)
1427 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1430 #endif /* CONFIG_NUMA */
1433 * get_page_from_freelist goes through the zonelist trying to allocate
1436 static struct page
*
1437 get_page_from_freelist(gfp_t gfp_mask
, nodemask_t
*nodemask
, unsigned int order
,
1438 struct zonelist
*zonelist
, int high_zoneidx
, int alloc_flags
,
1439 struct zone
*preferred_zone
, int migratetype
)
1442 struct page
*page
= NULL
;
1445 nodemask_t
*allowednodes
= NULL
;/* zonelist_cache approximation */
1446 int zlc_active
= 0; /* set if using zonelist_cache */
1447 int did_zlc_setup
= 0; /* just call zlc_setup() one time */
1449 classzone_idx
= zone_idx(preferred_zone
);
1452 * Scan zonelist, looking for a zone with enough free.
1453 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1455 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
1456 high_zoneidx
, nodemask
) {
1457 if (NUMA_BUILD
&& zlc_active
&&
1458 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
1460 if ((alloc_flags
& ALLOC_CPUSET
) &&
1461 !cpuset_zone_allowed_softwall(zone
, gfp_mask
))
1464 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
1465 if (!(alloc_flags
& ALLOC_NO_WATERMARKS
)) {
1467 mark
= zone
->watermark
[alloc_flags
& ALLOC_WMARK_MASK
];
1468 if (!zone_watermark_ok(zone
, order
, mark
,
1469 classzone_idx
, alloc_flags
)) {
1470 if (!zone_reclaim_mode
||
1471 !zone_reclaim(zone
, gfp_mask
, order
))
1472 goto this_zone_full
;
1476 page
= buffered_rmqueue(preferred_zone
, zone
, order
,
1477 gfp_mask
, migratetype
);
1482 zlc_mark_zone_full(zonelist
, z
);
1484 if (NUMA_BUILD
&& !did_zlc_setup
&& nr_online_nodes
> 1) {
1486 * we do zlc_setup after the first zone is tried but only
1487 * if there are multiple nodes make it worthwhile
1489 allowednodes
= zlc_setup(zonelist
, alloc_flags
);
1495 if (unlikely(NUMA_BUILD
&& page
== NULL
&& zlc_active
)) {
1496 /* Disable zlc cache for second zonelist scan */
1504 should_alloc_retry(gfp_t gfp_mask
, unsigned int order
,
1505 unsigned long pages_reclaimed
)
1507 /* Do not loop if specifically requested */
1508 if (gfp_mask
& __GFP_NORETRY
)
1512 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
1513 * means __GFP_NOFAIL, but that may not be true in other
1516 if (order
<= PAGE_ALLOC_COSTLY_ORDER
)
1520 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
1521 * specified, then we retry until we no longer reclaim any pages
1522 * (above), or we've reclaimed an order of pages at least as
1523 * large as the allocation's order. In both cases, if the
1524 * allocation still fails, we stop retrying.
1526 if (gfp_mask
& __GFP_REPEAT
&& pages_reclaimed
< (1 << order
))
1530 * Don't let big-order allocations loop unless the caller
1531 * explicitly requests that.
1533 if (gfp_mask
& __GFP_NOFAIL
)
1539 static inline struct page
*
1540 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
1541 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1542 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
1547 /* Acquire the OOM killer lock for the zones in zonelist */
1548 if (!try_set_zone_oom(zonelist
, gfp_mask
)) {
1549 schedule_timeout_uninterruptible(1);
1554 * Go through the zonelist yet one more time, keep very high watermark
1555 * here, this is only to catch a parallel oom killing, we must fail if
1556 * we're still under heavy pressure.
1558 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
,
1559 order
, zonelist
, high_zoneidx
,
1560 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
,
1561 preferred_zone
, migratetype
);
1565 /* The OOM killer will not help higher order allocs */
1566 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
1569 /* Exhausted what can be done so it's blamo time */
1570 out_of_memory(zonelist
, gfp_mask
, order
);
1573 clear_zonelist_oom(zonelist
, gfp_mask
);
1577 /* The really slow allocator path where we enter direct reclaim */
1578 static inline struct page
*
1579 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
1580 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1581 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
1582 int migratetype
, unsigned long *did_some_progress
)
1584 struct page
*page
= NULL
;
1585 struct reclaim_state reclaim_state
;
1586 struct task_struct
*p
= current
;
1590 /* We now go into synchronous reclaim */
1591 cpuset_memory_pressure_bump();
1594 * The task's cpuset might have expanded its set of allowable nodes
1596 p
->flags
|= PF_MEMALLOC
;
1597 lockdep_set_current_reclaim_state(gfp_mask
);
1598 reclaim_state
.reclaimed_slab
= 0;
1599 p
->reclaim_state
= &reclaim_state
;
1601 *did_some_progress
= try_to_free_pages(zonelist
, order
, gfp_mask
, nodemask
);
1603 p
->reclaim_state
= NULL
;
1604 lockdep_clear_current_reclaim_state();
1605 p
->flags
&= ~PF_MEMALLOC
;
1612 if (likely(*did_some_progress
))
1613 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
1614 zonelist
, high_zoneidx
,
1615 alloc_flags
, preferred_zone
,
1621 * This is called in the allocator slow-path if the allocation request is of
1622 * sufficient urgency to ignore watermarks and take other desperate measures
1624 static inline struct page
*
1625 __alloc_pages_high_priority(gfp_t gfp_mask
, unsigned int order
,
1626 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1627 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
1633 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
1634 zonelist
, high_zoneidx
, ALLOC_NO_WATERMARKS
,
1635 preferred_zone
, migratetype
);
1637 if (!page
&& gfp_mask
& __GFP_NOFAIL
)
1638 congestion_wait(WRITE
, HZ
/50);
1639 } while (!page
&& (gfp_mask
& __GFP_NOFAIL
));
1645 void wake_all_kswapd(unsigned int order
, struct zonelist
*zonelist
,
1646 enum zone_type high_zoneidx
)
1651 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
)
1652 wakeup_kswapd(zone
, order
);
1656 gfp_to_alloc_flags(gfp_t gfp_mask
)
1658 struct task_struct
*p
= current
;
1659 int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
1660 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
1662 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
1663 BUILD_BUG_ON(__GFP_HIGH
!= ALLOC_HIGH
);
1666 * The caller may dip into page reserves a bit more if the caller
1667 * cannot run direct reclaim, or if the caller has realtime scheduling
1668 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
1669 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
1671 alloc_flags
|= (gfp_mask
& __GFP_HIGH
);
1674 alloc_flags
|= ALLOC_HARDER
;
1676 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
1677 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1679 alloc_flags
&= ~ALLOC_CPUSET
;
1680 } else if (unlikely(rt_task(p
)))
1681 alloc_flags
|= ALLOC_HARDER
;
1683 if (likely(!(gfp_mask
& __GFP_NOMEMALLOC
))) {
1684 if (!in_interrupt() &&
1685 ((p
->flags
& PF_MEMALLOC
) ||
1686 unlikely(test_thread_flag(TIF_MEMDIE
))))
1687 alloc_flags
|= ALLOC_NO_WATERMARKS
;
1693 static inline struct page
*
1694 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
1695 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1696 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
1699 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
1700 struct page
*page
= NULL
;
1702 unsigned long pages_reclaimed
= 0;
1703 unsigned long did_some_progress
;
1704 struct task_struct
*p
= current
;
1707 * In the slowpath, we sanity check order to avoid ever trying to
1708 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
1709 * be using allocators in order of preference for an area that is
1712 if (WARN_ON_ONCE(order
>= MAX_ORDER
))
1716 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
1717 * __GFP_NOWARN set) should not cause reclaim since the subsystem
1718 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
1719 * using a larger set of nodes after it has established that the
1720 * allowed per node queues are empty and that nodes are
1723 if (NUMA_BUILD
&& (gfp_mask
& GFP_THISNODE
) == GFP_THISNODE
)
1726 wake_all_kswapd(order
, zonelist
, high_zoneidx
);
1729 * OK, we're below the kswapd watermark and have kicked background
1730 * reclaim. Now things get more complex, so set up alloc_flags according
1731 * to how we want to proceed.
1733 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
1736 /* This is the last chance, in general, before the goto nopage. */
1737 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
, zonelist
,
1738 high_zoneidx
, alloc_flags
& ~ALLOC_NO_WATERMARKS
,
1739 preferred_zone
, migratetype
);
1744 /* Allocate without watermarks if the context allows */
1745 if (alloc_flags
& ALLOC_NO_WATERMARKS
) {
1746 page
= __alloc_pages_high_priority(gfp_mask
, order
,
1747 zonelist
, high_zoneidx
, nodemask
,
1748 preferred_zone
, migratetype
);
1753 /* Atomic allocations - we can't balance anything */
1757 /* Avoid recursion of direct reclaim */
1758 if (p
->flags
& PF_MEMALLOC
)
1761 /* Try direct reclaim and then allocating */
1762 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
,
1763 zonelist
, high_zoneidx
,
1765 alloc_flags
, preferred_zone
,
1766 migratetype
, &did_some_progress
);
1771 * If we failed to make any progress reclaiming, then we are
1772 * running out of options and have to consider going OOM
1774 if (!did_some_progress
) {
1775 if ((gfp_mask
& __GFP_FS
) && !(gfp_mask
& __GFP_NORETRY
)) {
1776 page
= __alloc_pages_may_oom(gfp_mask
, order
,
1777 zonelist
, high_zoneidx
,
1778 nodemask
, preferred_zone
,
1784 * The OOM killer does not trigger for high-order allocations
1785 * but if no progress is being made, there are no other
1786 * options and retrying is unlikely to help
1788 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
1795 /* Check if we should retry the allocation */
1796 pages_reclaimed
+= did_some_progress
;
1797 if (should_alloc_retry(gfp_mask
, order
, pages_reclaimed
)) {
1798 /* Wait for some write requests to complete then retry */
1799 congestion_wait(WRITE
, HZ
/50);
1804 if (!(gfp_mask
& __GFP_NOWARN
) && printk_ratelimit()) {
1805 printk(KERN_WARNING
"%s: page allocation failure."
1806 " order:%d, mode:0x%x\n",
1807 p
->comm
, order
, gfp_mask
);
1817 * This is the 'heart' of the zoned buddy allocator.
1820 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
,
1821 struct zonelist
*zonelist
, nodemask_t
*nodemask
)
1823 enum zone_type high_zoneidx
= gfp_zone(gfp_mask
);
1824 struct zone
*preferred_zone
;
1826 int migratetype
= allocflags_to_migratetype(gfp_mask
);
1828 lockdep_trace_alloc(gfp_mask
);
1830 might_sleep_if(gfp_mask
& __GFP_WAIT
);
1832 if (should_fail_alloc_page(gfp_mask
, order
))
1836 * Check the zones suitable for the gfp_mask contain at least one
1837 * valid zone. It's possible to have an empty zonelist as a result
1838 * of GFP_THISNODE and a memoryless node
1840 if (unlikely(!zonelist
->_zonerefs
->zone
))
1843 /* The preferred zone is used for statistics later */
1844 first_zones_zonelist(zonelist
, high_zoneidx
, nodemask
, &preferred_zone
);
1845 if (!preferred_zone
)
1848 /* First allocation attempt */
1849 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
, order
,
1850 zonelist
, high_zoneidx
, ALLOC_WMARK_LOW
|ALLOC_CPUSET
,
1851 preferred_zone
, migratetype
);
1852 if (unlikely(!page
))
1853 page
= __alloc_pages_slowpath(gfp_mask
, order
,
1854 zonelist
, high_zoneidx
, nodemask
,
1855 preferred_zone
, migratetype
);
1859 EXPORT_SYMBOL(__alloc_pages_nodemask
);
1862 * Common helper functions.
1864 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
1867 page
= alloc_pages(gfp_mask
, order
);
1870 return (unsigned long) page_address(page
);
1873 EXPORT_SYMBOL(__get_free_pages
);
1875 unsigned long get_zeroed_page(gfp_t gfp_mask
)
1880 * get_zeroed_page() returns a 32-bit address, which cannot represent
1883 VM_BUG_ON((gfp_mask
& __GFP_HIGHMEM
) != 0);
1885 page
= alloc_pages(gfp_mask
| __GFP_ZERO
, 0);
1887 return (unsigned long) page_address(page
);
1891 EXPORT_SYMBOL(get_zeroed_page
);
1893 void __pagevec_free(struct pagevec
*pvec
)
1895 int i
= pagevec_count(pvec
);
1898 free_hot_cold_page(pvec
->pages
[i
], pvec
->cold
);
1901 void __free_pages(struct page
*page
, unsigned int order
)
1903 if (put_page_testzero(page
)) {
1905 free_hot_page(page
);
1907 __free_pages_ok(page
, order
);
1911 EXPORT_SYMBOL(__free_pages
);
1913 void free_pages(unsigned long addr
, unsigned int order
)
1916 VM_BUG_ON(!virt_addr_valid((void *)addr
));
1917 __free_pages(virt_to_page((void *)addr
), order
);
1921 EXPORT_SYMBOL(free_pages
);
1924 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
1925 * @size: the number of bytes to allocate
1926 * @gfp_mask: GFP flags for the allocation
1928 * This function is similar to alloc_pages(), except that it allocates the
1929 * minimum number of pages to satisfy the request. alloc_pages() can only
1930 * allocate memory in power-of-two pages.
1932 * This function is also limited by MAX_ORDER.
1934 * Memory allocated by this function must be released by free_pages_exact().
1936 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
1938 unsigned int order
= get_order(size
);
1941 addr
= __get_free_pages(gfp_mask
, order
);
1943 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
1944 unsigned long used
= addr
+ PAGE_ALIGN(size
);
1946 split_page(virt_to_page(addr
), order
);
1947 while (used
< alloc_end
) {
1953 return (void *)addr
;
1955 EXPORT_SYMBOL(alloc_pages_exact
);
1958 * free_pages_exact - release memory allocated via alloc_pages_exact()
1959 * @virt: the value returned by alloc_pages_exact.
1960 * @size: size of allocation, same value as passed to alloc_pages_exact().
1962 * Release the memory allocated by a previous call to alloc_pages_exact.
1964 void free_pages_exact(void *virt
, size_t size
)
1966 unsigned long addr
= (unsigned long)virt
;
1967 unsigned long end
= addr
+ PAGE_ALIGN(size
);
1969 while (addr
< end
) {
1974 EXPORT_SYMBOL(free_pages_exact
);
1976 static unsigned int nr_free_zone_pages(int offset
)
1981 /* Just pick one node, since fallback list is circular */
1982 unsigned int sum
= 0;
1984 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
1986 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
1987 unsigned long size
= zone
->present_pages
;
1988 unsigned long high
= high_wmark_pages(zone
);
1997 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
1999 unsigned int nr_free_buffer_pages(void)
2001 return nr_free_zone_pages(gfp_zone(GFP_USER
));
2003 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
2006 * Amount of free RAM allocatable within all zones
2008 unsigned int nr_free_pagecache_pages(void)
2010 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
2013 static inline void show_node(struct zone
*zone
)
2016 printk("Node %d ", zone_to_nid(zone
));
2019 void si_meminfo(struct sysinfo
*val
)
2021 val
->totalram
= totalram_pages
;
2023 val
->freeram
= global_page_state(NR_FREE_PAGES
);
2024 val
->bufferram
= nr_blockdev_pages();
2025 val
->totalhigh
= totalhigh_pages
;
2026 val
->freehigh
= nr_free_highpages();
2027 val
->mem_unit
= PAGE_SIZE
;
2030 EXPORT_SYMBOL(si_meminfo
);
2033 void si_meminfo_node(struct sysinfo
*val
, int nid
)
2035 pg_data_t
*pgdat
= NODE_DATA(nid
);
2037 val
->totalram
= pgdat
->node_present_pages
;
2038 val
->freeram
= node_page_state(nid
, NR_FREE_PAGES
);
2039 #ifdef CONFIG_HIGHMEM
2040 val
->totalhigh
= pgdat
->node_zones
[ZONE_HIGHMEM
].present_pages
;
2041 val
->freehigh
= zone_page_state(&pgdat
->node_zones
[ZONE_HIGHMEM
],
2047 val
->mem_unit
= PAGE_SIZE
;
2051 #define K(x) ((x) << (PAGE_SHIFT-10))
2054 * Show free area list (used inside shift_scroll-lock stuff)
2055 * We also calculate the percentage fragmentation. We do this by counting the
2056 * memory on each free list with the exception of the first item on the list.
2058 void show_free_areas(void)
2063 for_each_populated_zone(zone
) {
2065 printk("%s per-cpu:\n", zone
->name
);
2067 for_each_online_cpu(cpu
) {
2068 struct per_cpu_pageset
*pageset
;
2070 pageset
= zone_pcp(zone
, cpu
);
2072 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2073 cpu
, pageset
->pcp
.high
,
2074 pageset
->pcp
.batch
, pageset
->pcp
.count
);
2078 printk("Active_anon:%lu active_file:%lu inactive_anon:%lu\n"
2079 " inactive_file:%lu"
2080 //TODO: check/adjust line lengths
2081 #ifdef CONFIG_UNEVICTABLE_LRU
2084 " dirty:%lu writeback:%lu unstable:%lu\n"
2085 " free:%lu slab:%lu mapped:%lu pagetables:%lu bounce:%lu\n",
2086 global_page_state(NR_ACTIVE_ANON
),
2087 global_page_state(NR_ACTIVE_FILE
),
2088 global_page_state(NR_INACTIVE_ANON
),
2089 global_page_state(NR_INACTIVE_FILE
),
2090 #ifdef CONFIG_UNEVICTABLE_LRU
2091 global_page_state(NR_UNEVICTABLE
),
2093 global_page_state(NR_FILE_DIRTY
),
2094 global_page_state(NR_WRITEBACK
),
2095 global_page_state(NR_UNSTABLE_NFS
),
2096 global_page_state(NR_FREE_PAGES
),
2097 global_page_state(NR_SLAB_RECLAIMABLE
) +
2098 global_page_state(NR_SLAB_UNRECLAIMABLE
),
2099 global_page_state(NR_FILE_MAPPED
),
2100 global_page_state(NR_PAGETABLE
),
2101 global_page_state(NR_BOUNCE
));
2103 for_each_populated_zone(zone
) {
2112 " active_anon:%lukB"
2113 " inactive_anon:%lukB"
2114 " active_file:%lukB"
2115 " inactive_file:%lukB"
2116 #ifdef CONFIG_UNEVICTABLE_LRU
2117 " unevictable:%lukB"
2120 " pages_scanned:%lu"
2121 " all_unreclaimable? %s"
2124 K(zone_page_state(zone
, NR_FREE_PAGES
)),
2125 K(min_wmark_pages(zone
)),
2126 K(low_wmark_pages(zone
)),
2127 K(high_wmark_pages(zone
)),
2128 K(zone_page_state(zone
, NR_ACTIVE_ANON
)),
2129 K(zone_page_state(zone
, NR_INACTIVE_ANON
)),
2130 K(zone_page_state(zone
, NR_ACTIVE_FILE
)),
2131 K(zone_page_state(zone
, NR_INACTIVE_FILE
)),
2132 #ifdef CONFIG_UNEVICTABLE_LRU
2133 K(zone_page_state(zone
, NR_UNEVICTABLE
)),
2135 K(zone
->present_pages
),
2136 zone
->pages_scanned
,
2137 (zone_is_all_unreclaimable(zone
) ? "yes" : "no")
2139 printk("lowmem_reserve[]:");
2140 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
2141 printk(" %lu", zone
->lowmem_reserve
[i
]);
2145 for_each_populated_zone(zone
) {
2146 unsigned long nr
[MAX_ORDER
], flags
, order
, total
= 0;
2149 printk("%s: ", zone
->name
);
2151 spin_lock_irqsave(&zone
->lock
, flags
);
2152 for (order
= 0; order
< MAX_ORDER
; order
++) {
2153 nr
[order
] = zone
->free_area
[order
].nr_free
;
2154 total
+= nr
[order
] << order
;
2156 spin_unlock_irqrestore(&zone
->lock
, flags
);
2157 for (order
= 0; order
< MAX_ORDER
; order
++)
2158 printk("%lu*%lukB ", nr
[order
], K(1UL) << order
);
2159 printk("= %lukB\n", K(total
));
2162 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES
));
2164 show_swap_cache_info();
2167 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
2169 zoneref
->zone
= zone
;
2170 zoneref
->zone_idx
= zone_idx(zone
);
2174 * Builds allocation fallback zone lists.
2176 * Add all populated zones of a node to the zonelist.
2178 static int build_zonelists_node(pg_data_t
*pgdat
, struct zonelist
*zonelist
,
2179 int nr_zones
, enum zone_type zone_type
)
2183 BUG_ON(zone_type
>= MAX_NR_ZONES
);
2188 zone
= pgdat
->node_zones
+ zone_type
;
2189 if (populated_zone(zone
)) {
2190 zoneref_set_zone(zone
,
2191 &zonelist
->_zonerefs
[nr_zones
++]);
2192 check_highest_zone(zone_type
);
2195 } while (zone_type
);
2202 * 0 = automatic detection of better ordering.
2203 * 1 = order by ([node] distance, -zonetype)
2204 * 2 = order by (-zonetype, [node] distance)
2206 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
2207 * the same zonelist. So only NUMA can configure this param.
2209 #define ZONELIST_ORDER_DEFAULT 0
2210 #define ZONELIST_ORDER_NODE 1
2211 #define ZONELIST_ORDER_ZONE 2
2213 /* zonelist order in the kernel.
2214 * set_zonelist_order() will set this to NODE or ZONE.
2216 static int current_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
2217 static char zonelist_order_name
[3][8] = {"Default", "Node", "Zone"};
2221 /* The value user specified ....changed by config */
2222 static int user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
2223 /* string for sysctl */
2224 #define NUMA_ZONELIST_ORDER_LEN 16
2225 char numa_zonelist_order
[16] = "default";
2228 * interface for configure zonelist ordering.
2229 * command line option "numa_zonelist_order"
2230 * = "[dD]efault - default, automatic configuration.
2231 * = "[nN]ode - order by node locality, then by zone within node
2232 * = "[zZ]one - order by zone, then by locality within zone
2235 static int __parse_numa_zonelist_order(char *s
)
2237 if (*s
== 'd' || *s
== 'D') {
2238 user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
2239 } else if (*s
== 'n' || *s
== 'N') {
2240 user_zonelist_order
= ZONELIST_ORDER_NODE
;
2241 } else if (*s
== 'z' || *s
== 'Z') {
2242 user_zonelist_order
= ZONELIST_ORDER_ZONE
;
2245 "Ignoring invalid numa_zonelist_order value: "
2252 static __init
int setup_numa_zonelist_order(char *s
)
2255 return __parse_numa_zonelist_order(s
);
2258 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
2261 * sysctl handler for numa_zonelist_order
2263 int numa_zonelist_order_handler(ctl_table
*table
, int write
,
2264 struct file
*file
, void __user
*buffer
, size_t *length
,
2267 char saved_string
[NUMA_ZONELIST_ORDER_LEN
];
2271 strncpy(saved_string
, (char*)table
->data
,
2272 NUMA_ZONELIST_ORDER_LEN
);
2273 ret
= proc_dostring(table
, write
, file
, buffer
, length
, ppos
);
2277 int oldval
= user_zonelist_order
;
2278 if (__parse_numa_zonelist_order((char*)table
->data
)) {
2280 * bogus value. restore saved string
2282 strncpy((char*)table
->data
, saved_string
,
2283 NUMA_ZONELIST_ORDER_LEN
);
2284 user_zonelist_order
= oldval
;
2285 } else if (oldval
!= user_zonelist_order
)
2286 build_all_zonelists();
2292 #define MAX_NODE_LOAD (nr_online_nodes)
2293 static int node_load
[MAX_NUMNODES
];
2296 * find_next_best_node - find the next node that should appear in a given node's fallback list
2297 * @node: node whose fallback list we're appending
2298 * @used_node_mask: nodemask_t of already used nodes
2300 * We use a number of factors to determine which is the next node that should
2301 * appear on a given node's fallback list. The node should not have appeared
2302 * already in @node's fallback list, and it should be the next closest node
2303 * according to the distance array (which contains arbitrary distance values
2304 * from each node to each node in the system), and should also prefer nodes
2305 * with no CPUs, since presumably they'll have very little allocation pressure
2306 * on them otherwise.
2307 * It returns -1 if no node is found.
2309 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
2312 int min_val
= INT_MAX
;
2314 const struct cpumask
*tmp
= cpumask_of_node(0);
2316 /* Use the local node if we haven't already */
2317 if (!node_isset(node
, *used_node_mask
)) {
2318 node_set(node
, *used_node_mask
);
2322 for_each_node_state(n
, N_HIGH_MEMORY
) {
2324 /* Don't want a node to appear more than once */
2325 if (node_isset(n
, *used_node_mask
))
2328 /* Use the distance array to find the distance */
2329 val
= node_distance(node
, n
);
2331 /* Penalize nodes under us ("prefer the next node") */
2334 /* Give preference to headless and unused nodes */
2335 tmp
= cpumask_of_node(n
);
2336 if (!cpumask_empty(tmp
))
2337 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
2339 /* Slight preference for less loaded node */
2340 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
2341 val
+= node_load
[n
];
2343 if (val
< min_val
) {
2350 node_set(best_node
, *used_node_mask
);
2357 * Build zonelists ordered by node and zones within node.
2358 * This results in maximum locality--normal zone overflows into local
2359 * DMA zone, if any--but risks exhausting DMA zone.
2361 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int node
)
2364 struct zonelist
*zonelist
;
2366 zonelist
= &pgdat
->node_zonelists
[0];
2367 for (j
= 0; zonelist
->_zonerefs
[j
].zone
!= NULL
; j
++)
2369 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
2371 zonelist
->_zonerefs
[j
].zone
= NULL
;
2372 zonelist
->_zonerefs
[j
].zone_idx
= 0;
2376 * Build gfp_thisnode zonelists
2378 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
2381 struct zonelist
*zonelist
;
2383 zonelist
= &pgdat
->node_zonelists
[1];
2384 j
= build_zonelists_node(pgdat
, zonelist
, 0, MAX_NR_ZONES
- 1);
2385 zonelist
->_zonerefs
[j
].zone
= NULL
;
2386 zonelist
->_zonerefs
[j
].zone_idx
= 0;
2390 * Build zonelists ordered by zone and nodes within zones.
2391 * This results in conserving DMA zone[s] until all Normal memory is
2392 * exhausted, but results in overflowing to remote node while memory
2393 * may still exist in local DMA zone.
2395 static int node_order
[MAX_NUMNODES
];
2397 static void build_zonelists_in_zone_order(pg_data_t
*pgdat
, int nr_nodes
)
2400 int zone_type
; /* needs to be signed */
2402 struct zonelist
*zonelist
;
2404 zonelist
= &pgdat
->node_zonelists
[0];
2406 for (zone_type
= MAX_NR_ZONES
- 1; zone_type
>= 0; zone_type
--) {
2407 for (j
= 0; j
< nr_nodes
; j
++) {
2408 node
= node_order
[j
];
2409 z
= &NODE_DATA(node
)->node_zones
[zone_type
];
2410 if (populated_zone(z
)) {
2412 &zonelist
->_zonerefs
[pos
++]);
2413 check_highest_zone(zone_type
);
2417 zonelist
->_zonerefs
[pos
].zone
= NULL
;
2418 zonelist
->_zonerefs
[pos
].zone_idx
= 0;
2421 static int default_zonelist_order(void)
2424 unsigned long low_kmem_size
,total_size
;
2428 * ZONE_DMA and ZONE_DMA32 can be very small area in the sytem.
2429 * If they are really small and used heavily, the system can fall
2430 * into OOM very easily.
2431 * This function detect ZONE_DMA/DMA32 size and confgigures zone order.
2433 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
2436 for_each_online_node(nid
) {
2437 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
2438 z
= &NODE_DATA(nid
)->node_zones
[zone_type
];
2439 if (populated_zone(z
)) {
2440 if (zone_type
< ZONE_NORMAL
)
2441 low_kmem_size
+= z
->present_pages
;
2442 total_size
+= z
->present_pages
;
2446 if (!low_kmem_size
|| /* there are no DMA area. */
2447 low_kmem_size
> total_size
/2) /* DMA/DMA32 is big. */
2448 return ZONELIST_ORDER_NODE
;
2450 * look into each node's config.
2451 * If there is a node whose DMA/DMA32 memory is very big area on
2452 * local memory, NODE_ORDER may be suitable.
2454 average_size
= total_size
/
2455 (nodes_weight(node_states
[N_HIGH_MEMORY
]) + 1);
2456 for_each_online_node(nid
) {
2459 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
2460 z
= &NODE_DATA(nid
)->node_zones
[zone_type
];
2461 if (populated_zone(z
)) {
2462 if (zone_type
< ZONE_NORMAL
)
2463 low_kmem_size
+= z
->present_pages
;
2464 total_size
+= z
->present_pages
;
2467 if (low_kmem_size
&&
2468 total_size
> average_size
&& /* ignore small node */
2469 low_kmem_size
> total_size
* 70/100)
2470 return ZONELIST_ORDER_NODE
;
2472 return ZONELIST_ORDER_ZONE
;
2475 static void set_zonelist_order(void)
2477 if (user_zonelist_order
== ZONELIST_ORDER_DEFAULT
)
2478 current_zonelist_order
= default_zonelist_order();
2480 current_zonelist_order
= user_zonelist_order
;
2483 static void build_zonelists(pg_data_t
*pgdat
)
2487 nodemask_t used_mask
;
2488 int local_node
, prev_node
;
2489 struct zonelist
*zonelist
;
2490 int order
= current_zonelist_order
;
2492 /* initialize zonelists */
2493 for (i
= 0; i
< MAX_ZONELISTS
; i
++) {
2494 zonelist
= pgdat
->node_zonelists
+ i
;
2495 zonelist
->_zonerefs
[0].zone
= NULL
;
2496 zonelist
->_zonerefs
[0].zone_idx
= 0;
2499 /* NUMA-aware ordering of nodes */
2500 local_node
= pgdat
->node_id
;
2501 load
= nr_online_nodes
;
2502 prev_node
= local_node
;
2503 nodes_clear(used_mask
);
2505 memset(node_load
, 0, sizeof(node_load
));
2506 memset(node_order
, 0, sizeof(node_order
));
2509 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
2510 int distance
= node_distance(local_node
, node
);
2513 * If another node is sufficiently far away then it is better
2514 * to reclaim pages in a zone before going off node.
2516 if (distance
> RECLAIM_DISTANCE
)
2517 zone_reclaim_mode
= 1;
2520 * We don't want to pressure a particular node.
2521 * So adding penalty to the first node in same
2522 * distance group to make it round-robin.
2524 if (distance
!= node_distance(local_node
, prev_node
))
2525 node_load
[node
] = load
;
2529 if (order
== ZONELIST_ORDER_NODE
)
2530 build_zonelists_in_node_order(pgdat
, node
);
2532 node_order
[j
++] = node
; /* remember order */
2535 if (order
== ZONELIST_ORDER_ZONE
) {
2536 /* calculate node order -- i.e., DMA last! */
2537 build_zonelists_in_zone_order(pgdat
, j
);
2540 build_thisnode_zonelists(pgdat
);
2543 /* Construct the zonelist performance cache - see further mmzone.h */
2544 static void build_zonelist_cache(pg_data_t
*pgdat
)
2546 struct zonelist
*zonelist
;
2547 struct zonelist_cache
*zlc
;
2550 zonelist
= &pgdat
->node_zonelists
[0];
2551 zonelist
->zlcache_ptr
= zlc
= &zonelist
->zlcache
;
2552 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
2553 for (z
= zonelist
->_zonerefs
; z
->zone
; z
++)
2554 zlc
->z_to_n
[z
- zonelist
->_zonerefs
] = zonelist_node_idx(z
);
2558 #else /* CONFIG_NUMA */
2560 static void set_zonelist_order(void)
2562 current_zonelist_order
= ZONELIST_ORDER_ZONE
;
2565 static void build_zonelists(pg_data_t
*pgdat
)
2567 int node
, local_node
;
2569 struct zonelist
*zonelist
;
2571 local_node
= pgdat
->node_id
;
2573 zonelist
= &pgdat
->node_zonelists
[0];
2574 j
= build_zonelists_node(pgdat
, zonelist
, 0, MAX_NR_ZONES
- 1);
2577 * Now we build the zonelist so that it contains the zones
2578 * of all the other nodes.
2579 * We don't want to pressure a particular node, so when
2580 * building the zones for node N, we make sure that the
2581 * zones coming right after the local ones are those from
2582 * node N+1 (modulo N)
2584 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
2585 if (!node_online(node
))
2587 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
2590 for (node
= 0; node
< local_node
; node
++) {
2591 if (!node_online(node
))
2593 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
2597 zonelist
->_zonerefs
[j
].zone
= NULL
;
2598 zonelist
->_zonerefs
[j
].zone_idx
= 0;
2601 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
2602 static void build_zonelist_cache(pg_data_t
*pgdat
)
2604 pgdat
->node_zonelists
[0].zlcache_ptr
= NULL
;
2607 #endif /* CONFIG_NUMA */
2609 /* return values int ....just for stop_machine() */
2610 static int __build_all_zonelists(void *dummy
)
2614 for_each_online_node(nid
) {
2615 pg_data_t
*pgdat
= NODE_DATA(nid
);
2617 build_zonelists(pgdat
);
2618 build_zonelist_cache(pgdat
);
2623 void build_all_zonelists(void)
2625 set_zonelist_order();
2627 if (system_state
== SYSTEM_BOOTING
) {
2628 __build_all_zonelists(NULL
);
2629 mminit_verify_zonelist();
2630 cpuset_init_current_mems_allowed();
2632 /* we have to stop all cpus to guarantee there is no user
2634 stop_machine(__build_all_zonelists
, NULL
, NULL
);
2635 /* cpuset refresh routine should be here */
2637 vm_total_pages
= nr_free_pagecache_pages();
2639 * Disable grouping by mobility if the number of pages in the
2640 * system is too low to allow the mechanism to work. It would be
2641 * more accurate, but expensive to check per-zone. This check is
2642 * made on memory-hotadd so a system can start with mobility
2643 * disabled and enable it later
2645 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
2646 page_group_by_mobility_disabled
= 1;
2648 page_group_by_mobility_disabled
= 0;
2650 printk("Built %i zonelists in %s order, mobility grouping %s. "
2651 "Total pages: %ld\n",
2653 zonelist_order_name
[current_zonelist_order
],
2654 page_group_by_mobility_disabled
? "off" : "on",
2657 printk("Policy zone: %s\n", zone_names
[policy_zone
]);
2662 * Helper functions to size the waitqueue hash table.
2663 * Essentially these want to choose hash table sizes sufficiently
2664 * large so that collisions trying to wait on pages are rare.
2665 * But in fact, the number of active page waitqueues on typical
2666 * systems is ridiculously low, less than 200. So this is even
2667 * conservative, even though it seems large.
2669 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
2670 * waitqueues, i.e. the size of the waitq table given the number of pages.
2672 #define PAGES_PER_WAITQUEUE 256
2674 #ifndef CONFIG_MEMORY_HOTPLUG
2675 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
2677 unsigned long size
= 1;
2679 pages
/= PAGES_PER_WAITQUEUE
;
2681 while (size
< pages
)
2685 * Once we have dozens or even hundreds of threads sleeping
2686 * on IO we've got bigger problems than wait queue collision.
2687 * Limit the size of the wait table to a reasonable size.
2689 size
= min(size
, 4096UL);
2691 return max(size
, 4UL);
2695 * A zone's size might be changed by hot-add, so it is not possible to determine
2696 * a suitable size for its wait_table. So we use the maximum size now.
2698 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
2700 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
2701 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
2702 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
2704 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
2705 * or more by the traditional way. (See above). It equals:
2707 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
2708 * ia64(16K page size) : = ( 8G + 4M)byte.
2709 * powerpc (64K page size) : = (32G +16M)byte.
2711 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
2718 * This is an integer logarithm so that shifts can be used later
2719 * to extract the more random high bits from the multiplicative
2720 * hash function before the remainder is taken.
2722 static inline unsigned long wait_table_bits(unsigned long size
)
2727 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
2730 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
2731 * of blocks reserved is based on min_wmark_pages(zone). The memory within
2732 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
2733 * higher will lead to a bigger reserve which will get freed as contiguous
2734 * blocks as reclaim kicks in
2736 static void setup_zone_migrate_reserve(struct zone
*zone
)
2738 unsigned long start_pfn
, pfn
, end_pfn
;
2740 unsigned long reserve
, block_migratetype
;
2742 /* Get the start pfn, end pfn and the number of blocks to reserve */
2743 start_pfn
= zone
->zone_start_pfn
;
2744 end_pfn
= start_pfn
+ zone
->spanned_pages
;
2745 reserve
= roundup(min_wmark_pages(zone
), pageblock_nr_pages
) >>
2748 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= pageblock_nr_pages
) {
2749 if (!pfn_valid(pfn
))
2751 page
= pfn_to_page(pfn
);
2753 /* Watch out for overlapping nodes */
2754 if (page_to_nid(page
) != zone_to_nid(zone
))
2757 /* Blocks with reserved pages will never free, skip them. */
2758 if (PageReserved(page
))
2761 block_migratetype
= get_pageblock_migratetype(page
);
2763 /* If this block is reserved, account for it */
2764 if (reserve
> 0 && block_migratetype
== MIGRATE_RESERVE
) {
2769 /* Suitable for reserving if this block is movable */
2770 if (reserve
> 0 && block_migratetype
== MIGRATE_MOVABLE
) {
2771 set_pageblock_migratetype(page
, MIGRATE_RESERVE
);
2772 move_freepages_block(zone
, page
, MIGRATE_RESERVE
);
2778 * If the reserve is met and this is a previous reserved block,
2781 if (block_migratetype
== MIGRATE_RESERVE
) {
2782 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
2783 move_freepages_block(zone
, page
, MIGRATE_MOVABLE
);
2789 * Initially all pages are reserved - free ones are freed
2790 * up by free_all_bootmem() once the early boot process is
2791 * done. Non-atomic initialization, single-pass.
2793 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
2794 unsigned long start_pfn
, enum memmap_context context
)
2797 unsigned long end_pfn
= start_pfn
+ size
;
2801 if (highest_memmap_pfn
< end_pfn
- 1)
2802 highest_memmap_pfn
= end_pfn
- 1;
2804 z
= &NODE_DATA(nid
)->node_zones
[zone
];
2805 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
2807 * There can be holes in boot-time mem_map[]s
2808 * handed to this function. They do not
2809 * exist on hotplugged memory.
2811 if (context
== MEMMAP_EARLY
) {
2812 if (!early_pfn_valid(pfn
))
2814 if (!early_pfn_in_nid(pfn
, nid
))
2817 page
= pfn_to_page(pfn
);
2818 set_page_links(page
, zone
, nid
, pfn
);
2819 mminit_verify_page_links(page
, zone
, nid
, pfn
);
2820 init_page_count(page
);
2821 reset_page_mapcount(page
);
2822 SetPageReserved(page
);
2824 * Mark the block movable so that blocks are reserved for
2825 * movable at startup. This will force kernel allocations
2826 * to reserve their blocks rather than leaking throughout
2827 * the address space during boot when many long-lived
2828 * kernel allocations are made. Later some blocks near
2829 * the start are marked MIGRATE_RESERVE by
2830 * setup_zone_migrate_reserve()
2832 * bitmap is created for zone's valid pfn range. but memmap
2833 * can be created for invalid pages (for alignment)
2834 * check here not to call set_pageblock_migratetype() against
2837 if ((z
->zone_start_pfn
<= pfn
)
2838 && (pfn
< z
->zone_start_pfn
+ z
->spanned_pages
)
2839 && !(pfn
& (pageblock_nr_pages
- 1)))
2840 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
2842 INIT_LIST_HEAD(&page
->lru
);
2843 #ifdef WANT_PAGE_VIRTUAL
2844 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
2845 if (!is_highmem_idx(zone
))
2846 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
2851 static void __meminit
zone_init_free_lists(struct zone
*zone
)
2854 for_each_migratetype_order(order
, t
) {
2855 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
2856 zone
->free_area
[order
].nr_free
= 0;
2860 #ifndef __HAVE_ARCH_MEMMAP_INIT
2861 #define memmap_init(size, nid, zone, start_pfn) \
2862 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
2865 static int zone_batchsize(struct zone
*zone
)
2871 * The per-cpu-pages pools are set to around 1000th of the
2872 * size of the zone. But no more than 1/2 of a meg.
2874 * OK, so we don't know how big the cache is. So guess.
2876 batch
= zone
->present_pages
/ 1024;
2877 if (batch
* PAGE_SIZE
> 512 * 1024)
2878 batch
= (512 * 1024) / PAGE_SIZE
;
2879 batch
/= 4; /* We effectively *= 4 below */
2884 * Clamp the batch to a 2^n - 1 value. Having a power
2885 * of 2 value was found to be more likely to have
2886 * suboptimal cache aliasing properties in some cases.
2888 * For example if 2 tasks are alternately allocating
2889 * batches of pages, one task can end up with a lot
2890 * of pages of one half of the possible page colors
2891 * and the other with pages of the other colors.
2893 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
2898 /* The deferral and batching of frees should be suppressed under NOMMU
2901 * The problem is that NOMMU needs to be able to allocate large chunks
2902 * of contiguous memory as there's no hardware page translation to
2903 * assemble apparent contiguous memory from discontiguous pages.
2905 * Queueing large contiguous runs of pages for batching, however,
2906 * causes the pages to actually be freed in smaller chunks. As there
2907 * can be a significant delay between the individual batches being
2908 * recycled, this leads to the once large chunks of space being
2909 * fragmented and becoming unavailable for high-order allocations.
2915 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
2917 struct per_cpu_pages
*pcp
;
2919 memset(p
, 0, sizeof(*p
));
2923 pcp
->high
= 6 * batch
;
2924 pcp
->batch
= max(1UL, 1 * batch
);
2925 INIT_LIST_HEAD(&pcp
->list
);
2929 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
2930 * to the value high for the pageset p.
2933 static void setup_pagelist_highmark(struct per_cpu_pageset
*p
,
2936 struct per_cpu_pages
*pcp
;
2940 pcp
->batch
= max(1UL, high
/4);
2941 if ((high
/4) > (PAGE_SHIFT
* 8))
2942 pcp
->batch
= PAGE_SHIFT
* 8;
2948 * Boot pageset table. One per cpu which is going to be used for all
2949 * zones and all nodes. The parameters will be set in such a way
2950 * that an item put on a list will immediately be handed over to
2951 * the buddy list. This is safe since pageset manipulation is done
2952 * with interrupts disabled.
2954 * Some NUMA counter updates may also be caught by the boot pagesets.
2956 * The boot_pagesets must be kept even after bootup is complete for
2957 * unused processors and/or zones. They do play a role for bootstrapping
2958 * hotplugged processors.
2960 * zoneinfo_show() and maybe other functions do
2961 * not check if the processor is online before following the pageset pointer.
2962 * Other parts of the kernel may not check if the zone is available.
2964 static struct per_cpu_pageset boot_pageset
[NR_CPUS
];
2967 * Dynamically allocate memory for the
2968 * per cpu pageset array in struct zone.
2970 static int __cpuinit
process_zones(int cpu
)
2972 struct zone
*zone
, *dzone
;
2973 int node
= cpu_to_node(cpu
);
2975 node_set_state(node
, N_CPU
); /* this node has a cpu */
2977 for_each_populated_zone(zone
) {
2978 zone_pcp(zone
, cpu
) = kmalloc_node(sizeof(struct per_cpu_pageset
),
2980 if (!zone_pcp(zone
, cpu
))
2983 setup_pageset(zone_pcp(zone
, cpu
), zone_batchsize(zone
));
2985 if (percpu_pagelist_fraction
)
2986 setup_pagelist_highmark(zone_pcp(zone
, cpu
),
2987 (zone
->present_pages
/ percpu_pagelist_fraction
));
2992 for_each_zone(dzone
) {
2993 if (!populated_zone(dzone
))
2997 kfree(zone_pcp(dzone
, cpu
));
2998 zone_pcp(dzone
, cpu
) = NULL
;
3003 static inline void free_zone_pagesets(int cpu
)
3007 for_each_zone(zone
) {
3008 struct per_cpu_pageset
*pset
= zone_pcp(zone
, cpu
);
3010 /* Free per_cpu_pageset if it is slab allocated */
3011 if (pset
!= &boot_pageset
[cpu
])
3013 zone_pcp(zone
, cpu
) = NULL
;
3017 static int __cpuinit
pageset_cpuup_callback(struct notifier_block
*nfb
,
3018 unsigned long action
,
3021 int cpu
= (long)hcpu
;
3022 int ret
= NOTIFY_OK
;
3025 case CPU_UP_PREPARE
:
3026 case CPU_UP_PREPARE_FROZEN
:
3027 if (process_zones(cpu
))
3030 case CPU_UP_CANCELED
:
3031 case CPU_UP_CANCELED_FROZEN
:
3033 case CPU_DEAD_FROZEN
:
3034 free_zone_pagesets(cpu
);
3042 static struct notifier_block __cpuinitdata pageset_notifier
=
3043 { &pageset_cpuup_callback
, NULL
, 0 };
3045 void __init
setup_per_cpu_pageset(void)
3049 /* Initialize per_cpu_pageset for cpu 0.
3050 * A cpuup callback will do this for every cpu
3051 * as it comes online
3053 err
= process_zones(smp_processor_id());
3055 register_cpu_notifier(&pageset_notifier
);
3060 static noinline __init_refok
3061 int zone_wait_table_init(struct zone
*zone
, unsigned long zone_size_pages
)
3064 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
3068 * The per-page waitqueue mechanism uses hashed waitqueues
3071 zone
->wait_table_hash_nr_entries
=
3072 wait_table_hash_nr_entries(zone_size_pages
);
3073 zone
->wait_table_bits
=
3074 wait_table_bits(zone
->wait_table_hash_nr_entries
);
3075 alloc_size
= zone
->wait_table_hash_nr_entries
3076 * sizeof(wait_queue_head_t
);
3078 if (!slab_is_available()) {
3079 zone
->wait_table
= (wait_queue_head_t
*)
3080 alloc_bootmem_node(pgdat
, alloc_size
);
3083 * This case means that a zone whose size was 0 gets new memory
3084 * via memory hot-add.
3085 * But it may be the case that a new node was hot-added. In
3086 * this case vmalloc() will not be able to use this new node's
3087 * memory - this wait_table must be initialized to use this new
3088 * node itself as well.
3089 * To use this new node's memory, further consideration will be
3092 zone
->wait_table
= vmalloc(alloc_size
);
3094 if (!zone
->wait_table
)
3097 for(i
= 0; i
< zone
->wait_table_hash_nr_entries
; ++i
)
3098 init_waitqueue_head(zone
->wait_table
+ i
);
3103 static __meminit
void zone_pcp_init(struct zone
*zone
)
3106 unsigned long batch
= zone_batchsize(zone
);
3108 for (cpu
= 0; cpu
< NR_CPUS
; cpu
++) {
3110 /* Early boot. Slab allocator not functional yet */
3111 zone_pcp(zone
, cpu
) = &boot_pageset
[cpu
];
3112 setup_pageset(&boot_pageset
[cpu
],0);
3114 setup_pageset(zone_pcp(zone
,cpu
), batch
);
3117 if (zone
->present_pages
)
3118 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%lu\n",
3119 zone
->name
, zone
->present_pages
, batch
);
3122 __meminit
int init_currently_empty_zone(struct zone
*zone
,
3123 unsigned long zone_start_pfn
,
3125 enum memmap_context context
)
3127 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
3129 ret
= zone_wait_table_init(zone
, size
);
3132 pgdat
->nr_zones
= zone_idx(zone
) + 1;
3134 zone
->zone_start_pfn
= zone_start_pfn
;
3136 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
3137 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
3139 (unsigned long)zone_idx(zone
),
3140 zone_start_pfn
, (zone_start_pfn
+ size
));
3142 zone_init_free_lists(zone
);
3147 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3149 * Basic iterator support. Return the first range of PFNs for a node
3150 * Note: nid == MAX_NUMNODES returns first region regardless of node
3152 static int __meminit
first_active_region_index_in_nid(int nid
)
3156 for (i
= 0; i
< nr_nodemap_entries
; i
++)
3157 if (nid
== MAX_NUMNODES
|| early_node_map
[i
].nid
== nid
)
3164 * Basic iterator support. Return the next active range of PFNs for a node
3165 * Note: nid == MAX_NUMNODES returns next region regardless of node
3167 static int __meminit
next_active_region_index_in_nid(int index
, int nid
)
3169 for (index
= index
+ 1; index
< nr_nodemap_entries
; index
++)
3170 if (nid
== MAX_NUMNODES
|| early_node_map
[index
].nid
== nid
)
3176 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
3178 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
3179 * Architectures may implement their own version but if add_active_range()
3180 * was used and there are no special requirements, this is a convenient
3183 int __meminit
__early_pfn_to_nid(unsigned long pfn
)
3187 for (i
= 0; i
< nr_nodemap_entries
; i
++) {
3188 unsigned long start_pfn
= early_node_map
[i
].start_pfn
;
3189 unsigned long end_pfn
= early_node_map
[i
].end_pfn
;
3191 if (start_pfn
<= pfn
&& pfn
< end_pfn
)
3192 return early_node_map
[i
].nid
;
3194 /* This is a memory hole */
3197 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
3199 int __meminit
early_pfn_to_nid(unsigned long pfn
)
3203 nid
= __early_pfn_to_nid(pfn
);
3206 /* just returns 0 */
3210 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
3211 bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
3215 nid
= __early_pfn_to_nid(pfn
);
3216 if (nid
>= 0 && nid
!= node
)
3222 /* Basic iterator support to walk early_node_map[] */
3223 #define for_each_active_range_index_in_nid(i, nid) \
3224 for (i = first_active_region_index_in_nid(nid); i != -1; \
3225 i = next_active_region_index_in_nid(i, nid))
3228 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
3229 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
3230 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
3232 * If an architecture guarantees that all ranges registered with
3233 * add_active_ranges() contain no holes and may be freed, this
3234 * this function may be used instead of calling free_bootmem() manually.
3236 void __init
free_bootmem_with_active_regions(int nid
,
3237 unsigned long max_low_pfn
)
3241 for_each_active_range_index_in_nid(i
, nid
) {
3242 unsigned long size_pages
= 0;
3243 unsigned long end_pfn
= early_node_map
[i
].end_pfn
;
3245 if (early_node_map
[i
].start_pfn
>= max_low_pfn
)
3248 if (end_pfn
> max_low_pfn
)
3249 end_pfn
= max_low_pfn
;
3251 size_pages
= end_pfn
- early_node_map
[i
].start_pfn
;
3252 free_bootmem_node(NODE_DATA(early_node_map
[i
].nid
),
3253 PFN_PHYS(early_node_map
[i
].start_pfn
),
3254 size_pages
<< PAGE_SHIFT
);
3258 void __init
work_with_active_regions(int nid
, work_fn_t work_fn
, void *data
)
3263 for_each_active_range_index_in_nid(i
, nid
) {
3264 ret
= work_fn(early_node_map
[i
].start_pfn
,
3265 early_node_map
[i
].end_pfn
, data
);
3271 * sparse_memory_present_with_active_regions - Call memory_present for each active range
3272 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
3274 * If an architecture guarantees that all ranges registered with
3275 * add_active_ranges() contain no holes and may be freed, this
3276 * function may be used instead of calling memory_present() manually.
3278 void __init
sparse_memory_present_with_active_regions(int nid
)
3282 for_each_active_range_index_in_nid(i
, nid
)
3283 memory_present(early_node_map
[i
].nid
,
3284 early_node_map
[i
].start_pfn
,
3285 early_node_map
[i
].end_pfn
);
3289 * get_pfn_range_for_nid - Return the start and end page frames for a node
3290 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
3291 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
3292 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
3294 * It returns the start and end page frame of a node based on information
3295 * provided by an arch calling add_active_range(). If called for a node
3296 * with no available memory, a warning is printed and the start and end
3299 void __meminit
get_pfn_range_for_nid(unsigned int nid
,
3300 unsigned long *start_pfn
, unsigned long *end_pfn
)
3306 for_each_active_range_index_in_nid(i
, nid
) {
3307 *start_pfn
= min(*start_pfn
, early_node_map
[i
].start_pfn
);
3308 *end_pfn
= max(*end_pfn
, early_node_map
[i
].end_pfn
);
3311 if (*start_pfn
== -1UL)
3316 * This finds a zone that can be used for ZONE_MOVABLE pages. The
3317 * assumption is made that zones within a node are ordered in monotonic
3318 * increasing memory addresses so that the "highest" populated zone is used
3320 static void __init
find_usable_zone_for_movable(void)
3323 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
3324 if (zone_index
== ZONE_MOVABLE
)
3327 if (arch_zone_highest_possible_pfn
[zone_index
] >
3328 arch_zone_lowest_possible_pfn
[zone_index
])
3332 VM_BUG_ON(zone_index
== -1);
3333 movable_zone
= zone_index
;
3337 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
3338 * because it is sized independant of architecture. Unlike the other zones,
3339 * the starting point for ZONE_MOVABLE is not fixed. It may be different
3340 * in each node depending on the size of each node and how evenly kernelcore
3341 * is distributed. This helper function adjusts the zone ranges
3342 * provided by the architecture for a given node by using the end of the
3343 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
3344 * zones within a node are in order of monotonic increases memory addresses
3346 static void __meminit
adjust_zone_range_for_zone_movable(int nid
,
3347 unsigned long zone_type
,
3348 unsigned long node_start_pfn
,
3349 unsigned long node_end_pfn
,
3350 unsigned long *zone_start_pfn
,
3351 unsigned long *zone_end_pfn
)
3353 /* Only adjust if ZONE_MOVABLE is on this node */
3354 if (zone_movable_pfn
[nid
]) {
3355 /* Size ZONE_MOVABLE */
3356 if (zone_type
== ZONE_MOVABLE
) {
3357 *zone_start_pfn
= zone_movable_pfn
[nid
];
3358 *zone_end_pfn
= min(node_end_pfn
,
3359 arch_zone_highest_possible_pfn
[movable_zone
]);
3361 /* Adjust for ZONE_MOVABLE starting within this range */
3362 } else if (*zone_start_pfn
< zone_movable_pfn
[nid
] &&
3363 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
3364 *zone_end_pfn
= zone_movable_pfn
[nid
];
3366 /* Check if this whole range is within ZONE_MOVABLE */
3367 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
3368 *zone_start_pfn
= *zone_end_pfn
;
3373 * Return the number of pages a zone spans in a node, including holes
3374 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
3376 static unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
3377 unsigned long zone_type
,
3378 unsigned long *ignored
)
3380 unsigned long node_start_pfn
, node_end_pfn
;
3381 unsigned long zone_start_pfn
, zone_end_pfn
;
3383 /* Get the start and end of the node and zone */
3384 get_pfn_range_for_nid(nid
, &node_start_pfn
, &node_end_pfn
);
3385 zone_start_pfn
= arch_zone_lowest_possible_pfn
[zone_type
];
3386 zone_end_pfn
= arch_zone_highest_possible_pfn
[zone_type
];
3387 adjust_zone_range_for_zone_movable(nid
, zone_type
,
3388 node_start_pfn
, node_end_pfn
,
3389 &zone_start_pfn
, &zone_end_pfn
);
3391 /* Check that this node has pages within the zone's required range */
3392 if (zone_end_pfn
< node_start_pfn
|| zone_start_pfn
> node_end_pfn
)
3395 /* Move the zone boundaries inside the node if necessary */
3396 zone_end_pfn
= min(zone_end_pfn
, node_end_pfn
);
3397 zone_start_pfn
= max(zone_start_pfn
, node_start_pfn
);
3399 /* Return the spanned pages */
3400 return zone_end_pfn
- zone_start_pfn
;
3404 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
3405 * then all holes in the requested range will be accounted for.
3407 static unsigned long __meminit
__absent_pages_in_range(int nid
,
3408 unsigned long range_start_pfn
,
3409 unsigned long range_end_pfn
)
3412 unsigned long prev_end_pfn
= 0, hole_pages
= 0;
3413 unsigned long start_pfn
;
3415 /* Find the end_pfn of the first active range of pfns in the node */
3416 i
= first_active_region_index_in_nid(nid
);
3420 prev_end_pfn
= min(early_node_map
[i
].start_pfn
, range_end_pfn
);
3422 /* Account for ranges before physical memory on this node */
3423 if (early_node_map
[i
].start_pfn
> range_start_pfn
)
3424 hole_pages
= prev_end_pfn
- range_start_pfn
;
3426 /* Find all holes for the zone within the node */
3427 for (; i
!= -1; i
= next_active_region_index_in_nid(i
, nid
)) {
3429 /* No need to continue if prev_end_pfn is outside the zone */
3430 if (prev_end_pfn
>= range_end_pfn
)
3433 /* Make sure the end of the zone is not within the hole */
3434 start_pfn
= min(early_node_map
[i
].start_pfn
, range_end_pfn
);
3435 prev_end_pfn
= max(prev_end_pfn
, range_start_pfn
);
3437 /* Update the hole size cound and move on */
3438 if (start_pfn
> range_start_pfn
) {
3439 BUG_ON(prev_end_pfn
> start_pfn
);
3440 hole_pages
+= start_pfn
- prev_end_pfn
;
3442 prev_end_pfn
= early_node_map
[i
].end_pfn
;
3445 /* Account for ranges past physical memory on this node */
3446 if (range_end_pfn
> prev_end_pfn
)
3447 hole_pages
+= range_end_pfn
-
3448 max(range_start_pfn
, prev_end_pfn
);
3454 * absent_pages_in_range - Return number of page frames in holes within a range
3455 * @start_pfn: The start PFN to start searching for holes
3456 * @end_pfn: The end PFN to stop searching for holes
3458 * It returns the number of pages frames in memory holes within a range.
3460 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
3461 unsigned long end_pfn
)
3463 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
3466 /* Return the number of page frames in holes in a zone on a node */
3467 static unsigned long __meminit
zone_absent_pages_in_node(int nid
,
3468 unsigned long zone_type
,
3469 unsigned long *ignored
)
3471 unsigned long node_start_pfn
, node_end_pfn
;
3472 unsigned long zone_start_pfn
, zone_end_pfn
;
3474 get_pfn_range_for_nid(nid
, &node_start_pfn
, &node_end_pfn
);
3475 zone_start_pfn
= max(arch_zone_lowest_possible_pfn
[zone_type
],
3477 zone_end_pfn
= min(arch_zone_highest_possible_pfn
[zone_type
],
3480 adjust_zone_range_for_zone_movable(nid
, zone_type
,
3481 node_start_pfn
, node_end_pfn
,
3482 &zone_start_pfn
, &zone_end_pfn
);
3483 return __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
3487 static inline unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
3488 unsigned long zone_type
,
3489 unsigned long *zones_size
)
3491 return zones_size
[zone_type
];
3494 static inline unsigned long __meminit
zone_absent_pages_in_node(int nid
,
3495 unsigned long zone_type
,
3496 unsigned long *zholes_size
)
3501 return zholes_size
[zone_type
];
3506 static void __meminit
calculate_node_totalpages(struct pglist_data
*pgdat
,
3507 unsigned long *zones_size
, unsigned long *zholes_size
)
3509 unsigned long realtotalpages
, totalpages
= 0;
3512 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
3513 totalpages
+= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
3515 pgdat
->node_spanned_pages
= totalpages
;
3517 realtotalpages
= totalpages
;
3518 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
3520 zone_absent_pages_in_node(pgdat
->node_id
, i
,
3522 pgdat
->node_present_pages
= realtotalpages
;
3523 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
3527 #ifndef CONFIG_SPARSEMEM
3529 * Calculate the size of the zone->blockflags rounded to an unsigned long
3530 * Start by making sure zonesize is a multiple of pageblock_order by rounding
3531 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
3532 * round what is now in bits to nearest long in bits, then return it in
3535 static unsigned long __init
usemap_size(unsigned long zonesize
)
3537 unsigned long usemapsize
;
3539 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
3540 usemapsize
= usemapsize
>> pageblock_order
;
3541 usemapsize
*= NR_PAGEBLOCK_BITS
;
3542 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
3544 return usemapsize
/ 8;
3547 static void __init
setup_usemap(struct pglist_data
*pgdat
,
3548 struct zone
*zone
, unsigned long zonesize
)
3550 unsigned long usemapsize
= usemap_size(zonesize
);
3551 zone
->pageblock_flags
= NULL
;
3553 zone
->pageblock_flags
= alloc_bootmem_node(pgdat
, usemapsize
);
3556 static void inline setup_usemap(struct pglist_data
*pgdat
,
3557 struct zone
*zone
, unsigned long zonesize
) {}
3558 #endif /* CONFIG_SPARSEMEM */
3560 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
3562 /* Return a sensible default order for the pageblock size. */
3563 static inline int pageblock_default_order(void)
3565 if (HPAGE_SHIFT
> PAGE_SHIFT
)
3566 return HUGETLB_PAGE_ORDER
;
3571 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
3572 static inline void __init
set_pageblock_order(unsigned int order
)
3574 /* Check that pageblock_nr_pages has not already been setup */
3575 if (pageblock_order
)
3579 * Assume the largest contiguous order of interest is a huge page.
3580 * This value may be variable depending on boot parameters on IA64
3582 pageblock_order
= order
;
3584 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
3587 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
3588 * and pageblock_default_order() are unused as pageblock_order is set
3589 * at compile-time. See include/linux/pageblock-flags.h for the values of
3590 * pageblock_order based on the kernel config
3592 static inline int pageblock_default_order(unsigned int order
)
3596 #define set_pageblock_order(x) do {} while (0)
3598 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
3601 * Set up the zone data structures:
3602 * - mark all pages reserved
3603 * - mark all memory queues empty
3604 * - clear the memory bitmaps
3606 static void __paginginit
free_area_init_core(struct pglist_data
*pgdat
,
3607 unsigned long *zones_size
, unsigned long *zholes_size
)
3610 int nid
= pgdat
->node_id
;
3611 unsigned long zone_start_pfn
= pgdat
->node_start_pfn
;
3614 pgdat_resize_init(pgdat
);
3615 pgdat
->nr_zones
= 0;
3616 init_waitqueue_head(&pgdat
->kswapd_wait
);
3617 pgdat
->kswapd_max_order
= 0;
3618 pgdat_page_cgroup_init(pgdat
);
3620 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
3621 struct zone
*zone
= pgdat
->node_zones
+ j
;
3622 unsigned long size
, realsize
, memmap_pages
;
3625 size
= zone_spanned_pages_in_node(nid
, j
, zones_size
);
3626 realsize
= size
- zone_absent_pages_in_node(nid
, j
,
3630 * Adjust realsize so that it accounts for how much memory
3631 * is used by this zone for memmap. This affects the watermark
3632 * and per-cpu initialisations
3635 PAGE_ALIGN(size
* sizeof(struct page
)) >> PAGE_SHIFT
;
3636 if (realsize
>= memmap_pages
) {
3637 realsize
-= memmap_pages
;
3640 " %s zone: %lu pages used for memmap\n",
3641 zone_names
[j
], memmap_pages
);
3644 " %s zone: %lu pages exceeds realsize %lu\n",
3645 zone_names
[j
], memmap_pages
, realsize
);
3647 /* Account for reserved pages */
3648 if (j
== 0 && realsize
> dma_reserve
) {
3649 realsize
-= dma_reserve
;
3650 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
3651 zone_names
[0], dma_reserve
);
3654 if (!is_highmem_idx(j
))
3655 nr_kernel_pages
+= realsize
;
3656 nr_all_pages
+= realsize
;
3658 zone
->spanned_pages
= size
;
3659 zone
->present_pages
= realsize
;
3662 zone
->min_unmapped_pages
= (realsize
*sysctl_min_unmapped_ratio
)
3664 zone
->min_slab_pages
= (realsize
* sysctl_min_slab_ratio
) / 100;
3666 zone
->name
= zone_names
[j
];
3667 spin_lock_init(&zone
->lock
);
3668 spin_lock_init(&zone
->lru_lock
);
3669 zone_seqlock_init(zone
);
3670 zone
->zone_pgdat
= pgdat
;
3672 zone
->prev_priority
= DEF_PRIORITY
;
3674 zone_pcp_init(zone
);
3676 INIT_LIST_HEAD(&zone
->lru
[l
].list
);
3677 zone
->lru
[l
].nr_scan
= 0;
3679 zone
->reclaim_stat
.recent_rotated
[0] = 0;
3680 zone
->reclaim_stat
.recent_rotated
[1] = 0;
3681 zone
->reclaim_stat
.recent_scanned
[0] = 0;
3682 zone
->reclaim_stat
.recent_scanned
[1] = 0;
3683 zap_zone_vm_stats(zone
);
3688 set_pageblock_order(pageblock_default_order());
3689 setup_usemap(pgdat
, zone
, size
);
3690 ret
= init_currently_empty_zone(zone
, zone_start_pfn
,
3691 size
, MEMMAP_EARLY
);
3693 memmap_init(size
, nid
, j
, zone_start_pfn
);
3694 zone_start_pfn
+= size
;
3698 static void __init_refok
alloc_node_mem_map(struct pglist_data
*pgdat
)
3700 /* Skip empty nodes */
3701 if (!pgdat
->node_spanned_pages
)
3704 #ifdef CONFIG_FLAT_NODE_MEM_MAP
3705 /* ia64 gets its own node_mem_map, before this, without bootmem */
3706 if (!pgdat
->node_mem_map
) {
3707 unsigned long size
, start
, end
;
3711 * The zone's endpoints aren't required to be MAX_ORDER
3712 * aligned but the node_mem_map endpoints must be in order
3713 * for the buddy allocator to function correctly.
3715 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
3716 end
= pgdat
->node_start_pfn
+ pgdat
->node_spanned_pages
;
3717 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
3718 size
= (end
- start
) * sizeof(struct page
);
3719 map
= alloc_remap(pgdat
->node_id
, size
);
3721 map
= alloc_bootmem_node(pgdat
, size
);
3722 pgdat
->node_mem_map
= map
+ (pgdat
->node_start_pfn
- start
);
3724 #ifndef CONFIG_NEED_MULTIPLE_NODES
3726 * With no DISCONTIG, the global mem_map is just set as node 0's
3728 if (pgdat
== NODE_DATA(0)) {
3729 mem_map
= NODE_DATA(0)->node_mem_map
;
3730 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3731 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
3732 mem_map
-= (pgdat
->node_start_pfn
- ARCH_PFN_OFFSET
);
3733 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
3736 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
3739 void __paginginit
free_area_init_node(int nid
, unsigned long *zones_size
,
3740 unsigned long node_start_pfn
, unsigned long *zholes_size
)
3742 pg_data_t
*pgdat
= NODE_DATA(nid
);
3744 pgdat
->node_id
= nid
;
3745 pgdat
->node_start_pfn
= node_start_pfn
;
3746 calculate_node_totalpages(pgdat
, zones_size
, zholes_size
);
3748 alloc_node_mem_map(pgdat
);
3749 #ifdef CONFIG_FLAT_NODE_MEM_MAP
3750 printk(KERN_DEBUG
"free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
3751 nid
, (unsigned long)pgdat
,
3752 (unsigned long)pgdat
->node_mem_map
);
3755 free_area_init_core(pgdat
, zones_size
, zholes_size
);
3758 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3760 #if MAX_NUMNODES > 1
3762 * Figure out the number of possible node ids.
3764 static void __init
setup_nr_node_ids(void)
3767 unsigned int highest
= 0;
3769 for_each_node_mask(node
, node_possible_map
)
3771 nr_node_ids
= highest
+ 1;
3774 static inline void setup_nr_node_ids(void)
3780 * add_active_range - Register a range of PFNs backed by physical memory
3781 * @nid: The node ID the range resides on
3782 * @start_pfn: The start PFN of the available physical memory
3783 * @end_pfn: The end PFN of the available physical memory
3785 * These ranges are stored in an early_node_map[] and later used by
3786 * free_area_init_nodes() to calculate zone sizes and holes. If the
3787 * range spans a memory hole, it is up to the architecture to ensure
3788 * the memory is not freed by the bootmem allocator. If possible
3789 * the range being registered will be merged with existing ranges.
3791 void __init
add_active_range(unsigned int nid
, unsigned long start_pfn
,
3792 unsigned long end_pfn
)
3796 mminit_dprintk(MMINIT_TRACE
, "memory_register",
3797 "Entering add_active_range(%d, %#lx, %#lx) "
3798 "%d entries of %d used\n",
3799 nid
, start_pfn
, end_pfn
,
3800 nr_nodemap_entries
, MAX_ACTIVE_REGIONS
);
3802 mminit_validate_memmodel_limits(&start_pfn
, &end_pfn
);
3804 /* Merge with existing active regions if possible */
3805 for (i
= 0; i
< nr_nodemap_entries
; i
++) {
3806 if (early_node_map
[i
].nid
!= nid
)
3809 /* Skip if an existing region covers this new one */
3810 if (start_pfn
>= early_node_map
[i
].start_pfn
&&
3811 end_pfn
<= early_node_map
[i
].end_pfn
)
3814 /* Merge forward if suitable */
3815 if (start_pfn
<= early_node_map
[i
].end_pfn
&&
3816 end_pfn
> early_node_map
[i
].end_pfn
) {
3817 early_node_map
[i
].end_pfn
= end_pfn
;
3821 /* Merge backward if suitable */
3822 if (start_pfn
< early_node_map
[i
].end_pfn
&&
3823 end_pfn
>= early_node_map
[i
].start_pfn
) {
3824 early_node_map
[i
].start_pfn
= start_pfn
;
3829 /* Check that early_node_map is large enough */
3830 if (i
>= MAX_ACTIVE_REGIONS
) {
3831 printk(KERN_CRIT
"More than %d memory regions, truncating\n",
3832 MAX_ACTIVE_REGIONS
);
3836 early_node_map
[i
].nid
= nid
;
3837 early_node_map
[i
].start_pfn
= start_pfn
;
3838 early_node_map
[i
].end_pfn
= end_pfn
;
3839 nr_nodemap_entries
= i
+ 1;
3843 * remove_active_range - Shrink an existing registered range of PFNs
3844 * @nid: The node id the range is on that should be shrunk
3845 * @start_pfn: The new PFN of the range
3846 * @end_pfn: The new PFN of the range
3848 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
3849 * The map is kept near the end physical page range that has already been
3850 * registered. This function allows an arch to shrink an existing registered
3853 void __init
remove_active_range(unsigned int nid
, unsigned long start_pfn
,
3854 unsigned long end_pfn
)
3859 printk(KERN_DEBUG
"remove_active_range (%d, %lu, %lu)\n",
3860 nid
, start_pfn
, end_pfn
);
3862 /* Find the old active region end and shrink */
3863 for_each_active_range_index_in_nid(i
, nid
) {
3864 if (early_node_map
[i
].start_pfn
>= start_pfn
&&
3865 early_node_map
[i
].end_pfn
<= end_pfn
) {
3867 early_node_map
[i
].start_pfn
= 0;
3868 early_node_map
[i
].end_pfn
= 0;
3872 if (early_node_map
[i
].start_pfn
< start_pfn
&&
3873 early_node_map
[i
].end_pfn
> start_pfn
) {
3874 unsigned long temp_end_pfn
= early_node_map
[i
].end_pfn
;
3875 early_node_map
[i
].end_pfn
= start_pfn
;
3876 if (temp_end_pfn
> end_pfn
)
3877 add_active_range(nid
, end_pfn
, temp_end_pfn
);
3880 if (early_node_map
[i
].start_pfn
>= start_pfn
&&
3881 early_node_map
[i
].end_pfn
> end_pfn
&&
3882 early_node_map
[i
].start_pfn
< end_pfn
) {
3883 early_node_map
[i
].start_pfn
= end_pfn
;
3891 /* remove the blank ones */
3892 for (i
= nr_nodemap_entries
- 1; i
> 0; i
--) {
3893 if (early_node_map
[i
].nid
!= nid
)
3895 if (early_node_map
[i
].end_pfn
)
3897 /* we found it, get rid of it */
3898 for (j
= i
; j
< nr_nodemap_entries
- 1; j
++)
3899 memcpy(&early_node_map
[j
], &early_node_map
[j
+1],
3900 sizeof(early_node_map
[j
]));
3901 j
= nr_nodemap_entries
- 1;
3902 memset(&early_node_map
[j
], 0, sizeof(early_node_map
[j
]));
3903 nr_nodemap_entries
--;
3908 * remove_all_active_ranges - Remove all currently registered regions
3910 * During discovery, it may be found that a table like SRAT is invalid
3911 * and an alternative discovery method must be used. This function removes
3912 * all currently registered regions.
3914 void __init
remove_all_active_ranges(void)
3916 memset(early_node_map
, 0, sizeof(early_node_map
));
3917 nr_nodemap_entries
= 0;
3920 /* Compare two active node_active_regions */
3921 static int __init
cmp_node_active_region(const void *a
, const void *b
)
3923 struct node_active_region
*arange
= (struct node_active_region
*)a
;
3924 struct node_active_region
*brange
= (struct node_active_region
*)b
;
3926 /* Done this way to avoid overflows */
3927 if (arange
->start_pfn
> brange
->start_pfn
)
3929 if (arange
->start_pfn
< brange
->start_pfn
)
3935 /* sort the node_map by start_pfn */
3936 static void __init
sort_node_map(void)
3938 sort(early_node_map
, (size_t)nr_nodemap_entries
,
3939 sizeof(struct node_active_region
),
3940 cmp_node_active_region
, NULL
);
3943 /* Find the lowest pfn for a node */
3944 static unsigned long __init
find_min_pfn_for_node(int nid
)
3947 unsigned long min_pfn
= ULONG_MAX
;
3949 /* Assuming a sorted map, the first range found has the starting pfn */
3950 for_each_active_range_index_in_nid(i
, nid
)
3951 min_pfn
= min(min_pfn
, early_node_map
[i
].start_pfn
);
3953 if (min_pfn
== ULONG_MAX
) {
3955 "Could not find start_pfn for node %d\n", nid
);
3963 * find_min_pfn_with_active_regions - Find the minimum PFN registered
3965 * It returns the minimum PFN based on information provided via
3966 * add_active_range().
3968 unsigned long __init
find_min_pfn_with_active_regions(void)
3970 return find_min_pfn_for_node(MAX_NUMNODES
);
3974 * early_calculate_totalpages()
3975 * Sum pages in active regions for movable zone.
3976 * Populate N_HIGH_MEMORY for calculating usable_nodes.
3978 static unsigned long __init
early_calculate_totalpages(void)
3981 unsigned long totalpages
= 0;
3983 for (i
= 0; i
< nr_nodemap_entries
; i
++) {
3984 unsigned long pages
= early_node_map
[i
].end_pfn
-
3985 early_node_map
[i
].start_pfn
;
3986 totalpages
+= pages
;
3988 node_set_state(early_node_map
[i
].nid
, N_HIGH_MEMORY
);
3994 * Find the PFN the Movable zone begins in each node. Kernel memory
3995 * is spread evenly between nodes as long as the nodes have enough
3996 * memory. When they don't, some nodes will have more kernelcore than
3999 static void __init
find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn
)
4002 unsigned long usable_startpfn
;
4003 unsigned long kernelcore_node
, kernelcore_remaining
;
4004 unsigned long totalpages
= early_calculate_totalpages();
4005 int usable_nodes
= nodes_weight(node_states
[N_HIGH_MEMORY
]);
4008 * If movablecore was specified, calculate what size of
4009 * kernelcore that corresponds so that memory usable for
4010 * any allocation type is evenly spread. If both kernelcore
4011 * and movablecore are specified, then the value of kernelcore
4012 * will be used for required_kernelcore if it's greater than
4013 * what movablecore would have allowed.
4015 if (required_movablecore
) {
4016 unsigned long corepages
;
4019 * Round-up so that ZONE_MOVABLE is at least as large as what
4020 * was requested by the user
4022 required_movablecore
=
4023 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
4024 corepages
= totalpages
- required_movablecore
;
4026 required_kernelcore
= max(required_kernelcore
, corepages
);
4029 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4030 if (!required_kernelcore
)
4033 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4034 find_usable_zone_for_movable();
4035 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
4038 /* Spread kernelcore memory as evenly as possible throughout nodes */
4039 kernelcore_node
= required_kernelcore
/ usable_nodes
;
4040 for_each_node_state(nid
, N_HIGH_MEMORY
) {
4042 * Recalculate kernelcore_node if the division per node
4043 * now exceeds what is necessary to satisfy the requested
4044 * amount of memory for the kernel
4046 if (required_kernelcore
< kernelcore_node
)
4047 kernelcore_node
= required_kernelcore
/ usable_nodes
;
4050 * As the map is walked, we track how much memory is usable
4051 * by the kernel using kernelcore_remaining. When it is
4052 * 0, the rest of the node is usable by ZONE_MOVABLE
4054 kernelcore_remaining
= kernelcore_node
;
4056 /* Go through each range of PFNs within this node */
4057 for_each_active_range_index_in_nid(i
, nid
) {
4058 unsigned long start_pfn
, end_pfn
;
4059 unsigned long size_pages
;
4061 start_pfn
= max(early_node_map
[i
].start_pfn
,
4062 zone_movable_pfn
[nid
]);
4063 end_pfn
= early_node_map
[i
].end_pfn
;
4064 if (start_pfn
>= end_pfn
)
4067 /* Account for what is only usable for kernelcore */
4068 if (start_pfn
< usable_startpfn
) {
4069 unsigned long kernel_pages
;
4070 kernel_pages
= min(end_pfn
, usable_startpfn
)
4073 kernelcore_remaining
-= min(kernel_pages
,
4074 kernelcore_remaining
);
4075 required_kernelcore
-= min(kernel_pages
,
4076 required_kernelcore
);
4078 /* Continue if range is now fully accounted */
4079 if (end_pfn
<= usable_startpfn
) {
4082 * Push zone_movable_pfn to the end so
4083 * that if we have to rebalance
4084 * kernelcore across nodes, we will
4085 * not double account here
4087 zone_movable_pfn
[nid
] = end_pfn
;
4090 start_pfn
= usable_startpfn
;
4094 * The usable PFN range for ZONE_MOVABLE is from
4095 * start_pfn->end_pfn. Calculate size_pages as the
4096 * number of pages used as kernelcore
4098 size_pages
= end_pfn
- start_pfn
;
4099 if (size_pages
> kernelcore_remaining
)
4100 size_pages
= kernelcore_remaining
;
4101 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
4104 * Some kernelcore has been met, update counts and
4105 * break if the kernelcore for this node has been
4108 required_kernelcore
-= min(required_kernelcore
,
4110 kernelcore_remaining
-= size_pages
;
4111 if (!kernelcore_remaining
)
4117 * If there is still required_kernelcore, we do another pass with one
4118 * less node in the count. This will push zone_movable_pfn[nid] further
4119 * along on the nodes that still have memory until kernelcore is
4123 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
4126 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4127 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
4128 zone_movable_pfn
[nid
] =
4129 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
4132 /* Any regular memory on that node ? */
4133 static void check_for_regular_memory(pg_data_t
*pgdat
)
4135 #ifdef CONFIG_HIGHMEM
4136 enum zone_type zone_type
;
4138 for (zone_type
= 0; zone_type
<= ZONE_NORMAL
; zone_type
++) {
4139 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
4140 if (zone
->present_pages
)
4141 node_set_state(zone_to_nid(zone
), N_NORMAL_MEMORY
);
4147 * free_area_init_nodes - Initialise all pg_data_t and zone data
4148 * @max_zone_pfn: an array of max PFNs for each zone
4150 * This will call free_area_init_node() for each active node in the system.
4151 * Using the page ranges provided by add_active_range(), the size of each
4152 * zone in each node and their holes is calculated. If the maximum PFN
4153 * between two adjacent zones match, it is assumed that the zone is empty.
4154 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4155 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4156 * starts where the previous one ended. For example, ZONE_DMA32 starts
4157 * at arch_max_dma_pfn.
4159 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
4164 /* Sort early_node_map as initialisation assumes it is sorted */
4167 /* Record where the zone boundaries are */
4168 memset(arch_zone_lowest_possible_pfn
, 0,
4169 sizeof(arch_zone_lowest_possible_pfn
));
4170 memset(arch_zone_highest_possible_pfn
, 0,
4171 sizeof(arch_zone_highest_possible_pfn
));
4172 arch_zone_lowest_possible_pfn
[0] = find_min_pfn_with_active_regions();
4173 arch_zone_highest_possible_pfn
[0] = max_zone_pfn
[0];
4174 for (i
= 1; i
< MAX_NR_ZONES
; i
++) {
4175 if (i
== ZONE_MOVABLE
)
4177 arch_zone_lowest_possible_pfn
[i
] =
4178 arch_zone_highest_possible_pfn
[i
-1];
4179 arch_zone_highest_possible_pfn
[i
] =
4180 max(max_zone_pfn
[i
], arch_zone_lowest_possible_pfn
[i
]);
4182 arch_zone_lowest_possible_pfn
[ZONE_MOVABLE
] = 0;
4183 arch_zone_highest_possible_pfn
[ZONE_MOVABLE
] = 0;
4185 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4186 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
4187 find_zone_movable_pfns_for_nodes(zone_movable_pfn
);
4189 /* Print out the zone ranges */
4190 printk("Zone PFN ranges:\n");
4191 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
4192 if (i
== ZONE_MOVABLE
)
4194 printk(" %-8s %0#10lx -> %0#10lx\n",
4196 arch_zone_lowest_possible_pfn
[i
],
4197 arch_zone_highest_possible_pfn
[i
]);
4200 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4201 printk("Movable zone start PFN for each node\n");
4202 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
4203 if (zone_movable_pfn
[i
])
4204 printk(" Node %d: %lu\n", i
, zone_movable_pfn
[i
]);
4207 /* Print out the early_node_map[] */
4208 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries
);
4209 for (i
= 0; i
< nr_nodemap_entries
; i
++)
4210 printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map
[i
].nid
,
4211 early_node_map
[i
].start_pfn
,
4212 early_node_map
[i
].end_pfn
);
4214 /* Initialise every node */
4215 mminit_verify_pageflags_layout();
4216 setup_nr_node_ids();
4217 for_each_online_node(nid
) {
4218 pg_data_t
*pgdat
= NODE_DATA(nid
);
4219 free_area_init_node(nid
, NULL
,
4220 find_min_pfn_for_node(nid
), NULL
);
4222 /* Any memory on that node */
4223 if (pgdat
->node_present_pages
)
4224 node_set_state(nid
, N_HIGH_MEMORY
);
4225 check_for_regular_memory(pgdat
);
4229 static int __init
cmdline_parse_core(char *p
, unsigned long *core
)
4231 unsigned long long coremem
;
4235 coremem
= memparse(p
, &p
);
4236 *core
= coremem
>> PAGE_SHIFT
;
4238 /* Paranoid check that UL is enough for the coremem value */
4239 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
4245 * kernelcore=size sets the amount of memory for use for allocations that
4246 * cannot be reclaimed or migrated.
4248 static int __init
cmdline_parse_kernelcore(char *p
)
4250 return cmdline_parse_core(p
, &required_kernelcore
);
4254 * movablecore=size sets the amount of memory for use for allocations that
4255 * can be reclaimed or migrated.
4257 static int __init
cmdline_parse_movablecore(char *p
)
4259 return cmdline_parse_core(p
, &required_movablecore
);
4262 early_param("kernelcore", cmdline_parse_kernelcore
);
4263 early_param("movablecore", cmdline_parse_movablecore
);
4265 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4268 * set_dma_reserve - set the specified number of pages reserved in the first zone
4269 * @new_dma_reserve: The number of pages to mark reserved
4271 * The per-cpu batchsize and zone watermarks are determined by present_pages.
4272 * In the DMA zone, a significant percentage may be consumed by kernel image
4273 * and other unfreeable allocations which can skew the watermarks badly. This
4274 * function may optionally be used to account for unfreeable pages in the
4275 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
4276 * smaller per-cpu batchsize.
4278 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
4280 dma_reserve
= new_dma_reserve
;
4283 #ifndef CONFIG_NEED_MULTIPLE_NODES
4284 struct pglist_data __refdata contig_page_data
= { .bdata
= &bootmem_node_data
[0] };
4285 EXPORT_SYMBOL(contig_page_data
);
4288 void __init
free_area_init(unsigned long *zones_size
)
4290 free_area_init_node(0, zones_size
,
4291 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
4294 static int page_alloc_cpu_notify(struct notifier_block
*self
,
4295 unsigned long action
, void *hcpu
)
4297 int cpu
= (unsigned long)hcpu
;
4299 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
) {
4303 * Spill the event counters of the dead processor
4304 * into the current processors event counters.
4305 * This artificially elevates the count of the current
4308 vm_events_fold_cpu(cpu
);
4311 * Zero the differential counters of the dead processor
4312 * so that the vm statistics are consistent.
4314 * This is only okay since the processor is dead and cannot
4315 * race with what we are doing.
4317 refresh_cpu_vm_stats(cpu
);
4322 void __init
page_alloc_init(void)
4324 hotcpu_notifier(page_alloc_cpu_notify
, 0);
4328 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
4329 * or min_free_kbytes changes.
4331 static void calculate_totalreserve_pages(void)
4333 struct pglist_data
*pgdat
;
4334 unsigned long reserve_pages
= 0;
4335 enum zone_type i
, j
;
4337 for_each_online_pgdat(pgdat
) {
4338 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
4339 struct zone
*zone
= pgdat
->node_zones
+ i
;
4340 unsigned long max
= 0;
4342 /* Find valid and maximum lowmem_reserve in the zone */
4343 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
4344 if (zone
->lowmem_reserve
[j
] > max
)
4345 max
= zone
->lowmem_reserve
[j
];
4348 /* we treat the high watermark as reserved pages. */
4349 max
+= high_wmark_pages(zone
);
4351 if (max
> zone
->present_pages
)
4352 max
= zone
->present_pages
;
4353 reserve_pages
+= max
;
4356 totalreserve_pages
= reserve_pages
;
4360 * setup_per_zone_lowmem_reserve - called whenever
4361 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
4362 * has a correct pages reserved value, so an adequate number of
4363 * pages are left in the zone after a successful __alloc_pages().
4365 static void setup_per_zone_lowmem_reserve(void)
4367 struct pglist_data
*pgdat
;
4368 enum zone_type j
, idx
;
4370 for_each_online_pgdat(pgdat
) {
4371 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
4372 struct zone
*zone
= pgdat
->node_zones
+ j
;
4373 unsigned long present_pages
= zone
->present_pages
;
4375 zone
->lowmem_reserve
[j
] = 0;
4379 struct zone
*lower_zone
;
4383 if (sysctl_lowmem_reserve_ratio
[idx
] < 1)
4384 sysctl_lowmem_reserve_ratio
[idx
] = 1;
4386 lower_zone
= pgdat
->node_zones
+ idx
;
4387 lower_zone
->lowmem_reserve
[j
] = present_pages
/
4388 sysctl_lowmem_reserve_ratio
[idx
];
4389 present_pages
+= lower_zone
->present_pages
;
4394 /* update totalreserve_pages */
4395 calculate_totalreserve_pages();
4399 * setup_per_zone_pages_min - called when min_free_kbytes changes.
4401 * Ensures that the pages_{min,low,high} values for each zone are set correctly
4402 * with respect to min_free_kbytes.
4404 void setup_per_zone_pages_min(void)
4406 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
4407 unsigned long lowmem_pages
= 0;
4409 unsigned long flags
;
4411 /* Calculate total number of !ZONE_HIGHMEM pages */
4412 for_each_zone(zone
) {
4413 if (!is_highmem(zone
))
4414 lowmem_pages
+= zone
->present_pages
;
4417 for_each_zone(zone
) {
4420 spin_lock_irqsave(&zone
->lock
, flags
);
4421 tmp
= (u64
)pages_min
* zone
->present_pages
;
4422 do_div(tmp
, lowmem_pages
);
4423 if (is_highmem(zone
)) {
4425 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
4426 * need highmem pages, so cap pages_min to a small
4429 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
4430 * deltas controls asynch page reclaim, and so should
4431 * not be capped for highmem.
4435 min_pages
= zone
->present_pages
/ 1024;
4436 if (min_pages
< SWAP_CLUSTER_MAX
)
4437 min_pages
= SWAP_CLUSTER_MAX
;
4438 if (min_pages
> 128)
4440 zone
->watermark
[WMARK_MIN
] = min_pages
;
4443 * If it's a lowmem zone, reserve a number of pages
4444 * proportionate to the zone's size.
4446 zone
->watermark
[WMARK_MIN
] = tmp
;
4449 zone
->watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + (tmp
>> 2);
4450 zone
->watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + (tmp
>> 1);
4451 setup_zone_migrate_reserve(zone
);
4452 spin_unlock_irqrestore(&zone
->lock
, flags
);
4455 /* update totalreserve_pages */
4456 calculate_totalreserve_pages();
4460 * setup_per_zone_inactive_ratio - called when min_free_kbytes changes.
4462 * The inactive anon list should be small enough that the VM never has to
4463 * do too much work, but large enough that each inactive page has a chance
4464 * to be referenced again before it is swapped out.
4466 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
4467 * INACTIVE_ANON pages on this zone's LRU, maintained by the
4468 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
4469 * the anonymous pages are kept on the inactive list.
4472 * memory ratio inactive anon
4473 * -------------------------------------
4482 static void setup_per_zone_inactive_ratio(void)
4486 for_each_zone(zone
) {
4487 unsigned int gb
, ratio
;
4489 /* Zone size in gigabytes */
4490 gb
= zone
->present_pages
>> (30 - PAGE_SHIFT
);
4491 ratio
= int_sqrt(10 * gb
);
4495 zone
->inactive_ratio
= ratio
;
4500 * Initialise min_free_kbytes.
4502 * For small machines we want it small (128k min). For large machines
4503 * we want it large (64MB max). But it is not linear, because network
4504 * bandwidth does not increase linearly with machine size. We use
4506 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
4507 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
4523 static int __init
init_per_zone_pages_min(void)
4525 unsigned long lowmem_kbytes
;
4527 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
4529 min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
4530 if (min_free_kbytes
< 128)
4531 min_free_kbytes
= 128;
4532 if (min_free_kbytes
> 65536)
4533 min_free_kbytes
= 65536;
4534 setup_per_zone_pages_min();
4535 setup_per_zone_lowmem_reserve();
4536 setup_per_zone_inactive_ratio();
4539 module_init(init_per_zone_pages_min
)
4542 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
4543 * that we can call two helper functions whenever min_free_kbytes
4546 int min_free_kbytes_sysctl_handler(ctl_table
*table
, int write
,
4547 struct file
*file
, void __user
*buffer
, size_t *length
, loff_t
*ppos
)
4549 proc_dointvec(table
, write
, file
, buffer
, length
, ppos
);
4551 setup_per_zone_pages_min();
4556 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table
*table
, int write
,
4557 struct file
*file
, void __user
*buffer
, size_t *length
, loff_t
*ppos
)
4562 rc
= proc_dointvec_minmax(table
, write
, file
, buffer
, length
, ppos
);
4567 zone
->min_unmapped_pages
= (zone
->present_pages
*
4568 sysctl_min_unmapped_ratio
) / 100;
4572 int sysctl_min_slab_ratio_sysctl_handler(ctl_table
*table
, int write
,
4573 struct file
*file
, void __user
*buffer
, size_t *length
, loff_t
*ppos
)
4578 rc
= proc_dointvec_minmax(table
, write
, file
, buffer
, length
, ppos
);
4583 zone
->min_slab_pages
= (zone
->present_pages
*
4584 sysctl_min_slab_ratio
) / 100;
4590 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
4591 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
4592 * whenever sysctl_lowmem_reserve_ratio changes.
4594 * The reserve ratio obviously has absolutely no relation with the
4595 * minimum watermarks. The lowmem reserve ratio can only make sense
4596 * if in function of the boot time zone sizes.
4598 int lowmem_reserve_ratio_sysctl_handler(ctl_table
*table
, int write
,
4599 struct file
*file
, void __user
*buffer
, size_t *length
, loff_t
*ppos
)
4601 proc_dointvec_minmax(table
, write
, file
, buffer
, length
, ppos
);
4602 setup_per_zone_lowmem_reserve();
4607 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
4608 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
4609 * can have before it gets flushed back to buddy allocator.
4612 int percpu_pagelist_fraction_sysctl_handler(ctl_table
*table
, int write
,
4613 struct file
*file
, void __user
*buffer
, size_t *length
, loff_t
*ppos
)
4619 ret
= proc_dointvec_minmax(table
, write
, file
, buffer
, length
, ppos
);
4620 if (!write
|| (ret
== -EINVAL
))
4622 for_each_zone(zone
) {
4623 for_each_online_cpu(cpu
) {
4625 high
= zone
->present_pages
/ percpu_pagelist_fraction
;
4626 setup_pagelist_highmark(zone_pcp(zone
, cpu
), high
);
4632 int hashdist
= HASHDIST_DEFAULT
;
4635 static int __init
set_hashdist(char *str
)
4639 hashdist
= simple_strtoul(str
, &str
, 0);
4642 __setup("hashdist=", set_hashdist
);
4646 * allocate a large system hash table from bootmem
4647 * - it is assumed that the hash table must contain an exact power-of-2
4648 * quantity of entries
4649 * - limit is the number of hash buckets, not the total allocation size
4651 void *__init
alloc_large_system_hash(const char *tablename
,
4652 unsigned long bucketsize
,
4653 unsigned long numentries
,
4656 unsigned int *_hash_shift
,
4657 unsigned int *_hash_mask
,
4658 unsigned long limit
)
4660 unsigned long long max
= limit
;
4661 unsigned long log2qty
, size
;
4664 /* allow the kernel cmdline to have a say */
4666 /* round applicable memory size up to nearest megabyte */
4667 numentries
= nr_kernel_pages
;
4668 numentries
+= (1UL << (20 - PAGE_SHIFT
)) - 1;
4669 numentries
>>= 20 - PAGE_SHIFT
;
4670 numentries
<<= 20 - PAGE_SHIFT
;
4672 /* limit to 1 bucket per 2^scale bytes of low memory */
4673 if (scale
> PAGE_SHIFT
)
4674 numentries
>>= (scale
- PAGE_SHIFT
);
4676 numentries
<<= (PAGE_SHIFT
- scale
);
4678 /* Make sure we've got at least a 0-order allocation.. */
4679 if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
4680 numentries
= PAGE_SIZE
/ bucketsize
;
4682 numentries
= roundup_pow_of_two(numentries
);
4684 /* limit allocation size to 1/16 total memory by default */
4686 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
4687 do_div(max
, bucketsize
);
4690 if (numentries
> max
)
4693 log2qty
= ilog2(numentries
);
4696 size
= bucketsize
<< log2qty
;
4697 if (flags
& HASH_EARLY
)
4698 table
= alloc_bootmem_nopanic(size
);
4700 table
= __vmalloc(size
, GFP_ATOMIC
, PAGE_KERNEL
);
4703 * If bucketsize is not a power-of-two, we may free
4704 * some pages at the end of hash table which
4705 * alloc_pages_exact() automatically does
4707 if (get_order(size
) < MAX_ORDER
)
4708 table
= alloc_pages_exact(size
, GFP_ATOMIC
);
4710 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
4713 panic("Failed to allocate %s hash table\n", tablename
);
4715 printk(KERN_INFO
"%s hash table entries: %d (order: %d, %lu bytes)\n",
4718 ilog2(size
) - PAGE_SHIFT
,
4722 *_hash_shift
= log2qty
;
4724 *_hash_mask
= (1 << log2qty
) - 1;
4727 * If hashdist is set, the table allocation is done with __vmalloc()
4728 * which invokes the kmemleak_alloc() callback. This function may also
4729 * be called before the slab and kmemleak are initialised when
4730 * kmemleak simply buffers the request to be executed later
4731 * (GFP_ATOMIC flag ignored in this case).
4734 kmemleak_alloc(table
, size
, 1, GFP_ATOMIC
);
4739 /* Return a pointer to the bitmap storing bits affecting a block of pages */
4740 static inline unsigned long *get_pageblock_bitmap(struct zone
*zone
,
4743 #ifdef CONFIG_SPARSEMEM
4744 return __pfn_to_section(pfn
)->pageblock_flags
;
4746 return zone
->pageblock_flags
;
4747 #endif /* CONFIG_SPARSEMEM */
4750 static inline int pfn_to_bitidx(struct zone
*zone
, unsigned long pfn
)
4752 #ifdef CONFIG_SPARSEMEM
4753 pfn
&= (PAGES_PER_SECTION
-1);
4754 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
4756 pfn
= pfn
- zone
->zone_start_pfn
;
4757 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
4758 #endif /* CONFIG_SPARSEMEM */
4762 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
4763 * @page: The page within the block of interest
4764 * @start_bitidx: The first bit of interest to retrieve
4765 * @end_bitidx: The last bit of interest
4766 * returns pageblock_bits flags
4768 unsigned long get_pageblock_flags_group(struct page
*page
,
4769 int start_bitidx
, int end_bitidx
)
4772 unsigned long *bitmap
;
4773 unsigned long pfn
, bitidx
;
4774 unsigned long flags
= 0;
4775 unsigned long value
= 1;
4777 zone
= page_zone(page
);
4778 pfn
= page_to_pfn(page
);
4779 bitmap
= get_pageblock_bitmap(zone
, pfn
);
4780 bitidx
= pfn_to_bitidx(zone
, pfn
);
4782 for (; start_bitidx
<= end_bitidx
; start_bitidx
++, value
<<= 1)
4783 if (test_bit(bitidx
+ start_bitidx
, bitmap
))
4790 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
4791 * @page: The page within the block of interest
4792 * @start_bitidx: The first bit of interest
4793 * @end_bitidx: The last bit of interest
4794 * @flags: The flags to set
4796 void set_pageblock_flags_group(struct page
*page
, unsigned long flags
,
4797 int start_bitidx
, int end_bitidx
)
4800 unsigned long *bitmap
;
4801 unsigned long pfn
, bitidx
;
4802 unsigned long value
= 1;
4804 zone
= page_zone(page
);
4805 pfn
= page_to_pfn(page
);
4806 bitmap
= get_pageblock_bitmap(zone
, pfn
);
4807 bitidx
= pfn_to_bitidx(zone
, pfn
);
4808 VM_BUG_ON(pfn
< zone
->zone_start_pfn
);
4809 VM_BUG_ON(pfn
>= zone
->zone_start_pfn
+ zone
->spanned_pages
);
4811 for (; start_bitidx
<= end_bitidx
; start_bitidx
++, value
<<= 1)
4813 __set_bit(bitidx
+ start_bitidx
, bitmap
);
4815 __clear_bit(bitidx
+ start_bitidx
, bitmap
);
4819 * This is designed as sub function...plz see page_isolation.c also.
4820 * set/clear page block's type to be ISOLATE.
4821 * page allocater never alloc memory from ISOLATE block.
4824 int set_migratetype_isolate(struct page
*page
)
4827 unsigned long flags
;
4830 zone
= page_zone(page
);
4831 spin_lock_irqsave(&zone
->lock
, flags
);
4833 * In future, more migrate types will be able to be isolation target.
4835 if (get_pageblock_migratetype(page
) != MIGRATE_MOVABLE
)
4837 set_pageblock_migratetype(page
, MIGRATE_ISOLATE
);
4838 move_freepages_block(zone
, page
, MIGRATE_ISOLATE
);
4841 spin_unlock_irqrestore(&zone
->lock
, flags
);
4847 void unset_migratetype_isolate(struct page
*page
)
4850 unsigned long flags
;
4851 zone
= page_zone(page
);
4852 spin_lock_irqsave(&zone
->lock
, flags
);
4853 if (get_pageblock_migratetype(page
) != MIGRATE_ISOLATE
)
4855 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
4856 move_freepages_block(zone
, page
, MIGRATE_MOVABLE
);
4858 spin_unlock_irqrestore(&zone
->lock
, flags
);
4861 #ifdef CONFIG_MEMORY_HOTREMOVE
4863 * All pages in the range must be isolated before calling this.
4866 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
4872 unsigned long flags
;
4873 /* find the first valid pfn */
4874 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
4879 zone
= page_zone(pfn_to_page(pfn
));
4880 spin_lock_irqsave(&zone
->lock
, flags
);
4882 while (pfn
< end_pfn
) {
4883 if (!pfn_valid(pfn
)) {
4887 page
= pfn_to_page(pfn
);
4888 BUG_ON(page_count(page
));
4889 BUG_ON(!PageBuddy(page
));
4890 order
= page_order(page
);
4891 #ifdef CONFIG_DEBUG_VM
4892 printk(KERN_INFO
"remove from free list %lx %d %lx\n",
4893 pfn
, 1 << order
, end_pfn
);
4895 list_del(&page
->lru
);
4896 rmv_page_order(page
);
4897 zone
->free_area
[order
].nr_free
--;
4898 __mod_zone_page_state(zone
, NR_FREE_PAGES
,
4900 for (i
= 0; i
< (1 << order
); i
++)
4901 SetPageReserved((page
+i
));
4902 pfn
+= (1 << order
);
4904 spin_unlock_irqrestore(&zone
->lock
, flags
);