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 EXPORT_SYMBOL(nr_node_ids
);
167 int page_group_by_mobility_disabled __read_mostly
;
169 static void set_pageblock_migratetype(struct page
*page
, int migratetype
)
172 if (unlikely(page_group_by_mobility_disabled
))
173 migratetype
= MIGRATE_UNMOVABLE
;
175 set_pageblock_flags_group(page
, (unsigned long)migratetype
,
176 PB_migrate
, PB_migrate_end
);
179 #ifdef CONFIG_DEBUG_VM
180 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
184 unsigned long pfn
= page_to_pfn(page
);
187 seq
= zone_span_seqbegin(zone
);
188 if (pfn
>= zone
->zone_start_pfn
+ zone
->spanned_pages
)
190 else if (pfn
< zone
->zone_start_pfn
)
192 } while (zone_span_seqretry(zone
, seq
));
197 static int page_is_consistent(struct zone
*zone
, struct page
*page
)
199 if (!pfn_valid_within(page_to_pfn(page
)))
201 if (zone
!= page_zone(page
))
207 * Temporary debugging check for pages not lying within a given zone.
209 static int bad_range(struct zone
*zone
, struct page
*page
)
211 if (page_outside_zone_boundaries(zone
, page
))
213 if (!page_is_consistent(zone
, page
))
219 static inline int bad_range(struct zone
*zone
, struct page
*page
)
225 static void bad_page(struct page
*page
)
227 static unsigned long resume
;
228 static unsigned long nr_shown
;
229 static unsigned long nr_unshown
;
232 * Allow a burst of 60 reports, then keep quiet for that minute;
233 * or allow a steady drip of one report per second.
235 if (nr_shown
== 60) {
236 if (time_before(jiffies
, resume
)) {
242 "BUG: Bad page state: %lu messages suppressed\n",
249 resume
= jiffies
+ 60 * HZ
;
251 printk(KERN_ALERT
"BUG: Bad page state in process %s pfn:%05lx\n",
252 current
->comm
, page_to_pfn(page
));
254 "page:%p flags:%p count:%d mapcount:%d mapping:%p index:%lx\n",
255 page
, (void *)page
->flags
, page_count(page
),
256 page_mapcount(page
), page
->mapping
, page
->index
);
260 /* Leave bad fields for debug, except PageBuddy could make trouble */
261 __ClearPageBuddy(page
);
262 add_taint(TAINT_BAD_PAGE
);
266 * Higher-order pages are called "compound pages". They are structured thusly:
268 * The first PAGE_SIZE page is called the "head page".
270 * The remaining PAGE_SIZE pages are called "tail pages".
272 * All pages have PG_compound set. All pages have their ->private pointing at
273 * the head page (even the head page has this).
275 * The first tail page's ->lru.next holds the address of the compound page's
276 * put_page() function. Its ->lru.prev holds the order of allocation.
277 * This usage means that zero-order pages may not be compound.
280 static void free_compound_page(struct page
*page
)
282 __free_pages_ok(page
, compound_order(page
));
285 void prep_compound_page(struct page
*page
, unsigned long order
)
288 int nr_pages
= 1 << order
;
290 set_compound_page_dtor(page
, free_compound_page
);
291 set_compound_order(page
, order
);
293 for (i
= 1; i
< nr_pages
; i
++) {
294 struct page
*p
= page
+ i
;
297 p
->first_page
= page
;
301 #ifdef CONFIG_HUGETLBFS
302 void prep_compound_gigantic_page(struct page
*page
, unsigned long order
)
305 int nr_pages
= 1 << order
;
306 struct page
*p
= page
+ 1;
308 set_compound_page_dtor(page
, free_compound_page
);
309 set_compound_order(page
, order
);
311 for (i
= 1; i
< nr_pages
; i
++, p
= mem_map_next(p
, page
, i
)) {
313 p
->first_page
= page
;
318 static int destroy_compound_page(struct page
*page
, unsigned long order
)
321 int nr_pages
= 1 << order
;
324 if (unlikely(compound_order(page
) != order
) ||
325 unlikely(!PageHead(page
))) {
330 __ClearPageHead(page
);
332 for (i
= 1; i
< nr_pages
; i
++) {
333 struct page
*p
= page
+ i
;
335 if (unlikely(!PageTail(p
) || (p
->first_page
!= page
))) {
345 static inline void prep_zero_page(struct page
*page
, int order
, gfp_t gfp_flags
)
350 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
351 * and __GFP_HIGHMEM from hard or soft interrupt context.
353 VM_BUG_ON((gfp_flags
& __GFP_HIGHMEM
) && in_interrupt());
354 for (i
= 0; i
< (1 << order
); i
++)
355 clear_highpage(page
+ i
);
358 static inline void set_page_order(struct page
*page
, int order
)
360 set_page_private(page
, order
);
361 __SetPageBuddy(page
);
364 static inline void rmv_page_order(struct page
*page
)
366 __ClearPageBuddy(page
);
367 set_page_private(page
, 0);
371 * Locate the struct page for both the matching buddy in our
372 * pair (buddy1) and the combined O(n+1) page they form (page).
374 * 1) Any buddy B1 will have an order O twin B2 which satisfies
375 * the following equation:
377 * For example, if the starting buddy (buddy2) is #8 its order
379 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
381 * 2) Any buddy B will have an order O+1 parent P which
382 * satisfies the following equation:
385 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
387 static inline struct page
*
388 __page_find_buddy(struct page
*page
, unsigned long page_idx
, unsigned int order
)
390 unsigned long buddy_idx
= page_idx
^ (1 << order
);
392 return page
+ (buddy_idx
- page_idx
);
395 static inline unsigned long
396 __find_combined_index(unsigned long page_idx
, unsigned int order
)
398 return (page_idx
& ~(1 << order
));
402 * This function checks whether a page is free && is the buddy
403 * we can do coalesce a page and its buddy if
404 * (a) the buddy is not in a hole &&
405 * (b) the buddy is in the buddy system &&
406 * (c) a page and its buddy have the same order &&
407 * (d) a page and its buddy are in the same zone.
409 * For recording whether a page is in the buddy system, we use PG_buddy.
410 * Setting, clearing, and testing PG_buddy is serialized by zone->lock.
412 * For recording page's order, we use page_private(page).
414 static inline int page_is_buddy(struct page
*page
, struct page
*buddy
,
417 if (!pfn_valid_within(page_to_pfn(buddy
)))
420 if (page_zone_id(page
) != page_zone_id(buddy
))
423 if (PageBuddy(buddy
) && page_order(buddy
) == order
) {
424 VM_BUG_ON(page_count(buddy
) != 0);
431 * Freeing function for a buddy system allocator.
433 * The concept of a buddy system is to maintain direct-mapped table
434 * (containing bit values) for memory blocks of various "orders".
435 * The bottom level table contains the map for the smallest allocatable
436 * units of memory (here, pages), and each level above it describes
437 * pairs of units from the levels below, hence, "buddies".
438 * At a high level, all that happens here is marking the table entry
439 * at the bottom level available, and propagating the changes upward
440 * as necessary, plus some accounting needed to play nicely with other
441 * parts of the VM system.
442 * At each level, we keep a list of pages, which are heads of continuous
443 * free pages of length of (1 << order) and marked with PG_buddy. Page's
444 * order is recorded in page_private(page) field.
445 * So when we are allocating or freeing one, we can derive the state of the
446 * other. That is, if we allocate a small block, and both were
447 * free, the remainder of the region must be split into blocks.
448 * If a block is freed, and its buddy is also free, then this
449 * triggers coalescing into a block of larger size.
454 static inline void __free_one_page(struct page
*page
,
455 struct zone
*zone
, unsigned int order
,
458 unsigned long page_idx
;
459 int order_size
= 1 << order
;
461 if (unlikely(PageCompound(page
)))
462 if (unlikely(destroy_compound_page(page
, order
)))
465 VM_BUG_ON(migratetype
== -1);
467 page_idx
= page_to_pfn(page
) & ((1 << MAX_ORDER
) - 1);
469 VM_BUG_ON(page_idx
& (order_size
- 1));
470 VM_BUG_ON(bad_range(zone
, page
));
472 __mod_zone_page_state(zone
, NR_FREE_PAGES
, order_size
);
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 static inline int free_pages_check(struct page
*page
)
498 if (unlikely(page_mapcount(page
) |
499 (page
->mapping
!= NULL
) |
500 (atomic_read(&page
->_count
) != 0) |
501 (page
->flags
& PAGE_FLAGS_CHECK_AT_FREE
))) {
505 if (page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)
506 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
511 * Frees a list of pages.
512 * Assumes all pages on list are in same zone, and of same order.
513 * count is the number of pages to free.
515 * If the zone was previously in an "all pages pinned" state then look to
516 * see if this freeing clears that state.
518 * And clear the zone's pages_scanned counter, to hold off the "all pages are
519 * pinned" detection logic.
521 static void free_pages_bulk(struct zone
*zone
, int count
,
522 struct list_head
*list
, int order
)
524 spin_lock(&zone
->lock
);
525 zone_clear_flag(zone
, ZONE_ALL_UNRECLAIMABLE
);
526 zone
->pages_scanned
= 0;
530 VM_BUG_ON(list_empty(list
));
531 page
= list_entry(list
->prev
, struct page
, lru
);
532 /* have to delete it as __free_one_page list manipulates */
533 list_del(&page
->lru
);
534 __free_one_page(page
, zone
, order
, page_private(page
));
536 spin_unlock(&zone
->lock
);
539 static void free_one_page(struct zone
*zone
, struct page
*page
, int order
,
542 spin_lock(&zone
->lock
);
543 zone_clear_flag(zone
, ZONE_ALL_UNRECLAIMABLE
);
544 zone
->pages_scanned
= 0;
545 __free_one_page(page
, zone
, order
, migratetype
);
546 spin_unlock(&zone
->lock
);
549 static void __free_pages_ok(struct page
*page
, unsigned int order
)
554 int clearMlocked
= PageMlocked(page
);
556 for (i
= 0 ; i
< (1 << order
) ; ++i
)
557 bad
+= free_pages_check(page
+ i
);
561 if (!PageHighMem(page
)) {
562 debug_check_no_locks_freed(page_address(page
),PAGE_SIZE
<<order
);
563 debug_check_no_obj_freed(page_address(page
),
566 arch_free_page(page
, order
);
567 kernel_map_pages(page
, 1 << order
, 0);
569 local_irq_save(flags
);
570 if (unlikely(clearMlocked
))
571 free_page_mlock(page
);
572 __count_vm_events(PGFREE
, 1 << order
);
573 free_one_page(page_zone(page
), page
, order
,
574 get_pageblock_migratetype(page
));
575 local_irq_restore(flags
);
579 * permit the bootmem allocator to evade page validation on high-order frees
581 void __meminit
__free_pages_bootmem(struct page
*page
, unsigned int order
)
584 __ClearPageReserved(page
);
585 set_page_count(page
, 0);
586 set_page_refcounted(page
);
592 for (loop
= 0; loop
< BITS_PER_LONG
; loop
++) {
593 struct page
*p
= &page
[loop
];
595 if (loop
+ 1 < BITS_PER_LONG
)
597 __ClearPageReserved(p
);
598 set_page_count(p
, 0);
601 set_page_refcounted(page
);
602 __free_pages(page
, order
);
608 * The order of subdivision here is critical for the IO subsystem.
609 * Please do not alter this order without good reasons and regression
610 * testing. Specifically, as large blocks of memory are subdivided,
611 * the order in which smaller blocks are delivered depends on the order
612 * they're subdivided in this function. This is the primary factor
613 * influencing the order in which pages are delivered to the IO
614 * subsystem according to empirical testing, and this is also justified
615 * by considering the behavior of a buddy system containing a single
616 * large block of memory acted on by a series of small allocations.
617 * This behavior is a critical factor in sglist merging's success.
621 static inline void expand(struct zone
*zone
, struct page
*page
,
622 int low
, int high
, struct free_area
*area
,
625 unsigned long size
= 1 << high
;
631 VM_BUG_ON(bad_range(zone
, &page
[size
]));
632 list_add(&page
[size
].lru
, &area
->free_list
[migratetype
]);
634 set_page_order(&page
[size
], high
);
639 * This page is about to be returned from the page allocator
641 static int prep_new_page(struct page
*page
, int order
, gfp_t gfp_flags
)
643 if (unlikely(page_mapcount(page
) |
644 (page
->mapping
!= NULL
) |
645 (atomic_read(&page
->_count
) != 0) |
646 (page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
))) {
651 set_page_private(page
, 0);
652 set_page_refcounted(page
);
654 arch_alloc_page(page
, order
);
655 kernel_map_pages(page
, 1 << order
, 1);
657 if (gfp_flags
& __GFP_ZERO
)
658 prep_zero_page(page
, order
, gfp_flags
);
660 if (order
&& (gfp_flags
& __GFP_COMP
))
661 prep_compound_page(page
, order
);
667 * Go through the free lists for the given migratetype and remove
668 * the smallest available page from the freelists
671 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
674 unsigned int current_order
;
675 struct free_area
* area
;
678 /* Find a page of the appropriate size in the preferred list */
679 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
680 area
= &(zone
->free_area
[current_order
]);
681 if (list_empty(&area
->free_list
[migratetype
]))
684 page
= list_entry(area
->free_list
[migratetype
].next
,
686 list_del(&page
->lru
);
687 rmv_page_order(page
);
689 __mod_zone_page_state(zone
, NR_FREE_PAGES
, - (1UL << order
));
690 expand(zone
, page
, order
, current_order
, area
, migratetype
);
699 * This array describes the order lists are fallen back to when
700 * the free lists for the desirable migrate type are depleted
702 static int fallbacks
[MIGRATE_TYPES
][MIGRATE_TYPES
-1] = {
703 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
704 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
705 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_RESERVE
},
706 [MIGRATE_RESERVE
] = { MIGRATE_RESERVE
, MIGRATE_RESERVE
, MIGRATE_RESERVE
}, /* Never used */
710 * Move the free pages in a range to the free lists of the requested type.
711 * Note that start_page and end_pages are not aligned on a pageblock
712 * boundary. If alignment is required, use move_freepages_block()
714 static int move_freepages(struct zone
*zone
,
715 struct page
*start_page
, struct page
*end_page
,
722 #ifndef CONFIG_HOLES_IN_ZONE
724 * page_zone is not safe to call in this context when
725 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
726 * anyway as we check zone boundaries in move_freepages_block().
727 * Remove at a later date when no bug reports exist related to
728 * grouping pages by mobility
730 BUG_ON(page_zone(start_page
) != page_zone(end_page
));
733 for (page
= start_page
; page
<= end_page
;) {
734 /* Make sure we are not inadvertently changing nodes */
735 VM_BUG_ON(page_to_nid(page
) != zone_to_nid(zone
));
737 if (!pfn_valid_within(page_to_pfn(page
))) {
742 if (!PageBuddy(page
)) {
747 order
= page_order(page
);
748 list_del(&page
->lru
);
750 &zone
->free_area
[order
].free_list
[migratetype
]);
752 pages_moved
+= 1 << order
;
758 static int move_freepages_block(struct zone
*zone
, struct page
*page
,
761 unsigned long start_pfn
, end_pfn
;
762 struct page
*start_page
, *end_page
;
764 start_pfn
= page_to_pfn(page
);
765 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
766 start_page
= pfn_to_page(start_pfn
);
767 end_page
= start_page
+ pageblock_nr_pages
- 1;
768 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
770 /* Do not cross zone boundaries */
771 if (start_pfn
< zone
->zone_start_pfn
)
773 if (end_pfn
>= zone
->zone_start_pfn
+ zone
->spanned_pages
)
776 return move_freepages(zone
, start_page
, end_page
, migratetype
);
779 /* Remove an element from the buddy allocator from the fallback list */
780 static inline struct page
*
781 __rmqueue_fallback(struct zone
*zone
, int order
, int start_migratetype
)
783 struct free_area
* area
;
788 /* Find the largest possible block of pages in the other list */
789 for (current_order
= MAX_ORDER
-1; current_order
>= order
;
791 for (i
= 0; i
< MIGRATE_TYPES
- 1; i
++) {
792 migratetype
= fallbacks
[start_migratetype
][i
];
794 /* MIGRATE_RESERVE handled later if necessary */
795 if (migratetype
== MIGRATE_RESERVE
)
798 area
= &(zone
->free_area
[current_order
]);
799 if (list_empty(&area
->free_list
[migratetype
]))
802 page
= list_entry(area
->free_list
[migratetype
].next
,
807 * If breaking a large block of pages, move all free
808 * pages to the preferred allocation list. If falling
809 * back for a reclaimable kernel allocation, be more
810 * agressive about taking ownership of free pages
812 if (unlikely(current_order
>= (pageblock_order
>> 1)) ||
813 start_migratetype
== MIGRATE_RECLAIMABLE
) {
815 pages
= move_freepages_block(zone
, page
,
818 /* Claim the whole block if over half of it is free */
819 if (pages
>= (1 << (pageblock_order
-1)))
820 set_pageblock_migratetype(page
,
823 migratetype
= start_migratetype
;
826 /* Remove the page from the freelists */
827 list_del(&page
->lru
);
828 rmv_page_order(page
);
829 __mod_zone_page_state(zone
, NR_FREE_PAGES
,
832 if (current_order
== pageblock_order
)
833 set_pageblock_migratetype(page
,
836 expand(zone
, page
, order
, current_order
, area
, migratetype
);
845 * Do the hard work of removing an element from the buddy allocator.
846 * Call me with the zone->lock already held.
848 static struct page
*__rmqueue(struct zone
*zone
, unsigned int order
,
854 page
= __rmqueue_smallest(zone
, order
, migratetype
);
856 if (unlikely(!page
) && migratetype
!= MIGRATE_RESERVE
) {
857 page
= __rmqueue_fallback(zone
, order
, migratetype
);
860 * Use MIGRATE_RESERVE rather than fail an allocation. goto
861 * is used because __rmqueue_smallest is an inline function
862 * and we want just one call site
865 migratetype
= MIGRATE_RESERVE
;
874 * Obtain a specified number of elements from the buddy allocator, all under
875 * a single hold of the lock, for efficiency. Add them to the supplied list.
876 * Returns the number of new pages which were placed at *list.
878 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
879 unsigned long count
, struct list_head
*list
,
884 spin_lock(&zone
->lock
);
885 for (i
= 0; i
< count
; ++i
) {
886 struct page
*page
= __rmqueue(zone
, order
, migratetype
);
887 if (unlikely(page
== NULL
))
891 * Split buddy pages returned by expand() are received here
892 * in physical page order. The page is added to the callers and
893 * list and the list head then moves forward. From the callers
894 * perspective, the linked list is ordered by page number in
895 * some conditions. This is useful for IO devices that can
896 * merge IO requests if the physical pages are ordered
899 list_add(&page
->lru
, list
);
900 set_page_private(page
, migratetype
);
903 spin_unlock(&zone
->lock
);
909 * Called from the vmstat counter updater to drain pagesets of this
910 * currently executing processor on remote nodes after they have
913 * Note that this function must be called with the thread pinned to
914 * a single processor.
916 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
921 local_irq_save(flags
);
922 if (pcp
->count
>= pcp
->batch
)
923 to_drain
= pcp
->batch
;
925 to_drain
= pcp
->count
;
926 free_pages_bulk(zone
, to_drain
, &pcp
->list
, 0);
927 pcp
->count
-= to_drain
;
928 local_irq_restore(flags
);
933 * Drain pages of the indicated processor.
935 * The processor must either be the current processor and the
936 * thread pinned to the current processor or a processor that
939 static void drain_pages(unsigned int cpu
)
944 for_each_populated_zone(zone
) {
945 struct per_cpu_pageset
*pset
;
946 struct per_cpu_pages
*pcp
;
948 pset
= zone_pcp(zone
, cpu
);
951 local_irq_save(flags
);
952 free_pages_bulk(zone
, pcp
->count
, &pcp
->list
, 0);
954 local_irq_restore(flags
);
959 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
961 void drain_local_pages(void *arg
)
963 drain_pages(smp_processor_id());
967 * Spill all the per-cpu pages from all CPUs back into the buddy allocator
969 void drain_all_pages(void)
971 on_each_cpu(drain_local_pages
, NULL
, 1);
974 #ifdef CONFIG_HIBERNATION
976 void mark_free_pages(struct zone
*zone
)
978 unsigned long pfn
, max_zone_pfn
;
981 struct list_head
*curr
;
983 if (!zone
->spanned_pages
)
986 spin_lock_irqsave(&zone
->lock
, flags
);
988 max_zone_pfn
= zone
->zone_start_pfn
+ zone
->spanned_pages
;
989 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
990 if (pfn_valid(pfn
)) {
991 struct page
*page
= pfn_to_page(pfn
);
993 if (!swsusp_page_is_forbidden(page
))
994 swsusp_unset_page_free(page
);
997 for_each_migratetype_order(order
, t
) {
998 list_for_each(curr
, &zone
->free_area
[order
].free_list
[t
]) {
1001 pfn
= page_to_pfn(list_entry(curr
, struct page
, lru
));
1002 for (i
= 0; i
< (1UL << order
); i
++)
1003 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
1006 spin_unlock_irqrestore(&zone
->lock
, flags
);
1008 #endif /* CONFIG_PM */
1011 * Free a 0-order page
1013 static void free_hot_cold_page(struct page
*page
, int cold
)
1015 struct zone
*zone
= page_zone(page
);
1016 struct per_cpu_pages
*pcp
;
1017 unsigned long flags
;
1018 int clearMlocked
= PageMlocked(page
);
1021 page
->mapping
= NULL
;
1022 if (free_pages_check(page
))
1025 if (!PageHighMem(page
)) {
1026 debug_check_no_locks_freed(page_address(page
), PAGE_SIZE
);
1027 debug_check_no_obj_freed(page_address(page
), PAGE_SIZE
);
1029 arch_free_page(page
, 0);
1030 kernel_map_pages(page
, 1, 0);
1032 pcp
= &zone_pcp(zone
, get_cpu())->pcp
;
1033 local_irq_save(flags
);
1034 if (unlikely(clearMlocked
))
1035 free_page_mlock(page
);
1036 __count_vm_event(PGFREE
);
1039 list_add_tail(&page
->lru
, &pcp
->list
);
1041 list_add(&page
->lru
, &pcp
->list
);
1042 set_page_private(page
, get_pageblock_migratetype(page
));
1044 if (pcp
->count
>= pcp
->high
) {
1045 free_pages_bulk(zone
, pcp
->batch
, &pcp
->list
, 0);
1046 pcp
->count
-= pcp
->batch
;
1048 local_irq_restore(flags
);
1052 void free_hot_page(struct page
*page
)
1054 free_hot_cold_page(page
, 0);
1057 void free_cold_page(struct page
*page
)
1059 free_hot_cold_page(page
, 1);
1063 * split_page takes a non-compound higher-order page, and splits it into
1064 * n (1<<order) sub-pages: page[0..n]
1065 * Each sub-page must be freed individually.
1067 * Note: this is probably too low level an operation for use in drivers.
1068 * Please consult with lkml before using this in your driver.
1070 void split_page(struct page
*page
, unsigned int order
)
1074 VM_BUG_ON(PageCompound(page
));
1075 VM_BUG_ON(!page_count(page
));
1076 for (i
= 1; i
< (1 << order
); i
++)
1077 set_page_refcounted(page
+ i
);
1081 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1082 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1086 struct page
*buffered_rmqueue(struct zone
*preferred_zone
,
1087 struct zone
*zone
, int order
, gfp_t gfp_flags
,
1090 unsigned long flags
;
1092 int cold
= !!(gfp_flags
& __GFP_COLD
);
1097 if (likely(order
== 0)) {
1098 struct per_cpu_pages
*pcp
;
1100 pcp
= &zone_pcp(zone
, cpu
)->pcp
;
1101 local_irq_save(flags
);
1103 pcp
->count
= rmqueue_bulk(zone
, 0,
1104 pcp
->batch
, &pcp
->list
, migratetype
);
1105 if (unlikely(!pcp
->count
))
1109 /* Find a page of the appropriate migrate type */
1111 list_for_each_entry_reverse(page
, &pcp
->list
, lru
)
1112 if (page_private(page
) == migratetype
)
1115 list_for_each_entry(page
, &pcp
->list
, lru
)
1116 if (page_private(page
) == migratetype
)
1120 /* Allocate more to the pcp list if necessary */
1121 if (unlikely(&page
->lru
== &pcp
->list
)) {
1122 pcp
->count
+= rmqueue_bulk(zone
, 0,
1123 pcp
->batch
, &pcp
->list
, migratetype
);
1124 page
= list_entry(pcp
->list
.next
, struct page
, lru
);
1127 list_del(&page
->lru
);
1130 spin_lock_irqsave(&zone
->lock
, flags
);
1131 page
= __rmqueue(zone
, order
, migratetype
);
1132 spin_unlock(&zone
->lock
);
1137 __count_zone_vm_events(PGALLOC
, zone
, 1 << order
);
1138 zone_statistics(preferred_zone
, zone
);
1139 local_irq_restore(flags
);
1142 VM_BUG_ON(bad_range(zone
, page
));
1143 if (prep_new_page(page
, order
, gfp_flags
))
1148 local_irq_restore(flags
);
1153 /* The ALLOC_WMARK bits are used as an index to zone->watermark */
1154 #define ALLOC_WMARK_MIN WMARK_MIN
1155 #define ALLOC_WMARK_LOW WMARK_LOW
1156 #define ALLOC_WMARK_HIGH WMARK_HIGH
1157 #define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */
1159 /* Mask to get the watermark bits */
1160 #define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1)
1162 #define ALLOC_HARDER 0x10 /* try to alloc harder */
1163 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
1164 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
1166 #ifdef CONFIG_FAIL_PAGE_ALLOC
1168 static struct fail_page_alloc_attr
{
1169 struct fault_attr attr
;
1171 u32 ignore_gfp_highmem
;
1172 u32 ignore_gfp_wait
;
1175 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1177 struct dentry
*ignore_gfp_highmem_file
;
1178 struct dentry
*ignore_gfp_wait_file
;
1179 struct dentry
*min_order_file
;
1181 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1183 } fail_page_alloc
= {
1184 .attr
= FAULT_ATTR_INITIALIZER
,
1185 .ignore_gfp_wait
= 1,
1186 .ignore_gfp_highmem
= 1,
1190 static int __init
setup_fail_page_alloc(char *str
)
1192 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
1194 __setup("fail_page_alloc=", setup_fail_page_alloc
);
1196 static int should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1198 if (order
< fail_page_alloc
.min_order
)
1200 if (gfp_mask
& __GFP_NOFAIL
)
1202 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
1204 if (fail_page_alloc
.ignore_gfp_wait
&& (gfp_mask
& __GFP_WAIT
))
1207 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
1210 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1212 static int __init
fail_page_alloc_debugfs(void)
1214 mode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
1218 err
= init_fault_attr_dentries(&fail_page_alloc
.attr
,
1222 dir
= fail_page_alloc
.attr
.dentries
.dir
;
1224 fail_page_alloc
.ignore_gfp_wait_file
=
1225 debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
1226 &fail_page_alloc
.ignore_gfp_wait
);
1228 fail_page_alloc
.ignore_gfp_highmem_file
=
1229 debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
1230 &fail_page_alloc
.ignore_gfp_highmem
);
1231 fail_page_alloc
.min_order_file
=
1232 debugfs_create_u32("min-order", mode
, dir
,
1233 &fail_page_alloc
.min_order
);
1235 if (!fail_page_alloc
.ignore_gfp_wait_file
||
1236 !fail_page_alloc
.ignore_gfp_highmem_file
||
1237 !fail_page_alloc
.min_order_file
) {
1239 debugfs_remove(fail_page_alloc
.ignore_gfp_wait_file
);
1240 debugfs_remove(fail_page_alloc
.ignore_gfp_highmem_file
);
1241 debugfs_remove(fail_page_alloc
.min_order_file
);
1242 cleanup_fault_attr_dentries(&fail_page_alloc
.attr
);
1248 late_initcall(fail_page_alloc_debugfs
);
1250 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1252 #else /* CONFIG_FAIL_PAGE_ALLOC */
1254 static inline int should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1259 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1262 * Return 1 if free pages are above 'mark'. This takes into account the order
1263 * of the allocation.
1265 int zone_watermark_ok(struct zone
*z
, int order
, unsigned long mark
,
1266 int classzone_idx
, int alloc_flags
)
1268 /* free_pages my go negative - that's OK */
1270 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
) - (1 << order
) + 1;
1273 if (alloc_flags
& ALLOC_HIGH
)
1275 if (alloc_flags
& ALLOC_HARDER
)
1278 if (free_pages
<= min
+ z
->lowmem_reserve
[classzone_idx
])
1280 for (o
= 0; o
< order
; o
++) {
1281 /* At the next order, this order's pages become unavailable */
1282 free_pages
-= z
->free_area
[o
].nr_free
<< o
;
1284 /* Require fewer higher order pages to be free */
1287 if (free_pages
<= min
)
1295 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1296 * skip over zones that are not allowed by the cpuset, or that have
1297 * been recently (in last second) found to be nearly full. See further
1298 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1299 * that have to skip over a lot of full or unallowed zones.
1301 * If the zonelist cache is present in the passed in zonelist, then
1302 * returns a pointer to the allowed node mask (either the current
1303 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1305 * If the zonelist cache is not available for this zonelist, does
1306 * nothing and returns NULL.
1308 * If the fullzones BITMAP in the zonelist cache is stale (more than
1309 * a second since last zap'd) then we zap it out (clear its bits.)
1311 * We hold off even calling zlc_setup, until after we've checked the
1312 * first zone in the zonelist, on the theory that most allocations will
1313 * be satisfied from that first zone, so best to examine that zone as
1314 * quickly as we can.
1316 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1318 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1319 nodemask_t
*allowednodes
; /* zonelist_cache approximation */
1321 zlc
= zonelist
->zlcache_ptr
;
1325 if (time_after(jiffies
, zlc
->last_full_zap
+ HZ
)) {
1326 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
1327 zlc
->last_full_zap
= jiffies
;
1330 allowednodes
= !in_interrupt() && (alloc_flags
& ALLOC_CPUSET
) ?
1331 &cpuset_current_mems_allowed
:
1332 &node_states
[N_HIGH_MEMORY
];
1333 return allowednodes
;
1337 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1338 * if it is worth looking at further for free memory:
1339 * 1) Check that the zone isn't thought to be full (doesn't have its
1340 * bit set in the zonelist_cache fullzones BITMAP).
1341 * 2) Check that the zones node (obtained from the zonelist_cache
1342 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1343 * Return true (non-zero) if zone is worth looking at further, or
1344 * else return false (zero) if it is not.
1346 * This check -ignores- the distinction between various watermarks,
1347 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1348 * found to be full for any variation of these watermarks, it will
1349 * be considered full for up to one second by all requests, unless
1350 * we are so low on memory on all allowed nodes that we are forced
1351 * into the second scan of the zonelist.
1353 * In the second scan we ignore this zonelist cache and exactly
1354 * apply the watermarks to all zones, even it is slower to do so.
1355 * We are low on memory in the second scan, and should leave no stone
1356 * unturned looking for a free page.
1358 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1359 nodemask_t
*allowednodes
)
1361 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1362 int i
; /* index of *z in zonelist zones */
1363 int n
; /* node that zone *z is on */
1365 zlc
= zonelist
->zlcache_ptr
;
1369 i
= z
- zonelist
->_zonerefs
;
1372 /* This zone is worth trying if it is allowed but not full */
1373 return node_isset(n
, *allowednodes
) && !test_bit(i
, zlc
->fullzones
);
1377 * Given 'z' scanning a zonelist, set the corresponding bit in
1378 * zlc->fullzones, so that subsequent attempts to allocate a page
1379 * from that zone don't waste time re-examining it.
1381 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1383 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1384 int i
; /* index of *z in zonelist zones */
1386 zlc
= zonelist
->zlcache_ptr
;
1390 i
= z
- zonelist
->_zonerefs
;
1392 set_bit(i
, zlc
->fullzones
);
1395 #else /* CONFIG_NUMA */
1397 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1402 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1403 nodemask_t
*allowednodes
)
1408 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1411 #endif /* CONFIG_NUMA */
1414 * get_page_from_freelist goes through the zonelist trying to allocate
1417 static struct page
*
1418 get_page_from_freelist(gfp_t gfp_mask
, nodemask_t
*nodemask
, unsigned int order
,
1419 struct zonelist
*zonelist
, int high_zoneidx
, int alloc_flags
,
1420 struct zone
*preferred_zone
, int migratetype
)
1423 struct page
*page
= NULL
;
1426 nodemask_t
*allowednodes
= NULL
;/* zonelist_cache approximation */
1427 int zlc_active
= 0; /* set if using zonelist_cache */
1428 int did_zlc_setup
= 0; /* just call zlc_setup() one time */
1430 if (WARN_ON_ONCE(order
>= MAX_ORDER
))
1433 classzone_idx
= zone_idx(preferred_zone
);
1436 * Scan zonelist, looking for a zone with enough free.
1437 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1439 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
1440 high_zoneidx
, nodemask
) {
1441 if (NUMA_BUILD
&& zlc_active
&&
1442 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
1444 if ((alloc_flags
& ALLOC_CPUSET
) &&
1445 !cpuset_zone_allowed_softwall(zone
, gfp_mask
))
1448 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
1449 if (!(alloc_flags
& ALLOC_NO_WATERMARKS
)) {
1451 mark
= zone
->watermark
[alloc_flags
& ALLOC_WMARK_MASK
];
1452 if (!zone_watermark_ok(zone
, order
, mark
,
1453 classzone_idx
, alloc_flags
)) {
1454 if (!zone_reclaim_mode
||
1455 !zone_reclaim(zone
, gfp_mask
, order
))
1456 goto this_zone_full
;
1460 page
= buffered_rmqueue(preferred_zone
, zone
, order
,
1461 gfp_mask
, migratetype
);
1466 zlc_mark_zone_full(zonelist
, z
);
1468 if (NUMA_BUILD
&& !did_zlc_setup
&& num_online_nodes() > 1) {
1470 * we do zlc_setup after the first zone is tried but only
1471 * if there are multiple nodes make it worthwhile
1473 allowednodes
= zlc_setup(zonelist
, alloc_flags
);
1479 if (unlikely(NUMA_BUILD
&& page
== NULL
&& zlc_active
)) {
1480 /* Disable zlc cache for second zonelist scan */
1488 should_alloc_retry(gfp_t gfp_mask
, unsigned int order
,
1489 unsigned long pages_reclaimed
)
1491 /* Do not loop if specifically requested */
1492 if (gfp_mask
& __GFP_NORETRY
)
1496 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
1497 * means __GFP_NOFAIL, but that may not be true in other
1500 if (order
<= PAGE_ALLOC_COSTLY_ORDER
)
1504 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
1505 * specified, then we retry until we no longer reclaim any pages
1506 * (above), or we've reclaimed an order of pages at least as
1507 * large as the allocation's order. In both cases, if the
1508 * allocation still fails, we stop retrying.
1510 if (gfp_mask
& __GFP_REPEAT
&& pages_reclaimed
< (1 << order
))
1514 * Don't let big-order allocations loop unless the caller
1515 * explicitly requests that.
1517 if (gfp_mask
& __GFP_NOFAIL
)
1523 static inline struct page
*
1524 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
1525 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1526 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
1531 /* Acquire the OOM killer lock for the zones in zonelist */
1532 if (!try_set_zone_oom(zonelist
, gfp_mask
)) {
1533 schedule_timeout_uninterruptible(1);
1538 * Go through the zonelist yet one more time, keep very high watermark
1539 * here, this is only to catch a parallel oom killing, we must fail if
1540 * we're still under heavy pressure.
1542 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
,
1543 order
, zonelist
, high_zoneidx
,
1544 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
,
1545 preferred_zone
, migratetype
);
1549 /* The OOM killer will not help higher order allocs */
1550 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
1553 /* Exhausted what can be done so it's blamo time */
1554 out_of_memory(zonelist
, gfp_mask
, order
);
1557 clear_zonelist_oom(zonelist
, gfp_mask
);
1561 /* The really slow allocator path where we enter direct reclaim */
1562 static inline struct page
*
1563 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
1564 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1565 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
1566 int migratetype
, unsigned long *did_some_progress
)
1568 struct page
*page
= NULL
;
1569 struct reclaim_state reclaim_state
;
1570 struct task_struct
*p
= current
;
1574 /* We now go into synchronous reclaim */
1575 cpuset_memory_pressure_bump();
1578 * The task's cpuset might have expanded its set of allowable nodes
1580 p
->flags
|= PF_MEMALLOC
;
1581 lockdep_set_current_reclaim_state(gfp_mask
);
1582 reclaim_state
.reclaimed_slab
= 0;
1583 p
->reclaim_state
= &reclaim_state
;
1585 *did_some_progress
= try_to_free_pages(zonelist
, order
, gfp_mask
, nodemask
);
1587 p
->reclaim_state
= NULL
;
1588 lockdep_clear_current_reclaim_state();
1589 p
->flags
&= ~PF_MEMALLOC
;
1596 if (likely(*did_some_progress
))
1597 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
1598 zonelist
, high_zoneidx
,
1599 alloc_flags
, preferred_zone
,
1605 * This is called in the allocator slow-path if the allocation request is of
1606 * sufficient urgency to ignore watermarks and take other desperate measures
1608 static inline struct page
*
1609 __alloc_pages_high_priority(gfp_t gfp_mask
, unsigned int order
,
1610 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1611 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
1617 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
1618 zonelist
, high_zoneidx
, ALLOC_NO_WATERMARKS
,
1619 preferred_zone
, migratetype
);
1621 if (!page
&& gfp_mask
& __GFP_NOFAIL
)
1622 congestion_wait(WRITE
, HZ
/50);
1623 } while (!page
&& (gfp_mask
& __GFP_NOFAIL
));
1629 void wake_all_kswapd(unsigned int order
, struct zonelist
*zonelist
,
1630 enum zone_type high_zoneidx
)
1635 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
)
1636 wakeup_kswapd(zone
, order
);
1640 gfp_to_alloc_flags(gfp_t gfp_mask
)
1642 struct task_struct
*p
= current
;
1643 int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
1644 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
1646 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
1647 BUILD_BUG_ON(__GFP_HIGH
!= ALLOC_HIGH
);
1650 * The caller may dip into page reserves a bit more if the caller
1651 * cannot run direct reclaim, or if the caller has realtime scheduling
1652 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
1653 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
1655 alloc_flags
|= (gfp_mask
& __GFP_HIGH
);
1658 alloc_flags
|= ALLOC_HARDER
;
1660 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
1661 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1663 alloc_flags
&= ~ALLOC_CPUSET
;
1664 } else if (unlikely(rt_task(p
)))
1665 alloc_flags
|= ALLOC_HARDER
;
1667 if (likely(!(gfp_mask
& __GFP_NOMEMALLOC
))) {
1668 if (!in_interrupt() &&
1669 ((p
->flags
& PF_MEMALLOC
) ||
1670 unlikely(test_thread_flag(TIF_MEMDIE
))))
1671 alloc_flags
|= ALLOC_NO_WATERMARKS
;
1677 static inline struct page
*
1678 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
1679 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1680 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
1683 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
1684 struct page
*page
= NULL
;
1686 unsigned long pages_reclaimed
= 0;
1687 unsigned long did_some_progress
;
1688 struct task_struct
*p
= current
;
1691 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
1692 * __GFP_NOWARN set) should not cause reclaim since the subsystem
1693 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
1694 * using a larger set of nodes after it has established that the
1695 * allowed per node queues are empty and that nodes are
1698 if (NUMA_BUILD
&& (gfp_mask
& GFP_THISNODE
) == GFP_THISNODE
)
1701 wake_all_kswapd(order
, zonelist
, high_zoneidx
);
1704 * OK, we're below the kswapd watermark and have kicked background
1705 * reclaim. Now things get more complex, so set up alloc_flags according
1706 * to how we want to proceed.
1708 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
1711 /* This is the last chance, in general, before the goto nopage. */
1712 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
, zonelist
,
1713 high_zoneidx
, alloc_flags
& ~ALLOC_NO_WATERMARKS
,
1714 preferred_zone
, migratetype
);
1719 /* Allocate without watermarks if the context allows */
1720 if (alloc_flags
& ALLOC_NO_WATERMARKS
) {
1721 page
= __alloc_pages_high_priority(gfp_mask
, order
,
1722 zonelist
, high_zoneidx
, nodemask
,
1723 preferred_zone
, migratetype
);
1728 /* Atomic allocations - we can't balance anything */
1732 /* Avoid recursion of direct reclaim */
1733 if (p
->flags
& PF_MEMALLOC
)
1736 /* Try direct reclaim and then allocating */
1737 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
,
1738 zonelist
, high_zoneidx
,
1740 alloc_flags
, preferred_zone
,
1741 migratetype
, &did_some_progress
);
1746 * If we failed to make any progress reclaiming, then we are
1747 * running out of options and have to consider going OOM
1749 if (!did_some_progress
) {
1750 if ((gfp_mask
& __GFP_FS
) && !(gfp_mask
& __GFP_NORETRY
)) {
1751 page
= __alloc_pages_may_oom(gfp_mask
, order
,
1752 zonelist
, high_zoneidx
,
1753 nodemask
, preferred_zone
,
1759 * The OOM killer does not trigger for high-order allocations
1760 * but if no progress is being made, there are no other
1761 * options and retrying is unlikely to help
1763 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
1770 /* Check if we should retry the allocation */
1771 pages_reclaimed
+= did_some_progress
;
1772 if (should_alloc_retry(gfp_mask
, order
, pages_reclaimed
)) {
1773 /* Wait for some write requests to complete then retry */
1774 congestion_wait(WRITE
, HZ
/50);
1779 if (!(gfp_mask
& __GFP_NOWARN
) && printk_ratelimit()) {
1780 printk(KERN_WARNING
"%s: page allocation failure."
1781 " order:%d, mode:0x%x\n",
1782 p
->comm
, order
, gfp_mask
);
1792 * This is the 'heart' of the zoned buddy allocator.
1795 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
,
1796 struct zonelist
*zonelist
, nodemask_t
*nodemask
)
1798 enum zone_type high_zoneidx
= gfp_zone(gfp_mask
);
1799 struct zone
*preferred_zone
;
1801 int migratetype
= allocflags_to_migratetype(gfp_mask
);
1803 lockdep_trace_alloc(gfp_mask
);
1805 might_sleep_if(gfp_mask
& __GFP_WAIT
);
1807 if (should_fail_alloc_page(gfp_mask
, order
))
1811 * Check the zones suitable for the gfp_mask contain at least one
1812 * valid zone. It's possible to have an empty zonelist as a result
1813 * of GFP_THISNODE and a memoryless node
1815 if (unlikely(!zonelist
->_zonerefs
->zone
))
1818 /* The preferred zone is used for statistics later */
1819 first_zones_zonelist(zonelist
, high_zoneidx
, nodemask
, &preferred_zone
);
1820 if (!preferred_zone
)
1823 /* First allocation attempt */
1824 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
, order
,
1825 zonelist
, high_zoneidx
, ALLOC_WMARK_LOW
|ALLOC_CPUSET
,
1826 preferred_zone
, migratetype
);
1827 if (unlikely(!page
))
1828 page
= __alloc_pages_slowpath(gfp_mask
, order
,
1829 zonelist
, high_zoneidx
, nodemask
,
1830 preferred_zone
, migratetype
);
1834 EXPORT_SYMBOL(__alloc_pages_nodemask
);
1837 * Common helper functions.
1839 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
1842 page
= alloc_pages(gfp_mask
, order
);
1845 return (unsigned long) page_address(page
);
1848 EXPORT_SYMBOL(__get_free_pages
);
1850 unsigned long get_zeroed_page(gfp_t gfp_mask
)
1855 * get_zeroed_page() returns a 32-bit address, which cannot represent
1858 VM_BUG_ON((gfp_mask
& __GFP_HIGHMEM
) != 0);
1860 page
= alloc_pages(gfp_mask
| __GFP_ZERO
, 0);
1862 return (unsigned long) page_address(page
);
1866 EXPORT_SYMBOL(get_zeroed_page
);
1868 void __pagevec_free(struct pagevec
*pvec
)
1870 int i
= pagevec_count(pvec
);
1873 free_hot_cold_page(pvec
->pages
[i
], pvec
->cold
);
1876 void __free_pages(struct page
*page
, unsigned int order
)
1878 if (put_page_testzero(page
)) {
1880 free_hot_page(page
);
1882 __free_pages_ok(page
, order
);
1886 EXPORT_SYMBOL(__free_pages
);
1888 void free_pages(unsigned long addr
, unsigned int order
)
1891 VM_BUG_ON(!virt_addr_valid((void *)addr
));
1892 __free_pages(virt_to_page((void *)addr
), order
);
1896 EXPORT_SYMBOL(free_pages
);
1899 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
1900 * @size: the number of bytes to allocate
1901 * @gfp_mask: GFP flags for the allocation
1903 * This function is similar to alloc_pages(), except that it allocates the
1904 * minimum number of pages to satisfy the request. alloc_pages() can only
1905 * allocate memory in power-of-two pages.
1907 * This function is also limited by MAX_ORDER.
1909 * Memory allocated by this function must be released by free_pages_exact().
1911 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
1913 unsigned int order
= get_order(size
);
1916 addr
= __get_free_pages(gfp_mask
, order
);
1918 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
1919 unsigned long used
= addr
+ PAGE_ALIGN(size
);
1921 split_page(virt_to_page(addr
), order
);
1922 while (used
< alloc_end
) {
1928 return (void *)addr
;
1930 EXPORT_SYMBOL(alloc_pages_exact
);
1933 * free_pages_exact - release memory allocated via alloc_pages_exact()
1934 * @virt: the value returned by alloc_pages_exact.
1935 * @size: size of allocation, same value as passed to alloc_pages_exact().
1937 * Release the memory allocated by a previous call to alloc_pages_exact.
1939 void free_pages_exact(void *virt
, size_t size
)
1941 unsigned long addr
= (unsigned long)virt
;
1942 unsigned long end
= addr
+ PAGE_ALIGN(size
);
1944 while (addr
< end
) {
1949 EXPORT_SYMBOL(free_pages_exact
);
1951 static unsigned int nr_free_zone_pages(int offset
)
1956 /* Just pick one node, since fallback list is circular */
1957 unsigned int sum
= 0;
1959 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
1961 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
1962 unsigned long size
= zone
->present_pages
;
1963 unsigned long high
= high_wmark_pages(zone
);
1972 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
1974 unsigned int nr_free_buffer_pages(void)
1976 return nr_free_zone_pages(gfp_zone(GFP_USER
));
1978 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
1981 * Amount of free RAM allocatable within all zones
1983 unsigned int nr_free_pagecache_pages(void)
1985 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
1988 static inline void show_node(struct zone
*zone
)
1991 printk("Node %d ", zone_to_nid(zone
));
1994 void si_meminfo(struct sysinfo
*val
)
1996 val
->totalram
= totalram_pages
;
1998 val
->freeram
= global_page_state(NR_FREE_PAGES
);
1999 val
->bufferram
= nr_blockdev_pages();
2000 val
->totalhigh
= totalhigh_pages
;
2001 val
->freehigh
= nr_free_highpages();
2002 val
->mem_unit
= PAGE_SIZE
;
2005 EXPORT_SYMBOL(si_meminfo
);
2008 void si_meminfo_node(struct sysinfo
*val
, int nid
)
2010 pg_data_t
*pgdat
= NODE_DATA(nid
);
2012 val
->totalram
= pgdat
->node_present_pages
;
2013 val
->freeram
= node_page_state(nid
, NR_FREE_PAGES
);
2014 #ifdef CONFIG_HIGHMEM
2015 val
->totalhigh
= pgdat
->node_zones
[ZONE_HIGHMEM
].present_pages
;
2016 val
->freehigh
= zone_page_state(&pgdat
->node_zones
[ZONE_HIGHMEM
],
2022 val
->mem_unit
= PAGE_SIZE
;
2026 #define K(x) ((x) << (PAGE_SHIFT-10))
2029 * Show free area list (used inside shift_scroll-lock stuff)
2030 * We also calculate the percentage fragmentation. We do this by counting the
2031 * memory on each free list with the exception of the first item on the list.
2033 void show_free_areas(void)
2038 for_each_populated_zone(zone
) {
2040 printk("%s per-cpu:\n", zone
->name
);
2042 for_each_online_cpu(cpu
) {
2043 struct per_cpu_pageset
*pageset
;
2045 pageset
= zone_pcp(zone
, cpu
);
2047 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2048 cpu
, pageset
->pcp
.high
,
2049 pageset
->pcp
.batch
, pageset
->pcp
.count
);
2053 printk("Active_anon:%lu active_file:%lu inactive_anon:%lu\n"
2054 " inactive_file:%lu"
2055 //TODO: check/adjust line lengths
2056 #ifdef CONFIG_UNEVICTABLE_LRU
2059 " dirty:%lu writeback:%lu unstable:%lu\n"
2060 " free:%lu slab:%lu mapped:%lu pagetables:%lu bounce:%lu\n",
2061 global_page_state(NR_ACTIVE_ANON
),
2062 global_page_state(NR_ACTIVE_FILE
),
2063 global_page_state(NR_INACTIVE_ANON
),
2064 global_page_state(NR_INACTIVE_FILE
),
2065 #ifdef CONFIG_UNEVICTABLE_LRU
2066 global_page_state(NR_UNEVICTABLE
),
2068 global_page_state(NR_FILE_DIRTY
),
2069 global_page_state(NR_WRITEBACK
),
2070 global_page_state(NR_UNSTABLE_NFS
),
2071 global_page_state(NR_FREE_PAGES
),
2072 global_page_state(NR_SLAB_RECLAIMABLE
) +
2073 global_page_state(NR_SLAB_UNRECLAIMABLE
),
2074 global_page_state(NR_FILE_MAPPED
),
2075 global_page_state(NR_PAGETABLE
),
2076 global_page_state(NR_BOUNCE
));
2078 for_each_populated_zone(zone
) {
2087 " active_anon:%lukB"
2088 " inactive_anon:%lukB"
2089 " active_file:%lukB"
2090 " inactive_file:%lukB"
2091 #ifdef CONFIG_UNEVICTABLE_LRU
2092 " unevictable:%lukB"
2095 " pages_scanned:%lu"
2096 " all_unreclaimable? %s"
2099 K(zone_page_state(zone
, NR_FREE_PAGES
)),
2100 K(min_wmark_pages(zone
)),
2101 K(low_wmark_pages(zone
)),
2102 K(high_wmark_pages(zone
)),
2103 K(zone_page_state(zone
, NR_ACTIVE_ANON
)),
2104 K(zone_page_state(zone
, NR_INACTIVE_ANON
)),
2105 K(zone_page_state(zone
, NR_ACTIVE_FILE
)),
2106 K(zone_page_state(zone
, NR_INACTIVE_FILE
)),
2107 #ifdef CONFIG_UNEVICTABLE_LRU
2108 K(zone_page_state(zone
, NR_UNEVICTABLE
)),
2110 K(zone
->present_pages
),
2111 zone
->pages_scanned
,
2112 (zone_is_all_unreclaimable(zone
) ? "yes" : "no")
2114 printk("lowmem_reserve[]:");
2115 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
2116 printk(" %lu", zone
->lowmem_reserve
[i
]);
2120 for_each_populated_zone(zone
) {
2121 unsigned long nr
[MAX_ORDER
], flags
, order
, total
= 0;
2124 printk("%s: ", zone
->name
);
2126 spin_lock_irqsave(&zone
->lock
, flags
);
2127 for (order
= 0; order
< MAX_ORDER
; order
++) {
2128 nr
[order
] = zone
->free_area
[order
].nr_free
;
2129 total
+= nr
[order
] << order
;
2131 spin_unlock_irqrestore(&zone
->lock
, flags
);
2132 for (order
= 0; order
< MAX_ORDER
; order
++)
2133 printk("%lu*%lukB ", nr
[order
], K(1UL) << order
);
2134 printk("= %lukB\n", K(total
));
2137 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES
));
2139 show_swap_cache_info();
2142 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
2144 zoneref
->zone
= zone
;
2145 zoneref
->zone_idx
= zone_idx(zone
);
2149 * Builds allocation fallback zone lists.
2151 * Add all populated zones of a node to the zonelist.
2153 static int build_zonelists_node(pg_data_t
*pgdat
, struct zonelist
*zonelist
,
2154 int nr_zones
, enum zone_type zone_type
)
2158 BUG_ON(zone_type
>= MAX_NR_ZONES
);
2163 zone
= pgdat
->node_zones
+ zone_type
;
2164 if (populated_zone(zone
)) {
2165 zoneref_set_zone(zone
,
2166 &zonelist
->_zonerefs
[nr_zones
++]);
2167 check_highest_zone(zone_type
);
2170 } while (zone_type
);
2177 * 0 = automatic detection of better ordering.
2178 * 1 = order by ([node] distance, -zonetype)
2179 * 2 = order by (-zonetype, [node] distance)
2181 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
2182 * the same zonelist. So only NUMA can configure this param.
2184 #define ZONELIST_ORDER_DEFAULT 0
2185 #define ZONELIST_ORDER_NODE 1
2186 #define ZONELIST_ORDER_ZONE 2
2188 /* zonelist order in the kernel.
2189 * set_zonelist_order() will set this to NODE or ZONE.
2191 static int current_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
2192 static char zonelist_order_name
[3][8] = {"Default", "Node", "Zone"};
2196 /* The value user specified ....changed by config */
2197 static int user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
2198 /* string for sysctl */
2199 #define NUMA_ZONELIST_ORDER_LEN 16
2200 char numa_zonelist_order
[16] = "default";
2203 * interface for configure zonelist ordering.
2204 * command line option "numa_zonelist_order"
2205 * = "[dD]efault - default, automatic configuration.
2206 * = "[nN]ode - order by node locality, then by zone within node
2207 * = "[zZ]one - order by zone, then by locality within zone
2210 static int __parse_numa_zonelist_order(char *s
)
2212 if (*s
== 'd' || *s
== 'D') {
2213 user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
2214 } else if (*s
== 'n' || *s
== 'N') {
2215 user_zonelist_order
= ZONELIST_ORDER_NODE
;
2216 } else if (*s
== 'z' || *s
== 'Z') {
2217 user_zonelist_order
= ZONELIST_ORDER_ZONE
;
2220 "Ignoring invalid numa_zonelist_order value: "
2227 static __init
int setup_numa_zonelist_order(char *s
)
2230 return __parse_numa_zonelist_order(s
);
2233 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
2236 * sysctl handler for numa_zonelist_order
2238 int numa_zonelist_order_handler(ctl_table
*table
, int write
,
2239 struct file
*file
, void __user
*buffer
, size_t *length
,
2242 char saved_string
[NUMA_ZONELIST_ORDER_LEN
];
2246 strncpy(saved_string
, (char*)table
->data
,
2247 NUMA_ZONELIST_ORDER_LEN
);
2248 ret
= proc_dostring(table
, write
, file
, buffer
, length
, ppos
);
2252 int oldval
= user_zonelist_order
;
2253 if (__parse_numa_zonelist_order((char*)table
->data
)) {
2255 * bogus value. restore saved string
2257 strncpy((char*)table
->data
, saved_string
,
2258 NUMA_ZONELIST_ORDER_LEN
);
2259 user_zonelist_order
= oldval
;
2260 } else if (oldval
!= user_zonelist_order
)
2261 build_all_zonelists();
2267 #define MAX_NODE_LOAD (num_online_nodes())
2268 static int node_load
[MAX_NUMNODES
];
2271 * find_next_best_node - find the next node that should appear in a given node's fallback list
2272 * @node: node whose fallback list we're appending
2273 * @used_node_mask: nodemask_t of already used nodes
2275 * We use a number of factors to determine which is the next node that should
2276 * appear on a given node's fallback list. The node should not have appeared
2277 * already in @node's fallback list, and it should be the next closest node
2278 * according to the distance array (which contains arbitrary distance values
2279 * from each node to each node in the system), and should also prefer nodes
2280 * with no CPUs, since presumably they'll have very little allocation pressure
2281 * on them otherwise.
2282 * It returns -1 if no node is found.
2284 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
2287 int min_val
= INT_MAX
;
2289 const struct cpumask
*tmp
= cpumask_of_node(0);
2291 /* Use the local node if we haven't already */
2292 if (!node_isset(node
, *used_node_mask
)) {
2293 node_set(node
, *used_node_mask
);
2297 for_each_node_state(n
, N_HIGH_MEMORY
) {
2299 /* Don't want a node to appear more than once */
2300 if (node_isset(n
, *used_node_mask
))
2303 /* Use the distance array to find the distance */
2304 val
= node_distance(node
, n
);
2306 /* Penalize nodes under us ("prefer the next node") */
2309 /* Give preference to headless and unused nodes */
2310 tmp
= cpumask_of_node(n
);
2311 if (!cpumask_empty(tmp
))
2312 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
2314 /* Slight preference for less loaded node */
2315 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
2316 val
+= node_load
[n
];
2318 if (val
< min_val
) {
2325 node_set(best_node
, *used_node_mask
);
2332 * Build zonelists ordered by node and zones within node.
2333 * This results in maximum locality--normal zone overflows into local
2334 * DMA zone, if any--but risks exhausting DMA zone.
2336 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int node
)
2339 struct zonelist
*zonelist
;
2341 zonelist
= &pgdat
->node_zonelists
[0];
2342 for (j
= 0; zonelist
->_zonerefs
[j
].zone
!= NULL
; j
++)
2344 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
2346 zonelist
->_zonerefs
[j
].zone
= NULL
;
2347 zonelist
->_zonerefs
[j
].zone_idx
= 0;
2351 * Build gfp_thisnode zonelists
2353 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
2356 struct zonelist
*zonelist
;
2358 zonelist
= &pgdat
->node_zonelists
[1];
2359 j
= build_zonelists_node(pgdat
, zonelist
, 0, MAX_NR_ZONES
- 1);
2360 zonelist
->_zonerefs
[j
].zone
= NULL
;
2361 zonelist
->_zonerefs
[j
].zone_idx
= 0;
2365 * Build zonelists ordered by zone and nodes within zones.
2366 * This results in conserving DMA zone[s] until all Normal memory is
2367 * exhausted, but results in overflowing to remote node while memory
2368 * may still exist in local DMA zone.
2370 static int node_order
[MAX_NUMNODES
];
2372 static void build_zonelists_in_zone_order(pg_data_t
*pgdat
, int nr_nodes
)
2375 int zone_type
; /* needs to be signed */
2377 struct zonelist
*zonelist
;
2379 zonelist
= &pgdat
->node_zonelists
[0];
2381 for (zone_type
= MAX_NR_ZONES
- 1; zone_type
>= 0; zone_type
--) {
2382 for (j
= 0; j
< nr_nodes
; j
++) {
2383 node
= node_order
[j
];
2384 z
= &NODE_DATA(node
)->node_zones
[zone_type
];
2385 if (populated_zone(z
)) {
2387 &zonelist
->_zonerefs
[pos
++]);
2388 check_highest_zone(zone_type
);
2392 zonelist
->_zonerefs
[pos
].zone
= NULL
;
2393 zonelist
->_zonerefs
[pos
].zone_idx
= 0;
2396 static int default_zonelist_order(void)
2399 unsigned long low_kmem_size
,total_size
;
2403 * ZONE_DMA and ZONE_DMA32 can be very small area in the sytem.
2404 * If they are really small and used heavily, the system can fall
2405 * into OOM very easily.
2406 * This function detect ZONE_DMA/DMA32 size and confgigures zone order.
2408 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
2411 for_each_online_node(nid
) {
2412 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
2413 z
= &NODE_DATA(nid
)->node_zones
[zone_type
];
2414 if (populated_zone(z
)) {
2415 if (zone_type
< ZONE_NORMAL
)
2416 low_kmem_size
+= z
->present_pages
;
2417 total_size
+= z
->present_pages
;
2421 if (!low_kmem_size
|| /* there are no DMA area. */
2422 low_kmem_size
> total_size
/2) /* DMA/DMA32 is big. */
2423 return ZONELIST_ORDER_NODE
;
2425 * look into each node's config.
2426 * If there is a node whose DMA/DMA32 memory is very big area on
2427 * local memory, NODE_ORDER may be suitable.
2429 average_size
= total_size
/
2430 (nodes_weight(node_states
[N_HIGH_MEMORY
]) + 1);
2431 for_each_online_node(nid
) {
2434 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
2435 z
= &NODE_DATA(nid
)->node_zones
[zone_type
];
2436 if (populated_zone(z
)) {
2437 if (zone_type
< ZONE_NORMAL
)
2438 low_kmem_size
+= z
->present_pages
;
2439 total_size
+= z
->present_pages
;
2442 if (low_kmem_size
&&
2443 total_size
> average_size
&& /* ignore small node */
2444 low_kmem_size
> total_size
* 70/100)
2445 return ZONELIST_ORDER_NODE
;
2447 return ZONELIST_ORDER_ZONE
;
2450 static void set_zonelist_order(void)
2452 if (user_zonelist_order
== ZONELIST_ORDER_DEFAULT
)
2453 current_zonelist_order
= default_zonelist_order();
2455 current_zonelist_order
= user_zonelist_order
;
2458 static void build_zonelists(pg_data_t
*pgdat
)
2462 nodemask_t used_mask
;
2463 int local_node
, prev_node
;
2464 struct zonelist
*zonelist
;
2465 int order
= current_zonelist_order
;
2467 /* initialize zonelists */
2468 for (i
= 0; i
< MAX_ZONELISTS
; i
++) {
2469 zonelist
= pgdat
->node_zonelists
+ i
;
2470 zonelist
->_zonerefs
[0].zone
= NULL
;
2471 zonelist
->_zonerefs
[0].zone_idx
= 0;
2474 /* NUMA-aware ordering of nodes */
2475 local_node
= pgdat
->node_id
;
2476 load
= num_online_nodes();
2477 prev_node
= local_node
;
2478 nodes_clear(used_mask
);
2480 memset(node_load
, 0, sizeof(node_load
));
2481 memset(node_order
, 0, sizeof(node_order
));
2484 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
2485 int distance
= node_distance(local_node
, node
);
2488 * If another node is sufficiently far away then it is better
2489 * to reclaim pages in a zone before going off node.
2491 if (distance
> RECLAIM_DISTANCE
)
2492 zone_reclaim_mode
= 1;
2495 * We don't want to pressure a particular node.
2496 * So adding penalty to the first node in same
2497 * distance group to make it round-robin.
2499 if (distance
!= node_distance(local_node
, prev_node
))
2500 node_load
[node
] = load
;
2504 if (order
== ZONELIST_ORDER_NODE
)
2505 build_zonelists_in_node_order(pgdat
, node
);
2507 node_order
[j
++] = node
; /* remember order */
2510 if (order
== ZONELIST_ORDER_ZONE
) {
2511 /* calculate node order -- i.e., DMA last! */
2512 build_zonelists_in_zone_order(pgdat
, j
);
2515 build_thisnode_zonelists(pgdat
);
2518 /* Construct the zonelist performance cache - see further mmzone.h */
2519 static void build_zonelist_cache(pg_data_t
*pgdat
)
2521 struct zonelist
*zonelist
;
2522 struct zonelist_cache
*zlc
;
2525 zonelist
= &pgdat
->node_zonelists
[0];
2526 zonelist
->zlcache_ptr
= zlc
= &zonelist
->zlcache
;
2527 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
2528 for (z
= zonelist
->_zonerefs
; z
->zone
; z
++)
2529 zlc
->z_to_n
[z
- zonelist
->_zonerefs
] = zonelist_node_idx(z
);
2533 #else /* CONFIG_NUMA */
2535 static void set_zonelist_order(void)
2537 current_zonelist_order
= ZONELIST_ORDER_ZONE
;
2540 static void build_zonelists(pg_data_t
*pgdat
)
2542 int node
, local_node
;
2544 struct zonelist
*zonelist
;
2546 local_node
= pgdat
->node_id
;
2548 zonelist
= &pgdat
->node_zonelists
[0];
2549 j
= build_zonelists_node(pgdat
, zonelist
, 0, MAX_NR_ZONES
- 1);
2552 * Now we build the zonelist so that it contains the zones
2553 * of all the other nodes.
2554 * We don't want to pressure a particular node, so when
2555 * building the zones for node N, we make sure that the
2556 * zones coming right after the local ones are those from
2557 * node N+1 (modulo N)
2559 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
2560 if (!node_online(node
))
2562 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
2565 for (node
= 0; node
< local_node
; node
++) {
2566 if (!node_online(node
))
2568 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
2572 zonelist
->_zonerefs
[j
].zone
= NULL
;
2573 zonelist
->_zonerefs
[j
].zone_idx
= 0;
2576 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
2577 static void build_zonelist_cache(pg_data_t
*pgdat
)
2579 pgdat
->node_zonelists
[0].zlcache_ptr
= NULL
;
2582 #endif /* CONFIG_NUMA */
2584 /* return values int ....just for stop_machine() */
2585 static int __build_all_zonelists(void *dummy
)
2589 for_each_online_node(nid
) {
2590 pg_data_t
*pgdat
= NODE_DATA(nid
);
2592 build_zonelists(pgdat
);
2593 build_zonelist_cache(pgdat
);
2598 void build_all_zonelists(void)
2600 set_zonelist_order();
2602 if (system_state
== SYSTEM_BOOTING
) {
2603 __build_all_zonelists(NULL
);
2604 mminit_verify_zonelist();
2605 cpuset_init_current_mems_allowed();
2607 /* we have to stop all cpus to guarantee there is no user
2609 stop_machine(__build_all_zonelists
, NULL
, NULL
);
2610 /* cpuset refresh routine should be here */
2612 vm_total_pages
= nr_free_pagecache_pages();
2614 * Disable grouping by mobility if the number of pages in the
2615 * system is too low to allow the mechanism to work. It would be
2616 * more accurate, but expensive to check per-zone. This check is
2617 * made on memory-hotadd so a system can start with mobility
2618 * disabled and enable it later
2620 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
2621 page_group_by_mobility_disabled
= 1;
2623 page_group_by_mobility_disabled
= 0;
2625 printk("Built %i zonelists in %s order, mobility grouping %s. "
2626 "Total pages: %ld\n",
2628 zonelist_order_name
[current_zonelist_order
],
2629 page_group_by_mobility_disabled
? "off" : "on",
2632 printk("Policy zone: %s\n", zone_names
[policy_zone
]);
2637 * Helper functions to size the waitqueue hash table.
2638 * Essentially these want to choose hash table sizes sufficiently
2639 * large so that collisions trying to wait on pages are rare.
2640 * But in fact, the number of active page waitqueues on typical
2641 * systems is ridiculously low, less than 200. So this is even
2642 * conservative, even though it seems large.
2644 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
2645 * waitqueues, i.e. the size of the waitq table given the number of pages.
2647 #define PAGES_PER_WAITQUEUE 256
2649 #ifndef CONFIG_MEMORY_HOTPLUG
2650 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
2652 unsigned long size
= 1;
2654 pages
/= PAGES_PER_WAITQUEUE
;
2656 while (size
< pages
)
2660 * Once we have dozens or even hundreds of threads sleeping
2661 * on IO we've got bigger problems than wait queue collision.
2662 * Limit the size of the wait table to a reasonable size.
2664 size
= min(size
, 4096UL);
2666 return max(size
, 4UL);
2670 * A zone's size might be changed by hot-add, so it is not possible to determine
2671 * a suitable size for its wait_table. So we use the maximum size now.
2673 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
2675 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
2676 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
2677 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
2679 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
2680 * or more by the traditional way. (See above). It equals:
2682 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
2683 * ia64(16K page size) : = ( 8G + 4M)byte.
2684 * powerpc (64K page size) : = (32G +16M)byte.
2686 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
2693 * This is an integer logarithm so that shifts can be used later
2694 * to extract the more random high bits from the multiplicative
2695 * hash function before the remainder is taken.
2697 static inline unsigned long wait_table_bits(unsigned long size
)
2702 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
2705 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
2706 * of blocks reserved is based on min_wmark_pages(zone). The memory within
2707 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
2708 * higher will lead to a bigger reserve which will get freed as contiguous
2709 * blocks as reclaim kicks in
2711 static void setup_zone_migrate_reserve(struct zone
*zone
)
2713 unsigned long start_pfn
, pfn
, end_pfn
;
2715 unsigned long reserve
, block_migratetype
;
2717 /* Get the start pfn, end pfn and the number of blocks to reserve */
2718 start_pfn
= zone
->zone_start_pfn
;
2719 end_pfn
= start_pfn
+ zone
->spanned_pages
;
2720 reserve
= roundup(min_wmark_pages(zone
), pageblock_nr_pages
) >>
2723 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= pageblock_nr_pages
) {
2724 if (!pfn_valid(pfn
))
2726 page
= pfn_to_page(pfn
);
2728 /* Watch out for overlapping nodes */
2729 if (page_to_nid(page
) != zone_to_nid(zone
))
2732 /* Blocks with reserved pages will never free, skip them. */
2733 if (PageReserved(page
))
2736 block_migratetype
= get_pageblock_migratetype(page
);
2738 /* If this block is reserved, account for it */
2739 if (reserve
> 0 && block_migratetype
== MIGRATE_RESERVE
) {
2744 /* Suitable for reserving if this block is movable */
2745 if (reserve
> 0 && block_migratetype
== MIGRATE_MOVABLE
) {
2746 set_pageblock_migratetype(page
, MIGRATE_RESERVE
);
2747 move_freepages_block(zone
, page
, MIGRATE_RESERVE
);
2753 * If the reserve is met and this is a previous reserved block,
2756 if (block_migratetype
== MIGRATE_RESERVE
) {
2757 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
2758 move_freepages_block(zone
, page
, MIGRATE_MOVABLE
);
2764 * Initially all pages are reserved - free ones are freed
2765 * up by free_all_bootmem() once the early boot process is
2766 * done. Non-atomic initialization, single-pass.
2768 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
2769 unsigned long start_pfn
, enum memmap_context context
)
2772 unsigned long end_pfn
= start_pfn
+ size
;
2776 if (highest_memmap_pfn
< end_pfn
- 1)
2777 highest_memmap_pfn
= end_pfn
- 1;
2779 z
= &NODE_DATA(nid
)->node_zones
[zone
];
2780 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
2782 * There can be holes in boot-time mem_map[]s
2783 * handed to this function. They do not
2784 * exist on hotplugged memory.
2786 if (context
== MEMMAP_EARLY
) {
2787 if (!early_pfn_valid(pfn
))
2789 if (!early_pfn_in_nid(pfn
, nid
))
2792 page
= pfn_to_page(pfn
);
2793 set_page_links(page
, zone
, nid
, pfn
);
2794 mminit_verify_page_links(page
, zone
, nid
, pfn
);
2795 init_page_count(page
);
2796 reset_page_mapcount(page
);
2797 SetPageReserved(page
);
2799 * Mark the block movable so that blocks are reserved for
2800 * movable at startup. This will force kernel allocations
2801 * to reserve their blocks rather than leaking throughout
2802 * the address space during boot when many long-lived
2803 * kernel allocations are made. Later some blocks near
2804 * the start are marked MIGRATE_RESERVE by
2805 * setup_zone_migrate_reserve()
2807 * bitmap is created for zone's valid pfn range. but memmap
2808 * can be created for invalid pages (for alignment)
2809 * check here not to call set_pageblock_migratetype() against
2812 if ((z
->zone_start_pfn
<= pfn
)
2813 && (pfn
< z
->zone_start_pfn
+ z
->spanned_pages
)
2814 && !(pfn
& (pageblock_nr_pages
- 1)))
2815 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
2817 INIT_LIST_HEAD(&page
->lru
);
2818 #ifdef WANT_PAGE_VIRTUAL
2819 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
2820 if (!is_highmem_idx(zone
))
2821 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
2826 static void __meminit
zone_init_free_lists(struct zone
*zone
)
2829 for_each_migratetype_order(order
, t
) {
2830 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
2831 zone
->free_area
[order
].nr_free
= 0;
2835 #ifndef __HAVE_ARCH_MEMMAP_INIT
2836 #define memmap_init(size, nid, zone, start_pfn) \
2837 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
2840 static int zone_batchsize(struct zone
*zone
)
2846 * The per-cpu-pages pools are set to around 1000th of the
2847 * size of the zone. But no more than 1/2 of a meg.
2849 * OK, so we don't know how big the cache is. So guess.
2851 batch
= zone
->present_pages
/ 1024;
2852 if (batch
* PAGE_SIZE
> 512 * 1024)
2853 batch
= (512 * 1024) / PAGE_SIZE
;
2854 batch
/= 4; /* We effectively *= 4 below */
2859 * Clamp the batch to a 2^n - 1 value. Having a power
2860 * of 2 value was found to be more likely to have
2861 * suboptimal cache aliasing properties in some cases.
2863 * For example if 2 tasks are alternately allocating
2864 * batches of pages, one task can end up with a lot
2865 * of pages of one half of the possible page colors
2866 * and the other with pages of the other colors.
2868 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
2873 /* The deferral and batching of frees should be suppressed under NOMMU
2876 * The problem is that NOMMU needs to be able to allocate large chunks
2877 * of contiguous memory as there's no hardware page translation to
2878 * assemble apparent contiguous memory from discontiguous pages.
2880 * Queueing large contiguous runs of pages for batching, however,
2881 * causes the pages to actually be freed in smaller chunks. As there
2882 * can be a significant delay between the individual batches being
2883 * recycled, this leads to the once large chunks of space being
2884 * fragmented and becoming unavailable for high-order allocations.
2890 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
2892 struct per_cpu_pages
*pcp
;
2894 memset(p
, 0, sizeof(*p
));
2898 pcp
->high
= 6 * batch
;
2899 pcp
->batch
= max(1UL, 1 * batch
);
2900 INIT_LIST_HEAD(&pcp
->list
);
2904 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
2905 * to the value high for the pageset p.
2908 static void setup_pagelist_highmark(struct per_cpu_pageset
*p
,
2911 struct per_cpu_pages
*pcp
;
2915 pcp
->batch
= max(1UL, high
/4);
2916 if ((high
/4) > (PAGE_SHIFT
* 8))
2917 pcp
->batch
= PAGE_SHIFT
* 8;
2923 * Boot pageset table. One per cpu which is going to be used for all
2924 * zones and all nodes. The parameters will be set in such a way
2925 * that an item put on a list will immediately be handed over to
2926 * the buddy list. This is safe since pageset manipulation is done
2927 * with interrupts disabled.
2929 * Some NUMA counter updates may also be caught by the boot pagesets.
2931 * The boot_pagesets must be kept even after bootup is complete for
2932 * unused processors and/or zones. They do play a role for bootstrapping
2933 * hotplugged processors.
2935 * zoneinfo_show() and maybe other functions do
2936 * not check if the processor is online before following the pageset pointer.
2937 * Other parts of the kernel may not check if the zone is available.
2939 static struct per_cpu_pageset boot_pageset
[NR_CPUS
];
2942 * Dynamically allocate memory for the
2943 * per cpu pageset array in struct zone.
2945 static int __cpuinit
process_zones(int cpu
)
2947 struct zone
*zone
, *dzone
;
2948 int node
= cpu_to_node(cpu
);
2950 node_set_state(node
, N_CPU
); /* this node has a cpu */
2952 for_each_populated_zone(zone
) {
2953 zone_pcp(zone
, cpu
) = kmalloc_node(sizeof(struct per_cpu_pageset
),
2955 if (!zone_pcp(zone
, cpu
))
2958 setup_pageset(zone_pcp(zone
, cpu
), zone_batchsize(zone
));
2960 if (percpu_pagelist_fraction
)
2961 setup_pagelist_highmark(zone_pcp(zone
, cpu
),
2962 (zone
->present_pages
/ percpu_pagelist_fraction
));
2967 for_each_zone(dzone
) {
2968 if (!populated_zone(dzone
))
2972 kfree(zone_pcp(dzone
, cpu
));
2973 zone_pcp(dzone
, cpu
) = NULL
;
2978 static inline void free_zone_pagesets(int cpu
)
2982 for_each_zone(zone
) {
2983 struct per_cpu_pageset
*pset
= zone_pcp(zone
, cpu
);
2985 /* Free per_cpu_pageset if it is slab allocated */
2986 if (pset
!= &boot_pageset
[cpu
])
2988 zone_pcp(zone
, cpu
) = NULL
;
2992 static int __cpuinit
pageset_cpuup_callback(struct notifier_block
*nfb
,
2993 unsigned long action
,
2996 int cpu
= (long)hcpu
;
2997 int ret
= NOTIFY_OK
;
3000 case CPU_UP_PREPARE
:
3001 case CPU_UP_PREPARE_FROZEN
:
3002 if (process_zones(cpu
))
3005 case CPU_UP_CANCELED
:
3006 case CPU_UP_CANCELED_FROZEN
:
3008 case CPU_DEAD_FROZEN
:
3009 free_zone_pagesets(cpu
);
3017 static struct notifier_block __cpuinitdata pageset_notifier
=
3018 { &pageset_cpuup_callback
, NULL
, 0 };
3020 void __init
setup_per_cpu_pageset(void)
3024 /* Initialize per_cpu_pageset for cpu 0.
3025 * A cpuup callback will do this for every cpu
3026 * as it comes online
3028 err
= process_zones(smp_processor_id());
3030 register_cpu_notifier(&pageset_notifier
);
3035 static noinline __init_refok
3036 int zone_wait_table_init(struct zone
*zone
, unsigned long zone_size_pages
)
3039 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
3043 * The per-page waitqueue mechanism uses hashed waitqueues
3046 zone
->wait_table_hash_nr_entries
=
3047 wait_table_hash_nr_entries(zone_size_pages
);
3048 zone
->wait_table_bits
=
3049 wait_table_bits(zone
->wait_table_hash_nr_entries
);
3050 alloc_size
= zone
->wait_table_hash_nr_entries
3051 * sizeof(wait_queue_head_t
);
3053 if (!slab_is_available()) {
3054 zone
->wait_table
= (wait_queue_head_t
*)
3055 alloc_bootmem_node(pgdat
, alloc_size
);
3058 * This case means that a zone whose size was 0 gets new memory
3059 * via memory hot-add.
3060 * But it may be the case that a new node was hot-added. In
3061 * this case vmalloc() will not be able to use this new node's
3062 * memory - this wait_table must be initialized to use this new
3063 * node itself as well.
3064 * To use this new node's memory, further consideration will be
3067 zone
->wait_table
= vmalloc(alloc_size
);
3069 if (!zone
->wait_table
)
3072 for(i
= 0; i
< zone
->wait_table_hash_nr_entries
; ++i
)
3073 init_waitqueue_head(zone
->wait_table
+ i
);
3078 static __meminit
void zone_pcp_init(struct zone
*zone
)
3081 unsigned long batch
= zone_batchsize(zone
);
3083 for (cpu
= 0; cpu
< NR_CPUS
; cpu
++) {
3085 /* Early boot. Slab allocator not functional yet */
3086 zone_pcp(zone
, cpu
) = &boot_pageset
[cpu
];
3087 setup_pageset(&boot_pageset
[cpu
],0);
3089 setup_pageset(zone_pcp(zone
,cpu
), batch
);
3092 if (zone
->present_pages
)
3093 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%lu\n",
3094 zone
->name
, zone
->present_pages
, batch
);
3097 __meminit
int init_currently_empty_zone(struct zone
*zone
,
3098 unsigned long zone_start_pfn
,
3100 enum memmap_context context
)
3102 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
3104 ret
= zone_wait_table_init(zone
, size
);
3107 pgdat
->nr_zones
= zone_idx(zone
) + 1;
3109 zone
->zone_start_pfn
= zone_start_pfn
;
3111 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
3112 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
3114 (unsigned long)zone_idx(zone
),
3115 zone_start_pfn
, (zone_start_pfn
+ size
));
3117 zone_init_free_lists(zone
);
3122 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3124 * Basic iterator support. Return the first range of PFNs for a node
3125 * Note: nid == MAX_NUMNODES returns first region regardless of node
3127 static int __meminit
first_active_region_index_in_nid(int nid
)
3131 for (i
= 0; i
< nr_nodemap_entries
; i
++)
3132 if (nid
== MAX_NUMNODES
|| early_node_map
[i
].nid
== nid
)
3139 * Basic iterator support. Return the next active range of PFNs for a node
3140 * Note: nid == MAX_NUMNODES returns next region regardless of node
3142 static int __meminit
next_active_region_index_in_nid(int index
, int nid
)
3144 for (index
= index
+ 1; index
< nr_nodemap_entries
; index
++)
3145 if (nid
== MAX_NUMNODES
|| early_node_map
[index
].nid
== nid
)
3151 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
3153 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
3154 * Architectures may implement their own version but if add_active_range()
3155 * was used and there are no special requirements, this is a convenient
3158 int __meminit
__early_pfn_to_nid(unsigned long pfn
)
3162 for (i
= 0; i
< nr_nodemap_entries
; i
++) {
3163 unsigned long start_pfn
= early_node_map
[i
].start_pfn
;
3164 unsigned long end_pfn
= early_node_map
[i
].end_pfn
;
3166 if (start_pfn
<= pfn
&& pfn
< end_pfn
)
3167 return early_node_map
[i
].nid
;
3169 /* This is a memory hole */
3172 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
3174 int __meminit
early_pfn_to_nid(unsigned long pfn
)
3178 nid
= __early_pfn_to_nid(pfn
);
3181 /* just returns 0 */
3185 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
3186 bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
3190 nid
= __early_pfn_to_nid(pfn
);
3191 if (nid
>= 0 && nid
!= node
)
3197 /* Basic iterator support to walk early_node_map[] */
3198 #define for_each_active_range_index_in_nid(i, nid) \
3199 for (i = first_active_region_index_in_nid(nid); i != -1; \
3200 i = next_active_region_index_in_nid(i, nid))
3203 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
3204 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
3205 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
3207 * If an architecture guarantees that all ranges registered with
3208 * add_active_ranges() contain no holes and may be freed, this
3209 * this function may be used instead of calling free_bootmem() manually.
3211 void __init
free_bootmem_with_active_regions(int nid
,
3212 unsigned long max_low_pfn
)
3216 for_each_active_range_index_in_nid(i
, nid
) {
3217 unsigned long size_pages
= 0;
3218 unsigned long end_pfn
= early_node_map
[i
].end_pfn
;
3220 if (early_node_map
[i
].start_pfn
>= max_low_pfn
)
3223 if (end_pfn
> max_low_pfn
)
3224 end_pfn
= max_low_pfn
;
3226 size_pages
= end_pfn
- early_node_map
[i
].start_pfn
;
3227 free_bootmem_node(NODE_DATA(early_node_map
[i
].nid
),
3228 PFN_PHYS(early_node_map
[i
].start_pfn
),
3229 size_pages
<< PAGE_SHIFT
);
3233 void __init
work_with_active_regions(int nid
, work_fn_t work_fn
, void *data
)
3238 for_each_active_range_index_in_nid(i
, nid
) {
3239 ret
= work_fn(early_node_map
[i
].start_pfn
,
3240 early_node_map
[i
].end_pfn
, data
);
3246 * sparse_memory_present_with_active_regions - Call memory_present for each active range
3247 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
3249 * If an architecture guarantees that all ranges registered with
3250 * add_active_ranges() contain no holes and may be freed, this
3251 * function may be used instead of calling memory_present() manually.
3253 void __init
sparse_memory_present_with_active_regions(int nid
)
3257 for_each_active_range_index_in_nid(i
, nid
)
3258 memory_present(early_node_map
[i
].nid
,
3259 early_node_map
[i
].start_pfn
,
3260 early_node_map
[i
].end_pfn
);
3264 * get_pfn_range_for_nid - Return the start and end page frames for a node
3265 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
3266 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
3267 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
3269 * It returns the start and end page frame of a node based on information
3270 * provided by an arch calling add_active_range(). If called for a node
3271 * with no available memory, a warning is printed and the start and end
3274 void __meminit
get_pfn_range_for_nid(unsigned int nid
,
3275 unsigned long *start_pfn
, unsigned long *end_pfn
)
3281 for_each_active_range_index_in_nid(i
, nid
) {
3282 *start_pfn
= min(*start_pfn
, early_node_map
[i
].start_pfn
);
3283 *end_pfn
= max(*end_pfn
, early_node_map
[i
].end_pfn
);
3286 if (*start_pfn
== -1UL)
3291 * This finds a zone that can be used for ZONE_MOVABLE pages. The
3292 * assumption is made that zones within a node are ordered in monotonic
3293 * increasing memory addresses so that the "highest" populated zone is used
3295 static void __init
find_usable_zone_for_movable(void)
3298 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
3299 if (zone_index
== ZONE_MOVABLE
)
3302 if (arch_zone_highest_possible_pfn
[zone_index
] >
3303 arch_zone_lowest_possible_pfn
[zone_index
])
3307 VM_BUG_ON(zone_index
== -1);
3308 movable_zone
= zone_index
;
3312 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
3313 * because it is sized independant of architecture. Unlike the other zones,
3314 * the starting point for ZONE_MOVABLE is not fixed. It may be different
3315 * in each node depending on the size of each node and how evenly kernelcore
3316 * is distributed. This helper function adjusts the zone ranges
3317 * provided by the architecture for a given node by using the end of the
3318 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
3319 * zones within a node are in order of monotonic increases memory addresses
3321 static void __meminit
adjust_zone_range_for_zone_movable(int nid
,
3322 unsigned long zone_type
,
3323 unsigned long node_start_pfn
,
3324 unsigned long node_end_pfn
,
3325 unsigned long *zone_start_pfn
,
3326 unsigned long *zone_end_pfn
)
3328 /* Only adjust if ZONE_MOVABLE is on this node */
3329 if (zone_movable_pfn
[nid
]) {
3330 /* Size ZONE_MOVABLE */
3331 if (zone_type
== ZONE_MOVABLE
) {
3332 *zone_start_pfn
= zone_movable_pfn
[nid
];
3333 *zone_end_pfn
= min(node_end_pfn
,
3334 arch_zone_highest_possible_pfn
[movable_zone
]);
3336 /* Adjust for ZONE_MOVABLE starting within this range */
3337 } else if (*zone_start_pfn
< zone_movable_pfn
[nid
] &&
3338 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
3339 *zone_end_pfn
= zone_movable_pfn
[nid
];
3341 /* Check if this whole range is within ZONE_MOVABLE */
3342 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
3343 *zone_start_pfn
= *zone_end_pfn
;
3348 * Return the number of pages a zone spans in a node, including holes
3349 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
3351 static unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
3352 unsigned long zone_type
,
3353 unsigned long *ignored
)
3355 unsigned long node_start_pfn
, node_end_pfn
;
3356 unsigned long zone_start_pfn
, zone_end_pfn
;
3358 /* Get the start and end of the node and zone */
3359 get_pfn_range_for_nid(nid
, &node_start_pfn
, &node_end_pfn
);
3360 zone_start_pfn
= arch_zone_lowest_possible_pfn
[zone_type
];
3361 zone_end_pfn
= arch_zone_highest_possible_pfn
[zone_type
];
3362 adjust_zone_range_for_zone_movable(nid
, zone_type
,
3363 node_start_pfn
, node_end_pfn
,
3364 &zone_start_pfn
, &zone_end_pfn
);
3366 /* Check that this node has pages within the zone's required range */
3367 if (zone_end_pfn
< node_start_pfn
|| zone_start_pfn
> node_end_pfn
)
3370 /* Move the zone boundaries inside the node if necessary */
3371 zone_end_pfn
= min(zone_end_pfn
, node_end_pfn
);
3372 zone_start_pfn
= max(zone_start_pfn
, node_start_pfn
);
3374 /* Return the spanned pages */
3375 return zone_end_pfn
- zone_start_pfn
;
3379 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
3380 * then all holes in the requested range will be accounted for.
3382 static unsigned long __meminit
__absent_pages_in_range(int nid
,
3383 unsigned long range_start_pfn
,
3384 unsigned long range_end_pfn
)
3387 unsigned long prev_end_pfn
= 0, hole_pages
= 0;
3388 unsigned long start_pfn
;
3390 /* Find the end_pfn of the first active range of pfns in the node */
3391 i
= first_active_region_index_in_nid(nid
);
3395 prev_end_pfn
= min(early_node_map
[i
].start_pfn
, range_end_pfn
);
3397 /* Account for ranges before physical memory on this node */
3398 if (early_node_map
[i
].start_pfn
> range_start_pfn
)
3399 hole_pages
= prev_end_pfn
- range_start_pfn
;
3401 /* Find all holes for the zone within the node */
3402 for (; i
!= -1; i
= next_active_region_index_in_nid(i
, nid
)) {
3404 /* No need to continue if prev_end_pfn is outside the zone */
3405 if (prev_end_pfn
>= range_end_pfn
)
3408 /* Make sure the end of the zone is not within the hole */
3409 start_pfn
= min(early_node_map
[i
].start_pfn
, range_end_pfn
);
3410 prev_end_pfn
= max(prev_end_pfn
, range_start_pfn
);
3412 /* Update the hole size cound and move on */
3413 if (start_pfn
> range_start_pfn
) {
3414 BUG_ON(prev_end_pfn
> start_pfn
);
3415 hole_pages
+= start_pfn
- prev_end_pfn
;
3417 prev_end_pfn
= early_node_map
[i
].end_pfn
;
3420 /* Account for ranges past physical memory on this node */
3421 if (range_end_pfn
> prev_end_pfn
)
3422 hole_pages
+= range_end_pfn
-
3423 max(range_start_pfn
, prev_end_pfn
);
3429 * absent_pages_in_range - Return number of page frames in holes within a range
3430 * @start_pfn: The start PFN to start searching for holes
3431 * @end_pfn: The end PFN to stop searching for holes
3433 * It returns the number of pages frames in memory holes within a range.
3435 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
3436 unsigned long end_pfn
)
3438 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
3441 /* Return the number of page frames in holes in a zone on a node */
3442 static unsigned long __meminit
zone_absent_pages_in_node(int nid
,
3443 unsigned long zone_type
,
3444 unsigned long *ignored
)
3446 unsigned long node_start_pfn
, node_end_pfn
;
3447 unsigned long zone_start_pfn
, zone_end_pfn
;
3449 get_pfn_range_for_nid(nid
, &node_start_pfn
, &node_end_pfn
);
3450 zone_start_pfn
= max(arch_zone_lowest_possible_pfn
[zone_type
],
3452 zone_end_pfn
= min(arch_zone_highest_possible_pfn
[zone_type
],
3455 adjust_zone_range_for_zone_movable(nid
, zone_type
,
3456 node_start_pfn
, node_end_pfn
,
3457 &zone_start_pfn
, &zone_end_pfn
);
3458 return __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
3462 static inline unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
3463 unsigned long zone_type
,
3464 unsigned long *zones_size
)
3466 return zones_size
[zone_type
];
3469 static inline unsigned long __meminit
zone_absent_pages_in_node(int nid
,
3470 unsigned long zone_type
,
3471 unsigned long *zholes_size
)
3476 return zholes_size
[zone_type
];
3481 static void __meminit
calculate_node_totalpages(struct pglist_data
*pgdat
,
3482 unsigned long *zones_size
, unsigned long *zholes_size
)
3484 unsigned long realtotalpages
, totalpages
= 0;
3487 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
3488 totalpages
+= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
3490 pgdat
->node_spanned_pages
= totalpages
;
3492 realtotalpages
= totalpages
;
3493 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
3495 zone_absent_pages_in_node(pgdat
->node_id
, i
,
3497 pgdat
->node_present_pages
= realtotalpages
;
3498 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
3502 #ifndef CONFIG_SPARSEMEM
3504 * Calculate the size of the zone->blockflags rounded to an unsigned long
3505 * Start by making sure zonesize is a multiple of pageblock_order by rounding
3506 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
3507 * round what is now in bits to nearest long in bits, then return it in
3510 static unsigned long __init
usemap_size(unsigned long zonesize
)
3512 unsigned long usemapsize
;
3514 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
3515 usemapsize
= usemapsize
>> pageblock_order
;
3516 usemapsize
*= NR_PAGEBLOCK_BITS
;
3517 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
3519 return usemapsize
/ 8;
3522 static void __init
setup_usemap(struct pglist_data
*pgdat
,
3523 struct zone
*zone
, unsigned long zonesize
)
3525 unsigned long usemapsize
= usemap_size(zonesize
);
3526 zone
->pageblock_flags
= NULL
;
3528 zone
->pageblock_flags
= alloc_bootmem_node(pgdat
, usemapsize
);
3531 static void inline setup_usemap(struct pglist_data
*pgdat
,
3532 struct zone
*zone
, unsigned long zonesize
) {}
3533 #endif /* CONFIG_SPARSEMEM */
3535 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
3537 /* Return a sensible default order for the pageblock size. */
3538 static inline int pageblock_default_order(void)
3540 if (HPAGE_SHIFT
> PAGE_SHIFT
)
3541 return HUGETLB_PAGE_ORDER
;
3546 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
3547 static inline void __init
set_pageblock_order(unsigned int order
)
3549 /* Check that pageblock_nr_pages has not already been setup */
3550 if (pageblock_order
)
3554 * Assume the largest contiguous order of interest is a huge page.
3555 * This value may be variable depending on boot parameters on IA64
3557 pageblock_order
= order
;
3559 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
3562 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
3563 * and pageblock_default_order() are unused as pageblock_order is set
3564 * at compile-time. See include/linux/pageblock-flags.h for the values of
3565 * pageblock_order based on the kernel config
3567 static inline int pageblock_default_order(unsigned int order
)
3571 #define set_pageblock_order(x) do {} while (0)
3573 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
3576 * Set up the zone data structures:
3577 * - mark all pages reserved
3578 * - mark all memory queues empty
3579 * - clear the memory bitmaps
3581 static void __paginginit
free_area_init_core(struct pglist_data
*pgdat
,
3582 unsigned long *zones_size
, unsigned long *zholes_size
)
3585 int nid
= pgdat
->node_id
;
3586 unsigned long zone_start_pfn
= pgdat
->node_start_pfn
;
3589 pgdat_resize_init(pgdat
);
3590 pgdat
->nr_zones
= 0;
3591 init_waitqueue_head(&pgdat
->kswapd_wait
);
3592 pgdat
->kswapd_max_order
= 0;
3593 pgdat_page_cgroup_init(pgdat
);
3595 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
3596 struct zone
*zone
= pgdat
->node_zones
+ j
;
3597 unsigned long size
, realsize
, memmap_pages
;
3600 size
= zone_spanned_pages_in_node(nid
, j
, zones_size
);
3601 realsize
= size
- zone_absent_pages_in_node(nid
, j
,
3605 * Adjust realsize so that it accounts for how much memory
3606 * is used by this zone for memmap. This affects the watermark
3607 * and per-cpu initialisations
3610 PAGE_ALIGN(size
* sizeof(struct page
)) >> PAGE_SHIFT
;
3611 if (realsize
>= memmap_pages
) {
3612 realsize
-= memmap_pages
;
3615 " %s zone: %lu pages used for memmap\n",
3616 zone_names
[j
], memmap_pages
);
3619 " %s zone: %lu pages exceeds realsize %lu\n",
3620 zone_names
[j
], memmap_pages
, realsize
);
3622 /* Account for reserved pages */
3623 if (j
== 0 && realsize
> dma_reserve
) {
3624 realsize
-= dma_reserve
;
3625 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
3626 zone_names
[0], dma_reserve
);
3629 if (!is_highmem_idx(j
))
3630 nr_kernel_pages
+= realsize
;
3631 nr_all_pages
+= realsize
;
3633 zone
->spanned_pages
= size
;
3634 zone
->present_pages
= realsize
;
3637 zone
->min_unmapped_pages
= (realsize
*sysctl_min_unmapped_ratio
)
3639 zone
->min_slab_pages
= (realsize
* sysctl_min_slab_ratio
) / 100;
3641 zone
->name
= zone_names
[j
];
3642 spin_lock_init(&zone
->lock
);
3643 spin_lock_init(&zone
->lru_lock
);
3644 zone_seqlock_init(zone
);
3645 zone
->zone_pgdat
= pgdat
;
3647 zone
->prev_priority
= DEF_PRIORITY
;
3649 zone_pcp_init(zone
);
3651 INIT_LIST_HEAD(&zone
->lru
[l
].list
);
3652 zone
->lru
[l
].nr_scan
= 0;
3654 zone
->reclaim_stat
.recent_rotated
[0] = 0;
3655 zone
->reclaim_stat
.recent_rotated
[1] = 0;
3656 zone
->reclaim_stat
.recent_scanned
[0] = 0;
3657 zone
->reclaim_stat
.recent_scanned
[1] = 0;
3658 zap_zone_vm_stats(zone
);
3663 set_pageblock_order(pageblock_default_order());
3664 setup_usemap(pgdat
, zone
, size
);
3665 ret
= init_currently_empty_zone(zone
, zone_start_pfn
,
3666 size
, MEMMAP_EARLY
);
3668 memmap_init(size
, nid
, j
, zone_start_pfn
);
3669 zone_start_pfn
+= size
;
3673 static void __init_refok
alloc_node_mem_map(struct pglist_data
*pgdat
)
3675 /* Skip empty nodes */
3676 if (!pgdat
->node_spanned_pages
)
3679 #ifdef CONFIG_FLAT_NODE_MEM_MAP
3680 /* ia64 gets its own node_mem_map, before this, without bootmem */
3681 if (!pgdat
->node_mem_map
) {
3682 unsigned long size
, start
, end
;
3686 * The zone's endpoints aren't required to be MAX_ORDER
3687 * aligned but the node_mem_map endpoints must be in order
3688 * for the buddy allocator to function correctly.
3690 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
3691 end
= pgdat
->node_start_pfn
+ pgdat
->node_spanned_pages
;
3692 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
3693 size
= (end
- start
) * sizeof(struct page
);
3694 map
= alloc_remap(pgdat
->node_id
, size
);
3696 map
= alloc_bootmem_node(pgdat
, size
);
3697 pgdat
->node_mem_map
= map
+ (pgdat
->node_start_pfn
- start
);
3699 #ifndef CONFIG_NEED_MULTIPLE_NODES
3701 * With no DISCONTIG, the global mem_map is just set as node 0's
3703 if (pgdat
== NODE_DATA(0)) {
3704 mem_map
= NODE_DATA(0)->node_mem_map
;
3705 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3706 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
3707 mem_map
-= (pgdat
->node_start_pfn
- ARCH_PFN_OFFSET
);
3708 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
3711 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
3714 void __paginginit
free_area_init_node(int nid
, unsigned long *zones_size
,
3715 unsigned long node_start_pfn
, unsigned long *zholes_size
)
3717 pg_data_t
*pgdat
= NODE_DATA(nid
);
3719 pgdat
->node_id
= nid
;
3720 pgdat
->node_start_pfn
= node_start_pfn
;
3721 calculate_node_totalpages(pgdat
, zones_size
, zholes_size
);
3723 alloc_node_mem_map(pgdat
);
3724 #ifdef CONFIG_FLAT_NODE_MEM_MAP
3725 printk(KERN_DEBUG
"free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
3726 nid
, (unsigned long)pgdat
,
3727 (unsigned long)pgdat
->node_mem_map
);
3730 free_area_init_core(pgdat
, zones_size
, zholes_size
);
3733 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3735 #if MAX_NUMNODES > 1
3737 * Figure out the number of possible node ids.
3739 static void __init
setup_nr_node_ids(void)
3742 unsigned int highest
= 0;
3744 for_each_node_mask(node
, node_possible_map
)
3746 nr_node_ids
= highest
+ 1;
3749 static inline void setup_nr_node_ids(void)
3755 * add_active_range - Register a range of PFNs backed by physical memory
3756 * @nid: The node ID the range resides on
3757 * @start_pfn: The start PFN of the available physical memory
3758 * @end_pfn: The end PFN of the available physical memory
3760 * These ranges are stored in an early_node_map[] and later used by
3761 * free_area_init_nodes() to calculate zone sizes and holes. If the
3762 * range spans a memory hole, it is up to the architecture to ensure
3763 * the memory is not freed by the bootmem allocator. If possible
3764 * the range being registered will be merged with existing ranges.
3766 void __init
add_active_range(unsigned int nid
, unsigned long start_pfn
,
3767 unsigned long end_pfn
)
3771 mminit_dprintk(MMINIT_TRACE
, "memory_register",
3772 "Entering add_active_range(%d, %#lx, %#lx) "
3773 "%d entries of %d used\n",
3774 nid
, start_pfn
, end_pfn
,
3775 nr_nodemap_entries
, MAX_ACTIVE_REGIONS
);
3777 mminit_validate_memmodel_limits(&start_pfn
, &end_pfn
);
3779 /* Merge with existing active regions if possible */
3780 for (i
= 0; i
< nr_nodemap_entries
; i
++) {
3781 if (early_node_map
[i
].nid
!= nid
)
3784 /* Skip if an existing region covers this new one */
3785 if (start_pfn
>= early_node_map
[i
].start_pfn
&&
3786 end_pfn
<= early_node_map
[i
].end_pfn
)
3789 /* Merge forward if suitable */
3790 if (start_pfn
<= early_node_map
[i
].end_pfn
&&
3791 end_pfn
> early_node_map
[i
].end_pfn
) {
3792 early_node_map
[i
].end_pfn
= end_pfn
;
3796 /* Merge backward if suitable */
3797 if (start_pfn
< early_node_map
[i
].end_pfn
&&
3798 end_pfn
>= early_node_map
[i
].start_pfn
) {
3799 early_node_map
[i
].start_pfn
= start_pfn
;
3804 /* Check that early_node_map is large enough */
3805 if (i
>= MAX_ACTIVE_REGIONS
) {
3806 printk(KERN_CRIT
"More than %d memory regions, truncating\n",
3807 MAX_ACTIVE_REGIONS
);
3811 early_node_map
[i
].nid
= nid
;
3812 early_node_map
[i
].start_pfn
= start_pfn
;
3813 early_node_map
[i
].end_pfn
= end_pfn
;
3814 nr_nodemap_entries
= i
+ 1;
3818 * remove_active_range - Shrink an existing registered range of PFNs
3819 * @nid: The node id the range is on that should be shrunk
3820 * @start_pfn: The new PFN of the range
3821 * @end_pfn: The new PFN of the range
3823 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
3824 * The map is kept near the end physical page range that has already been
3825 * registered. This function allows an arch to shrink an existing registered
3828 void __init
remove_active_range(unsigned int nid
, unsigned long start_pfn
,
3829 unsigned long end_pfn
)
3834 printk(KERN_DEBUG
"remove_active_range (%d, %lu, %lu)\n",
3835 nid
, start_pfn
, end_pfn
);
3837 /* Find the old active region end and shrink */
3838 for_each_active_range_index_in_nid(i
, nid
) {
3839 if (early_node_map
[i
].start_pfn
>= start_pfn
&&
3840 early_node_map
[i
].end_pfn
<= end_pfn
) {
3842 early_node_map
[i
].start_pfn
= 0;
3843 early_node_map
[i
].end_pfn
= 0;
3847 if (early_node_map
[i
].start_pfn
< start_pfn
&&
3848 early_node_map
[i
].end_pfn
> start_pfn
) {
3849 unsigned long temp_end_pfn
= early_node_map
[i
].end_pfn
;
3850 early_node_map
[i
].end_pfn
= start_pfn
;
3851 if (temp_end_pfn
> end_pfn
)
3852 add_active_range(nid
, end_pfn
, temp_end_pfn
);
3855 if (early_node_map
[i
].start_pfn
>= start_pfn
&&
3856 early_node_map
[i
].end_pfn
> end_pfn
&&
3857 early_node_map
[i
].start_pfn
< end_pfn
) {
3858 early_node_map
[i
].start_pfn
= end_pfn
;
3866 /* remove the blank ones */
3867 for (i
= nr_nodemap_entries
- 1; i
> 0; i
--) {
3868 if (early_node_map
[i
].nid
!= nid
)
3870 if (early_node_map
[i
].end_pfn
)
3872 /* we found it, get rid of it */
3873 for (j
= i
; j
< nr_nodemap_entries
- 1; j
++)
3874 memcpy(&early_node_map
[j
], &early_node_map
[j
+1],
3875 sizeof(early_node_map
[j
]));
3876 j
= nr_nodemap_entries
- 1;
3877 memset(&early_node_map
[j
], 0, sizeof(early_node_map
[j
]));
3878 nr_nodemap_entries
--;
3883 * remove_all_active_ranges - Remove all currently registered regions
3885 * During discovery, it may be found that a table like SRAT is invalid
3886 * and an alternative discovery method must be used. This function removes
3887 * all currently registered regions.
3889 void __init
remove_all_active_ranges(void)
3891 memset(early_node_map
, 0, sizeof(early_node_map
));
3892 nr_nodemap_entries
= 0;
3895 /* Compare two active node_active_regions */
3896 static int __init
cmp_node_active_region(const void *a
, const void *b
)
3898 struct node_active_region
*arange
= (struct node_active_region
*)a
;
3899 struct node_active_region
*brange
= (struct node_active_region
*)b
;
3901 /* Done this way to avoid overflows */
3902 if (arange
->start_pfn
> brange
->start_pfn
)
3904 if (arange
->start_pfn
< brange
->start_pfn
)
3910 /* sort the node_map by start_pfn */
3911 static void __init
sort_node_map(void)
3913 sort(early_node_map
, (size_t)nr_nodemap_entries
,
3914 sizeof(struct node_active_region
),
3915 cmp_node_active_region
, NULL
);
3918 /* Find the lowest pfn for a node */
3919 static unsigned long __init
find_min_pfn_for_node(int nid
)
3922 unsigned long min_pfn
= ULONG_MAX
;
3924 /* Assuming a sorted map, the first range found has the starting pfn */
3925 for_each_active_range_index_in_nid(i
, nid
)
3926 min_pfn
= min(min_pfn
, early_node_map
[i
].start_pfn
);
3928 if (min_pfn
== ULONG_MAX
) {
3930 "Could not find start_pfn for node %d\n", nid
);
3938 * find_min_pfn_with_active_regions - Find the minimum PFN registered
3940 * It returns the minimum PFN based on information provided via
3941 * add_active_range().
3943 unsigned long __init
find_min_pfn_with_active_regions(void)
3945 return find_min_pfn_for_node(MAX_NUMNODES
);
3949 * early_calculate_totalpages()
3950 * Sum pages in active regions for movable zone.
3951 * Populate N_HIGH_MEMORY for calculating usable_nodes.
3953 static unsigned long __init
early_calculate_totalpages(void)
3956 unsigned long totalpages
= 0;
3958 for (i
= 0; i
< nr_nodemap_entries
; i
++) {
3959 unsigned long pages
= early_node_map
[i
].end_pfn
-
3960 early_node_map
[i
].start_pfn
;
3961 totalpages
+= pages
;
3963 node_set_state(early_node_map
[i
].nid
, N_HIGH_MEMORY
);
3969 * Find the PFN the Movable zone begins in each node. Kernel memory
3970 * is spread evenly between nodes as long as the nodes have enough
3971 * memory. When they don't, some nodes will have more kernelcore than
3974 static void __init
find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn
)
3977 unsigned long usable_startpfn
;
3978 unsigned long kernelcore_node
, kernelcore_remaining
;
3979 unsigned long totalpages
= early_calculate_totalpages();
3980 int usable_nodes
= nodes_weight(node_states
[N_HIGH_MEMORY
]);
3983 * If movablecore was specified, calculate what size of
3984 * kernelcore that corresponds so that memory usable for
3985 * any allocation type is evenly spread. If both kernelcore
3986 * and movablecore are specified, then the value of kernelcore
3987 * will be used for required_kernelcore if it's greater than
3988 * what movablecore would have allowed.
3990 if (required_movablecore
) {
3991 unsigned long corepages
;
3994 * Round-up so that ZONE_MOVABLE is at least as large as what
3995 * was requested by the user
3997 required_movablecore
=
3998 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
3999 corepages
= totalpages
- required_movablecore
;
4001 required_kernelcore
= max(required_kernelcore
, corepages
);
4004 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4005 if (!required_kernelcore
)
4008 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4009 find_usable_zone_for_movable();
4010 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
4013 /* Spread kernelcore memory as evenly as possible throughout nodes */
4014 kernelcore_node
= required_kernelcore
/ usable_nodes
;
4015 for_each_node_state(nid
, N_HIGH_MEMORY
) {
4017 * Recalculate kernelcore_node if the division per node
4018 * now exceeds what is necessary to satisfy the requested
4019 * amount of memory for the kernel
4021 if (required_kernelcore
< kernelcore_node
)
4022 kernelcore_node
= required_kernelcore
/ usable_nodes
;
4025 * As the map is walked, we track how much memory is usable
4026 * by the kernel using kernelcore_remaining. When it is
4027 * 0, the rest of the node is usable by ZONE_MOVABLE
4029 kernelcore_remaining
= kernelcore_node
;
4031 /* Go through each range of PFNs within this node */
4032 for_each_active_range_index_in_nid(i
, nid
) {
4033 unsigned long start_pfn
, end_pfn
;
4034 unsigned long size_pages
;
4036 start_pfn
= max(early_node_map
[i
].start_pfn
,
4037 zone_movable_pfn
[nid
]);
4038 end_pfn
= early_node_map
[i
].end_pfn
;
4039 if (start_pfn
>= end_pfn
)
4042 /* Account for what is only usable for kernelcore */
4043 if (start_pfn
< usable_startpfn
) {
4044 unsigned long kernel_pages
;
4045 kernel_pages
= min(end_pfn
, usable_startpfn
)
4048 kernelcore_remaining
-= min(kernel_pages
,
4049 kernelcore_remaining
);
4050 required_kernelcore
-= min(kernel_pages
,
4051 required_kernelcore
);
4053 /* Continue if range is now fully accounted */
4054 if (end_pfn
<= usable_startpfn
) {
4057 * Push zone_movable_pfn to the end so
4058 * that if we have to rebalance
4059 * kernelcore across nodes, we will
4060 * not double account here
4062 zone_movable_pfn
[nid
] = end_pfn
;
4065 start_pfn
= usable_startpfn
;
4069 * The usable PFN range for ZONE_MOVABLE is from
4070 * start_pfn->end_pfn. Calculate size_pages as the
4071 * number of pages used as kernelcore
4073 size_pages
= end_pfn
- start_pfn
;
4074 if (size_pages
> kernelcore_remaining
)
4075 size_pages
= kernelcore_remaining
;
4076 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
4079 * Some kernelcore has been met, update counts and
4080 * break if the kernelcore for this node has been
4083 required_kernelcore
-= min(required_kernelcore
,
4085 kernelcore_remaining
-= size_pages
;
4086 if (!kernelcore_remaining
)
4092 * If there is still required_kernelcore, we do another pass with one
4093 * less node in the count. This will push zone_movable_pfn[nid] further
4094 * along on the nodes that still have memory until kernelcore is
4098 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
4101 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4102 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
4103 zone_movable_pfn
[nid
] =
4104 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
4107 /* Any regular memory on that node ? */
4108 static void check_for_regular_memory(pg_data_t
*pgdat
)
4110 #ifdef CONFIG_HIGHMEM
4111 enum zone_type zone_type
;
4113 for (zone_type
= 0; zone_type
<= ZONE_NORMAL
; zone_type
++) {
4114 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
4115 if (zone
->present_pages
)
4116 node_set_state(zone_to_nid(zone
), N_NORMAL_MEMORY
);
4122 * free_area_init_nodes - Initialise all pg_data_t and zone data
4123 * @max_zone_pfn: an array of max PFNs for each zone
4125 * This will call free_area_init_node() for each active node in the system.
4126 * Using the page ranges provided by add_active_range(), the size of each
4127 * zone in each node and their holes is calculated. If the maximum PFN
4128 * between two adjacent zones match, it is assumed that the zone is empty.
4129 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4130 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4131 * starts where the previous one ended. For example, ZONE_DMA32 starts
4132 * at arch_max_dma_pfn.
4134 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
4139 /* Sort early_node_map as initialisation assumes it is sorted */
4142 /* Record where the zone boundaries are */
4143 memset(arch_zone_lowest_possible_pfn
, 0,
4144 sizeof(arch_zone_lowest_possible_pfn
));
4145 memset(arch_zone_highest_possible_pfn
, 0,
4146 sizeof(arch_zone_highest_possible_pfn
));
4147 arch_zone_lowest_possible_pfn
[0] = find_min_pfn_with_active_regions();
4148 arch_zone_highest_possible_pfn
[0] = max_zone_pfn
[0];
4149 for (i
= 1; i
< MAX_NR_ZONES
; i
++) {
4150 if (i
== ZONE_MOVABLE
)
4152 arch_zone_lowest_possible_pfn
[i
] =
4153 arch_zone_highest_possible_pfn
[i
-1];
4154 arch_zone_highest_possible_pfn
[i
] =
4155 max(max_zone_pfn
[i
], arch_zone_lowest_possible_pfn
[i
]);
4157 arch_zone_lowest_possible_pfn
[ZONE_MOVABLE
] = 0;
4158 arch_zone_highest_possible_pfn
[ZONE_MOVABLE
] = 0;
4160 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4161 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
4162 find_zone_movable_pfns_for_nodes(zone_movable_pfn
);
4164 /* Print out the zone ranges */
4165 printk("Zone PFN ranges:\n");
4166 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
4167 if (i
== ZONE_MOVABLE
)
4169 printk(" %-8s %0#10lx -> %0#10lx\n",
4171 arch_zone_lowest_possible_pfn
[i
],
4172 arch_zone_highest_possible_pfn
[i
]);
4175 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4176 printk("Movable zone start PFN for each node\n");
4177 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
4178 if (zone_movable_pfn
[i
])
4179 printk(" Node %d: %lu\n", i
, zone_movable_pfn
[i
]);
4182 /* Print out the early_node_map[] */
4183 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries
);
4184 for (i
= 0; i
< nr_nodemap_entries
; i
++)
4185 printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map
[i
].nid
,
4186 early_node_map
[i
].start_pfn
,
4187 early_node_map
[i
].end_pfn
);
4189 /* Initialise every node */
4190 mminit_verify_pageflags_layout();
4191 setup_nr_node_ids();
4192 for_each_online_node(nid
) {
4193 pg_data_t
*pgdat
= NODE_DATA(nid
);
4194 free_area_init_node(nid
, NULL
,
4195 find_min_pfn_for_node(nid
), NULL
);
4197 /* Any memory on that node */
4198 if (pgdat
->node_present_pages
)
4199 node_set_state(nid
, N_HIGH_MEMORY
);
4200 check_for_regular_memory(pgdat
);
4204 static int __init
cmdline_parse_core(char *p
, unsigned long *core
)
4206 unsigned long long coremem
;
4210 coremem
= memparse(p
, &p
);
4211 *core
= coremem
>> PAGE_SHIFT
;
4213 /* Paranoid check that UL is enough for the coremem value */
4214 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
4220 * kernelcore=size sets the amount of memory for use for allocations that
4221 * cannot be reclaimed or migrated.
4223 static int __init
cmdline_parse_kernelcore(char *p
)
4225 return cmdline_parse_core(p
, &required_kernelcore
);
4229 * movablecore=size sets the amount of memory for use for allocations that
4230 * can be reclaimed or migrated.
4232 static int __init
cmdline_parse_movablecore(char *p
)
4234 return cmdline_parse_core(p
, &required_movablecore
);
4237 early_param("kernelcore", cmdline_parse_kernelcore
);
4238 early_param("movablecore", cmdline_parse_movablecore
);
4240 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4243 * set_dma_reserve - set the specified number of pages reserved in the first zone
4244 * @new_dma_reserve: The number of pages to mark reserved
4246 * The per-cpu batchsize and zone watermarks are determined by present_pages.
4247 * In the DMA zone, a significant percentage may be consumed by kernel image
4248 * and other unfreeable allocations which can skew the watermarks badly. This
4249 * function may optionally be used to account for unfreeable pages in the
4250 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
4251 * smaller per-cpu batchsize.
4253 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
4255 dma_reserve
= new_dma_reserve
;
4258 #ifndef CONFIG_NEED_MULTIPLE_NODES
4259 struct pglist_data __refdata contig_page_data
= { .bdata
= &bootmem_node_data
[0] };
4260 EXPORT_SYMBOL(contig_page_data
);
4263 void __init
free_area_init(unsigned long *zones_size
)
4265 free_area_init_node(0, zones_size
,
4266 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
4269 static int page_alloc_cpu_notify(struct notifier_block
*self
,
4270 unsigned long action
, void *hcpu
)
4272 int cpu
= (unsigned long)hcpu
;
4274 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
) {
4278 * Spill the event counters of the dead processor
4279 * into the current processors event counters.
4280 * This artificially elevates the count of the current
4283 vm_events_fold_cpu(cpu
);
4286 * Zero the differential counters of the dead processor
4287 * so that the vm statistics are consistent.
4289 * This is only okay since the processor is dead and cannot
4290 * race with what we are doing.
4292 refresh_cpu_vm_stats(cpu
);
4297 void __init
page_alloc_init(void)
4299 hotcpu_notifier(page_alloc_cpu_notify
, 0);
4303 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
4304 * or min_free_kbytes changes.
4306 static void calculate_totalreserve_pages(void)
4308 struct pglist_data
*pgdat
;
4309 unsigned long reserve_pages
= 0;
4310 enum zone_type i
, j
;
4312 for_each_online_pgdat(pgdat
) {
4313 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
4314 struct zone
*zone
= pgdat
->node_zones
+ i
;
4315 unsigned long max
= 0;
4317 /* Find valid and maximum lowmem_reserve in the zone */
4318 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
4319 if (zone
->lowmem_reserve
[j
] > max
)
4320 max
= zone
->lowmem_reserve
[j
];
4323 /* we treat the high watermark as reserved pages. */
4324 max
+= high_wmark_pages(zone
);
4326 if (max
> zone
->present_pages
)
4327 max
= zone
->present_pages
;
4328 reserve_pages
+= max
;
4331 totalreserve_pages
= reserve_pages
;
4335 * setup_per_zone_lowmem_reserve - called whenever
4336 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
4337 * has a correct pages reserved value, so an adequate number of
4338 * pages are left in the zone after a successful __alloc_pages().
4340 static void setup_per_zone_lowmem_reserve(void)
4342 struct pglist_data
*pgdat
;
4343 enum zone_type j
, idx
;
4345 for_each_online_pgdat(pgdat
) {
4346 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
4347 struct zone
*zone
= pgdat
->node_zones
+ j
;
4348 unsigned long present_pages
= zone
->present_pages
;
4350 zone
->lowmem_reserve
[j
] = 0;
4354 struct zone
*lower_zone
;
4358 if (sysctl_lowmem_reserve_ratio
[idx
] < 1)
4359 sysctl_lowmem_reserve_ratio
[idx
] = 1;
4361 lower_zone
= pgdat
->node_zones
+ idx
;
4362 lower_zone
->lowmem_reserve
[j
] = present_pages
/
4363 sysctl_lowmem_reserve_ratio
[idx
];
4364 present_pages
+= lower_zone
->present_pages
;
4369 /* update totalreserve_pages */
4370 calculate_totalreserve_pages();
4374 * setup_per_zone_pages_min - called when min_free_kbytes changes.
4376 * Ensures that the pages_{min,low,high} values for each zone are set correctly
4377 * with respect to min_free_kbytes.
4379 void setup_per_zone_pages_min(void)
4381 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
4382 unsigned long lowmem_pages
= 0;
4384 unsigned long flags
;
4386 /* Calculate total number of !ZONE_HIGHMEM pages */
4387 for_each_zone(zone
) {
4388 if (!is_highmem(zone
))
4389 lowmem_pages
+= zone
->present_pages
;
4392 for_each_zone(zone
) {
4395 spin_lock_irqsave(&zone
->lock
, flags
);
4396 tmp
= (u64
)pages_min
* zone
->present_pages
;
4397 do_div(tmp
, lowmem_pages
);
4398 if (is_highmem(zone
)) {
4400 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
4401 * need highmem pages, so cap pages_min to a small
4404 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
4405 * deltas controls asynch page reclaim, and so should
4406 * not be capped for highmem.
4410 min_pages
= zone
->present_pages
/ 1024;
4411 if (min_pages
< SWAP_CLUSTER_MAX
)
4412 min_pages
= SWAP_CLUSTER_MAX
;
4413 if (min_pages
> 128)
4415 zone
->watermark
[WMARK_MIN
] = min_pages
;
4418 * If it's a lowmem zone, reserve a number of pages
4419 * proportionate to the zone's size.
4421 zone
->watermark
[WMARK_MIN
] = tmp
;
4424 zone
->watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + (tmp
>> 2);
4425 zone
->watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + (tmp
>> 1);
4426 setup_zone_migrate_reserve(zone
);
4427 spin_unlock_irqrestore(&zone
->lock
, flags
);
4430 /* update totalreserve_pages */
4431 calculate_totalreserve_pages();
4435 * setup_per_zone_inactive_ratio - called when min_free_kbytes changes.
4437 * The inactive anon list should be small enough that the VM never has to
4438 * do too much work, but large enough that each inactive page has a chance
4439 * to be referenced again before it is swapped out.
4441 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
4442 * INACTIVE_ANON pages on this zone's LRU, maintained by the
4443 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
4444 * the anonymous pages are kept on the inactive list.
4447 * memory ratio inactive anon
4448 * -------------------------------------
4457 static void setup_per_zone_inactive_ratio(void)
4461 for_each_zone(zone
) {
4462 unsigned int gb
, ratio
;
4464 /* Zone size in gigabytes */
4465 gb
= zone
->present_pages
>> (30 - PAGE_SHIFT
);
4466 ratio
= int_sqrt(10 * gb
);
4470 zone
->inactive_ratio
= ratio
;
4475 * Initialise min_free_kbytes.
4477 * For small machines we want it small (128k min). For large machines
4478 * we want it large (64MB max). But it is not linear, because network
4479 * bandwidth does not increase linearly with machine size. We use
4481 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
4482 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
4498 static int __init
init_per_zone_pages_min(void)
4500 unsigned long lowmem_kbytes
;
4502 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
4504 min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
4505 if (min_free_kbytes
< 128)
4506 min_free_kbytes
= 128;
4507 if (min_free_kbytes
> 65536)
4508 min_free_kbytes
= 65536;
4509 setup_per_zone_pages_min();
4510 setup_per_zone_lowmem_reserve();
4511 setup_per_zone_inactive_ratio();
4514 module_init(init_per_zone_pages_min
)
4517 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
4518 * that we can call two helper functions whenever min_free_kbytes
4521 int min_free_kbytes_sysctl_handler(ctl_table
*table
, int write
,
4522 struct file
*file
, void __user
*buffer
, size_t *length
, loff_t
*ppos
)
4524 proc_dointvec(table
, write
, file
, buffer
, length
, ppos
);
4526 setup_per_zone_pages_min();
4531 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table
*table
, int write
,
4532 struct file
*file
, void __user
*buffer
, size_t *length
, loff_t
*ppos
)
4537 rc
= proc_dointvec_minmax(table
, write
, file
, buffer
, length
, ppos
);
4542 zone
->min_unmapped_pages
= (zone
->present_pages
*
4543 sysctl_min_unmapped_ratio
) / 100;
4547 int sysctl_min_slab_ratio_sysctl_handler(ctl_table
*table
, int write
,
4548 struct file
*file
, void __user
*buffer
, size_t *length
, loff_t
*ppos
)
4553 rc
= proc_dointvec_minmax(table
, write
, file
, buffer
, length
, ppos
);
4558 zone
->min_slab_pages
= (zone
->present_pages
*
4559 sysctl_min_slab_ratio
) / 100;
4565 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
4566 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
4567 * whenever sysctl_lowmem_reserve_ratio changes.
4569 * The reserve ratio obviously has absolutely no relation with the
4570 * minimum watermarks. The lowmem reserve ratio can only make sense
4571 * if in function of the boot time zone sizes.
4573 int lowmem_reserve_ratio_sysctl_handler(ctl_table
*table
, int write
,
4574 struct file
*file
, void __user
*buffer
, size_t *length
, loff_t
*ppos
)
4576 proc_dointvec_minmax(table
, write
, file
, buffer
, length
, ppos
);
4577 setup_per_zone_lowmem_reserve();
4582 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
4583 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
4584 * can have before it gets flushed back to buddy allocator.
4587 int percpu_pagelist_fraction_sysctl_handler(ctl_table
*table
, int write
,
4588 struct file
*file
, void __user
*buffer
, size_t *length
, loff_t
*ppos
)
4594 ret
= proc_dointvec_minmax(table
, write
, file
, buffer
, length
, ppos
);
4595 if (!write
|| (ret
== -EINVAL
))
4597 for_each_zone(zone
) {
4598 for_each_online_cpu(cpu
) {
4600 high
= zone
->present_pages
/ percpu_pagelist_fraction
;
4601 setup_pagelist_highmark(zone_pcp(zone
, cpu
), high
);
4607 int hashdist
= HASHDIST_DEFAULT
;
4610 static int __init
set_hashdist(char *str
)
4614 hashdist
= simple_strtoul(str
, &str
, 0);
4617 __setup("hashdist=", set_hashdist
);
4621 * allocate a large system hash table from bootmem
4622 * - it is assumed that the hash table must contain an exact power-of-2
4623 * quantity of entries
4624 * - limit is the number of hash buckets, not the total allocation size
4626 void *__init
alloc_large_system_hash(const char *tablename
,
4627 unsigned long bucketsize
,
4628 unsigned long numentries
,
4631 unsigned int *_hash_shift
,
4632 unsigned int *_hash_mask
,
4633 unsigned long limit
)
4635 unsigned long long max
= limit
;
4636 unsigned long log2qty
, size
;
4639 /* allow the kernel cmdline to have a say */
4641 /* round applicable memory size up to nearest megabyte */
4642 numentries
= nr_kernel_pages
;
4643 numentries
+= (1UL << (20 - PAGE_SHIFT
)) - 1;
4644 numentries
>>= 20 - PAGE_SHIFT
;
4645 numentries
<<= 20 - PAGE_SHIFT
;
4647 /* limit to 1 bucket per 2^scale bytes of low memory */
4648 if (scale
> PAGE_SHIFT
)
4649 numentries
>>= (scale
- PAGE_SHIFT
);
4651 numentries
<<= (PAGE_SHIFT
- scale
);
4653 /* Make sure we've got at least a 0-order allocation.. */
4654 if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
4655 numentries
= PAGE_SIZE
/ bucketsize
;
4657 numentries
= roundup_pow_of_two(numentries
);
4659 /* limit allocation size to 1/16 total memory by default */
4661 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
4662 do_div(max
, bucketsize
);
4665 if (numentries
> max
)
4668 log2qty
= ilog2(numentries
);
4671 size
= bucketsize
<< log2qty
;
4672 if (flags
& HASH_EARLY
)
4673 table
= alloc_bootmem_nopanic(size
);
4675 table
= __vmalloc(size
, GFP_ATOMIC
, PAGE_KERNEL
);
4677 unsigned long order
= get_order(size
);
4679 if (order
< MAX_ORDER
)
4680 table
= (void *)__get_free_pages(GFP_ATOMIC
,
4683 * If bucketsize is not a power-of-two, we may free
4684 * some pages at the end of hash table.
4687 unsigned long alloc_end
= (unsigned long)table
+
4688 (PAGE_SIZE
<< order
);
4689 unsigned long used
= (unsigned long)table
+
4691 split_page(virt_to_page(table
), order
);
4692 while (used
< alloc_end
) {
4698 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
4701 panic("Failed to allocate %s hash table\n", tablename
);
4703 printk(KERN_INFO
"%s hash table entries: %d (order: %d, %lu bytes)\n",
4706 ilog2(size
) - PAGE_SHIFT
,
4710 *_hash_shift
= log2qty
;
4712 *_hash_mask
= (1 << log2qty
) - 1;
4715 * If hashdist is set, the table allocation is done with __vmalloc()
4716 * which invokes the kmemleak_alloc() callback. This function may also
4717 * be called before the slab and kmemleak are initialised when
4718 * kmemleak simply buffers the request to be executed later
4719 * (GFP_ATOMIC flag ignored in this case).
4722 kmemleak_alloc(table
, size
, 1, GFP_ATOMIC
);
4727 /* Return a pointer to the bitmap storing bits affecting a block of pages */
4728 static inline unsigned long *get_pageblock_bitmap(struct zone
*zone
,
4731 #ifdef CONFIG_SPARSEMEM
4732 return __pfn_to_section(pfn
)->pageblock_flags
;
4734 return zone
->pageblock_flags
;
4735 #endif /* CONFIG_SPARSEMEM */
4738 static inline int pfn_to_bitidx(struct zone
*zone
, unsigned long pfn
)
4740 #ifdef CONFIG_SPARSEMEM
4741 pfn
&= (PAGES_PER_SECTION
-1);
4742 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
4744 pfn
= pfn
- zone
->zone_start_pfn
;
4745 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
4746 #endif /* CONFIG_SPARSEMEM */
4750 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
4751 * @page: The page within the block of interest
4752 * @start_bitidx: The first bit of interest to retrieve
4753 * @end_bitidx: The last bit of interest
4754 * returns pageblock_bits flags
4756 unsigned long get_pageblock_flags_group(struct page
*page
,
4757 int start_bitidx
, int end_bitidx
)
4760 unsigned long *bitmap
;
4761 unsigned long pfn
, bitidx
;
4762 unsigned long flags
= 0;
4763 unsigned long value
= 1;
4765 zone
= page_zone(page
);
4766 pfn
= page_to_pfn(page
);
4767 bitmap
= get_pageblock_bitmap(zone
, pfn
);
4768 bitidx
= pfn_to_bitidx(zone
, pfn
);
4770 for (; start_bitidx
<= end_bitidx
; start_bitidx
++, value
<<= 1)
4771 if (test_bit(bitidx
+ start_bitidx
, bitmap
))
4778 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
4779 * @page: The page within the block of interest
4780 * @start_bitidx: The first bit of interest
4781 * @end_bitidx: The last bit of interest
4782 * @flags: The flags to set
4784 void set_pageblock_flags_group(struct page
*page
, unsigned long flags
,
4785 int start_bitidx
, int end_bitidx
)
4788 unsigned long *bitmap
;
4789 unsigned long pfn
, bitidx
;
4790 unsigned long value
= 1;
4792 zone
= page_zone(page
);
4793 pfn
= page_to_pfn(page
);
4794 bitmap
= get_pageblock_bitmap(zone
, pfn
);
4795 bitidx
= pfn_to_bitidx(zone
, pfn
);
4796 VM_BUG_ON(pfn
< zone
->zone_start_pfn
);
4797 VM_BUG_ON(pfn
>= zone
->zone_start_pfn
+ zone
->spanned_pages
);
4799 for (; start_bitidx
<= end_bitidx
; start_bitidx
++, value
<<= 1)
4801 __set_bit(bitidx
+ start_bitidx
, bitmap
);
4803 __clear_bit(bitidx
+ start_bitidx
, bitmap
);
4807 * This is designed as sub function...plz see page_isolation.c also.
4808 * set/clear page block's type to be ISOLATE.
4809 * page allocater never alloc memory from ISOLATE block.
4812 int set_migratetype_isolate(struct page
*page
)
4815 unsigned long flags
;
4818 zone
= page_zone(page
);
4819 spin_lock_irqsave(&zone
->lock
, flags
);
4821 * In future, more migrate types will be able to be isolation target.
4823 if (get_pageblock_migratetype(page
) != MIGRATE_MOVABLE
)
4825 set_pageblock_migratetype(page
, MIGRATE_ISOLATE
);
4826 move_freepages_block(zone
, page
, MIGRATE_ISOLATE
);
4829 spin_unlock_irqrestore(&zone
->lock
, flags
);
4835 void unset_migratetype_isolate(struct page
*page
)
4838 unsigned long flags
;
4839 zone
= page_zone(page
);
4840 spin_lock_irqsave(&zone
->lock
, flags
);
4841 if (get_pageblock_migratetype(page
) != MIGRATE_ISOLATE
)
4843 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
4844 move_freepages_block(zone
, page
, MIGRATE_MOVABLE
);
4846 spin_unlock_irqrestore(&zone
->lock
, flags
);
4849 #ifdef CONFIG_MEMORY_HOTREMOVE
4851 * All pages in the range must be isolated before calling this.
4854 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
4860 unsigned long flags
;
4861 /* find the first valid pfn */
4862 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
4867 zone
= page_zone(pfn_to_page(pfn
));
4868 spin_lock_irqsave(&zone
->lock
, flags
);
4870 while (pfn
< end_pfn
) {
4871 if (!pfn_valid(pfn
)) {
4875 page
= pfn_to_page(pfn
);
4876 BUG_ON(page_count(page
));
4877 BUG_ON(!PageBuddy(page
));
4878 order
= page_order(page
);
4879 #ifdef CONFIG_DEBUG_VM
4880 printk(KERN_INFO
"remove from free list %lx %d %lx\n",
4881 pfn
, 1 << order
, end_pfn
);
4883 list_del(&page
->lru
);
4884 rmv_page_order(page
);
4885 zone
->free_area
[order
].nr_free
--;
4886 __mod_zone_page_state(zone
, NR_FREE_PAGES
,
4888 for (i
= 0; i
< (1 << order
); i
++)
4889 SetPageReserved((page
+i
));
4890 pfn
+= (1 << order
);
4892 spin_unlock_irqrestore(&zone
->lock
, flags
);