2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/compiler.h>
25 #include <linux/kernel.h>
26 #include <linux/kmemcheck.h>
27 #include <linux/module.h>
28 #include <linux/suspend.h>
29 #include <linux/pagevec.h>
30 #include <linux/blkdev.h>
31 #include <linux/slab.h>
32 #include <linux/oom.h>
33 #include <linux/notifier.h>
34 #include <linux/topology.h>
35 #include <linux/sysctl.h>
36 #include <linux/cpu.h>
37 #include <linux/cpuset.h>
38 #include <linux/memory_hotplug.h>
39 #include <linux/nodemask.h>
40 #include <linux/vmalloc.h>
41 #include <linux/mempolicy.h>
42 #include <linux/stop_machine.h>
43 #include <linux/sort.h>
44 #include <linux/pfn.h>
45 #include <linux/backing-dev.h>
46 #include <linux/fault-inject.h>
47 #include <linux/page-isolation.h>
48 #include <linux/page_cgroup.h>
49 #include <linux/debugobjects.h>
50 #include <linux/kmemleak.h>
52 #include <asm/tlbflush.h>
53 #include <asm/div64.h>
57 * Array of node states.
59 nodemask_t node_states
[NR_NODE_STATES
] __read_mostly
= {
60 [N_POSSIBLE
] = NODE_MASK_ALL
,
61 [N_ONLINE
] = { { [0] = 1UL } },
63 [N_NORMAL_MEMORY
] = { { [0] = 1UL } },
65 [N_HIGH_MEMORY
] = { { [0] = 1UL } },
67 [N_CPU
] = { { [0] = 1UL } },
70 EXPORT_SYMBOL(node_states
);
72 unsigned long totalram_pages __read_mostly
;
73 unsigned long totalreserve_pages __read_mostly
;
74 unsigned long highest_memmap_pfn __read_mostly
;
75 int percpu_pagelist_fraction
;
76 gfp_t gfp_allowed_mask __read_mostly
= GFP_BOOT_MASK
;
78 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
79 int pageblock_order __read_mostly
;
82 static void __free_pages_ok(struct page
*page
, unsigned int order
);
85 * results with 256, 32 in the lowmem_reserve sysctl:
86 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
87 * 1G machine -> (16M dma, 784M normal, 224M high)
88 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
89 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
90 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
92 * TBD: should special case ZONE_DMA32 machines here - in those we normally
93 * don't need any ZONE_NORMAL reservation
95 int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
-1] = {
96 #ifdef CONFIG_ZONE_DMA
99 #ifdef CONFIG_ZONE_DMA32
102 #ifdef CONFIG_HIGHMEM
108 EXPORT_SYMBOL(totalram_pages
);
110 static char * const zone_names
[MAX_NR_ZONES
] = {
111 #ifdef CONFIG_ZONE_DMA
114 #ifdef CONFIG_ZONE_DMA32
118 #ifdef CONFIG_HIGHMEM
124 int min_free_kbytes
= 1024;
126 unsigned long __meminitdata nr_kernel_pages
;
127 unsigned long __meminitdata nr_all_pages
;
128 static unsigned long __meminitdata dma_reserve
;
130 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
132 * MAX_ACTIVE_REGIONS determines the maximum number of distinct
133 * ranges of memory (RAM) that may be registered with add_active_range().
134 * Ranges passed to add_active_range() will be merged if possible
135 * so the number of times add_active_range() can be called is
136 * related to the number of nodes and the number of holes
138 #ifdef CONFIG_MAX_ACTIVE_REGIONS
139 /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */
140 #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS
142 #if MAX_NUMNODES >= 32
143 /* If there can be many nodes, allow up to 50 holes per node */
144 #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50)
146 /* By default, allow up to 256 distinct regions */
147 #define MAX_ACTIVE_REGIONS 256
151 static struct node_active_region __meminitdata early_node_map
[MAX_ACTIVE_REGIONS
];
152 static int __meminitdata nr_nodemap_entries
;
153 static unsigned long __meminitdata arch_zone_lowest_possible_pfn
[MAX_NR_ZONES
];
154 static unsigned long __meminitdata arch_zone_highest_possible_pfn
[MAX_NR_ZONES
];
155 static unsigned long __initdata required_kernelcore
;
156 static unsigned long __initdata required_movablecore
;
157 static unsigned long __meminitdata zone_movable_pfn
[MAX_NUMNODES
];
159 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
161 EXPORT_SYMBOL(movable_zone
);
162 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
165 int nr_node_ids __read_mostly
= MAX_NUMNODES
;
166 int nr_online_nodes __read_mostly
= 1;
167 EXPORT_SYMBOL(nr_node_ids
);
168 EXPORT_SYMBOL(nr_online_nodes
);
171 int page_group_by_mobility_disabled __read_mostly
;
173 static void set_pageblock_migratetype(struct page
*page
, int migratetype
)
176 if (unlikely(page_group_by_mobility_disabled
))
177 migratetype
= MIGRATE_UNMOVABLE
;
179 set_pageblock_flags_group(page
, (unsigned long)migratetype
,
180 PB_migrate
, PB_migrate_end
);
183 bool oom_killer_disabled __read_mostly
;
185 #ifdef CONFIG_DEBUG_VM
186 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
190 unsigned long pfn
= page_to_pfn(page
);
193 seq
= zone_span_seqbegin(zone
);
194 if (pfn
>= zone
->zone_start_pfn
+ zone
->spanned_pages
)
196 else if (pfn
< zone
->zone_start_pfn
)
198 } while (zone_span_seqretry(zone
, seq
));
203 static int page_is_consistent(struct zone
*zone
, struct page
*page
)
205 if (!pfn_valid_within(page_to_pfn(page
)))
207 if (zone
!= page_zone(page
))
213 * Temporary debugging check for pages not lying within a given zone.
215 static int bad_range(struct zone
*zone
, struct page
*page
)
217 if (page_outside_zone_boundaries(zone
, page
))
219 if (!page_is_consistent(zone
, page
))
225 static inline int bad_range(struct zone
*zone
, struct page
*page
)
231 static void bad_page(struct page
*page
)
233 static unsigned long resume
;
234 static unsigned long nr_shown
;
235 static unsigned long nr_unshown
;
237 /* Don't complain about poisoned pages */
238 if (PageHWPoison(page
)) {
239 __ClearPageBuddy(page
);
244 * Allow a burst of 60 reports, then keep quiet for that minute;
245 * or allow a steady drip of one report per second.
247 if (nr_shown
== 60) {
248 if (time_before(jiffies
, resume
)) {
254 "BUG: Bad page state: %lu messages suppressed\n",
261 resume
= jiffies
+ 60 * HZ
;
263 printk(KERN_ALERT
"BUG: Bad page state in process %s pfn:%05lx\n",
264 current
->comm
, page_to_pfn(page
));
266 "page:%p flags:%p count:%d mapcount:%d mapping:%p index:%lx\n",
267 page
, (void *)page
->flags
, page_count(page
),
268 page_mapcount(page
), page
->mapping
, page
->index
);
272 /* Leave bad fields for debug, except PageBuddy could make trouble */
273 __ClearPageBuddy(page
);
274 add_taint(TAINT_BAD_PAGE
);
278 * Higher-order pages are called "compound pages". They are structured thusly:
280 * The first PAGE_SIZE page is called the "head page".
282 * The remaining PAGE_SIZE pages are called "tail pages".
284 * All pages have PG_compound set. All pages have their ->private pointing at
285 * the head page (even the head page has this).
287 * The first tail page's ->lru.next holds the address of the compound page's
288 * put_page() function. Its ->lru.prev holds the order of allocation.
289 * This usage means that zero-order pages may not be compound.
292 static void free_compound_page(struct page
*page
)
294 __free_pages_ok(page
, compound_order(page
));
297 void prep_compound_page(struct page
*page
, unsigned long order
)
300 int nr_pages
= 1 << order
;
302 set_compound_page_dtor(page
, free_compound_page
);
303 set_compound_order(page
, order
);
305 for (i
= 1; i
< nr_pages
; i
++) {
306 struct page
*p
= page
+ i
;
309 p
->first_page
= page
;
313 static int destroy_compound_page(struct page
*page
, unsigned long order
)
316 int nr_pages
= 1 << order
;
319 if (unlikely(compound_order(page
) != order
) ||
320 unlikely(!PageHead(page
))) {
325 __ClearPageHead(page
);
327 for (i
= 1; i
< nr_pages
; i
++) {
328 struct page
*p
= page
+ i
;
330 if (unlikely(!PageTail(p
) || (p
->first_page
!= page
))) {
340 static inline void prep_zero_page(struct page
*page
, int order
, gfp_t gfp_flags
)
345 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
346 * and __GFP_HIGHMEM from hard or soft interrupt context.
348 VM_BUG_ON((gfp_flags
& __GFP_HIGHMEM
) && in_interrupt());
349 for (i
= 0; i
< (1 << order
); i
++)
350 clear_highpage(page
+ i
);
353 static inline void set_page_order(struct page
*page
, int order
)
355 set_page_private(page
, order
);
356 __SetPageBuddy(page
);
359 static inline void rmv_page_order(struct page
*page
)
361 __ClearPageBuddy(page
);
362 set_page_private(page
, 0);
366 * Locate the struct page for both the matching buddy in our
367 * pair (buddy1) and the combined O(n+1) page they form (page).
369 * 1) Any buddy B1 will have an order O twin B2 which satisfies
370 * the following equation:
372 * For example, if the starting buddy (buddy2) is #8 its order
374 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
376 * 2) Any buddy B will have an order O+1 parent P which
377 * satisfies the following equation:
380 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
382 static inline struct page
*
383 __page_find_buddy(struct page
*page
, unsigned long page_idx
, unsigned int order
)
385 unsigned long buddy_idx
= page_idx
^ (1 << order
);
387 return page
+ (buddy_idx
- page_idx
);
390 static inline unsigned long
391 __find_combined_index(unsigned long page_idx
, unsigned int order
)
393 return (page_idx
& ~(1 << order
));
397 * This function checks whether a page is free && is the buddy
398 * we can do coalesce a page and its buddy if
399 * (a) the buddy is not in a hole &&
400 * (b) the buddy is in the buddy system &&
401 * (c) a page and its buddy have the same order &&
402 * (d) a page and its buddy are in the same zone.
404 * For recording whether a page is in the buddy system, we use PG_buddy.
405 * Setting, clearing, and testing PG_buddy is serialized by zone->lock.
407 * For recording page's order, we use page_private(page).
409 static inline int page_is_buddy(struct page
*page
, struct page
*buddy
,
412 if (!pfn_valid_within(page_to_pfn(buddy
)))
415 if (page_zone_id(page
) != page_zone_id(buddy
))
418 if (PageBuddy(buddy
) && page_order(buddy
) == order
) {
419 VM_BUG_ON(page_count(buddy
) != 0);
426 * Freeing function for a buddy system allocator.
428 * The concept of a buddy system is to maintain direct-mapped table
429 * (containing bit values) for memory blocks of various "orders".
430 * The bottom level table contains the map for the smallest allocatable
431 * units of memory (here, pages), and each level above it describes
432 * pairs of units from the levels below, hence, "buddies".
433 * At a high level, all that happens here is marking the table entry
434 * at the bottom level available, and propagating the changes upward
435 * as necessary, plus some accounting needed to play nicely with other
436 * parts of the VM system.
437 * At each level, we keep a list of pages, which are heads of continuous
438 * free pages of length of (1 << order) and marked with PG_buddy. Page's
439 * order is recorded in page_private(page) field.
440 * So when we are allocating or freeing one, we can derive the state of the
441 * other. That is, if we allocate a small block, and both were
442 * free, the remainder of the region must be split into blocks.
443 * If a block is freed, and its buddy is also free, then this
444 * triggers coalescing into a block of larger size.
449 static inline void __free_one_page(struct page
*page
,
450 struct zone
*zone
, unsigned int order
,
453 unsigned long page_idx
;
455 if (unlikely(PageCompound(page
)))
456 if (unlikely(destroy_compound_page(page
, order
)))
459 VM_BUG_ON(migratetype
== -1);
461 page_idx
= page_to_pfn(page
) & ((1 << MAX_ORDER
) - 1);
463 VM_BUG_ON(page_idx
& ((1 << order
) - 1));
464 VM_BUG_ON(bad_range(zone
, page
));
466 while (order
< MAX_ORDER
-1) {
467 unsigned long combined_idx
;
470 buddy
= __page_find_buddy(page
, page_idx
, order
);
471 if (!page_is_buddy(page
, buddy
, order
))
474 /* Our buddy is free, merge with it and move up one order. */
475 list_del(&buddy
->lru
);
476 zone
->free_area
[order
].nr_free
--;
477 rmv_page_order(buddy
);
478 combined_idx
= __find_combined_index(page_idx
, order
);
479 page
= page
+ (combined_idx
- page_idx
);
480 page_idx
= combined_idx
;
483 set_page_order(page
, order
);
485 &zone
->free_area
[order
].free_list
[migratetype
]);
486 zone
->free_area
[order
].nr_free
++;
489 #ifdef CONFIG_HAVE_MLOCKED_PAGE_BIT
491 * free_page_mlock() -- clean up attempts to free and mlocked() page.
492 * Page should not be on lru, so no need to fix that up.
493 * free_pages_check() will verify...
495 static inline void free_page_mlock(struct page
*page
)
497 __dec_zone_page_state(page
, NR_MLOCK
);
498 __count_vm_event(UNEVICTABLE_MLOCKFREED
);
501 static void free_page_mlock(struct page
*page
) { }
504 static inline int free_pages_check(struct page
*page
)
506 if (unlikely(page_mapcount(page
) |
507 (page
->mapping
!= NULL
) |
508 (atomic_read(&page
->_count
) != 0) |
509 (page
->flags
& PAGE_FLAGS_CHECK_AT_FREE
))) {
513 if (page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)
514 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
519 * Frees a list of pages.
520 * Assumes all pages on list are in same zone, and of same order.
521 * count is the number of pages to free.
523 * If the zone was previously in an "all pages pinned" state then look to
524 * see if this freeing clears that state.
526 * And clear the zone's pages_scanned counter, to hold off the "all pages are
527 * pinned" detection logic.
529 static void free_pages_bulk(struct zone
*zone
, int count
,
530 struct list_head
*list
, int order
)
532 spin_lock(&zone
->lock
);
533 zone_clear_flag(zone
, ZONE_ALL_UNRECLAIMABLE
);
534 zone
->pages_scanned
= 0;
536 __mod_zone_page_state(zone
, NR_FREE_PAGES
, count
<< order
);
540 VM_BUG_ON(list_empty(list
));
541 page
= list_entry(list
->prev
, struct page
, lru
);
542 /* have to delete it as __free_one_page list manipulates */
543 list_del(&page
->lru
);
544 __free_one_page(page
, zone
, order
, page_private(page
));
546 spin_unlock(&zone
->lock
);
549 static void free_one_page(struct zone
*zone
, struct page
*page
, int order
,
552 spin_lock(&zone
->lock
);
553 zone_clear_flag(zone
, ZONE_ALL_UNRECLAIMABLE
);
554 zone
->pages_scanned
= 0;
556 __mod_zone_page_state(zone
, NR_FREE_PAGES
, 1 << order
);
557 __free_one_page(page
, zone
, order
, migratetype
);
558 spin_unlock(&zone
->lock
);
561 static void __free_pages_ok(struct page
*page
, unsigned int order
)
566 int wasMlocked
= TestClearPageMlocked(page
);
568 kmemcheck_free_shadow(page
, order
);
570 for (i
= 0 ; i
< (1 << order
) ; ++i
)
571 bad
+= free_pages_check(page
+ i
);
575 if (!PageHighMem(page
)) {
576 debug_check_no_locks_freed(page_address(page
),PAGE_SIZE
<<order
);
577 debug_check_no_obj_freed(page_address(page
),
580 arch_free_page(page
, order
);
581 kernel_map_pages(page
, 1 << order
, 0);
583 local_irq_save(flags
);
584 if (unlikely(wasMlocked
))
585 free_page_mlock(page
);
586 __count_vm_events(PGFREE
, 1 << order
);
587 free_one_page(page_zone(page
), page
, order
,
588 get_pageblock_migratetype(page
));
589 local_irq_restore(flags
);
593 * permit the bootmem allocator to evade page validation on high-order frees
595 void __meminit
__free_pages_bootmem(struct page
*page
, unsigned int order
)
598 __ClearPageReserved(page
);
599 set_page_count(page
, 0);
600 set_page_refcounted(page
);
606 for (loop
= 0; loop
< BITS_PER_LONG
; loop
++) {
607 struct page
*p
= &page
[loop
];
609 if (loop
+ 1 < BITS_PER_LONG
)
611 __ClearPageReserved(p
);
612 set_page_count(p
, 0);
615 set_page_refcounted(page
);
616 __free_pages(page
, order
);
622 * The order of subdivision here is critical for the IO subsystem.
623 * Please do not alter this order without good reasons and regression
624 * testing. Specifically, as large blocks of memory are subdivided,
625 * the order in which smaller blocks are delivered depends on the order
626 * they're subdivided in this function. This is the primary factor
627 * influencing the order in which pages are delivered to the IO
628 * subsystem according to empirical testing, and this is also justified
629 * by considering the behavior of a buddy system containing a single
630 * large block of memory acted on by a series of small allocations.
631 * This behavior is a critical factor in sglist merging's success.
635 static inline void expand(struct zone
*zone
, struct page
*page
,
636 int low
, int high
, struct free_area
*area
,
639 unsigned long size
= 1 << high
;
645 VM_BUG_ON(bad_range(zone
, &page
[size
]));
646 list_add(&page
[size
].lru
, &area
->free_list
[migratetype
]);
648 set_page_order(&page
[size
], high
);
653 * This page is about to be returned from the page allocator
655 static inline int check_new_page(struct page
*page
)
657 if (unlikely(page_mapcount(page
) |
658 (page
->mapping
!= NULL
) |
659 (atomic_read(&page
->_count
) != 0) |
660 (page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
))) {
667 static int prep_new_page(struct page
*page
, int order
, gfp_t gfp_flags
)
671 for (i
= 0; i
< (1 << order
); i
++) {
672 struct page
*p
= page
+ i
;
673 if (unlikely(check_new_page(p
)))
677 set_page_private(page
, 0);
678 set_page_refcounted(page
);
680 arch_alloc_page(page
, order
);
681 kernel_map_pages(page
, 1 << order
, 1);
683 if (gfp_flags
& __GFP_ZERO
)
684 prep_zero_page(page
, order
, gfp_flags
);
686 if (order
&& (gfp_flags
& __GFP_COMP
))
687 prep_compound_page(page
, order
);
693 * Go through the free lists for the given migratetype and remove
694 * the smallest available page from the freelists
697 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
700 unsigned int current_order
;
701 struct free_area
* area
;
704 /* Find a page of the appropriate size in the preferred list */
705 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
706 area
= &(zone
->free_area
[current_order
]);
707 if (list_empty(&area
->free_list
[migratetype
]))
710 page
= list_entry(area
->free_list
[migratetype
].next
,
712 list_del(&page
->lru
);
713 rmv_page_order(page
);
715 expand(zone
, page
, order
, current_order
, area
, migratetype
);
724 * This array describes the order lists are fallen back to when
725 * the free lists for the desirable migrate type are depleted
727 static int fallbacks
[MIGRATE_TYPES
][MIGRATE_TYPES
-1] = {
728 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
729 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
730 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_RESERVE
},
731 [MIGRATE_RESERVE
] = { MIGRATE_RESERVE
, MIGRATE_RESERVE
, MIGRATE_RESERVE
}, /* Never used */
735 * Move the free pages in a range to the free lists of the requested type.
736 * Note that start_page and end_pages are not aligned on a pageblock
737 * boundary. If alignment is required, use move_freepages_block()
739 static int move_freepages(struct zone
*zone
,
740 struct page
*start_page
, struct page
*end_page
,
747 #ifndef CONFIG_HOLES_IN_ZONE
749 * page_zone is not safe to call in this context when
750 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
751 * anyway as we check zone boundaries in move_freepages_block().
752 * Remove at a later date when no bug reports exist related to
753 * grouping pages by mobility
755 BUG_ON(page_zone(start_page
) != page_zone(end_page
));
758 for (page
= start_page
; page
<= end_page
;) {
759 /* Make sure we are not inadvertently changing nodes */
760 VM_BUG_ON(page_to_nid(page
) != zone_to_nid(zone
));
762 if (!pfn_valid_within(page_to_pfn(page
))) {
767 if (!PageBuddy(page
)) {
772 order
= page_order(page
);
773 list_del(&page
->lru
);
775 &zone
->free_area
[order
].free_list
[migratetype
]);
777 pages_moved
+= 1 << order
;
783 static int move_freepages_block(struct zone
*zone
, struct page
*page
,
786 unsigned long start_pfn
, end_pfn
;
787 struct page
*start_page
, *end_page
;
789 start_pfn
= page_to_pfn(page
);
790 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
791 start_page
= pfn_to_page(start_pfn
);
792 end_page
= start_page
+ pageblock_nr_pages
- 1;
793 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
795 /* Do not cross zone boundaries */
796 if (start_pfn
< zone
->zone_start_pfn
)
798 if (end_pfn
>= zone
->zone_start_pfn
+ zone
->spanned_pages
)
801 return move_freepages(zone
, start_page
, end_page
, migratetype
);
804 /* Remove an element from the buddy allocator from the fallback list */
805 static inline struct page
*
806 __rmqueue_fallback(struct zone
*zone
, int order
, int start_migratetype
)
808 struct free_area
* area
;
813 /* Find the largest possible block of pages in the other list */
814 for (current_order
= MAX_ORDER
-1; current_order
>= order
;
816 for (i
= 0; i
< MIGRATE_TYPES
- 1; i
++) {
817 migratetype
= fallbacks
[start_migratetype
][i
];
819 /* MIGRATE_RESERVE handled later if necessary */
820 if (migratetype
== MIGRATE_RESERVE
)
823 area
= &(zone
->free_area
[current_order
]);
824 if (list_empty(&area
->free_list
[migratetype
]))
827 page
= list_entry(area
->free_list
[migratetype
].next
,
832 * If breaking a large block of pages, move all free
833 * pages to the preferred allocation list. If falling
834 * back for a reclaimable kernel allocation, be more
835 * agressive about taking ownership of free pages
837 if (unlikely(current_order
>= (pageblock_order
>> 1)) ||
838 start_migratetype
== MIGRATE_RECLAIMABLE
||
839 page_group_by_mobility_disabled
) {
841 pages
= move_freepages_block(zone
, page
,
844 /* Claim the whole block if over half of it is free */
845 if (pages
>= (1 << (pageblock_order
-1)) ||
846 page_group_by_mobility_disabled
)
847 set_pageblock_migratetype(page
,
850 migratetype
= start_migratetype
;
853 /* Remove the page from the freelists */
854 list_del(&page
->lru
);
855 rmv_page_order(page
);
857 if (current_order
== pageblock_order
)
858 set_pageblock_migratetype(page
,
861 expand(zone
, page
, order
, current_order
, area
, migratetype
);
870 * Do the hard work of removing an element from the buddy allocator.
871 * Call me with the zone->lock already held.
873 static struct page
*__rmqueue(struct zone
*zone
, unsigned int order
,
879 page
= __rmqueue_smallest(zone
, order
, migratetype
);
881 if (unlikely(!page
) && migratetype
!= MIGRATE_RESERVE
) {
882 page
= __rmqueue_fallback(zone
, order
, migratetype
);
885 * Use MIGRATE_RESERVE rather than fail an allocation. goto
886 * is used because __rmqueue_smallest is an inline function
887 * and we want just one call site
890 migratetype
= MIGRATE_RESERVE
;
899 * Obtain a specified number of elements from the buddy allocator, all under
900 * a single hold of the lock, for efficiency. Add them to the supplied list.
901 * Returns the number of new pages which were placed at *list.
903 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
904 unsigned long count
, struct list_head
*list
,
905 int migratetype
, int cold
)
909 spin_lock(&zone
->lock
);
910 for (i
= 0; i
< count
; ++i
) {
911 struct page
*page
= __rmqueue(zone
, order
, migratetype
);
912 if (unlikely(page
== NULL
))
916 * Split buddy pages returned by expand() are received here
917 * in physical page order. The page is added to the callers and
918 * list and the list head then moves forward. From the callers
919 * perspective, the linked list is ordered by page number in
920 * some conditions. This is useful for IO devices that can
921 * merge IO requests if the physical pages are ordered
924 if (likely(cold
== 0))
925 list_add(&page
->lru
, list
);
927 list_add_tail(&page
->lru
, list
);
928 set_page_private(page
, migratetype
);
931 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
932 spin_unlock(&zone
->lock
);
938 * Called from the vmstat counter updater to drain pagesets of this
939 * currently executing processor on remote nodes after they have
942 * Note that this function must be called with the thread pinned to
943 * a single processor.
945 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
950 local_irq_save(flags
);
951 if (pcp
->count
>= pcp
->batch
)
952 to_drain
= pcp
->batch
;
954 to_drain
= pcp
->count
;
955 free_pages_bulk(zone
, to_drain
, &pcp
->list
, 0);
956 pcp
->count
-= to_drain
;
957 local_irq_restore(flags
);
962 * Drain pages of the indicated processor.
964 * The processor must either be the current processor and the
965 * thread pinned to the current processor or a processor that
968 static void drain_pages(unsigned int cpu
)
973 for_each_populated_zone(zone
) {
974 struct per_cpu_pageset
*pset
;
975 struct per_cpu_pages
*pcp
;
977 pset
= zone_pcp(zone
, cpu
);
980 local_irq_save(flags
);
981 free_pages_bulk(zone
, pcp
->count
, &pcp
->list
, 0);
983 local_irq_restore(flags
);
988 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
990 void drain_local_pages(void *arg
)
992 drain_pages(smp_processor_id());
996 * Spill all the per-cpu pages from all CPUs back into the buddy allocator
998 void drain_all_pages(void)
1000 on_each_cpu(drain_local_pages
, NULL
, 1);
1003 #ifdef CONFIG_HIBERNATION
1005 void mark_free_pages(struct zone
*zone
)
1007 unsigned long pfn
, max_zone_pfn
;
1008 unsigned long flags
;
1010 struct list_head
*curr
;
1012 if (!zone
->spanned_pages
)
1015 spin_lock_irqsave(&zone
->lock
, flags
);
1017 max_zone_pfn
= zone
->zone_start_pfn
+ zone
->spanned_pages
;
1018 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
1019 if (pfn_valid(pfn
)) {
1020 struct page
*page
= pfn_to_page(pfn
);
1022 if (!swsusp_page_is_forbidden(page
))
1023 swsusp_unset_page_free(page
);
1026 for_each_migratetype_order(order
, t
) {
1027 list_for_each(curr
, &zone
->free_area
[order
].free_list
[t
]) {
1030 pfn
= page_to_pfn(list_entry(curr
, struct page
, lru
));
1031 for (i
= 0; i
< (1UL << order
); i
++)
1032 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
1035 spin_unlock_irqrestore(&zone
->lock
, flags
);
1037 #endif /* CONFIG_PM */
1040 * Free a 0-order page
1042 static void free_hot_cold_page(struct page
*page
, int cold
)
1044 struct zone
*zone
= page_zone(page
);
1045 struct per_cpu_pages
*pcp
;
1046 unsigned long flags
;
1047 int wasMlocked
= TestClearPageMlocked(page
);
1049 kmemcheck_free_shadow(page
, 0);
1052 page
->mapping
= NULL
;
1053 if (free_pages_check(page
))
1056 if (!PageHighMem(page
)) {
1057 debug_check_no_locks_freed(page_address(page
), PAGE_SIZE
);
1058 debug_check_no_obj_freed(page_address(page
), PAGE_SIZE
);
1060 arch_free_page(page
, 0);
1061 kernel_map_pages(page
, 1, 0);
1063 pcp
= &zone_pcp(zone
, get_cpu())->pcp
;
1064 set_page_private(page
, get_pageblock_migratetype(page
));
1065 local_irq_save(flags
);
1066 if (unlikely(wasMlocked
))
1067 free_page_mlock(page
);
1068 __count_vm_event(PGFREE
);
1071 list_add_tail(&page
->lru
, &pcp
->list
);
1073 list_add(&page
->lru
, &pcp
->list
);
1075 if (pcp
->count
>= pcp
->high
) {
1076 free_pages_bulk(zone
, pcp
->batch
, &pcp
->list
, 0);
1077 pcp
->count
-= pcp
->batch
;
1079 local_irq_restore(flags
);
1083 void free_hot_page(struct page
*page
)
1085 free_hot_cold_page(page
, 0);
1088 void free_cold_page(struct page
*page
)
1090 free_hot_cold_page(page
, 1);
1094 * split_page takes a non-compound higher-order page, and splits it into
1095 * n (1<<order) sub-pages: page[0..n]
1096 * Each sub-page must be freed individually.
1098 * Note: this is probably too low level an operation for use in drivers.
1099 * Please consult with lkml before using this in your driver.
1101 void split_page(struct page
*page
, unsigned int order
)
1105 VM_BUG_ON(PageCompound(page
));
1106 VM_BUG_ON(!page_count(page
));
1108 #ifdef CONFIG_KMEMCHECK
1110 * Split shadow pages too, because free(page[0]) would
1111 * otherwise free the whole shadow.
1113 if (kmemcheck_page_is_tracked(page
))
1114 split_page(virt_to_page(page
[0].shadow
), order
);
1117 for (i
= 1; i
< (1 << order
); i
++)
1118 set_page_refcounted(page
+ i
);
1122 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1123 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1127 struct page
*buffered_rmqueue(struct zone
*preferred_zone
,
1128 struct zone
*zone
, int order
, gfp_t gfp_flags
,
1131 unsigned long flags
;
1133 int cold
= !!(gfp_flags
& __GFP_COLD
);
1138 if (likely(order
== 0)) {
1139 struct per_cpu_pages
*pcp
;
1141 pcp
= &zone_pcp(zone
, cpu
)->pcp
;
1142 local_irq_save(flags
);
1144 pcp
->count
= rmqueue_bulk(zone
, 0,
1145 pcp
->batch
, &pcp
->list
,
1147 if (unlikely(!pcp
->count
))
1151 /* Find a page of the appropriate migrate type */
1153 list_for_each_entry_reverse(page
, &pcp
->list
, lru
)
1154 if (page_private(page
) == migratetype
)
1157 list_for_each_entry(page
, &pcp
->list
, lru
)
1158 if (page_private(page
) == migratetype
)
1162 /* Allocate more to the pcp list if necessary */
1163 if (unlikely(&page
->lru
== &pcp
->list
)) {
1164 pcp
->count
+= rmqueue_bulk(zone
, 0,
1165 pcp
->batch
, &pcp
->list
,
1167 page
= list_entry(pcp
->list
.next
, struct page
, lru
);
1170 list_del(&page
->lru
);
1173 if (unlikely(gfp_flags
& __GFP_NOFAIL
)) {
1175 * __GFP_NOFAIL is not to be used in new code.
1177 * All __GFP_NOFAIL callers should be fixed so that they
1178 * properly detect and handle allocation failures.
1180 * We most definitely don't want callers attempting to
1181 * allocate greater than order-1 page units with
1184 WARN_ON_ONCE(order
> 1);
1186 spin_lock_irqsave(&zone
->lock
, flags
);
1187 page
= __rmqueue(zone
, order
, migratetype
);
1188 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(1 << order
));
1189 spin_unlock(&zone
->lock
);
1194 __count_zone_vm_events(PGALLOC
, zone
, 1 << order
);
1195 zone_statistics(preferred_zone
, zone
);
1196 local_irq_restore(flags
);
1199 VM_BUG_ON(bad_range(zone
, page
));
1200 if (prep_new_page(page
, order
, gfp_flags
))
1205 local_irq_restore(flags
);
1210 /* The ALLOC_WMARK bits are used as an index to zone->watermark */
1211 #define ALLOC_WMARK_MIN WMARK_MIN
1212 #define ALLOC_WMARK_LOW WMARK_LOW
1213 #define ALLOC_WMARK_HIGH WMARK_HIGH
1214 #define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */
1216 /* Mask to get the watermark bits */
1217 #define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1)
1219 #define ALLOC_HARDER 0x10 /* try to alloc harder */
1220 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
1221 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
1223 #ifdef CONFIG_FAIL_PAGE_ALLOC
1225 static struct fail_page_alloc_attr
{
1226 struct fault_attr attr
;
1228 u32 ignore_gfp_highmem
;
1229 u32 ignore_gfp_wait
;
1232 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1234 struct dentry
*ignore_gfp_highmem_file
;
1235 struct dentry
*ignore_gfp_wait_file
;
1236 struct dentry
*min_order_file
;
1238 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1240 } fail_page_alloc
= {
1241 .attr
= FAULT_ATTR_INITIALIZER
,
1242 .ignore_gfp_wait
= 1,
1243 .ignore_gfp_highmem
= 1,
1247 static int __init
setup_fail_page_alloc(char *str
)
1249 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
1251 __setup("fail_page_alloc=", setup_fail_page_alloc
);
1253 static int should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1255 if (order
< fail_page_alloc
.min_order
)
1257 if (gfp_mask
& __GFP_NOFAIL
)
1259 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
1261 if (fail_page_alloc
.ignore_gfp_wait
&& (gfp_mask
& __GFP_WAIT
))
1264 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
1267 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1269 static int __init
fail_page_alloc_debugfs(void)
1271 mode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
1275 err
= init_fault_attr_dentries(&fail_page_alloc
.attr
,
1279 dir
= fail_page_alloc
.attr
.dentries
.dir
;
1281 fail_page_alloc
.ignore_gfp_wait_file
=
1282 debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
1283 &fail_page_alloc
.ignore_gfp_wait
);
1285 fail_page_alloc
.ignore_gfp_highmem_file
=
1286 debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
1287 &fail_page_alloc
.ignore_gfp_highmem
);
1288 fail_page_alloc
.min_order_file
=
1289 debugfs_create_u32("min-order", mode
, dir
,
1290 &fail_page_alloc
.min_order
);
1292 if (!fail_page_alloc
.ignore_gfp_wait_file
||
1293 !fail_page_alloc
.ignore_gfp_highmem_file
||
1294 !fail_page_alloc
.min_order_file
) {
1296 debugfs_remove(fail_page_alloc
.ignore_gfp_wait_file
);
1297 debugfs_remove(fail_page_alloc
.ignore_gfp_highmem_file
);
1298 debugfs_remove(fail_page_alloc
.min_order_file
);
1299 cleanup_fault_attr_dentries(&fail_page_alloc
.attr
);
1305 late_initcall(fail_page_alloc_debugfs
);
1307 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1309 #else /* CONFIG_FAIL_PAGE_ALLOC */
1311 static inline int should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1316 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1319 * Return 1 if free pages are above 'mark'. This takes into account the order
1320 * of the allocation.
1322 int zone_watermark_ok(struct zone
*z
, int order
, unsigned long mark
,
1323 int classzone_idx
, int alloc_flags
)
1325 /* free_pages my go negative - that's OK */
1327 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
) - (1 << order
) + 1;
1330 if (alloc_flags
& ALLOC_HIGH
)
1332 if (alloc_flags
& ALLOC_HARDER
)
1335 if (free_pages
<= min
+ z
->lowmem_reserve
[classzone_idx
])
1337 for (o
= 0; o
< order
; o
++) {
1338 /* At the next order, this order's pages become unavailable */
1339 free_pages
-= z
->free_area
[o
].nr_free
<< o
;
1341 /* Require fewer higher order pages to be free */
1344 if (free_pages
<= min
)
1352 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1353 * skip over zones that are not allowed by the cpuset, or that have
1354 * been recently (in last second) found to be nearly full. See further
1355 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1356 * that have to skip over a lot of full or unallowed zones.
1358 * If the zonelist cache is present in the passed in zonelist, then
1359 * returns a pointer to the allowed node mask (either the current
1360 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1362 * If the zonelist cache is not available for this zonelist, does
1363 * nothing and returns NULL.
1365 * If the fullzones BITMAP in the zonelist cache is stale (more than
1366 * a second since last zap'd) then we zap it out (clear its bits.)
1368 * We hold off even calling zlc_setup, until after we've checked the
1369 * first zone in the zonelist, on the theory that most allocations will
1370 * be satisfied from that first zone, so best to examine that zone as
1371 * quickly as we can.
1373 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1375 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1376 nodemask_t
*allowednodes
; /* zonelist_cache approximation */
1378 zlc
= zonelist
->zlcache_ptr
;
1382 if (time_after(jiffies
, zlc
->last_full_zap
+ HZ
)) {
1383 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
1384 zlc
->last_full_zap
= jiffies
;
1387 allowednodes
= !in_interrupt() && (alloc_flags
& ALLOC_CPUSET
) ?
1388 &cpuset_current_mems_allowed
:
1389 &node_states
[N_HIGH_MEMORY
];
1390 return allowednodes
;
1394 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1395 * if it is worth looking at further for free memory:
1396 * 1) Check that the zone isn't thought to be full (doesn't have its
1397 * bit set in the zonelist_cache fullzones BITMAP).
1398 * 2) Check that the zones node (obtained from the zonelist_cache
1399 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1400 * Return true (non-zero) if zone is worth looking at further, or
1401 * else return false (zero) if it is not.
1403 * This check -ignores- the distinction between various watermarks,
1404 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1405 * found to be full for any variation of these watermarks, it will
1406 * be considered full for up to one second by all requests, unless
1407 * we are so low on memory on all allowed nodes that we are forced
1408 * into the second scan of the zonelist.
1410 * In the second scan we ignore this zonelist cache and exactly
1411 * apply the watermarks to all zones, even it is slower to do so.
1412 * We are low on memory in the second scan, and should leave no stone
1413 * unturned looking for a free page.
1415 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1416 nodemask_t
*allowednodes
)
1418 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1419 int i
; /* index of *z in zonelist zones */
1420 int n
; /* node that zone *z is on */
1422 zlc
= zonelist
->zlcache_ptr
;
1426 i
= z
- zonelist
->_zonerefs
;
1429 /* This zone is worth trying if it is allowed but not full */
1430 return node_isset(n
, *allowednodes
) && !test_bit(i
, zlc
->fullzones
);
1434 * Given 'z' scanning a zonelist, set the corresponding bit in
1435 * zlc->fullzones, so that subsequent attempts to allocate a page
1436 * from that zone don't waste time re-examining it.
1438 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1440 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1441 int i
; /* index of *z in zonelist zones */
1443 zlc
= zonelist
->zlcache_ptr
;
1447 i
= z
- zonelist
->_zonerefs
;
1449 set_bit(i
, zlc
->fullzones
);
1452 #else /* CONFIG_NUMA */
1454 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1459 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1460 nodemask_t
*allowednodes
)
1465 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1468 #endif /* CONFIG_NUMA */
1471 * get_page_from_freelist goes through the zonelist trying to allocate
1474 static struct page
*
1475 get_page_from_freelist(gfp_t gfp_mask
, nodemask_t
*nodemask
, unsigned int order
,
1476 struct zonelist
*zonelist
, int high_zoneidx
, int alloc_flags
,
1477 struct zone
*preferred_zone
, int migratetype
)
1480 struct page
*page
= NULL
;
1483 nodemask_t
*allowednodes
= NULL
;/* zonelist_cache approximation */
1484 int zlc_active
= 0; /* set if using zonelist_cache */
1485 int did_zlc_setup
= 0; /* just call zlc_setup() one time */
1487 classzone_idx
= zone_idx(preferred_zone
);
1490 * Scan zonelist, looking for a zone with enough free.
1491 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1493 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
1494 high_zoneidx
, nodemask
) {
1495 if (NUMA_BUILD
&& zlc_active
&&
1496 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
1498 if ((alloc_flags
& ALLOC_CPUSET
) &&
1499 !cpuset_zone_allowed_softwall(zone
, gfp_mask
))
1502 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
1503 if (!(alloc_flags
& ALLOC_NO_WATERMARKS
)) {
1507 mark
= zone
->watermark
[alloc_flags
& ALLOC_WMARK_MASK
];
1508 if (zone_watermark_ok(zone
, order
, mark
,
1509 classzone_idx
, alloc_flags
))
1512 if (zone_reclaim_mode
== 0)
1513 goto this_zone_full
;
1515 ret
= zone_reclaim(zone
, gfp_mask
, order
);
1517 case ZONE_RECLAIM_NOSCAN
:
1520 case ZONE_RECLAIM_FULL
:
1521 /* scanned but unreclaimable */
1522 goto this_zone_full
;
1524 /* did we reclaim enough */
1525 if (!zone_watermark_ok(zone
, order
, mark
,
1526 classzone_idx
, alloc_flags
))
1527 goto this_zone_full
;
1532 page
= buffered_rmqueue(preferred_zone
, zone
, order
,
1533 gfp_mask
, migratetype
);
1538 zlc_mark_zone_full(zonelist
, z
);
1540 if (NUMA_BUILD
&& !did_zlc_setup
&& nr_online_nodes
> 1) {
1542 * we do zlc_setup after the first zone is tried but only
1543 * if there are multiple nodes make it worthwhile
1545 allowednodes
= zlc_setup(zonelist
, alloc_flags
);
1551 if (unlikely(NUMA_BUILD
&& page
== NULL
&& zlc_active
)) {
1552 /* Disable zlc cache for second zonelist scan */
1560 should_alloc_retry(gfp_t gfp_mask
, unsigned int order
,
1561 unsigned long pages_reclaimed
)
1563 /* Do not loop if specifically requested */
1564 if (gfp_mask
& __GFP_NORETRY
)
1568 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
1569 * means __GFP_NOFAIL, but that may not be true in other
1572 if (order
<= PAGE_ALLOC_COSTLY_ORDER
)
1576 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
1577 * specified, then we retry until we no longer reclaim any pages
1578 * (above), or we've reclaimed an order of pages at least as
1579 * large as the allocation's order. In both cases, if the
1580 * allocation still fails, we stop retrying.
1582 if (gfp_mask
& __GFP_REPEAT
&& pages_reclaimed
< (1 << order
))
1586 * Don't let big-order allocations loop unless the caller
1587 * explicitly requests that.
1589 if (gfp_mask
& __GFP_NOFAIL
)
1595 static inline struct page
*
1596 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
1597 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1598 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
1603 /* Acquire the OOM killer lock for the zones in zonelist */
1604 if (!try_set_zone_oom(zonelist
, gfp_mask
)) {
1605 schedule_timeout_uninterruptible(1);
1610 * Go through the zonelist yet one more time, keep very high watermark
1611 * here, this is only to catch a parallel oom killing, we must fail if
1612 * we're still under heavy pressure.
1614 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
,
1615 order
, zonelist
, high_zoneidx
,
1616 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
,
1617 preferred_zone
, migratetype
);
1621 /* The OOM killer will not help higher order allocs */
1622 if (order
> PAGE_ALLOC_COSTLY_ORDER
&& !(gfp_mask
& __GFP_NOFAIL
))
1625 /* Exhausted what can be done so it's blamo time */
1626 out_of_memory(zonelist
, gfp_mask
, order
);
1629 clear_zonelist_oom(zonelist
, gfp_mask
);
1633 /* The really slow allocator path where we enter direct reclaim */
1634 static inline struct page
*
1635 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
1636 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1637 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
1638 int migratetype
, unsigned long *did_some_progress
)
1640 struct page
*page
= NULL
;
1641 struct reclaim_state reclaim_state
;
1642 struct task_struct
*p
= current
;
1646 /* We now go into synchronous reclaim */
1647 cpuset_memory_pressure_bump();
1650 * The task's cpuset might have expanded its set of allowable nodes
1652 p
->flags
|= PF_MEMALLOC
;
1653 lockdep_set_current_reclaim_state(gfp_mask
);
1654 reclaim_state
.reclaimed_slab
= 0;
1655 p
->reclaim_state
= &reclaim_state
;
1657 *did_some_progress
= try_to_free_pages(zonelist
, order
, gfp_mask
, nodemask
);
1659 p
->reclaim_state
= NULL
;
1660 lockdep_clear_current_reclaim_state();
1661 p
->flags
&= ~PF_MEMALLOC
;
1668 if (likely(*did_some_progress
))
1669 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
1670 zonelist
, high_zoneidx
,
1671 alloc_flags
, preferred_zone
,
1677 * This is called in the allocator slow-path if the allocation request is of
1678 * sufficient urgency to ignore watermarks and take other desperate measures
1680 static inline struct page
*
1681 __alloc_pages_high_priority(gfp_t gfp_mask
, unsigned int order
,
1682 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1683 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
1689 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
1690 zonelist
, high_zoneidx
, ALLOC_NO_WATERMARKS
,
1691 preferred_zone
, migratetype
);
1693 if (!page
&& gfp_mask
& __GFP_NOFAIL
)
1694 congestion_wait(BLK_RW_ASYNC
, HZ
/50);
1695 } while (!page
&& (gfp_mask
& __GFP_NOFAIL
));
1701 void wake_all_kswapd(unsigned int order
, struct zonelist
*zonelist
,
1702 enum zone_type high_zoneidx
)
1707 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
)
1708 wakeup_kswapd(zone
, order
);
1712 gfp_to_alloc_flags(gfp_t gfp_mask
)
1714 struct task_struct
*p
= current
;
1715 int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
1716 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
1718 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
1719 BUILD_BUG_ON(__GFP_HIGH
!= ALLOC_HIGH
);
1722 * The caller may dip into page reserves a bit more if the caller
1723 * cannot run direct reclaim, or if the caller has realtime scheduling
1724 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
1725 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
1727 alloc_flags
|= (gfp_mask
& __GFP_HIGH
);
1730 alloc_flags
|= ALLOC_HARDER
;
1732 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
1733 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1735 alloc_flags
&= ~ALLOC_CPUSET
;
1736 } else if (unlikely(rt_task(p
)))
1737 alloc_flags
|= ALLOC_HARDER
;
1739 if (likely(!(gfp_mask
& __GFP_NOMEMALLOC
))) {
1740 if (!in_interrupt() &&
1741 ((p
->flags
& PF_MEMALLOC
) ||
1742 unlikely(test_thread_flag(TIF_MEMDIE
))))
1743 alloc_flags
|= ALLOC_NO_WATERMARKS
;
1749 static inline struct page
*
1750 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
1751 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1752 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
1755 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
1756 struct page
*page
= NULL
;
1758 unsigned long pages_reclaimed
= 0;
1759 unsigned long did_some_progress
;
1760 struct task_struct
*p
= current
;
1763 * In the slowpath, we sanity check order to avoid ever trying to
1764 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
1765 * be using allocators in order of preference for an area that is
1768 if (order
>= MAX_ORDER
) {
1769 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
1774 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
1775 * __GFP_NOWARN set) should not cause reclaim since the subsystem
1776 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
1777 * using a larger set of nodes after it has established that the
1778 * allowed per node queues are empty and that nodes are
1781 if (NUMA_BUILD
&& (gfp_mask
& GFP_THISNODE
) == GFP_THISNODE
)
1784 wake_all_kswapd(order
, zonelist
, high_zoneidx
);
1787 * OK, we're below the kswapd watermark and have kicked background
1788 * reclaim. Now things get more complex, so set up alloc_flags according
1789 * to how we want to proceed.
1791 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
1794 /* This is the last chance, in general, before the goto nopage. */
1795 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
, zonelist
,
1796 high_zoneidx
, alloc_flags
& ~ALLOC_NO_WATERMARKS
,
1797 preferred_zone
, migratetype
);
1802 /* Allocate without watermarks if the context allows */
1803 if (alloc_flags
& ALLOC_NO_WATERMARKS
) {
1804 page
= __alloc_pages_high_priority(gfp_mask
, order
,
1805 zonelist
, high_zoneidx
, nodemask
,
1806 preferred_zone
, migratetype
);
1811 /* Atomic allocations - we can't balance anything */
1815 /* Avoid recursion of direct reclaim */
1816 if (p
->flags
& PF_MEMALLOC
)
1819 /* Avoid allocations with no watermarks from looping endlessly */
1820 if (test_thread_flag(TIF_MEMDIE
) && !(gfp_mask
& __GFP_NOFAIL
))
1823 /* Try direct reclaim and then allocating */
1824 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
,
1825 zonelist
, high_zoneidx
,
1827 alloc_flags
, preferred_zone
,
1828 migratetype
, &did_some_progress
);
1833 * If we failed to make any progress reclaiming, then we are
1834 * running out of options and have to consider going OOM
1836 if (!did_some_progress
) {
1837 if ((gfp_mask
& __GFP_FS
) && !(gfp_mask
& __GFP_NORETRY
)) {
1838 if (oom_killer_disabled
)
1840 page
= __alloc_pages_may_oom(gfp_mask
, order
,
1841 zonelist
, high_zoneidx
,
1842 nodemask
, preferred_zone
,
1848 * The OOM killer does not trigger for high-order
1849 * ~__GFP_NOFAIL allocations so if no progress is being
1850 * made, there are no other options and retrying is
1853 if (order
> PAGE_ALLOC_COSTLY_ORDER
&&
1854 !(gfp_mask
& __GFP_NOFAIL
))
1861 /* Check if we should retry the allocation */
1862 pages_reclaimed
+= did_some_progress
;
1863 if (should_alloc_retry(gfp_mask
, order
, pages_reclaimed
)) {
1864 /* Wait for some write requests to complete then retry */
1865 congestion_wait(BLK_RW_ASYNC
, HZ
/50);
1870 if (!(gfp_mask
& __GFP_NOWARN
) && printk_ratelimit()) {
1871 printk(KERN_WARNING
"%s: page allocation failure."
1872 " order:%d, mode:0x%x\n",
1873 p
->comm
, order
, gfp_mask
);
1879 if (kmemcheck_enabled
)
1880 kmemcheck_pagealloc_alloc(page
, order
, gfp_mask
);
1886 * This is the 'heart' of the zoned buddy allocator.
1889 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
,
1890 struct zonelist
*zonelist
, nodemask_t
*nodemask
)
1892 enum zone_type high_zoneidx
= gfp_zone(gfp_mask
);
1893 struct zone
*preferred_zone
;
1895 int migratetype
= allocflags_to_migratetype(gfp_mask
);
1897 gfp_mask
&= gfp_allowed_mask
;
1899 lockdep_trace_alloc(gfp_mask
);
1901 might_sleep_if(gfp_mask
& __GFP_WAIT
);
1903 if (should_fail_alloc_page(gfp_mask
, order
))
1907 * Check the zones suitable for the gfp_mask contain at least one
1908 * valid zone. It's possible to have an empty zonelist as a result
1909 * of GFP_THISNODE and a memoryless node
1911 if (unlikely(!zonelist
->_zonerefs
->zone
))
1914 /* The preferred zone is used for statistics later */
1915 first_zones_zonelist(zonelist
, high_zoneidx
, nodemask
, &preferred_zone
);
1916 if (!preferred_zone
)
1919 /* First allocation attempt */
1920 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
, order
,
1921 zonelist
, high_zoneidx
, ALLOC_WMARK_LOW
|ALLOC_CPUSET
,
1922 preferred_zone
, migratetype
);
1923 if (unlikely(!page
))
1924 page
= __alloc_pages_slowpath(gfp_mask
, order
,
1925 zonelist
, high_zoneidx
, nodemask
,
1926 preferred_zone
, migratetype
);
1930 EXPORT_SYMBOL(__alloc_pages_nodemask
);
1933 * Common helper functions.
1935 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
1938 page
= alloc_pages(gfp_mask
, order
);
1941 return (unsigned long) page_address(page
);
1944 EXPORT_SYMBOL(__get_free_pages
);
1946 unsigned long get_zeroed_page(gfp_t gfp_mask
)
1951 * get_zeroed_page() returns a 32-bit address, which cannot represent
1954 VM_BUG_ON((gfp_mask
& __GFP_HIGHMEM
) != 0);
1956 page
= alloc_pages(gfp_mask
| __GFP_ZERO
, 0);
1958 return (unsigned long) page_address(page
);
1962 EXPORT_SYMBOL(get_zeroed_page
);
1964 void __pagevec_free(struct pagevec
*pvec
)
1966 int i
= pagevec_count(pvec
);
1969 free_hot_cold_page(pvec
->pages
[i
], pvec
->cold
);
1972 void __free_pages(struct page
*page
, unsigned int order
)
1974 if (put_page_testzero(page
)) {
1976 free_hot_page(page
);
1978 __free_pages_ok(page
, order
);
1982 EXPORT_SYMBOL(__free_pages
);
1984 void free_pages(unsigned long addr
, unsigned int order
)
1987 VM_BUG_ON(!virt_addr_valid((void *)addr
));
1988 __free_pages(virt_to_page((void *)addr
), order
);
1992 EXPORT_SYMBOL(free_pages
);
1995 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
1996 * @size: the number of bytes to allocate
1997 * @gfp_mask: GFP flags for the allocation
1999 * This function is similar to alloc_pages(), except that it allocates the
2000 * minimum number of pages to satisfy the request. alloc_pages() can only
2001 * allocate memory in power-of-two pages.
2003 * This function is also limited by MAX_ORDER.
2005 * Memory allocated by this function must be released by free_pages_exact().
2007 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
2009 unsigned int order
= get_order(size
);
2012 addr
= __get_free_pages(gfp_mask
, order
);
2014 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
2015 unsigned long used
= addr
+ PAGE_ALIGN(size
);
2017 split_page(virt_to_page((void *)addr
), order
);
2018 while (used
< alloc_end
) {
2024 return (void *)addr
;
2026 EXPORT_SYMBOL(alloc_pages_exact
);
2029 * free_pages_exact - release memory allocated via alloc_pages_exact()
2030 * @virt: the value returned by alloc_pages_exact.
2031 * @size: size of allocation, same value as passed to alloc_pages_exact().
2033 * Release the memory allocated by a previous call to alloc_pages_exact.
2035 void free_pages_exact(void *virt
, size_t size
)
2037 unsigned long addr
= (unsigned long)virt
;
2038 unsigned long end
= addr
+ PAGE_ALIGN(size
);
2040 while (addr
< end
) {
2045 EXPORT_SYMBOL(free_pages_exact
);
2047 static unsigned int nr_free_zone_pages(int offset
)
2052 /* Just pick one node, since fallback list is circular */
2053 unsigned int sum
= 0;
2055 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
2057 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
2058 unsigned long size
= zone
->present_pages
;
2059 unsigned long high
= high_wmark_pages(zone
);
2068 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2070 unsigned int nr_free_buffer_pages(void)
2072 return nr_free_zone_pages(gfp_zone(GFP_USER
));
2074 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
2077 * Amount of free RAM allocatable within all zones
2079 unsigned int nr_free_pagecache_pages(void)
2081 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
2084 static inline void show_node(struct zone
*zone
)
2087 printk("Node %d ", zone_to_nid(zone
));
2090 void si_meminfo(struct sysinfo
*val
)
2092 val
->totalram
= totalram_pages
;
2094 val
->freeram
= global_page_state(NR_FREE_PAGES
);
2095 val
->bufferram
= nr_blockdev_pages();
2096 val
->totalhigh
= totalhigh_pages
;
2097 val
->freehigh
= nr_free_highpages();
2098 val
->mem_unit
= PAGE_SIZE
;
2101 EXPORT_SYMBOL(si_meminfo
);
2104 void si_meminfo_node(struct sysinfo
*val
, int nid
)
2106 pg_data_t
*pgdat
= NODE_DATA(nid
);
2108 val
->totalram
= pgdat
->node_present_pages
;
2109 val
->freeram
= node_page_state(nid
, NR_FREE_PAGES
);
2110 #ifdef CONFIG_HIGHMEM
2111 val
->totalhigh
= pgdat
->node_zones
[ZONE_HIGHMEM
].present_pages
;
2112 val
->freehigh
= zone_page_state(&pgdat
->node_zones
[ZONE_HIGHMEM
],
2118 val
->mem_unit
= PAGE_SIZE
;
2122 #define K(x) ((x) << (PAGE_SHIFT-10))
2125 * Show free area list (used inside shift_scroll-lock stuff)
2126 * We also calculate the percentage fragmentation. We do this by counting the
2127 * memory on each free list with the exception of the first item on the list.
2129 void show_free_areas(void)
2134 for_each_populated_zone(zone
) {
2136 printk("%s per-cpu:\n", zone
->name
);
2138 for_each_online_cpu(cpu
) {
2139 struct per_cpu_pageset
*pageset
;
2141 pageset
= zone_pcp(zone
, cpu
);
2143 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2144 cpu
, pageset
->pcp
.high
,
2145 pageset
->pcp
.batch
, pageset
->pcp
.count
);
2149 printk("Active_anon:%lu active_file:%lu inactive_anon:%lu\n"
2150 " inactive_file:%lu"
2152 " dirty:%lu writeback:%lu unstable:%lu\n"
2153 " free:%lu slab:%lu mapped:%lu pagetables:%lu bounce:%lu\n",
2154 global_page_state(NR_ACTIVE_ANON
),
2155 global_page_state(NR_ACTIVE_FILE
),
2156 global_page_state(NR_INACTIVE_ANON
),
2157 global_page_state(NR_INACTIVE_FILE
),
2158 global_page_state(NR_UNEVICTABLE
),
2159 global_page_state(NR_FILE_DIRTY
),
2160 global_page_state(NR_WRITEBACK
),
2161 global_page_state(NR_UNSTABLE_NFS
),
2162 global_page_state(NR_FREE_PAGES
),
2163 global_page_state(NR_SLAB_RECLAIMABLE
) +
2164 global_page_state(NR_SLAB_UNRECLAIMABLE
),
2165 global_page_state(NR_FILE_MAPPED
),
2166 global_page_state(NR_PAGETABLE
),
2167 global_page_state(NR_BOUNCE
));
2169 for_each_populated_zone(zone
) {
2178 " active_anon:%lukB"
2179 " inactive_anon:%lukB"
2180 " active_file:%lukB"
2181 " inactive_file:%lukB"
2182 " unevictable:%lukB"
2184 " pages_scanned:%lu"
2185 " all_unreclaimable? %s"
2188 K(zone_page_state(zone
, NR_FREE_PAGES
)),
2189 K(min_wmark_pages(zone
)),
2190 K(low_wmark_pages(zone
)),
2191 K(high_wmark_pages(zone
)),
2192 K(zone_page_state(zone
, NR_ACTIVE_ANON
)),
2193 K(zone_page_state(zone
, NR_INACTIVE_ANON
)),
2194 K(zone_page_state(zone
, NR_ACTIVE_FILE
)),
2195 K(zone_page_state(zone
, NR_INACTIVE_FILE
)),
2196 K(zone_page_state(zone
, NR_UNEVICTABLE
)),
2197 K(zone
->present_pages
),
2198 zone
->pages_scanned
,
2199 (zone_is_all_unreclaimable(zone
) ? "yes" : "no")
2201 printk("lowmem_reserve[]:");
2202 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
2203 printk(" %lu", zone
->lowmem_reserve
[i
]);
2207 for_each_populated_zone(zone
) {
2208 unsigned long nr
[MAX_ORDER
], flags
, order
, total
= 0;
2211 printk("%s: ", zone
->name
);
2213 spin_lock_irqsave(&zone
->lock
, flags
);
2214 for (order
= 0; order
< MAX_ORDER
; order
++) {
2215 nr
[order
] = zone
->free_area
[order
].nr_free
;
2216 total
+= nr
[order
] << order
;
2218 spin_unlock_irqrestore(&zone
->lock
, flags
);
2219 for (order
= 0; order
< MAX_ORDER
; order
++)
2220 printk("%lu*%lukB ", nr
[order
], K(1UL) << order
);
2221 printk("= %lukB\n", K(total
));
2224 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES
));
2226 show_swap_cache_info();
2229 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
2231 zoneref
->zone
= zone
;
2232 zoneref
->zone_idx
= zone_idx(zone
);
2236 * Builds allocation fallback zone lists.
2238 * Add all populated zones of a node to the zonelist.
2240 static int build_zonelists_node(pg_data_t
*pgdat
, struct zonelist
*zonelist
,
2241 int nr_zones
, enum zone_type zone_type
)
2245 BUG_ON(zone_type
>= MAX_NR_ZONES
);
2250 zone
= pgdat
->node_zones
+ zone_type
;
2251 if (populated_zone(zone
)) {
2252 zoneref_set_zone(zone
,
2253 &zonelist
->_zonerefs
[nr_zones
++]);
2254 check_highest_zone(zone_type
);
2257 } while (zone_type
);
2264 * 0 = automatic detection of better ordering.
2265 * 1 = order by ([node] distance, -zonetype)
2266 * 2 = order by (-zonetype, [node] distance)
2268 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
2269 * the same zonelist. So only NUMA can configure this param.
2271 #define ZONELIST_ORDER_DEFAULT 0
2272 #define ZONELIST_ORDER_NODE 1
2273 #define ZONELIST_ORDER_ZONE 2
2275 /* zonelist order in the kernel.
2276 * set_zonelist_order() will set this to NODE or ZONE.
2278 static int current_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
2279 static char zonelist_order_name
[3][8] = {"Default", "Node", "Zone"};
2283 /* The value user specified ....changed by config */
2284 static int user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
2285 /* string for sysctl */
2286 #define NUMA_ZONELIST_ORDER_LEN 16
2287 char numa_zonelist_order
[16] = "default";
2290 * interface for configure zonelist ordering.
2291 * command line option "numa_zonelist_order"
2292 * = "[dD]efault - default, automatic configuration.
2293 * = "[nN]ode - order by node locality, then by zone within node
2294 * = "[zZ]one - order by zone, then by locality within zone
2297 static int __parse_numa_zonelist_order(char *s
)
2299 if (*s
== 'd' || *s
== 'D') {
2300 user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
2301 } else if (*s
== 'n' || *s
== 'N') {
2302 user_zonelist_order
= ZONELIST_ORDER_NODE
;
2303 } else if (*s
== 'z' || *s
== 'Z') {
2304 user_zonelist_order
= ZONELIST_ORDER_ZONE
;
2307 "Ignoring invalid numa_zonelist_order value: "
2314 static __init
int setup_numa_zonelist_order(char *s
)
2317 return __parse_numa_zonelist_order(s
);
2320 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
2323 * sysctl handler for numa_zonelist_order
2325 int numa_zonelist_order_handler(ctl_table
*table
, int write
,
2326 struct file
*file
, void __user
*buffer
, size_t *length
,
2329 char saved_string
[NUMA_ZONELIST_ORDER_LEN
];
2333 strncpy(saved_string
, (char*)table
->data
,
2334 NUMA_ZONELIST_ORDER_LEN
);
2335 ret
= proc_dostring(table
, write
, file
, buffer
, length
, ppos
);
2339 int oldval
= user_zonelist_order
;
2340 if (__parse_numa_zonelist_order((char*)table
->data
)) {
2342 * bogus value. restore saved string
2344 strncpy((char*)table
->data
, saved_string
,
2345 NUMA_ZONELIST_ORDER_LEN
);
2346 user_zonelist_order
= oldval
;
2347 } else if (oldval
!= user_zonelist_order
)
2348 build_all_zonelists();
2354 #define MAX_NODE_LOAD (nr_online_nodes)
2355 static int node_load
[MAX_NUMNODES
];
2358 * find_next_best_node - find the next node that should appear in a given node's fallback list
2359 * @node: node whose fallback list we're appending
2360 * @used_node_mask: nodemask_t of already used nodes
2362 * We use a number of factors to determine which is the next node that should
2363 * appear on a given node's fallback list. The node should not have appeared
2364 * already in @node's fallback list, and it should be the next closest node
2365 * according to the distance array (which contains arbitrary distance values
2366 * from each node to each node in the system), and should also prefer nodes
2367 * with no CPUs, since presumably they'll have very little allocation pressure
2368 * on them otherwise.
2369 * It returns -1 if no node is found.
2371 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
2374 int min_val
= INT_MAX
;
2376 const struct cpumask
*tmp
= cpumask_of_node(0);
2378 /* Use the local node if we haven't already */
2379 if (!node_isset(node
, *used_node_mask
)) {
2380 node_set(node
, *used_node_mask
);
2384 for_each_node_state(n
, N_HIGH_MEMORY
) {
2386 /* Don't want a node to appear more than once */
2387 if (node_isset(n
, *used_node_mask
))
2390 /* Use the distance array to find the distance */
2391 val
= node_distance(node
, n
);
2393 /* Penalize nodes under us ("prefer the next node") */
2396 /* Give preference to headless and unused nodes */
2397 tmp
= cpumask_of_node(n
);
2398 if (!cpumask_empty(tmp
))
2399 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
2401 /* Slight preference for less loaded node */
2402 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
2403 val
+= node_load
[n
];
2405 if (val
< min_val
) {
2412 node_set(best_node
, *used_node_mask
);
2419 * Build zonelists ordered by node and zones within node.
2420 * This results in maximum locality--normal zone overflows into local
2421 * DMA zone, if any--but risks exhausting DMA zone.
2423 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int node
)
2426 struct zonelist
*zonelist
;
2428 zonelist
= &pgdat
->node_zonelists
[0];
2429 for (j
= 0; zonelist
->_zonerefs
[j
].zone
!= NULL
; j
++)
2431 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
2433 zonelist
->_zonerefs
[j
].zone
= NULL
;
2434 zonelist
->_zonerefs
[j
].zone_idx
= 0;
2438 * Build gfp_thisnode zonelists
2440 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
2443 struct zonelist
*zonelist
;
2445 zonelist
= &pgdat
->node_zonelists
[1];
2446 j
= build_zonelists_node(pgdat
, zonelist
, 0, MAX_NR_ZONES
- 1);
2447 zonelist
->_zonerefs
[j
].zone
= NULL
;
2448 zonelist
->_zonerefs
[j
].zone_idx
= 0;
2452 * Build zonelists ordered by zone and nodes within zones.
2453 * This results in conserving DMA zone[s] until all Normal memory is
2454 * exhausted, but results in overflowing to remote node while memory
2455 * may still exist in local DMA zone.
2457 static int node_order
[MAX_NUMNODES
];
2459 static void build_zonelists_in_zone_order(pg_data_t
*pgdat
, int nr_nodes
)
2462 int zone_type
; /* needs to be signed */
2464 struct zonelist
*zonelist
;
2466 zonelist
= &pgdat
->node_zonelists
[0];
2468 for (zone_type
= MAX_NR_ZONES
- 1; zone_type
>= 0; zone_type
--) {
2469 for (j
= 0; j
< nr_nodes
; j
++) {
2470 node
= node_order
[j
];
2471 z
= &NODE_DATA(node
)->node_zones
[zone_type
];
2472 if (populated_zone(z
)) {
2474 &zonelist
->_zonerefs
[pos
++]);
2475 check_highest_zone(zone_type
);
2479 zonelist
->_zonerefs
[pos
].zone
= NULL
;
2480 zonelist
->_zonerefs
[pos
].zone_idx
= 0;
2483 static int default_zonelist_order(void)
2486 unsigned long low_kmem_size
,total_size
;
2490 * ZONE_DMA and ZONE_DMA32 can be very small area in the sytem.
2491 * If they are really small and used heavily, the system can fall
2492 * into OOM very easily.
2493 * This function detect ZONE_DMA/DMA32 size and confgigures zone order.
2495 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
2498 for_each_online_node(nid
) {
2499 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
2500 z
= &NODE_DATA(nid
)->node_zones
[zone_type
];
2501 if (populated_zone(z
)) {
2502 if (zone_type
< ZONE_NORMAL
)
2503 low_kmem_size
+= z
->present_pages
;
2504 total_size
+= z
->present_pages
;
2508 if (!low_kmem_size
|| /* there are no DMA area. */
2509 low_kmem_size
> total_size
/2) /* DMA/DMA32 is big. */
2510 return ZONELIST_ORDER_NODE
;
2512 * look into each node's config.
2513 * If there is a node whose DMA/DMA32 memory is very big area on
2514 * local memory, NODE_ORDER may be suitable.
2516 average_size
= total_size
/
2517 (nodes_weight(node_states
[N_HIGH_MEMORY
]) + 1);
2518 for_each_online_node(nid
) {
2521 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
2522 z
= &NODE_DATA(nid
)->node_zones
[zone_type
];
2523 if (populated_zone(z
)) {
2524 if (zone_type
< ZONE_NORMAL
)
2525 low_kmem_size
+= z
->present_pages
;
2526 total_size
+= z
->present_pages
;
2529 if (low_kmem_size
&&
2530 total_size
> average_size
&& /* ignore small node */
2531 low_kmem_size
> total_size
* 70/100)
2532 return ZONELIST_ORDER_NODE
;
2534 return ZONELIST_ORDER_ZONE
;
2537 static void set_zonelist_order(void)
2539 if (user_zonelist_order
== ZONELIST_ORDER_DEFAULT
)
2540 current_zonelist_order
= default_zonelist_order();
2542 current_zonelist_order
= user_zonelist_order
;
2545 static void build_zonelists(pg_data_t
*pgdat
)
2549 nodemask_t used_mask
;
2550 int local_node
, prev_node
;
2551 struct zonelist
*zonelist
;
2552 int order
= current_zonelist_order
;
2554 /* initialize zonelists */
2555 for (i
= 0; i
< MAX_ZONELISTS
; i
++) {
2556 zonelist
= pgdat
->node_zonelists
+ i
;
2557 zonelist
->_zonerefs
[0].zone
= NULL
;
2558 zonelist
->_zonerefs
[0].zone_idx
= 0;
2561 /* NUMA-aware ordering of nodes */
2562 local_node
= pgdat
->node_id
;
2563 load
= nr_online_nodes
;
2564 prev_node
= local_node
;
2565 nodes_clear(used_mask
);
2567 memset(node_order
, 0, sizeof(node_order
));
2570 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
2571 int distance
= node_distance(local_node
, node
);
2574 * If another node is sufficiently far away then it is better
2575 * to reclaim pages in a zone before going off node.
2577 if (distance
> RECLAIM_DISTANCE
)
2578 zone_reclaim_mode
= 1;
2581 * We don't want to pressure a particular node.
2582 * So adding penalty to the first node in same
2583 * distance group to make it round-robin.
2585 if (distance
!= node_distance(local_node
, prev_node
))
2586 node_load
[node
] = load
;
2590 if (order
== ZONELIST_ORDER_NODE
)
2591 build_zonelists_in_node_order(pgdat
, node
);
2593 node_order
[j
++] = node
; /* remember order */
2596 if (order
== ZONELIST_ORDER_ZONE
) {
2597 /* calculate node order -- i.e., DMA last! */
2598 build_zonelists_in_zone_order(pgdat
, j
);
2601 build_thisnode_zonelists(pgdat
);
2604 /* Construct the zonelist performance cache - see further mmzone.h */
2605 static void build_zonelist_cache(pg_data_t
*pgdat
)
2607 struct zonelist
*zonelist
;
2608 struct zonelist_cache
*zlc
;
2611 zonelist
= &pgdat
->node_zonelists
[0];
2612 zonelist
->zlcache_ptr
= zlc
= &zonelist
->zlcache
;
2613 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
2614 for (z
= zonelist
->_zonerefs
; z
->zone
; z
++)
2615 zlc
->z_to_n
[z
- zonelist
->_zonerefs
] = zonelist_node_idx(z
);
2619 #else /* CONFIG_NUMA */
2621 static void set_zonelist_order(void)
2623 current_zonelist_order
= ZONELIST_ORDER_ZONE
;
2626 static void build_zonelists(pg_data_t
*pgdat
)
2628 int node
, local_node
;
2630 struct zonelist
*zonelist
;
2632 local_node
= pgdat
->node_id
;
2634 zonelist
= &pgdat
->node_zonelists
[0];
2635 j
= build_zonelists_node(pgdat
, zonelist
, 0, MAX_NR_ZONES
- 1);
2638 * Now we build the zonelist so that it contains the zones
2639 * of all the other nodes.
2640 * We don't want to pressure a particular node, so when
2641 * building the zones for node N, we make sure that the
2642 * zones coming right after the local ones are those from
2643 * node N+1 (modulo N)
2645 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
2646 if (!node_online(node
))
2648 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
2651 for (node
= 0; node
< local_node
; node
++) {
2652 if (!node_online(node
))
2654 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
2658 zonelist
->_zonerefs
[j
].zone
= NULL
;
2659 zonelist
->_zonerefs
[j
].zone_idx
= 0;
2662 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
2663 static void build_zonelist_cache(pg_data_t
*pgdat
)
2665 pgdat
->node_zonelists
[0].zlcache_ptr
= NULL
;
2668 #endif /* CONFIG_NUMA */
2670 /* return values int ....just for stop_machine() */
2671 static int __build_all_zonelists(void *dummy
)
2676 memset(node_load
, 0, sizeof(node_load
));
2678 for_each_online_node(nid
) {
2679 pg_data_t
*pgdat
= NODE_DATA(nid
);
2681 build_zonelists(pgdat
);
2682 build_zonelist_cache(pgdat
);
2687 void build_all_zonelists(void)
2689 set_zonelist_order();
2691 if (system_state
== SYSTEM_BOOTING
) {
2692 __build_all_zonelists(NULL
);
2693 mminit_verify_zonelist();
2694 cpuset_init_current_mems_allowed();
2696 /* we have to stop all cpus to guarantee there is no user
2698 stop_machine(__build_all_zonelists
, NULL
, NULL
);
2699 /* cpuset refresh routine should be here */
2701 vm_total_pages
= nr_free_pagecache_pages();
2703 * Disable grouping by mobility if the number of pages in the
2704 * system is too low to allow the mechanism to work. It would be
2705 * more accurate, but expensive to check per-zone. This check is
2706 * made on memory-hotadd so a system can start with mobility
2707 * disabled and enable it later
2709 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
2710 page_group_by_mobility_disabled
= 1;
2712 page_group_by_mobility_disabled
= 0;
2714 printk("Built %i zonelists in %s order, mobility grouping %s. "
2715 "Total pages: %ld\n",
2717 zonelist_order_name
[current_zonelist_order
],
2718 page_group_by_mobility_disabled
? "off" : "on",
2721 printk("Policy zone: %s\n", zone_names
[policy_zone
]);
2726 * Helper functions to size the waitqueue hash table.
2727 * Essentially these want to choose hash table sizes sufficiently
2728 * large so that collisions trying to wait on pages are rare.
2729 * But in fact, the number of active page waitqueues on typical
2730 * systems is ridiculously low, less than 200. So this is even
2731 * conservative, even though it seems large.
2733 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
2734 * waitqueues, i.e. the size of the waitq table given the number of pages.
2736 #define PAGES_PER_WAITQUEUE 256
2738 #ifndef CONFIG_MEMORY_HOTPLUG
2739 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
2741 unsigned long size
= 1;
2743 pages
/= PAGES_PER_WAITQUEUE
;
2745 while (size
< pages
)
2749 * Once we have dozens or even hundreds of threads sleeping
2750 * on IO we've got bigger problems than wait queue collision.
2751 * Limit the size of the wait table to a reasonable size.
2753 size
= min(size
, 4096UL);
2755 return max(size
, 4UL);
2759 * A zone's size might be changed by hot-add, so it is not possible to determine
2760 * a suitable size for its wait_table. So we use the maximum size now.
2762 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
2764 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
2765 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
2766 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
2768 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
2769 * or more by the traditional way. (See above). It equals:
2771 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
2772 * ia64(16K page size) : = ( 8G + 4M)byte.
2773 * powerpc (64K page size) : = (32G +16M)byte.
2775 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
2782 * This is an integer logarithm so that shifts can be used later
2783 * to extract the more random high bits from the multiplicative
2784 * hash function before the remainder is taken.
2786 static inline unsigned long wait_table_bits(unsigned long size
)
2791 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
2794 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
2795 * of blocks reserved is based on min_wmark_pages(zone). The memory within
2796 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
2797 * higher will lead to a bigger reserve which will get freed as contiguous
2798 * blocks as reclaim kicks in
2800 static void setup_zone_migrate_reserve(struct zone
*zone
)
2802 unsigned long start_pfn
, pfn
, end_pfn
;
2804 unsigned long reserve
, block_migratetype
;
2806 /* Get the start pfn, end pfn and the number of blocks to reserve */
2807 start_pfn
= zone
->zone_start_pfn
;
2808 end_pfn
= start_pfn
+ zone
->spanned_pages
;
2809 reserve
= roundup(min_wmark_pages(zone
), pageblock_nr_pages
) >>
2812 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= pageblock_nr_pages
) {
2813 if (!pfn_valid(pfn
))
2815 page
= pfn_to_page(pfn
);
2817 /* Watch out for overlapping nodes */
2818 if (page_to_nid(page
) != zone_to_nid(zone
))
2821 /* Blocks with reserved pages will never free, skip them. */
2822 if (PageReserved(page
))
2825 block_migratetype
= get_pageblock_migratetype(page
);
2827 /* If this block is reserved, account for it */
2828 if (reserve
> 0 && block_migratetype
== MIGRATE_RESERVE
) {
2833 /* Suitable for reserving if this block is movable */
2834 if (reserve
> 0 && block_migratetype
== MIGRATE_MOVABLE
) {
2835 set_pageblock_migratetype(page
, MIGRATE_RESERVE
);
2836 move_freepages_block(zone
, page
, MIGRATE_RESERVE
);
2842 * If the reserve is met and this is a previous reserved block,
2845 if (block_migratetype
== MIGRATE_RESERVE
) {
2846 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
2847 move_freepages_block(zone
, page
, MIGRATE_MOVABLE
);
2853 * Initially all pages are reserved - free ones are freed
2854 * up by free_all_bootmem() once the early boot process is
2855 * done. Non-atomic initialization, single-pass.
2857 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
2858 unsigned long start_pfn
, enum memmap_context context
)
2861 unsigned long end_pfn
= start_pfn
+ size
;
2865 if (highest_memmap_pfn
< end_pfn
- 1)
2866 highest_memmap_pfn
= end_pfn
- 1;
2868 z
= &NODE_DATA(nid
)->node_zones
[zone
];
2869 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
2871 * There can be holes in boot-time mem_map[]s
2872 * handed to this function. They do not
2873 * exist on hotplugged memory.
2875 if (context
== MEMMAP_EARLY
) {
2876 if (!early_pfn_valid(pfn
))
2878 if (!early_pfn_in_nid(pfn
, nid
))
2881 page
= pfn_to_page(pfn
);
2882 set_page_links(page
, zone
, nid
, pfn
);
2883 mminit_verify_page_links(page
, zone
, nid
, pfn
);
2884 init_page_count(page
);
2885 reset_page_mapcount(page
);
2886 SetPageReserved(page
);
2888 * Mark the block movable so that blocks are reserved for
2889 * movable at startup. This will force kernel allocations
2890 * to reserve their blocks rather than leaking throughout
2891 * the address space during boot when many long-lived
2892 * kernel allocations are made. Later some blocks near
2893 * the start are marked MIGRATE_RESERVE by
2894 * setup_zone_migrate_reserve()
2896 * bitmap is created for zone's valid pfn range. but memmap
2897 * can be created for invalid pages (for alignment)
2898 * check here not to call set_pageblock_migratetype() against
2901 if ((z
->zone_start_pfn
<= pfn
)
2902 && (pfn
< z
->zone_start_pfn
+ z
->spanned_pages
)
2903 && !(pfn
& (pageblock_nr_pages
- 1)))
2904 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
2906 INIT_LIST_HEAD(&page
->lru
);
2907 #ifdef WANT_PAGE_VIRTUAL
2908 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
2909 if (!is_highmem_idx(zone
))
2910 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
2915 static void __meminit
zone_init_free_lists(struct zone
*zone
)
2918 for_each_migratetype_order(order
, t
) {
2919 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
2920 zone
->free_area
[order
].nr_free
= 0;
2924 #ifndef __HAVE_ARCH_MEMMAP_INIT
2925 #define memmap_init(size, nid, zone, start_pfn) \
2926 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
2929 static int zone_batchsize(struct zone
*zone
)
2935 * The per-cpu-pages pools are set to around 1000th of the
2936 * size of the zone. But no more than 1/2 of a meg.
2938 * OK, so we don't know how big the cache is. So guess.
2940 batch
= zone
->present_pages
/ 1024;
2941 if (batch
* PAGE_SIZE
> 512 * 1024)
2942 batch
= (512 * 1024) / PAGE_SIZE
;
2943 batch
/= 4; /* We effectively *= 4 below */
2948 * Clamp the batch to a 2^n - 1 value. Having a power
2949 * of 2 value was found to be more likely to have
2950 * suboptimal cache aliasing properties in some cases.
2952 * For example if 2 tasks are alternately allocating
2953 * batches of pages, one task can end up with a lot
2954 * of pages of one half of the possible page colors
2955 * and the other with pages of the other colors.
2957 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
2962 /* The deferral and batching of frees should be suppressed under NOMMU
2965 * The problem is that NOMMU needs to be able to allocate large chunks
2966 * of contiguous memory as there's no hardware page translation to
2967 * assemble apparent contiguous memory from discontiguous pages.
2969 * Queueing large contiguous runs of pages for batching, however,
2970 * causes the pages to actually be freed in smaller chunks. As there
2971 * can be a significant delay between the individual batches being
2972 * recycled, this leads to the once large chunks of space being
2973 * fragmented and becoming unavailable for high-order allocations.
2979 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
2981 struct per_cpu_pages
*pcp
;
2983 memset(p
, 0, sizeof(*p
));
2987 pcp
->high
= 6 * batch
;
2988 pcp
->batch
= max(1UL, 1 * batch
);
2989 INIT_LIST_HEAD(&pcp
->list
);
2993 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
2994 * to the value high for the pageset p.
2997 static void setup_pagelist_highmark(struct per_cpu_pageset
*p
,
3000 struct per_cpu_pages
*pcp
;
3004 pcp
->batch
= max(1UL, high
/4);
3005 if ((high
/4) > (PAGE_SHIFT
* 8))
3006 pcp
->batch
= PAGE_SHIFT
* 8;
3012 * Boot pageset table. One per cpu which is going to be used for all
3013 * zones and all nodes. The parameters will be set in such a way
3014 * that an item put on a list will immediately be handed over to
3015 * the buddy list. This is safe since pageset manipulation is done
3016 * with interrupts disabled.
3018 * Some NUMA counter updates may also be caught by the boot pagesets.
3020 * The boot_pagesets must be kept even after bootup is complete for
3021 * unused processors and/or zones. They do play a role for bootstrapping
3022 * hotplugged processors.
3024 * zoneinfo_show() and maybe other functions do
3025 * not check if the processor is online before following the pageset pointer.
3026 * Other parts of the kernel may not check if the zone is available.
3028 static struct per_cpu_pageset boot_pageset
[NR_CPUS
];
3031 * Dynamically allocate memory for the
3032 * per cpu pageset array in struct zone.
3034 static int __cpuinit
process_zones(int cpu
)
3036 struct zone
*zone
, *dzone
;
3037 int node
= cpu_to_node(cpu
);
3039 node_set_state(node
, N_CPU
); /* this node has a cpu */
3041 for_each_populated_zone(zone
) {
3042 zone_pcp(zone
, cpu
) = kmalloc_node(sizeof(struct per_cpu_pageset
),
3044 if (!zone_pcp(zone
, cpu
))
3047 setup_pageset(zone_pcp(zone
, cpu
), zone_batchsize(zone
));
3049 if (percpu_pagelist_fraction
)
3050 setup_pagelist_highmark(zone_pcp(zone
, cpu
),
3051 (zone
->present_pages
/ percpu_pagelist_fraction
));
3056 for_each_zone(dzone
) {
3057 if (!populated_zone(dzone
))
3061 kfree(zone_pcp(dzone
, cpu
));
3062 zone_pcp(dzone
, cpu
) = &boot_pageset
[cpu
];
3067 static inline void free_zone_pagesets(int cpu
)
3071 for_each_zone(zone
) {
3072 struct per_cpu_pageset
*pset
= zone_pcp(zone
, cpu
);
3074 /* Free per_cpu_pageset if it is slab allocated */
3075 if (pset
!= &boot_pageset
[cpu
])
3077 zone_pcp(zone
, cpu
) = &boot_pageset
[cpu
];
3081 static int __cpuinit
pageset_cpuup_callback(struct notifier_block
*nfb
,
3082 unsigned long action
,
3085 int cpu
= (long)hcpu
;
3086 int ret
= NOTIFY_OK
;
3089 case CPU_UP_PREPARE
:
3090 case CPU_UP_PREPARE_FROZEN
:
3091 if (process_zones(cpu
))
3094 case CPU_UP_CANCELED
:
3095 case CPU_UP_CANCELED_FROZEN
:
3097 case CPU_DEAD_FROZEN
:
3098 free_zone_pagesets(cpu
);
3106 static struct notifier_block __cpuinitdata pageset_notifier
=
3107 { &pageset_cpuup_callback
, NULL
, 0 };
3109 void __init
setup_per_cpu_pageset(void)
3113 /* Initialize per_cpu_pageset for cpu 0.
3114 * A cpuup callback will do this for every cpu
3115 * as it comes online
3117 err
= process_zones(smp_processor_id());
3119 register_cpu_notifier(&pageset_notifier
);
3124 static noinline __init_refok
3125 int zone_wait_table_init(struct zone
*zone
, unsigned long zone_size_pages
)
3128 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
3132 * The per-page waitqueue mechanism uses hashed waitqueues
3135 zone
->wait_table_hash_nr_entries
=
3136 wait_table_hash_nr_entries(zone_size_pages
);
3137 zone
->wait_table_bits
=
3138 wait_table_bits(zone
->wait_table_hash_nr_entries
);
3139 alloc_size
= zone
->wait_table_hash_nr_entries
3140 * sizeof(wait_queue_head_t
);
3142 if (!slab_is_available()) {
3143 zone
->wait_table
= (wait_queue_head_t
*)
3144 alloc_bootmem_node(pgdat
, alloc_size
);
3147 * This case means that a zone whose size was 0 gets new memory
3148 * via memory hot-add.
3149 * But it may be the case that a new node was hot-added. In
3150 * this case vmalloc() will not be able to use this new node's
3151 * memory - this wait_table must be initialized to use this new
3152 * node itself as well.
3153 * To use this new node's memory, further consideration will be
3156 zone
->wait_table
= vmalloc(alloc_size
);
3158 if (!zone
->wait_table
)
3161 for(i
= 0; i
< zone
->wait_table_hash_nr_entries
; ++i
)
3162 init_waitqueue_head(zone
->wait_table
+ i
);
3167 static __meminit
void zone_pcp_init(struct zone
*zone
)
3170 unsigned long batch
= zone_batchsize(zone
);
3172 for (cpu
= 0; cpu
< NR_CPUS
; cpu
++) {
3174 /* Early boot. Slab allocator not functional yet */
3175 zone_pcp(zone
, cpu
) = &boot_pageset
[cpu
];
3176 setup_pageset(&boot_pageset
[cpu
],0);
3178 setup_pageset(zone_pcp(zone
,cpu
), batch
);
3181 if (zone
->present_pages
)
3182 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%lu\n",
3183 zone
->name
, zone
->present_pages
, batch
);
3186 __meminit
int init_currently_empty_zone(struct zone
*zone
,
3187 unsigned long zone_start_pfn
,
3189 enum memmap_context context
)
3191 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
3193 ret
= zone_wait_table_init(zone
, size
);
3196 pgdat
->nr_zones
= zone_idx(zone
) + 1;
3198 zone
->zone_start_pfn
= zone_start_pfn
;
3200 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
3201 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
3203 (unsigned long)zone_idx(zone
),
3204 zone_start_pfn
, (zone_start_pfn
+ size
));
3206 zone_init_free_lists(zone
);
3211 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3213 * Basic iterator support. Return the first range of PFNs for a node
3214 * Note: nid == MAX_NUMNODES returns first region regardless of node
3216 static int __meminit
first_active_region_index_in_nid(int nid
)
3220 for (i
= 0; i
< nr_nodemap_entries
; i
++)
3221 if (nid
== MAX_NUMNODES
|| early_node_map
[i
].nid
== nid
)
3228 * Basic iterator support. Return the next active range of PFNs for a node
3229 * Note: nid == MAX_NUMNODES returns next region regardless of node
3231 static int __meminit
next_active_region_index_in_nid(int index
, int nid
)
3233 for (index
= index
+ 1; index
< nr_nodemap_entries
; index
++)
3234 if (nid
== MAX_NUMNODES
|| early_node_map
[index
].nid
== nid
)
3240 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
3242 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
3243 * Architectures may implement their own version but if add_active_range()
3244 * was used and there are no special requirements, this is a convenient
3247 int __meminit
__early_pfn_to_nid(unsigned long pfn
)
3251 for (i
= 0; i
< nr_nodemap_entries
; i
++) {
3252 unsigned long start_pfn
= early_node_map
[i
].start_pfn
;
3253 unsigned long end_pfn
= early_node_map
[i
].end_pfn
;
3255 if (start_pfn
<= pfn
&& pfn
< end_pfn
)
3256 return early_node_map
[i
].nid
;
3258 /* This is a memory hole */
3261 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
3263 int __meminit
early_pfn_to_nid(unsigned long pfn
)
3267 nid
= __early_pfn_to_nid(pfn
);
3270 /* just returns 0 */
3274 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
3275 bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
3279 nid
= __early_pfn_to_nid(pfn
);
3280 if (nid
>= 0 && nid
!= node
)
3286 /* Basic iterator support to walk early_node_map[] */
3287 #define for_each_active_range_index_in_nid(i, nid) \
3288 for (i = first_active_region_index_in_nid(nid); i != -1; \
3289 i = next_active_region_index_in_nid(i, nid))
3292 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
3293 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
3294 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
3296 * If an architecture guarantees that all ranges registered with
3297 * add_active_ranges() contain no holes and may be freed, this
3298 * this function may be used instead of calling free_bootmem() manually.
3300 void __init
free_bootmem_with_active_regions(int nid
,
3301 unsigned long max_low_pfn
)
3305 for_each_active_range_index_in_nid(i
, nid
) {
3306 unsigned long size_pages
= 0;
3307 unsigned long end_pfn
= early_node_map
[i
].end_pfn
;
3309 if (early_node_map
[i
].start_pfn
>= max_low_pfn
)
3312 if (end_pfn
> max_low_pfn
)
3313 end_pfn
= max_low_pfn
;
3315 size_pages
= end_pfn
- early_node_map
[i
].start_pfn
;
3316 free_bootmem_node(NODE_DATA(early_node_map
[i
].nid
),
3317 PFN_PHYS(early_node_map
[i
].start_pfn
),
3318 size_pages
<< PAGE_SHIFT
);
3322 void __init
work_with_active_regions(int nid
, work_fn_t work_fn
, void *data
)
3327 for_each_active_range_index_in_nid(i
, nid
) {
3328 ret
= work_fn(early_node_map
[i
].start_pfn
,
3329 early_node_map
[i
].end_pfn
, data
);
3335 * sparse_memory_present_with_active_regions - Call memory_present for each active range
3336 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
3338 * If an architecture guarantees that all ranges registered with
3339 * add_active_ranges() contain no holes and may be freed, this
3340 * function may be used instead of calling memory_present() manually.
3342 void __init
sparse_memory_present_with_active_regions(int nid
)
3346 for_each_active_range_index_in_nid(i
, nid
)
3347 memory_present(early_node_map
[i
].nid
,
3348 early_node_map
[i
].start_pfn
,
3349 early_node_map
[i
].end_pfn
);
3353 * get_pfn_range_for_nid - Return the start and end page frames for a node
3354 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
3355 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
3356 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
3358 * It returns the start and end page frame of a node based on information
3359 * provided by an arch calling add_active_range(). If called for a node
3360 * with no available memory, a warning is printed and the start and end
3363 void __meminit
get_pfn_range_for_nid(unsigned int nid
,
3364 unsigned long *start_pfn
, unsigned long *end_pfn
)
3370 for_each_active_range_index_in_nid(i
, nid
) {
3371 *start_pfn
= min(*start_pfn
, early_node_map
[i
].start_pfn
);
3372 *end_pfn
= max(*end_pfn
, early_node_map
[i
].end_pfn
);
3375 if (*start_pfn
== -1UL)
3380 * This finds a zone that can be used for ZONE_MOVABLE pages. The
3381 * assumption is made that zones within a node are ordered in monotonic
3382 * increasing memory addresses so that the "highest" populated zone is used
3384 static void __init
find_usable_zone_for_movable(void)
3387 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
3388 if (zone_index
== ZONE_MOVABLE
)
3391 if (arch_zone_highest_possible_pfn
[zone_index
] >
3392 arch_zone_lowest_possible_pfn
[zone_index
])
3396 VM_BUG_ON(zone_index
== -1);
3397 movable_zone
= zone_index
;
3401 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
3402 * because it is sized independant of architecture. Unlike the other zones,
3403 * the starting point for ZONE_MOVABLE is not fixed. It may be different
3404 * in each node depending on the size of each node and how evenly kernelcore
3405 * is distributed. This helper function adjusts the zone ranges
3406 * provided by the architecture for a given node by using the end of the
3407 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
3408 * zones within a node are in order of monotonic increases memory addresses
3410 static void __meminit
adjust_zone_range_for_zone_movable(int nid
,
3411 unsigned long zone_type
,
3412 unsigned long node_start_pfn
,
3413 unsigned long node_end_pfn
,
3414 unsigned long *zone_start_pfn
,
3415 unsigned long *zone_end_pfn
)
3417 /* Only adjust if ZONE_MOVABLE is on this node */
3418 if (zone_movable_pfn
[nid
]) {
3419 /* Size ZONE_MOVABLE */
3420 if (zone_type
== ZONE_MOVABLE
) {
3421 *zone_start_pfn
= zone_movable_pfn
[nid
];
3422 *zone_end_pfn
= min(node_end_pfn
,
3423 arch_zone_highest_possible_pfn
[movable_zone
]);
3425 /* Adjust for ZONE_MOVABLE starting within this range */
3426 } else if (*zone_start_pfn
< zone_movable_pfn
[nid
] &&
3427 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
3428 *zone_end_pfn
= zone_movable_pfn
[nid
];
3430 /* Check if this whole range is within ZONE_MOVABLE */
3431 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
3432 *zone_start_pfn
= *zone_end_pfn
;
3437 * Return the number of pages a zone spans in a node, including holes
3438 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
3440 static unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
3441 unsigned long zone_type
,
3442 unsigned long *ignored
)
3444 unsigned long node_start_pfn
, node_end_pfn
;
3445 unsigned long zone_start_pfn
, zone_end_pfn
;
3447 /* Get the start and end of the node and zone */
3448 get_pfn_range_for_nid(nid
, &node_start_pfn
, &node_end_pfn
);
3449 zone_start_pfn
= arch_zone_lowest_possible_pfn
[zone_type
];
3450 zone_end_pfn
= arch_zone_highest_possible_pfn
[zone_type
];
3451 adjust_zone_range_for_zone_movable(nid
, zone_type
,
3452 node_start_pfn
, node_end_pfn
,
3453 &zone_start_pfn
, &zone_end_pfn
);
3455 /* Check that this node has pages within the zone's required range */
3456 if (zone_end_pfn
< node_start_pfn
|| zone_start_pfn
> node_end_pfn
)
3459 /* Move the zone boundaries inside the node if necessary */
3460 zone_end_pfn
= min(zone_end_pfn
, node_end_pfn
);
3461 zone_start_pfn
= max(zone_start_pfn
, node_start_pfn
);
3463 /* Return the spanned pages */
3464 return zone_end_pfn
- zone_start_pfn
;
3468 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
3469 * then all holes in the requested range will be accounted for.
3471 static unsigned long __meminit
__absent_pages_in_range(int nid
,
3472 unsigned long range_start_pfn
,
3473 unsigned long range_end_pfn
)
3476 unsigned long prev_end_pfn
= 0, hole_pages
= 0;
3477 unsigned long start_pfn
;
3479 /* Find the end_pfn of the first active range of pfns in the node */
3480 i
= first_active_region_index_in_nid(nid
);
3484 prev_end_pfn
= min(early_node_map
[i
].start_pfn
, range_end_pfn
);
3486 /* Account for ranges before physical memory on this node */
3487 if (early_node_map
[i
].start_pfn
> range_start_pfn
)
3488 hole_pages
= prev_end_pfn
- range_start_pfn
;
3490 /* Find all holes for the zone within the node */
3491 for (; i
!= -1; i
= next_active_region_index_in_nid(i
, nid
)) {
3493 /* No need to continue if prev_end_pfn is outside the zone */
3494 if (prev_end_pfn
>= range_end_pfn
)
3497 /* Make sure the end of the zone is not within the hole */
3498 start_pfn
= min(early_node_map
[i
].start_pfn
, range_end_pfn
);
3499 prev_end_pfn
= max(prev_end_pfn
, range_start_pfn
);
3501 /* Update the hole size cound and move on */
3502 if (start_pfn
> range_start_pfn
) {
3503 BUG_ON(prev_end_pfn
> start_pfn
);
3504 hole_pages
+= start_pfn
- prev_end_pfn
;
3506 prev_end_pfn
= early_node_map
[i
].end_pfn
;
3509 /* Account for ranges past physical memory on this node */
3510 if (range_end_pfn
> prev_end_pfn
)
3511 hole_pages
+= range_end_pfn
-
3512 max(range_start_pfn
, prev_end_pfn
);
3518 * absent_pages_in_range - Return number of page frames in holes within a range
3519 * @start_pfn: The start PFN to start searching for holes
3520 * @end_pfn: The end PFN to stop searching for holes
3522 * It returns the number of pages frames in memory holes within a range.
3524 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
3525 unsigned long end_pfn
)
3527 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
3530 /* Return the number of page frames in holes in a zone on a node */
3531 static unsigned long __meminit
zone_absent_pages_in_node(int nid
,
3532 unsigned long zone_type
,
3533 unsigned long *ignored
)
3535 unsigned long node_start_pfn
, node_end_pfn
;
3536 unsigned long zone_start_pfn
, zone_end_pfn
;
3538 get_pfn_range_for_nid(nid
, &node_start_pfn
, &node_end_pfn
);
3539 zone_start_pfn
= max(arch_zone_lowest_possible_pfn
[zone_type
],
3541 zone_end_pfn
= min(arch_zone_highest_possible_pfn
[zone_type
],
3544 adjust_zone_range_for_zone_movable(nid
, zone_type
,
3545 node_start_pfn
, node_end_pfn
,
3546 &zone_start_pfn
, &zone_end_pfn
);
3547 return __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
3551 static inline unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
3552 unsigned long zone_type
,
3553 unsigned long *zones_size
)
3555 return zones_size
[zone_type
];
3558 static inline unsigned long __meminit
zone_absent_pages_in_node(int nid
,
3559 unsigned long zone_type
,
3560 unsigned long *zholes_size
)
3565 return zholes_size
[zone_type
];
3570 static void __meminit
calculate_node_totalpages(struct pglist_data
*pgdat
,
3571 unsigned long *zones_size
, unsigned long *zholes_size
)
3573 unsigned long realtotalpages
, totalpages
= 0;
3576 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
3577 totalpages
+= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
3579 pgdat
->node_spanned_pages
= totalpages
;
3581 realtotalpages
= totalpages
;
3582 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
3584 zone_absent_pages_in_node(pgdat
->node_id
, i
,
3586 pgdat
->node_present_pages
= realtotalpages
;
3587 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
3591 #ifndef CONFIG_SPARSEMEM
3593 * Calculate the size of the zone->blockflags rounded to an unsigned long
3594 * Start by making sure zonesize is a multiple of pageblock_order by rounding
3595 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
3596 * round what is now in bits to nearest long in bits, then return it in
3599 static unsigned long __init
usemap_size(unsigned long zonesize
)
3601 unsigned long usemapsize
;
3603 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
3604 usemapsize
= usemapsize
>> pageblock_order
;
3605 usemapsize
*= NR_PAGEBLOCK_BITS
;
3606 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
3608 return usemapsize
/ 8;
3611 static void __init
setup_usemap(struct pglist_data
*pgdat
,
3612 struct zone
*zone
, unsigned long zonesize
)
3614 unsigned long usemapsize
= usemap_size(zonesize
);
3615 zone
->pageblock_flags
= NULL
;
3617 zone
->pageblock_flags
= alloc_bootmem_node(pgdat
, usemapsize
);
3620 static void inline setup_usemap(struct pglist_data
*pgdat
,
3621 struct zone
*zone
, unsigned long zonesize
) {}
3622 #endif /* CONFIG_SPARSEMEM */
3624 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
3626 /* Return a sensible default order for the pageblock size. */
3627 static inline int pageblock_default_order(void)
3629 if (HPAGE_SHIFT
> PAGE_SHIFT
)
3630 return HUGETLB_PAGE_ORDER
;
3635 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
3636 static inline void __init
set_pageblock_order(unsigned int order
)
3638 /* Check that pageblock_nr_pages has not already been setup */
3639 if (pageblock_order
)
3643 * Assume the largest contiguous order of interest is a huge page.
3644 * This value may be variable depending on boot parameters on IA64
3646 pageblock_order
= order
;
3648 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
3651 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
3652 * and pageblock_default_order() are unused as pageblock_order is set
3653 * at compile-time. See include/linux/pageblock-flags.h for the values of
3654 * pageblock_order based on the kernel config
3656 static inline int pageblock_default_order(unsigned int order
)
3660 #define set_pageblock_order(x) do {} while (0)
3662 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
3665 * Set up the zone data structures:
3666 * - mark all pages reserved
3667 * - mark all memory queues empty
3668 * - clear the memory bitmaps
3670 static void __paginginit
free_area_init_core(struct pglist_data
*pgdat
,
3671 unsigned long *zones_size
, unsigned long *zholes_size
)
3674 int nid
= pgdat
->node_id
;
3675 unsigned long zone_start_pfn
= pgdat
->node_start_pfn
;
3678 pgdat_resize_init(pgdat
);
3679 pgdat
->nr_zones
= 0;
3680 init_waitqueue_head(&pgdat
->kswapd_wait
);
3681 pgdat
->kswapd_max_order
= 0;
3682 pgdat_page_cgroup_init(pgdat
);
3684 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
3685 struct zone
*zone
= pgdat
->node_zones
+ j
;
3686 unsigned long size
, realsize
, memmap_pages
;
3689 size
= zone_spanned_pages_in_node(nid
, j
, zones_size
);
3690 realsize
= size
- zone_absent_pages_in_node(nid
, j
,
3694 * Adjust realsize so that it accounts for how much memory
3695 * is used by this zone for memmap. This affects the watermark
3696 * and per-cpu initialisations
3699 PAGE_ALIGN(size
* sizeof(struct page
)) >> PAGE_SHIFT
;
3700 if (realsize
>= memmap_pages
) {
3701 realsize
-= memmap_pages
;
3704 " %s zone: %lu pages used for memmap\n",
3705 zone_names
[j
], memmap_pages
);
3708 " %s zone: %lu pages exceeds realsize %lu\n",
3709 zone_names
[j
], memmap_pages
, realsize
);
3711 /* Account for reserved pages */
3712 if (j
== 0 && realsize
> dma_reserve
) {
3713 realsize
-= dma_reserve
;
3714 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
3715 zone_names
[0], dma_reserve
);
3718 if (!is_highmem_idx(j
))
3719 nr_kernel_pages
+= realsize
;
3720 nr_all_pages
+= realsize
;
3722 zone
->spanned_pages
= size
;
3723 zone
->present_pages
= realsize
;
3726 zone
->min_unmapped_pages
= (realsize
*sysctl_min_unmapped_ratio
)
3728 zone
->min_slab_pages
= (realsize
* sysctl_min_slab_ratio
) / 100;
3730 zone
->name
= zone_names
[j
];
3731 spin_lock_init(&zone
->lock
);
3732 spin_lock_init(&zone
->lru_lock
);
3733 zone_seqlock_init(zone
);
3734 zone
->zone_pgdat
= pgdat
;
3736 zone
->prev_priority
= DEF_PRIORITY
;
3738 zone_pcp_init(zone
);
3740 INIT_LIST_HEAD(&zone
->lru
[l
].list
);
3741 zone
->lru
[l
].nr_saved_scan
= 0;
3743 zone
->reclaim_stat
.recent_rotated
[0] = 0;
3744 zone
->reclaim_stat
.recent_rotated
[1] = 0;
3745 zone
->reclaim_stat
.recent_scanned
[0] = 0;
3746 zone
->reclaim_stat
.recent_scanned
[1] = 0;
3747 zap_zone_vm_stats(zone
);
3752 set_pageblock_order(pageblock_default_order());
3753 setup_usemap(pgdat
, zone
, size
);
3754 ret
= init_currently_empty_zone(zone
, zone_start_pfn
,
3755 size
, MEMMAP_EARLY
);
3757 memmap_init(size
, nid
, j
, zone_start_pfn
);
3758 zone_start_pfn
+= size
;
3762 static void __init_refok
alloc_node_mem_map(struct pglist_data
*pgdat
)
3764 /* Skip empty nodes */
3765 if (!pgdat
->node_spanned_pages
)
3768 #ifdef CONFIG_FLAT_NODE_MEM_MAP
3769 /* ia64 gets its own node_mem_map, before this, without bootmem */
3770 if (!pgdat
->node_mem_map
) {
3771 unsigned long size
, start
, end
;
3775 * The zone's endpoints aren't required to be MAX_ORDER
3776 * aligned but the node_mem_map endpoints must be in order
3777 * for the buddy allocator to function correctly.
3779 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
3780 end
= pgdat
->node_start_pfn
+ pgdat
->node_spanned_pages
;
3781 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
3782 size
= (end
- start
) * sizeof(struct page
);
3783 map
= alloc_remap(pgdat
->node_id
, size
);
3785 map
= alloc_bootmem_node(pgdat
, size
);
3786 pgdat
->node_mem_map
= map
+ (pgdat
->node_start_pfn
- start
);
3788 #ifndef CONFIG_NEED_MULTIPLE_NODES
3790 * With no DISCONTIG, the global mem_map is just set as node 0's
3792 if (pgdat
== NODE_DATA(0)) {
3793 mem_map
= NODE_DATA(0)->node_mem_map
;
3794 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3795 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
3796 mem_map
-= (pgdat
->node_start_pfn
- ARCH_PFN_OFFSET
);
3797 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
3800 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
3803 void __paginginit
free_area_init_node(int nid
, unsigned long *zones_size
,
3804 unsigned long node_start_pfn
, unsigned long *zholes_size
)
3806 pg_data_t
*pgdat
= NODE_DATA(nid
);
3808 pgdat
->node_id
= nid
;
3809 pgdat
->node_start_pfn
= node_start_pfn
;
3810 calculate_node_totalpages(pgdat
, zones_size
, zholes_size
);
3812 alloc_node_mem_map(pgdat
);
3813 #ifdef CONFIG_FLAT_NODE_MEM_MAP
3814 printk(KERN_DEBUG
"free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
3815 nid
, (unsigned long)pgdat
,
3816 (unsigned long)pgdat
->node_mem_map
);
3819 free_area_init_core(pgdat
, zones_size
, zholes_size
);
3822 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3824 #if MAX_NUMNODES > 1
3826 * Figure out the number of possible node ids.
3828 static void __init
setup_nr_node_ids(void)
3831 unsigned int highest
= 0;
3833 for_each_node_mask(node
, node_possible_map
)
3835 nr_node_ids
= highest
+ 1;
3838 static inline void setup_nr_node_ids(void)
3844 * add_active_range - Register a range of PFNs backed by physical memory
3845 * @nid: The node ID the range resides on
3846 * @start_pfn: The start PFN of the available physical memory
3847 * @end_pfn: The end PFN of the available physical memory
3849 * These ranges are stored in an early_node_map[] and later used by
3850 * free_area_init_nodes() to calculate zone sizes and holes. If the
3851 * range spans a memory hole, it is up to the architecture to ensure
3852 * the memory is not freed by the bootmem allocator. If possible
3853 * the range being registered will be merged with existing ranges.
3855 void __init
add_active_range(unsigned int nid
, unsigned long start_pfn
,
3856 unsigned long end_pfn
)
3860 mminit_dprintk(MMINIT_TRACE
, "memory_register",
3861 "Entering add_active_range(%d, %#lx, %#lx) "
3862 "%d entries of %d used\n",
3863 nid
, start_pfn
, end_pfn
,
3864 nr_nodemap_entries
, MAX_ACTIVE_REGIONS
);
3866 mminit_validate_memmodel_limits(&start_pfn
, &end_pfn
);
3868 /* Merge with existing active regions if possible */
3869 for (i
= 0; i
< nr_nodemap_entries
; i
++) {
3870 if (early_node_map
[i
].nid
!= nid
)
3873 /* Skip if an existing region covers this new one */
3874 if (start_pfn
>= early_node_map
[i
].start_pfn
&&
3875 end_pfn
<= early_node_map
[i
].end_pfn
)
3878 /* Merge forward if suitable */
3879 if (start_pfn
<= early_node_map
[i
].end_pfn
&&
3880 end_pfn
> early_node_map
[i
].end_pfn
) {
3881 early_node_map
[i
].end_pfn
= end_pfn
;
3885 /* Merge backward if suitable */
3886 if (start_pfn
< early_node_map
[i
].end_pfn
&&
3887 end_pfn
>= early_node_map
[i
].start_pfn
) {
3888 early_node_map
[i
].start_pfn
= start_pfn
;
3893 /* Check that early_node_map is large enough */
3894 if (i
>= MAX_ACTIVE_REGIONS
) {
3895 printk(KERN_CRIT
"More than %d memory regions, truncating\n",
3896 MAX_ACTIVE_REGIONS
);
3900 early_node_map
[i
].nid
= nid
;
3901 early_node_map
[i
].start_pfn
= start_pfn
;
3902 early_node_map
[i
].end_pfn
= end_pfn
;
3903 nr_nodemap_entries
= i
+ 1;
3907 * remove_active_range - Shrink an existing registered range of PFNs
3908 * @nid: The node id the range is on that should be shrunk
3909 * @start_pfn: The new PFN of the range
3910 * @end_pfn: The new PFN of the range
3912 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
3913 * The map is kept near the end physical page range that has already been
3914 * registered. This function allows an arch to shrink an existing registered
3917 void __init
remove_active_range(unsigned int nid
, unsigned long start_pfn
,
3918 unsigned long end_pfn
)
3923 printk(KERN_DEBUG
"remove_active_range (%d, %lu, %lu)\n",
3924 nid
, start_pfn
, end_pfn
);
3926 /* Find the old active region end and shrink */
3927 for_each_active_range_index_in_nid(i
, nid
) {
3928 if (early_node_map
[i
].start_pfn
>= start_pfn
&&
3929 early_node_map
[i
].end_pfn
<= end_pfn
) {
3931 early_node_map
[i
].start_pfn
= 0;
3932 early_node_map
[i
].end_pfn
= 0;
3936 if (early_node_map
[i
].start_pfn
< start_pfn
&&
3937 early_node_map
[i
].end_pfn
> start_pfn
) {
3938 unsigned long temp_end_pfn
= early_node_map
[i
].end_pfn
;
3939 early_node_map
[i
].end_pfn
= start_pfn
;
3940 if (temp_end_pfn
> end_pfn
)
3941 add_active_range(nid
, end_pfn
, temp_end_pfn
);
3944 if (early_node_map
[i
].start_pfn
>= start_pfn
&&
3945 early_node_map
[i
].end_pfn
> end_pfn
&&
3946 early_node_map
[i
].start_pfn
< end_pfn
) {
3947 early_node_map
[i
].start_pfn
= end_pfn
;
3955 /* remove the blank ones */
3956 for (i
= nr_nodemap_entries
- 1; i
> 0; i
--) {
3957 if (early_node_map
[i
].nid
!= nid
)
3959 if (early_node_map
[i
].end_pfn
)
3961 /* we found it, get rid of it */
3962 for (j
= i
; j
< nr_nodemap_entries
- 1; j
++)
3963 memcpy(&early_node_map
[j
], &early_node_map
[j
+1],
3964 sizeof(early_node_map
[j
]));
3965 j
= nr_nodemap_entries
- 1;
3966 memset(&early_node_map
[j
], 0, sizeof(early_node_map
[j
]));
3967 nr_nodemap_entries
--;
3972 * remove_all_active_ranges - Remove all currently registered regions
3974 * During discovery, it may be found that a table like SRAT is invalid
3975 * and an alternative discovery method must be used. This function removes
3976 * all currently registered regions.
3978 void __init
remove_all_active_ranges(void)
3980 memset(early_node_map
, 0, sizeof(early_node_map
));
3981 nr_nodemap_entries
= 0;
3984 /* Compare two active node_active_regions */
3985 static int __init
cmp_node_active_region(const void *a
, const void *b
)
3987 struct node_active_region
*arange
= (struct node_active_region
*)a
;
3988 struct node_active_region
*brange
= (struct node_active_region
*)b
;
3990 /* Done this way to avoid overflows */
3991 if (arange
->start_pfn
> brange
->start_pfn
)
3993 if (arange
->start_pfn
< brange
->start_pfn
)
3999 /* sort the node_map by start_pfn */
4000 static void __init
sort_node_map(void)
4002 sort(early_node_map
, (size_t)nr_nodemap_entries
,
4003 sizeof(struct node_active_region
),
4004 cmp_node_active_region
, NULL
);
4007 /* Find the lowest pfn for a node */
4008 static unsigned long __init
find_min_pfn_for_node(int nid
)
4011 unsigned long min_pfn
= ULONG_MAX
;
4013 /* Assuming a sorted map, the first range found has the starting pfn */
4014 for_each_active_range_index_in_nid(i
, nid
)
4015 min_pfn
= min(min_pfn
, early_node_map
[i
].start_pfn
);
4017 if (min_pfn
== ULONG_MAX
) {
4019 "Could not find start_pfn for node %d\n", nid
);
4027 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4029 * It returns the minimum PFN based on information provided via
4030 * add_active_range().
4032 unsigned long __init
find_min_pfn_with_active_regions(void)
4034 return find_min_pfn_for_node(MAX_NUMNODES
);
4038 * early_calculate_totalpages()
4039 * Sum pages in active regions for movable zone.
4040 * Populate N_HIGH_MEMORY for calculating usable_nodes.
4042 static unsigned long __init
early_calculate_totalpages(void)
4045 unsigned long totalpages
= 0;
4047 for (i
= 0; i
< nr_nodemap_entries
; i
++) {
4048 unsigned long pages
= early_node_map
[i
].end_pfn
-
4049 early_node_map
[i
].start_pfn
;
4050 totalpages
+= pages
;
4052 node_set_state(early_node_map
[i
].nid
, N_HIGH_MEMORY
);
4058 * Find the PFN the Movable zone begins in each node. Kernel memory
4059 * is spread evenly between nodes as long as the nodes have enough
4060 * memory. When they don't, some nodes will have more kernelcore than
4063 static void __init
find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn
)
4066 unsigned long usable_startpfn
;
4067 unsigned long kernelcore_node
, kernelcore_remaining
;
4068 /* save the state before borrow the nodemask */
4069 nodemask_t saved_node_state
= node_states
[N_HIGH_MEMORY
];
4070 unsigned long totalpages
= early_calculate_totalpages();
4071 int usable_nodes
= nodes_weight(node_states
[N_HIGH_MEMORY
]);
4074 * If movablecore was specified, calculate what size of
4075 * kernelcore that corresponds so that memory usable for
4076 * any allocation type is evenly spread. If both kernelcore
4077 * and movablecore are specified, then the value of kernelcore
4078 * will be used for required_kernelcore if it's greater than
4079 * what movablecore would have allowed.
4081 if (required_movablecore
) {
4082 unsigned long corepages
;
4085 * Round-up so that ZONE_MOVABLE is at least as large as what
4086 * was requested by the user
4088 required_movablecore
=
4089 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
4090 corepages
= totalpages
- required_movablecore
;
4092 required_kernelcore
= max(required_kernelcore
, corepages
);
4095 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4096 if (!required_kernelcore
)
4099 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4100 find_usable_zone_for_movable();
4101 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
4104 /* Spread kernelcore memory as evenly as possible throughout nodes */
4105 kernelcore_node
= required_kernelcore
/ usable_nodes
;
4106 for_each_node_state(nid
, N_HIGH_MEMORY
) {
4108 * Recalculate kernelcore_node if the division per node
4109 * now exceeds what is necessary to satisfy the requested
4110 * amount of memory for the kernel
4112 if (required_kernelcore
< kernelcore_node
)
4113 kernelcore_node
= required_kernelcore
/ usable_nodes
;
4116 * As the map is walked, we track how much memory is usable
4117 * by the kernel using kernelcore_remaining. When it is
4118 * 0, the rest of the node is usable by ZONE_MOVABLE
4120 kernelcore_remaining
= kernelcore_node
;
4122 /* Go through each range of PFNs within this node */
4123 for_each_active_range_index_in_nid(i
, nid
) {
4124 unsigned long start_pfn
, end_pfn
;
4125 unsigned long size_pages
;
4127 start_pfn
= max(early_node_map
[i
].start_pfn
,
4128 zone_movable_pfn
[nid
]);
4129 end_pfn
= early_node_map
[i
].end_pfn
;
4130 if (start_pfn
>= end_pfn
)
4133 /* Account for what is only usable for kernelcore */
4134 if (start_pfn
< usable_startpfn
) {
4135 unsigned long kernel_pages
;
4136 kernel_pages
= min(end_pfn
, usable_startpfn
)
4139 kernelcore_remaining
-= min(kernel_pages
,
4140 kernelcore_remaining
);
4141 required_kernelcore
-= min(kernel_pages
,
4142 required_kernelcore
);
4144 /* Continue if range is now fully accounted */
4145 if (end_pfn
<= usable_startpfn
) {
4148 * Push zone_movable_pfn to the end so
4149 * that if we have to rebalance
4150 * kernelcore across nodes, we will
4151 * not double account here
4153 zone_movable_pfn
[nid
] = end_pfn
;
4156 start_pfn
= usable_startpfn
;
4160 * The usable PFN range for ZONE_MOVABLE is from
4161 * start_pfn->end_pfn. Calculate size_pages as the
4162 * number of pages used as kernelcore
4164 size_pages
= end_pfn
- start_pfn
;
4165 if (size_pages
> kernelcore_remaining
)
4166 size_pages
= kernelcore_remaining
;
4167 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
4170 * Some kernelcore has been met, update counts and
4171 * break if the kernelcore for this node has been
4174 required_kernelcore
-= min(required_kernelcore
,
4176 kernelcore_remaining
-= size_pages
;
4177 if (!kernelcore_remaining
)
4183 * If there is still required_kernelcore, we do another pass with one
4184 * less node in the count. This will push zone_movable_pfn[nid] further
4185 * along on the nodes that still have memory until kernelcore is
4189 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
4192 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4193 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
4194 zone_movable_pfn
[nid
] =
4195 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
4198 /* restore the node_state */
4199 node_states
[N_HIGH_MEMORY
] = saved_node_state
;
4202 /* Any regular memory on that node ? */
4203 static void check_for_regular_memory(pg_data_t
*pgdat
)
4205 #ifdef CONFIG_HIGHMEM
4206 enum zone_type zone_type
;
4208 for (zone_type
= 0; zone_type
<= ZONE_NORMAL
; zone_type
++) {
4209 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
4210 if (zone
->present_pages
)
4211 node_set_state(zone_to_nid(zone
), N_NORMAL_MEMORY
);
4217 * free_area_init_nodes - Initialise all pg_data_t and zone data
4218 * @max_zone_pfn: an array of max PFNs for each zone
4220 * This will call free_area_init_node() for each active node in the system.
4221 * Using the page ranges provided by add_active_range(), the size of each
4222 * zone in each node and their holes is calculated. If the maximum PFN
4223 * between two adjacent zones match, it is assumed that the zone is empty.
4224 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4225 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4226 * starts where the previous one ended. For example, ZONE_DMA32 starts
4227 * at arch_max_dma_pfn.
4229 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
4234 /* Sort early_node_map as initialisation assumes it is sorted */
4237 /* Record where the zone boundaries are */
4238 memset(arch_zone_lowest_possible_pfn
, 0,
4239 sizeof(arch_zone_lowest_possible_pfn
));
4240 memset(arch_zone_highest_possible_pfn
, 0,
4241 sizeof(arch_zone_highest_possible_pfn
));
4242 arch_zone_lowest_possible_pfn
[0] = find_min_pfn_with_active_regions();
4243 arch_zone_highest_possible_pfn
[0] = max_zone_pfn
[0];
4244 for (i
= 1; i
< MAX_NR_ZONES
; i
++) {
4245 if (i
== ZONE_MOVABLE
)
4247 arch_zone_lowest_possible_pfn
[i
] =
4248 arch_zone_highest_possible_pfn
[i
-1];
4249 arch_zone_highest_possible_pfn
[i
] =
4250 max(max_zone_pfn
[i
], arch_zone_lowest_possible_pfn
[i
]);
4252 arch_zone_lowest_possible_pfn
[ZONE_MOVABLE
] = 0;
4253 arch_zone_highest_possible_pfn
[ZONE_MOVABLE
] = 0;
4255 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4256 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
4257 find_zone_movable_pfns_for_nodes(zone_movable_pfn
);
4259 /* Print out the zone ranges */
4260 printk("Zone PFN ranges:\n");
4261 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
4262 if (i
== ZONE_MOVABLE
)
4264 printk(" %-8s %0#10lx -> %0#10lx\n",
4266 arch_zone_lowest_possible_pfn
[i
],
4267 arch_zone_highest_possible_pfn
[i
]);
4270 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4271 printk("Movable zone start PFN for each node\n");
4272 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
4273 if (zone_movable_pfn
[i
])
4274 printk(" Node %d: %lu\n", i
, zone_movable_pfn
[i
]);
4277 /* Print out the early_node_map[] */
4278 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries
);
4279 for (i
= 0; i
< nr_nodemap_entries
; i
++)
4280 printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map
[i
].nid
,
4281 early_node_map
[i
].start_pfn
,
4282 early_node_map
[i
].end_pfn
);
4284 /* Initialise every node */
4285 mminit_verify_pageflags_layout();
4286 setup_nr_node_ids();
4287 for_each_online_node(nid
) {
4288 pg_data_t
*pgdat
= NODE_DATA(nid
);
4289 free_area_init_node(nid
, NULL
,
4290 find_min_pfn_for_node(nid
), NULL
);
4292 /* Any memory on that node */
4293 if (pgdat
->node_present_pages
)
4294 node_set_state(nid
, N_HIGH_MEMORY
);
4295 check_for_regular_memory(pgdat
);
4299 static int __init
cmdline_parse_core(char *p
, unsigned long *core
)
4301 unsigned long long coremem
;
4305 coremem
= memparse(p
, &p
);
4306 *core
= coremem
>> PAGE_SHIFT
;
4308 /* Paranoid check that UL is enough for the coremem value */
4309 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
4315 * kernelcore=size sets the amount of memory for use for allocations that
4316 * cannot be reclaimed or migrated.
4318 static int __init
cmdline_parse_kernelcore(char *p
)
4320 return cmdline_parse_core(p
, &required_kernelcore
);
4324 * movablecore=size sets the amount of memory for use for allocations that
4325 * can be reclaimed or migrated.
4327 static int __init
cmdline_parse_movablecore(char *p
)
4329 return cmdline_parse_core(p
, &required_movablecore
);
4332 early_param("kernelcore", cmdline_parse_kernelcore
);
4333 early_param("movablecore", cmdline_parse_movablecore
);
4335 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4338 * set_dma_reserve - set the specified number of pages reserved in the first zone
4339 * @new_dma_reserve: The number of pages to mark reserved
4341 * The per-cpu batchsize and zone watermarks are determined by present_pages.
4342 * In the DMA zone, a significant percentage may be consumed by kernel image
4343 * and other unfreeable allocations which can skew the watermarks badly. This
4344 * function may optionally be used to account for unfreeable pages in the
4345 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
4346 * smaller per-cpu batchsize.
4348 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
4350 dma_reserve
= new_dma_reserve
;
4353 #ifndef CONFIG_NEED_MULTIPLE_NODES
4354 struct pglist_data __refdata contig_page_data
= { .bdata
= &bootmem_node_data
[0] };
4355 EXPORT_SYMBOL(contig_page_data
);
4358 void __init
free_area_init(unsigned long *zones_size
)
4360 free_area_init_node(0, zones_size
,
4361 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
4364 static int page_alloc_cpu_notify(struct notifier_block
*self
,
4365 unsigned long action
, void *hcpu
)
4367 int cpu
= (unsigned long)hcpu
;
4369 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
) {
4373 * Spill the event counters of the dead processor
4374 * into the current processors event counters.
4375 * This artificially elevates the count of the current
4378 vm_events_fold_cpu(cpu
);
4381 * Zero the differential counters of the dead processor
4382 * so that the vm statistics are consistent.
4384 * This is only okay since the processor is dead and cannot
4385 * race with what we are doing.
4387 refresh_cpu_vm_stats(cpu
);
4392 void __init
page_alloc_init(void)
4394 hotcpu_notifier(page_alloc_cpu_notify
, 0);
4398 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
4399 * or min_free_kbytes changes.
4401 static void calculate_totalreserve_pages(void)
4403 struct pglist_data
*pgdat
;
4404 unsigned long reserve_pages
= 0;
4405 enum zone_type i
, j
;
4407 for_each_online_pgdat(pgdat
) {
4408 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
4409 struct zone
*zone
= pgdat
->node_zones
+ i
;
4410 unsigned long max
= 0;
4412 /* Find valid and maximum lowmem_reserve in the zone */
4413 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
4414 if (zone
->lowmem_reserve
[j
] > max
)
4415 max
= zone
->lowmem_reserve
[j
];
4418 /* we treat the high watermark as reserved pages. */
4419 max
+= high_wmark_pages(zone
);
4421 if (max
> zone
->present_pages
)
4422 max
= zone
->present_pages
;
4423 reserve_pages
+= max
;
4426 totalreserve_pages
= reserve_pages
;
4430 * setup_per_zone_lowmem_reserve - called whenever
4431 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
4432 * has a correct pages reserved value, so an adequate number of
4433 * pages are left in the zone after a successful __alloc_pages().
4435 static void setup_per_zone_lowmem_reserve(void)
4437 struct pglist_data
*pgdat
;
4438 enum zone_type j
, idx
;
4440 for_each_online_pgdat(pgdat
) {
4441 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
4442 struct zone
*zone
= pgdat
->node_zones
+ j
;
4443 unsigned long present_pages
= zone
->present_pages
;
4445 zone
->lowmem_reserve
[j
] = 0;
4449 struct zone
*lower_zone
;
4453 if (sysctl_lowmem_reserve_ratio
[idx
] < 1)
4454 sysctl_lowmem_reserve_ratio
[idx
] = 1;
4456 lower_zone
= pgdat
->node_zones
+ idx
;
4457 lower_zone
->lowmem_reserve
[j
] = present_pages
/
4458 sysctl_lowmem_reserve_ratio
[idx
];
4459 present_pages
+= lower_zone
->present_pages
;
4464 /* update totalreserve_pages */
4465 calculate_totalreserve_pages();
4469 * setup_per_zone_wmarks - called when min_free_kbytes changes
4470 * or when memory is hot-{added|removed}
4472 * Ensures that the watermark[min,low,high] values for each zone are set
4473 * correctly with respect to min_free_kbytes.
4475 void setup_per_zone_wmarks(void)
4477 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
4478 unsigned long lowmem_pages
= 0;
4480 unsigned long flags
;
4482 /* Calculate total number of !ZONE_HIGHMEM pages */
4483 for_each_zone(zone
) {
4484 if (!is_highmem(zone
))
4485 lowmem_pages
+= zone
->present_pages
;
4488 for_each_zone(zone
) {
4491 spin_lock_irqsave(&zone
->lock
, flags
);
4492 tmp
= (u64
)pages_min
* zone
->present_pages
;
4493 do_div(tmp
, lowmem_pages
);
4494 if (is_highmem(zone
)) {
4496 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
4497 * need highmem pages, so cap pages_min to a small
4500 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
4501 * deltas controls asynch page reclaim, and so should
4502 * not be capped for highmem.
4506 min_pages
= zone
->present_pages
/ 1024;
4507 if (min_pages
< SWAP_CLUSTER_MAX
)
4508 min_pages
= SWAP_CLUSTER_MAX
;
4509 if (min_pages
> 128)
4511 zone
->watermark
[WMARK_MIN
] = min_pages
;
4514 * If it's a lowmem zone, reserve a number of pages
4515 * proportionate to the zone's size.
4517 zone
->watermark
[WMARK_MIN
] = tmp
;
4520 zone
->watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + (tmp
>> 2);
4521 zone
->watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + (tmp
>> 1);
4522 setup_zone_migrate_reserve(zone
);
4523 spin_unlock_irqrestore(&zone
->lock
, flags
);
4526 /* update totalreserve_pages */
4527 calculate_totalreserve_pages();
4531 * The inactive anon list should be small enough that the VM never has to
4532 * do too much work, but large enough that each inactive page has a chance
4533 * to be referenced again before it is swapped out.
4535 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
4536 * INACTIVE_ANON pages on this zone's LRU, maintained by the
4537 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
4538 * the anonymous pages are kept on the inactive list.
4541 * memory ratio inactive anon
4542 * -------------------------------------
4551 void calculate_zone_inactive_ratio(struct zone
*zone
)
4553 unsigned int gb
, ratio
;
4555 /* Zone size in gigabytes */
4556 gb
= zone
->present_pages
>> (30 - PAGE_SHIFT
);
4558 ratio
= int_sqrt(10 * gb
);
4562 zone
->inactive_ratio
= ratio
;
4565 static void __init
setup_per_zone_inactive_ratio(void)
4570 calculate_zone_inactive_ratio(zone
);
4574 * Initialise min_free_kbytes.
4576 * For small machines we want it small (128k min). For large machines
4577 * we want it large (64MB max). But it is not linear, because network
4578 * bandwidth does not increase linearly with machine size. We use
4580 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
4581 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
4597 static int __init
init_per_zone_wmark_min(void)
4599 unsigned long lowmem_kbytes
;
4601 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
4603 min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
4604 if (min_free_kbytes
< 128)
4605 min_free_kbytes
= 128;
4606 if (min_free_kbytes
> 65536)
4607 min_free_kbytes
= 65536;
4608 setup_per_zone_wmarks();
4609 setup_per_zone_lowmem_reserve();
4610 setup_per_zone_inactive_ratio();
4613 module_init(init_per_zone_wmark_min
)
4616 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
4617 * that we can call two helper functions whenever min_free_kbytes
4620 int min_free_kbytes_sysctl_handler(ctl_table
*table
, int write
,
4621 struct file
*file
, void __user
*buffer
, size_t *length
, loff_t
*ppos
)
4623 proc_dointvec(table
, write
, file
, buffer
, length
, ppos
);
4625 setup_per_zone_wmarks();
4630 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table
*table
, int write
,
4631 struct file
*file
, void __user
*buffer
, size_t *length
, loff_t
*ppos
)
4636 rc
= proc_dointvec_minmax(table
, write
, file
, buffer
, length
, ppos
);
4641 zone
->min_unmapped_pages
= (zone
->present_pages
*
4642 sysctl_min_unmapped_ratio
) / 100;
4646 int sysctl_min_slab_ratio_sysctl_handler(ctl_table
*table
, int write
,
4647 struct file
*file
, void __user
*buffer
, size_t *length
, loff_t
*ppos
)
4652 rc
= proc_dointvec_minmax(table
, write
, file
, buffer
, length
, ppos
);
4657 zone
->min_slab_pages
= (zone
->present_pages
*
4658 sysctl_min_slab_ratio
) / 100;
4664 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
4665 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
4666 * whenever sysctl_lowmem_reserve_ratio changes.
4668 * The reserve ratio obviously has absolutely no relation with the
4669 * minimum watermarks. The lowmem reserve ratio can only make sense
4670 * if in function of the boot time zone sizes.
4672 int lowmem_reserve_ratio_sysctl_handler(ctl_table
*table
, int write
,
4673 struct file
*file
, void __user
*buffer
, size_t *length
, loff_t
*ppos
)
4675 proc_dointvec_minmax(table
, write
, file
, buffer
, length
, ppos
);
4676 setup_per_zone_lowmem_reserve();
4681 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
4682 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
4683 * can have before it gets flushed back to buddy allocator.
4686 int percpu_pagelist_fraction_sysctl_handler(ctl_table
*table
, int write
,
4687 struct file
*file
, void __user
*buffer
, size_t *length
, loff_t
*ppos
)
4693 ret
= proc_dointvec_minmax(table
, write
, file
, buffer
, length
, ppos
);
4694 if (!write
|| (ret
== -EINVAL
))
4696 for_each_populated_zone(zone
) {
4697 for_each_online_cpu(cpu
) {
4699 high
= zone
->present_pages
/ percpu_pagelist_fraction
;
4700 setup_pagelist_highmark(zone_pcp(zone
, cpu
), high
);
4706 int hashdist
= HASHDIST_DEFAULT
;
4709 static int __init
set_hashdist(char *str
)
4713 hashdist
= simple_strtoul(str
, &str
, 0);
4716 __setup("hashdist=", set_hashdist
);
4720 * allocate a large system hash table from bootmem
4721 * - it is assumed that the hash table must contain an exact power-of-2
4722 * quantity of entries
4723 * - limit is the number of hash buckets, not the total allocation size
4725 void *__init
alloc_large_system_hash(const char *tablename
,
4726 unsigned long bucketsize
,
4727 unsigned long numentries
,
4730 unsigned int *_hash_shift
,
4731 unsigned int *_hash_mask
,
4732 unsigned long limit
)
4734 unsigned long long max
= limit
;
4735 unsigned long log2qty
, size
;
4738 /* allow the kernel cmdline to have a say */
4740 /* round applicable memory size up to nearest megabyte */
4741 numentries
= nr_kernel_pages
;
4742 numentries
+= (1UL << (20 - PAGE_SHIFT
)) - 1;
4743 numentries
>>= 20 - PAGE_SHIFT
;
4744 numentries
<<= 20 - PAGE_SHIFT
;
4746 /* limit to 1 bucket per 2^scale bytes of low memory */
4747 if (scale
> PAGE_SHIFT
)
4748 numentries
>>= (scale
- PAGE_SHIFT
);
4750 numentries
<<= (PAGE_SHIFT
- scale
);
4752 /* Make sure we've got at least a 0-order allocation.. */
4753 if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
4754 numentries
= PAGE_SIZE
/ bucketsize
;
4756 numentries
= roundup_pow_of_two(numentries
);
4758 /* limit allocation size to 1/16 total memory by default */
4760 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
4761 do_div(max
, bucketsize
);
4764 if (numentries
> max
)
4767 log2qty
= ilog2(numentries
);
4770 size
= bucketsize
<< log2qty
;
4771 if (flags
& HASH_EARLY
)
4772 table
= alloc_bootmem_nopanic(size
);
4774 table
= __vmalloc(size
, GFP_ATOMIC
, PAGE_KERNEL
);
4777 * If bucketsize is not a power-of-two, we may free
4778 * some pages at the end of hash table which
4779 * alloc_pages_exact() automatically does
4781 if (get_order(size
) < MAX_ORDER
) {
4782 table
= alloc_pages_exact(size
, GFP_ATOMIC
);
4783 kmemleak_alloc(table
, size
, 1, GFP_ATOMIC
);
4786 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
4789 panic("Failed to allocate %s hash table\n", tablename
);
4791 printk(KERN_INFO
"%s hash table entries: %d (order: %d, %lu bytes)\n",
4794 ilog2(size
) - PAGE_SHIFT
,
4798 *_hash_shift
= log2qty
;
4800 *_hash_mask
= (1 << log2qty
) - 1;
4805 /* Return a pointer to the bitmap storing bits affecting a block of pages */
4806 static inline unsigned long *get_pageblock_bitmap(struct zone
*zone
,
4809 #ifdef CONFIG_SPARSEMEM
4810 return __pfn_to_section(pfn
)->pageblock_flags
;
4812 return zone
->pageblock_flags
;
4813 #endif /* CONFIG_SPARSEMEM */
4816 static inline int pfn_to_bitidx(struct zone
*zone
, unsigned long pfn
)
4818 #ifdef CONFIG_SPARSEMEM
4819 pfn
&= (PAGES_PER_SECTION
-1);
4820 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
4822 pfn
= pfn
- zone
->zone_start_pfn
;
4823 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
4824 #endif /* CONFIG_SPARSEMEM */
4828 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
4829 * @page: The page within the block of interest
4830 * @start_bitidx: The first bit of interest to retrieve
4831 * @end_bitidx: The last bit of interest
4832 * returns pageblock_bits flags
4834 unsigned long get_pageblock_flags_group(struct page
*page
,
4835 int start_bitidx
, int end_bitidx
)
4838 unsigned long *bitmap
;
4839 unsigned long pfn
, bitidx
;
4840 unsigned long flags
= 0;
4841 unsigned long value
= 1;
4843 zone
= page_zone(page
);
4844 pfn
= page_to_pfn(page
);
4845 bitmap
= get_pageblock_bitmap(zone
, pfn
);
4846 bitidx
= pfn_to_bitidx(zone
, pfn
);
4848 for (; start_bitidx
<= end_bitidx
; start_bitidx
++, value
<<= 1)
4849 if (test_bit(bitidx
+ start_bitidx
, bitmap
))
4856 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
4857 * @page: The page within the block of interest
4858 * @start_bitidx: The first bit of interest
4859 * @end_bitidx: The last bit of interest
4860 * @flags: The flags to set
4862 void set_pageblock_flags_group(struct page
*page
, unsigned long flags
,
4863 int start_bitidx
, int end_bitidx
)
4866 unsigned long *bitmap
;
4867 unsigned long pfn
, bitidx
;
4868 unsigned long value
= 1;
4870 zone
= page_zone(page
);
4871 pfn
= page_to_pfn(page
);
4872 bitmap
= get_pageblock_bitmap(zone
, pfn
);
4873 bitidx
= pfn_to_bitidx(zone
, pfn
);
4874 VM_BUG_ON(pfn
< zone
->zone_start_pfn
);
4875 VM_BUG_ON(pfn
>= zone
->zone_start_pfn
+ zone
->spanned_pages
);
4877 for (; start_bitidx
<= end_bitidx
; start_bitidx
++, value
<<= 1)
4879 __set_bit(bitidx
+ start_bitidx
, bitmap
);
4881 __clear_bit(bitidx
+ start_bitidx
, bitmap
);
4885 * This is designed as sub function...plz see page_isolation.c also.
4886 * set/clear page block's type to be ISOLATE.
4887 * page allocater never alloc memory from ISOLATE block.
4890 int set_migratetype_isolate(struct page
*page
)
4893 unsigned long flags
;
4896 zone
= page_zone(page
);
4897 spin_lock_irqsave(&zone
->lock
, flags
);
4899 * In future, more migrate types will be able to be isolation target.
4901 if (get_pageblock_migratetype(page
) != MIGRATE_MOVABLE
)
4903 set_pageblock_migratetype(page
, MIGRATE_ISOLATE
);
4904 move_freepages_block(zone
, page
, MIGRATE_ISOLATE
);
4907 spin_unlock_irqrestore(&zone
->lock
, flags
);
4913 void unset_migratetype_isolate(struct page
*page
)
4916 unsigned long flags
;
4917 zone
= page_zone(page
);
4918 spin_lock_irqsave(&zone
->lock
, flags
);
4919 if (get_pageblock_migratetype(page
) != MIGRATE_ISOLATE
)
4921 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
4922 move_freepages_block(zone
, page
, MIGRATE_MOVABLE
);
4924 spin_unlock_irqrestore(&zone
->lock
, flags
);
4927 #ifdef CONFIG_MEMORY_HOTREMOVE
4929 * All pages in the range must be isolated before calling this.
4932 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
4938 unsigned long flags
;
4939 /* find the first valid pfn */
4940 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
4945 zone
= page_zone(pfn_to_page(pfn
));
4946 spin_lock_irqsave(&zone
->lock
, flags
);
4948 while (pfn
< end_pfn
) {
4949 if (!pfn_valid(pfn
)) {
4953 page
= pfn_to_page(pfn
);
4954 BUG_ON(page_count(page
));
4955 BUG_ON(!PageBuddy(page
));
4956 order
= page_order(page
);
4957 #ifdef CONFIG_DEBUG_VM
4958 printk(KERN_INFO
"remove from free list %lx %d %lx\n",
4959 pfn
, 1 << order
, end_pfn
);
4961 list_del(&page
->lru
);
4962 rmv_page_order(page
);
4963 zone
->free_area
[order
].nr_free
--;
4964 __mod_zone_page_state(zone
, NR_FREE_PAGES
,
4966 for (i
= 0; i
< (1 << order
); i
++)
4967 SetPageReserved((page
+i
));
4968 pfn
+= (1 << order
);
4970 spin_unlock_irqrestore(&zone
->lock
, flags
);