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>
56 #define CREATE_TRACE_POINTS
57 #include <trace/events/kmem.h>
60 * Array of node states.
62 nodemask_t node_states
[NR_NODE_STATES
] __read_mostly
= {
63 [N_POSSIBLE
] = NODE_MASK_ALL
,
64 [N_ONLINE
] = { { [0] = 1UL } },
66 [N_NORMAL_MEMORY
] = { { [0] = 1UL } },
68 [N_HIGH_MEMORY
] = { { [0] = 1UL } },
70 [N_CPU
] = { { [0] = 1UL } },
73 EXPORT_SYMBOL(node_states
);
75 unsigned long totalram_pages __read_mostly
;
76 unsigned long totalreserve_pages __read_mostly
;
77 int percpu_pagelist_fraction
;
78 gfp_t gfp_allowed_mask __read_mostly
= GFP_BOOT_MASK
;
80 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
81 int pageblock_order __read_mostly
;
84 static void __free_pages_ok(struct page
*page
, unsigned int order
);
87 * results with 256, 32 in the lowmem_reserve sysctl:
88 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
89 * 1G machine -> (16M dma, 784M normal, 224M high)
90 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
91 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
92 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
94 * TBD: should special case ZONE_DMA32 machines here - in those we normally
95 * don't need any ZONE_NORMAL reservation
97 int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
-1] = {
98 #ifdef CONFIG_ZONE_DMA
101 #ifdef CONFIG_ZONE_DMA32
104 #ifdef CONFIG_HIGHMEM
110 EXPORT_SYMBOL(totalram_pages
);
112 static char * const zone_names
[MAX_NR_ZONES
] = {
113 #ifdef CONFIG_ZONE_DMA
116 #ifdef CONFIG_ZONE_DMA32
120 #ifdef CONFIG_HIGHMEM
126 int min_free_kbytes
= 1024;
128 static unsigned long __meminitdata nr_kernel_pages
;
129 static unsigned long __meminitdata nr_all_pages
;
130 static unsigned long __meminitdata dma_reserve
;
132 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
134 * MAX_ACTIVE_REGIONS determines the maximum number of distinct
135 * ranges of memory (RAM) that may be registered with add_active_range().
136 * Ranges passed to add_active_range() will be merged if possible
137 * so the number of times add_active_range() can be called is
138 * related to the number of nodes and the number of holes
140 #ifdef CONFIG_MAX_ACTIVE_REGIONS
141 /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */
142 #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS
144 #if MAX_NUMNODES >= 32
145 /* If there can be many nodes, allow up to 50 holes per node */
146 #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50)
148 /* By default, allow up to 256 distinct regions */
149 #define MAX_ACTIVE_REGIONS 256
153 static struct node_active_region __meminitdata early_node_map
[MAX_ACTIVE_REGIONS
];
154 static int __meminitdata nr_nodemap_entries
;
155 static unsigned long __meminitdata arch_zone_lowest_possible_pfn
[MAX_NR_ZONES
];
156 static unsigned long __meminitdata arch_zone_highest_possible_pfn
[MAX_NR_ZONES
];
157 static unsigned long __initdata required_kernelcore
;
158 static unsigned long __initdata required_movablecore
;
159 static unsigned long __meminitdata zone_movable_pfn
[MAX_NUMNODES
];
161 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
163 EXPORT_SYMBOL(movable_zone
);
164 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
167 int nr_node_ids __read_mostly
= MAX_NUMNODES
;
168 int nr_online_nodes __read_mostly
= 1;
169 EXPORT_SYMBOL(nr_node_ids
);
170 EXPORT_SYMBOL(nr_online_nodes
);
173 int page_group_by_mobility_disabled __read_mostly
;
175 static void set_pageblock_migratetype(struct page
*page
, int migratetype
)
178 if (unlikely(page_group_by_mobility_disabled
))
179 migratetype
= MIGRATE_UNMOVABLE
;
181 set_pageblock_flags_group(page
, (unsigned long)migratetype
,
182 PB_migrate
, PB_migrate_end
);
185 bool oom_killer_disabled __read_mostly
;
187 #ifdef CONFIG_DEBUG_VM
188 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
192 unsigned long pfn
= page_to_pfn(page
);
195 seq
= zone_span_seqbegin(zone
);
196 if (pfn
>= zone
->zone_start_pfn
+ zone
->spanned_pages
)
198 else if (pfn
< zone
->zone_start_pfn
)
200 } while (zone_span_seqretry(zone
, seq
));
205 static int page_is_consistent(struct zone
*zone
, struct page
*page
)
207 if (!pfn_valid_within(page_to_pfn(page
)))
209 if (zone
!= page_zone(page
))
215 * Temporary debugging check for pages not lying within a given zone.
217 static int bad_range(struct zone
*zone
, struct page
*page
)
219 if (page_outside_zone_boundaries(zone
, page
))
221 if (!page_is_consistent(zone
, page
))
227 static inline int bad_range(struct zone
*zone
, struct page
*page
)
233 static void bad_page(struct page
*page
)
235 static unsigned long resume
;
236 static unsigned long nr_shown
;
237 static unsigned long nr_unshown
;
239 /* Don't complain about poisoned pages */
240 if (PageHWPoison(page
)) {
241 __ClearPageBuddy(page
);
246 * Allow a burst of 60 reports, then keep quiet for that minute;
247 * or allow a steady drip of one report per second.
249 if (nr_shown
== 60) {
250 if (time_before(jiffies
, resume
)) {
256 "BUG: Bad page state: %lu messages suppressed\n",
263 resume
= jiffies
+ 60 * HZ
;
265 printk(KERN_ALERT
"BUG: Bad page state in process %s pfn:%05lx\n",
266 current
->comm
, page_to_pfn(page
));
268 "page:%p flags:%p count:%d mapcount:%d mapping:%p index:%lx\n",
269 page
, (void *)page
->flags
, page_count(page
),
270 page_mapcount(page
), page
->mapping
, page
->index
);
274 /* Leave bad fields for debug, except PageBuddy could make trouble */
275 __ClearPageBuddy(page
);
276 add_taint(TAINT_BAD_PAGE
);
280 * Higher-order pages are called "compound pages". They are structured thusly:
282 * The first PAGE_SIZE page is called the "head page".
284 * The remaining PAGE_SIZE pages are called "tail pages".
286 * All pages have PG_compound set. All pages have their ->private pointing at
287 * the head page (even the head page has this).
289 * The first tail page's ->lru.next holds the address of the compound page's
290 * put_page() function. Its ->lru.prev holds the order of allocation.
291 * This usage means that zero-order pages may not be compound.
294 static void free_compound_page(struct page
*page
)
296 __free_pages_ok(page
, compound_order(page
));
299 void prep_compound_page(struct page
*page
, unsigned long order
)
302 int nr_pages
= 1 << order
;
304 set_compound_page_dtor(page
, free_compound_page
);
305 set_compound_order(page
, order
);
307 for (i
= 1; i
< nr_pages
; i
++) {
308 struct page
*p
= page
+ i
;
311 p
->first_page
= page
;
315 static int destroy_compound_page(struct page
*page
, unsigned long order
)
318 int nr_pages
= 1 << order
;
321 if (unlikely(compound_order(page
) != order
) ||
322 unlikely(!PageHead(page
))) {
327 __ClearPageHead(page
);
329 for (i
= 1; i
< nr_pages
; i
++) {
330 struct page
*p
= page
+ i
;
332 if (unlikely(!PageTail(p
) || (p
->first_page
!= page
))) {
342 static inline void prep_zero_page(struct page
*page
, int order
, gfp_t gfp_flags
)
347 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
348 * and __GFP_HIGHMEM from hard or soft interrupt context.
350 VM_BUG_ON((gfp_flags
& __GFP_HIGHMEM
) && in_interrupt());
351 for (i
= 0; i
< (1 << order
); i
++)
352 clear_highpage(page
+ i
);
355 static inline void set_page_order(struct page
*page
, int order
)
357 set_page_private(page
, order
);
358 __SetPageBuddy(page
);
361 static inline void rmv_page_order(struct page
*page
)
363 __ClearPageBuddy(page
);
364 set_page_private(page
, 0);
368 * Locate the struct page for both the matching buddy in our
369 * pair (buddy1) and the combined O(n+1) page they form (page).
371 * 1) Any buddy B1 will have an order O twin B2 which satisfies
372 * the following equation:
374 * For example, if the starting buddy (buddy2) is #8 its order
376 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
378 * 2) Any buddy B will have an order O+1 parent P which
379 * satisfies the following equation:
382 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
384 static inline struct page
*
385 __page_find_buddy(struct page
*page
, unsigned long page_idx
, unsigned int order
)
387 unsigned long buddy_idx
= page_idx
^ (1 << order
);
389 return page
+ (buddy_idx
- page_idx
);
392 static inline unsigned long
393 __find_combined_index(unsigned long page_idx
, unsigned int order
)
395 return (page_idx
& ~(1 << order
));
399 * This function checks whether a page is free && is the buddy
400 * we can do coalesce a page and its buddy if
401 * (a) the buddy is not in a hole &&
402 * (b) the buddy is in the buddy system &&
403 * (c) a page and its buddy have the same order &&
404 * (d) a page and its buddy are in the same zone.
406 * For recording whether a page is in the buddy system, we use PG_buddy.
407 * Setting, clearing, and testing PG_buddy is serialized by zone->lock.
409 * For recording page's order, we use page_private(page).
411 static inline int page_is_buddy(struct page
*page
, struct page
*buddy
,
414 if (!pfn_valid_within(page_to_pfn(buddy
)))
417 if (page_zone_id(page
) != page_zone_id(buddy
))
420 if (PageBuddy(buddy
) && page_order(buddy
) == order
) {
421 VM_BUG_ON(page_count(buddy
) != 0);
428 * Freeing function for a buddy system allocator.
430 * The concept of a buddy system is to maintain direct-mapped table
431 * (containing bit values) for memory blocks of various "orders".
432 * The bottom level table contains the map for the smallest allocatable
433 * units of memory (here, pages), and each level above it describes
434 * pairs of units from the levels below, hence, "buddies".
435 * At a high level, all that happens here is marking the table entry
436 * at the bottom level available, and propagating the changes upward
437 * as necessary, plus some accounting needed to play nicely with other
438 * parts of the VM system.
439 * At each level, we keep a list of pages, which are heads of continuous
440 * free pages of length of (1 << order) and marked with PG_buddy. Page's
441 * order is recorded in page_private(page) field.
442 * So when we are allocating or freeing one, we can derive the state of the
443 * other. That is, if we allocate a small block, and both were
444 * free, the remainder of the region must be split into blocks.
445 * If a block is freed, and its buddy is also free, then this
446 * triggers coalescing into a block of larger size.
451 static inline void __free_one_page(struct page
*page
,
452 struct zone
*zone
, unsigned int order
,
455 unsigned long page_idx
;
457 if (unlikely(PageCompound(page
)))
458 if (unlikely(destroy_compound_page(page
, order
)))
461 VM_BUG_ON(migratetype
== -1);
463 page_idx
= page_to_pfn(page
) & ((1 << MAX_ORDER
) - 1);
465 VM_BUG_ON(page_idx
& ((1 << order
) - 1));
466 VM_BUG_ON(bad_range(zone
, page
));
468 while (order
< MAX_ORDER
-1) {
469 unsigned long combined_idx
;
472 buddy
= __page_find_buddy(page
, page_idx
, order
);
473 if (!page_is_buddy(page
, buddy
, order
))
476 /* Our buddy is free, merge with it and move up one order. */
477 list_del(&buddy
->lru
);
478 zone
->free_area
[order
].nr_free
--;
479 rmv_page_order(buddy
);
480 combined_idx
= __find_combined_index(page_idx
, order
);
481 page
= page
+ (combined_idx
- page_idx
);
482 page_idx
= combined_idx
;
485 set_page_order(page
, order
);
487 &zone
->free_area
[order
].free_list
[migratetype
]);
488 zone
->free_area
[order
].nr_free
++;
491 #ifdef CONFIG_HAVE_MLOCKED_PAGE_BIT
493 * free_page_mlock() -- clean up attempts to free and mlocked() page.
494 * Page should not be on lru, so no need to fix that up.
495 * free_pages_check() will verify...
497 static inline void free_page_mlock(struct page
*page
)
499 __dec_zone_page_state(page
, NR_MLOCK
);
500 __count_vm_event(UNEVICTABLE_MLOCKFREED
);
503 static void free_page_mlock(struct page
*page
) { }
506 static inline int free_pages_check(struct page
*page
)
508 if (unlikely(page_mapcount(page
) |
509 (page
->mapping
!= NULL
) |
510 (atomic_read(&page
->_count
) != 0) |
511 (page
->flags
& PAGE_FLAGS_CHECK_AT_FREE
))) {
515 if (page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)
516 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
521 * Frees a number of pages from the PCP lists
522 * Assumes all pages on list are in same zone, and of same order.
523 * count is the number of pages to free.
525 * If the zone was previously in an "all pages pinned" state then look to
526 * see if this freeing clears that state.
528 * And clear the zone's pages_scanned counter, to hold off the "all pages are
529 * pinned" detection logic.
531 static void free_pcppages_bulk(struct zone
*zone
, int count
,
532 struct per_cpu_pages
*pcp
)
537 spin_lock(&zone
->lock
);
538 zone_clear_flag(zone
, ZONE_ALL_UNRECLAIMABLE
);
539 zone
->pages_scanned
= 0;
541 __mod_zone_page_state(zone
, NR_FREE_PAGES
, count
);
544 struct list_head
*list
;
547 * Remove pages from lists in a round-robin fashion. A
548 * batch_free count is maintained that is incremented when an
549 * empty list is encountered. This is so more pages are freed
550 * off fuller lists instead of spinning excessively around empty
555 if (++migratetype
== MIGRATE_PCPTYPES
)
557 list
= &pcp
->lists
[migratetype
];
558 } while (list_empty(list
));
561 page
= list_entry(list
->prev
, struct page
, lru
);
562 /* must delete as __free_one_page list manipulates */
563 list_del(&page
->lru
);
564 __free_one_page(page
, zone
, 0, migratetype
);
565 trace_mm_page_pcpu_drain(page
, 0, migratetype
);
566 } while (--count
&& --batch_free
&& !list_empty(list
));
568 spin_unlock(&zone
->lock
);
571 static void free_one_page(struct zone
*zone
, struct page
*page
, int order
,
574 spin_lock(&zone
->lock
);
575 zone_clear_flag(zone
, ZONE_ALL_UNRECLAIMABLE
);
576 zone
->pages_scanned
= 0;
578 __mod_zone_page_state(zone
, NR_FREE_PAGES
, 1 << order
);
579 __free_one_page(page
, zone
, order
, migratetype
);
580 spin_unlock(&zone
->lock
);
583 static void __free_pages_ok(struct page
*page
, unsigned int order
)
588 int wasMlocked
= __TestClearPageMlocked(page
);
590 kmemcheck_free_shadow(page
, order
);
592 for (i
= 0 ; i
< (1 << order
) ; ++i
)
593 bad
+= free_pages_check(page
+ i
);
597 if (!PageHighMem(page
)) {
598 debug_check_no_locks_freed(page_address(page
),PAGE_SIZE
<<order
);
599 debug_check_no_obj_freed(page_address(page
),
602 arch_free_page(page
, order
);
603 kernel_map_pages(page
, 1 << order
, 0);
605 local_irq_save(flags
);
606 if (unlikely(wasMlocked
))
607 free_page_mlock(page
);
608 __count_vm_events(PGFREE
, 1 << order
);
609 free_one_page(page_zone(page
), page
, order
,
610 get_pageblock_migratetype(page
));
611 local_irq_restore(flags
);
615 * permit the bootmem allocator to evade page validation on high-order frees
617 void __meminit
__free_pages_bootmem(struct page
*page
, unsigned int order
)
620 __ClearPageReserved(page
);
621 set_page_count(page
, 0);
622 set_page_refcounted(page
);
628 for (loop
= 0; loop
< BITS_PER_LONG
; loop
++) {
629 struct page
*p
= &page
[loop
];
631 if (loop
+ 1 < BITS_PER_LONG
)
633 __ClearPageReserved(p
);
634 set_page_count(p
, 0);
637 set_page_refcounted(page
);
638 __free_pages(page
, order
);
644 * The order of subdivision here is critical for the IO subsystem.
645 * Please do not alter this order without good reasons and regression
646 * testing. Specifically, as large blocks of memory are subdivided,
647 * the order in which smaller blocks are delivered depends on the order
648 * they're subdivided in this function. This is the primary factor
649 * influencing the order in which pages are delivered to the IO
650 * subsystem according to empirical testing, and this is also justified
651 * by considering the behavior of a buddy system containing a single
652 * large block of memory acted on by a series of small allocations.
653 * This behavior is a critical factor in sglist merging's success.
657 static inline void expand(struct zone
*zone
, struct page
*page
,
658 int low
, int high
, struct free_area
*area
,
661 unsigned long size
= 1 << high
;
667 VM_BUG_ON(bad_range(zone
, &page
[size
]));
668 list_add(&page
[size
].lru
, &area
->free_list
[migratetype
]);
670 set_page_order(&page
[size
], high
);
675 * This page is about to be returned from the page allocator
677 static inline int check_new_page(struct page
*page
)
679 if (unlikely(page_mapcount(page
) |
680 (page
->mapping
!= NULL
) |
681 (atomic_read(&page
->_count
) != 0) |
682 (page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
))) {
689 static int prep_new_page(struct page
*page
, int order
, gfp_t gfp_flags
)
693 for (i
= 0; i
< (1 << order
); i
++) {
694 struct page
*p
= page
+ i
;
695 if (unlikely(check_new_page(p
)))
699 set_page_private(page
, 0);
700 set_page_refcounted(page
);
702 arch_alloc_page(page
, order
);
703 kernel_map_pages(page
, 1 << order
, 1);
705 if (gfp_flags
& __GFP_ZERO
)
706 prep_zero_page(page
, order
, gfp_flags
);
708 if (order
&& (gfp_flags
& __GFP_COMP
))
709 prep_compound_page(page
, order
);
715 * Go through the free lists for the given migratetype and remove
716 * the smallest available page from the freelists
719 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
722 unsigned int current_order
;
723 struct free_area
* area
;
726 /* Find a page of the appropriate size in the preferred list */
727 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
728 area
= &(zone
->free_area
[current_order
]);
729 if (list_empty(&area
->free_list
[migratetype
]))
732 page
= list_entry(area
->free_list
[migratetype
].next
,
734 list_del(&page
->lru
);
735 rmv_page_order(page
);
737 expand(zone
, page
, order
, current_order
, area
, migratetype
);
746 * This array describes the order lists are fallen back to when
747 * the free lists for the desirable migrate type are depleted
749 static int fallbacks
[MIGRATE_TYPES
][MIGRATE_TYPES
-1] = {
750 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
751 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
752 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_RESERVE
},
753 [MIGRATE_RESERVE
] = { MIGRATE_RESERVE
, MIGRATE_RESERVE
, MIGRATE_RESERVE
}, /* Never used */
757 * Move the free pages in a range to the free lists of the requested type.
758 * Note that start_page and end_pages are not aligned on a pageblock
759 * boundary. If alignment is required, use move_freepages_block()
761 static int move_freepages(struct zone
*zone
,
762 struct page
*start_page
, struct page
*end_page
,
769 #ifndef CONFIG_HOLES_IN_ZONE
771 * page_zone is not safe to call in this context when
772 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
773 * anyway as we check zone boundaries in move_freepages_block().
774 * Remove at a later date when no bug reports exist related to
775 * grouping pages by mobility
777 BUG_ON(page_zone(start_page
) != page_zone(end_page
));
780 for (page
= start_page
; page
<= end_page
;) {
781 /* Make sure we are not inadvertently changing nodes */
782 VM_BUG_ON(page_to_nid(page
) != zone_to_nid(zone
));
784 if (!pfn_valid_within(page_to_pfn(page
))) {
789 if (!PageBuddy(page
)) {
794 order
= page_order(page
);
795 list_del(&page
->lru
);
797 &zone
->free_area
[order
].free_list
[migratetype
]);
799 pages_moved
+= 1 << order
;
805 static int move_freepages_block(struct zone
*zone
, struct page
*page
,
808 unsigned long start_pfn
, end_pfn
;
809 struct page
*start_page
, *end_page
;
811 start_pfn
= page_to_pfn(page
);
812 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
813 start_page
= pfn_to_page(start_pfn
);
814 end_page
= start_page
+ pageblock_nr_pages
- 1;
815 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
817 /* Do not cross zone boundaries */
818 if (start_pfn
< zone
->zone_start_pfn
)
820 if (end_pfn
>= zone
->zone_start_pfn
+ zone
->spanned_pages
)
823 return move_freepages(zone
, start_page
, end_page
, migratetype
);
826 static void change_pageblock_range(struct page
*pageblock_page
,
827 int start_order
, int migratetype
)
829 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
831 while (nr_pageblocks
--) {
832 set_pageblock_migratetype(pageblock_page
, migratetype
);
833 pageblock_page
+= pageblock_nr_pages
;
837 /* Remove an element from the buddy allocator from the fallback list */
838 static inline struct page
*
839 __rmqueue_fallback(struct zone
*zone
, int order
, int start_migratetype
)
841 struct free_area
* area
;
846 /* Find the largest possible block of pages in the other list */
847 for (current_order
= MAX_ORDER
-1; current_order
>= order
;
849 for (i
= 0; i
< MIGRATE_TYPES
- 1; i
++) {
850 migratetype
= fallbacks
[start_migratetype
][i
];
852 /* MIGRATE_RESERVE handled later if necessary */
853 if (migratetype
== MIGRATE_RESERVE
)
856 area
= &(zone
->free_area
[current_order
]);
857 if (list_empty(&area
->free_list
[migratetype
]))
860 page
= list_entry(area
->free_list
[migratetype
].next
,
865 * If breaking a large block of pages, move all free
866 * pages to the preferred allocation list. If falling
867 * back for a reclaimable kernel allocation, be more
868 * agressive about taking ownership of free pages
870 if (unlikely(current_order
>= (pageblock_order
>> 1)) ||
871 start_migratetype
== MIGRATE_RECLAIMABLE
||
872 page_group_by_mobility_disabled
) {
874 pages
= move_freepages_block(zone
, page
,
877 /* Claim the whole block if over half of it is free */
878 if (pages
>= (1 << (pageblock_order
-1)) ||
879 page_group_by_mobility_disabled
)
880 set_pageblock_migratetype(page
,
883 migratetype
= start_migratetype
;
886 /* Remove the page from the freelists */
887 list_del(&page
->lru
);
888 rmv_page_order(page
);
890 /* Take ownership for orders >= pageblock_order */
891 if (current_order
>= pageblock_order
)
892 change_pageblock_range(page
, current_order
,
895 expand(zone
, page
, order
, current_order
, area
, migratetype
);
897 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
898 start_migratetype
, migratetype
);
908 * Do the hard work of removing an element from the buddy allocator.
909 * Call me with the zone->lock already held.
911 static struct page
*__rmqueue(struct zone
*zone
, unsigned int order
,
917 page
= __rmqueue_smallest(zone
, order
, migratetype
);
919 if (unlikely(!page
) && migratetype
!= MIGRATE_RESERVE
) {
920 page
= __rmqueue_fallback(zone
, order
, migratetype
);
923 * Use MIGRATE_RESERVE rather than fail an allocation. goto
924 * is used because __rmqueue_smallest is an inline function
925 * and we want just one call site
928 migratetype
= MIGRATE_RESERVE
;
933 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
938 * Obtain a specified number of elements from the buddy allocator, all under
939 * a single hold of the lock, for efficiency. Add them to the supplied list.
940 * Returns the number of new pages which were placed at *list.
942 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
943 unsigned long count
, struct list_head
*list
,
944 int migratetype
, int cold
)
948 spin_lock(&zone
->lock
);
949 for (i
= 0; i
< count
; ++i
) {
950 struct page
*page
= __rmqueue(zone
, order
, migratetype
);
951 if (unlikely(page
== NULL
))
955 * Split buddy pages returned by expand() are received here
956 * in physical page order. The page is added to the callers and
957 * list and the list head then moves forward. From the callers
958 * perspective, the linked list is ordered by page number in
959 * some conditions. This is useful for IO devices that can
960 * merge IO requests if the physical pages are ordered
963 if (likely(cold
== 0))
964 list_add(&page
->lru
, list
);
966 list_add_tail(&page
->lru
, list
);
967 set_page_private(page
, migratetype
);
970 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
971 spin_unlock(&zone
->lock
);
977 * Called from the vmstat counter updater to drain pagesets of this
978 * currently executing processor on remote nodes after they have
981 * Note that this function must be called with the thread pinned to
982 * a single processor.
984 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
989 local_irq_save(flags
);
990 if (pcp
->count
>= pcp
->batch
)
991 to_drain
= pcp
->batch
;
993 to_drain
= pcp
->count
;
994 free_pcppages_bulk(zone
, to_drain
, pcp
);
995 pcp
->count
-= to_drain
;
996 local_irq_restore(flags
);
1001 * Drain pages of the indicated processor.
1003 * The processor must either be the current processor and the
1004 * thread pinned to the current processor or a processor that
1007 static void drain_pages(unsigned int cpu
)
1009 unsigned long flags
;
1012 for_each_populated_zone(zone
) {
1013 struct per_cpu_pageset
*pset
;
1014 struct per_cpu_pages
*pcp
;
1016 pset
= zone_pcp(zone
, cpu
);
1019 local_irq_save(flags
);
1020 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
1022 local_irq_restore(flags
);
1027 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1029 void drain_local_pages(void *arg
)
1031 drain_pages(smp_processor_id());
1035 * Spill all the per-cpu pages from all CPUs back into the buddy allocator
1037 void drain_all_pages(void)
1039 on_each_cpu(drain_local_pages
, NULL
, 1);
1042 #ifdef CONFIG_HIBERNATION
1044 void mark_free_pages(struct zone
*zone
)
1046 unsigned long pfn
, max_zone_pfn
;
1047 unsigned long flags
;
1049 struct list_head
*curr
;
1051 if (!zone
->spanned_pages
)
1054 spin_lock_irqsave(&zone
->lock
, flags
);
1056 max_zone_pfn
= zone
->zone_start_pfn
+ zone
->spanned_pages
;
1057 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
1058 if (pfn_valid(pfn
)) {
1059 struct page
*page
= pfn_to_page(pfn
);
1061 if (!swsusp_page_is_forbidden(page
))
1062 swsusp_unset_page_free(page
);
1065 for_each_migratetype_order(order
, t
) {
1066 list_for_each(curr
, &zone
->free_area
[order
].free_list
[t
]) {
1069 pfn
= page_to_pfn(list_entry(curr
, struct page
, lru
));
1070 for (i
= 0; i
< (1UL << order
); i
++)
1071 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
1074 spin_unlock_irqrestore(&zone
->lock
, flags
);
1076 #endif /* CONFIG_PM */
1079 * Free a 0-order page
1081 static void free_hot_cold_page(struct page
*page
, int cold
)
1083 struct zone
*zone
= page_zone(page
);
1084 struct per_cpu_pages
*pcp
;
1085 unsigned long flags
;
1087 int wasMlocked
= __TestClearPageMlocked(page
);
1089 kmemcheck_free_shadow(page
, 0);
1092 page
->mapping
= NULL
;
1093 if (free_pages_check(page
))
1096 if (!PageHighMem(page
)) {
1097 debug_check_no_locks_freed(page_address(page
), PAGE_SIZE
);
1098 debug_check_no_obj_freed(page_address(page
), PAGE_SIZE
);
1100 arch_free_page(page
, 0);
1101 kernel_map_pages(page
, 1, 0);
1103 pcp
= &zone_pcp(zone
, get_cpu())->pcp
;
1104 migratetype
= get_pageblock_migratetype(page
);
1105 set_page_private(page
, migratetype
);
1106 local_irq_save(flags
);
1107 if (unlikely(wasMlocked
))
1108 free_page_mlock(page
);
1109 __count_vm_event(PGFREE
);
1112 * We only track unmovable, reclaimable and movable on pcp lists.
1113 * Free ISOLATE pages back to the allocator because they are being
1114 * offlined but treat RESERVE as movable pages so we can get those
1115 * areas back if necessary. Otherwise, we may have to free
1116 * excessively into the page allocator
1118 if (migratetype
>= MIGRATE_PCPTYPES
) {
1119 if (unlikely(migratetype
== MIGRATE_ISOLATE
)) {
1120 free_one_page(zone
, page
, 0, migratetype
);
1123 migratetype
= MIGRATE_MOVABLE
;
1127 list_add_tail(&page
->lru
, &pcp
->lists
[migratetype
]);
1129 list_add(&page
->lru
, &pcp
->lists
[migratetype
]);
1131 if (pcp
->count
>= pcp
->high
) {
1132 free_pcppages_bulk(zone
, pcp
->batch
, pcp
);
1133 pcp
->count
-= pcp
->batch
;
1137 local_irq_restore(flags
);
1141 void free_hot_page(struct page
*page
)
1143 trace_mm_page_free_direct(page
, 0);
1144 free_hot_cold_page(page
, 0);
1148 * split_page takes a non-compound higher-order page, and splits it into
1149 * n (1<<order) sub-pages: page[0..n]
1150 * Each sub-page must be freed individually.
1152 * Note: this is probably too low level an operation for use in drivers.
1153 * Please consult with lkml before using this in your driver.
1155 void split_page(struct page
*page
, unsigned int order
)
1159 VM_BUG_ON(PageCompound(page
));
1160 VM_BUG_ON(!page_count(page
));
1162 #ifdef CONFIG_KMEMCHECK
1164 * Split shadow pages too, because free(page[0]) would
1165 * otherwise free the whole shadow.
1167 if (kmemcheck_page_is_tracked(page
))
1168 split_page(virt_to_page(page
[0].shadow
), order
);
1171 for (i
= 1; i
< (1 << order
); i
++)
1172 set_page_refcounted(page
+ i
);
1176 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1177 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1181 struct page
*buffered_rmqueue(struct zone
*preferred_zone
,
1182 struct zone
*zone
, int order
, gfp_t gfp_flags
,
1185 unsigned long flags
;
1187 int cold
= !!(gfp_flags
& __GFP_COLD
);
1192 if (likely(order
== 0)) {
1193 struct per_cpu_pages
*pcp
;
1194 struct list_head
*list
;
1196 pcp
= &zone_pcp(zone
, cpu
)->pcp
;
1197 list
= &pcp
->lists
[migratetype
];
1198 local_irq_save(flags
);
1199 if (list_empty(list
)) {
1200 pcp
->count
+= rmqueue_bulk(zone
, 0,
1203 if (unlikely(list_empty(list
)))
1208 page
= list_entry(list
->prev
, struct page
, lru
);
1210 page
= list_entry(list
->next
, struct page
, lru
);
1212 list_del(&page
->lru
);
1215 if (unlikely(gfp_flags
& __GFP_NOFAIL
)) {
1217 * __GFP_NOFAIL is not to be used in new code.
1219 * All __GFP_NOFAIL callers should be fixed so that they
1220 * properly detect and handle allocation failures.
1222 * We most definitely don't want callers attempting to
1223 * allocate greater than order-1 page units with
1226 WARN_ON_ONCE(order
> 1);
1228 spin_lock_irqsave(&zone
->lock
, flags
);
1229 page
= __rmqueue(zone
, order
, migratetype
);
1230 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(1 << order
));
1231 spin_unlock(&zone
->lock
);
1236 __count_zone_vm_events(PGALLOC
, zone
, 1 << order
);
1237 zone_statistics(preferred_zone
, zone
);
1238 local_irq_restore(flags
);
1241 VM_BUG_ON(bad_range(zone
, page
));
1242 if (prep_new_page(page
, order
, gfp_flags
))
1247 local_irq_restore(flags
);
1252 /* The ALLOC_WMARK bits are used as an index to zone->watermark */
1253 #define ALLOC_WMARK_MIN WMARK_MIN
1254 #define ALLOC_WMARK_LOW WMARK_LOW
1255 #define ALLOC_WMARK_HIGH WMARK_HIGH
1256 #define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */
1258 /* Mask to get the watermark bits */
1259 #define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1)
1261 #define ALLOC_HARDER 0x10 /* try to alloc harder */
1262 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
1263 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
1265 #ifdef CONFIG_FAIL_PAGE_ALLOC
1267 static struct fail_page_alloc_attr
{
1268 struct fault_attr attr
;
1270 u32 ignore_gfp_highmem
;
1271 u32 ignore_gfp_wait
;
1274 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1276 struct dentry
*ignore_gfp_highmem_file
;
1277 struct dentry
*ignore_gfp_wait_file
;
1278 struct dentry
*min_order_file
;
1280 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1282 } fail_page_alloc
= {
1283 .attr
= FAULT_ATTR_INITIALIZER
,
1284 .ignore_gfp_wait
= 1,
1285 .ignore_gfp_highmem
= 1,
1289 static int __init
setup_fail_page_alloc(char *str
)
1291 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
1293 __setup("fail_page_alloc=", setup_fail_page_alloc
);
1295 static int should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1297 if (order
< fail_page_alloc
.min_order
)
1299 if (gfp_mask
& __GFP_NOFAIL
)
1301 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
1303 if (fail_page_alloc
.ignore_gfp_wait
&& (gfp_mask
& __GFP_WAIT
))
1306 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
1309 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1311 static int __init
fail_page_alloc_debugfs(void)
1313 mode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
1317 err
= init_fault_attr_dentries(&fail_page_alloc
.attr
,
1321 dir
= fail_page_alloc
.attr
.dentries
.dir
;
1323 fail_page_alloc
.ignore_gfp_wait_file
=
1324 debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
1325 &fail_page_alloc
.ignore_gfp_wait
);
1327 fail_page_alloc
.ignore_gfp_highmem_file
=
1328 debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
1329 &fail_page_alloc
.ignore_gfp_highmem
);
1330 fail_page_alloc
.min_order_file
=
1331 debugfs_create_u32("min-order", mode
, dir
,
1332 &fail_page_alloc
.min_order
);
1334 if (!fail_page_alloc
.ignore_gfp_wait_file
||
1335 !fail_page_alloc
.ignore_gfp_highmem_file
||
1336 !fail_page_alloc
.min_order_file
) {
1338 debugfs_remove(fail_page_alloc
.ignore_gfp_wait_file
);
1339 debugfs_remove(fail_page_alloc
.ignore_gfp_highmem_file
);
1340 debugfs_remove(fail_page_alloc
.min_order_file
);
1341 cleanup_fault_attr_dentries(&fail_page_alloc
.attr
);
1347 late_initcall(fail_page_alloc_debugfs
);
1349 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1351 #else /* CONFIG_FAIL_PAGE_ALLOC */
1353 static inline int should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1358 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1361 * Return 1 if free pages are above 'mark'. This takes into account the order
1362 * of the allocation.
1364 int zone_watermark_ok(struct zone
*z
, int order
, unsigned long mark
,
1365 int classzone_idx
, int alloc_flags
)
1367 /* free_pages my go negative - that's OK */
1369 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
) - (1 << order
) + 1;
1372 if (alloc_flags
& ALLOC_HIGH
)
1374 if (alloc_flags
& ALLOC_HARDER
)
1377 if (free_pages
<= min
+ z
->lowmem_reserve
[classzone_idx
])
1379 for (o
= 0; o
< order
; o
++) {
1380 /* At the next order, this order's pages become unavailable */
1381 free_pages
-= z
->free_area
[o
].nr_free
<< o
;
1383 /* Require fewer higher order pages to be free */
1386 if (free_pages
<= min
)
1394 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1395 * skip over zones that are not allowed by the cpuset, or that have
1396 * been recently (in last second) found to be nearly full. See further
1397 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1398 * that have to skip over a lot of full or unallowed zones.
1400 * If the zonelist cache is present in the passed in zonelist, then
1401 * returns a pointer to the allowed node mask (either the current
1402 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1404 * If the zonelist cache is not available for this zonelist, does
1405 * nothing and returns NULL.
1407 * If the fullzones BITMAP in the zonelist cache is stale (more than
1408 * a second since last zap'd) then we zap it out (clear its bits.)
1410 * We hold off even calling zlc_setup, until after we've checked the
1411 * first zone in the zonelist, on the theory that most allocations will
1412 * be satisfied from that first zone, so best to examine that zone as
1413 * quickly as we can.
1415 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1417 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1418 nodemask_t
*allowednodes
; /* zonelist_cache approximation */
1420 zlc
= zonelist
->zlcache_ptr
;
1424 if (time_after(jiffies
, zlc
->last_full_zap
+ HZ
)) {
1425 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
1426 zlc
->last_full_zap
= jiffies
;
1429 allowednodes
= !in_interrupt() && (alloc_flags
& ALLOC_CPUSET
) ?
1430 &cpuset_current_mems_allowed
:
1431 &node_states
[N_HIGH_MEMORY
];
1432 return allowednodes
;
1436 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1437 * if it is worth looking at further for free memory:
1438 * 1) Check that the zone isn't thought to be full (doesn't have its
1439 * bit set in the zonelist_cache fullzones BITMAP).
1440 * 2) Check that the zones node (obtained from the zonelist_cache
1441 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1442 * Return true (non-zero) if zone is worth looking at further, or
1443 * else return false (zero) if it is not.
1445 * This check -ignores- the distinction between various watermarks,
1446 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1447 * found to be full for any variation of these watermarks, it will
1448 * be considered full for up to one second by all requests, unless
1449 * we are so low on memory on all allowed nodes that we are forced
1450 * into the second scan of the zonelist.
1452 * In the second scan we ignore this zonelist cache and exactly
1453 * apply the watermarks to all zones, even it is slower to do so.
1454 * We are low on memory in the second scan, and should leave no stone
1455 * unturned looking for a free page.
1457 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1458 nodemask_t
*allowednodes
)
1460 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1461 int i
; /* index of *z in zonelist zones */
1462 int n
; /* node that zone *z is on */
1464 zlc
= zonelist
->zlcache_ptr
;
1468 i
= z
- zonelist
->_zonerefs
;
1471 /* This zone is worth trying if it is allowed but not full */
1472 return node_isset(n
, *allowednodes
) && !test_bit(i
, zlc
->fullzones
);
1476 * Given 'z' scanning a zonelist, set the corresponding bit in
1477 * zlc->fullzones, so that subsequent attempts to allocate a page
1478 * from that zone don't waste time re-examining it.
1480 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1482 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1483 int i
; /* index of *z in zonelist zones */
1485 zlc
= zonelist
->zlcache_ptr
;
1489 i
= z
- zonelist
->_zonerefs
;
1491 set_bit(i
, zlc
->fullzones
);
1494 #else /* CONFIG_NUMA */
1496 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1501 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1502 nodemask_t
*allowednodes
)
1507 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1510 #endif /* CONFIG_NUMA */
1513 * get_page_from_freelist goes through the zonelist trying to allocate
1516 static struct page
*
1517 get_page_from_freelist(gfp_t gfp_mask
, nodemask_t
*nodemask
, unsigned int order
,
1518 struct zonelist
*zonelist
, int high_zoneidx
, int alloc_flags
,
1519 struct zone
*preferred_zone
, int migratetype
)
1522 struct page
*page
= NULL
;
1525 nodemask_t
*allowednodes
= NULL
;/* zonelist_cache approximation */
1526 int zlc_active
= 0; /* set if using zonelist_cache */
1527 int did_zlc_setup
= 0; /* just call zlc_setup() one time */
1529 classzone_idx
= zone_idx(preferred_zone
);
1532 * Scan zonelist, looking for a zone with enough free.
1533 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1535 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
1536 high_zoneidx
, nodemask
) {
1537 if (NUMA_BUILD
&& zlc_active
&&
1538 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
1540 if ((alloc_flags
& ALLOC_CPUSET
) &&
1541 !cpuset_zone_allowed_softwall(zone
, gfp_mask
))
1544 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
1545 if (!(alloc_flags
& ALLOC_NO_WATERMARKS
)) {
1549 mark
= zone
->watermark
[alloc_flags
& ALLOC_WMARK_MASK
];
1550 if (zone_watermark_ok(zone
, order
, mark
,
1551 classzone_idx
, alloc_flags
))
1554 if (zone_reclaim_mode
== 0)
1555 goto this_zone_full
;
1557 ret
= zone_reclaim(zone
, gfp_mask
, order
);
1559 case ZONE_RECLAIM_NOSCAN
:
1562 case ZONE_RECLAIM_FULL
:
1563 /* scanned but unreclaimable */
1564 goto this_zone_full
;
1566 /* did we reclaim enough */
1567 if (!zone_watermark_ok(zone
, order
, mark
,
1568 classzone_idx
, alloc_flags
))
1569 goto this_zone_full
;
1574 page
= buffered_rmqueue(preferred_zone
, zone
, order
,
1575 gfp_mask
, migratetype
);
1580 zlc_mark_zone_full(zonelist
, z
);
1582 if (NUMA_BUILD
&& !did_zlc_setup
&& nr_online_nodes
> 1) {
1584 * we do zlc_setup after the first zone is tried but only
1585 * if there are multiple nodes make it worthwhile
1587 allowednodes
= zlc_setup(zonelist
, alloc_flags
);
1593 if (unlikely(NUMA_BUILD
&& page
== NULL
&& zlc_active
)) {
1594 /* Disable zlc cache for second zonelist scan */
1602 should_alloc_retry(gfp_t gfp_mask
, unsigned int order
,
1603 unsigned long pages_reclaimed
)
1605 /* Do not loop if specifically requested */
1606 if (gfp_mask
& __GFP_NORETRY
)
1610 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
1611 * means __GFP_NOFAIL, but that may not be true in other
1614 if (order
<= PAGE_ALLOC_COSTLY_ORDER
)
1618 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
1619 * specified, then we retry until we no longer reclaim any pages
1620 * (above), or we've reclaimed an order of pages at least as
1621 * large as the allocation's order. In both cases, if the
1622 * allocation still fails, we stop retrying.
1624 if (gfp_mask
& __GFP_REPEAT
&& pages_reclaimed
< (1 << order
))
1628 * Don't let big-order allocations loop unless the caller
1629 * explicitly requests that.
1631 if (gfp_mask
& __GFP_NOFAIL
)
1637 static inline struct page
*
1638 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
1639 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1640 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
1645 /* Acquire the OOM killer lock for the zones in zonelist */
1646 if (!try_set_zone_oom(zonelist
, gfp_mask
)) {
1647 schedule_timeout_uninterruptible(1);
1652 * Go through the zonelist yet one more time, keep very high watermark
1653 * here, this is only to catch a parallel oom killing, we must fail if
1654 * we're still under heavy pressure.
1656 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
,
1657 order
, zonelist
, high_zoneidx
,
1658 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
,
1659 preferred_zone
, migratetype
);
1663 /* The OOM killer will not help higher order allocs */
1664 if (order
> PAGE_ALLOC_COSTLY_ORDER
&& !(gfp_mask
& __GFP_NOFAIL
))
1667 /* Exhausted what can be done so it's blamo time */
1668 out_of_memory(zonelist
, gfp_mask
, order
);
1671 clear_zonelist_oom(zonelist
, gfp_mask
);
1675 /* The really slow allocator path where we enter direct reclaim */
1676 static inline struct page
*
1677 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
1678 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1679 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
1680 int migratetype
, unsigned long *did_some_progress
)
1682 struct page
*page
= NULL
;
1683 struct reclaim_state reclaim_state
;
1684 struct task_struct
*p
= current
;
1688 /* We now go into synchronous reclaim */
1689 cpuset_memory_pressure_bump();
1690 p
->flags
|= PF_MEMALLOC
;
1691 lockdep_set_current_reclaim_state(gfp_mask
);
1692 reclaim_state
.reclaimed_slab
= 0;
1693 p
->reclaim_state
= &reclaim_state
;
1695 *did_some_progress
= try_to_free_pages(zonelist
, order
, gfp_mask
, nodemask
);
1697 p
->reclaim_state
= NULL
;
1698 lockdep_clear_current_reclaim_state();
1699 p
->flags
&= ~PF_MEMALLOC
;
1706 if (likely(*did_some_progress
))
1707 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
1708 zonelist
, high_zoneidx
,
1709 alloc_flags
, preferred_zone
,
1715 * This is called in the allocator slow-path if the allocation request is of
1716 * sufficient urgency to ignore watermarks and take other desperate measures
1718 static inline struct page
*
1719 __alloc_pages_high_priority(gfp_t gfp_mask
, unsigned int order
,
1720 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1721 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
1727 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
1728 zonelist
, high_zoneidx
, ALLOC_NO_WATERMARKS
,
1729 preferred_zone
, migratetype
);
1731 if (!page
&& gfp_mask
& __GFP_NOFAIL
)
1732 congestion_wait(BLK_RW_ASYNC
, HZ
/50);
1733 } while (!page
&& (gfp_mask
& __GFP_NOFAIL
));
1739 void wake_all_kswapd(unsigned int order
, struct zonelist
*zonelist
,
1740 enum zone_type high_zoneidx
)
1745 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
)
1746 wakeup_kswapd(zone
, order
);
1750 gfp_to_alloc_flags(gfp_t gfp_mask
)
1752 struct task_struct
*p
= current
;
1753 int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
1754 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
1756 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
1757 BUILD_BUG_ON(__GFP_HIGH
!= ALLOC_HIGH
);
1760 * The caller may dip into page reserves a bit more if the caller
1761 * cannot run direct reclaim, or if the caller has realtime scheduling
1762 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
1763 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
1765 alloc_flags
|= (gfp_mask
& __GFP_HIGH
);
1768 alloc_flags
|= ALLOC_HARDER
;
1770 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
1771 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1773 alloc_flags
&= ~ALLOC_CPUSET
;
1774 } else if (unlikely(rt_task(p
)) && !in_interrupt())
1775 alloc_flags
|= ALLOC_HARDER
;
1777 if (likely(!(gfp_mask
& __GFP_NOMEMALLOC
))) {
1778 if (!in_interrupt() &&
1779 ((p
->flags
& PF_MEMALLOC
) ||
1780 unlikely(test_thread_flag(TIF_MEMDIE
))))
1781 alloc_flags
|= ALLOC_NO_WATERMARKS
;
1787 static inline struct page
*
1788 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
1789 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1790 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
1793 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
1794 struct page
*page
= NULL
;
1796 unsigned long pages_reclaimed
= 0;
1797 unsigned long did_some_progress
;
1798 struct task_struct
*p
= current
;
1801 * In the slowpath, we sanity check order to avoid ever trying to
1802 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
1803 * be using allocators in order of preference for an area that is
1806 if (order
>= MAX_ORDER
) {
1807 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
1812 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
1813 * __GFP_NOWARN set) should not cause reclaim since the subsystem
1814 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
1815 * using a larger set of nodes after it has established that the
1816 * allowed per node queues are empty and that nodes are
1819 if (NUMA_BUILD
&& (gfp_mask
& GFP_THISNODE
) == GFP_THISNODE
)
1823 wake_all_kswapd(order
, zonelist
, high_zoneidx
);
1826 * OK, we're below the kswapd watermark and have kicked background
1827 * reclaim. Now things get more complex, so set up alloc_flags according
1828 * to how we want to proceed.
1830 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
1832 /* This is the last chance, in general, before the goto nopage. */
1833 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
, zonelist
,
1834 high_zoneidx
, alloc_flags
& ~ALLOC_NO_WATERMARKS
,
1835 preferred_zone
, migratetype
);
1840 /* Allocate without watermarks if the context allows */
1841 if (alloc_flags
& ALLOC_NO_WATERMARKS
) {
1842 page
= __alloc_pages_high_priority(gfp_mask
, order
,
1843 zonelist
, high_zoneidx
, nodemask
,
1844 preferred_zone
, migratetype
);
1849 /* Atomic allocations - we can't balance anything */
1853 /* Avoid recursion of direct reclaim */
1854 if (p
->flags
& PF_MEMALLOC
)
1857 /* Avoid allocations with no watermarks from looping endlessly */
1858 if (test_thread_flag(TIF_MEMDIE
) && !(gfp_mask
& __GFP_NOFAIL
))
1861 /* Try direct reclaim and then allocating */
1862 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
,
1863 zonelist
, high_zoneidx
,
1865 alloc_flags
, preferred_zone
,
1866 migratetype
, &did_some_progress
);
1871 * If we failed to make any progress reclaiming, then we are
1872 * running out of options and have to consider going OOM
1874 if (!did_some_progress
) {
1875 if ((gfp_mask
& __GFP_FS
) && !(gfp_mask
& __GFP_NORETRY
)) {
1876 if (oom_killer_disabled
)
1878 page
= __alloc_pages_may_oom(gfp_mask
, order
,
1879 zonelist
, high_zoneidx
,
1880 nodemask
, preferred_zone
,
1886 * The OOM killer does not trigger for high-order
1887 * ~__GFP_NOFAIL allocations so if no progress is being
1888 * made, there are no other options and retrying is
1891 if (order
> PAGE_ALLOC_COSTLY_ORDER
&&
1892 !(gfp_mask
& __GFP_NOFAIL
))
1899 /* Check if we should retry the allocation */
1900 pages_reclaimed
+= did_some_progress
;
1901 if (should_alloc_retry(gfp_mask
, order
, pages_reclaimed
)) {
1902 /* Wait for some write requests to complete then retry */
1903 congestion_wait(BLK_RW_ASYNC
, HZ
/50);
1908 if (!(gfp_mask
& __GFP_NOWARN
) && printk_ratelimit()) {
1909 printk(KERN_WARNING
"%s: page allocation failure."
1910 " order:%d, mode:0x%x\n",
1911 p
->comm
, order
, gfp_mask
);
1917 if (kmemcheck_enabled
)
1918 kmemcheck_pagealloc_alloc(page
, order
, gfp_mask
);
1924 * This is the 'heart' of the zoned buddy allocator.
1927 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
,
1928 struct zonelist
*zonelist
, nodemask_t
*nodemask
)
1930 enum zone_type high_zoneidx
= gfp_zone(gfp_mask
);
1931 struct zone
*preferred_zone
;
1933 int migratetype
= allocflags_to_migratetype(gfp_mask
);
1935 gfp_mask
&= gfp_allowed_mask
;
1937 lockdep_trace_alloc(gfp_mask
);
1939 might_sleep_if(gfp_mask
& __GFP_WAIT
);
1941 if (should_fail_alloc_page(gfp_mask
, order
))
1945 * Check the zones suitable for the gfp_mask contain at least one
1946 * valid zone. It's possible to have an empty zonelist as a result
1947 * of GFP_THISNODE and a memoryless node
1949 if (unlikely(!zonelist
->_zonerefs
->zone
))
1952 /* The preferred zone is used for statistics later */
1953 first_zones_zonelist(zonelist
, high_zoneidx
, nodemask
, &preferred_zone
);
1954 if (!preferred_zone
)
1957 /* First allocation attempt */
1958 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
, order
,
1959 zonelist
, high_zoneidx
, ALLOC_WMARK_LOW
|ALLOC_CPUSET
,
1960 preferred_zone
, migratetype
);
1961 if (unlikely(!page
))
1962 page
= __alloc_pages_slowpath(gfp_mask
, order
,
1963 zonelist
, high_zoneidx
, nodemask
,
1964 preferred_zone
, migratetype
);
1966 trace_mm_page_alloc(page
, order
, gfp_mask
, migratetype
);
1969 EXPORT_SYMBOL(__alloc_pages_nodemask
);
1972 * Common helper functions.
1974 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
1979 * __get_free_pages() returns a 32-bit address, which cannot represent
1982 VM_BUG_ON((gfp_mask
& __GFP_HIGHMEM
) != 0);
1984 page
= alloc_pages(gfp_mask
, order
);
1987 return (unsigned long) page_address(page
);
1989 EXPORT_SYMBOL(__get_free_pages
);
1991 unsigned long get_zeroed_page(gfp_t gfp_mask
)
1993 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
1995 EXPORT_SYMBOL(get_zeroed_page
);
1997 void __pagevec_free(struct pagevec
*pvec
)
1999 int i
= pagevec_count(pvec
);
2002 trace_mm_pagevec_free(pvec
->pages
[i
], pvec
->cold
);
2003 free_hot_cold_page(pvec
->pages
[i
], pvec
->cold
);
2007 void __free_pages(struct page
*page
, unsigned int order
)
2009 if (put_page_testzero(page
)) {
2010 trace_mm_page_free_direct(page
, order
);
2012 free_hot_page(page
);
2014 __free_pages_ok(page
, order
);
2018 EXPORT_SYMBOL(__free_pages
);
2020 void free_pages(unsigned long addr
, unsigned int order
)
2023 VM_BUG_ON(!virt_addr_valid((void *)addr
));
2024 __free_pages(virt_to_page((void *)addr
), order
);
2028 EXPORT_SYMBOL(free_pages
);
2031 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2032 * @size: the number of bytes to allocate
2033 * @gfp_mask: GFP flags for the allocation
2035 * This function is similar to alloc_pages(), except that it allocates the
2036 * minimum number of pages to satisfy the request. alloc_pages() can only
2037 * allocate memory in power-of-two pages.
2039 * This function is also limited by MAX_ORDER.
2041 * Memory allocated by this function must be released by free_pages_exact().
2043 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
2045 unsigned int order
= get_order(size
);
2048 addr
= __get_free_pages(gfp_mask
, order
);
2050 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
2051 unsigned long used
= addr
+ PAGE_ALIGN(size
);
2053 split_page(virt_to_page((void *)addr
), order
);
2054 while (used
< alloc_end
) {
2060 return (void *)addr
;
2062 EXPORT_SYMBOL(alloc_pages_exact
);
2065 * free_pages_exact - release memory allocated via alloc_pages_exact()
2066 * @virt: the value returned by alloc_pages_exact.
2067 * @size: size of allocation, same value as passed to alloc_pages_exact().
2069 * Release the memory allocated by a previous call to alloc_pages_exact.
2071 void free_pages_exact(void *virt
, size_t size
)
2073 unsigned long addr
= (unsigned long)virt
;
2074 unsigned long end
= addr
+ PAGE_ALIGN(size
);
2076 while (addr
< end
) {
2081 EXPORT_SYMBOL(free_pages_exact
);
2083 static unsigned int nr_free_zone_pages(int offset
)
2088 /* Just pick one node, since fallback list is circular */
2089 unsigned int sum
= 0;
2091 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
2093 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
2094 unsigned long size
= zone
->present_pages
;
2095 unsigned long high
= high_wmark_pages(zone
);
2104 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2106 unsigned int nr_free_buffer_pages(void)
2108 return nr_free_zone_pages(gfp_zone(GFP_USER
));
2110 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
2113 * Amount of free RAM allocatable within all zones
2115 unsigned int nr_free_pagecache_pages(void)
2117 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
2120 static inline void show_node(struct zone
*zone
)
2123 printk("Node %d ", zone_to_nid(zone
));
2126 void si_meminfo(struct sysinfo
*val
)
2128 val
->totalram
= totalram_pages
;
2130 val
->freeram
= global_page_state(NR_FREE_PAGES
);
2131 val
->bufferram
= nr_blockdev_pages();
2132 val
->totalhigh
= totalhigh_pages
;
2133 val
->freehigh
= nr_free_highpages();
2134 val
->mem_unit
= PAGE_SIZE
;
2137 EXPORT_SYMBOL(si_meminfo
);
2140 void si_meminfo_node(struct sysinfo
*val
, int nid
)
2142 pg_data_t
*pgdat
= NODE_DATA(nid
);
2144 val
->totalram
= pgdat
->node_present_pages
;
2145 val
->freeram
= node_page_state(nid
, NR_FREE_PAGES
);
2146 #ifdef CONFIG_HIGHMEM
2147 val
->totalhigh
= pgdat
->node_zones
[ZONE_HIGHMEM
].present_pages
;
2148 val
->freehigh
= zone_page_state(&pgdat
->node_zones
[ZONE_HIGHMEM
],
2154 val
->mem_unit
= PAGE_SIZE
;
2158 #define K(x) ((x) << (PAGE_SHIFT-10))
2161 * Show free area list (used inside shift_scroll-lock stuff)
2162 * We also calculate the percentage fragmentation. We do this by counting the
2163 * memory on each free list with the exception of the first item on the list.
2165 void show_free_areas(void)
2170 for_each_populated_zone(zone
) {
2172 printk("%s per-cpu:\n", zone
->name
);
2174 for_each_online_cpu(cpu
) {
2175 struct per_cpu_pageset
*pageset
;
2177 pageset
= zone_pcp(zone
, cpu
);
2179 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2180 cpu
, pageset
->pcp
.high
,
2181 pageset
->pcp
.batch
, pageset
->pcp
.count
);
2185 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2186 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2188 " dirty:%lu writeback:%lu unstable:%lu\n"
2189 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2190 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n",
2191 global_page_state(NR_ACTIVE_ANON
),
2192 global_page_state(NR_INACTIVE_ANON
),
2193 global_page_state(NR_ISOLATED_ANON
),
2194 global_page_state(NR_ACTIVE_FILE
),
2195 global_page_state(NR_INACTIVE_FILE
),
2196 global_page_state(NR_ISOLATED_FILE
),
2197 global_page_state(NR_UNEVICTABLE
),
2198 global_page_state(NR_FILE_DIRTY
),
2199 global_page_state(NR_WRITEBACK
),
2200 global_page_state(NR_UNSTABLE_NFS
),
2201 global_page_state(NR_FREE_PAGES
),
2202 global_page_state(NR_SLAB_RECLAIMABLE
),
2203 global_page_state(NR_SLAB_UNRECLAIMABLE
),
2204 global_page_state(NR_FILE_MAPPED
),
2205 global_page_state(NR_SHMEM
),
2206 global_page_state(NR_PAGETABLE
),
2207 global_page_state(NR_BOUNCE
));
2209 for_each_populated_zone(zone
) {
2218 " active_anon:%lukB"
2219 " inactive_anon:%lukB"
2220 " active_file:%lukB"
2221 " inactive_file:%lukB"
2222 " unevictable:%lukB"
2223 " isolated(anon):%lukB"
2224 " isolated(file):%lukB"
2231 " slab_reclaimable:%lukB"
2232 " slab_unreclaimable:%lukB"
2233 " kernel_stack:%lukB"
2237 " writeback_tmp:%lukB"
2238 " pages_scanned:%lu"
2239 " all_unreclaimable? %s"
2242 K(zone_page_state(zone
, NR_FREE_PAGES
)),
2243 K(min_wmark_pages(zone
)),
2244 K(low_wmark_pages(zone
)),
2245 K(high_wmark_pages(zone
)),
2246 K(zone_page_state(zone
, NR_ACTIVE_ANON
)),
2247 K(zone_page_state(zone
, NR_INACTIVE_ANON
)),
2248 K(zone_page_state(zone
, NR_ACTIVE_FILE
)),
2249 K(zone_page_state(zone
, NR_INACTIVE_FILE
)),
2250 K(zone_page_state(zone
, NR_UNEVICTABLE
)),
2251 K(zone_page_state(zone
, NR_ISOLATED_ANON
)),
2252 K(zone_page_state(zone
, NR_ISOLATED_FILE
)),
2253 K(zone
->present_pages
),
2254 K(zone_page_state(zone
, NR_MLOCK
)),
2255 K(zone_page_state(zone
, NR_FILE_DIRTY
)),
2256 K(zone_page_state(zone
, NR_WRITEBACK
)),
2257 K(zone_page_state(zone
, NR_FILE_MAPPED
)),
2258 K(zone_page_state(zone
, NR_SHMEM
)),
2259 K(zone_page_state(zone
, NR_SLAB_RECLAIMABLE
)),
2260 K(zone_page_state(zone
, NR_SLAB_UNRECLAIMABLE
)),
2261 zone_page_state(zone
, NR_KERNEL_STACK
) *
2263 K(zone_page_state(zone
, NR_PAGETABLE
)),
2264 K(zone_page_state(zone
, NR_UNSTABLE_NFS
)),
2265 K(zone_page_state(zone
, NR_BOUNCE
)),
2266 K(zone_page_state(zone
, NR_WRITEBACK_TEMP
)),
2267 zone
->pages_scanned
,
2268 (zone_is_all_unreclaimable(zone
) ? "yes" : "no")
2270 printk("lowmem_reserve[]:");
2271 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
2272 printk(" %lu", zone
->lowmem_reserve
[i
]);
2276 for_each_populated_zone(zone
) {
2277 unsigned long nr
[MAX_ORDER
], flags
, order
, total
= 0;
2280 printk("%s: ", zone
->name
);
2282 spin_lock_irqsave(&zone
->lock
, flags
);
2283 for (order
= 0; order
< MAX_ORDER
; order
++) {
2284 nr
[order
] = zone
->free_area
[order
].nr_free
;
2285 total
+= nr
[order
] << order
;
2287 spin_unlock_irqrestore(&zone
->lock
, flags
);
2288 for (order
= 0; order
< MAX_ORDER
; order
++)
2289 printk("%lu*%lukB ", nr
[order
], K(1UL) << order
);
2290 printk("= %lukB\n", K(total
));
2293 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES
));
2295 show_swap_cache_info();
2298 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
2300 zoneref
->zone
= zone
;
2301 zoneref
->zone_idx
= zone_idx(zone
);
2305 * Builds allocation fallback zone lists.
2307 * Add all populated zones of a node to the zonelist.
2309 static int build_zonelists_node(pg_data_t
*pgdat
, struct zonelist
*zonelist
,
2310 int nr_zones
, enum zone_type zone_type
)
2314 BUG_ON(zone_type
>= MAX_NR_ZONES
);
2319 zone
= pgdat
->node_zones
+ zone_type
;
2320 if (populated_zone(zone
)) {
2321 zoneref_set_zone(zone
,
2322 &zonelist
->_zonerefs
[nr_zones
++]);
2323 check_highest_zone(zone_type
);
2326 } while (zone_type
);
2333 * 0 = automatic detection of better ordering.
2334 * 1 = order by ([node] distance, -zonetype)
2335 * 2 = order by (-zonetype, [node] distance)
2337 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
2338 * the same zonelist. So only NUMA can configure this param.
2340 #define ZONELIST_ORDER_DEFAULT 0
2341 #define ZONELIST_ORDER_NODE 1
2342 #define ZONELIST_ORDER_ZONE 2
2344 /* zonelist order in the kernel.
2345 * set_zonelist_order() will set this to NODE or ZONE.
2347 static int current_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
2348 static char zonelist_order_name
[3][8] = {"Default", "Node", "Zone"};
2352 /* The value user specified ....changed by config */
2353 static int user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
2354 /* string for sysctl */
2355 #define NUMA_ZONELIST_ORDER_LEN 16
2356 char numa_zonelist_order
[16] = "default";
2359 * interface for configure zonelist ordering.
2360 * command line option "numa_zonelist_order"
2361 * = "[dD]efault - default, automatic configuration.
2362 * = "[nN]ode - order by node locality, then by zone within node
2363 * = "[zZ]one - order by zone, then by locality within zone
2366 static int __parse_numa_zonelist_order(char *s
)
2368 if (*s
== 'd' || *s
== 'D') {
2369 user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
2370 } else if (*s
== 'n' || *s
== 'N') {
2371 user_zonelist_order
= ZONELIST_ORDER_NODE
;
2372 } else if (*s
== 'z' || *s
== 'Z') {
2373 user_zonelist_order
= ZONELIST_ORDER_ZONE
;
2376 "Ignoring invalid numa_zonelist_order value: "
2383 static __init
int setup_numa_zonelist_order(char *s
)
2386 return __parse_numa_zonelist_order(s
);
2389 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
2392 * sysctl handler for numa_zonelist_order
2394 int numa_zonelist_order_handler(ctl_table
*table
, int write
,
2395 void __user
*buffer
, size_t *length
,
2398 char saved_string
[NUMA_ZONELIST_ORDER_LEN
];
2402 strncpy(saved_string
, (char*)table
->data
,
2403 NUMA_ZONELIST_ORDER_LEN
);
2404 ret
= proc_dostring(table
, write
, buffer
, length
, ppos
);
2408 int oldval
= user_zonelist_order
;
2409 if (__parse_numa_zonelist_order((char*)table
->data
)) {
2411 * bogus value. restore saved string
2413 strncpy((char*)table
->data
, saved_string
,
2414 NUMA_ZONELIST_ORDER_LEN
);
2415 user_zonelist_order
= oldval
;
2416 } else if (oldval
!= user_zonelist_order
)
2417 build_all_zonelists();
2423 #define MAX_NODE_LOAD (nr_online_nodes)
2424 static int node_load
[MAX_NUMNODES
];
2427 * find_next_best_node - find the next node that should appear in a given node's fallback list
2428 * @node: node whose fallback list we're appending
2429 * @used_node_mask: nodemask_t of already used nodes
2431 * We use a number of factors to determine which is the next node that should
2432 * appear on a given node's fallback list. The node should not have appeared
2433 * already in @node's fallback list, and it should be the next closest node
2434 * according to the distance array (which contains arbitrary distance values
2435 * from each node to each node in the system), and should also prefer nodes
2436 * with no CPUs, since presumably they'll have very little allocation pressure
2437 * on them otherwise.
2438 * It returns -1 if no node is found.
2440 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
2443 int min_val
= INT_MAX
;
2445 const struct cpumask
*tmp
= cpumask_of_node(0);
2447 /* Use the local node if we haven't already */
2448 if (!node_isset(node
, *used_node_mask
)) {
2449 node_set(node
, *used_node_mask
);
2453 for_each_node_state(n
, N_HIGH_MEMORY
) {
2455 /* Don't want a node to appear more than once */
2456 if (node_isset(n
, *used_node_mask
))
2459 /* Use the distance array to find the distance */
2460 val
= node_distance(node
, n
);
2462 /* Penalize nodes under us ("prefer the next node") */
2465 /* Give preference to headless and unused nodes */
2466 tmp
= cpumask_of_node(n
);
2467 if (!cpumask_empty(tmp
))
2468 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
2470 /* Slight preference for less loaded node */
2471 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
2472 val
+= node_load
[n
];
2474 if (val
< min_val
) {
2481 node_set(best_node
, *used_node_mask
);
2488 * Build zonelists ordered by node and zones within node.
2489 * This results in maximum locality--normal zone overflows into local
2490 * DMA zone, if any--but risks exhausting DMA zone.
2492 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int node
)
2495 struct zonelist
*zonelist
;
2497 zonelist
= &pgdat
->node_zonelists
[0];
2498 for (j
= 0; zonelist
->_zonerefs
[j
].zone
!= NULL
; j
++)
2500 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
2502 zonelist
->_zonerefs
[j
].zone
= NULL
;
2503 zonelist
->_zonerefs
[j
].zone_idx
= 0;
2507 * Build gfp_thisnode zonelists
2509 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
2512 struct zonelist
*zonelist
;
2514 zonelist
= &pgdat
->node_zonelists
[1];
2515 j
= build_zonelists_node(pgdat
, zonelist
, 0, MAX_NR_ZONES
- 1);
2516 zonelist
->_zonerefs
[j
].zone
= NULL
;
2517 zonelist
->_zonerefs
[j
].zone_idx
= 0;
2521 * Build zonelists ordered by zone and nodes within zones.
2522 * This results in conserving DMA zone[s] until all Normal memory is
2523 * exhausted, but results in overflowing to remote node while memory
2524 * may still exist in local DMA zone.
2526 static int node_order
[MAX_NUMNODES
];
2528 static void build_zonelists_in_zone_order(pg_data_t
*pgdat
, int nr_nodes
)
2531 int zone_type
; /* needs to be signed */
2533 struct zonelist
*zonelist
;
2535 zonelist
= &pgdat
->node_zonelists
[0];
2537 for (zone_type
= MAX_NR_ZONES
- 1; zone_type
>= 0; zone_type
--) {
2538 for (j
= 0; j
< nr_nodes
; j
++) {
2539 node
= node_order
[j
];
2540 z
= &NODE_DATA(node
)->node_zones
[zone_type
];
2541 if (populated_zone(z
)) {
2543 &zonelist
->_zonerefs
[pos
++]);
2544 check_highest_zone(zone_type
);
2548 zonelist
->_zonerefs
[pos
].zone
= NULL
;
2549 zonelist
->_zonerefs
[pos
].zone_idx
= 0;
2552 static int default_zonelist_order(void)
2555 unsigned long low_kmem_size
,total_size
;
2559 * ZONE_DMA and ZONE_DMA32 can be very small area in the sytem.
2560 * If they are really small and used heavily, the system can fall
2561 * into OOM very easily.
2562 * This function detect ZONE_DMA/DMA32 size and confgigures zone order.
2564 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
2567 for_each_online_node(nid
) {
2568 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
2569 z
= &NODE_DATA(nid
)->node_zones
[zone_type
];
2570 if (populated_zone(z
)) {
2571 if (zone_type
< ZONE_NORMAL
)
2572 low_kmem_size
+= z
->present_pages
;
2573 total_size
+= z
->present_pages
;
2577 if (!low_kmem_size
|| /* there are no DMA area. */
2578 low_kmem_size
> total_size
/2) /* DMA/DMA32 is big. */
2579 return ZONELIST_ORDER_NODE
;
2581 * look into each node's config.
2582 * If there is a node whose DMA/DMA32 memory is very big area on
2583 * local memory, NODE_ORDER may be suitable.
2585 average_size
= total_size
/
2586 (nodes_weight(node_states
[N_HIGH_MEMORY
]) + 1);
2587 for_each_online_node(nid
) {
2590 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
2591 z
= &NODE_DATA(nid
)->node_zones
[zone_type
];
2592 if (populated_zone(z
)) {
2593 if (zone_type
< ZONE_NORMAL
)
2594 low_kmem_size
+= z
->present_pages
;
2595 total_size
+= z
->present_pages
;
2598 if (low_kmem_size
&&
2599 total_size
> average_size
&& /* ignore small node */
2600 low_kmem_size
> total_size
* 70/100)
2601 return ZONELIST_ORDER_NODE
;
2603 return ZONELIST_ORDER_ZONE
;
2606 static void set_zonelist_order(void)
2608 if (user_zonelist_order
== ZONELIST_ORDER_DEFAULT
)
2609 current_zonelist_order
= default_zonelist_order();
2611 current_zonelist_order
= user_zonelist_order
;
2614 static void build_zonelists(pg_data_t
*pgdat
)
2618 nodemask_t used_mask
;
2619 int local_node
, prev_node
;
2620 struct zonelist
*zonelist
;
2621 int order
= current_zonelist_order
;
2623 /* initialize zonelists */
2624 for (i
= 0; i
< MAX_ZONELISTS
; i
++) {
2625 zonelist
= pgdat
->node_zonelists
+ i
;
2626 zonelist
->_zonerefs
[0].zone
= NULL
;
2627 zonelist
->_zonerefs
[0].zone_idx
= 0;
2630 /* NUMA-aware ordering of nodes */
2631 local_node
= pgdat
->node_id
;
2632 load
= nr_online_nodes
;
2633 prev_node
= local_node
;
2634 nodes_clear(used_mask
);
2636 memset(node_order
, 0, sizeof(node_order
));
2639 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
2640 int distance
= node_distance(local_node
, node
);
2643 * If another node is sufficiently far away then it is better
2644 * to reclaim pages in a zone before going off node.
2646 if (distance
> RECLAIM_DISTANCE
)
2647 zone_reclaim_mode
= 1;
2650 * We don't want to pressure a particular node.
2651 * So adding penalty to the first node in same
2652 * distance group to make it round-robin.
2654 if (distance
!= node_distance(local_node
, prev_node
))
2655 node_load
[node
] = load
;
2659 if (order
== ZONELIST_ORDER_NODE
)
2660 build_zonelists_in_node_order(pgdat
, node
);
2662 node_order
[j
++] = node
; /* remember order */
2665 if (order
== ZONELIST_ORDER_ZONE
) {
2666 /* calculate node order -- i.e., DMA last! */
2667 build_zonelists_in_zone_order(pgdat
, j
);
2670 build_thisnode_zonelists(pgdat
);
2673 /* Construct the zonelist performance cache - see further mmzone.h */
2674 static void build_zonelist_cache(pg_data_t
*pgdat
)
2676 struct zonelist
*zonelist
;
2677 struct zonelist_cache
*zlc
;
2680 zonelist
= &pgdat
->node_zonelists
[0];
2681 zonelist
->zlcache_ptr
= zlc
= &zonelist
->zlcache
;
2682 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
2683 for (z
= zonelist
->_zonerefs
; z
->zone
; z
++)
2684 zlc
->z_to_n
[z
- zonelist
->_zonerefs
] = zonelist_node_idx(z
);
2688 #else /* CONFIG_NUMA */
2690 static void set_zonelist_order(void)
2692 current_zonelist_order
= ZONELIST_ORDER_ZONE
;
2695 static void build_zonelists(pg_data_t
*pgdat
)
2697 int node
, local_node
;
2699 struct zonelist
*zonelist
;
2701 local_node
= pgdat
->node_id
;
2703 zonelist
= &pgdat
->node_zonelists
[0];
2704 j
= build_zonelists_node(pgdat
, zonelist
, 0, MAX_NR_ZONES
- 1);
2707 * Now we build the zonelist so that it contains the zones
2708 * of all the other nodes.
2709 * We don't want to pressure a particular node, so when
2710 * building the zones for node N, we make sure that the
2711 * zones coming right after the local ones are those from
2712 * node N+1 (modulo N)
2714 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
2715 if (!node_online(node
))
2717 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
2720 for (node
= 0; node
< local_node
; node
++) {
2721 if (!node_online(node
))
2723 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
2727 zonelist
->_zonerefs
[j
].zone
= NULL
;
2728 zonelist
->_zonerefs
[j
].zone_idx
= 0;
2731 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
2732 static void build_zonelist_cache(pg_data_t
*pgdat
)
2734 pgdat
->node_zonelists
[0].zlcache_ptr
= NULL
;
2737 #endif /* CONFIG_NUMA */
2739 /* return values int ....just for stop_machine() */
2740 static int __build_all_zonelists(void *dummy
)
2745 memset(node_load
, 0, sizeof(node_load
));
2747 for_each_online_node(nid
) {
2748 pg_data_t
*pgdat
= NODE_DATA(nid
);
2750 build_zonelists(pgdat
);
2751 build_zonelist_cache(pgdat
);
2756 void build_all_zonelists(void)
2758 set_zonelist_order();
2760 if (system_state
== SYSTEM_BOOTING
) {
2761 __build_all_zonelists(NULL
);
2762 mminit_verify_zonelist();
2763 cpuset_init_current_mems_allowed();
2765 /* we have to stop all cpus to guarantee there is no user
2767 stop_machine(__build_all_zonelists
, NULL
, NULL
);
2768 /* cpuset refresh routine should be here */
2770 vm_total_pages
= nr_free_pagecache_pages();
2772 * Disable grouping by mobility if the number of pages in the
2773 * system is too low to allow the mechanism to work. It would be
2774 * more accurate, but expensive to check per-zone. This check is
2775 * made on memory-hotadd so a system can start with mobility
2776 * disabled and enable it later
2778 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
2779 page_group_by_mobility_disabled
= 1;
2781 page_group_by_mobility_disabled
= 0;
2783 printk("Built %i zonelists in %s order, mobility grouping %s. "
2784 "Total pages: %ld\n",
2786 zonelist_order_name
[current_zonelist_order
],
2787 page_group_by_mobility_disabled
? "off" : "on",
2790 printk("Policy zone: %s\n", zone_names
[policy_zone
]);
2795 * Helper functions to size the waitqueue hash table.
2796 * Essentially these want to choose hash table sizes sufficiently
2797 * large so that collisions trying to wait on pages are rare.
2798 * But in fact, the number of active page waitqueues on typical
2799 * systems is ridiculously low, less than 200. So this is even
2800 * conservative, even though it seems large.
2802 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
2803 * waitqueues, i.e. the size of the waitq table given the number of pages.
2805 #define PAGES_PER_WAITQUEUE 256
2807 #ifndef CONFIG_MEMORY_HOTPLUG
2808 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
2810 unsigned long size
= 1;
2812 pages
/= PAGES_PER_WAITQUEUE
;
2814 while (size
< pages
)
2818 * Once we have dozens or even hundreds of threads sleeping
2819 * on IO we've got bigger problems than wait queue collision.
2820 * Limit the size of the wait table to a reasonable size.
2822 size
= min(size
, 4096UL);
2824 return max(size
, 4UL);
2828 * A zone's size might be changed by hot-add, so it is not possible to determine
2829 * a suitable size for its wait_table. So we use the maximum size now.
2831 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
2833 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
2834 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
2835 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
2837 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
2838 * or more by the traditional way. (See above). It equals:
2840 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
2841 * ia64(16K page size) : = ( 8G + 4M)byte.
2842 * powerpc (64K page size) : = (32G +16M)byte.
2844 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
2851 * This is an integer logarithm so that shifts can be used later
2852 * to extract the more random high bits from the multiplicative
2853 * hash function before the remainder is taken.
2855 static inline unsigned long wait_table_bits(unsigned long size
)
2860 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
2863 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
2864 * of blocks reserved is based on min_wmark_pages(zone). The memory within
2865 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
2866 * higher will lead to a bigger reserve which will get freed as contiguous
2867 * blocks as reclaim kicks in
2869 static void setup_zone_migrate_reserve(struct zone
*zone
)
2871 unsigned long start_pfn
, pfn
, end_pfn
;
2873 unsigned long block_migratetype
;
2876 /* Get the start pfn, end pfn and the number of blocks to reserve */
2877 start_pfn
= zone
->zone_start_pfn
;
2878 end_pfn
= start_pfn
+ zone
->spanned_pages
;
2879 reserve
= roundup(min_wmark_pages(zone
), pageblock_nr_pages
) >>
2883 * Reserve blocks are generally in place to help high-order atomic
2884 * allocations that are short-lived. A min_free_kbytes value that
2885 * would result in more than 2 reserve blocks for atomic allocations
2886 * is assumed to be in place to help anti-fragmentation for the
2887 * future allocation of hugepages at runtime.
2889 reserve
= min(2, reserve
);
2891 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= pageblock_nr_pages
) {
2892 if (!pfn_valid(pfn
))
2894 page
= pfn_to_page(pfn
);
2896 /* Watch out for overlapping nodes */
2897 if (page_to_nid(page
) != zone_to_nid(zone
))
2900 /* Blocks with reserved pages will never free, skip them. */
2901 if (PageReserved(page
))
2904 block_migratetype
= get_pageblock_migratetype(page
);
2906 /* If this block is reserved, account for it */
2907 if (reserve
> 0 && block_migratetype
== MIGRATE_RESERVE
) {
2912 /* Suitable for reserving if this block is movable */
2913 if (reserve
> 0 && block_migratetype
== MIGRATE_MOVABLE
) {
2914 set_pageblock_migratetype(page
, MIGRATE_RESERVE
);
2915 move_freepages_block(zone
, page
, MIGRATE_RESERVE
);
2921 * If the reserve is met and this is a previous reserved block,
2924 if (block_migratetype
== MIGRATE_RESERVE
) {
2925 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
2926 move_freepages_block(zone
, page
, MIGRATE_MOVABLE
);
2932 * Initially all pages are reserved - free ones are freed
2933 * up by free_all_bootmem() once the early boot process is
2934 * done. Non-atomic initialization, single-pass.
2936 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
2937 unsigned long start_pfn
, enum memmap_context context
)
2940 unsigned long end_pfn
= start_pfn
+ size
;
2944 if (highest_memmap_pfn
< end_pfn
- 1)
2945 highest_memmap_pfn
= end_pfn
- 1;
2947 z
= &NODE_DATA(nid
)->node_zones
[zone
];
2948 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
2950 * There can be holes in boot-time mem_map[]s
2951 * handed to this function. They do not
2952 * exist on hotplugged memory.
2954 if (context
== MEMMAP_EARLY
) {
2955 if (!early_pfn_valid(pfn
))
2957 if (!early_pfn_in_nid(pfn
, nid
))
2960 page
= pfn_to_page(pfn
);
2961 set_page_links(page
, zone
, nid
, pfn
);
2962 mminit_verify_page_links(page
, zone
, nid
, pfn
);
2963 init_page_count(page
);
2964 reset_page_mapcount(page
);
2965 SetPageReserved(page
);
2967 * Mark the block movable so that blocks are reserved for
2968 * movable at startup. This will force kernel allocations
2969 * to reserve their blocks rather than leaking throughout
2970 * the address space during boot when many long-lived
2971 * kernel allocations are made. Later some blocks near
2972 * the start are marked MIGRATE_RESERVE by
2973 * setup_zone_migrate_reserve()
2975 * bitmap is created for zone's valid pfn range. but memmap
2976 * can be created for invalid pages (for alignment)
2977 * check here not to call set_pageblock_migratetype() against
2980 if ((z
->zone_start_pfn
<= pfn
)
2981 && (pfn
< z
->zone_start_pfn
+ z
->spanned_pages
)
2982 && !(pfn
& (pageblock_nr_pages
- 1)))
2983 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
2985 INIT_LIST_HEAD(&page
->lru
);
2986 #ifdef WANT_PAGE_VIRTUAL
2987 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
2988 if (!is_highmem_idx(zone
))
2989 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
2994 static void __meminit
zone_init_free_lists(struct zone
*zone
)
2997 for_each_migratetype_order(order
, t
) {
2998 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
2999 zone
->free_area
[order
].nr_free
= 0;
3003 #ifndef __HAVE_ARCH_MEMMAP_INIT
3004 #define memmap_init(size, nid, zone, start_pfn) \
3005 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3008 static int zone_batchsize(struct zone
*zone
)
3014 * The per-cpu-pages pools are set to around 1000th of the
3015 * size of the zone. But no more than 1/2 of a meg.
3017 * OK, so we don't know how big the cache is. So guess.
3019 batch
= zone
->present_pages
/ 1024;
3020 if (batch
* PAGE_SIZE
> 512 * 1024)
3021 batch
= (512 * 1024) / PAGE_SIZE
;
3022 batch
/= 4; /* We effectively *= 4 below */
3027 * Clamp the batch to a 2^n - 1 value. Having a power
3028 * of 2 value was found to be more likely to have
3029 * suboptimal cache aliasing properties in some cases.
3031 * For example if 2 tasks are alternately allocating
3032 * batches of pages, one task can end up with a lot
3033 * of pages of one half of the possible page colors
3034 * and the other with pages of the other colors.
3036 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
3041 /* The deferral and batching of frees should be suppressed under NOMMU
3044 * The problem is that NOMMU needs to be able to allocate large chunks
3045 * of contiguous memory as there's no hardware page translation to
3046 * assemble apparent contiguous memory from discontiguous pages.
3048 * Queueing large contiguous runs of pages for batching, however,
3049 * causes the pages to actually be freed in smaller chunks. As there
3050 * can be a significant delay between the individual batches being
3051 * recycled, this leads to the once large chunks of space being
3052 * fragmented and becoming unavailable for high-order allocations.
3058 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
3060 struct per_cpu_pages
*pcp
;
3063 memset(p
, 0, sizeof(*p
));
3067 pcp
->high
= 6 * batch
;
3068 pcp
->batch
= max(1UL, 1 * batch
);
3069 for (migratetype
= 0; migratetype
< MIGRATE_PCPTYPES
; migratetype
++)
3070 INIT_LIST_HEAD(&pcp
->lists
[migratetype
]);
3074 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3075 * to the value high for the pageset p.
3078 static void setup_pagelist_highmark(struct per_cpu_pageset
*p
,
3081 struct per_cpu_pages
*pcp
;
3085 pcp
->batch
= max(1UL, high
/4);
3086 if ((high
/4) > (PAGE_SHIFT
* 8))
3087 pcp
->batch
= PAGE_SHIFT
* 8;
3093 * Boot pageset table. One per cpu which is going to be used for all
3094 * zones and all nodes. The parameters will be set in such a way
3095 * that an item put on a list will immediately be handed over to
3096 * the buddy list. This is safe since pageset manipulation is done
3097 * with interrupts disabled.
3099 * Some NUMA counter updates may also be caught by the boot pagesets.
3101 * The boot_pagesets must be kept even after bootup is complete for
3102 * unused processors and/or zones. They do play a role for bootstrapping
3103 * hotplugged processors.
3105 * zoneinfo_show() and maybe other functions do
3106 * not check if the processor is online before following the pageset pointer.
3107 * Other parts of the kernel may not check if the zone is available.
3109 static struct per_cpu_pageset boot_pageset
[NR_CPUS
];
3112 * Dynamically allocate memory for the
3113 * per cpu pageset array in struct zone.
3115 static int __cpuinit
process_zones(int cpu
)
3117 struct zone
*zone
, *dzone
;
3118 int node
= cpu_to_node(cpu
);
3120 node_set_state(node
, N_CPU
); /* this node has a cpu */
3122 for_each_populated_zone(zone
) {
3123 zone_pcp(zone
, cpu
) = kmalloc_node(sizeof(struct per_cpu_pageset
),
3125 if (!zone_pcp(zone
, cpu
))
3128 setup_pageset(zone_pcp(zone
, cpu
), zone_batchsize(zone
));
3130 if (percpu_pagelist_fraction
)
3131 setup_pagelist_highmark(zone_pcp(zone
, cpu
),
3132 (zone
->present_pages
/ percpu_pagelist_fraction
));
3137 for_each_zone(dzone
) {
3138 if (!populated_zone(dzone
))
3142 kfree(zone_pcp(dzone
, cpu
));
3143 zone_pcp(dzone
, cpu
) = &boot_pageset
[cpu
];
3148 static inline void free_zone_pagesets(int cpu
)
3152 for_each_zone(zone
) {
3153 struct per_cpu_pageset
*pset
= zone_pcp(zone
, cpu
);
3155 /* Free per_cpu_pageset if it is slab allocated */
3156 if (pset
!= &boot_pageset
[cpu
])
3158 zone_pcp(zone
, cpu
) = &boot_pageset
[cpu
];
3162 static int __cpuinit
pageset_cpuup_callback(struct notifier_block
*nfb
,
3163 unsigned long action
,
3166 int cpu
= (long)hcpu
;
3167 int ret
= NOTIFY_OK
;
3170 case CPU_UP_PREPARE
:
3171 case CPU_UP_PREPARE_FROZEN
:
3172 if (process_zones(cpu
))
3175 case CPU_UP_CANCELED
:
3176 case CPU_UP_CANCELED_FROZEN
:
3178 case CPU_DEAD_FROZEN
:
3179 free_zone_pagesets(cpu
);
3187 static struct notifier_block __cpuinitdata pageset_notifier
=
3188 { &pageset_cpuup_callback
, NULL
, 0 };
3190 void __init
setup_per_cpu_pageset(void)
3194 /* Initialize per_cpu_pageset for cpu 0.
3195 * A cpuup callback will do this for every cpu
3196 * as it comes online
3198 err
= process_zones(smp_processor_id());
3200 register_cpu_notifier(&pageset_notifier
);
3205 static noinline __init_refok
3206 int zone_wait_table_init(struct zone
*zone
, unsigned long zone_size_pages
)
3209 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
3213 * The per-page waitqueue mechanism uses hashed waitqueues
3216 zone
->wait_table_hash_nr_entries
=
3217 wait_table_hash_nr_entries(zone_size_pages
);
3218 zone
->wait_table_bits
=
3219 wait_table_bits(zone
->wait_table_hash_nr_entries
);
3220 alloc_size
= zone
->wait_table_hash_nr_entries
3221 * sizeof(wait_queue_head_t
);
3223 if (!slab_is_available()) {
3224 zone
->wait_table
= (wait_queue_head_t
*)
3225 alloc_bootmem_node(pgdat
, alloc_size
);
3228 * This case means that a zone whose size was 0 gets new memory
3229 * via memory hot-add.
3230 * But it may be the case that a new node was hot-added. In
3231 * this case vmalloc() will not be able to use this new node's
3232 * memory - this wait_table must be initialized to use this new
3233 * node itself as well.
3234 * To use this new node's memory, further consideration will be
3237 zone
->wait_table
= vmalloc(alloc_size
);
3239 if (!zone
->wait_table
)
3242 for(i
= 0; i
< zone
->wait_table_hash_nr_entries
; ++i
)
3243 init_waitqueue_head(zone
->wait_table
+ i
);
3248 static int __zone_pcp_update(void *data
)
3250 struct zone
*zone
= data
;
3252 unsigned long batch
= zone_batchsize(zone
), flags
;
3254 for (cpu
= 0; cpu
< NR_CPUS
; cpu
++) {
3255 struct per_cpu_pageset
*pset
;
3256 struct per_cpu_pages
*pcp
;
3258 pset
= zone_pcp(zone
, cpu
);
3261 local_irq_save(flags
);
3262 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
3263 setup_pageset(pset
, batch
);
3264 local_irq_restore(flags
);
3269 void zone_pcp_update(struct zone
*zone
)
3271 stop_machine(__zone_pcp_update
, zone
, NULL
);
3274 static __meminit
void zone_pcp_init(struct zone
*zone
)
3277 unsigned long batch
= zone_batchsize(zone
);
3279 for (cpu
= 0; cpu
< NR_CPUS
; cpu
++) {
3281 /* Early boot. Slab allocator not functional yet */
3282 zone_pcp(zone
, cpu
) = &boot_pageset
[cpu
];
3283 setup_pageset(&boot_pageset
[cpu
],0);
3285 setup_pageset(zone_pcp(zone
,cpu
), batch
);
3288 if (zone
->present_pages
)
3289 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%lu\n",
3290 zone
->name
, zone
->present_pages
, batch
);
3293 __meminit
int init_currently_empty_zone(struct zone
*zone
,
3294 unsigned long zone_start_pfn
,
3296 enum memmap_context context
)
3298 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
3300 ret
= zone_wait_table_init(zone
, size
);
3303 pgdat
->nr_zones
= zone_idx(zone
) + 1;
3305 zone
->zone_start_pfn
= zone_start_pfn
;
3307 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
3308 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
3310 (unsigned long)zone_idx(zone
),
3311 zone_start_pfn
, (zone_start_pfn
+ size
));
3313 zone_init_free_lists(zone
);
3318 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3320 * Basic iterator support. Return the first range of PFNs for a node
3321 * Note: nid == MAX_NUMNODES returns first region regardless of node
3323 static int __meminit
first_active_region_index_in_nid(int nid
)
3327 for (i
= 0; i
< nr_nodemap_entries
; i
++)
3328 if (nid
== MAX_NUMNODES
|| early_node_map
[i
].nid
== nid
)
3335 * Basic iterator support. Return the next active range of PFNs for a node
3336 * Note: nid == MAX_NUMNODES returns next region regardless of node
3338 static int __meminit
next_active_region_index_in_nid(int index
, int nid
)
3340 for (index
= index
+ 1; index
< nr_nodemap_entries
; index
++)
3341 if (nid
== MAX_NUMNODES
|| early_node_map
[index
].nid
== nid
)
3347 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
3349 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
3350 * Architectures may implement their own version but if add_active_range()
3351 * was used and there are no special requirements, this is a convenient
3354 int __meminit
__early_pfn_to_nid(unsigned long pfn
)
3358 for (i
= 0; i
< nr_nodemap_entries
; i
++) {
3359 unsigned long start_pfn
= early_node_map
[i
].start_pfn
;
3360 unsigned long end_pfn
= early_node_map
[i
].end_pfn
;
3362 if (start_pfn
<= pfn
&& pfn
< end_pfn
)
3363 return early_node_map
[i
].nid
;
3365 /* This is a memory hole */
3368 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
3370 int __meminit
early_pfn_to_nid(unsigned long pfn
)
3374 nid
= __early_pfn_to_nid(pfn
);
3377 /* just returns 0 */
3381 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
3382 bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
3386 nid
= __early_pfn_to_nid(pfn
);
3387 if (nid
>= 0 && nid
!= node
)
3393 /* Basic iterator support to walk early_node_map[] */
3394 #define for_each_active_range_index_in_nid(i, nid) \
3395 for (i = first_active_region_index_in_nid(nid); i != -1; \
3396 i = next_active_region_index_in_nid(i, nid))
3399 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
3400 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
3401 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
3403 * If an architecture guarantees that all ranges registered with
3404 * add_active_ranges() contain no holes and may be freed, this
3405 * this function may be used instead of calling free_bootmem() manually.
3407 void __init
free_bootmem_with_active_regions(int nid
,
3408 unsigned long max_low_pfn
)
3412 for_each_active_range_index_in_nid(i
, nid
) {
3413 unsigned long size_pages
= 0;
3414 unsigned long end_pfn
= early_node_map
[i
].end_pfn
;
3416 if (early_node_map
[i
].start_pfn
>= max_low_pfn
)
3419 if (end_pfn
> max_low_pfn
)
3420 end_pfn
= max_low_pfn
;
3422 size_pages
= end_pfn
- early_node_map
[i
].start_pfn
;
3423 free_bootmem_node(NODE_DATA(early_node_map
[i
].nid
),
3424 PFN_PHYS(early_node_map
[i
].start_pfn
),
3425 size_pages
<< PAGE_SHIFT
);
3429 void __init
work_with_active_regions(int nid
, work_fn_t work_fn
, void *data
)
3434 for_each_active_range_index_in_nid(i
, nid
) {
3435 ret
= work_fn(early_node_map
[i
].start_pfn
,
3436 early_node_map
[i
].end_pfn
, data
);
3442 * sparse_memory_present_with_active_regions - Call memory_present for each active range
3443 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
3445 * If an architecture guarantees that all ranges registered with
3446 * add_active_ranges() contain no holes and may be freed, this
3447 * function may be used instead of calling memory_present() manually.
3449 void __init
sparse_memory_present_with_active_regions(int nid
)
3453 for_each_active_range_index_in_nid(i
, nid
)
3454 memory_present(early_node_map
[i
].nid
,
3455 early_node_map
[i
].start_pfn
,
3456 early_node_map
[i
].end_pfn
);
3460 * get_pfn_range_for_nid - Return the start and end page frames for a node
3461 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
3462 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
3463 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
3465 * It returns the start and end page frame of a node based on information
3466 * provided by an arch calling add_active_range(). If called for a node
3467 * with no available memory, a warning is printed and the start and end
3470 void __meminit
get_pfn_range_for_nid(unsigned int nid
,
3471 unsigned long *start_pfn
, unsigned long *end_pfn
)
3477 for_each_active_range_index_in_nid(i
, nid
) {
3478 *start_pfn
= min(*start_pfn
, early_node_map
[i
].start_pfn
);
3479 *end_pfn
= max(*end_pfn
, early_node_map
[i
].end_pfn
);
3482 if (*start_pfn
== -1UL)
3487 * This finds a zone that can be used for ZONE_MOVABLE pages. The
3488 * assumption is made that zones within a node are ordered in monotonic
3489 * increasing memory addresses so that the "highest" populated zone is used
3491 static void __init
find_usable_zone_for_movable(void)
3494 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
3495 if (zone_index
== ZONE_MOVABLE
)
3498 if (arch_zone_highest_possible_pfn
[zone_index
] >
3499 arch_zone_lowest_possible_pfn
[zone_index
])
3503 VM_BUG_ON(zone_index
== -1);
3504 movable_zone
= zone_index
;
3508 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
3509 * because it is sized independant of architecture. Unlike the other zones,
3510 * the starting point for ZONE_MOVABLE is not fixed. It may be different
3511 * in each node depending on the size of each node and how evenly kernelcore
3512 * is distributed. This helper function adjusts the zone ranges
3513 * provided by the architecture for a given node by using the end of the
3514 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
3515 * zones within a node are in order of monotonic increases memory addresses
3517 static void __meminit
adjust_zone_range_for_zone_movable(int nid
,
3518 unsigned long zone_type
,
3519 unsigned long node_start_pfn
,
3520 unsigned long node_end_pfn
,
3521 unsigned long *zone_start_pfn
,
3522 unsigned long *zone_end_pfn
)
3524 /* Only adjust if ZONE_MOVABLE is on this node */
3525 if (zone_movable_pfn
[nid
]) {
3526 /* Size ZONE_MOVABLE */
3527 if (zone_type
== ZONE_MOVABLE
) {
3528 *zone_start_pfn
= zone_movable_pfn
[nid
];
3529 *zone_end_pfn
= min(node_end_pfn
,
3530 arch_zone_highest_possible_pfn
[movable_zone
]);
3532 /* Adjust for ZONE_MOVABLE starting within this range */
3533 } else if (*zone_start_pfn
< zone_movable_pfn
[nid
] &&
3534 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
3535 *zone_end_pfn
= zone_movable_pfn
[nid
];
3537 /* Check if this whole range is within ZONE_MOVABLE */
3538 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
3539 *zone_start_pfn
= *zone_end_pfn
;
3544 * Return the number of pages a zone spans in a node, including holes
3545 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
3547 static unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
3548 unsigned long zone_type
,
3549 unsigned long *ignored
)
3551 unsigned long node_start_pfn
, node_end_pfn
;
3552 unsigned long zone_start_pfn
, zone_end_pfn
;
3554 /* Get the start and end of the node and zone */
3555 get_pfn_range_for_nid(nid
, &node_start_pfn
, &node_end_pfn
);
3556 zone_start_pfn
= arch_zone_lowest_possible_pfn
[zone_type
];
3557 zone_end_pfn
= arch_zone_highest_possible_pfn
[zone_type
];
3558 adjust_zone_range_for_zone_movable(nid
, zone_type
,
3559 node_start_pfn
, node_end_pfn
,
3560 &zone_start_pfn
, &zone_end_pfn
);
3562 /* Check that this node has pages within the zone's required range */
3563 if (zone_end_pfn
< node_start_pfn
|| zone_start_pfn
> node_end_pfn
)
3566 /* Move the zone boundaries inside the node if necessary */
3567 zone_end_pfn
= min(zone_end_pfn
, node_end_pfn
);
3568 zone_start_pfn
= max(zone_start_pfn
, node_start_pfn
);
3570 /* Return the spanned pages */
3571 return zone_end_pfn
- zone_start_pfn
;
3575 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
3576 * then all holes in the requested range will be accounted for.
3578 static unsigned long __meminit
__absent_pages_in_range(int nid
,
3579 unsigned long range_start_pfn
,
3580 unsigned long range_end_pfn
)
3583 unsigned long prev_end_pfn
= 0, hole_pages
= 0;
3584 unsigned long start_pfn
;
3586 /* Find the end_pfn of the first active range of pfns in the node */
3587 i
= first_active_region_index_in_nid(nid
);
3591 prev_end_pfn
= min(early_node_map
[i
].start_pfn
, range_end_pfn
);
3593 /* Account for ranges before physical memory on this node */
3594 if (early_node_map
[i
].start_pfn
> range_start_pfn
)
3595 hole_pages
= prev_end_pfn
- range_start_pfn
;
3597 /* Find all holes for the zone within the node */
3598 for (; i
!= -1; i
= next_active_region_index_in_nid(i
, nid
)) {
3600 /* No need to continue if prev_end_pfn is outside the zone */
3601 if (prev_end_pfn
>= range_end_pfn
)
3604 /* Make sure the end of the zone is not within the hole */
3605 start_pfn
= min(early_node_map
[i
].start_pfn
, range_end_pfn
);
3606 prev_end_pfn
= max(prev_end_pfn
, range_start_pfn
);
3608 /* Update the hole size cound and move on */
3609 if (start_pfn
> range_start_pfn
) {
3610 BUG_ON(prev_end_pfn
> start_pfn
);
3611 hole_pages
+= start_pfn
- prev_end_pfn
;
3613 prev_end_pfn
= early_node_map
[i
].end_pfn
;
3616 /* Account for ranges past physical memory on this node */
3617 if (range_end_pfn
> prev_end_pfn
)
3618 hole_pages
+= range_end_pfn
-
3619 max(range_start_pfn
, prev_end_pfn
);
3625 * absent_pages_in_range - Return number of page frames in holes within a range
3626 * @start_pfn: The start PFN to start searching for holes
3627 * @end_pfn: The end PFN to stop searching for holes
3629 * It returns the number of pages frames in memory holes within a range.
3631 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
3632 unsigned long end_pfn
)
3634 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
3637 /* Return the number of page frames in holes in a zone on a node */
3638 static unsigned long __meminit
zone_absent_pages_in_node(int nid
,
3639 unsigned long zone_type
,
3640 unsigned long *ignored
)
3642 unsigned long node_start_pfn
, node_end_pfn
;
3643 unsigned long zone_start_pfn
, zone_end_pfn
;
3645 get_pfn_range_for_nid(nid
, &node_start_pfn
, &node_end_pfn
);
3646 zone_start_pfn
= max(arch_zone_lowest_possible_pfn
[zone_type
],
3648 zone_end_pfn
= min(arch_zone_highest_possible_pfn
[zone_type
],
3651 adjust_zone_range_for_zone_movable(nid
, zone_type
,
3652 node_start_pfn
, node_end_pfn
,
3653 &zone_start_pfn
, &zone_end_pfn
);
3654 return __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
3658 static inline unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
3659 unsigned long zone_type
,
3660 unsigned long *zones_size
)
3662 return zones_size
[zone_type
];
3665 static inline unsigned long __meminit
zone_absent_pages_in_node(int nid
,
3666 unsigned long zone_type
,
3667 unsigned long *zholes_size
)
3672 return zholes_size
[zone_type
];
3677 static void __meminit
calculate_node_totalpages(struct pglist_data
*pgdat
,
3678 unsigned long *zones_size
, unsigned long *zholes_size
)
3680 unsigned long realtotalpages
, totalpages
= 0;
3683 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
3684 totalpages
+= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
3686 pgdat
->node_spanned_pages
= totalpages
;
3688 realtotalpages
= totalpages
;
3689 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
3691 zone_absent_pages_in_node(pgdat
->node_id
, i
,
3693 pgdat
->node_present_pages
= realtotalpages
;
3694 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
3698 #ifndef CONFIG_SPARSEMEM
3700 * Calculate the size of the zone->blockflags rounded to an unsigned long
3701 * Start by making sure zonesize is a multiple of pageblock_order by rounding
3702 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
3703 * round what is now in bits to nearest long in bits, then return it in
3706 static unsigned long __init
usemap_size(unsigned long zonesize
)
3708 unsigned long usemapsize
;
3710 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
3711 usemapsize
= usemapsize
>> pageblock_order
;
3712 usemapsize
*= NR_PAGEBLOCK_BITS
;
3713 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
3715 return usemapsize
/ 8;
3718 static void __init
setup_usemap(struct pglist_data
*pgdat
,
3719 struct zone
*zone
, unsigned long zonesize
)
3721 unsigned long usemapsize
= usemap_size(zonesize
);
3722 zone
->pageblock_flags
= NULL
;
3724 zone
->pageblock_flags
= alloc_bootmem_node(pgdat
, usemapsize
);
3727 static void inline setup_usemap(struct pglist_data
*pgdat
,
3728 struct zone
*zone
, unsigned long zonesize
) {}
3729 #endif /* CONFIG_SPARSEMEM */
3731 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
3733 /* Return a sensible default order for the pageblock size. */
3734 static inline int pageblock_default_order(void)
3736 if (HPAGE_SHIFT
> PAGE_SHIFT
)
3737 return HUGETLB_PAGE_ORDER
;
3742 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
3743 static inline void __init
set_pageblock_order(unsigned int order
)
3745 /* Check that pageblock_nr_pages has not already been setup */
3746 if (pageblock_order
)
3750 * Assume the largest contiguous order of interest is a huge page.
3751 * This value may be variable depending on boot parameters on IA64
3753 pageblock_order
= order
;
3755 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
3758 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
3759 * and pageblock_default_order() are unused as pageblock_order is set
3760 * at compile-time. See include/linux/pageblock-flags.h for the values of
3761 * pageblock_order based on the kernel config
3763 static inline int pageblock_default_order(unsigned int order
)
3767 #define set_pageblock_order(x) do {} while (0)
3769 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
3772 * Set up the zone data structures:
3773 * - mark all pages reserved
3774 * - mark all memory queues empty
3775 * - clear the memory bitmaps
3777 static void __paginginit
free_area_init_core(struct pglist_data
*pgdat
,
3778 unsigned long *zones_size
, unsigned long *zholes_size
)
3781 int nid
= pgdat
->node_id
;
3782 unsigned long zone_start_pfn
= pgdat
->node_start_pfn
;
3785 pgdat_resize_init(pgdat
);
3786 pgdat
->nr_zones
= 0;
3787 init_waitqueue_head(&pgdat
->kswapd_wait
);
3788 pgdat
->kswapd_max_order
= 0;
3789 pgdat_page_cgroup_init(pgdat
);
3791 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
3792 struct zone
*zone
= pgdat
->node_zones
+ j
;
3793 unsigned long size
, realsize
, memmap_pages
;
3796 size
= zone_spanned_pages_in_node(nid
, j
, zones_size
);
3797 realsize
= size
- zone_absent_pages_in_node(nid
, j
,
3801 * Adjust realsize so that it accounts for how much memory
3802 * is used by this zone for memmap. This affects the watermark
3803 * and per-cpu initialisations
3806 PAGE_ALIGN(size
* sizeof(struct page
)) >> PAGE_SHIFT
;
3807 if (realsize
>= memmap_pages
) {
3808 realsize
-= memmap_pages
;
3811 " %s zone: %lu pages used for memmap\n",
3812 zone_names
[j
], memmap_pages
);
3815 " %s zone: %lu pages exceeds realsize %lu\n",
3816 zone_names
[j
], memmap_pages
, realsize
);
3818 /* Account for reserved pages */
3819 if (j
== 0 && realsize
> dma_reserve
) {
3820 realsize
-= dma_reserve
;
3821 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
3822 zone_names
[0], dma_reserve
);
3825 if (!is_highmem_idx(j
))
3826 nr_kernel_pages
+= realsize
;
3827 nr_all_pages
+= realsize
;
3829 zone
->spanned_pages
= size
;
3830 zone
->present_pages
= realsize
;
3833 zone
->min_unmapped_pages
= (realsize
*sysctl_min_unmapped_ratio
)
3835 zone
->min_slab_pages
= (realsize
* sysctl_min_slab_ratio
) / 100;
3837 zone
->name
= zone_names
[j
];
3838 spin_lock_init(&zone
->lock
);
3839 spin_lock_init(&zone
->lru_lock
);
3840 zone_seqlock_init(zone
);
3841 zone
->zone_pgdat
= pgdat
;
3843 zone
->prev_priority
= DEF_PRIORITY
;
3845 zone_pcp_init(zone
);
3847 INIT_LIST_HEAD(&zone
->lru
[l
].list
);
3848 zone
->reclaim_stat
.nr_saved_scan
[l
] = 0;
3850 zone
->reclaim_stat
.recent_rotated
[0] = 0;
3851 zone
->reclaim_stat
.recent_rotated
[1] = 0;
3852 zone
->reclaim_stat
.recent_scanned
[0] = 0;
3853 zone
->reclaim_stat
.recent_scanned
[1] = 0;
3854 zap_zone_vm_stats(zone
);
3859 set_pageblock_order(pageblock_default_order());
3860 setup_usemap(pgdat
, zone
, size
);
3861 ret
= init_currently_empty_zone(zone
, zone_start_pfn
,
3862 size
, MEMMAP_EARLY
);
3864 memmap_init(size
, nid
, j
, zone_start_pfn
);
3865 zone_start_pfn
+= size
;
3869 static void __init_refok
alloc_node_mem_map(struct pglist_data
*pgdat
)
3871 /* Skip empty nodes */
3872 if (!pgdat
->node_spanned_pages
)
3875 #ifdef CONFIG_FLAT_NODE_MEM_MAP
3876 /* ia64 gets its own node_mem_map, before this, without bootmem */
3877 if (!pgdat
->node_mem_map
) {
3878 unsigned long size
, start
, end
;
3882 * The zone's endpoints aren't required to be MAX_ORDER
3883 * aligned but the node_mem_map endpoints must be in order
3884 * for the buddy allocator to function correctly.
3886 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
3887 end
= pgdat
->node_start_pfn
+ pgdat
->node_spanned_pages
;
3888 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
3889 size
= (end
- start
) * sizeof(struct page
);
3890 map
= alloc_remap(pgdat
->node_id
, size
);
3892 map
= alloc_bootmem_node(pgdat
, size
);
3893 pgdat
->node_mem_map
= map
+ (pgdat
->node_start_pfn
- start
);
3895 #ifndef CONFIG_NEED_MULTIPLE_NODES
3897 * With no DISCONTIG, the global mem_map is just set as node 0's
3899 if (pgdat
== NODE_DATA(0)) {
3900 mem_map
= NODE_DATA(0)->node_mem_map
;
3901 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3902 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
3903 mem_map
-= (pgdat
->node_start_pfn
- ARCH_PFN_OFFSET
);
3904 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
3907 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
3910 void __paginginit
free_area_init_node(int nid
, unsigned long *zones_size
,
3911 unsigned long node_start_pfn
, unsigned long *zholes_size
)
3913 pg_data_t
*pgdat
= NODE_DATA(nid
);
3915 pgdat
->node_id
= nid
;
3916 pgdat
->node_start_pfn
= node_start_pfn
;
3917 calculate_node_totalpages(pgdat
, zones_size
, zholes_size
);
3919 alloc_node_mem_map(pgdat
);
3920 #ifdef CONFIG_FLAT_NODE_MEM_MAP
3921 printk(KERN_DEBUG
"free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
3922 nid
, (unsigned long)pgdat
,
3923 (unsigned long)pgdat
->node_mem_map
);
3926 free_area_init_core(pgdat
, zones_size
, zholes_size
);
3929 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3931 #if MAX_NUMNODES > 1
3933 * Figure out the number of possible node ids.
3935 static void __init
setup_nr_node_ids(void)
3938 unsigned int highest
= 0;
3940 for_each_node_mask(node
, node_possible_map
)
3942 nr_node_ids
= highest
+ 1;
3945 static inline void setup_nr_node_ids(void)
3951 * add_active_range - Register a range of PFNs backed by physical memory
3952 * @nid: The node ID the range resides on
3953 * @start_pfn: The start PFN of the available physical memory
3954 * @end_pfn: The end PFN of the available physical memory
3956 * These ranges are stored in an early_node_map[] and later used by
3957 * free_area_init_nodes() to calculate zone sizes and holes. If the
3958 * range spans a memory hole, it is up to the architecture to ensure
3959 * the memory is not freed by the bootmem allocator. If possible
3960 * the range being registered will be merged with existing ranges.
3962 void __init
add_active_range(unsigned int nid
, unsigned long start_pfn
,
3963 unsigned long end_pfn
)
3967 mminit_dprintk(MMINIT_TRACE
, "memory_register",
3968 "Entering add_active_range(%d, %#lx, %#lx) "
3969 "%d entries of %d used\n",
3970 nid
, start_pfn
, end_pfn
,
3971 nr_nodemap_entries
, MAX_ACTIVE_REGIONS
);
3973 mminit_validate_memmodel_limits(&start_pfn
, &end_pfn
);
3975 /* Merge with existing active regions if possible */
3976 for (i
= 0; i
< nr_nodemap_entries
; i
++) {
3977 if (early_node_map
[i
].nid
!= nid
)
3980 /* Skip if an existing region covers this new one */
3981 if (start_pfn
>= early_node_map
[i
].start_pfn
&&
3982 end_pfn
<= early_node_map
[i
].end_pfn
)
3985 /* Merge forward if suitable */
3986 if (start_pfn
<= early_node_map
[i
].end_pfn
&&
3987 end_pfn
> early_node_map
[i
].end_pfn
) {
3988 early_node_map
[i
].end_pfn
= end_pfn
;
3992 /* Merge backward if suitable */
3993 if (start_pfn
< early_node_map
[i
].end_pfn
&&
3994 end_pfn
>= early_node_map
[i
].start_pfn
) {
3995 early_node_map
[i
].start_pfn
= start_pfn
;
4000 /* Check that early_node_map is large enough */
4001 if (i
>= MAX_ACTIVE_REGIONS
) {
4002 printk(KERN_CRIT
"More than %d memory regions, truncating\n",
4003 MAX_ACTIVE_REGIONS
);
4007 early_node_map
[i
].nid
= nid
;
4008 early_node_map
[i
].start_pfn
= start_pfn
;
4009 early_node_map
[i
].end_pfn
= end_pfn
;
4010 nr_nodemap_entries
= i
+ 1;
4014 * remove_active_range - Shrink an existing registered range of PFNs
4015 * @nid: The node id the range is on that should be shrunk
4016 * @start_pfn: The new PFN of the range
4017 * @end_pfn: The new PFN of the range
4019 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
4020 * The map is kept near the end physical page range that has already been
4021 * registered. This function allows an arch to shrink an existing registered
4024 void __init
remove_active_range(unsigned int nid
, unsigned long start_pfn
,
4025 unsigned long end_pfn
)
4030 printk(KERN_DEBUG
"remove_active_range (%d, %lu, %lu)\n",
4031 nid
, start_pfn
, end_pfn
);
4033 /* Find the old active region end and shrink */
4034 for_each_active_range_index_in_nid(i
, nid
) {
4035 if (early_node_map
[i
].start_pfn
>= start_pfn
&&
4036 early_node_map
[i
].end_pfn
<= end_pfn
) {
4038 early_node_map
[i
].start_pfn
= 0;
4039 early_node_map
[i
].end_pfn
= 0;
4043 if (early_node_map
[i
].start_pfn
< start_pfn
&&
4044 early_node_map
[i
].end_pfn
> start_pfn
) {
4045 unsigned long temp_end_pfn
= early_node_map
[i
].end_pfn
;
4046 early_node_map
[i
].end_pfn
= start_pfn
;
4047 if (temp_end_pfn
> end_pfn
)
4048 add_active_range(nid
, end_pfn
, temp_end_pfn
);
4051 if (early_node_map
[i
].start_pfn
>= start_pfn
&&
4052 early_node_map
[i
].end_pfn
> end_pfn
&&
4053 early_node_map
[i
].start_pfn
< end_pfn
) {
4054 early_node_map
[i
].start_pfn
= end_pfn
;
4062 /* remove the blank ones */
4063 for (i
= nr_nodemap_entries
- 1; i
> 0; i
--) {
4064 if (early_node_map
[i
].nid
!= nid
)
4066 if (early_node_map
[i
].end_pfn
)
4068 /* we found it, get rid of it */
4069 for (j
= i
; j
< nr_nodemap_entries
- 1; j
++)
4070 memcpy(&early_node_map
[j
], &early_node_map
[j
+1],
4071 sizeof(early_node_map
[j
]));
4072 j
= nr_nodemap_entries
- 1;
4073 memset(&early_node_map
[j
], 0, sizeof(early_node_map
[j
]));
4074 nr_nodemap_entries
--;
4079 * remove_all_active_ranges - Remove all currently registered regions
4081 * During discovery, it may be found that a table like SRAT is invalid
4082 * and an alternative discovery method must be used. This function removes
4083 * all currently registered regions.
4085 void __init
remove_all_active_ranges(void)
4087 memset(early_node_map
, 0, sizeof(early_node_map
));
4088 nr_nodemap_entries
= 0;
4091 /* Compare two active node_active_regions */
4092 static int __init
cmp_node_active_region(const void *a
, const void *b
)
4094 struct node_active_region
*arange
= (struct node_active_region
*)a
;
4095 struct node_active_region
*brange
= (struct node_active_region
*)b
;
4097 /* Done this way to avoid overflows */
4098 if (arange
->start_pfn
> brange
->start_pfn
)
4100 if (arange
->start_pfn
< brange
->start_pfn
)
4106 /* sort the node_map by start_pfn */
4107 static void __init
sort_node_map(void)
4109 sort(early_node_map
, (size_t)nr_nodemap_entries
,
4110 sizeof(struct node_active_region
),
4111 cmp_node_active_region
, NULL
);
4114 /* Find the lowest pfn for a node */
4115 static unsigned long __init
find_min_pfn_for_node(int nid
)
4118 unsigned long min_pfn
= ULONG_MAX
;
4120 /* Assuming a sorted map, the first range found has the starting pfn */
4121 for_each_active_range_index_in_nid(i
, nid
)
4122 min_pfn
= min(min_pfn
, early_node_map
[i
].start_pfn
);
4124 if (min_pfn
== ULONG_MAX
) {
4126 "Could not find start_pfn for node %d\n", nid
);
4134 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4136 * It returns the minimum PFN based on information provided via
4137 * add_active_range().
4139 unsigned long __init
find_min_pfn_with_active_regions(void)
4141 return find_min_pfn_for_node(MAX_NUMNODES
);
4145 * early_calculate_totalpages()
4146 * Sum pages in active regions for movable zone.
4147 * Populate N_HIGH_MEMORY for calculating usable_nodes.
4149 static unsigned long __init
early_calculate_totalpages(void)
4152 unsigned long totalpages
= 0;
4154 for (i
= 0; i
< nr_nodemap_entries
; i
++) {
4155 unsigned long pages
= early_node_map
[i
].end_pfn
-
4156 early_node_map
[i
].start_pfn
;
4157 totalpages
+= pages
;
4159 node_set_state(early_node_map
[i
].nid
, N_HIGH_MEMORY
);
4165 * Find the PFN the Movable zone begins in each node. Kernel memory
4166 * is spread evenly between nodes as long as the nodes have enough
4167 * memory. When they don't, some nodes will have more kernelcore than
4170 static void __init
find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn
)
4173 unsigned long usable_startpfn
;
4174 unsigned long kernelcore_node
, kernelcore_remaining
;
4175 /* save the state before borrow the nodemask */
4176 nodemask_t saved_node_state
= node_states
[N_HIGH_MEMORY
];
4177 unsigned long totalpages
= early_calculate_totalpages();
4178 int usable_nodes
= nodes_weight(node_states
[N_HIGH_MEMORY
]);
4181 * If movablecore was specified, calculate what size of
4182 * kernelcore that corresponds so that memory usable for
4183 * any allocation type is evenly spread. If both kernelcore
4184 * and movablecore are specified, then the value of kernelcore
4185 * will be used for required_kernelcore if it's greater than
4186 * what movablecore would have allowed.
4188 if (required_movablecore
) {
4189 unsigned long corepages
;
4192 * Round-up so that ZONE_MOVABLE is at least as large as what
4193 * was requested by the user
4195 required_movablecore
=
4196 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
4197 corepages
= totalpages
- required_movablecore
;
4199 required_kernelcore
= max(required_kernelcore
, corepages
);
4202 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4203 if (!required_kernelcore
)
4206 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4207 find_usable_zone_for_movable();
4208 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
4211 /* Spread kernelcore memory as evenly as possible throughout nodes */
4212 kernelcore_node
= required_kernelcore
/ usable_nodes
;
4213 for_each_node_state(nid
, N_HIGH_MEMORY
) {
4215 * Recalculate kernelcore_node if the division per node
4216 * now exceeds what is necessary to satisfy the requested
4217 * amount of memory for the kernel
4219 if (required_kernelcore
< kernelcore_node
)
4220 kernelcore_node
= required_kernelcore
/ usable_nodes
;
4223 * As the map is walked, we track how much memory is usable
4224 * by the kernel using kernelcore_remaining. When it is
4225 * 0, the rest of the node is usable by ZONE_MOVABLE
4227 kernelcore_remaining
= kernelcore_node
;
4229 /* Go through each range of PFNs within this node */
4230 for_each_active_range_index_in_nid(i
, nid
) {
4231 unsigned long start_pfn
, end_pfn
;
4232 unsigned long size_pages
;
4234 start_pfn
= max(early_node_map
[i
].start_pfn
,
4235 zone_movable_pfn
[nid
]);
4236 end_pfn
= early_node_map
[i
].end_pfn
;
4237 if (start_pfn
>= end_pfn
)
4240 /* Account for what is only usable for kernelcore */
4241 if (start_pfn
< usable_startpfn
) {
4242 unsigned long kernel_pages
;
4243 kernel_pages
= min(end_pfn
, usable_startpfn
)
4246 kernelcore_remaining
-= min(kernel_pages
,
4247 kernelcore_remaining
);
4248 required_kernelcore
-= min(kernel_pages
,
4249 required_kernelcore
);
4251 /* Continue if range is now fully accounted */
4252 if (end_pfn
<= usable_startpfn
) {
4255 * Push zone_movable_pfn to the end so
4256 * that if we have to rebalance
4257 * kernelcore across nodes, we will
4258 * not double account here
4260 zone_movable_pfn
[nid
] = end_pfn
;
4263 start_pfn
= usable_startpfn
;
4267 * The usable PFN range for ZONE_MOVABLE is from
4268 * start_pfn->end_pfn. Calculate size_pages as the
4269 * number of pages used as kernelcore
4271 size_pages
= end_pfn
- start_pfn
;
4272 if (size_pages
> kernelcore_remaining
)
4273 size_pages
= kernelcore_remaining
;
4274 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
4277 * Some kernelcore has been met, update counts and
4278 * break if the kernelcore for this node has been
4281 required_kernelcore
-= min(required_kernelcore
,
4283 kernelcore_remaining
-= size_pages
;
4284 if (!kernelcore_remaining
)
4290 * If there is still required_kernelcore, we do another pass with one
4291 * less node in the count. This will push zone_movable_pfn[nid] further
4292 * along on the nodes that still have memory until kernelcore is
4296 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
4299 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4300 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
4301 zone_movable_pfn
[nid
] =
4302 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
4305 /* restore the node_state */
4306 node_states
[N_HIGH_MEMORY
] = saved_node_state
;
4309 /* Any regular memory on that node ? */
4310 static void check_for_regular_memory(pg_data_t
*pgdat
)
4312 #ifdef CONFIG_HIGHMEM
4313 enum zone_type zone_type
;
4315 for (zone_type
= 0; zone_type
<= ZONE_NORMAL
; zone_type
++) {
4316 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
4317 if (zone
->present_pages
)
4318 node_set_state(zone_to_nid(zone
), N_NORMAL_MEMORY
);
4324 * free_area_init_nodes - Initialise all pg_data_t and zone data
4325 * @max_zone_pfn: an array of max PFNs for each zone
4327 * This will call free_area_init_node() for each active node in the system.
4328 * Using the page ranges provided by add_active_range(), the size of each
4329 * zone in each node and their holes is calculated. If the maximum PFN
4330 * between two adjacent zones match, it is assumed that the zone is empty.
4331 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4332 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4333 * starts where the previous one ended. For example, ZONE_DMA32 starts
4334 * at arch_max_dma_pfn.
4336 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
4341 /* Sort early_node_map as initialisation assumes it is sorted */
4344 /* Record where the zone boundaries are */
4345 memset(arch_zone_lowest_possible_pfn
, 0,
4346 sizeof(arch_zone_lowest_possible_pfn
));
4347 memset(arch_zone_highest_possible_pfn
, 0,
4348 sizeof(arch_zone_highest_possible_pfn
));
4349 arch_zone_lowest_possible_pfn
[0] = find_min_pfn_with_active_regions();
4350 arch_zone_highest_possible_pfn
[0] = max_zone_pfn
[0];
4351 for (i
= 1; i
< MAX_NR_ZONES
; i
++) {
4352 if (i
== ZONE_MOVABLE
)
4354 arch_zone_lowest_possible_pfn
[i
] =
4355 arch_zone_highest_possible_pfn
[i
-1];
4356 arch_zone_highest_possible_pfn
[i
] =
4357 max(max_zone_pfn
[i
], arch_zone_lowest_possible_pfn
[i
]);
4359 arch_zone_lowest_possible_pfn
[ZONE_MOVABLE
] = 0;
4360 arch_zone_highest_possible_pfn
[ZONE_MOVABLE
] = 0;
4362 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4363 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
4364 find_zone_movable_pfns_for_nodes(zone_movable_pfn
);
4366 /* Print out the zone ranges */
4367 printk("Zone PFN ranges:\n");
4368 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
4369 if (i
== ZONE_MOVABLE
)
4371 printk(" %-8s %0#10lx -> %0#10lx\n",
4373 arch_zone_lowest_possible_pfn
[i
],
4374 arch_zone_highest_possible_pfn
[i
]);
4377 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4378 printk("Movable zone start PFN for each node\n");
4379 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
4380 if (zone_movable_pfn
[i
])
4381 printk(" Node %d: %lu\n", i
, zone_movable_pfn
[i
]);
4384 /* Print out the early_node_map[] */
4385 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries
);
4386 for (i
= 0; i
< nr_nodemap_entries
; i
++)
4387 printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map
[i
].nid
,
4388 early_node_map
[i
].start_pfn
,
4389 early_node_map
[i
].end_pfn
);
4391 /* Initialise every node */
4392 mminit_verify_pageflags_layout();
4393 setup_nr_node_ids();
4394 for_each_online_node(nid
) {
4395 pg_data_t
*pgdat
= NODE_DATA(nid
);
4396 free_area_init_node(nid
, NULL
,
4397 find_min_pfn_for_node(nid
), NULL
);
4399 /* Any memory on that node */
4400 if (pgdat
->node_present_pages
)
4401 node_set_state(nid
, N_HIGH_MEMORY
);
4402 check_for_regular_memory(pgdat
);
4406 static int __init
cmdline_parse_core(char *p
, unsigned long *core
)
4408 unsigned long long coremem
;
4412 coremem
= memparse(p
, &p
);
4413 *core
= coremem
>> PAGE_SHIFT
;
4415 /* Paranoid check that UL is enough for the coremem value */
4416 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
4422 * kernelcore=size sets the amount of memory for use for allocations that
4423 * cannot be reclaimed or migrated.
4425 static int __init
cmdline_parse_kernelcore(char *p
)
4427 return cmdline_parse_core(p
, &required_kernelcore
);
4431 * movablecore=size sets the amount of memory for use for allocations that
4432 * can be reclaimed or migrated.
4434 static int __init
cmdline_parse_movablecore(char *p
)
4436 return cmdline_parse_core(p
, &required_movablecore
);
4439 early_param("kernelcore", cmdline_parse_kernelcore
);
4440 early_param("movablecore", cmdline_parse_movablecore
);
4442 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4445 * set_dma_reserve - set the specified number of pages reserved in the first zone
4446 * @new_dma_reserve: The number of pages to mark reserved
4448 * The per-cpu batchsize and zone watermarks are determined by present_pages.
4449 * In the DMA zone, a significant percentage may be consumed by kernel image
4450 * and other unfreeable allocations which can skew the watermarks badly. This
4451 * function may optionally be used to account for unfreeable pages in the
4452 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
4453 * smaller per-cpu batchsize.
4455 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
4457 dma_reserve
= new_dma_reserve
;
4460 #ifndef CONFIG_NEED_MULTIPLE_NODES
4461 struct pglist_data __refdata contig_page_data
= { .bdata
= &bootmem_node_data
[0] };
4462 EXPORT_SYMBOL(contig_page_data
);
4465 void __init
free_area_init(unsigned long *zones_size
)
4467 free_area_init_node(0, zones_size
,
4468 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
4471 static int page_alloc_cpu_notify(struct notifier_block
*self
,
4472 unsigned long action
, void *hcpu
)
4474 int cpu
= (unsigned long)hcpu
;
4476 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
) {
4480 * Spill the event counters of the dead processor
4481 * into the current processors event counters.
4482 * This artificially elevates the count of the current
4485 vm_events_fold_cpu(cpu
);
4488 * Zero the differential counters of the dead processor
4489 * so that the vm statistics are consistent.
4491 * This is only okay since the processor is dead and cannot
4492 * race with what we are doing.
4494 refresh_cpu_vm_stats(cpu
);
4499 void __init
page_alloc_init(void)
4501 hotcpu_notifier(page_alloc_cpu_notify
, 0);
4505 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
4506 * or min_free_kbytes changes.
4508 static void calculate_totalreserve_pages(void)
4510 struct pglist_data
*pgdat
;
4511 unsigned long reserve_pages
= 0;
4512 enum zone_type i
, j
;
4514 for_each_online_pgdat(pgdat
) {
4515 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
4516 struct zone
*zone
= pgdat
->node_zones
+ i
;
4517 unsigned long max
= 0;
4519 /* Find valid and maximum lowmem_reserve in the zone */
4520 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
4521 if (zone
->lowmem_reserve
[j
] > max
)
4522 max
= zone
->lowmem_reserve
[j
];
4525 /* we treat the high watermark as reserved pages. */
4526 max
+= high_wmark_pages(zone
);
4528 if (max
> zone
->present_pages
)
4529 max
= zone
->present_pages
;
4530 reserve_pages
+= max
;
4533 totalreserve_pages
= reserve_pages
;
4537 * setup_per_zone_lowmem_reserve - called whenever
4538 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
4539 * has a correct pages reserved value, so an adequate number of
4540 * pages are left in the zone after a successful __alloc_pages().
4542 static void setup_per_zone_lowmem_reserve(void)
4544 struct pglist_data
*pgdat
;
4545 enum zone_type j
, idx
;
4547 for_each_online_pgdat(pgdat
) {
4548 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
4549 struct zone
*zone
= pgdat
->node_zones
+ j
;
4550 unsigned long present_pages
= zone
->present_pages
;
4552 zone
->lowmem_reserve
[j
] = 0;
4556 struct zone
*lower_zone
;
4560 if (sysctl_lowmem_reserve_ratio
[idx
] < 1)
4561 sysctl_lowmem_reserve_ratio
[idx
] = 1;
4563 lower_zone
= pgdat
->node_zones
+ idx
;
4564 lower_zone
->lowmem_reserve
[j
] = present_pages
/
4565 sysctl_lowmem_reserve_ratio
[idx
];
4566 present_pages
+= lower_zone
->present_pages
;
4571 /* update totalreserve_pages */
4572 calculate_totalreserve_pages();
4576 * setup_per_zone_wmarks - called when min_free_kbytes changes
4577 * or when memory is hot-{added|removed}
4579 * Ensures that the watermark[min,low,high] values for each zone are set
4580 * correctly with respect to min_free_kbytes.
4582 void setup_per_zone_wmarks(void)
4584 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
4585 unsigned long lowmem_pages
= 0;
4587 unsigned long flags
;
4589 /* Calculate total number of !ZONE_HIGHMEM pages */
4590 for_each_zone(zone
) {
4591 if (!is_highmem(zone
))
4592 lowmem_pages
+= zone
->present_pages
;
4595 for_each_zone(zone
) {
4598 spin_lock_irqsave(&zone
->lock
, flags
);
4599 tmp
= (u64
)pages_min
* zone
->present_pages
;
4600 do_div(tmp
, lowmem_pages
);
4601 if (is_highmem(zone
)) {
4603 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
4604 * need highmem pages, so cap pages_min to a small
4607 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
4608 * deltas controls asynch page reclaim, and so should
4609 * not be capped for highmem.
4613 min_pages
= zone
->present_pages
/ 1024;
4614 if (min_pages
< SWAP_CLUSTER_MAX
)
4615 min_pages
= SWAP_CLUSTER_MAX
;
4616 if (min_pages
> 128)
4618 zone
->watermark
[WMARK_MIN
] = min_pages
;
4621 * If it's a lowmem zone, reserve a number of pages
4622 * proportionate to the zone's size.
4624 zone
->watermark
[WMARK_MIN
] = tmp
;
4627 zone
->watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + (tmp
>> 2);
4628 zone
->watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + (tmp
>> 1);
4629 setup_zone_migrate_reserve(zone
);
4630 spin_unlock_irqrestore(&zone
->lock
, flags
);
4633 /* update totalreserve_pages */
4634 calculate_totalreserve_pages();
4638 * The inactive anon list should be small enough that the VM never has to
4639 * do too much work, but large enough that each inactive page has a chance
4640 * to be referenced again before it is swapped out.
4642 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
4643 * INACTIVE_ANON pages on this zone's LRU, maintained by the
4644 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
4645 * the anonymous pages are kept on the inactive list.
4648 * memory ratio inactive anon
4649 * -------------------------------------
4658 void calculate_zone_inactive_ratio(struct zone
*zone
)
4660 unsigned int gb
, ratio
;
4662 /* Zone size in gigabytes */
4663 gb
= zone
->present_pages
>> (30 - PAGE_SHIFT
);
4665 ratio
= int_sqrt(10 * gb
);
4669 zone
->inactive_ratio
= ratio
;
4672 static void __init
setup_per_zone_inactive_ratio(void)
4677 calculate_zone_inactive_ratio(zone
);
4681 * Initialise min_free_kbytes.
4683 * For small machines we want it small (128k min). For large machines
4684 * we want it large (64MB max). But it is not linear, because network
4685 * bandwidth does not increase linearly with machine size. We use
4687 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
4688 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
4704 static int __init
init_per_zone_wmark_min(void)
4706 unsigned long lowmem_kbytes
;
4708 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
4710 min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
4711 if (min_free_kbytes
< 128)
4712 min_free_kbytes
= 128;
4713 if (min_free_kbytes
> 65536)
4714 min_free_kbytes
= 65536;
4715 setup_per_zone_wmarks();
4716 setup_per_zone_lowmem_reserve();
4717 setup_per_zone_inactive_ratio();
4720 module_init(init_per_zone_wmark_min
)
4723 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
4724 * that we can call two helper functions whenever min_free_kbytes
4727 int min_free_kbytes_sysctl_handler(ctl_table
*table
, int write
,
4728 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
4730 proc_dointvec(table
, write
, buffer
, length
, ppos
);
4732 setup_per_zone_wmarks();
4737 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table
*table
, int write
,
4738 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
4743 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
4748 zone
->min_unmapped_pages
= (zone
->present_pages
*
4749 sysctl_min_unmapped_ratio
) / 100;
4753 int sysctl_min_slab_ratio_sysctl_handler(ctl_table
*table
, int write
,
4754 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
4759 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
4764 zone
->min_slab_pages
= (zone
->present_pages
*
4765 sysctl_min_slab_ratio
) / 100;
4771 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
4772 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
4773 * whenever sysctl_lowmem_reserve_ratio changes.
4775 * The reserve ratio obviously has absolutely no relation with the
4776 * minimum watermarks. The lowmem reserve ratio can only make sense
4777 * if in function of the boot time zone sizes.
4779 int lowmem_reserve_ratio_sysctl_handler(ctl_table
*table
, int write
,
4780 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
4782 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
4783 setup_per_zone_lowmem_reserve();
4788 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
4789 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
4790 * can have before it gets flushed back to buddy allocator.
4793 int percpu_pagelist_fraction_sysctl_handler(ctl_table
*table
, int write
,
4794 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
4800 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
4801 if (!write
|| (ret
== -EINVAL
))
4803 for_each_populated_zone(zone
) {
4804 for_each_online_cpu(cpu
) {
4806 high
= zone
->present_pages
/ percpu_pagelist_fraction
;
4807 setup_pagelist_highmark(zone_pcp(zone
, cpu
), high
);
4813 int hashdist
= HASHDIST_DEFAULT
;
4816 static int __init
set_hashdist(char *str
)
4820 hashdist
= simple_strtoul(str
, &str
, 0);
4823 __setup("hashdist=", set_hashdist
);
4827 * allocate a large system hash table from bootmem
4828 * - it is assumed that the hash table must contain an exact power-of-2
4829 * quantity of entries
4830 * - limit is the number of hash buckets, not the total allocation size
4832 void *__init
alloc_large_system_hash(const char *tablename
,
4833 unsigned long bucketsize
,
4834 unsigned long numentries
,
4837 unsigned int *_hash_shift
,
4838 unsigned int *_hash_mask
,
4839 unsigned long limit
)
4841 unsigned long long max
= limit
;
4842 unsigned long log2qty
, size
;
4845 /* allow the kernel cmdline to have a say */
4847 /* round applicable memory size up to nearest megabyte */
4848 numentries
= nr_kernel_pages
;
4849 numentries
+= (1UL << (20 - PAGE_SHIFT
)) - 1;
4850 numentries
>>= 20 - PAGE_SHIFT
;
4851 numentries
<<= 20 - PAGE_SHIFT
;
4853 /* limit to 1 bucket per 2^scale bytes of low memory */
4854 if (scale
> PAGE_SHIFT
)
4855 numentries
>>= (scale
- PAGE_SHIFT
);
4857 numentries
<<= (PAGE_SHIFT
- scale
);
4859 /* Make sure we've got at least a 0-order allocation.. */
4860 if (unlikely(flags
& HASH_SMALL
)) {
4861 /* Makes no sense without HASH_EARLY */
4862 WARN_ON(!(flags
& HASH_EARLY
));
4863 if (!(numentries
>> *_hash_shift
)) {
4864 numentries
= 1UL << *_hash_shift
;
4865 BUG_ON(!numentries
);
4867 } else if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
4868 numentries
= PAGE_SIZE
/ bucketsize
;
4870 numentries
= roundup_pow_of_two(numentries
);
4872 /* limit allocation size to 1/16 total memory by default */
4874 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
4875 do_div(max
, bucketsize
);
4878 if (numentries
> max
)
4881 log2qty
= ilog2(numentries
);
4884 size
= bucketsize
<< log2qty
;
4885 if (flags
& HASH_EARLY
)
4886 table
= alloc_bootmem_nopanic(size
);
4888 table
= __vmalloc(size
, GFP_ATOMIC
, PAGE_KERNEL
);
4891 * If bucketsize is not a power-of-two, we may free
4892 * some pages at the end of hash table which
4893 * alloc_pages_exact() automatically does
4895 if (get_order(size
) < MAX_ORDER
) {
4896 table
= alloc_pages_exact(size
, GFP_ATOMIC
);
4897 kmemleak_alloc(table
, size
, 1, GFP_ATOMIC
);
4900 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
4903 panic("Failed to allocate %s hash table\n", tablename
);
4905 printk(KERN_INFO
"%s hash table entries: %d (order: %d, %lu bytes)\n",
4908 ilog2(size
) - PAGE_SHIFT
,
4912 *_hash_shift
= log2qty
;
4914 *_hash_mask
= (1 << log2qty
) - 1;
4919 /* Return a pointer to the bitmap storing bits affecting a block of pages */
4920 static inline unsigned long *get_pageblock_bitmap(struct zone
*zone
,
4923 #ifdef CONFIG_SPARSEMEM
4924 return __pfn_to_section(pfn
)->pageblock_flags
;
4926 return zone
->pageblock_flags
;
4927 #endif /* CONFIG_SPARSEMEM */
4930 static inline int pfn_to_bitidx(struct zone
*zone
, unsigned long pfn
)
4932 #ifdef CONFIG_SPARSEMEM
4933 pfn
&= (PAGES_PER_SECTION
-1);
4934 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
4936 pfn
= pfn
- zone
->zone_start_pfn
;
4937 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
4938 #endif /* CONFIG_SPARSEMEM */
4942 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
4943 * @page: The page within the block of interest
4944 * @start_bitidx: The first bit of interest to retrieve
4945 * @end_bitidx: The last bit of interest
4946 * returns pageblock_bits flags
4948 unsigned long get_pageblock_flags_group(struct page
*page
,
4949 int start_bitidx
, int end_bitidx
)
4952 unsigned long *bitmap
;
4953 unsigned long pfn
, bitidx
;
4954 unsigned long flags
= 0;
4955 unsigned long value
= 1;
4957 zone
= page_zone(page
);
4958 pfn
= page_to_pfn(page
);
4959 bitmap
= get_pageblock_bitmap(zone
, pfn
);
4960 bitidx
= pfn_to_bitidx(zone
, pfn
);
4962 for (; start_bitidx
<= end_bitidx
; start_bitidx
++, value
<<= 1)
4963 if (test_bit(bitidx
+ start_bitidx
, bitmap
))
4970 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
4971 * @page: The page within the block of interest
4972 * @start_bitidx: The first bit of interest
4973 * @end_bitidx: The last bit of interest
4974 * @flags: The flags to set
4976 void set_pageblock_flags_group(struct page
*page
, unsigned long flags
,
4977 int start_bitidx
, int end_bitidx
)
4980 unsigned long *bitmap
;
4981 unsigned long pfn
, bitidx
;
4982 unsigned long value
= 1;
4984 zone
= page_zone(page
);
4985 pfn
= page_to_pfn(page
);
4986 bitmap
= get_pageblock_bitmap(zone
, pfn
);
4987 bitidx
= pfn_to_bitidx(zone
, pfn
);
4988 VM_BUG_ON(pfn
< zone
->zone_start_pfn
);
4989 VM_BUG_ON(pfn
>= zone
->zone_start_pfn
+ zone
->spanned_pages
);
4991 for (; start_bitidx
<= end_bitidx
; start_bitidx
++, value
<<= 1)
4993 __set_bit(bitidx
+ start_bitidx
, bitmap
);
4995 __clear_bit(bitidx
+ start_bitidx
, bitmap
);
4999 * This is designed as sub function...plz see page_isolation.c also.
5000 * set/clear page block's type to be ISOLATE.
5001 * page allocater never alloc memory from ISOLATE block.
5004 int set_migratetype_isolate(struct page
*page
)
5007 unsigned long flags
;
5011 zone
= page_zone(page
);
5012 zone_idx
= zone_idx(zone
);
5013 spin_lock_irqsave(&zone
->lock
, flags
);
5015 * In future, more migrate types will be able to be isolation target.
5017 if (get_pageblock_migratetype(page
) != MIGRATE_MOVABLE
&&
5018 zone_idx
!= ZONE_MOVABLE
)
5020 set_pageblock_migratetype(page
, MIGRATE_ISOLATE
);
5021 move_freepages_block(zone
, page
, MIGRATE_ISOLATE
);
5024 spin_unlock_irqrestore(&zone
->lock
, flags
);
5030 void unset_migratetype_isolate(struct page
*page
)
5033 unsigned long flags
;
5034 zone
= page_zone(page
);
5035 spin_lock_irqsave(&zone
->lock
, flags
);
5036 if (get_pageblock_migratetype(page
) != MIGRATE_ISOLATE
)
5038 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
5039 move_freepages_block(zone
, page
, MIGRATE_MOVABLE
);
5041 spin_unlock_irqrestore(&zone
->lock
, flags
);
5044 #ifdef CONFIG_MEMORY_HOTREMOVE
5046 * All pages in the range must be isolated before calling this.
5049 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
5055 unsigned long flags
;
5056 /* find the first valid pfn */
5057 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
5062 zone
= page_zone(pfn_to_page(pfn
));
5063 spin_lock_irqsave(&zone
->lock
, flags
);
5065 while (pfn
< end_pfn
) {
5066 if (!pfn_valid(pfn
)) {
5070 page
= pfn_to_page(pfn
);
5071 BUG_ON(page_count(page
));
5072 BUG_ON(!PageBuddy(page
));
5073 order
= page_order(page
);
5074 #ifdef CONFIG_DEBUG_VM
5075 printk(KERN_INFO
"remove from free list %lx %d %lx\n",
5076 pfn
, 1 << order
, end_pfn
);
5078 list_del(&page
->lru
);
5079 rmv_page_order(page
);
5080 zone
->free_area
[order
].nr_free
--;
5081 __mod_zone_page_state(zone
, NR_FREE_PAGES
,
5083 for (i
= 0; i
< (1 << order
); i
++)
5084 SetPageReserved((page
+i
));
5085 pfn
+= (1 << order
);
5087 spin_unlock_irqrestore(&zone
->lock
, flags
);