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/bootmem.h>
23 #include <linux/compiler.h>
24 #include <linux/kernel.h>
25 #include <linux/module.h>
26 #include <linux/suspend.h>
27 #include <linux/pagevec.h>
28 #include <linux/blkdev.h>
29 #include <linux/slab.h>
30 #include <linux/notifier.h>
31 #include <linux/topology.h>
32 #include <linux/sysctl.h>
33 #include <linux/cpu.h>
34 #include <linux/cpuset.h>
35 #include <linux/memory_hotplug.h>
36 #include <linux/nodemask.h>
37 #include <linux/vmalloc.h>
38 #include <linux/mempolicy.h>
39 #include <linux/stop_machine.h>
40 #include <linux/sort.h>
41 #include <linux/pfn.h>
42 #include <linux/backing-dev.h>
43 #include <linux/fault-inject.h>
45 #include <asm/tlbflush.h>
46 #include <asm/div64.h>
50 * MCD - HACK: Find somewhere to initialize this EARLY, or make this
53 nodemask_t node_online_map __read_mostly
= { { [0] = 1UL } };
54 EXPORT_SYMBOL(node_online_map
);
55 nodemask_t node_possible_map __read_mostly
= NODE_MASK_ALL
;
56 EXPORT_SYMBOL(node_possible_map
);
57 unsigned long totalram_pages __read_mostly
;
58 unsigned long totalreserve_pages __read_mostly
;
60 int percpu_pagelist_fraction
;
62 static void __free_pages_ok(struct page
*page
, unsigned int order
);
65 * results with 256, 32 in the lowmem_reserve sysctl:
66 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
67 * 1G machine -> (16M dma, 784M normal, 224M high)
68 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
69 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
70 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
72 * TBD: should special case ZONE_DMA32 machines here - in those we normally
73 * don't need any ZONE_NORMAL reservation
75 int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
-1] = {
76 #ifdef CONFIG_ZONE_DMA
79 #ifdef CONFIG_ZONE_DMA32
87 EXPORT_SYMBOL(totalram_pages
);
89 static char * const zone_names
[MAX_NR_ZONES
] = {
90 #ifdef CONFIG_ZONE_DMA
93 #ifdef CONFIG_ZONE_DMA32
102 int min_free_kbytes
= 1024;
104 unsigned long __meminitdata nr_kernel_pages
;
105 unsigned long __meminitdata nr_all_pages
;
106 static unsigned long __meminitdata dma_reserve
;
108 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
110 * MAX_ACTIVE_REGIONS determines the maxmimum number of distinct
111 * ranges of memory (RAM) that may be registered with add_active_range().
112 * Ranges passed to add_active_range() will be merged if possible
113 * so the number of times add_active_range() can be called is
114 * related to the number of nodes and the number of holes
116 #ifdef CONFIG_MAX_ACTIVE_REGIONS
117 /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */
118 #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS
120 #if MAX_NUMNODES >= 32
121 /* If there can be many nodes, allow up to 50 holes per node */
122 #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50)
124 /* By default, allow up to 256 distinct regions */
125 #define MAX_ACTIVE_REGIONS 256
129 struct node_active_region __meminitdata early_node_map
[MAX_ACTIVE_REGIONS
];
130 int __meminitdata nr_nodemap_entries
;
131 unsigned long __meminitdata arch_zone_lowest_possible_pfn
[MAX_NR_ZONES
];
132 unsigned long __meminitdata arch_zone_highest_possible_pfn
[MAX_NR_ZONES
];
133 #ifdef CONFIG_MEMORY_HOTPLUG_RESERVE
134 unsigned long __initdata node_boundary_start_pfn
[MAX_NUMNODES
];
135 unsigned long __initdata node_boundary_end_pfn
[MAX_NUMNODES
];
136 #endif /* CONFIG_MEMORY_HOTPLUG_RESERVE */
137 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
140 int nr_node_ids __read_mostly
= MAX_NUMNODES
;
141 EXPORT_SYMBOL(nr_node_ids
);
144 #ifdef CONFIG_DEBUG_VM
145 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
149 unsigned long pfn
= page_to_pfn(page
);
152 seq
= zone_span_seqbegin(zone
);
153 if (pfn
>= zone
->zone_start_pfn
+ zone
->spanned_pages
)
155 else if (pfn
< zone
->zone_start_pfn
)
157 } while (zone_span_seqretry(zone
, seq
));
162 static int page_is_consistent(struct zone
*zone
, struct page
*page
)
164 if (!pfn_valid_within(page_to_pfn(page
)))
166 if (zone
!= page_zone(page
))
172 * Temporary debugging check for pages not lying within a given zone.
174 static int bad_range(struct zone
*zone
, struct page
*page
)
176 if (page_outside_zone_boundaries(zone
, page
))
178 if (!page_is_consistent(zone
, page
))
184 static inline int bad_range(struct zone
*zone
, struct page
*page
)
190 static void bad_page(struct page
*page
)
192 printk(KERN_EMERG
"Bad page state in process '%s'\n"
193 KERN_EMERG
"page:%p flags:0x%0*lx mapping:%p mapcount:%d count:%d\n"
194 KERN_EMERG
"Trying to fix it up, but a reboot is needed\n"
195 KERN_EMERG
"Backtrace:\n",
196 current
->comm
, page
, (int)(2*sizeof(unsigned long)),
197 (unsigned long)page
->flags
, page
->mapping
,
198 page_mapcount(page
), page_count(page
));
200 page
->flags
&= ~(1 << PG_lru
|
210 set_page_count(page
, 0);
211 reset_page_mapcount(page
);
212 page
->mapping
= NULL
;
213 add_taint(TAINT_BAD_PAGE
);
217 * Higher-order pages are called "compound pages". They are structured thusly:
219 * The first PAGE_SIZE page is called the "head page".
221 * The remaining PAGE_SIZE pages are called "tail pages".
223 * All pages have PG_compound set. All pages have their ->private pointing at
224 * the head page (even the head page has this).
226 * The first tail page's ->lru.next holds the address of the compound page's
227 * put_page() function. Its ->lru.prev holds the order of allocation.
228 * This usage means that zero-order pages may not be compound.
231 static void free_compound_page(struct page
*page
)
233 __free_pages_ok(page
, compound_order(page
));
236 static void prep_compound_page(struct page
*page
, unsigned long order
)
239 int nr_pages
= 1 << order
;
241 set_compound_page_dtor(page
, free_compound_page
);
242 set_compound_order(page
, order
);
244 for (i
= 1; i
< nr_pages
; i
++) {
245 struct page
*p
= page
+ i
;
248 p
->first_page
= page
;
252 static void destroy_compound_page(struct page
*page
, unsigned long order
)
255 int nr_pages
= 1 << order
;
257 if (unlikely(compound_order(page
) != order
))
260 if (unlikely(!PageHead(page
)))
262 __ClearPageHead(page
);
263 for (i
= 1; i
< nr_pages
; i
++) {
264 struct page
*p
= page
+ i
;
266 if (unlikely(!PageTail(p
) |
267 (p
->first_page
!= page
)))
273 static inline void prep_zero_page(struct page
*page
, int order
, gfp_t gfp_flags
)
277 VM_BUG_ON((gfp_flags
& (__GFP_WAIT
| __GFP_HIGHMEM
)) == __GFP_HIGHMEM
);
279 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
280 * and __GFP_HIGHMEM from hard or soft interrupt context.
282 VM_BUG_ON((gfp_flags
& __GFP_HIGHMEM
) && in_interrupt());
283 for (i
= 0; i
< (1 << order
); i
++)
284 clear_highpage(page
+ i
);
288 * function for dealing with page's order in buddy system.
289 * zone->lock is already acquired when we use these.
290 * So, we don't need atomic page->flags operations here.
292 static inline unsigned long page_order(struct page
*page
)
294 return page_private(page
);
297 static inline void set_page_order(struct page
*page
, int order
)
299 set_page_private(page
, order
);
300 __SetPageBuddy(page
);
303 static inline void rmv_page_order(struct page
*page
)
305 __ClearPageBuddy(page
);
306 set_page_private(page
, 0);
310 * Locate the struct page for both the matching buddy in our
311 * pair (buddy1) and the combined O(n+1) page they form (page).
313 * 1) Any buddy B1 will have an order O twin B2 which satisfies
314 * the following equation:
316 * For example, if the starting buddy (buddy2) is #8 its order
318 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
320 * 2) Any buddy B will have an order O+1 parent P which
321 * satisfies the following equation:
324 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
326 static inline struct page
*
327 __page_find_buddy(struct page
*page
, unsigned long page_idx
, unsigned int order
)
329 unsigned long buddy_idx
= page_idx
^ (1 << order
);
331 return page
+ (buddy_idx
- page_idx
);
334 static inline unsigned long
335 __find_combined_index(unsigned long page_idx
, unsigned int order
)
337 return (page_idx
& ~(1 << order
));
341 * This function checks whether a page is free && is the buddy
342 * we can do coalesce a page and its buddy if
343 * (a) the buddy is not in a hole &&
344 * (b) the buddy is in the buddy system &&
345 * (c) a page and its buddy have the same order &&
346 * (d) a page and its buddy are in the same zone.
348 * For recording whether a page is in the buddy system, we use PG_buddy.
349 * Setting, clearing, and testing PG_buddy is serialized by zone->lock.
351 * For recording page's order, we use page_private(page).
353 static inline int page_is_buddy(struct page
*page
, struct page
*buddy
,
356 if (!pfn_valid_within(page_to_pfn(buddy
)))
359 if (page_zone_id(page
) != page_zone_id(buddy
))
362 if (PageBuddy(buddy
) && page_order(buddy
) == order
) {
363 BUG_ON(page_count(buddy
) != 0);
370 * Freeing function for a buddy system allocator.
372 * The concept of a buddy system is to maintain direct-mapped table
373 * (containing bit values) for memory blocks of various "orders".
374 * The bottom level table contains the map for the smallest allocatable
375 * units of memory (here, pages), and each level above it describes
376 * pairs of units from the levels below, hence, "buddies".
377 * At a high level, all that happens here is marking the table entry
378 * at the bottom level available, and propagating the changes upward
379 * as necessary, plus some accounting needed to play nicely with other
380 * parts of the VM system.
381 * At each level, we keep a list of pages, which are heads of continuous
382 * free pages of length of (1 << order) and marked with PG_buddy. Page's
383 * order is recorded in page_private(page) field.
384 * So when we are allocating or freeing one, we can derive the state of the
385 * other. That is, if we allocate a small block, and both were
386 * free, the remainder of the region must be split into blocks.
387 * If a block is freed, and its buddy is also free, then this
388 * triggers coalescing into a block of larger size.
393 static inline void __free_one_page(struct page
*page
,
394 struct zone
*zone
, unsigned int order
)
396 unsigned long page_idx
;
397 int order_size
= 1 << order
;
399 if (unlikely(PageCompound(page
)))
400 destroy_compound_page(page
, order
);
402 page_idx
= page_to_pfn(page
) & ((1 << MAX_ORDER
) - 1);
404 VM_BUG_ON(page_idx
& (order_size
- 1));
405 VM_BUG_ON(bad_range(zone
, page
));
407 __mod_zone_page_state(zone
, NR_FREE_PAGES
, order_size
);
408 while (order
< MAX_ORDER
-1) {
409 unsigned long combined_idx
;
410 struct free_area
*area
;
413 buddy
= __page_find_buddy(page
, page_idx
, order
);
414 if (!page_is_buddy(page
, buddy
, order
))
415 break; /* Move the buddy up one level. */
417 list_del(&buddy
->lru
);
418 area
= zone
->free_area
+ order
;
420 rmv_page_order(buddy
);
421 combined_idx
= __find_combined_index(page_idx
, order
);
422 page
= page
+ (combined_idx
- page_idx
);
423 page_idx
= combined_idx
;
426 set_page_order(page
, order
);
427 list_add(&page
->lru
, &zone
->free_area
[order
].free_list
);
428 zone
->free_area
[order
].nr_free
++;
431 static inline int free_pages_check(struct page
*page
)
433 if (unlikely(page_mapcount(page
) |
434 (page
->mapping
!= NULL
) |
435 (page_count(page
) != 0) |
448 * PageReclaim == PageTail. It is only an error
449 * for PageReclaim to be set if PageCompound is clear.
451 if (unlikely(!PageCompound(page
) && PageReclaim(page
)))
454 __ClearPageDirty(page
);
456 * For now, we report if PG_reserved was found set, but do not
457 * clear it, and do not free the page. But we shall soon need
458 * to do more, for when the ZERO_PAGE count wraps negative.
460 return PageReserved(page
);
464 * Frees a list of pages.
465 * Assumes all pages on list are in same zone, and of same order.
466 * count is the number of pages to free.
468 * If the zone was previously in an "all pages pinned" state then look to
469 * see if this freeing clears that state.
471 * And clear the zone's pages_scanned counter, to hold off the "all pages are
472 * pinned" detection logic.
474 static void free_pages_bulk(struct zone
*zone
, int count
,
475 struct list_head
*list
, int order
)
477 spin_lock(&zone
->lock
);
478 zone
->all_unreclaimable
= 0;
479 zone
->pages_scanned
= 0;
483 VM_BUG_ON(list_empty(list
));
484 page
= list_entry(list
->prev
, struct page
, lru
);
485 /* have to delete it as __free_one_page list manipulates */
486 list_del(&page
->lru
);
487 __free_one_page(page
, zone
, order
);
489 spin_unlock(&zone
->lock
);
492 static void free_one_page(struct zone
*zone
, struct page
*page
, int order
)
494 spin_lock(&zone
->lock
);
495 zone
->all_unreclaimable
= 0;
496 zone
->pages_scanned
= 0;
497 __free_one_page(page
, zone
, order
);
498 spin_unlock(&zone
->lock
);
501 static void __free_pages_ok(struct page
*page
, unsigned int order
)
507 for (i
= 0 ; i
< (1 << order
) ; ++i
)
508 reserved
+= free_pages_check(page
+ i
);
512 if (!PageHighMem(page
))
513 debug_check_no_locks_freed(page_address(page
),PAGE_SIZE
<<order
);
514 arch_free_page(page
, order
);
515 kernel_map_pages(page
, 1 << order
, 0);
517 local_irq_save(flags
);
518 __count_vm_events(PGFREE
, 1 << order
);
519 free_one_page(page_zone(page
), page
, order
);
520 local_irq_restore(flags
);
524 * permit the bootmem allocator to evade page validation on high-order frees
526 void fastcall __init
__free_pages_bootmem(struct page
*page
, unsigned int order
)
529 __ClearPageReserved(page
);
530 set_page_count(page
, 0);
531 set_page_refcounted(page
);
537 for (loop
= 0; loop
< BITS_PER_LONG
; loop
++) {
538 struct page
*p
= &page
[loop
];
540 if (loop
+ 1 < BITS_PER_LONG
)
542 __ClearPageReserved(p
);
543 set_page_count(p
, 0);
546 set_page_refcounted(page
);
547 __free_pages(page
, order
);
553 * The order of subdivision here is critical for the IO subsystem.
554 * Please do not alter this order without good reasons and regression
555 * testing. Specifically, as large blocks of memory are subdivided,
556 * the order in which smaller blocks are delivered depends on the order
557 * they're subdivided in this function. This is the primary factor
558 * influencing the order in which pages are delivered to the IO
559 * subsystem according to empirical testing, and this is also justified
560 * by considering the behavior of a buddy system containing a single
561 * large block of memory acted on by a series of small allocations.
562 * This behavior is a critical factor in sglist merging's success.
566 static inline void expand(struct zone
*zone
, struct page
*page
,
567 int low
, int high
, struct free_area
*area
)
569 unsigned long size
= 1 << high
;
575 VM_BUG_ON(bad_range(zone
, &page
[size
]));
576 list_add(&page
[size
].lru
, &area
->free_list
);
578 set_page_order(&page
[size
], high
);
583 * This page is about to be returned from the page allocator
585 static int prep_new_page(struct page
*page
, int order
, gfp_t gfp_flags
)
587 if (unlikely(page_mapcount(page
) |
588 (page
->mapping
!= NULL
) |
589 (page_count(page
) != 0) |
605 * For now, we report if PG_reserved was found set, but do not
606 * clear it, and do not allocate the page: as a safety net.
608 if (PageReserved(page
))
611 page
->flags
&= ~(1 << PG_uptodate
| 1 << PG_error
|
612 1 << PG_referenced
| 1 << PG_arch_1
|
613 1 << PG_owner_priv_1
| 1 << PG_mappedtodisk
);
614 set_page_private(page
, 0);
615 set_page_refcounted(page
);
617 arch_alloc_page(page
, order
);
618 kernel_map_pages(page
, 1 << order
, 1);
620 if (gfp_flags
& __GFP_ZERO
)
621 prep_zero_page(page
, order
, gfp_flags
);
623 if (order
&& (gfp_flags
& __GFP_COMP
))
624 prep_compound_page(page
, order
);
630 * Do the hard work of removing an element from the buddy allocator.
631 * Call me with the zone->lock already held.
633 static struct page
*__rmqueue(struct zone
*zone
, unsigned int order
)
635 struct free_area
* area
;
636 unsigned int current_order
;
639 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
640 area
= zone
->free_area
+ current_order
;
641 if (list_empty(&area
->free_list
))
644 page
= list_entry(area
->free_list
.next
, struct page
, lru
);
645 list_del(&page
->lru
);
646 rmv_page_order(page
);
648 __mod_zone_page_state(zone
, NR_FREE_PAGES
, - (1UL << order
));
649 expand(zone
, page
, order
, current_order
, area
);
657 * Obtain a specified number of elements from the buddy allocator, all under
658 * a single hold of the lock, for efficiency. Add them to the supplied list.
659 * Returns the number of new pages which were placed at *list.
661 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
662 unsigned long count
, struct list_head
*list
)
666 spin_lock(&zone
->lock
);
667 for (i
= 0; i
< count
; ++i
) {
668 struct page
*page
= __rmqueue(zone
, order
);
669 if (unlikely(page
== NULL
))
671 list_add_tail(&page
->lru
, list
);
673 spin_unlock(&zone
->lock
);
679 * Called from the vmstat counter updater to drain pagesets of this
680 * currently executing processor on remote nodes after they have
683 * Note that this function must be called with the thread pinned to
684 * a single processor.
686 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
691 local_irq_save(flags
);
692 if (pcp
->count
>= pcp
->batch
)
693 to_drain
= pcp
->batch
;
695 to_drain
= pcp
->count
;
696 free_pages_bulk(zone
, to_drain
, &pcp
->list
, 0);
697 pcp
->count
-= to_drain
;
698 local_irq_restore(flags
);
702 static void __drain_pages(unsigned int cpu
)
708 for_each_zone(zone
) {
709 struct per_cpu_pageset
*pset
;
711 if (!populated_zone(zone
))
714 pset
= zone_pcp(zone
, cpu
);
715 for (i
= 0; i
< ARRAY_SIZE(pset
->pcp
); i
++) {
716 struct per_cpu_pages
*pcp
;
719 local_irq_save(flags
);
720 free_pages_bulk(zone
, pcp
->count
, &pcp
->list
, 0);
722 local_irq_restore(flags
);
729 void mark_free_pages(struct zone
*zone
)
731 unsigned long pfn
, max_zone_pfn
;
734 struct list_head
*curr
;
736 if (!zone
->spanned_pages
)
739 spin_lock_irqsave(&zone
->lock
, flags
);
741 max_zone_pfn
= zone
->zone_start_pfn
+ zone
->spanned_pages
;
742 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
743 if (pfn_valid(pfn
)) {
744 struct page
*page
= pfn_to_page(pfn
);
746 if (!swsusp_page_is_forbidden(page
))
747 swsusp_unset_page_free(page
);
750 for (order
= MAX_ORDER
- 1; order
>= 0; --order
)
751 list_for_each(curr
, &zone
->free_area
[order
].free_list
) {
754 pfn
= page_to_pfn(list_entry(curr
, struct page
, lru
));
755 for (i
= 0; i
< (1UL << order
); i
++)
756 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
759 spin_unlock_irqrestore(&zone
->lock
, flags
);
763 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
765 void drain_local_pages(void)
769 local_irq_save(flags
);
770 __drain_pages(smp_processor_id());
771 local_irq_restore(flags
);
773 #endif /* CONFIG_PM */
776 * Free a 0-order page
778 static void fastcall
free_hot_cold_page(struct page
*page
, int cold
)
780 struct zone
*zone
= page_zone(page
);
781 struct per_cpu_pages
*pcp
;
785 page
->mapping
= NULL
;
786 if (free_pages_check(page
))
789 if (!PageHighMem(page
))
790 debug_check_no_locks_freed(page_address(page
), PAGE_SIZE
);
791 arch_free_page(page
, 0);
792 kernel_map_pages(page
, 1, 0);
794 pcp
= &zone_pcp(zone
, get_cpu())->pcp
[cold
];
795 local_irq_save(flags
);
796 __count_vm_event(PGFREE
);
797 list_add(&page
->lru
, &pcp
->list
);
799 if (pcp
->count
>= pcp
->high
) {
800 free_pages_bulk(zone
, pcp
->batch
, &pcp
->list
, 0);
801 pcp
->count
-= pcp
->batch
;
803 local_irq_restore(flags
);
807 void fastcall
free_hot_page(struct page
*page
)
809 free_hot_cold_page(page
, 0);
812 void fastcall
free_cold_page(struct page
*page
)
814 free_hot_cold_page(page
, 1);
818 * split_page takes a non-compound higher-order page, and splits it into
819 * n (1<<order) sub-pages: page[0..n]
820 * Each sub-page must be freed individually.
822 * Note: this is probably too low level an operation for use in drivers.
823 * Please consult with lkml before using this in your driver.
825 void split_page(struct page
*page
, unsigned int order
)
829 VM_BUG_ON(PageCompound(page
));
830 VM_BUG_ON(!page_count(page
));
831 for (i
= 1; i
< (1 << order
); i
++)
832 set_page_refcounted(page
+ i
);
836 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
837 * we cheat by calling it from here, in the order > 0 path. Saves a branch
840 static struct page
*buffered_rmqueue(struct zonelist
*zonelist
,
841 struct zone
*zone
, int order
, gfp_t gfp_flags
)
845 int cold
= !!(gfp_flags
& __GFP_COLD
);
850 if (likely(order
== 0)) {
851 struct per_cpu_pages
*pcp
;
853 pcp
= &zone_pcp(zone
, cpu
)->pcp
[cold
];
854 local_irq_save(flags
);
856 pcp
->count
= rmqueue_bulk(zone
, 0,
857 pcp
->batch
, &pcp
->list
);
858 if (unlikely(!pcp
->count
))
861 page
= list_entry(pcp
->list
.next
, struct page
, lru
);
862 list_del(&page
->lru
);
865 spin_lock_irqsave(&zone
->lock
, flags
);
866 page
= __rmqueue(zone
, order
);
867 spin_unlock(&zone
->lock
);
872 __count_zone_vm_events(PGALLOC
, zone
, 1 << order
);
873 zone_statistics(zonelist
, zone
);
874 local_irq_restore(flags
);
877 VM_BUG_ON(bad_range(zone
, page
));
878 if (prep_new_page(page
, order
, gfp_flags
))
883 local_irq_restore(flags
);
888 #define ALLOC_NO_WATERMARKS 0x01 /* don't check watermarks at all */
889 #define ALLOC_WMARK_MIN 0x02 /* use pages_min watermark */
890 #define ALLOC_WMARK_LOW 0x04 /* use pages_low watermark */
891 #define ALLOC_WMARK_HIGH 0x08 /* use pages_high watermark */
892 #define ALLOC_HARDER 0x10 /* try to alloc harder */
893 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
894 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
896 #ifdef CONFIG_FAIL_PAGE_ALLOC
898 static struct fail_page_alloc_attr
{
899 struct fault_attr attr
;
901 u32 ignore_gfp_highmem
;
904 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
906 struct dentry
*ignore_gfp_highmem_file
;
907 struct dentry
*ignore_gfp_wait_file
;
909 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
911 } fail_page_alloc
= {
912 .attr
= FAULT_ATTR_INITIALIZER
,
913 .ignore_gfp_wait
= 1,
914 .ignore_gfp_highmem
= 1,
917 static int __init
setup_fail_page_alloc(char *str
)
919 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
921 __setup("fail_page_alloc=", setup_fail_page_alloc
);
923 static int should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
925 if (gfp_mask
& __GFP_NOFAIL
)
927 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
929 if (fail_page_alloc
.ignore_gfp_wait
&& (gfp_mask
& __GFP_WAIT
))
932 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
935 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
937 static int __init
fail_page_alloc_debugfs(void)
939 mode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
943 err
= init_fault_attr_dentries(&fail_page_alloc
.attr
,
947 dir
= fail_page_alloc
.attr
.dentries
.dir
;
949 fail_page_alloc
.ignore_gfp_wait_file
=
950 debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
951 &fail_page_alloc
.ignore_gfp_wait
);
953 fail_page_alloc
.ignore_gfp_highmem_file
=
954 debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
955 &fail_page_alloc
.ignore_gfp_highmem
);
957 if (!fail_page_alloc
.ignore_gfp_wait_file
||
958 !fail_page_alloc
.ignore_gfp_highmem_file
) {
960 debugfs_remove(fail_page_alloc
.ignore_gfp_wait_file
);
961 debugfs_remove(fail_page_alloc
.ignore_gfp_highmem_file
);
962 cleanup_fault_attr_dentries(&fail_page_alloc
.attr
);
968 late_initcall(fail_page_alloc_debugfs
);
970 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
972 #else /* CONFIG_FAIL_PAGE_ALLOC */
974 static inline int should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
979 #endif /* CONFIG_FAIL_PAGE_ALLOC */
982 * Return 1 if free pages are above 'mark'. This takes into account the order
985 int zone_watermark_ok(struct zone
*z
, int order
, unsigned long mark
,
986 int classzone_idx
, int alloc_flags
)
988 /* free_pages my go negative - that's OK */
990 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
) - (1 << order
) + 1;
993 if (alloc_flags
& ALLOC_HIGH
)
995 if (alloc_flags
& ALLOC_HARDER
)
998 if (free_pages
<= min
+ z
->lowmem_reserve
[classzone_idx
])
1000 for (o
= 0; o
< order
; o
++) {
1001 /* At the next order, this order's pages become unavailable */
1002 free_pages
-= z
->free_area
[o
].nr_free
<< o
;
1004 /* Require fewer higher order pages to be free */
1007 if (free_pages
<= min
)
1015 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1016 * skip over zones that are not allowed by the cpuset, or that have
1017 * been recently (in last second) found to be nearly full. See further
1018 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1019 * that have to skip over alot of full or unallowed zones.
1021 * If the zonelist cache is present in the passed in zonelist, then
1022 * returns a pointer to the allowed node mask (either the current
1023 * tasks mems_allowed, or node_online_map.)
1025 * If the zonelist cache is not available for this zonelist, does
1026 * nothing and returns NULL.
1028 * If the fullzones BITMAP in the zonelist cache is stale (more than
1029 * a second since last zap'd) then we zap it out (clear its bits.)
1031 * We hold off even calling zlc_setup, until after we've checked the
1032 * first zone in the zonelist, on the theory that most allocations will
1033 * be satisfied from that first zone, so best to examine that zone as
1034 * quickly as we can.
1036 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1038 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1039 nodemask_t
*allowednodes
; /* zonelist_cache approximation */
1041 zlc
= zonelist
->zlcache_ptr
;
1045 if (jiffies
- zlc
->last_full_zap
> 1 * HZ
) {
1046 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
1047 zlc
->last_full_zap
= jiffies
;
1050 allowednodes
= !in_interrupt() && (alloc_flags
& ALLOC_CPUSET
) ?
1051 &cpuset_current_mems_allowed
:
1053 return allowednodes
;
1057 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1058 * if it is worth looking at further for free memory:
1059 * 1) Check that the zone isn't thought to be full (doesn't have its
1060 * bit set in the zonelist_cache fullzones BITMAP).
1061 * 2) Check that the zones node (obtained from the zonelist_cache
1062 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1063 * Return true (non-zero) if zone is worth looking at further, or
1064 * else return false (zero) if it is not.
1066 * This check -ignores- the distinction between various watermarks,
1067 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1068 * found to be full for any variation of these watermarks, it will
1069 * be considered full for up to one second by all requests, unless
1070 * we are so low on memory on all allowed nodes that we are forced
1071 * into the second scan of the zonelist.
1073 * In the second scan we ignore this zonelist cache and exactly
1074 * apply the watermarks to all zones, even it is slower to do so.
1075 * We are low on memory in the second scan, and should leave no stone
1076 * unturned looking for a free page.
1078 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zone
**z
,
1079 nodemask_t
*allowednodes
)
1081 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1082 int i
; /* index of *z in zonelist zones */
1083 int n
; /* node that zone *z is on */
1085 zlc
= zonelist
->zlcache_ptr
;
1089 i
= z
- zonelist
->zones
;
1092 /* This zone is worth trying if it is allowed but not full */
1093 return node_isset(n
, *allowednodes
) && !test_bit(i
, zlc
->fullzones
);
1097 * Given 'z' scanning a zonelist, set the corresponding bit in
1098 * zlc->fullzones, so that subsequent attempts to allocate a page
1099 * from that zone don't waste time re-examining it.
1101 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zone
**z
)
1103 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1104 int i
; /* index of *z in zonelist zones */
1106 zlc
= zonelist
->zlcache_ptr
;
1110 i
= z
- zonelist
->zones
;
1112 set_bit(i
, zlc
->fullzones
);
1115 #else /* CONFIG_NUMA */
1117 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1122 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zone
**z
,
1123 nodemask_t
*allowednodes
)
1128 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zone
**z
)
1131 #endif /* CONFIG_NUMA */
1134 * get_page_from_freelist goes through the zonelist trying to allocate
1137 static struct page
*
1138 get_page_from_freelist(gfp_t gfp_mask
, unsigned int order
,
1139 struct zonelist
*zonelist
, int alloc_flags
)
1142 struct page
*page
= NULL
;
1143 int classzone_idx
= zone_idx(zonelist
->zones
[0]);
1145 nodemask_t
*allowednodes
= NULL
;/* zonelist_cache approximation */
1146 int zlc_active
= 0; /* set if using zonelist_cache */
1147 int did_zlc_setup
= 0; /* just call zlc_setup() one time */
1151 * Scan zonelist, looking for a zone with enough free.
1152 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1154 z
= zonelist
->zones
;
1157 if (NUMA_BUILD
&& zlc_active
&&
1158 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
1161 if (unlikely(NUMA_BUILD
&& (gfp_mask
& __GFP_THISNODE
) &&
1162 zone
->zone_pgdat
!= zonelist
->zones
[0]->zone_pgdat
))
1164 if ((alloc_flags
& ALLOC_CPUSET
) &&
1165 !cpuset_zone_allowed_softwall(zone
, gfp_mask
))
1168 if (!(alloc_flags
& ALLOC_NO_WATERMARKS
)) {
1170 if (alloc_flags
& ALLOC_WMARK_MIN
)
1171 mark
= zone
->pages_min
;
1172 else if (alloc_flags
& ALLOC_WMARK_LOW
)
1173 mark
= zone
->pages_low
;
1175 mark
= zone
->pages_high
;
1176 if (!zone_watermark_ok(zone
, order
, mark
,
1177 classzone_idx
, alloc_flags
)) {
1178 if (!zone_reclaim_mode
||
1179 !zone_reclaim(zone
, gfp_mask
, order
))
1180 goto this_zone_full
;
1184 page
= buffered_rmqueue(zonelist
, zone
, order
, gfp_mask
);
1189 zlc_mark_zone_full(zonelist
, z
);
1191 if (NUMA_BUILD
&& !did_zlc_setup
) {
1192 /* we do zlc_setup after the first zone is tried */
1193 allowednodes
= zlc_setup(zonelist
, alloc_flags
);
1197 } while (*(++z
) != NULL
);
1199 if (unlikely(NUMA_BUILD
&& page
== NULL
&& zlc_active
)) {
1200 /* Disable zlc cache for second zonelist scan */
1208 * This is the 'heart' of the zoned buddy allocator.
1210 struct page
* fastcall
1211 __alloc_pages(gfp_t gfp_mask
, unsigned int order
,
1212 struct zonelist
*zonelist
)
1214 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
1217 struct reclaim_state reclaim_state
;
1218 struct task_struct
*p
= current
;
1221 int did_some_progress
;
1223 might_sleep_if(wait
);
1225 if (should_fail_alloc_page(gfp_mask
, order
))
1229 z
= zonelist
->zones
; /* the list of zones suitable for gfp_mask */
1231 if (unlikely(*z
== NULL
)) {
1232 /* Should this ever happen?? */
1236 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, order
,
1237 zonelist
, ALLOC_WMARK_LOW
|ALLOC_CPUSET
);
1242 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
1243 * __GFP_NOWARN set) should not cause reclaim since the subsystem
1244 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
1245 * using a larger set of nodes after it has established that the
1246 * allowed per node queues are empty and that nodes are
1249 if (NUMA_BUILD
&& (gfp_mask
& GFP_THISNODE
) == GFP_THISNODE
)
1252 for (z
= zonelist
->zones
; *z
; z
++)
1253 wakeup_kswapd(*z
, order
);
1256 * OK, we're below the kswapd watermark and have kicked background
1257 * reclaim. Now things get more complex, so set up alloc_flags according
1258 * to how we want to proceed.
1260 * The caller may dip into page reserves a bit more if the caller
1261 * cannot run direct reclaim, or if the caller has realtime scheduling
1262 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
1263 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
1265 alloc_flags
= ALLOC_WMARK_MIN
;
1266 if ((unlikely(rt_task(p
)) && !in_interrupt()) || !wait
)
1267 alloc_flags
|= ALLOC_HARDER
;
1268 if (gfp_mask
& __GFP_HIGH
)
1269 alloc_flags
|= ALLOC_HIGH
;
1271 alloc_flags
|= ALLOC_CPUSET
;
1274 * Go through the zonelist again. Let __GFP_HIGH and allocations
1275 * coming from realtime tasks go deeper into reserves.
1277 * This is the last chance, in general, before the goto nopage.
1278 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
1279 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1281 page
= get_page_from_freelist(gfp_mask
, order
, zonelist
, alloc_flags
);
1285 /* This allocation should allow future memory freeing. */
1288 if (((p
->flags
& PF_MEMALLOC
) || unlikely(test_thread_flag(TIF_MEMDIE
)))
1289 && !in_interrupt()) {
1290 if (!(gfp_mask
& __GFP_NOMEMALLOC
)) {
1292 /* go through the zonelist yet again, ignoring mins */
1293 page
= get_page_from_freelist(gfp_mask
, order
,
1294 zonelist
, ALLOC_NO_WATERMARKS
);
1297 if (gfp_mask
& __GFP_NOFAIL
) {
1298 congestion_wait(WRITE
, HZ
/50);
1305 /* Atomic allocations - we can't balance anything */
1311 /* We now go into synchronous reclaim */
1312 cpuset_memory_pressure_bump();
1313 p
->flags
|= PF_MEMALLOC
;
1314 reclaim_state
.reclaimed_slab
= 0;
1315 p
->reclaim_state
= &reclaim_state
;
1317 did_some_progress
= try_to_free_pages(zonelist
->zones
, gfp_mask
);
1319 p
->reclaim_state
= NULL
;
1320 p
->flags
&= ~PF_MEMALLOC
;
1324 if (likely(did_some_progress
)) {
1325 page
= get_page_from_freelist(gfp_mask
, order
,
1326 zonelist
, alloc_flags
);
1329 } else if ((gfp_mask
& __GFP_FS
) && !(gfp_mask
& __GFP_NORETRY
)) {
1331 * Go through the zonelist yet one more time, keep
1332 * very high watermark here, this is only to catch
1333 * a parallel oom killing, we must fail if we're still
1334 * under heavy pressure.
1336 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, order
,
1337 zonelist
, ALLOC_WMARK_HIGH
|ALLOC_CPUSET
);
1341 out_of_memory(zonelist
, gfp_mask
, order
);
1346 * Don't let big-order allocations loop unless the caller explicitly
1347 * requests that. Wait for some write requests to complete then retry.
1349 * In this implementation, __GFP_REPEAT means __GFP_NOFAIL for order
1350 * <= 3, but that may not be true in other implementations.
1353 if (!(gfp_mask
& __GFP_NORETRY
)) {
1354 if ((order
<= 3) || (gfp_mask
& __GFP_REPEAT
))
1356 if (gfp_mask
& __GFP_NOFAIL
)
1360 congestion_wait(WRITE
, HZ
/50);
1365 if (!(gfp_mask
& __GFP_NOWARN
) && printk_ratelimit()) {
1366 printk(KERN_WARNING
"%s: page allocation failure."
1367 " order:%d, mode:0x%x\n",
1368 p
->comm
, order
, gfp_mask
);
1376 EXPORT_SYMBOL(__alloc_pages
);
1379 * Common helper functions.
1381 fastcall
unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
1384 page
= alloc_pages(gfp_mask
, order
);
1387 return (unsigned long) page_address(page
);
1390 EXPORT_SYMBOL(__get_free_pages
);
1392 fastcall
unsigned long get_zeroed_page(gfp_t gfp_mask
)
1397 * get_zeroed_page() returns a 32-bit address, which cannot represent
1400 VM_BUG_ON((gfp_mask
& __GFP_HIGHMEM
) != 0);
1402 page
= alloc_pages(gfp_mask
| __GFP_ZERO
, 0);
1404 return (unsigned long) page_address(page
);
1408 EXPORT_SYMBOL(get_zeroed_page
);
1410 void __pagevec_free(struct pagevec
*pvec
)
1412 int i
= pagevec_count(pvec
);
1415 free_hot_cold_page(pvec
->pages
[i
], pvec
->cold
);
1418 fastcall
void __free_pages(struct page
*page
, unsigned int order
)
1420 if (put_page_testzero(page
)) {
1422 free_hot_page(page
);
1424 __free_pages_ok(page
, order
);
1428 EXPORT_SYMBOL(__free_pages
);
1430 fastcall
void free_pages(unsigned long addr
, unsigned int order
)
1433 VM_BUG_ON(!virt_addr_valid((void *)addr
));
1434 __free_pages(virt_to_page((void *)addr
), order
);
1438 EXPORT_SYMBOL(free_pages
);
1440 static unsigned int nr_free_zone_pages(int offset
)
1442 /* Just pick one node, since fallback list is circular */
1443 pg_data_t
*pgdat
= NODE_DATA(numa_node_id());
1444 unsigned int sum
= 0;
1446 struct zonelist
*zonelist
= pgdat
->node_zonelists
+ offset
;
1447 struct zone
**zonep
= zonelist
->zones
;
1450 for (zone
= *zonep
++; zone
; zone
= *zonep
++) {
1451 unsigned long size
= zone
->present_pages
;
1452 unsigned long high
= zone
->pages_high
;
1461 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
1463 unsigned int nr_free_buffer_pages(void)
1465 return nr_free_zone_pages(gfp_zone(GFP_USER
));
1469 * Amount of free RAM allocatable within all zones
1471 unsigned int nr_free_pagecache_pages(void)
1473 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER
));
1476 static inline void show_node(struct zone
*zone
)
1479 printk("Node %d ", zone_to_nid(zone
));
1482 void si_meminfo(struct sysinfo
*val
)
1484 val
->totalram
= totalram_pages
;
1486 val
->freeram
= global_page_state(NR_FREE_PAGES
);
1487 val
->bufferram
= nr_blockdev_pages();
1488 val
->totalhigh
= totalhigh_pages
;
1489 val
->freehigh
= nr_free_highpages();
1490 val
->mem_unit
= PAGE_SIZE
;
1493 EXPORT_SYMBOL(si_meminfo
);
1496 void si_meminfo_node(struct sysinfo
*val
, int nid
)
1498 pg_data_t
*pgdat
= NODE_DATA(nid
);
1500 val
->totalram
= pgdat
->node_present_pages
;
1501 val
->freeram
= node_page_state(nid
, NR_FREE_PAGES
);
1502 #ifdef CONFIG_HIGHMEM
1503 val
->totalhigh
= pgdat
->node_zones
[ZONE_HIGHMEM
].present_pages
;
1504 val
->freehigh
= zone_page_state(&pgdat
->node_zones
[ZONE_HIGHMEM
],
1510 val
->mem_unit
= PAGE_SIZE
;
1514 #define K(x) ((x) << (PAGE_SHIFT-10))
1517 * Show free area list (used inside shift_scroll-lock stuff)
1518 * We also calculate the percentage fragmentation. We do this by counting the
1519 * memory on each free list with the exception of the first item on the list.
1521 void show_free_areas(void)
1526 for_each_zone(zone
) {
1527 if (!populated_zone(zone
))
1531 printk("%s per-cpu:\n", zone
->name
);
1533 for_each_online_cpu(cpu
) {
1534 struct per_cpu_pageset
*pageset
;
1536 pageset
= zone_pcp(zone
, cpu
);
1538 printk("CPU %4d: Hot: hi:%5d, btch:%4d usd:%4d "
1539 "Cold: hi:%5d, btch:%4d usd:%4d\n",
1540 cpu
, pageset
->pcp
[0].high
,
1541 pageset
->pcp
[0].batch
, pageset
->pcp
[0].count
,
1542 pageset
->pcp
[1].high
, pageset
->pcp
[1].batch
,
1543 pageset
->pcp
[1].count
);
1547 printk("Active:%lu inactive:%lu dirty:%lu writeback:%lu unstable:%lu\n"
1548 " free:%lu slab:%lu mapped:%lu pagetables:%lu bounce:%lu\n",
1549 global_page_state(NR_ACTIVE
),
1550 global_page_state(NR_INACTIVE
),
1551 global_page_state(NR_FILE_DIRTY
),
1552 global_page_state(NR_WRITEBACK
),
1553 global_page_state(NR_UNSTABLE_NFS
),
1554 global_page_state(NR_FREE_PAGES
),
1555 global_page_state(NR_SLAB_RECLAIMABLE
) +
1556 global_page_state(NR_SLAB_UNRECLAIMABLE
),
1557 global_page_state(NR_FILE_MAPPED
),
1558 global_page_state(NR_PAGETABLE
),
1559 global_page_state(NR_BOUNCE
));
1561 for_each_zone(zone
) {
1564 if (!populated_zone(zone
))
1576 " pages_scanned:%lu"
1577 " all_unreclaimable? %s"
1580 K(zone_page_state(zone
, NR_FREE_PAGES
)),
1583 K(zone
->pages_high
),
1584 K(zone_page_state(zone
, NR_ACTIVE
)),
1585 K(zone_page_state(zone
, NR_INACTIVE
)),
1586 K(zone
->present_pages
),
1587 zone
->pages_scanned
,
1588 (zone
->all_unreclaimable
? "yes" : "no")
1590 printk("lowmem_reserve[]:");
1591 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
1592 printk(" %lu", zone
->lowmem_reserve
[i
]);
1596 for_each_zone(zone
) {
1597 unsigned long nr
[MAX_ORDER
], flags
, order
, total
= 0;
1599 if (!populated_zone(zone
))
1603 printk("%s: ", zone
->name
);
1605 spin_lock_irqsave(&zone
->lock
, flags
);
1606 for (order
= 0; order
< MAX_ORDER
; order
++) {
1607 nr
[order
] = zone
->free_area
[order
].nr_free
;
1608 total
+= nr
[order
] << order
;
1610 spin_unlock_irqrestore(&zone
->lock
, flags
);
1611 for (order
= 0; order
< MAX_ORDER
; order
++)
1612 printk("%lu*%lukB ", nr
[order
], K(1UL) << order
);
1613 printk("= %lukB\n", K(total
));
1616 show_swap_cache_info();
1620 * Builds allocation fallback zone lists.
1622 * Add all populated zones of a node to the zonelist.
1624 static int build_zonelists_node(pg_data_t
*pgdat
, struct zonelist
*zonelist
,
1625 int nr_zones
, enum zone_type zone_type
)
1629 BUG_ON(zone_type
>= MAX_NR_ZONES
);
1634 zone
= pgdat
->node_zones
+ zone_type
;
1635 if (populated_zone(zone
)) {
1636 zonelist
->zones
[nr_zones
++] = zone
;
1637 check_highest_zone(zone_type
);
1640 } while (zone_type
);
1647 * 0 = automatic detection of better ordering.
1648 * 1 = order by ([node] distance, -zonetype)
1649 * 2 = order by (-zonetype, [node] distance)
1651 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
1652 * the same zonelist. So only NUMA can configure this param.
1654 #define ZONELIST_ORDER_DEFAULT 0
1655 #define ZONELIST_ORDER_NODE 1
1656 #define ZONELIST_ORDER_ZONE 2
1658 /* zonelist order in the kernel.
1659 * set_zonelist_order() will set this to NODE or ZONE.
1661 static int current_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
1662 static char zonelist_order_name
[3][8] = {"Default", "Node", "Zone"};
1666 /* The value user specified ....changed by config */
1667 static int user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
1668 /* string for sysctl */
1669 #define NUMA_ZONELIST_ORDER_LEN 16
1670 char numa_zonelist_order
[16] = "default";
1673 * interface for configure zonelist ordering.
1674 * command line option "numa_zonelist_order"
1675 * = "[dD]efault - default, automatic configuration.
1676 * = "[nN]ode - order by node locality, then by zone within node
1677 * = "[zZ]one - order by zone, then by locality within zone
1680 static int __parse_numa_zonelist_order(char *s
)
1682 if (*s
== 'd' || *s
== 'D') {
1683 user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
1684 } else if (*s
== 'n' || *s
== 'N') {
1685 user_zonelist_order
= ZONELIST_ORDER_NODE
;
1686 } else if (*s
== 'z' || *s
== 'Z') {
1687 user_zonelist_order
= ZONELIST_ORDER_ZONE
;
1690 "Ignoring invalid numa_zonelist_order value: "
1697 static __init
int setup_numa_zonelist_order(char *s
)
1700 return __parse_numa_zonelist_order(s
);
1703 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
1706 * sysctl handler for numa_zonelist_order
1708 int numa_zonelist_order_handler(ctl_table
*table
, int write
,
1709 struct file
*file
, void __user
*buffer
, size_t *length
,
1712 char saved_string
[NUMA_ZONELIST_ORDER_LEN
];
1716 strncpy(saved_string
, (char*)table
->data
,
1717 NUMA_ZONELIST_ORDER_LEN
);
1718 ret
= proc_dostring(table
, write
, file
, buffer
, length
, ppos
);
1722 int oldval
= user_zonelist_order
;
1723 if (__parse_numa_zonelist_order((char*)table
->data
)) {
1725 * bogus value. restore saved string
1727 strncpy((char*)table
->data
, saved_string
,
1728 NUMA_ZONELIST_ORDER_LEN
);
1729 user_zonelist_order
= oldval
;
1730 } else if (oldval
!= user_zonelist_order
)
1731 build_all_zonelists();
1737 #define MAX_NODE_LOAD (num_online_nodes())
1738 static int node_load
[MAX_NUMNODES
];
1741 * find_next_best_node - find the next node that should appear in a given node's fallback list
1742 * @node: node whose fallback list we're appending
1743 * @used_node_mask: nodemask_t of already used nodes
1745 * We use a number of factors to determine which is the next node that should
1746 * appear on a given node's fallback list. The node should not have appeared
1747 * already in @node's fallback list, and it should be the next closest node
1748 * according to the distance array (which contains arbitrary distance values
1749 * from each node to each node in the system), and should also prefer nodes
1750 * with no CPUs, since presumably they'll have very little allocation pressure
1751 * on them otherwise.
1752 * It returns -1 if no node is found.
1754 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
1757 int min_val
= INT_MAX
;
1760 /* Use the local node if we haven't already */
1761 if (!node_isset(node
, *used_node_mask
)) {
1762 node_set(node
, *used_node_mask
);
1766 for_each_online_node(n
) {
1769 /* Don't want a node to appear more than once */
1770 if (node_isset(n
, *used_node_mask
))
1773 /* Use the distance array to find the distance */
1774 val
= node_distance(node
, n
);
1776 /* Penalize nodes under us ("prefer the next node") */
1779 /* Give preference to headless and unused nodes */
1780 tmp
= node_to_cpumask(n
);
1781 if (!cpus_empty(tmp
))
1782 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
1784 /* Slight preference for less loaded node */
1785 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
1786 val
+= node_load
[n
];
1788 if (val
< min_val
) {
1795 node_set(best_node
, *used_node_mask
);
1802 * Build zonelists ordered by node and zones within node.
1803 * This results in maximum locality--normal zone overflows into local
1804 * DMA zone, if any--but risks exhausting DMA zone.
1806 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int node
)
1810 struct zonelist
*zonelist
;
1812 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
1813 zonelist
= pgdat
->node_zonelists
+ i
;
1814 for (j
= 0; zonelist
->zones
[j
] != NULL
; j
++)
1816 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
, i
);
1817 zonelist
->zones
[j
] = NULL
;
1822 * Build zonelists ordered by zone and nodes within zones.
1823 * This results in conserving DMA zone[s] until all Normal memory is
1824 * exhausted, but results in overflowing to remote node while memory
1825 * may still exist in local DMA zone.
1827 static int node_order
[MAX_NUMNODES
];
1829 static void build_zonelists_in_zone_order(pg_data_t
*pgdat
, int nr_nodes
)
1833 int zone_type
; /* needs to be signed */
1835 struct zonelist
*zonelist
;
1837 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
1838 zonelist
= pgdat
->node_zonelists
+ i
;
1840 for (zone_type
= i
; zone_type
>= 0; zone_type
--) {
1841 for (j
= 0; j
< nr_nodes
; j
++) {
1842 node
= node_order
[j
];
1843 z
= &NODE_DATA(node
)->node_zones
[zone_type
];
1844 if (populated_zone(z
)) {
1845 zonelist
->zones
[pos
++] = z
;
1846 check_highest_zone(zone_type
);
1850 zonelist
->zones
[pos
] = NULL
;
1854 static int default_zonelist_order(void)
1857 unsigned long low_kmem_size
,total_size
;
1861 * ZONE_DMA and ZONE_DMA32 can be very small area in the sytem.
1862 * If they are really small and used heavily, the system can fall
1863 * into OOM very easily.
1864 * This function detect ZONE_DMA/DMA32 size and confgigures zone order.
1866 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
1869 for_each_online_node(nid
) {
1870 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
1871 z
= &NODE_DATA(nid
)->node_zones
[zone_type
];
1872 if (populated_zone(z
)) {
1873 if (zone_type
< ZONE_NORMAL
)
1874 low_kmem_size
+= z
->present_pages
;
1875 total_size
+= z
->present_pages
;
1879 if (!low_kmem_size
|| /* there are no DMA area. */
1880 low_kmem_size
> total_size
/2) /* DMA/DMA32 is big. */
1881 return ZONELIST_ORDER_NODE
;
1883 * look into each node's config.
1884 * If there is a node whose DMA/DMA32 memory is very big area on
1885 * local memory, NODE_ORDER may be suitable.
1887 average_size
= total_size
/ (num_online_nodes() + 1);
1888 for_each_online_node(nid
) {
1891 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
1892 z
= &NODE_DATA(nid
)->node_zones
[zone_type
];
1893 if (populated_zone(z
)) {
1894 if (zone_type
< ZONE_NORMAL
)
1895 low_kmem_size
+= z
->present_pages
;
1896 total_size
+= z
->present_pages
;
1899 if (low_kmem_size
&&
1900 total_size
> average_size
&& /* ignore small node */
1901 low_kmem_size
> total_size
* 70/100)
1902 return ZONELIST_ORDER_NODE
;
1904 return ZONELIST_ORDER_ZONE
;
1907 static void set_zonelist_order(void)
1909 if (user_zonelist_order
== ZONELIST_ORDER_DEFAULT
)
1910 current_zonelist_order
= default_zonelist_order();
1912 current_zonelist_order
= user_zonelist_order
;
1915 static void build_zonelists(pg_data_t
*pgdat
)
1919 nodemask_t used_mask
;
1920 int local_node
, prev_node
;
1921 struct zonelist
*zonelist
;
1922 int order
= current_zonelist_order
;
1924 /* initialize zonelists */
1925 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
1926 zonelist
= pgdat
->node_zonelists
+ i
;
1927 zonelist
->zones
[0] = NULL
;
1930 /* NUMA-aware ordering of nodes */
1931 local_node
= pgdat
->node_id
;
1932 load
= num_online_nodes();
1933 prev_node
= local_node
;
1934 nodes_clear(used_mask
);
1936 memset(node_load
, 0, sizeof(node_load
));
1937 memset(node_order
, 0, sizeof(node_order
));
1940 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
1941 int distance
= node_distance(local_node
, node
);
1944 * If another node is sufficiently far away then it is better
1945 * to reclaim pages in a zone before going off node.
1947 if (distance
> RECLAIM_DISTANCE
)
1948 zone_reclaim_mode
= 1;
1951 * We don't want to pressure a particular node.
1952 * So adding penalty to the first node in same
1953 * distance group to make it round-robin.
1955 if (distance
!= node_distance(local_node
, prev_node
))
1956 node_load
[node
] = load
;
1960 if (order
== ZONELIST_ORDER_NODE
)
1961 build_zonelists_in_node_order(pgdat
, node
);
1963 node_order
[j
++] = node
; /* remember order */
1966 if (order
== ZONELIST_ORDER_ZONE
) {
1967 /* calculate node order -- i.e., DMA last! */
1968 build_zonelists_in_zone_order(pgdat
, j
);
1972 /* Construct the zonelist performance cache - see further mmzone.h */
1973 static void build_zonelist_cache(pg_data_t
*pgdat
)
1977 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
1978 struct zonelist
*zonelist
;
1979 struct zonelist_cache
*zlc
;
1982 zonelist
= pgdat
->node_zonelists
+ i
;
1983 zonelist
->zlcache_ptr
= zlc
= &zonelist
->zlcache
;
1984 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
1985 for (z
= zonelist
->zones
; *z
; z
++)
1986 zlc
->z_to_n
[z
- zonelist
->zones
] = zone_to_nid(*z
);
1991 #else /* CONFIG_NUMA */
1993 static void set_zonelist_order(void)
1995 current_zonelist_order
= ZONELIST_ORDER_ZONE
;
1998 static void build_zonelists(pg_data_t
*pgdat
)
2000 int node
, local_node
;
2003 local_node
= pgdat
->node_id
;
2004 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
2005 struct zonelist
*zonelist
;
2007 zonelist
= pgdat
->node_zonelists
+ i
;
2009 j
= build_zonelists_node(pgdat
, zonelist
, 0, i
);
2011 * Now we build the zonelist so that it contains the zones
2012 * of all the other nodes.
2013 * We don't want to pressure a particular node, so when
2014 * building the zones for node N, we make sure that the
2015 * zones coming right after the local ones are those from
2016 * node N+1 (modulo N)
2018 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
2019 if (!node_online(node
))
2021 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
, i
);
2023 for (node
= 0; node
< local_node
; node
++) {
2024 if (!node_online(node
))
2026 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
, i
);
2029 zonelist
->zones
[j
] = NULL
;
2033 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
2034 static void build_zonelist_cache(pg_data_t
*pgdat
)
2038 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
2039 pgdat
->node_zonelists
[i
].zlcache_ptr
= NULL
;
2042 #endif /* CONFIG_NUMA */
2044 /* return values int ....just for stop_machine_run() */
2045 static int __build_all_zonelists(void *dummy
)
2049 for_each_online_node(nid
) {
2050 build_zonelists(NODE_DATA(nid
));
2051 build_zonelist_cache(NODE_DATA(nid
));
2056 void build_all_zonelists(void)
2058 set_zonelist_order();
2060 if (system_state
== SYSTEM_BOOTING
) {
2061 __build_all_zonelists(NULL
);
2062 cpuset_init_current_mems_allowed();
2064 /* we have to stop all cpus to guaranntee there is no user
2066 stop_machine_run(__build_all_zonelists
, NULL
, NR_CPUS
);
2067 /* cpuset refresh routine should be here */
2069 vm_total_pages
= nr_free_pagecache_pages();
2070 printk("Built %i zonelists in %s order. Total pages: %ld\n",
2072 zonelist_order_name
[current_zonelist_order
],
2075 printk("Policy zone: %s\n", zone_names
[policy_zone
]);
2080 * Helper functions to size the waitqueue hash table.
2081 * Essentially these want to choose hash table sizes sufficiently
2082 * large so that collisions trying to wait on pages are rare.
2083 * But in fact, the number of active page waitqueues on typical
2084 * systems is ridiculously low, less than 200. So this is even
2085 * conservative, even though it seems large.
2087 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
2088 * waitqueues, i.e. the size of the waitq table given the number of pages.
2090 #define PAGES_PER_WAITQUEUE 256
2092 #ifndef CONFIG_MEMORY_HOTPLUG
2093 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
2095 unsigned long size
= 1;
2097 pages
/= PAGES_PER_WAITQUEUE
;
2099 while (size
< pages
)
2103 * Once we have dozens or even hundreds of threads sleeping
2104 * on IO we've got bigger problems than wait queue collision.
2105 * Limit the size of the wait table to a reasonable size.
2107 size
= min(size
, 4096UL);
2109 return max(size
, 4UL);
2113 * A zone's size might be changed by hot-add, so it is not possible to determine
2114 * a suitable size for its wait_table. So we use the maximum size now.
2116 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
2118 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
2119 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
2120 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
2122 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
2123 * or more by the traditional way. (See above). It equals:
2125 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
2126 * ia64(16K page size) : = ( 8G + 4M)byte.
2127 * powerpc (64K page size) : = (32G +16M)byte.
2129 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
2136 * This is an integer logarithm so that shifts can be used later
2137 * to extract the more random high bits from the multiplicative
2138 * hash function before the remainder is taken.
2140 static inline unsigned long wait_table_bits(unsigned long size
)
2145 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
2148 * Initially all pages are reserved - free ones are freed
2149 * up by free_all_bootmem() once the early boot process is
2150 * done. Non-atomic initialization, single-pass.
2152 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
2153 unsigned long start_pfn
, enum memmap_context context
)
2156 unsigned long end_pfn
= start_pfn
+ size
;
2159 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
2161 * There can be holes in boot-time mem_map[]s
2162 * handed to this function. They do not
2163 * exist on hotplugged memory.
2165 if (context
== MEMMAP_EARLY
) {
2166 if (!early_pfn_valid(pfn
))
2168 if (!early_pfn_in_nid(pfn
, nid
))
2171 page
= pfn_to_page(pfn
);
2172 set_page_links(page
, zone
, nid
, pfn
);
2173 init_page_count(page
);
2174 reset_page_mapcount(page
);
2175 SetPageReserved(page
);
2176 INIT_LIST_HEAD(&page
->lru
);
2177 #ifdef WANT_PAGE_VIRTUAL
2178 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
2179 if (!is_highmem_idx(zone
))
2180 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
2185 void zone_init_free_lists(struct pglist_data
*pgdat
, struct zone
*zone
,
2189 for (order
= 0; order
< MAX_ORDER
; order
++) {
2190 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
);
2191 zone
->free_area
[order
].nr_free
= 0;
2195 #ifndef __HAVE_ARCH_MEMMAP_INIT
2196 #define memmap_init(size, nid, zone, start_pfn) \
2197 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
2200 static int __devinit
zone_batchsize(struct zone
*zone
)
2205 * The per-cpu-pages pools are set to around 1000th of the
2206 * size of the zone. But no more than 1/2 of a meg.
2208 * OK, so we don't know how big the cache is. So guess.
2210 batch
= zone
->present_pages
/ 1024;
2211 if (batch
* PAGE_SIZE
> 512 * 1024)
2212 batch
= (512 * 1024) / PAGE_SIZE
;
2213 batch
/= 4; /* We effectively *= 4 below */
2218 * Clamp the batch to a 2^n - 1 value. Having a power
2219 * of 2 value was found to be more likely to have
2220 * suboptimal cache aliasing properties in some cases.
2222 * For example if 2 tasks are alternately allocating
2223 * batches of pages, one task can end up with a lot
2224 * of pages of one half of the possible page colors
2225 * and the other with pages of the other colors.
2227 batch
= (1 << (fls(batch
+ batch
/2)-1)) - 1;
2232 inline void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
2234 struct per_cpu_pages
*pcp
;
2236 memset(p
, 0, sizeof(*p
));
2238 pcp
= &p
->pcp
[0]; /* hot */
2240 pcp
->high
= 6 * batch
;
2241 pcp
->batch
= max(1UL, 1 * batch
);
2242 INIT_LIST_HEAD(&pcp
->list
);
2244 pcp
= &p
->pcp
[1]; /* cold*/
2246 pcp
->high
= 2 * batch
;
2247 pcp
->batch
= max(1UL, batch
/2);
2248 INIT_LIST_HEAD(&pcp
->list
);
2252 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
2253 * to the value high for the pageset p.
2256 static void setup_pagelist_highmark(struct per_cpu_pageset
*p
,
2259 struct per_cpu_pages
*pcp
;
2261 pcp
= &p
->pcp
[0]; /* hot list */
2263 pcp
->batch
= max(1UL, high
/4);
2264 if ((high
/4) > (PAGE_SHIFT
* 8))
2265 pcp
->batch
= PAGE_SHIFT
* 8;
2271 * Boot pageset table. One per cpu which is going to be used for all
2272 * zones and all nodes. The parameters will be set in such a way
2273 * that an item put on a list will immediately be handed over to
2274 * the buddy list. This is safe since pageset manipulation is done
2275 * with interrupts disabled.
2277 * Some NUMA counter updates may also be caught by the boot pagesets.
2279 * The boot_pagesets must be kept even after bootup is complete for
2280 * unused processors and/or zones. They do play a role for bootstrapping
2281 * hotplugged processors.
2283 * zoneinfo_show() and maybe other functions do
2284 * not check if the processor is online before following the pageset pointer.
2285 * Other parts of the kernel may not check if the zone is available.
2287 static struct per_cpu_pageset boot_pageset
[NR_CPUS
];
2290 * Dynamically allocate memory for the
2291 * per cpu pageset array in struct zone.
2293 static int __cpuinit
process_zones(int cpu
)
2295 struct zone
*zone
, *dzone
;
2297 for_each_zone(zone
) {
2299 if (!populated_zone(zone
))
2302 zone_pcp(zone
, cpu
) = kmalloc_node(sizeof(struct per_cpu_pageset
),
2303 GFP_KERNEL
, cpu_to_node(cpu
));
2304 if (!zone_pcp(zone
, cpu
))
2307 setup_pageset(zone_pcp(zone
, cpu
), zone_batchsize(zone
));
2309 if (percpu_pagelist_fraction
)
2310 setup_pagelist_highmark(zone_pcp(zone
, cpu
),
2311 (zone
->present_pages
/ percpu_pagelist_fraction
));
2316 for_each_zone(dzone
) {
2319 kfree(zone_pcp(dzone
, cpu
));
2320 zone_pcp(dzone
, cpu
) = NULL
;
2325 static inline void free_zone_pagesets(int cpu
)
2329 for_each_zone(zone
) {
2330 struct per_cpu_pageset
*pset
= zone_pcp(zone
, cpu
);
2332 /* Free per_cpu_pageset if it is slab allocated */
2333 if (pset
!= &boot_pageset
[cpu
])
2335 zone_pcp(zone
, cpu
) = NULL
;
2339 static int __cpuinit
pageset_cpuup_callback(struct notifier_block
*nfb
,
2340 unsigned long action
,
2343 int cpu
= (long)hcpu
;
2344 int ret
= NOTIFY_OK
;
2347 case CPU_UP_PREPARE
:
2348 case CPU_UP_PREPARE_FROZEN
:
2349 if (process_zones(cpu
))
2352 case CPU_UP_CANCELED
:
2353 case CPU_UP_CANCELED_FROZEN
:
2355 case CPU_DEAD_FROZEN
:
2356 free_zone_pagesets(cpu
);
2364 static struct notifier_block __cpuinitdata pageset_notifier
=
2365 { &pageset_cpuup_callback
, NULL
, 0 };
2367 void __init
setup_per_cpu_pageset(void)
2371 /* Initialize per_cpu_pageset for cpu 0.
2372 * A cpuup callback will do this for every cpu
2373 * as it comes online
2375 err
= process_zones(smp_processor_id());
2377 register_cpu_notifier(&pageset_notifier
);
2382 static noinline __init_refok
2383 int zone_wait_table_init(struct zone
*zone
, unsigned long zone_size_pages
)
2386 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
2390 * The per-page waitqueue mechanism uses hashed waitqueues
2393 zone
->wait_table_hash_nr_entries
=
2394 wait_table_hash_nr_entries(zone_size_pages
);
2395 zone
->wait_table_bits
=
2396 wait_table_bits(zone
->wait_table_hash_nr_entries
);
2397 alloc_size
= zone
->wait_table_hash_nr_entries
2398 * sizeof(wait_queue_head_t
);
2400 if (system_state
== SYSTEM_BOOTING
) {
2401 zone
->wait_table
= (wait_queue_head_t
*)
2402 alloc_bootmem_node(pgdat
, alloc_size
);
2405 * This case means that a zone whose size was 0 gets new memory
2406 * via memory hot-add.
2407 * But it may be the case that a new node was hot-added. In
2408 * this case vmalloc() will not be able to use this new node's
2409 * memory - this wait_table must be initialized to use this new
2410 * node itself as well.
2411 * To use this new node's memory, further consideration will be
2414 zone
->wait_table
= (wait_queue_head_t
*)vmalloc(alloc_size
);
2416 if (!zone
->wait_table
)
2419 for(i
= 0; i
< zone
->wait_table_hash_nr_entries
; ++i
)
2420 init_waitqueue_head(zone
->wait_table
+ i
);
2425 static __meminit
void zone_pcp_init(struct zone
*zone
)
2428 unsigned long batch
= zone_batchsize(zone
);
2430 for (cpu
= 0; cpu
< NR_CPUS
; cpu
++) {
2432 /* Early boot. Slab allocator not functional yet */
2433 zone_pcp(zone
, cpu
) = &boot_pageset
[cpu
];
2434 setup_pageset(&boot_pageset
[cpu
],0);
2436 setup_pageset(zone_pcp(zone
,cpu
), batch
);
2439 if (zone
->present_pages
)
2440 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%lu\n",
2441 zone
->name
, zone
->present_pages
, batch
);
2444 __meminit
int init_currently_empty_zone(struct zone
*zone
,
2445 unsigned long zone_start_pfn
,
2447 enum memmap_context context
)
2449 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
2451 ret
= zone_wait_table_init(zone
, size
);
2454 pgdat
->nr_zones
= zone_idx(zone
) + 1;
2456 zone
->zone_start_pfn
= zone_start_pfn
;
2458 memmap_init(size
, pgdat
->node_id
, zone_idx(zone
), zone_start_pfn
);
2460 zone_init_free_lists(pgdat
, zone
, zone
->spanned_pages
);
2465 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
2467 * Basic iterator support. Return the first range of PFNs for a node
2468 * Note: nid == MAX_NUMNODES returns first region regardless of node
2470 static int __meminit
first_active_region_index_in_nid(int nid
)
2474 for (i
= 0; i
< nr_nodemap_entries
; i
++)
2475 if (nid
== MAX_NUMNODES
|| early_node_map
[i
].nid
== nid
)
2482 * Basic iterator support. Return the next active range of PFNs for a node
2483 * Note: nid == MAX_NUMNODES returns next region regardles of node
2485 static int __meminit
next_active_region_index_in_nid(int index
, int nid
)
2487 for (index
= index
+ 1; index
< nr_nodemap_entries
; index
++)
2488 if (nid
== MAX_NUMNODES
|| early_node_map
[index
].nid
== nid
)
2494 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
2496 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
2497 * Architectures may implement their own version but if add_active_range()
2498 * was used and there are no special requirements, this is a convenient
2501 int __meminit
early_pfn_to_nid(unsigned long pfn
)
2505 for (i
= 0; i
< nr_nodemap_entries
; i
++) {
2506 unsigned long start_pfn
= early_node_map
[i
].start_pfn
;
2507 unsigned long end_pfn
= early_node_map
[i
].end_pfn
;
2509 if (start_pfn
<= pfn
&& pfn
< end_pfn
)
2510 return early_node_map
[i
].nid
;
2515 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
2517 /* Basic iterator support to walk early_node_map[] */
2518 #define for_each_active_range_index_in_nid(i, nid) \
2519 for (i = first_active_region_index_in_nid(nid); i != -1; \
2520 i = next_active_region_index_in_nid(i, nid))
2523 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
2524 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
2525 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
2527 * If an architecture guarantees that all ranges registered with
2528 * add_active_ranges() contain no holes and may be freed, this
2529 * this function may be used instead of calling free_bootmem() manually.
2531 void __init
free_bootmem_with_active_regions(int nid
,
2532 unsigned long max_low_pfn
)
2536 for_each_active_range_index_in_nid(i
, nid
) {
2537 unsigned long size_pages
= 0;
2538 unsigned long end_pfn
= early_node_map
[i
].end_pfn
;
2540 if (early_node_map
[i
].start_pfn
>= max_low_pfn
)
2543 if (end_pfn
> max_low_pfn
)
2544 end_pfn
= max_low_pfn
;
2546 size_pages
= end_pfn
- early_node_map
[i
].start_pfn
;
2547 free_bootmem_node(NODE_DATA(early_node_map
[i
].nid
),
2548 PFN_PHYS(early_node_map
[i
].start_pfn
),
2549 size_pages
<< PAGE_SHIFT
);
2554 * sparse_memory_present_with_active_regions - Call memory_present for each active range
2555 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
2557 * If an architecture guarantees that all ranges registered with
2558 * add_active_ranges() contain no holes and may be freed, this
2559 * function may be used instead of calling memory_present() manually.
2561 void __init
sparse_memory_present_with_active_regions(int nid
)
2565 for_each_active_range_index_in_nid(i
, nid
)
2566 memory_present(early_node_map
[i
].nid
,
2567 early_node_map
[i
].start_pfn
,
2568 early_node_map
[i
].end_pfn
);
2572 * push_node_boundaries - Push node boundaries to at least the requested boundary
2573 * @nid: The nid of the node to push the boundary for
2574 * @start_pfn: The start pfn of the node
2575 * @end_pfn: The end pfn of the node
2577 * In reserve-based hot-add, mem_map is allocated that is unused until hotadd
2578 * time. Specifically, on x86_64, SRAT will report ranges that can potentially
2579 * be hotplugged even though no physical memory exists. This function allows
2580 * an arch to push out the node boundaries so mem_map is allocated that can
2583 #ifdef CONFIG_MEMORY_HOTPLUG_RESERVE
2584 void __init
push_node_boundaries(unsigned int nid
,
2585 unsigned long start_pfn
, unsigned long end_pfn
)
2587 printk(KERN_DEBUG
"Entering push_node_boundaries(%u, %lu, %lu)\n",
2588 nid
, start_pfn
, end_pfn
);
2590 /* Initialise the boundary for this node if necessary */
2591 if (node_boundary_end_pfn
[nid
] == 0)
2592 node_boundary_start_pfn
[nid
] = -1UL;
2594 /* Update the boundaries */
2595 if (node_boundary_start_pfn
[nid
] > start_pfn
)
2596 node_boundary_start_pfn
[nid
] = start_pfn
;
2597 if (node_boundary_end_pfn
[nid
] < end_pfn
)
2598 node_boundary_end_pfn
[nid
] = end_pfn
;
2601 /* If necessary, push the node boundary out for reserve hotadd */
2602 static void __init
account_node_boundary(unsigned int nid
,
2603 unsigned long *start_pfn
, unsigned long *end_pfn
)
2605 printk(KERN_DEBUG
"Entering account_node_boundary(%u, %lu, %lu)\n",
2606 nid
, *start_pfn
, *end_pfn
);
2608 /* Return if boundary information has not been provided */
2609 if (node_boundary_end_pfn
[nid
] == 0)
2612 /* Check the boundaries and update if necessary */
2613 if (node_boundary_start_pfn
[nid
] < *start_pfn
)
2614 *start_pfn
= node_boundary_start_pfn
[nid
];
2615 if (node_boundary_end_pfn
[nid
] > *end_pfn
)
2616 *end_pfn
= node_boundary_end_pfn
[nid
];
2619 void __init
push_node_boundaries(unsigned int nid
,
2620 unsigned long start_pfn
, unsigned long end_pfn
) {}
2622 static void __init
account_node_boundary(unsigned int nid
,
2623 unsigned long *start_pfn
, unsigned long *end_pfn
) {}
2628 * get_pfn_range_for_nid - Return the start and end page frames for a node
2629 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
2630 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
2631 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
2633 * It returns the start and end page frame of a node based on information
2634 * provided by an arch calling add_active_range(). If called for a node
2635 * with no available memory, a warning is printed and the start and end
2638 void __meminit
get_pfn_range_for_nid(unsigned int nid
,
2639 unsigned long *start_pfn
, unsigned long *end_pfn
)
2645 for_each_active_range_index_in_nid(i
, nid
) {
2646 *start_pfn
= min(*start_pfn
, early_node_map
[i
].start_pfn
);
2647 *end_pfn
= max(*end_pfn
, early_node_map
[i
].end_pfn
);
2650 if (*start_pfn
== -1UL) {
2651 printk(KERN_WARNING
"Node %u active with no memory\n", nid
);
2655 /* Push the node boundaries out if requested */
2656 account_node_boundary(nid
, start_pfn
, end_pfn
);
2660 * Return the number of pages a zone spans in a node, including holes
2661 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
2663 unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
2664 unsigned long zone_type
,
2665 unsigned long *ignored
)
2667 unsigned long node_start_pfn
, node_end_pfn
;
2668 unsigned long zone_start_pfn
, zone_end_pfn
;
2670 /* Get the start and end of the node and zone */
2671 get_pfn_range_for_nid(nid
, &node_start_pfn
, &node_end_pfn
);
2672 zone_start_pfn
= arch_zone_lowest_possible_pfn
[zone_type
];
2673 zone_end_pfn
= arch_zone_highest_possible_pfn
[zone_type
];
2675 /* Check that this node has pages within the zone's required range */
2676 if (zone_end_pfn
< node_start_pfn
|| zone_start_pfn
> node_end_pfn
)
2679 /* Move the zone boundaries inside the node if necessary */
2680 zone_end_pfn
= min(zone_end_pfn
, node_end_pfn
);
2681 zone_start_pfn
= max(zone_start_pfn
, node_start_pfn
);
2683 /* Return the spanned pages */
2684 return zone_end_pfn
- zone_start_pfn
;
2688 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
2689 * then all holes in the requested range will be accounted for.
2691 unsigned long __meminit
__absent_pages_in_range(int nid
,
2692 unsigned long range_start_pfn
,
2693 unsigned long range_end_pfn
)
2696 unsigned long prev_end_pfn
= 0, hole_pages
= 0;
2697 unsigned long start_pfn
;
2699 /* Find the end_pfn of the first active range of pfns in the node */
2700 i
= first_active_region_index_in_nid(nid
);
2704 /* Account for ranges before physical memory on this node */
2705 if (early_node_map
[i
].start_pfn
> range_start_pfn
)
2706 hole_pages
= early_node_map
[i
].start_pfn
- range_start_pfn
;
2708 prev_end_pfn
= early_node_map
[i
].start_pfn
;
2710 /* Find all holes for the zone within the node */
2711 for (; i
!= -1; i
= next_active_region_index_in_nid(i
, nid
)) {
2713 /* No need to continue if prev_end_pfn is outside the zone */
2714 if (prev_end_pfn
>= range_end_pfn
)
2717 /* Make sure the end of the zone is not within the hole */
2718 start_pfn
= min(early_node_map
[i
].start_pfn
, range_end_pfn
);
2719 prev_end_pfn
= max(prev_end_pfn
, range_start_pfn
);
2721 /* Update the hole size cound and move on */
2722 if (start_pfn
> range_start_pfn
) {
2723 BUG_ON(prev_end_pfn
> start_pfn
);
2724 hole_pages
+= start_pfn
- prev_end_pfn
;
2726 prev_end_pfn
= early_node_map
[i
].end_pfn
;
2729 /* Account for ranges past physical memory on this node */
2730 if (range_end_pfn
> prev_end_pfn
)
2731 hole_pages
+= range_end_pfn
-
2732 max(range_start_pfn
, prev_end_pfn
);
2738 * absent_pages_in_range - Return number of page frames in holes within a range
2739 * @start_pfn: The start PFN to start searching for holes
2740 * @end_pfn: The end PFN to stop searching for holes
2742 * It returns the number of pages frames in memory holes within a range.
2744 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
2745 unsigned long end_pfn
)
2747 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
2750 /* Return the number of page frames in holes in a zone on a node */
2751 unsigned long __meminit
zone_absent_pages_in_node(int nid
,
2752 unsigned long zone_type
,
2753 unsigned long *ignored
)
2755 unsigned long node_start_pfn
, node_end_pfn
;
2756 unsigned long zone_start_pfn
, zone_end_pfn
;
2758 get_pfn_range_for_nid(nid
, &node_start_pfn
, &node_end_pfn
);
2759 zone_start_pfn
= max(arch_zone_lowest_possible_pfn
[zone_type
],
2761 zone_end_pfn
= min(arch_zone_highest_possible_pfn
[zone_type
],
2764 return __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
2768 static inline unsigned long zone_spanned_pages_in_node(int nid
,
2769 unsigned long zone_type
,
2770 unsigned long *zones_size
)
2772 return zones_size
[zone_type
];
2775 static inline unsigned long zone_absent_pages_in_node(int nid
,
2776 unsigned long zone_type
,
2777 unsigned long *zholes_size
)
2782 return zholes_size
[zone_type
];
2787 static void __meminit
calculate_node_totalpages(struct pglist_data
*pgdat
,
2788 unsigned long *zones_size
, unsigned long *zholes_size
)
2790 unsigned long realtotalpages
, totalpages
= 0;
2793 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
2794 totalpages
+= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
2796 pgdat
->node_spanned_pages
= totalpages
;
2798 realtotalpages
= totalpages
;
2799 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
2801 zone_absent_pages_in_node(pgdat
->node_id
, i
,
2803 pgdat
->node_present_pages
= realtotalpages
;
2804 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
2809 * Set up the zone data structures:
2810 * - mark all pages reserved
2811 * - mark all memory queues empty
2812 * - clear the memory bitmaps
2814 static void __meminit
free_area_init_core(struct pglist_data
*pgdat
,
2815 unsigned long *zones_size
, unsigned long *zholes_size
)
2818 int nid
= pgdat
->node_id
;
2819 unsigned long zone_start_pfn
= pgdat
->node_start_pfn
;
2822 pgdat_resize_init(pgdat
);
2823 pgdat
->nr_zones
= 0;
2824 init_waitqueue_head(&pgdat
->kswapd_wait
);
2825 pgdat
->kswapd_max_order
= 0;
2827 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
2828 struct zone
*zone
= pgdat
->node_zones
+ j
;
2829 unsigned long size
, realsize
, memmap_pages
;
2831 size
= zone_spanned_pages_in_node(nid
, j
, zones_size
);
2832 realsize
= size
- zone_absent_pages_in_node(nid
, j
,
2836 * Adjust realsize so that it accounts for how much memory
2837 * is used by this zone for memmap. This affects the watermark
2838 * and per-cpu initialisations
2840 memmap_pages
= (size
* sizeof(struct page
)) >> PAGE_SHIFT
;
2841 if (realsize
>= memmap_pages
) {
2842 realsize
-= memmap_pages
;
2844 " %s zone: %lu pages used for memmap\n",
2845 zone_names
[j
], memmap_pages
);
2848 " %s zone: %lu pages exceeds realsize %lu\n",
2849 zone_names
[j
], memmap_pages
, realsize
);
2851 /* Account for reserved pages */
2852 if (j
== 0 && realsize
> dma_reserve
) {
2853 realsize
-= dma_reserve
;
2854 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
2855 zone_names
[0], dma_reserve
);
2858 if (!is_highmem_idx(j
))
2859 nr_kernel_pages
+= realsize
;
2860 nr_all_pages
+= realsize
;
2862 zone
->spanned_pages
= size
;
2863 zone
->present_pages
= realsize
;
2866 zone
->min_unmapped_pages
= (realsize
*sysctl_min_unmapped_ratio
)
2868 zone
->min_slab_pages
= (realsize
* sysctl_min_slab_ratio
) / 100;
2870 zone
->name
= zone_names
[j
];
2871 spin_lock_init(&zone
->lock
);
2872 spin_lock_init(&zone
->lru_lock
);
2873 zone_seqlock_init(zone
);
2874 zone
->zone_pgdat
= pgdat
;
2876 zone
->prev_priority
= DEF_PRIORITY
;
2878 zone_pcp_init(zone
);
2879 INIT_LIST_HEAD(&zone
->active_list
);
2880 INIT_LIST_HEAD(&zone
->inactive_list
);
2881 zone
->nr_scan_active
= 0;
2882 zone
->nr_scan_inactive
= 0;
2883 zap_zone_vm_stats(zone
);
2884 atomic_set(&zone
->reclaim_in_progress
, 0);
2888 ret
= init_currently_empty_zone(zone
, zone_start_pfn
,
2889 size
, MEMMAP_EARLY
);
2891 zone_start_pfn
+= size
;
2895 static void __init_refok
alloc_node_mem_map(struct pglist_data
*pgdat
)
2897 /* Skip empty nodes */
2898 if (!pgdat
->node_spanned_pages
)
2901 #ifdef CONFIG_FLAT_NODE_MEM_MAP
2902 /* ia64 gets its own node_mem_map, before this, without bootmem */
2903 if (!pgdat
->node_mem_map
) {
2904 unsigned long size
, start
, end
;
2908 * The zone's endpoints aren't required to be MAX_ORDER
2909 * aligned but the node_mem_map endpoints must be in order
2910 * for the buddy allocator to function correctly.
2912 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
2913 end
= pgdat
->node_start_pfn
+ pgdat
->node_spanned_pages
;
2914 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
2915 size
= (end
- start
) * sizeof(struct page
);
2916 map
= alloc_remap(pgdat
->node_id
, size
);
2918 map
= alloc_bootmem_node(pgdat
, size
);
2919 pgdat
->node_mem_map
= map
+ (pgdat
->node_start_pfn
- start
);
2921 #ifndef CONFIG_NEED_MULTIPLE_NODES
2923 * With no DISCONTIG, the global mem_map is just set as node 0's
2925 if (pgdat
== NODE_DATA(0)) {
2926 mem_map
= NODE_DATA(0)->node_mem_map
;
2927 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
2928 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
2929 mem_map
-= pgdat
->node_start_pfn
;
2930 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
2933 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
2936 void __meminit
free_area_init_node(int nid
, struct pglist_data
*pgdat
,
2937 unsigned long *zones_size
, unsigned long node_start_pfn
,
2938 unsigned long *zholes_size
)
2940 pgdat
->node_id
= nid
;
2941 pgdat
->node_start_pfn
= node_start_pfn
;
2942 calculate_node_totalpages(pgdat
, zones_size
, zholes_size
);
2944 alloc_node_mem_map(pgdat
);
2946 free_area_init_core(pgdat
, zones_size
, zholes_size
);
2949 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
2951 #if MAX_NUMNODES > 1
2953 * Figure out the number of possible node ids.
2955 static void __init
setup_nr_node_ids(void)
2958 unsigned int highest
= 0;
2960 for_each_node_mask(node
, node_possible_map
)
2962 nr_node_ids
= highest
+ 1;
2965 static inline void setup_nr_node_ids(void)
2971 * add_active_range - Register a range of PFNs backed by physical memory
2972 * @nid: The node ID the range resides on
2973 * @start_pfn: The start PFN of the available physical memory
2974 * @end_pfn: The end PFN of the available physical memory
2976 * These ranges are stored in an early_node_map[] and later used by
2977 * free_area_init_nodes() to calculate zone sizes and holes. If the
2978 * range spans a memory hole, it is up to the architecture to ensure
2979 * the memory is not freed by the bootmem allocator. If possible
2980 * the range being registered will be merged with existing ranges.
2982 void __init
add_active_range(unsigned int nid
, unsigned long start_pfn
,
2983 unsigned long end_pfn
)
2987 printk(KERN_DEBUG
"Entering add_active_range(%d, %lu, %lu) "
2988 "%d entries of %d used\n",
2989 nid
, start_pfn
, end_pfn
,
2990 nr_nodemap_entries
, MAX_ACTIVE_REGIONS
);
2992 /* Merge with existing active regions if possible */
2993 for (i
= 0; i
< nr_nodemap_entries
; i
++) {
2994 if (early_node_map
[i
].nid
!= nid
)
2997 /* Skip if an existing region covers this new one */
2998 if (start_pfn
>= early_node_map
[i
].start_pfn
&&
2999 end_pfn
<= early_node_map
[i
].end_pfn
)
3002 /* Merge forward if suitable */
3003 if (start_pfn
<= early_node_map
[i
].end_pfn
&&
3004 end_pfn
> early_node_map
[i
].end_pfn
) {
3005 early_node_map
[i
].end_pfn
= end_pfn
;
3009 /* Merge backward if suitable */
3010 if (start_pfn
< early_node_map
[i
].end_pfn
&&
3011 end_pfn
>= early_node_map
[i
].start_pfn
) {
3012 early_node_map
[i
].start_pfn
= start_pfn
;
3017 /* Check that early_node_map is large enough */
3018 if (i
>= MAX_ACTIVE_REGIONS
) {
3019 printk(KERN_CRIT
"More than %d memory regions, truncating\n",
3020 MAX_ACTIVE_REGIONS
);
3024 early_node_map
[i
].nid
= nid
;
3025 early_node_map
[i
].start_pfn
= start_pfn
;
3026 early_node_map
[i
].end_pfn
= end_pfn
;
3027 nr_nodemap_entries
= i
+ 1;
3031 * shrink_active_range - Shrink an existing registered range of PFNs
3032 * @nid: The node id the range is on that should be shrunk
3033 * @old_end_pfn: The old end PFN of the range
3034 * @new_end_pfn: The new PFN of the range
3036 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
3037 * The map is kept at the end physical page range that has already been
3038 * registered with add_active_range(). This function allows an arch to shrink
3039 * an existing registered range.
3041 void __init
shrink_active_range(unsigned int nid
, unsigned long old_end_pfn
,
3042 unsigned long new_end_pfn
)
3046 /* Find the old active region end and shrink */
3047 for_each_active_range_index_in_nid(i
, nid
)
3048 if (early_node_map
[i
].end_pfn
== old_end_pfn
) {
3049 early_node_map
[i
].end_pfn
= new_end_pfn
;
3055 * remove_all_active_ranges - Remove all currently registered regions
3057 * During discovery, it may be found that a table like SRAT is invalid
3058 * and an alternative discovery method must be used. This function removes
3059 * all currently registered regions.
3061 void __init
remove_all_active_ranges(void)
3063 memset(early_node_map
, 0, sizeof(early_node_map
));
3064 nr_nodemap_entries
= 0;
3065 #ifdef CONFIG_MEMORY_HOTPLUG_RESERVE
3066 memset(node_boundary_start_pfn
, 0, sizeof(node_boundary_start_pfn
));
3067 memset(node_boundary_end_pfn
, 0, sizeof(node_boundary_end_pfn
));
3068 #endif /* CONFIG_MEMORY_HOTPLUG_RESERVE */
3071 /* Compare two active node_active_regions */
3072 static int __init
cmp_node_active_region(const void *a
, const void *b
)
3074 struct node_active_region
*arange
= (struct node_active_region
*)a
;
3075 struct node_active_region
*brange
= (struct node_active_region
*)b
;
3077 /* Done this way to avoid overflows */
3078 if (arange
->start_pfn
> brange
->start_pfn
)
3080 if (arange
->start_pfn
< brange
->start_pfn
)
3086 /* sort the node_map by start_pfn */
3087 static void __init
sort_node_map(void)
3089 sort(early_node_map
, (size_t)nr_nodemap_entries
,
3090 sizeof(struct node_active_region
),
3091 cmp_node_active_region
, NULL
);
3094 /* Find the lowest pfn for a node */
3095 unsigned long __init
find_min_pfn_for_node(unsigned long nid
)
3098 unsigned long min_pfn
= ULONG_MAX
;
3100 /* Assuming a sorted map, the first range found has the starting pfn */
3101 for_each_active_range_index_in_nid(i
, nid
)
3102 min_pfn
= min(min_pfn
, early_node_map
[i
].start_pfn
);
3104 if (min_pfn
== ULONG_MAX
) {
3106 "Could not find start_pfn for node %lu\n", nid
);
3114 * find_min_pfn_with_active_regions - Find the minimum PFN registered
3116 * It returns the minimum PFN based on information provided via
3117 * add_active_range().
3119 unsigned long __init
find_min_pfn_with_active_regions(void)
3121 return find_min_pfn_for_node(MAX_NUMNODES
);
3125 * find_max_pfn_with_active_regions - Find the maximum PFN registered
3127 * It returns the maximum PFN based on information provided via
3128 * add_active_range().
3130 unsigned long __init
find_max_pfn_with_active_regions(void)
3133 unsigned long max_pfn
= 0;
3135 for (i
= 0; i
< nr_nodemap_entries
; i
++)
3136 max_pfn
= max(max_pfn
, early_node_map
[i
].end_pfn
);
3142 * free_area_init_nodes - Initialise all pg_data_t and zone data
3143 * @max_zone_pfn: an array of max PFNs for each zone
3145 * This will call free_area_init_node() for each active node in the system.
3146 * Using the page ranges provided by add_active_range(), the size of each
3147 * zone in each node and their holes is calculated. If the maximum PFN
3148 * between two adjacent zones match, it is assumed that the zone is empty.
3149 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
3150 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
3151 * starts where the previous one ended. For example, ZONE_DMA32 starts
3152 * at arch_max_dma_pfn.
3154 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
3159 /* Sort early_node_map as initialisation assumes it is sorted */
3162 /* Record where the zone boundaries are */
3163 memset(arch_zone_lowest_possible_pfn
, 0,
3164 sizeof(arch_zone_lowest_possible_pfn
));
3165 memset(arch_zone_highest_possible_pfn
, 0,
3166 sizeof(arch_zone_highest_possible_pfn
));
3167 arch_zone_lowest_possible_pfn
[0] = find_min_pfn_with_active_regions();
3168 arch_zone_highest_possible_pfn
[0] = max_zone_pfn
[0];
3169 for (i
= 1; i
< MAX_NR_ZONES
; i
++) {
3170 arch_zone_lowest_possible_pfn
[i
] =
3171 arch_zone_highest_possible_pfn
[i
-1];
3172 arch_zone_highest_possible_pfn
[i
] =
3173 max(max_zone_pfn
[i
], arch_zone_lowest_possible_pfn
[i
]);
3176 /* Print out the zone ranges */
3177 printk("Zone PFN ranges:\n");
3178 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
3179 printk(" %-8s %8lu -> %8lu\n",
3181 arch_zone_lowest_possible_pfn
[i
],
3182 arch_zone_highest_possible_pfn
[i
]);
3184 /* Print out the early_node_map[] */
3185 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries
);
3186 for (i
= 0; i
< nr_nodemap_entries
; i
++)
3187 printk(" %3d: %8lu -> %8lu\n", early_node_map
[i
].nid
,
3188 early_node_map
[i
].start_pfn
,
3189 early_node_map
[i
].end_pfn
);
3191 /* Initialise every node */
3192 setup_nr_node_ids();
3193 for_each_online_node(nid
) {
3194 pg_data_t
*pgdat
= NODE_DATA(nid
);
3195 free_area_init_node(nid
, pgdat
, NULL
,
3196 find_min_pfn_for_node(nid
), NULL
);
3199 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
3202 * set_dma_reserve - set the specified number of pages reserved in the first zone
3203 * @new_dma_reserve: The number of pages to mark reserved
3205 * The per-cpu batchsize and zone watermarks are determined by present_pages.
3206 * In the DMA zone, a significant percentage may be consumed by kernel image
3207 * and other unfreeable allocations which can skew the watermarks badly. This
3208 * function may optionally be used to account for unfreeable pages in the
3209 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
3210 * smaller per-cpu batchsize.
3212 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
3214 dma_reserve
= new_dma_reserve
;
3217 #ifndef CONFIG_NEED_MULTIPLE_NODES
3218 static bootmem_data_t contig_bootmem_data
;
3219 struct pglist_data contig_page_data
= { .bdata
= &contig_bootmem_data
};
3221 EXPORT_SYMBOL(contig_page_data
);
3224 void __init
free_area_init(unsigned long *zones_size
)
3226 free_area_init_node(0, NODE_DATA(0), zones_size
,
3227 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
3230 static int page_alloc_cpu_notify(struct notifier_block
*self
,
3231 unsigned long action
, void *hcpu
)
3233 int cpu
= (unsigned long)hcpu
;
3235 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
) {
3236 local_irq_disable();
3238 vm_events_fold_cpu(cpu
);
3240 refresh_cpu_vm_stats(cpu
);
3245 void __init
page_alloc_init(void)
3247 hotcpu_notifier(page_alloc_cpu_notify
, 0);
3251 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
3252 * or min_free_kbytes changes.
3254 static void calculate_totalreserve_pages(void)
3256 struct pglist_data
*pgdat
;
3257 unsigned long reserve_pages
= 0;
3258 enum zone_type i
, j
;
3260 for_each_online_pgdat(pgdat
) {
3261 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
3262 struct zone
*zone
= pgdat
->node_zones
+ i
;
3263 unsigned long max
= 0;
3265 /* Find valid and maximum lowmem_reserve in the zone */
3266 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
3267 if (zone
->lowmem_reserve
[j
] > max
)
3268 max
= zone
->lowmem_reserve
[j
];
3271 /* we treat pages_high as reserved pages. */
3272 max
+= zone
->pages_high
;
3274 if (max
> zone
->present_pages
)
3275 max
= zone
->present_pages
;
3276 reserve_pages
+= max
;
3279 totalreserve_pages
= reserve_pages
;
3283 * setup_per_zone_lowmem_reserve - called whenever
3284 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
3285 * has a correct pages reserved value, so an adequate number of
3286 * pages are left in the zone after a successful __alloc_pages().
3288 static void setup_per_zone_lowmem_reserve(void)
3290 struct pglist_data
*pgdat
;
3291 enum zone_type j
, idx
;
3293 for_each_online_pgdat(pgdat
) {
3294 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
3295 struct zone
*zone
= pgdat
->node_zones
+ j
;
3296 unsigned long present_pages
= zone
->present_pages
;
3298 zone
->lowmem_reserve
[j
] = 0;
3302 struct zone
*lower_zone
;
3306 if (sysctl_lowmem_reserve_ratio
[idx
] < 1)
3307 sysctl_lowmem_reserve_ratio
[idx
] = 1;
3309 lower_zone
= pgdat
->node_zones
+ idx
;
3310 lower_zone
->lowmem_reserve
[j
] = present_pages
/
3311 sysctl_lowmem_reserve_ratio
[idx
];
3312 present_pages
+= lower_zone
->present_pages
;
3317 /* update totalreserve_pages */
3318 calculate_totalreserve_pages();
3322 * setup_per_zone_pages_min - called when min_free_kbytes changes.
3324 * Ensures that the pages_{min,low,high} values for each zone are set correctly
3325 * with respect to min_free_kbytes.
3327 void setup_per_zone_pages_min(void)
3329 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
3330 unsigned long lowmem_pages
= 0;
3332 unsigned long flags
;
3334 /* Calculate total number of !ZONE_HIGHMEM pages */
3335 for_each_zone(zone
) {
3336 if (!is_highmem(zone
))
3337 lowmem_pages
+= zone
->present_pages
;
3340 for_each_zone(zone
) {
3343 spin_lock_irqsave(&zone
->lru_lock
, flags
);
3344 tmp
= (u64
)pages_min
* zone
->present_pages
;
3345 do_div(tmp
, lowmem_pages
);
3346 if (is_highmem(zone
)) {
3348 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
3349 * need highmem pages, so cap pages_min to a small
3352 * The (pages_high-pages_low) and (pages_low-pages_min)
3353 * deltas controls asynch page reclaim, and so should
3354 * not be capped for highmem.
3358 min_pages
= zone
->present_pages
/ 1024;
3359 if (min_pages
< SWAP_CLUSTER_MAX
)
3360 min_pages
= SWAP_CLUSTER_MAX
;
3361 if (min_pages
> 128)
3363 zone
->pages_min
= min_pages
;
3366 * If it's a lowmem zone, reserve a number of pages
3367 * proportionate to the zone's size.
3369 zone
->pages_min
= tmp
;
3372 zone
->pages_low
= zone
->pages_min
+ (tmp
>> 2);
3373 zone
->pages_high
= zone
->pages_min
+ (tmp
>> 1);
3374 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
3377 /* update totalreserve_pages */
3378 calculate_totalreserve_pages();
3382 * Initialise min_free_kbytes.
3384 * For small machines we want it small (128k min). For large machines
3385 * we want it large (64MB max). But it is not linear, because network
3386 * bandwidth does not increase linearly with machine size. We use
3388 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
3389 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
3405 static int __init
init_per_zone_pages_min(void)
3407 unsigned long lowmem_kbytes
;
3409 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
3411 min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
3412 if (min_free_kbytes
< 128)
3413 min_free_kbytes
= 128;
3414 if (min_free_kbytes
> 65536)
3415 min_free_kbytes
= 65536;
3416 setup_per_zone_pages_min();
3417 setup_per_zone_lowmem_reserve();
3420 module_init(init_per_zone_pages_min
)
3423 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
3424 * that we can call two helper functions whenever min_free_kbytes
3427 int min_free_kbytes_sysctl_handler(ctl_table
*table
, int write
,
3428 struct file
*file
, void __user
*buffer
, size_t *length
, loff_t
*ppos
)
3430 proc_dointvec(table
, write
, file
, buffer
, length
, ppos
);
3432 setup_per_zone_pages_min();
3437 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table
*table
, int write
,
3438 struct file
*file
, void __user
*buffer
, size_t *length
, loff_t
*ppos
)
3443 rc
= proc_dointvec_minmax(table
, write
, file
, buffer
, length
, ppos
);
3448 zone
->min_unmapped_pages
= (zone
->present_pages
*
3449 sysctl_min_unmapped_ratio
) / 100;
3453 int sysctl_min_slab_ratio_sysctl_handler(ctl_table
*table
, int write
,
3454 struct file
*file
, void __user
*buffer
, size_t *length
, loff_t
*ppos
)
3459 rc
= proc_dointvec_minmax(table
, write
, file
, buffer
, length
, ppos
);
3464 zone
->min_slab_pages
= (zone
->present_pages
*
3465 sysctl_min_slab_ratio
) / 100;
3471 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
3472 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
3473 * whenever sysctl_lowmem_reserve_ratio changes.
3475 * The reserve ratio obviously has absolutely no relation with the
3476 * pages_min watermarks. The lowmem reserve ratio can only make sense
3477 * if in function of the boot time zone sizes.
3479 int lowmem_reserve_ratio_sysctl_handler(ctl_table
*table
, int write
,
3480 struct file
*file
, void __user
*buffer
, size_t *length
, loff_t
*ppos
)
3482 proc_dointvec_minmax(table
, write
, file
, buffer
, length
, ppos
);
3483 setup_per_zone_lowmem_reserve();
3488 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
3489 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
3490 * can have before it gets flushed back to buddy allocator.
3493 int percpu_pagelist_fraction_sysctl_handler(ctl_table
*table
, int write
,
3494 struct file
*file
, void __user
*buffer
, size_t *length
, loff_t
*ppos
)
3500 ret
= proc_dointvec_minmax(table
, write
, file
, buffer
, length
, ppos
);
3501 if (!write
|| (ret
== -EINVAL
))
3503 for_each_zone(zone
) {
3504 for_each_online_cpu(cpu
) {
3506 high
= zone
->present_pages
/ percpu_pagelist_fraction
;
3507 setup_pagelist_highmark(zone_pcp(zone
, cpu
), high
);
3513 int hashdist
= HASHDIST_DEFAULT
;
3516 static int __init
set_hashdist(char *str
)
3520 hashdist
= simple_strtoul(str
, &str
, 0);
3523 __setup("hashdist=", set_hashdist
);
3527 * allocate a large system hash table from bootmem
3528 * - it is assumed that the hash table must contain an exact power-of-2
3529 * quantity of entries
3530 * - limit is the number of hash buckets, not the total allocation size
3532 void *__init
alloc_large_system_hash(const char *tablename
,
3533 unsigned long bucketsize
,
3534 unsigned long numentries
,
3537 unsigned int *_hash_shift
,
3538 unsigned int *_hash_mask
,
3539 unsigned long limit
)
3541 unsigned long long max
= limit
;
3542 unsigned long log2qty
, size
;
3545 /* allow the kernel cmdline to have a say */
3547 /* round applicable memory size up to nearest megabyte */
3548 numentries
= nr_kernel_pages
;
3549 numentries
+= (1UL << (20 - PAGE_SHIFT
)) - 1;
3550 numentries
>>= 20 - PAGE_SHIFT
;
3551 numentries
<<= 20 - PAGE_SHIFT
;
3553 /* limit to 1 bucket per 2^scale bytes of low memory */
3554 if (scale
> PAGE_SHIFT
)
3555 numentries
>>= (scale
- PAGE_SHIFT
);
3557 numentries
<<= (PAGE_SHIFT
- scale
);
3559 /* Make sure we've got at least a 0-order allocation.. */
3560 if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
3561 numentries
= PAGE_SIZE
/ bucketsize
;
3563 numentries
= roundup_pow_of_two(numentries
);
3565 /* limit allocation size to 1/16 total memory by default */
3567 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
3568 do_div(max
, bucketsize
);
3571 if (numentries
> max
)
3574 log2qty
= ilog2(numentries
);
3577 size
= bucketsize
<< log2qty
;
3578 if (flags
& HASH_EARLY
)
3579 table
= alloc_bootmem(size
);
3581 table
= __vmalloc(size
, GFP_ATOMIC
, PAGE_KERNEL
);
3583 unsigned long order
;
3584 for (order
= 0; ((1UL << order
) << PAGE_SHIFT
) < size
; order
++)
3586 table
= (void*) __get_free_pages(GFP_ATOMIC
, order
);
3588 * If bucketsize is not a power-of-two, we may free
3589 * some pages at the end of hash table.
3592 unsigned long alloc_end
= (unsigned long)table
+
3593 (PAGE_SIZE
<< order
);
3594 unsigned long used
= (unsigned long)table
+
3596 split_page(virt_to_page(table
), order
);
3597 while (used
< alloc_end
) {
3603 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
3606 panic("Failed to allocate %s hash table\n", tablename
);
3608 printk("%s hash table entries: %d (order: %d, %lu bytes)\n",
3611 ilog2(size
) - PAGE_SHIFT
,
3615 *_hash_shift
= log2qty
;
3617 *_hash_mask
= (1 << log2qty
) - 1;
3622 #ifdef CONFIG_OUT_OF_LINE_PFN_TO_PAGE
3623 struct page
*pfn_to_page(unsigned long pfn
)
3625 return __pfn_to_page(pfn
);
3627 unsigned long page_to_pfn(struct page
*page
)
3629 return __page_to_pfn(page
);
3631 EXPORT_SYMBOL(pfn_to_page
);
3632 EXPORT_SYMBOL(page_to_pfn
);
3633 #endif /* CONFIG_OUT_OF_LINE_PFN_TO_PAGE */