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>
41 #include <asm/tlbflush.h>
42 #include <asm/div64.h>
46 * MCD - HACK: Find somewhere to initialize this EARLY, or make this
49 nodemask_t node_online_map __read_mostly
= { { [0] = 1UL } };
50 EXPORT_SYMBOL(node_online_map
);
51 nodemask_t node_possible_map __read_mostly
= NODE_MASK_ALL
;
52 EXPORT_SYMBOL(node_possible_map
);
53 unsigned long totalram_pages __read_mostly
;
54 unsigned long totalreserve_pages __read_mostly
;
56 int percpu_pagelist_fraction
;
58 static void __free_pages_ok(struct page
*page
, unsigned int order
);
61 * results with 256, 32 in the lowmem_reserve sysctl:
62 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
63 * 1G machine -> (16M dma, 784M normal, 224M high)
64 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
65 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
66 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
68 * TBD: should special case ZONE_DMA32 machines here - in those we normally
69 * don't need any ZONE_NORMAL reservation
71 int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
-1] = {
73 #ifdef CONFIG_ZONE_DMA32
81 EXPORT_SYMBOL(totalram_pages
);
84 * Used by page_zone() to look up the address of the struct zone whose
85 * id is encoded in the upper bits of page->flags
87 struct zone
*zone_table
[1 << ZONETABLE_SHIFT
] __read_mostly
;
88 EXPORT_SYMBOL(zone_table
);
90 static char *zone_names
[MAX_NR_ZONES
] = {
92 #ifdef CONFIG_ZONE_DMA32
101 int min_free_kbytes
= 1024;
103 unsigned long __meminitdata nr_kernel_pages
;
104 unsigned long __meminitdata nr_all_pages
;
106 #ifdef CONFIG_DEBUG_VM
107 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
111 unsigned long pfn
= page_to_pfn(page
);
114 seq
= zone_span_seqbegin(zone
);
115 if (pfn
>= zone
->zone_start_pfn
+ zone
->spanned_pages
)
117 else if (pfn
< zone
->zone_start_pfn
)
119 } while (zone_span_seqretry(zone
, seq
));
124 static int page_is_consistent(struct zone
*zone
, struct page
*page
)
126 #ifdef CONFIG_HOLES_IN_ZONE
127 if (!pfn_valid(page_to_pfn(page
)))
130 if (zone
!= page_zone(page
))
136 * Temporary debugging check for pages not lying within a given zone.
138 static int bad_range(struct zone
*zone
, struct page
*page
)
140 if (page_outside_zone_boundaries(zone
, page
))
142 if (!page_is_consistent(zone
, page
))
148 static inline int bad_range(struct zone
*zone
, struct page
*page
)
154 static void bad_page(struct page
*page
)
156 printk(KERN_EMERG
"Bad page state in process '%s'\n"
157 KERN_EMERG
"page:%p flags:0x%0*lx mapping:%p mapcount:%d count:%d\n"
158 KERN_EMERG
"Trying to fix it up, but a reboot is needed\n"
159 KERN_EMERG
"Backtrace:\n",
160 current
->comm
, page
, (int)(2*sizeof(unsigned long)),
161 (unsigned long)page
->flags
, page
->mapping
,
162 page_mapcount(page
), page_count(page
));
164 page
->flags
&= ~(1 << PG_lru
|
174 set_page_count(page
, 0);
175 reset_page_mapcount(page
);
176 page
->mapping
= NULL
;
177 add_taint(TAINT_BAD_PAGE
);
181 * Higher-order pages are called "compound pages". They are structured thusly:
183 * The first PAGE_SIZE page is called the "head page".
185 * The remaining PAGE_SIZE pages are called "tail pages".
187 * All pages have PG_compound set. All pages have their ->private pointing at
188 * the head page (even the head page has this).
190 * The first tail page's ->lru.next holds the address of the compound page's
191 * put_page() function. Its ->lru.prev holds the order of allocation.
192 * This usage means that zero-order pages may not be compound.
195 static void free_compound_page(struct page
*page
)
197 __free_pages_ok(page
, (unsigned long)page
[1].lru
.prev
);
200 static void prep_compound_page(struct page
*page
, unsigned long order
)
203 int nr_pages
= 1 << order
;
205 page
[1].lru
.next
= (void *)free_compound_page
; /* set dtor */
206 page
[1].lru
.prev
= (void *)order
;
207 for (i
= 0; i
< nr_pages
; i
++) {
208 struct page
*p
= page
+ i
;
210 __SetPageCompound(p
);
211 set_page_private(p
, (unsigned long)page
);
215 static void destroy_compound_page(struct page
*page
, unsigned long order
)
218 int nr_pages
= 1 << order
;
220 if (unlikely((unsigned long)page
[1].lru
.prev
!= order
))
223 for (i
= 0; i
< nr_pages
; i
++) {
224 struct page
*p
= page
+ i
;
226 if (unlikely(!PageCompound(p
) |
227 (page_private(p
) != (unsigned long)page
)))
229 __ClearPageCompound(p
);
233 static inline void prep_zero_page(struct page
*page
, int order
, gfp_t gfp_flags
)
237 VM_BUG_ON((gfp_flags
& (__GFP_WAIT
| __GFP_HIGHMEM
)) == __GFP_HIGHMEM
);
239 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
240 * and __GFP_HIGHMEM from hard or soft interrupt context.
242 VM_BUG_ON((gfp_flags
& __GFP_HIGHMEM
) && in_interrupt());
243 for (i
= 0; i
< (1 << order
); i
++)
244 clear_highpage(page
+ i
);
248 * function for dealing with page's order in buddy system.
249 * zone->lock is already acquired when we use these.
250 * So, we don't need atomic page->flags operations here.
252 static inline unsigned long page_order(struct page
*page
)
254 return page_private(page
);
257 static inline void set_page_order(struct page
*page
, int order
)
259 set_page_private(page
, order
);
260 __SetPageBuddy(page
);
263 static inline void rmv_page_order(struct page
*page
)
265 __ClearPageBuddy(page
);
266 set_page_private(page
, 0);
270 * Locate the struct page for both the matching buddy in our
271 * pair (buddy1) and the combined O(n+1) page they form (page).
273 * 1) Any buddy B1 will have an order O twin B2 which satisfies
274 * the following equation:
276 * For example, if the starting buddy (buddy2) is #8 its order
278 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
280 * 2) Any buddy B will have an order O+1 parent P which
281 * satisfies the following equation:
284 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
286 static inline struct page
*
287 __page_find_buddy(struct page
*page
, unsigned long page_idx
, unsigned int order
)
289 unsigned long buddy_idx
= page_idx
^ (1 << order
);
291 return page
+ (buddy_idx
- page_idx
);
294 static inline unsigned long
295 __find_combined_index(unsigned long page_idx
, unsigned int order
)
297 return (page_idx
& ~(1 << order
));
301 * This function checks whether a page is free && is the buddy
302 * we can do coalesce a page and its buddy if
303 * (a) the buddy is not in a hole &&
304 * (b) the buddy is in the buddy system &&
305 * (c) a page and its buddy have the same order &&
306 * (d) a page and its buddy are in the same zone.
308 * For recording whether a page is in the buddy system, we use PG_buddy.
309 * Setting, clearing, and testing PG_buddy is serialized by zone->lock.
311 * For recording page's order, we use page_private(page).
313 static inline int page_is_buddy(struct page
*page
, struct page
*buddy
,
316 #ifdef CONFIG_HOLES_IN_ZONE
317 if (!pfn_valid(page_to_pfn(buddy
)))
321 if (page_zone_id(page
) != page_zone_id(buddy
))
324 if (PageBuddy(buddy
) && page_order(buddy
) == order
) {
325 BUG_ON(page_count(buddy
) != 0);
332 * Freeing function for a buddy system allocator.
334 * The concept of a buddy system is to maintain direct-mapped table
335 * (containing bit values) for memory blocks of various "orders".
336 * The bottom level table contains the map for the smallest allocatable
337 * units of memory (here, pages), and each level above it describes
338 * pairs of units from the levels below, hence, "buddies".
339 * At a high level, all that happens here is marking the table entry
340 * at the bottom level available, and propagating the changes upward
341 * as necessary, plus some accounting needed to play nicely with other
342 * parts of the VM system.
343 * At each level, we keep a list of pages, which are heads of continuous
344 * free pages of length of (1 << order) and marked with PG_buddy. Page's
345 * order is recorded in page_private(page) field.
346 * So when we are allocating or freeing one, we can derive the state of the
347 * other. That is, if we allocate a small block, and both were
348 * free, the remainder of the region must be split into blocks.
349 * If a block is freed, and its buddy is also free, then this
350 * triggers coalescing into a block of larger size.
355 static inline void __free_one_page(struct page
*page
,
356 struct zone
*zone
, unsigned int order
)
358 unsigned long page_idx
;
359 int order_size
= 1 << order
;
361 if (unlikely(PageCompound(page
)))
362 destroy_compound_page(page
, order
);
364 page_idx
= page_to_pfn(page
) & ((1 << MAX_ORDER
) - 1);
366 VM_BUG_ON(page_idx
& (order_size
- 1));
367 VM_BUG_ON(bad_range(zone
, page
));
369 zone
->free_pages
+= order_size
;
370 while (order
< MAX_ORDER
-1) {
371 unsigned long combined_idx
;
372 struct free_area
*area
;
375 buddy
= __page_find_buddy(page
, page_idx
, order
);
376 if (!page_is_buddy(page
, buddy
, order
))
377 break; /* Move the buddy up one level. */
379 list_del(&buddy
->lru
);
380 area
= zone
->free_area
+ order
;
382 rmv_page_order(buddy
);
383 combined_idx
= __find_combined_index(page_idx
, order
);
384 page
= page
+ (combined_idx
- page_idx
);
385 page_idx
= combined_idx
;
388 set_page_order(page
, order
);
389 list_add(&page
->lru
, &zone
->free_area
[order
].free_list
);
390 zone
->free_area
[order
].nr_free
++;
393 static inline int free_pages_check(struct page
*page
)
395 if (unlikely(page_mapcount(page
) |
396 (page
->mapping
!= NULL
) |
397 (page_count(page
) != 0) |
411 __ClearPageDirty(page
);
413 * For now, we report if PG_reserved was found set, but do not
414 * clear it, and do not free the page. But we shall soon need
415 * to do more, for when the ZERO_PAGE count wraps negative.
417 return PageReserved(page
);
421 * Frees a list of pages.
422 * Assumes all pages on list are in same zone, and of same order.
423 * count is the number of pages to free.
425 * If the zone was previously in an "all pages pinned" state then look to
426 * see if this freeing clears that state.
428 * And clear the zone's pages_scanned counter, to hold off the "all pages are
429 * pinned" detection logic.
431 static void free_pages_bulk(struct zone
*zone
, int count
,
432 struct list_head
*list
, int order
)
434 spin_lock(&zone
->lock
);
435 zone
->all_unreclaimable
= 0;
436 zone
->pages_scanned
= 0;
440 VM_BUG_ON(list_empty(list
));
441 page
= list_entry(list
->prev
, struct page
, lru
);
442 /* have to delete it as __free_one_page list manipulates */
443 list_del(&page
->lru
);
444 __free_one_page(page
, zone
, order
);
446 spin_unlock(&zone
->lock
);
449 static void free_one_page(struct zone
*zone
, struct page
*page
, int order
)
452 list_add(&page
->lru
, &list
);
453 free_pages_bulk(zone
, 1, &list
, order
);
456 static void __free_pages_ok(struct page
*page
, unsigned int order
)
462 arch_free_page(page
, order
);
463 if (!PageHighMem(page
))
464 debug_check_no_locks_freed(page_address(page
),
467 for (i
= 0 ; i
< (1 << order
) ; ++i
)
468 reserved
+= free_pages_check(page
+ i
);
472 kernel_map_pages(page
, 1 << order
, 0);
473 local_irq_save(flags
);
474 __count_vm_events(PGFREE
, 1 << order
);
475 free_one_page(page_zone(page
), page
, order
);
476 local_irq_restore(flags
);
480 * permit the bootmem allocator to evade page validation on high-order frees
482 void fastcall __init
__free_pages_bootmem(struct page
*page
, unsigned int order
)
485 __ClearPageReserved(page
);
486 set_page_count(page
, 0);
487 set_page_refcounted(page
);
493 for (loop
= 0; loop
< BITS_PER_LONG
; loop
++) {
494 struct page
*p
= &page
[loop
];
496 if (loop
+ 1 < BITS_PER_LONG
)
498 __ClearPageReserved(p
);
499 set_page_count(p
, 0);
502 set_page_refcounted(page
);
503 __free_pages(page
, order
);
509 * The order of subdivision here is critical for the IO subsystem.
510 * Please do not alter this order without good reasons and regression
511 * testing. Specifically, as large blocks of memory are subdivided,
512 * the order in which smaller blocks are delivered depends on the order
513 * they're subdivided in this function. This is the primary factor
514 * influencing the order in which pages are delivered to the IO
515 * subsystem according to empirical testing, and this is also justified
516 * by considering the behavior of a buddy system containing a single
517 * large block of memory acted on by a series of small allocations.
518 * This behavior is a critical factor in sglist merging's success.
522 static inline void expand(struct zone
*zone
, struct page
*page
,
523 int low
, int high
, struct free_area
*area
)
525 unsigned long size
= 1 << high
;
531 VM_BUG_ON(bad_range(zone
, &page
[size
]));
532 list_add(&page
[size
].lru
, &area
->free_list
);
534 set_page_order(&page
[size
], high
);
539 * This page is about to be returned from the page allocator
541 static int prep_new_page(struct page
*page
, int order
, gfp_t gfp_flags
)
543 if (unlikely(page_mapcount(page
) |
544 (page
->mapping
!= NULL
) |
545 (page_count(page
) != 0) |
561 * For now, we report if PG_reserved was found set, but do not
562 * clear it, and do not allocate the page: as a safety net.
564 if (PageReserved(page
))
567 page
->flags
&= ~(1 << PG_uptodate
| 1 << PG_error
|
568 1 << PG_referenced
| 1 << PG_arch_1
|
569 1 << PG_checked
| 1 << PG_mappedtodisk
);
570 set_page_private(page
, 0);
571 set_page_refcounted(page
);
572 kernel_map_pages(page
, 1 << order
, 1);
574 if (gfp_flags
& __GFP_ZERO
)
575 prep_zero_page(page
, order
, gfp_flags
);
577 if (order
&& (gfp_flags
& __GFP_COMP
))
578 prep_compound_page(page
, order
);
584 * Do the hard work of removing an element from the buddy allocator.
585 * Call me with the zone->lock already held.
587 static struct page
*__rmqueue(struct zone
*zone
, unsigned int order
)
589 struct free_area
* area
;
590 unsigned int current_order
;
593 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
594 area
= zone
->free_area
+ current_order
;
595 if (list_empty(&area
->free_list
))
598 page
= list_entry(area
->free_list
.next
, struct page
, lru
);
599 list_del(&page
->lru
);
600 rmv_page_order(page
);
602 zone
->free_pages
-= 1UL << order
;
603 expand(zone
, page
, order
, current_order
, area
);
611 * Obtain a specified number of elements from the buddy allocator, all under
612 * a single hold of the lock, for efficiency. Add them to the supplied list.
613 * Returns the number of new pages which were placed at *list.
615 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
616 unsigned long count
, struct list_head
*list
)
620 spin_lock(&zone
->lock
);
621 for (i
= 0; i
< count
; ++i
) {
622 struct page
*page
= __rmqueue(zone
, order
);
623 if (unlikely(page
== NULL
))
625 list_add_tail(&page
->lru
, list
);
627 spin_unlock(&zone
->lock
);
633 * Called from the slab reaper to drain pagesets on a particular node that
634 * belong to the currently executing processor.
635 * Note that this function must be called with the thread pinned to
636 * a single processor.
638 void drain_node_pages(int nodeid
)
644 for (z
= 0; z
< MAX_NR_ZONES
; z
++) {
645 struct zone
*zone
= NODE_DATA(nodeid
)->node_zones
+ z
;
646 struct per_cpu_pageset
*pset
;
648 pset
= zone_pcp(zone
, smp_processor_id());
649 for (i
= 0; i
< ARRAY_SIZE(pset
->pcp
); i
++) {
650 struct per_cpu_pages
*pcp
;
654 local_irq_save(flags
);
655 free_pages_bulk(zone
, pcp
->count
, &pcp
->list
, 0);
657 local_irq_restore(flags
);
664 #if defined(CONFIG_PM) || defined(CONFIG_HOTPLUG_CPU)
665 static void __drain_pages(unsigned int cpu
)
671 for_each_zone(zone
) {
672 struct per_cpu_pageset
*pset
;
674 pset
= zone_pcp(zone
, cpu
);
675 for (i
= 0; i
< ARRAY_SIZE(pset
->pcp
); i
++) {
676 struct per_cpu_pages
*pcp
;
679 local_irq_save(flags
);
680 free_pages_bulk(zone
, pcp
->count
, &pcp
->list
, 0);
682 local_irq_restore(flags
);
686 #endif /* CONFIG_PM || CONFIG_HOTPLUG_CPU */
690 void mark_free_pages(struct zone
*zone
)
692 unsigned long zone_pfn
, flags
;
694 struct list_head
*curr
;
696 if (!zone
->spanned_pages
)
699 spin_lock_irqsave(&zone
->lock
, flags
);
700 for (zone_pfn
= 0; zone_pfn
< zone
->spanned_pages
; ++zone_pfn
)
701 ClearPageNosaveFree(pfn_to_page(zone_pfn
+ zone
->zone_start_pfn
));
703 for (order
= MAX_ORDER
- 1; order
>= 0; --order
)
704 list_for_each(curr
, &zone
->free_area
[order
].free_list
) {
705 unsigned long start_pfn
, i
;
707 start_pfn
= page_to_pfn(list_entry(curr
, struct page
, lru
));
709 for (i
=0; i
< (1<<order
); i
++)
710 SetPageNosaveFree(pfn_to_page(start_pfn
+i
));
712 spin_unlock_irqrestore(&zone
->lock
, flags
);
716 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
718 void drain_local_pages(void)
722 local_irq_save(flags
);
723 __drain_pages(smp_processor_id());
724 local_irq_restore(flags
);
726 #endif /* CONFIG_PM */
729 * Free a 0-order page
731 static void fastcall
free_hot_cold_page(struct page
*page
, int cold
)
733 struct zone
*zone
= page_zone(page
);
734 struct per_cpu_pages
*pcp
;
737 arch_free_page(page
, 0);
740 page
->mapping
= NULL
;
741 if (free_pages_check(page
))
744 kernel_map_pages(page
, 1, 0);
746 pcp
= &zone_pcp(zone
, get_cpu())->pcp
[cold
];
747 local_irq_save(flags
);
748 __count_vm_event(PGFREE
);
749 list_add(&page
->lru
, &pcp
->list
);
751 if (pcp
->count
>= pcp
->high
) {
752 free_pages_bulk(zone
, pcp
->batch
, &pcp
->list
, 0);
753 pcp
->count
-= pcp
->batch
;
755 local_irq_restore(flags
);
759 void fastcall
free_hot_page(struct page
*page
)
761 free_hot_cold_page(page
, 0);
764 void fastcall
free_cold_page(struct page
*page
)
766 free_hot_cold_page(page
, 1);
770 * split_page takes a non-compound higher-order page, and splits it into
771 * n (1<<order) sub-pages: page[0..n]
772 * Each sub-page must be freed individually.
774 * Note: this is probably too low level an operation for use in drivers.
775 * Please consult with lkml before using this in your driver.
777 void split_page(struct page
*page
, unsigned int order
)
781 VM_BUG_ON(PageCompound(page
));
782 VM_BUG_ON(!page_count(page
));
783 for (i
= 1; i
< (1 << order
); i
++)
784 set_page_refcounted(page
+ i
);
788 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
789 * we cheat by calling it from here, in the order > 0 path. Saves a branch
792 static struct page
*buffered_rmqueue(struct zonelist
*zonelist
,
793 struct zone
*zone
, int order
, gfp_t gfp_flags
)
797 int cold
= !!(gfp_flags
& __GFP_COLD
);
802 if (likely(order
== 0)) {
803 struct per_cpu_pages
*pcp
;
805 pcp
= &zone_pcp(zone
, cpu
)->pcp
[cold
];
806 local_irq_save(flags
);
808 pcp
->count
+= rmqueue_bulk(zone
, 0,
809 pcp
->batch
, &pcp
->list
);
810 if (unlikely(!pcp
->count
))
813 page
= list_entry(pcp
->list
.next
, struct page
, lru
);
814 list_del(&page
->lru
);
817 spin_lock_irqsave(&zone
->lock
, flags
);
818 page
= __rmqueue(zone
, order
);
819 spin_unlock(&zone
->lock
);
824 __count_zone_vm_events(PGALLOC
, zone
, 1 << order
);
825 zone_statistics(zonelist
, zone
);
826 local_irq_restore(flags
);
829 VM_BUG_ON(bad_range(zone
, page
));
830 if (prep_new_page(page
, order
, gfp_flags
))
835 local_irq_restore(flags
);
840 #define ALLOC_NO_WATERMARKS 0x01 /* don't check watermarks at all */
841 #define ALLOC_WMARK_MIN 0x02 /* use pages_min watermark */
842 #define ALLOC_WMARK_LOW 0x04 /* use pages_low watermark */
843 #define ALLOC_WMARK_HIGH 0x08 /* use pages_high watermark */
844 #define ALLOC_HARDER 0x10 /* try to alloc harder */
845 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
846 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
849 * Return 1 if free pages are above 'mark'. This takes into account the order
852 int zone_watermark_ok(struct zone
*z
, int order
, unsigned long mark
,
853 int classzone_idx
, int alloc_flags
)
855 /* free_pages my go negative - that's OK */
856 long min
= mark
, free_pages
= z
->free_pages
- (1 << order
) + 1;
859 if (alloc_flags
& ALLOC_HIGH
)
861 if (alloc_flags
& ALLOC_HARDER
)
864 if (free_pages
<= min
+ z
->lowmem_reserve
[classzone_idx
])
866 for (o
= 0; o
< order
; o
++) {
867 /* At the next order, this order's pages become unavailable */
868 free_pages
-= z
->free_area
[o
].nr_free
<< o
;
870 /* Require fewer higher order pages to be free */
873 if (free_pages
<= min
)
880 * get_page_from_freeliest goes through the zonelist trying to allocate
884 get_page_from_freelist(gfp_t gfp_mask
, unsigned int order
,
885 struct zonelist
*zonelist
, int alloc_flags
)
887 struct zone
**z
= zonelist
->zones
;
888 struct page
*page
= NULL
;
889 int classzone_idx
= zone_idx(*z
);
892 * Go through the zonelist once, looking for a zone with enough free.
893 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
896 if (unlikely((gfp_mask
& __GFP_THISNODE
) &&
897 (*z
)->zone_pgdat
!= zonelist
->zones
[0]->zone_pgdat
))
899 if ((alloc_flags
& ALLOC_CPUSET
) &&
900 !cpuset_zone_allowed(*z
, gfp_mask
))
903 if (!(alloc_flags
& ALLOC_NO_WATERMARKS
)) {
905 if (alloc_flags
& ALLOC_WMARK_MIN
)
906 mark
= (*z
)->pages_min
;
907 else if (alloc_flags
& ALLOC_WMARK_LOW
)
908 mark
= (*z
)->pages_low
;
910 mark
= (*z
)->pages_high
;
911 if (!zone_watermark_ok(*z
, order
, mark
,
912 classzone_idx
, alloc_flags
))
913 if (!zone_reclaim_mode
||
914 !zone_reclaim(*z
, gfp_mask
, order
))
918 page
= buffered_rmqueue(zonelist
, *z
, order
, gfp_mask
);
922 } while (*(++z
) != NULL
);
927 * This is the 'heart' of the zoned buddy allocator.
929 struct page
* fastcall
930 __alloc_pages(gfp_t gfp_mask
, unsigned int order
,
931 struct zonelist
*zonelist
)
933 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
936 struct reclaim_state reclaim_state
;
937 struct task_struct
*p
= current
;
940 int did_some_progress
;
942 might_sleep_if(wait
);
945 z
= zonelist
->zones
; /* the list of zones suitable for gfp_mask */
947 if (unlikely(*z
== NULL
)) {
948 /* Should this ever happen?? */
952 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, order
,
953 zonelist
, ALLOC_WMARK_LOW
|ALLOC_CPUSET
);
958 wakeup_kswapd(*z
, order
);
962 * OK, we're below the kswapd watermark and have kicked background
963 * reclaim. Now things get more complex, so set up alloc_flags according
964 * to how we want to proceed.
966 * The caller may dip into page reserves a bit more if the caller
967 * cannot run direct reclaim, or if the caller has realtime scheduling
968 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
969 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
971 alloc_flags
= ALLOC_WMARK_MIN
;
972 if ((unlikely(rt_task(p
)) && !in_interrupt()) || !wait
)
973 alloc_flags
|= ALLOC_HARDER
;
974 if (gfp_mask
& __GFP_HIGH
)
975 alloc_flags
|= ALLOC_HIGH
;
977 alloc_flags
|= ALLOC_CPUSET
;
980 * Go through the zonelist again. Let __GFP_HIGH and allocations
981 * coming from realtime tasks go deeper into reserves.
983 * This is the last chance, in general, before the goto nopage.
984 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
985 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
987 page
= get_page_from_freelist(gfp_mask
, order
, zonelist
, alloc_flags
);
991 /* This allocation should allow future memory freeing. */
993 if (((p
->flags
& PF_MEMALLOC
) || unlikely(test_thread_flag(TIF_MEMDIE
)))
994 && !in_interrupt()) {
995 if (!(gfp_mask
& __GFP_NOMEMALLOC
)) {
997 /* go through the zonelist yet again, ignoring mins */
998 page
= get_page_from_freelist(gfp_mask
, order
,
999 zonelist
, ALLOC_NO_WATERMARKS
);
1002 if (gfp_mask
& __GFP_NOFAIL
) {
1003 blk_congestion_wait(WRITE
, HZ
/50);
1010 /* Atomic allocations - we can't balance anything */
1017 /* We now go into synchronous reclaim */
1018 cpuset_memory_pressure_bump();
1019 p
->flags
|= PF_MEMALLOC
;
1020 reclaim_state
.reclaimed_slab
= 0;
1021 p
->reclaim_state
= &reclaim_state
;
1023 did_some_progress
= try_to_free_pages(zonelist
->zones
, gfp_mask
);
1025 p
->reclaim_state
= NULL
;
1026 p
->flags
&= ~PF_MEMALLOC
;
1030 if (likely(did_some_progress
)) {
1031 page
= get_page_from_freelist(gfp_mask
, order
,
1032 zonelist
, alloc_flags
);
1035 } else if ((gfp_mask
& __GFP_FS
) && !(gfp_mask
& __GFP_NORETRY
)) {
1037 * Go through the zonelist yet one more time, keep
1038 * very high watermark here, this is only to catch
1039 * a parallel oom killing, we must fail if we're still
1040 * under heavy pressure.
1042 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, order
,
1043 zonelist
, ALLOC_WMARK_HIGH
|ALLOC_CPUSET
);
1047 out_of_memory(zonelist
, gfp_mask
, order
);
1052 * Don't let big-order allocations loop unless the caller explicitly
1053 * requests that. Wait for some write requests to complete then retry.
1055 * In this implementation, __GFP_REPEAT means __GFP_NOFAIL for order
1056 * <= 3, but that may not be true in other implementations.
1059 if (!(gfp_mask
& __GFP_NORETRY
)) {
1060 if ((order
<= 3) || (gfp_mask
& __GFP_REPEAT
))
1062 if (gfp_mask
& __GFP_NOFAIL
)
1066 blk_congestion_wait(WRITE
, HZ
/50);
1071 if (!(gfp_mask
& __GFP_NOWARN
) && printk_ratelimit()) {
1072 printk(KERN_WARNING
"%s: page allocation failure."
1073 " order:%d, mode:0x%x\n",
1074 p
->comm
, order
, gfp_mask
);
1082 EXPORT_SYMBOL(__alloc_pages
);
1085 * Common helper functions.
1087 fastcall
unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
1090 page
= alloc_pages(gfp_mask
, order
);
1093 return (unsigned long) page_address(page
);
1096 EXPORT_SYMBOL(__get_free_pages
);
1098 fastcall
unsigned long get_zeroed_page(gfp_t gfp_mask
)
1103 * get_zeroed_page() returns a 32-bit address, which cannot represent
1106 VM_BUG_ON((gfp_mask
& __GFP_HIGHMEM
) != 0);
1108 page
= alloc_pages(gfp_mask
| __GFP_ZERO
, 0);
1110 return (unsigned long) page_address(page
);
1114 EXPORT_SYMBOL(get_zeroed_page
);
1116 void __pagevec_free(struct pagevec
*pvec
)
1118 int i
= pagevec_count(pvec
);
1121 free_hot_cold_page(pvec
->pages
[i
], pvec
->cold
);
1124 fastcall
void __free_pages(struct page
*page
, unsigned int order
)
1126 if (put_page_testzero(page
)) {
1128 free_hot_page(page
);
1130 __free_pages_ok(page
, order
);
1134 EXPORT_SYMBOL(__free_pages
);
1136 fastcall
void free_pages(unsigned long addr
, unsigned int order
)
1139 VM_BUG_ON(!virt_addr_valid((void *)addr
));
1140 __free_pages(virt_to_page((void *)addr
), order
);
1144 EXPORT_SYMBOL(free_pages
);
1147 * Total amount of free (allocatable) RAM:
1149 unsigned int nr_free_pages(void)
1151 unsigned int sum
= 0;
1155 sum
+= zone
->free_pages
;
1160 EXPORT_SYMBOL(nr_free_pages
);
1163 unsigned int nr_free_pages_pgdat(pg_data_t
*pgdat
)
1165 unsigned int sum
= 0;
1168 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
1169 sum
+= pgdat
->node_zones
[i
].free_pages
;
1175 static unsigned int nr_free_zone_pages(int offset
)
1177 /* Just pick one node, since fallback list is circular */
1178 pg_data_t
*pgdat
= NODE_DATA(numa_node_id());
1179 unsigned int sum
= 0;
1181 struct zonelist
*zonelist
= pgdat
->node_zonelists
+ offset
;
1182 struct zone
**zonep
= zonelist
->zones
;
1185 for (zone
= *zonep
++; zone
; zone
= *zonep
++) {
1186 unsigned long size
= zone
->present_pages
;
1187 unsigned long high
= zone
->pages_high
;
1196 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
1198 unsigned int nr_free_buffer_pages(void)
1200 return nr_free_zone_pages(gfp_zone(GFP_USER
));
1204 * Amount of free RAM allocatable within all zones
1206 unsigned int nr_free_pagecache_pages(void)
1208 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER
));
1211 static void show_node(struct zone
*zone
)
1213 printk("Node %d ", zone
->zone_pgdat
->node_id
);
1216 #define show_node(zone) do { } while (0)
1219 void si_meminfo(struct sysinfo
*val
)
1221 val
->totalram
= totalram_pages
;
1223 val
->freeram
= nr_free_pages();
1224 val
->bufferram
= nr_blockdev_pages();
1225 val
->totalhigh
= totalhigh_pages
;
1226 val
->freehigh
= nr_free_highpages();
1227 val
->mem_unit
= PAGE_SIZE
;
1230 EXPORT_SYMBOL(si_meminfo
);
1233 void si_meminfo_node(struct sysinfo
*val
, int nid
)
1235 pg_data_t
*pgdat
= NODE_DATA(nid
);
1237 val
->totalram
= pgdat
->node_present_pages
;
1238 val
->freeram
= nr_free_pages_pgdat(pgdat
);
1239 #ifdef CONFIG_HIGHMEM
1240 val
->totalhigh
= pgdat
->node_zones
[ZONE_HIGHMEM
].present_pages
;
1241 val
->freehigh
= pgdat
->node_zones
[ZONE_HIGHMEM
].free_pages
;
1246 val
->mem_unit
= PAGE_SIZE
;
1250 #define K(x) ((x) << (PAGE_SHIFT-10))
1253 * Show free area list (used inside shift_scroll-lock stuff)
1254 * We also calculate the percentage fragmentation. We do this by counting the
1255 * memory on each free list with the exception of the first item on the list.
1257 void show_free_areas(void)
1259 int cpu
, temperature
;
1260 unsigned long active
;
1261 unsigned long inactive
;
1265 for_each_zone(zone
) {
1267 printk("%s per-cpu:", zone
->name
);
1269 if (!populated_zone(zone
)) {
1275 for_each_online_cpu(cpu
) {
1276 struct per_cpu_pageset
*pageset
;
1278 pageset
= zone_pcp(zone
, cpu
);
1280 for (temperature
= 0; temperature
< 2; temperature
++)
1281 printk("cpu %d %s: high %d, batch %d used:%d\n",
1283 temperature
? "cold" : "hot",
1284 pageset
->pcp
[temperature
].high
,
1285 pageset
->pcp
[temperature
].batch
,
1286 pageset
->pcp
[temperature
].count
);
1290 get_zone_counts(&active
, &inactive
, &free
);
1292 printk("Active:%lu inactive:%lu dirty:%lu writeback:%lu "
1293 "unstable:%lu free:%u slab:%lu mapped:%lu pagetables:%lu\n",
1296 global_page_state(NR_FILE_DIRTY
),
1297 global_page_state(NR_WRITEBACK
),
1298 global_page_state(NR_UNSTABLE_NFS
),
1300 global_page_state(NR_SLAB
),
1301 global_page_state(NR_FILE_MAPPED
),
1302 global_page_state(NR_PAGETABLE
));
1304 for_each_zone(zone
) {
1316 " pages_scanned:%lu"
1317 " all_unreclaimable? %s"
1320 K(zone
->free_pages
),
1323 K(zone
->pages_high
),
1325 K(zone
->nr_inactive
),
1326 K(zone
->present_pages
),
1327 zone
->pages_scanned
,
1328 (zone
->all_unreclaimable
? "yes" : "no")
1330 printk("lowmem_reserve[]:");
1331 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
1332 printk(" %lu", zone
->lowmem_reserve
[i
]);
1336 for_each_zone(zone
) {
1337 unsigned long nr
[MAX_ORDER
], flags
, order
, total
= 0;
1340 printk("%s: ", zone
->name
);
1341 if (!populated_zone(zone
)) {
1346 spin_lock_irqsave(&zone
->lock
, flags
);
1347 for (order
= 0; order
< MAX_ORDER
; order
++) {
1348 nr
[order
] = zone
->free_area
[order
].nr_free
;
1349 total
+= nr
[order
] << order
;
1351 spin_unlock_irqrestore(&zone
->lock
, flags
);
1352 for (order
= 0; order
< MAX_ORDER
; order
++)
1353 printk("%lu*%lukB ", nr
[order
], K(1UL) << order
);
1354 printk("= %lukB\n", K(total
));
1357 show_swap_cache_info();
1361 * Builds allocation fallback zone lists.
1363 * Add all populated zones of a node to the zonelist.
1365 static int __meminit
build_zonelists_node(pg_data_t
*pgdat
,
1366 struct zonelist
*zonelist
, int nr_zones
, enum zone_type zone_type
)
1370 BUG_ON(zone_type
>= MAX_NR_ZONES
);
1375 zone
= pgdat
->node_zones
+ zone_type
;
1376 if (populated_zone(zone
)) {
1377 zonelist
->zones
[nr_zones
++] = zone
;
1378 check_highest_zone(zone_type
);
1381 } while (zone_type
);
1386 #define MAX_NODE_LOAD (num_online_nodes())
1387 static int __meminitdata node_load
[MAX_NUMNODES
];
1389 * find_next_best_node - find the next node that should appear in a given node's fallback list
1390 * @node: node whose fallback list we're appending
1391 * @used_node_mask: nodemask_t of already used nodes
1393 * We use a number of factors to determine which is the next node that should
1394 * appear on a given node's fallback list. The node should not have appeared
1395 * already in @node's fallback list, and it should be the next closest node
1396 * according to the distance array (which contains arbitrary distance values
1397 * from each node to each node in the system), and should also prefer nodes
1398 * with no CPUs, since presumably they'll have very little allocation pressure
1399 * on them otherwise.
1400 * It returns -1 if no node is found.
1402 static int __meminit
find_next_best_node(int node
, nodemask_t
*used_node_mask
)
1405 int min_val
= INT_MAX
;
1408 /* Use the local node if we haven't already */
1409 if (!node_isset(node
, *used_node_mask
)) {
1410 node_set(node
, *used_node_mask
);
1414 for_each_online_node(n
) {
1417 /* Don't want a node to appear more than once */
1418 if (node_isset(n
, *used_node_mask
))
1421 /* Use the distance array to find the distance */
1422 val
= node_distance(node
, n
);
1424 /* Penalize nodes under us ("prefer the next node") */
1427 /* Give preference to headless and unused nodes */
1428 tmp
= node_to_cpumask(n
);
1429 if (!cpus_empty(tmp
))
1430 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
1432 /* Slight preference for less loaded node */
1433 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
1434 val
+= node_load
[n
];
1436 if (val
< min_val
) {
1443 node_set(best_node
, *used_node_mask
);
1448 static void __meminit
build_zonelists(pg_data_t
*pgdat
)
1450 int j
, node
, local_node
;
1452 int prev_node
, load
;
1453 struct zonelist
*zonelist
;
1454 nodemask_t used_mask
;
1456 /* initialize zonelists */
1457 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
1458 zonelist
= pgdat
->node_zonelists
+ i
;
1459 zonelist
->zones
[0] = NULL
;
1462 /* NUMA-aware ordering of nodes */
1463 local_node
= pgdat
->node_id
;
1464 load
= num_online_nodes();
1465 prev_node
= local_node
;
1466 nodes_clear(used_mask
);
1467 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
1468 int distance
= node_distance(local_node
, node
);
1471 * If another node is sufficiently far away then it is better
1472 * to reclaim pages in a zone before going off node.
1474 if (distance
> RECLAIM_DISTANCE
)
1475 zone_reclaim_mode
= 1;
1478 * We don't want to pressure a particular node.
1479 * So adding penalty to the first node in same
1480 * distance group to make it round-robin.
1483 if (distance
!= node_distance(local_node
, prev_node
))
1484 node_load
[node
] += load
;
1487 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
1488 zonelist
= pgdat
->node_zonelists
+ i
;
1489 for (j
= 0; zonelist
->zones
[j
] != NULL
; j
++);
1491 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
, i
);
1492 zonelist
->zones
[j
] = NULL
;
1497 #else /* CONFIG_NUMA */
1499 static void __meminit
build_zonelists(pg_data_t
*pgdat
)
1501 int node
, local_node
;
1504 local_node
= pgdat
->node_id
;
1505 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
1506 struct zonelist
*zonelist
;
1508 zonelist
= pgdat
->node_zonelists
+ i
;
1510 j
= build_zonelists_node(pgdat
, zonelist
, 0, i
);
1512 * Now we build the zonelist so that it contains the zones
1513 * of all the other nodes.
1514 * We don't want to pressure a particular node, so when
1515 * building the zones for node N, we make sure that the
1516 * zones coming right after the local ones are those from
1517 * node N+1 (modulo N)
1519 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
1520 if (!node_online(node
))
1522 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
, i
);
1524 for (node
= 0; node
< local_node
; node
++) {
1525 if (!node_online(node
))
1527 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
, i
);
1530 zonelist
->zones
[j
] = NULL
;
1534 #endif /* CONFIG_NUMA */
1536 /* return values int ....just for stop_machine_run() */
1537 static int __meminit
__build_all_zonelists(void *dummy
)
1540 for_each_online_node(nid
)
1541 build_zonelists(NODE_DATA(nid
));
1545 void __meminit
build_all_zonelists(void)
1547 if (system_state
== SYSTEM_BOOTING
) {
1548 __build_all_zonelists(0);
1549 cpuset_init_current_mems_allowed();
1551 /* we have to stop all cpus to guaranntee there is no user
1553 stop_machine_run(__build_all_zonelists
, NULL
, NR_CPUS
);
1554 /* cpuset refresh routine should be here */
1556 vm_total_pages
= nr_free_pagecache_pages();
1557 printk("Built %i zonelists. Total pages: %ld\n",
1558 num_online_nodes(), vm_total_pages
);
1562 * Helper functions to size the waitqueue hash table.
1563 * Essentially these want to choose hash table sizes sufficiently
1564 * large so that collisions trying to wait on pages are rare.
1565 * But in fact, the number of active page waitqueues on typical
1566 * systems is ridiculously low, less than 200. So this is even
1567 * conservative, even though it seems large.
1569 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
1570 * waitqueues, i.e. the size of the waitq table given the number of pages.
1572 #define PAGES_PER_WAITQUEUE 256
1574 #ifndef CONFIG_MEMORY_HOTPLUG
1575 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
1577 unsigned long size
= 1;
1579 pages
/= PAGES_PER_WAITQUEUE
;
1581 while (size
< pages
)
1585 * Once we have dozens or even hundreds of threads sleeping
1586 * on IO we've got bigger problems than wait queue collision.
1587 * Limit the size of the wait table to a reasonable size.
1589 size
= min(size
, 4096UL);
1591 return max(size
, 4UL);
1595 * A zone's size might be changed by hot-add, so it is not possible to determine
1596 * a suitable size for its wait_table. So we use the maximum size now.
1598 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
1600 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
1601 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
1602 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
1604 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
1605 * or more by the traditional way. (See above). It equals:
1607 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
1608 * ia64(16K page size) : = ( 8G + 4M)byte.
1609 * powerpc (64K page size) : = (32G +16M)byte.
1611 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
1618 * This is an integer logarithm so that shifts can be used later
1619 * to extract the more random high bits from the multiplicative
1620 * hash function before the remainder is taken.
1622 static inline unsigned long wait_table_bits(unsigned long size
)
1627 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
1629 static void __init
calculate_zone_totalpages(struct pglist_data
*pgdat
,
1630 unsigned long *zones_size
, unsigned long *zholes_size
)
1632 unsigned long realtotalpages
, totalpages
= 0;
1635 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
1636 totalpages
+= zones_size
[i
];
1637 pgdat
->node_spanned_pages
= totalpages
;
1639 realtotalpages
= totalpages
;
1641 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
1642 realtotalpages
-= zholes_size
[i
];
1643 pgdat
->node_present_pages
= realtotalpages
;
1644 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
, realtotalpages
);
1649 * Initially all pages are reserved - free ones are freed
1650 * up by free_all_bootmem() once the early boot process is
1651 * done. Non-atomic initialization, single-pass.
1653 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
1654 unsigned long start_pfn
)
1657 unsigned long end_pfn
= start_pfn
+ size
;
1660 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
1661 if (!early_pfn_valid(pfn
))
1663 page
= pfn_to_page(pfn
);
1664 set_page_links(page
, zone
, nid
, pfn
);
1665 init_page_count(page
);
1666 reset_page_mapcount(page
);
1667 SetPageReserved(page
);
1668 INIT_LIST_HEAD(&page
->lru
);
1669 #ifdef WANT_PAGE_VIRTUAL
1670 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1671 if (!is_highmem_idx(zone
))
1672 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
1677 void zone_init_free_lists(struct pglist_data
*pgdat
, struct zone
*zone
,
1681 for (order
= 0; order
< MAX_ORDER
; order
++) {
1682 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
);
1683 zone
->free_area
[order
].nr_free
= 0;
1687 #define ZONETABLE_INDEX(x, zone_nr) ((x << ZONES_SHIFT) | zone_nr)
1688 void zonetable_add(struct zone
*zone
, int nid
, enum zone_type zid
,
1689 unsigned long pfn
, unsigned long size
)
1691 unsigned long snum
= pfn_to_section_nr(pfn
);
1692 unsigned long end
= pfn_to_section_nr(pfn
+ size
);
1695 zone_table
[ZONETABLE_INDEX(nid
, zid
)] = zone
;
1697 for (; snum
<= end
; snum
++)
1698 zone_table
[ZONETABLE_INDEX(snum
, zid
)] = zone
;
1701 #ifndef __HAVE_ARCH_MEMMAP_INIT
1702 #define memmap_init(size, nid, zone, start_pfn) \
1703 memmap_init_zone((size), (nid), (zone), (start_pfn))
1706 static int __cpuinit
zone_batchsize(struct zone
*zone
)
1711 * The per-cpu-pages pools are set to around 1000th of the
1712 * size of the zone. But no more than 1/2 of a meg.
1714 * OK, so we don't know how big the cache is. So guess.
1716 batch
= zone
->present_pages
/ 1024;
1717 if (batch
* PAGE_SIZE
> 512 * 1024)
1718 batch
= (512 * 1024) / PAGE_SIZE
;
1719 batch
/= 4; /* We effectively *= 4 below */
1724 * Clamp the batch to a 2^n - 1 value. Having a power
1725 * of 2 value was found to be more likely to have
1726 * suboptimal cache aliasing properties in some cases.
1728 * For example if 2 tasks are alternately allocating
1729 * batches of pages, one task can end up with a lot
1730 * of pages of one half of the possible page colors
1731 * and the other with pages of the other colors.
1733 batch
= (1 << (fls(batch
+ batch
/2)-1)) - 1;
1738 inline void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
1740 struct per_cpu_pages
*pcp
;
1742 memset(p
, 0, sizeof(*p
));
1744 pcp
= &p
->pcp
[0]; /* hot */
1746 pcp
->high
= 6 * batch
;
1747 pcp
->batch
= max(1UL, 1 * batch
);
1748 INIT_LIST_HEAD(&pcp
->list
);
1750 pcp
= &p
->pcp
[1]; /* cold*/
1752 pcp
->high
= 2 * batch
;
1753 pcp
->batch
= max(1UL, batch
/2);
1754 INIT_LIST_HEAD(&pcp
->list
);
1758 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
1759 * to the value high for the pageset p.
1762 static void setup_pagelist_highmark(struct per_cpu_pageset
*p
,
1765 struct per_cpu_pages
*pcp
;
1767 pcp
= &p
->pcp
[0]; /* hot list */
1769 pcp
->batch
= max(1UL, high
/4);
1770 if ((high
/4) > (PAGE_SHIFT
* 8))
1771 pcp
->batch
= PAGE_SHIFT
* 8;
1777 * Boot pageset table. One per cpu which is going to be used for all
1778 * zones and all nodes. The parameters will be set in such a way
1779 * that an item put on a list will immediately be handed over to
1780 * the buddy list. This is safe since pageset manipulation is done
1781 * with interrupts disabled.
1783 * Some NUMA counter updates may also be caught by the boot pagesets.
1785 * The boot_pagesets must be kept even after bootup is complete for
1786 * unused processors and/or zones. They do play a role for bootstrapping
1787 * hotplugged processors.
1789 * zoneinfo_show() and maybe other functions do
1790 * not check if the processor is online before following the pageset pointer.
1791 * Other parts of the kernel may not check if the zone is available.
1793 static struct per_cpu_pageset boot_pageset
[NR_CPUS
];
1796 * Dynamically allocate memory for the
1797 * per cpu pageset array in struct zone.
1799 static int __cpuinit
process_zones(int cpu
)
1801 struct zone
*zone
, *dzone
;
1803 for_each_zone(zone
) {
1805 zone_pcp(zone
, cpu
) = kmalloc_node(sizeof(struct per_cpu_pageset
),
1806 GFP_KERNEL
, cpu_to_node(cpu
));
1807 if (!zone_pcp(zone
, cpu
))
1810 setup_pageset(zone_pcp(zone
, cpu
), zone_batchsize(zone
));
1812 if (percpu_pagelist_fraction
)
1813 setup_pagelist_highmark(zone_pcp(zone
, cpu
),
1814 (zone
->present_pages
/ percpu_pagelist_fraction
));
1819 for_each_zone(dzone
) {
1822 kfree(zone_pcp(dzone
, cpu
));
1823 zone_pcp(dzone
, cpu
) = NULL
;
1828 static inline void free_zone_pagesets(int cpu
)
1832 for_each_zone(zone
) {
1833 struct per_cpu_pageset
*pset
= zone_pcp(zone
, cpu
);
1835 /* Free per_cpu_pageset if it is slab allocated */
1836 if (pset
!= &boot_pageset
[cpu
])
1838 zone_pcp(zone
, cpu
) = NULL
;
1842 static int __cpuinit
pageset_cpuup_callback(struct notifier_block
*nfb
,
1843 unsigned long action
,
1846 int cpu
= (long)hcpu
;
1847 int ret
= NOTIFY_OK
;
1850 case CPU_UP_PREPARE
:
1851 if (process_zones(cpu
))
1854 case CPU_UP_CANCELED
:
1856 free_zone_pagesets(cpu
);
1864 static struct notifier_block __cpuinitdata pageset_notifier
=
1865 { &pageset_cpuup_callback
, NULL
, 0 };
1867 void __init
setup_per_cpu_pageset(void)
1871 /* Initialize per_cpu_pageset for cpu 0.
1872 * A cpuup callback will do this for every cpu
1873 * as it comes online
1875 err
= process_zones(smp_processor_id());
1877 register_cpu_notifier(&pageset_notifier
);
1883 int zone_wait_table_init(struct zone
*zone
, unsigned long zone_size_pages
)
1886 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
1890 * The per-page waitqueue mechanism uses hashed waitqueues
1893 zone
->wait_table_hash_nr_entries
=
1894 wait_table_hash_nr_entries(zone_size_pages
);
1895 zone
->wait_table_bits
=
1896 wait_table_bits(zone
->wait_table_hash_nr_entries
);
1897 alloc_size
= zone
->wait_table_hash_nr_entries
1898 * sizeof(wait_queue_head_t
);
1900 if (system_state
== SYSTEM_BOOTING
) {
1901 zone
->wait_table
= (wait_queue_head_t
*)
1902 alloc_bootmem_node(pgdat
, alloc_size
);
1905 * This case means that a zone whose size was 0 gets new memory
1906 * via memory hot-add.
1907 * But it may be the case that a new node was hot-added. In
1908 * this case vmalloc() will not be able to use this new node's
1909 * memory - this wait_table must be initialized to use this new
1910 * node itself as well.
1911 * To use this new node's memory, further consideration will be
1914 zone
->wait_table
= (wait_queue_head_t
*)vmalloc(alloc_size
);
1916 if (!zone
->wait_table
)
1919 for(i
= 0; i
< zone
->wait_table_hash_nr_entries
; ++i
)
1920 init_waitqueue_head(zone
->wait_table
+ i
);
1925 static __meminit
void zone_pcp_init(struct zone
*zone
)
1928 unsigned long batch
= zone_batchsize(zone
);
1930 for (cpu
= 0; cpu
< NR_CPUS
; cpu
++) {
1932 /* Early boot. Slab allocator not functional yet */
1933 zone_pcp(zone
, cpu
) = &boot_pageset
[cpu
];
1934 setup_pageset(&boot_pageset
[cpu
],0);
1936 setup_pageset(zone_pcp(zone
,cpu
), batch
);
1939 if (zone
->present_pages
)
1940 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%lu\n",
1941 zone
->name
, zone
->present_pages
, batch
);
1944 __meminit
int init_currently_empty_zone(struct zone
*zone
,
1945 unsigned long zone_start_pfn
,
1948 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
1950 ret
= zone_wait_table_init(zone
, size
);
1953 pgdat
->nr_zones
= zone_idx(zone
) + 1;
1955 zone
->zone_start_pfn
= zone_start_pfn
;
1957 memmap_init(size
, pgdat
->node_id
, zone_idx(zone
), zone_start_pfn
);
1959 zone_init_free_lists(pgdat
, zone
, zone
->spanned_pages
);
1965 * Set up the zone data structures:
1966 * - mark all pages reserved
1967 * - mark all memory queues empty
1968 * - clear the memory bitmaps
1970 static void __meminit
free_area_init_core(struct pglist_data
*pgdat
,
1971 unsigned long *zones_size
, unsigned long *zholes_size
)
1974 int nid
= pgdat
->node_id
;
1975 unsigned long zone_start_pfn
= pgdat
->node_start_pfn
;
1978 pgdat_resize_init(pgdat
);
1979 pgdat
->nr_zones
= 0;
1980 init_waitqueue_head(&pgdat
->kswapd_wait
);
1981 pgdat
->kswapd_max_order
= 0;
1983 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
1984 struct zone
*zone
= pgdat
->node_zones
+ j
;
1985 unsigned long size
, realsize
;
1987 realsize
= size
= zones_size
[j
];
1989 realsize
-= zholes_size
[j
];
1991 if (!is_highmem_idx(j
))
1992 nr_kernel_pages
+= realsize
;
1993 nr_all_pages
+= realsize
;
1995 zone
->spanned_pages
= size
;
1996 zone
->present_pages
= realsize
;
1998 zone
->min_unmapped_ratio
= (realsize
*sysctl_min_unmapped_ratio
)
2001 zone
->name
= zone_names
[j
];
2002 spin_lock_init(&zone
->lock
);
2003 spin_lock_init(&zone
->lru_lock
);
2004 zone_seqlock_init(zone
);
2005 zone
->zone_pgdat
= pgdat
;
2006 zone
->free_pages
= 0;
2008 zone
->temp_priority
= zone
->prev_priority
= DEF_PRIORITY
;
2010 zone_pcp_init(zone
);
2011 INIT_LIST_HEAD(&zone
->active_list
);
2012 INIT_LIST_HEAD(&zone
->inactive_list
);
2013 zone
->nr_scan_active
= 0;
2014 zone
->nr_scan_inactive
= 0;
2015 zone
->nr_active
= 0;
2016 zone
->nr_inactive
= 0;
2017 zap_zone_vm_stats(zone
);
2018 atomic_set(&zone
->reclaim_in_progress
, 0);
2022 zonetable_add(zone
, nid
, j
, zone_start_pfn
, size
);
2023 ret
= init_currently_empty_zone(zone
, zone_start_pfn
, size
);
2025 zone_start_pfn
+= size
;
2029 static void __init
alloc_node_mem_map(struct pglist_data
*pgdat
)
2031 /* Skip empty nodes */
2032 if (!pgdat
->node_spanned_pages
)
2035 #ifdef CONFIG_FLAT_NODE_MEM_MAP
2036 /* ia64 gets its own node_mem_map, before this, without bootmem */
2037 if (!pgdat
->node_mem_map
) {
2038 unsigned long size
, start
, end
;
2042 * The zone's endpoints aren't required to be MAX_ORDER
2043 * aligned but the node_mem_map endpoints must be in order
2044 * for the buddy allocator to function correctly.
2046 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
2047 end
= pgdat
->node_start_pfn
+ pgdat
->node_spanned_pages
;
2048 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
2049 size
= (end
- start
) * sizeof(struct page
);
2050 map
= alloc_remap(pgdat
->node_id
, size
);
2052 map
= alloc_bootmem_node(pgdat
, size
);
2053 pgdat
->node_mem_map
= map
+ (pgdat
->node_start_pfn
- start
);
2055 #ifdef CONFIG_FLATMEM
2057 * With no DISCONTIG, the global mem_map is just set as node 0's
2059 if (pgdat
== NODE_DATA(0))
2060 mem_map
= NODE_DATA(0)->node_mem_map
;
2062 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
2065 void __meminit
free_area_init_node(int nid
, struct pglist_data
*pgdat
,
2066 unsigned long *zones_size
, unsigned long node_start_pfn
,
2067 unsigned long *zholes_size
)
2069 pgdat
->node_id
= nid
;
2070 pgdat
->node_start_pfn
= node_start_pfn
;
2071 calculate_zone_totalpages(pgdat
, zones_size
, zholes_size
);
2073 alloc_node_mem_map(pgdat
);
2075 free_area_init_core(pgdat
, zones_size
, zholes_size
);
2078 #ifndef CONFIG_NEED_MULTIPLE_NODES
2079 static bootmem_data_t contig_bootmem_data
;
2080 struct pglist_data contig_page_data
= { .bdata
= &contig_bootmem_data
};
2082 EXPORT_SYMBOL(contig_page_data
);
2085 void __init
free_area_init(unsigned long *zones_size
)
2087 free_area_init_node(0, NODE_DATA(0), zones_size
,
2088 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
2091 #ifdef CONFIG_HOTPLUG_CPU
2092 static int page_alloc_cpu_notify(struct notifier_block
*self
,
2093 unsigned long action
, void *hcpu
)
2095 int cpu
= (unsigned long)hcpu
;
2097 if (action
== CPU_DEAD
) {
2098 local_irq_disable();
2100 vm_events_fold_cpu(cpu
);
2102 refresh_cpu_vm_stats(cpu
);
2106 #endif /* CONFIG_HOTPLUG_CPU */
2108 void __init
page_alloc_init(void)
2110 hotcpu_notifier(page_alloc_cpu_notify
, 0);
2114 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
2115 * or min_free_kbytes changes.
2117 static void calculate_totalreserve_pages(void)
2119 struct pglist_data
*pgdat
;
2120 unsigned long reserve_pages
= 0;
2121 enum zone_type i
, j
;
2123 for_each_online_pgdat(pgdat
) {
2124 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
2125 struct zone
*zone
= pgdat
->node_zones
+ i
;
2126 unsigned long max
= 0;
2128 /* Find valid and maximum lowmem_reserve in the zone */
2129 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
2130 if (zone
->lowmem_reserve
[j
] > max
)
2131 max
= zone
->lowmem_reserve
[j
];
2134 /* we treat pages_high as reserved pages. */
2135 max
+= zone
->pages_high
;
2137 if (max
> zone
->present_pages
)
2138 max
= zone
->present_pages
;
2139 reserve_pages
+= max
;
2142 totalreserve_pages
= reserve_pages
;
2146 * setup_per_zone_lowmem_reserve - called whenever
2147 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
2148 * has a correct pages reserved value, so an adequate number of
2149 * pages are left in the zone after a successful __alloc_pages().
2151 static void setup_per_zone_lowmem_reserve(void)
2153 struct pglist_data
*pgdat
;
2154 enum zone_type j
, idx
;
2156 for_each_online_pgdat(pgdat
) {
2157 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
2158 struct zone
*zone
= pgdat
->node_zones
+ j
;
2159 unsigned long present_pages
= zone
->present_pages
;
2161 zone
->lowmem_reserve
[j
] = 0;
2165 struct zone
*lower_zone
;
2169 if (sysctl_lowmem_reserve_ratio
[idx
] < 1)
2170 sysctl_lowmem_reserve_ratio
[idx
] = 1;
2172 lower_zone
= pgdat
->node_zones
+ idx
;
2173 lower_zone
->lowmem_reserve
[j
] = present_pages
/
2174 sysctl_lowmem_reserve_ratio
[idx
];
2175 present_pages
+= lower_zone
->present_pages
;
2180 /* update totalreserve_pages */
2181 calculate_totalreserve_pages();
2185 * setup_per_zone_pages_min - called when min_free_kbytes changes. Ensures
2186 * that the pages_{min,low,high} values for each zone are set correctly
2187 * with respect to min_free_kbytes.
2189 void setup_per_zone_pages_min(void)
2191 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
2192 unsigned long lowmem_pages
= 0;
2194 unsigned long flags
;
2196 /* Calculate total number of !ZONE_HIGHMEM pages */
2197 for_each_zone(zone
) {
2198 if (!is_highmem(zone
))
2199 lowmem_pages
+= zone
->present_pages
;
2202 for_each_zone(zone
) {
2205 spin_lock_irqsave(&zone
->lru_lock
, flags
);
2206 tmp
= (u64
)pages_min
* zone
->present_pages
;
2207 do_div(tmp
, lowmem_pages
);
2208 if (is_highmem(zone
)) {
2210 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
2211 * need highmem pages, so cap pages_min to a small
2214 * The (pages_high-pages_low) and (pages_low-pages_min)
2215 * deltas controls asynch page reclaim, and so should
2216 * not be capped for highmem.
2220 min_pages
= zone
->present_pages
/ 1024;
2221 if (min_pages
< SWAP_CLUSTER_MAX
)
2222 min_pages
= SWAP_CLUSTER_MAX
;
2223 if (min_pages
> 128)
2225 zone
->pages_min
= min_pages
;
2228 * If it's a lowmem zone, reserve a number of pages
2229 * proportionate to the zone's size.
2231 zone
->pages_min
= tmp
;
2234 zone
->pages_low
= zone
->pages_min
+ (tmp
>> 2);
2235 zone
->pages_high
= zone
->pages_min
+ (tmp
>> 1);
2236 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
2239 /* update totalreserve_pages */
2240 calculate_totalreserve_pages();
2244 * Initialise min_free_kbytes.
2246 * For small machines we want it small (128k min). For large machines
2247 * we want it large (64MB max). But it is not linear, because network
2248 * bandwidth does not increase linearly with machine size. We use
2250 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
2251 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
2267 static int __init
init_per_zone_pages_min(void)
2269 unsigned long lowmem_kbytes
;
2271 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
2273 min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
2274 if (min_free_kbytes
< 128)
2275 min_free_kbytes
= 128;
2276 if (min_free_kbytes
> 65536)
2277 min_free_kbytes
= 65536;
2278 setup_per_zone_pages_min();
2279 setup_per_zone_lowmem_reserve();
2282 module_init(init_per_zone_pages_min
)
2285 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
2286 * that we can call two helper functions whenever min_free_kbytes
2289 int min_free_kbytes_sysctl_handler(ctl_table
*table
, int write
,
2290 struct file
*file
, void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2292 proc_dointvec(table
, write
, file
, buffer
, length
, ppos
);
2293 setup_per_zone_pages_min();
2298 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table
*table
, int write
,
2299 struct file
*file
, void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2304 rc
= proc_dointvec_minmax(table
, write
, file
, buffer
, length
, ppos
);
2309 zone
->min_unmapped_ratio
= (zone
->present_pages
*
2310 sysctl_min_unmapped_ratio
) / 100;
2316 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
2317 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
2318 * whenever sysctl_lowmem_reserve_ratio changes.
2320 * The reserve ratio obviously has absolutely no relation with the
2321 * pages_min watermarks. The lowmem reserve ratio can only make sense
2322 * if in function of the boot time zone sizes.
2324 int lowmem_reserve_ratio_sysctl_handler(ctl_table
*table
, int write
,
2325 struct file
*file
, void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2327 proc_dointvec_minmax(table
, write
, file
, buffer
, length
, ppos
);
2328 setup_per_zone_lowmem_reserve();
2333 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
2334 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
2335 * can have before it gets flushed back to buddy allocator.
2338 int percpu_pagelist_fraction_sysctl_handler(ctl_table
*table
, int write
,
2339 struct file
*file
, void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2345 ret
= proc_dointvec_minmax(table
, write
, file
, buffer
, length
, ppos
);
2346 if (!write
|| (ret
== -EINVAL
))
2348 for_each_zone(zone
) {
2349 for_each_online_cpu(cpu
) {
2351 high
= zone
->present_pages
/ percpu_pagelist_fraction
;
2352 setup_pagelist_highmark(zone_pcp(zone
, cpu
), high
);
2358 int hashdist
= HASHDIST_DEFAULT
;
2361 static int __init
set_hashdist(char *str
)
2365 hashdist
= simple_strtoul(str
, &str
, 0);
2368 __setup("hashdist=", set_hashdist
);
2372 * allocate a large system hash table from bootmem
2373 * - it is assumed that the hash table must contain an exact power-of-2
2374 * quantity of entries
2375 * - limit is the number of hash buckets, not the total allocation size
2377 void *__init
alloc_large_system_hash(const char *tablename
,
2378 unsigned long bucketsize
,
2379 unsigned long numentries
,
2382 unsigned int *_hash_shift
,
2383 unsigned int *_hash_mask
,
2384 unsigned long limit
)
2386 unsigned long long max
= limit
;
2387 unsigned long log2qty
, size
;
2390 /* allow the kernel cmdline to have a say */
2392 /* round applicable memory size up to nearest megabyte */
2393 numentries
= (flags
& HASH_HIGHMEM
) ? nr_all_pages
: nr_kernel_pages
;
2394 numentries
+= (1UL << (20 - PAGE_SHIFT
)) - 1;
2395 numentries
>>= 20 - PAGE_SHIFT
;
2396 numentries
<<= 20 - PAGE_SHIFT
;
2398 /* limit to 1 bucket per 2^scale bytes of low memory */
2399 if (scale
> PAGE_SHIFT
)
2400 numentries
>>= (scale
- PAGE_SHIFT
);
2402 numentries
<<= (PAGE_SHIFT
- scale
);
2404 numentries
= roundup_pow_of_two(numentries
);
2406 /* limit allocation size to 1/16 total memory by default */
2408 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
2409 do_div(max
, bucketsize
);
2412 if (numentries
> max
)
2415 log2qty
= long_log2(numentries
);
2418 size
= bucketsize
<< log2qty
;
2419 if (flags
& HASH_EARLY
)
2420 table
= alloc_bootmem(size
);
2422 table
= __vmalloc(size
, GFP_ATOMIC
, PAGE_KERNEL
);
2424 unsigned long order
;
2425 for (order
= 0; ((1UL << order
) << PAGE_SHIFT
) < size
; order
++)
2427 table
= (void*) __get_free_pages(GFP_ATOMIC
, order
);
2429 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
2432 panic("Failed to allocate %s hash table\n", tablename
);
2434 printk("%s hash table entries: %d (order: %d, %lu bytes)\n",
2437 long_log2(size
) - PAGE_SHIFT
,
2441 *_hash_shift
= log2qty
;
2443 *_hash_mask
= (1 << log2qty
) - 1;
2448 #ifdef CONFIG_OUT_OF_LINE_PFN_TO_PAGE
2449 struct page
*pfn_to_page(unsigned long pfn
)
2451 return __pfn_to_page(pfn
);
2453 unsigned long page_to_pfn(struct page
*page
)
2455 return __page_to_pfn(page
);
2457 EXPORT_SYMBOL(pfn_to_page
);
2458 EXPORT_SYMBOL(page_to_pfn
);
2459 #endif /* CONFIG_OUT_OF_LINE_PFN_TO_PAGE */