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/config.h>
18 #include <linux/stddef.h>
20 #include <linux/swap.h>
21 #include <linux/interrupt.h>
22 #include <linux/pagemap.h>
23 #include <linux/bootmem.h>
24 #include <linux/compiler.h>
25 #include <linux/kernel.h>
26 #include <linux/module.h>
27 #include <linux/suspend.h>
28 #include <linux/pagevec.h>
29 #include <linux/blkdev.h>
30 #include <linux/slab.h>
31 #include <linux/notifier.h>
32 #include <linux/topology.h>
33 #include <linux/sysctl.h>
34 #include <linux/cpu.h>
35 #include <linux/cpuset.h>
36 #include <linux/memory_hotplug.h>
37 #include <linux/nodemask.h>
38 #include <linux/vmalloc.h>
39 #include <linux/mempolicy.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 totalhigh_pages __read_mostly
;
55 unsigned long totalreserve_pages __read_mostly
;
57 int percpu_pagelist_fraction
;
59 static void __free_pages_ok(struct page
*page
, unsigned int order
);
62 * results with 256, 32 in the lowmem_reserve sysctl:
63 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
64 * 1G machine -> (16M dma, 784M normal, 224M high)
65 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
66 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
67 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
69 * TBD: should special case ZONE_DMA32 machines here - in those we normally
70 * don't need any ZONE_NORMAL reservation
72 int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
-1] = { 256, 256, 32 };
74 EXPORT_SYMBOL(totalram_pages
);
77 * Used by page_zone() to look up the address of the struct zone whose
78 * id is encoded in the upper bits of page->flags
80 struct zone
*zone_table
[1 << ZONETABLE_SHIFT
] __read_mostly
;
81 EXPORT_SYMBOL(zone_table
);
83 static char *zone_names
[MAX_NR_ZONES
] = { "DMA", "DMA32", "Normal", "HighMem" };
84 int min_free_kbytes
= 1024;
86 unsigned long __meminitdata nr_kernel_pages
;
87 unsigned long __meminitdata nr_all_pages
;
89 #ifdef CONFIG_DEBUG_VM
90 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
94 unsigned long pfn
= page_to_pfn(page
);
97 seq
= zone_span_seqbegin(zone
);
98 if (pfn
>= zone
->zone_start_pfn
+ zone
->spanned_pages
)
100 else if (pfn
< zone
->zone_start_pfn
)
102 } while (zone_span_seqretry(zone
, seq
));
107 static int page_is_consistent(struct zone
*zone
, struct page
*page
)
109 #ifdef CONFIG_HOLES_IN_ZONE
110 if (!pfn_valid(page_to_pfn(page
)))
113 if (zone
!= page_zone(page
))
119 * Temporary debugging check for pages not lying within a given zone.
121 static int bad_range(struct zone
*zone
, struct page
*page
)
123 if (page_outside_zone_boundaries(zone
, page
))
125 if (!page_is_consistent(zone
, page
))
132 static inline int bad_range(struct zone
*zone
, struct page
*page
)
138 static void bad_page(struct page
*page
)
140 printk(KERN_EMERG
"Bad page state in process '%s'\n"
141 KERN_EMERG
"page:%p flags:0x%0*lx mapping:%p mapcount:%d count:%d\n"
142 KERN_EMERG
"Trying to fix it up, but a reboot is needed\n"
143 KERN_EMERG
"Backtrace:\n",
144 current
->comm
, page
, (int)(2*sizeof(unsigned long)),
145 (unsigned long)page
->flags
, page
->mapping
,
146 page_mapcount(page
), page_count(page
));
148 page
->flags
&= ~(1 << PG_lru
|
158 set_page_count(page
, 0);
159 reset_page_mapcount(page
);
160 page
->mapping
= NULL
;
161 add_taint(TAINT_BAD_PAGE
);
165 * Higher-order pages are called "compound pages". They are structured thusly:
167 * The first PAGE_SIZE page is called the "head page".
169 * The remaining PAGE_SIZE pages are called "tail pages".
171 * All pages have PG_compound set. All pages have their ->private pointing at
172 * the head page (even the head page has this).
174 * The first tail page's ->lru.next holds the address of the compound page's
175 * put_page() function. Its ->lru.prev holds the order of allocation.
176 * This usage means that zero-order pages may not be compound.
179 static void free_compound_page(struct page
*page
)
181 __free_pages_ok(page
, (unsigned long)page
[1].lru
.prev
);
184 static void prep_compound_page(struct page
*page
, unsigned long order
)
187 int nr_pages
= 1 << order
;
189 page
[1].lru
.next
= (void *)free_compound_page
; /* set dtor */
190 page
[1].lru
.prev
= (void *)order
;
191 for (i
= 0; i
< nr_pages
; i
++) {
192 struct page
*p
= page
+ i
;
194 __SetPageCompound(p
);
195 set_page_private(p
, (unsigned long)page
);
199 static void destroy_compound_page(struct page
*page
, unsigned long order
)
202 int nr_pages
= 1 << order
;
204 if (unlikely((unsigned long)page
[1].lru
.prev
!= order
))
207 for (i
= 0; i
< nr_pages
; i
++) {
208 struct page
*p
= page
+ i
;
210 if (unlikely(!PageCompound(p
) |
211 (page_private(p
) != (unsigned long)page
)))
213 __ClearPageCompound(p
);
217 static inline void prep_zero_page(struct page
*page
, int order
, gfp_t gfp_flags
)
221 BUG_ON((gfp_flags
& (__GFP_WAIT
| __GFP_HIGHMEM
)) == __GFP_HIGHMEM
);
223 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
224 * and __GFP_HIGHMEM from hard or soft interrupt context.
226 BUG_ON((gfp_flags
& __GFP_HIGHMEM
) && in_interrupt());
227 for (i
= 0; i
< (1 << order
); i
++)
228 clear_highpage(page
+ i
);
232 * function for dealing with page's order in buddy system.
233 * zone->lock is already acquired when we use these.
234 * So, we don't need atomic page->flags operations here.
236 static inline unsigned long page_order(struct page
*page
)
238 return page_private(page
);
241 static inline void set_page_order(struct page
*page
, int order
)
243 set_page_private(page
, order
);
244 __SetPageBuddy(page
);
247 static inline void rmv_page_order(struct page
*page
)
249 __ClearPageBuddy(page
);
250 set_page_private(page
, 0);
254 * Locate the struct page for both the matching buddy in our
255 * pair (buddy1) and the combined O(n+1) page they form (page).
257 * 1) Any buddy B1 will have an order O twin B2 which satisfies
258 * the following equation:
260 * For example, if the starting buddy (buddy2) is #8 its order
262 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
264 * 2) Any buddy B will have an order O+1 parent P which
265 * satisfies the following equation:
268 * Assumption: *_mem_map is contigious at least up to MAX_ORDER
270 static inline struct page
*
271 __page_find_buddy(struct page
*page
, unsigned long page_idx
, unsigned int order
)
273 unsigned long buddy_idx
= page_idx
^ (1 << order
);
275 return page
+ (buddy_idx
- page_idx
);
278 static inline unsigned long
279 __find_combined_index(unsigned long page_idx
, unsigned int order
)
281 return (page_idx
& ~(1 << order
));
285 * This function checks whether a page is free && is the buddy
286 * we can do coalesce a page and its buddy if
287 * (a) the buddy is not in a hole &&
288 * (b) the buddy is in the buddy system &&
289 * (c) a page and its buddy have the same order &&
290 * (d) a page and its buddy are in the same zone.
292 * For recording whether a page is in the buddy system, we use PG_buddy.
293 * Setting, clearing, and testing PG_buddy is serialized by zone->lock.
295 * For recording page's order, we use page_private(page).
297 static inline int page_is_buddy(struct page
*page
, struct page
*buddy
,
300 #ifdef CONFIG_HOLES_IN_ZONE
301 if (!pfn_valid(page_to_pfn(buddy
)))
305 if (page_zone_id(page
) != page_zone_id(buddy
))
308 if (PageBuddy(buddy
) && page_order(buddy
) == order
) {
309 BUG_ON(page_count(buddy
) != 0);
316 * Freeing function for a buddy system allocator.
318 * The concept of a buddy system is to maintain direct-mapped table
319 * (containing bit values) for memory blocks of various "orders".
320 * The bottom level table contains the map for the smallest allocatable
321 * units of memory (here, pages), and each level above it describes
322 * pairs of units from the levels below, hence, "buddies".
323 * At a high level, all that happens here is marking the table entry
324 * at the bottom level available, and propagating the changes upward
325 * as necessary, plus some accounting needed to play nicely with other
326 * parts of the VM system.
327 * At each level, we keep a list of pages, which are heads of continuous
328 * free pages of length of (1 << order) and marked with PG_buddy. Page's
329 * order is recorded in page_private(page) field.
330 * So when we are allocating or freeing one, we can derive the state of the
331 * other. That is, if we allocate a small block, and both were
332 * free, the remainder of the region must be split into blocks.
333 * If a block is freed, and its buddy is also free, then this
334 * triggers coalescing into a block of larger size.
339 static inline void __free_one_page(struct page
*page
,
340 struct zone
*zone
, unsigned int order
)
342 unsigned long page_idx
;
343 int order_size
= 1 << order
;
345 if (unlikely(PageCompound(page
)))
346 destroy_compound_page(page
, order
);
348 page_idx
= page_to_pfn(page
) & ((1 << MAX_ORDER
) - 1);
350 BUG_ON(page_idx
& (order_size
- 1));
351 BUG_ON(bad_range(zone
, page
));
353 zone
->free_pages
+= order_size
;
354 while (order
< MAX_ORDER
-1) {
355 unsigned long combined_idx
;
356 struct free_area
*area
;
359 buddy
= __page_find_buddy(page
, page_idx
, order
);
360 if (!page_is_buddy(page
, buddy
, order
))
361 break; /* Move the buddy up one level. */
363 list_del(&buddy
->lru
);
364 area
= zone
->free_area
+ order
;
366 rmv_page_order(buddy
);
367 combined_idx
= __find_combined_index(page_idx
, order
);
368 page
= page
+ (combined_idx
- page_idx
);
369 page_idx
= combined_idx
;
372 set_page_order(page
, order
);
373 list_add(&page
->lru
, &zone
->free_area
[order
].free_list
);
374 zone
->free_area
[order
].nr_free
++;
377 static inline int free_pages_check(struct page
*page
)
379 if (unlikely(page_mapcount(page
) |
380 (page
->mapping
!= NULL
) |
381 (page_count(page
) != 0) |
395 __ClearPageDirty(page
);
397 * For now, we report if PG_reserved was found set, but do not
398 * clear it, and do not free the page. But we shall soon need
399 * to do more, for when the ZERO_PAGE count wraps negative.
401 return PageReserved(page
);
405 * Frees a list of pages.
406 * Assumes all pages on list are in same zone, and of same order.
407 * count is the number of pages to free.
409 * If the zone was previously in an "all pages pinned" state then look to
410 * see if this freeing clears that state.
412 * And clear the zone's pages_scanned counter, to hold off the "all pages are
413 * pinned" detection logic.
415 static void free_pages_bulk(struct zone
*zone
, int count
,
416 struct list_head
*list
, int order
)
418 spin_lock(&zone
->lock
);
419 zone
->all_unreclaimable
= 0;
420 zone
->pages_scanned
= 0;
424 BUG_ON(list_empty(list
));
425 page
= list_entry(list
->prev
, struct page
, lru
);
426 /* have to delete it as __free_one_page list manipulates */
427 list_del(&page
->lru
);
428 __free_one_page(page
, zone
, order
);
430 spin_unlock(&zone
->lock
);
433 static void free_one_page(struct zone
*zone
, struct page
*page
, int order
)
436 list_add(&page
->lru
, &list
);
437 free_pages_bulk(zone
, 1, &list
, order
);
440 static void __free_pages_ok(struct page
*page
, unsigned int order
)
446 arch_free_page(page
, order
);
447 if (!PageHighMem(page
))
448 mutex_debug_check_no_locks_freed(page_address(page
),
451 for (i
= 0 ; i
< (1 << order
) ; ++i
)
452 reserved
+= free_pages_check(page
+ i
);
456 kernel_map_pages(page
, 1 << order
, 0);
457 local_irq_save(flags
);
458 __mod_page_state(pgfree
, 1 << order
);
459 free_one_page(page_zone(page
), page
, order
);
460 local_irq_restore(flags
);
464 * permit the bootmem allocator to evade page validation on high-order frees
466 void fastcall __init
__free_pages_bootmem(struct page
*page
, unsigned int order
)
469 __ClearPageReserved(page
);
470 set_page_count(page
, 0);
471 set_page_refcounted(page
);
477 for (loop
= 0; loop
< BITS_PER_LONG
; loop
++) {
478 struct page
*p
= &page
[loop
];
480 if (loop
+ 1 < BITS_PER_LONG
)
482 __ClearPageReserved(p
);
483 set_page_count(p
, 0);
486 set_page_refcounted(page
);
487 __free_pages(page
, order
);
493 * The order of subdivision here is critical for the IO subsystem.
494 * Please do not alter this order without good reasons and regression
495 * testing. Specifically, as large blocks of memory are subdivided,
496 * the order in which smaller blocks are delivered depends on the order
497 * they're subdivided in this function. This is the primary factor
498 * influencing the order in which pages are delivered to the IO
499 * subsystem according to empirical testing, and this is also justified
500 * by considering the behavior of a buddy system containing a single
501 * large block of memory acted on by a series of small allocations.
502 * This behavior is a critical factor in sglist merging's success.
506 static inline void expand(struct zone
*zone
, struct page
*page
,
507 int low
, int high
, struct free_area
*area
)
509 unsigned long size
= 1 << high
;
515 BUG_ON(bad_range(zone
, &page
[size
]));
516 list_add(&page
[size
].lru
, &area
->free_list
);
518 set_page_order(&page
[size
], high
);
523 * This page is about to be returned from the page allocator
525 static int prep_new_page(struct page
*page
, int order
, gfp_t gfp_flags
)
527 if (unlikely(page_mapcount(page
) |
528 (page
->mapping
!= NULL
) |
529 (page_count(page
) != 0) |
545 * For now, we report if PG_reserved was found set, but do not
546 * clear it, and do not allocate the page: as a safety net.
548 if (PageReserved(page
))
551 page
->flags
&= ~(1 << PG_uptodate
| 1 << PG_error
|
552 1 << PG_referenced
| 1 << PG_arch_1
|
553 1 << PG_checked
| 1 << PG_mappedtodisk
);
554 set_page_private(page
, 0);
555 set_page_refcounted(page
);
556 kernel_map_pages(page
, 1 << order
, 1);
558 if (gfp_flags
& __GFP_ZERO
)
559 prep_zero_page(page
, order
, gfp_flags
);
561 if (order
&& (gfp_flags
& __GFP_COMP
))
562 prep_compound_page(page
, order
);
568 * Do the hard work of removing an element from the buddy allocator.
569 * Call me with the zone->lock already held.
571 static struct page
*__rmqueue(struct zone
*zone
, unsigned int order
)
573 struct free_area
* area
;
574 unsigned int current_order
;
577 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
578 area
= zone
->free_area
+ current_order
;
579 if (list_empty(&area
->free_list
))
582 page
= list_entry(area
->free_list
.next
, struct page
, lru
);
583 list_del(&page
->lru
);
584 rmv_page_order(page
);
586 zone
->free_pages
-= 1UL << order
;
587 expand(zone
, page
, order
, current_order
, area
);
595 * Obtain a specified number of elements from the buddy allocator, all under
596 * a single hold of the lock, for efficiency. Add them to the supplied list.
597 * Returns the number of new pages which were placed at *list.
599 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
600 unsigned long count
, struct list_head
*list
)
604 spin_lock(&zone
->lock
);
605 for (i
= 0; i
< count
; ++i
) {
606 struct page
*page
= __rmqueue(zone
, order
);
607 if (unlikely(page
== NULL
))
609 list_add_tail(&page
->lru
, list
);
611 spin_unlock(&zone
->lock
);
617 * Called from the slab reaper to drain pagesets on a particular node that
618 * belong to the currently executing processor.
619 * Note that this function must be called with the thread pinned to
620 * a single processor.
622 void drain_node_pages(int nodeid
)
627 for (z
= 0; z
< MAX_NR_ZONES
; z
++) {
628 struct zone
*zone
= NODE_DATA(nodeid
)->node_zones
+ z
;
629 struct per_cpu_pageset
*pset
;
631 pset
= zone_pcp(zone
, smp_processor_id());
632 for (i
= 0; i
< ARRAY_SIZE(pset
->pcp
); i
++) {
633 struct per_cpu_pages
*pcp
;
637 local_irq_save(flags
);
638 free_pages_bulk(zone
, pcp
->count
, &pcp
->list
, 0);
640 local_irq_restore(flags
);
647 #if defined(CONFIG_PM) || defined(CONFIG_HOTPLUG_CPU)
648 static void __drain_pages(unsigned int cpu
)
654 for_each_zone(zone
) {
655 struct per_cpu_pageset
*pset
;
657 pset
= zone_pcp(zone
, cpu
);
658 for (i
= 0; i
< ARRAY_SIZE(pset
->pcp
); i
++) {
659 struct per_cpu_pages
*pcp
;
662 local_irq_save(flags
);
663 free_pages_bulk(zone
, pcp
->count
, &pcp
->list
, 0);
665 local_irq_restore(flags
);
669 #endif /* CONFIG_PM || CONFIG_HOTPLUG_CPU */
673 void mark_free_pages(struct zone
*zone
)
675 unsigned long zone_pfn
, flags
;
677 struct list_head
*curr
;
679 if (!zone
->spanned_pages
)
682 spin_lock_irqsave(&zone
->lock
, flags
);
683 for (zone_pfn
= 0; zone_pfn
< zone
->spanned_pages
; ++zone_pfn
)
684 ClearPageNosaveFree(pfn_to_page(zone_pfn
+ zone
->zone_start_pfn
));
686 for (order
= MAX_ORDER
- 1; order
>= 0; --order
)
687 list_for_each(curr
, &zone
->free_area
[order
].free_list
) {
688 unsigned long start_pfn
, i
;
690 start_pfn
= page_to_pfn(list_entry(curr
, struct page
, lru
));
692 for (i
=0; i
< (1<<order
); i
++)
693 SetPageNosaveFree(pfn_to_page(start_pfn
+i
));
695 spin_unlock_irqrestore(&zone
->lock
, flags
);
699 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
701 void drain_local_pages(void)
705 local_irq_save(flags
);
706 __drain_pages(smp_processor_id());
707 local_irq_restore(flags
);
709 #endif /* CONFIG_PM */
711 static void zone_statistics(struct zonelist
*zonelist
, struct zone
*z
, int cpu
)
714 pg_data_t
*pg
= z
->zone_pgdat
;
715 pg_data_t
*orig
= zonelist
->zones
[0]->zone_pgdat
;
716 struct per_cpu_pageset
*p
;
718 p
= zone_pcp(z
, cpu
);
723 zone_pcp(zonelist
->zones
[0], cpu
)->numa_foreign
++;
725 if (pg
== NODE_DATA(numa_node_id()))
733 * Free a 0-order page
735 static void fastcall
free_hot_cold_page(struct page
*page
, int cold
)
737 struct zone
*zone
= page_zone(page
);
738 struct per_cpu_pages
*pcp
;
741 arch_free_page(page
, 0);
744 page
->mapping
= NULL
;
745 if (free_pages_check(page
))
748 kernel_map_pages(page
, 1, 0);
750 pcp
= &zone_pcp(zone
, get_cpu())->pcp
[cold
];
751 local_irq_save(flags
);
752 __inc_page_state(pgfree
);
753 list_add(&page
->lru
, &pcp
->list
);
755 if (pcp
->count
>= pcp
->high
) {
756 free_pages_bulk(zone
, pcp
->batch
, &pcp
->list
, 0);
757 pcp
->count
-= pcp
->batch
;
759 local_irq_restore(flags
);
763 void fastcall
free_hot_page(struct page
*page
)
765 free_hot_cold_page(page
, 0);
768 void fastcall
free_cold_page(struct page
*page
)
770 free_hot_cold_page(page
, 1);
774 * split_page takes a non-compound higher-order page, and splits it into
775 * n (1<<order) sub-pages: page[0..n]
776 * Each sub-page must be freed individually.
778 * Note: this is probably too low level an operation for use in drivers.
779 * Please consult with lkml before using this in your driver.
781 void split_page(struct page
*page
, unsigned int order
)
785 BUG_ON(PageCompound(page
));
786 BUG_ON(!page_count(page
));
787 for (i
= 1; i
< (1 << order
); i
++)
788 set_page_refcounted(page
+ i
);
792 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
793 * we cheat by calling it from here, in the order > 0 path. Saves a branch
796 static struct page
*buffered_rmqueue(struct zonelist
*zonelist
,
797 struct zone
*zone
, int order
, gfp_t gfp_flags
)
801 int cold
= !!(gfp_flags
& __GFP_COLD
);
806 if (likely(order
== 0)) {
807 struct per_cpu_pages
*pcp
;
809 pcp
= &zone_pcp(zone
, cpu
)->pcp
[cold
];
810 local_irq_save(flags
);
812 pcp
->count
+= rmqueue_bulk(zone
, 0,
813 pcp
->batch
, &pcp
->list
);
814 if (unlikely(!pcp
->count
))
817 page
= list_entry(pcp
->list
.next
, struct page
, lru
);
818 list_del(&page
->lru
);
821 spin_lock_irqsave(&zone
->lock
, flags
);
822 page
= __rmqueue(zone
, order
);
823 spin_unlock(&zone
->lock
);
828 __mod_page_state_zone(zone
, pgalloc
, 1 << order
);
829 zone_statistics(zonelist
, zone
, cpu
);
830 local_irq_restore(flags
);
833 BUG_ON(bad_range(zone
, page
));
834 if (prep_new_page(page
, order
, gfp_flags
))
839 local_irq_restore(flags
);
844 #define ALLOC_NO_WATERMARKS 0x01 /* don't check watermarks at all */
845 #define ALLOC_WMARK_MIN 0x02 /* use pages_min watermark */
846 #define ALLOC_WMARK_LOW 0x04 /* use pages_low watermark */
847 #define ALLOC_WMARK_HIGH 0x08 /* use pages_high watermark */
848 #define ALLOC_HARDER 0x10 /* try to alloc harder */
849 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
850 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
853 * Return 1 if free pages are above 'mark'. This takes into account the order
856 int zone_watermark_ok(struct zone
*z
, int order
, unsigned long mark
,
857 int classzone_idx
, int alloc_flags
)
859 /* free_pages my go negative - that's OK */
860 long min
= mark
, free_pages
= z
->free_pages
- (1 << order
) + 1;
863 if (alloc_flags
& ALLOC_HIGH
)
865 if (alloc_flags
& ALLOC_HARDER
)
868 if (free_pages
<= min
+ z
->lowmem_reserve
[classzone_idx
])
870 for (o
= 0; o
< order
; o
++) {
871 /* At the next order, this order's pages become unavailable */
872 free_pages
-= z
->free_area
[o
].nr_free
<< o
;
874 /* Require fewer higher order pages to be free */
877 if (free_pages
<= min
)
884 * get_page_from_freeliest goes through the zonelist trying to allocate
888 get_page_from_freelist(gfp_t gfp_mask
, unsigned int order
,
889 struct zonelist
*zonelist
, int alloc_flags
)
891 struct zone
**z
= zonelist
->zones
;
892 struct page
*page
= NULL
;
893 int classzone_idx
= zone_idx(*z
);
896 * Go through the zonelist once, looking for a zone with enough free.
897 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
900 if ((alloc_flags
& ALLOC_CPUSET
) &&
901 !cpuset_zone_allowed(*z
, gfp_mask
))
904 if (!(alloc_flags
& ALLOC_NO_WATERMARKS
)) {
906 if (alloc_flags
& ALLOC_WMARK_MIN
)
907 mark
= (*z
)->pages_min
;
908 else if (alloc_flags
& ALLOC_WMARK_LOW
)
909 mark
= (*z
)->pages_low
;
911 mark
= (*z
)->pages_high
;
912 if (!zone_watermark_ok(*z
, order
, mark
,
913 classzone_idx
, alloc_flags
))
914 if (!zone_reclaim_mode
||
915 !zone_reclaim(*z
, gfp_mask
, order
))
919 page
= buffered_rmqueue(zonelist
, *z
, order
, gfp_mask
);
923 } while (*(++z
) != NULL
);
928 * This is the 'heart' of the zoned buddy allocator.
930 struct page
* fastcall
931 __alloc_pages(gfp_t gfp_mask
, unsigned int order
,
932 struct zonelist
*zonelist
)
934 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
937 struct reclaim_state reclaim_state
;
938 struct task_struct
*p
= current
;
941 int did_some_progress
;
943 might_sleep_if(wait
);
946 z
= zonelist
->zones
; /* the list of zones suitable for gfp_mask */
948 if (unlikely(*z
== NULL
)) {
949 /* Should this ever happen?? */
953 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, order
,
954 zonelist
, ALLOC_WMARK_LOW
|ALLOC_CPUSET
);
959 if (cpuset_zone_allowed(*z
, gfp_mask
|__GFP_HARDWALL
))
960 wakeup_kswapd(*z
, order
);
964 * OK, we're below the kswapd watermark and have kicked background
965 * reclaim. Now things get more complex, so set up alloc_flags according
966 * to how we want to proceed.
968 * The caller may dip into page reserves a bit more if the caller
969 * cannot run direct reclaim, or if the caller has realtime scheduling
970 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
971 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
973 alloc_flags
= ALLOC_WMARK_MIN
;
974 if ((unlikely(rt_task(p
)) && !in_interrupt()) || !wait
)
975 alloc_flags
|= ALLOC_HARDER
;
976 if (gfp_mask
& __GFP_HIGH
)
977 alloc_flags
|= ALLOC_HIGH
;
979 alloc_flags
|= ALLOC_CPUSET
;
982 * Go through the zonelist again. Let __GFP_HIGH and allocations
983 * coming from realtime tasks go deeper into reserves.
985 * This is the last chance, in general, before the goto nopage.
986 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
987 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
989 page
= get_page_from_freelist(gfp_mask
, order
, zonelist
, alloc_flags
);
993 /* This allocation should allow future memory freeing. */
995 if (((p
->flags
& PF_MEMALLOC
) || unlikely(test_thread_flag(TIF_MEMDIE
)))
996 && !in_interrupt()) {
997 if (!(gfp_mask
& __GFP_NOMEMALLOC
)) {
999 /* go through the zonelist yet again, ignoring mins */
1000 page
= get_page_from_freelist(gfp_mask
, order
,
1001 zonelist
, ALLOC_NO_WATERMARKS
);
1004 if (gfp_mask
& __GFP_NOFAIL
) {
1005 blk_congestion_wait(WRITE
, HZ
/50);
1012 /* Atomic allocations - we can't balance anything */
1019 /* We now go into synchronous reclaim */
1020 cpuset_memory_pressure_bump();
1021 p
->flags
|= PF_MEMALLOC
;
1022 reclaim_state
.reclaimed_slab
= 0;
1023 p
->reclaim_state
= &reclaim_state
;
1025 did_some_progress
= try_to_free_pages(zonelist
->zones
, gfp_mask
);
1027 p
->reclaim_state
= NULL
;
1028 p
->flags
&= ~PF_MEMALLOC
;
1032 if (likely(did_some_progress
)) {
1033 page
= get_page_from_freelist(gfp_mask
, order
,
1034 zonelist
, alloc_flags
);
1037 } else if ((gfp_mask
& __GFP_FS
) && !(gfp_mask
& __GFP_NORETRY
)) {
1039 * Go through the zonelist yet one more time, keep
1040 * very high watermark here, this is only to catch
1041 * a parallel oom killing, we must fail if we're still
1042 * under heavy pressure.
1044 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, order
,
1045 zonelist
, ALLOC_WMARK_HIGH
|ALLOC_CPUSET
);
1049 out_of_memory(zonelist
, gfp_mask
, order
);
1054 * Don't let big-order allocations loop unless the caller explicitly
1055 * requests that. Wait for some write requests to complete then retry.
1057 * In this implementation, __GFP_REPEAT means __GFP_NOFAIL for order
1058 * <= 3, but that may not be true in other implementations.
1061 if (!(gfp_mask
& __GFP_NORETRY
)) {
1062 if ((order
<= 3) || (gfp_mask
& __GFP_REPEAT
))
1064 if (gfp_mask
& __GFP_NOFAIL
)
1068 blk_congestion_wait(WRITE
, HZ
/50);
1073 if (!(gfp_mask
& __GFP_NOWARN
) && printk_ratelimit()) {
1074 printk(KERN_WARNING
"%s: page allocation failure."
1075 " order:%d, mode:0x%x\n",
1076 p
->comm
, order
, gfp_mask
);
1084 EXPORT_SYMBOL(__alloc_pages
);
1087 * Common helper functions.
1089 fastcall
unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
1092 page
= alloc_pages(gfp_mask
, order
);
1095 return (unsigned long) page_address(page
);
1098 EXPORT_SYMBOL(__get_free_pages
);
1100 fastcall
unsigned long get_zeroed_page(gfp_t gfp_mask
)
1105 * get_zeroed_page() returns a 32-bit address, which cannot represent
1108 BUG_ON((gfp_mask
& __GFP_HIGHMEM
) != 0);
1110 page
= alloc_pages(gfp_mask
| __GFP_ZERO
, 0);
1112 return (unsigned long) page_address(page
);
1116 EXPORT_SYMBOL(get_zeroed_page
);
1118 void __pagevec_free(struct pagevec
*pvec
)
1120 int i
= pagevec_count(pvec
);
1123 free_hot_cold_page(pvec
->pages
[i
], pvec
->cold
);
1126 fastcall
void __free_pages(struct page
*page
, unsigned int order
)
1128 if (put_page_testzero(page
)) {
1130 free_hot_page(page
);
1132 __free_pages_ok(page
, order
);
1136 EXPORT_SYMBOL(__free_pages
);
1138 fastcall
void free_pages(unsigned long addr
, unsigned int order
)
1141 BUG_ON(!virt_addr_valid((void *)addr
));
1142 __free_pages(virt_to_page((void *)addr
), order
);
1146 EXPORT_SYMBOL(free_pages
);
1149 * Total amount of free (allocatable) RAM:
1151 unsigned int nr_free_pages(void)
1153 unsigned int sum
= 0;
1157 sum
+= zone
->free_pages
;
1162 EXPORT_SYMBOL(nr_free_pages
);
1165 unsigned int nr_free_pages_pgdat(pg_data_t
*pgdat
)
1167 unsigned int i
, sum
= 0;
1169 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
1170 sum
+= pgdat
->node_zones
[i
].free_pages
;
1176 static unsigned int nr_free_zone_pages(int offset
)
1178 /* Just pick one node, since fallback list is circular */
1179 pg_data_t
*pgdat
= NODE_DATA(numa_node_id());
1180 unsigned int sum
= 0;
1182 struct zonelist
*zonelist
= pgdat
->node_zonelists
+ offset
;
1183 struct zone
**zonep
= zonelist
->zones
;
1186 for (zone
= *zonep
++; zone
; zone
= *zonep
++) {
1187 unsigned long size
= zone
->present_pages
;
1188 unsigned long high
= zone
->pages_high
;
1197 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
1199 unsigned int nr_free_buffer_pages(void)
1201 return nr_free_zone_pages(gfp_zone(GFP_USER
));
1205 * Amount of free RAM allocatable within all zones
1207 unsigned int nr_free_pagecache_pages(void)
1209 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER
));
1212 #ifdef CONFIG_HIGHMEM
1213 unsigned int nr_free_highpages (void)
1216 unsigned int pages
= 0;
1218 for_each_online_pgdat(pgdat
)
1219 pages
+= pgdat
->node_zones
[ZONE_HIGHMEM
].free_pages
;
1226 static void show_node(struct zone
*zone
)
1228 printk("Node %d ", zone
->zone_pgdat
->node_id
);
1231 #define show_node(zone) do { } while (0)
1235 * Accumulate the page_state information across all CPUs.
1236 * The result is unavoidably approximate - it can change
1237 * during and after execution of this function.
1239 static DEFINE_PER_CPU(struct page_state
, page_states
) = {0};
1241 atomic_t nr_pagecache
= ATOMIC_INIT(0);
1242 EXPORT_SYMBOL(nr_pagecache
);
1244 DEFINE_PER_CPU(long, nr_pagecache_local
) = 0;
1247 static void __get_page_state(struct page_state
*ret
, int nr
, cpumask_t
*cpumask
)
1251 memset(ret
, 0, nr
* sizeof(unsigned long));
1252 cpus_and(*cpumask
, *cpumask
, cpu_online_map
);
1254 for_each_cpu_mask(cpu
, *cpumask
) {
1260 in
= (unsigned long *)&per_cpu(page_states
, cpu
);
1262 next_cpu
= next_cpu(cpu
, *cpumask
);
1263 if (likely(next_cpu
< NR_CPUS
))
1264 prefetch(&per_cpu(page_states
, next_cpu
));
1266 out
= (unsigned long *)ret
;
1267 for (off
= 0; off
< nr
; off
++)
1272 void get_page_state_node(struct page_state
*ret
, int node
)
1275 cpumask_t mask
= node_to_cpumask(node
);
1277 nr
= offsetof(struct page_state
, GET_PAGE_STATE_LAST
);
1278 nr
/= sizeof(unsigned long);
1280 __get_page_state(ret
, nr
+1, &mask
);
1283 void get_page_state(struct page_state
*ret
)
1286 cpumask_t mask
= CPU_MASK_ALL
;
1288 nr
= offsetof(struct page_state
, GET_PAGE_STATE_LAST
);
1289 nr
/= sizeof(unsigned long);
1291 __get_page_state(ret
, nr
+ 1, &mask
);
1294 void get_full_page_state(struct page_state
*ret
)
1296 cpumask_t mask
= CPU_MASK_ALL
;
1298 __get_page_state(ret
, sizeof(*ret
) / sizeof(unsigned long), &mask
);
1301 unsigned long read_page_state_offset(unsigned long offset
)
1303 unsigned long ret
= 0;
1306 for_each_online_cpu(cpu
) {
1309 in
= (unsigned long)&per_cpu(page_states
, cpu
) + offset
;
1310 ret
+= *((unsigned long *)in
);
1315 void __mod_page_state_offset(unsigned long offset
, unsigned long delta
)
1319 ptr
= &__get_cpu_var(page_states
);
1320 *(unsigned long *)(ptr
+ offset
) += delta
;
1322 EXPORT_SYMBOL(__mod_page_state_offset
);
1324 void mod_page_state_offset(unsigned long offset
, unsigned long delta
)
1326 unsigned long flags
;
1329 local_irq_save(flags
);
1330 ptr
= &__get_cpu_var(page_states
);
1331 *(unsigned long *)(ptr
+ offset
) += delta
;
1332 local_irq_restore(flags
);
1334 EXPORT_SYMBOL(mod_page_state_offset
);
1336 void __get_zone_counts(unsigned long *active
, unsigned long *inactive
,
1337 unsigned long *free
, struct pglist_data
*pgdat
)
1339 struct zone
*zones
= pgdat
->node_zones
;
1345 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
1346 *active
+= zones
[i
].nr_active
;
1347 *inactive
+= zones
[i
].nr_inactive
;
1348 *free
+= zones
[i
].free_pages
;
1352 void get_zone_counts(unsigned long *active
,
1353 unsigned long *inactive
, unsigned long *free
)
1355 struct pglist_data
*pgdat
;
1360 for_each_online_pgdat(pgdat
) {
1361 unsigned long l
, m
, n
;
1362 __get_zone_counts(&l
, &m
, &n
, pgdat
);
1369 void si_meminfo(struct sysinfo
*val
)
1371 val
->totalram
= totalram_pages
;
1373 val
->freeram
= nr_free_pages();
1374 val
->bufferram
= nr_blockdev_pages();
1375 #ifdef CONFIG_HIGHMEM
1376 val
->totalhigh
= totalhigh_pages
;
1377 val
->freehigh
= nr_free_highpages();
1382 val
->mem_unit
= PAGE_SIZE
;
1385 EXPORT_SYMBOL(si_meminfo
);
1388 void si_meminfo_node(struct sysinfo
*val
, int nid
)
1390 pg_data_t
*pgdat
= NODE_DATA(nid
);
1392 val
->totalram
= pgdat
->node_present_pages
;
1393 val
->freeram
= nr_free_pages_pgdat(pgdat
);
1394 val
->totalhigh
= pgdat
->node_zones
[ZONE_HIGHMEM
].present_pages
;
1395 val
->freehigh
= pgdat
->node_zones
[ZONE_HIGHMEM
].free_pages
;
1396 val
->mem_unit
= PAGE_SIZE
;
1400 #define K(x) ((x) << (PAGE_SHIFT-10))
1403 * Show free area list (used inside shift_scroll-lock stuff)
1404 * We also calculate the percentage fragmentation. We do this by counting the
1405 * memory on each free list with the exception of the first item on the list.
1407 void show_free_areas(void)
1409 struct page_state ps
;
1410 int cpu
, temperature
;
1411 unsigned long active
;
1412 unsigned long inactive
;
1416 for_each_zone(zone
) {
1418 printk("%s per-cpu:", zone
->name
);
1420 if (!populated_zone(zone
)) {
1426 for_each_online_cpu(cpu
) {
1427 struct per_cpu_pageset
*pageset
;
1429 pageset
= zone_pcp(zone
, cpu
);
1431 for (temperature
= 0; temperature
< 2; temperature
++)
1432 printk("cpu %d %s: high %d, batch %d used:%d\n",
1434 temperature
? "cold" : "hot",
1435 pageset
->pcp
[temperature
].high
,
1436 pageset
->pcp
[temperature
].batch
,
1437 pageset
->pcp
[temperature
].count
);
1441 get_page_state(&ps
);
1442 get_zone_counts(&active
, &inactive
, &free
);
1444 printk("Free pages: %11ukB (%ukB HighMem)\n",
1446 K(nr_free_highpages()));
1448 printk("Active:%lu inactive:%lu dirty:%lu writeback:%lu "
1449 "unstable:%lu free:%u slab:%lu mapped:%lu pagetables:%lu\n",
1458 ps
.nr_page_table_pages
);
1460 for_each_zone(zone
) {
1472 " pages_scanned:%lu"
1473 " all_unreclaimable? %s"
1476 K(zone
->free_pages
),
1479 K(zone
->pages_high
),
1481 K(zone
->nr_inactive
),
1482 K(zone
->present_pages
),
1483 zone
->pages_scanned
,
1484 (zone
->all_unreclaimable
? "yes" : "no")
1486 printk("lowmem_reserve[]:");
1487 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
1488 printk(" %lu", zone
->lowmem_reserve
[i
]);
1492 for_each_zone(zone
) {
1493 unsigned long nr
, flags
, order
, total
= 0;
1496 printk("%s: ", zone
->name
);
1497 if (!populated_zone(zone
)) {
1502 spin_lock_irqsave(&zone
->lock
, flags
);
1503 for (order
= 0; order
< MAX_ORDER
; order
++) {
1504 nr
= zone
->free_area
[order
].nr_free
;
1505 total
+= nr
<< order
;
1506 printk("%lu*%lukB ", nr
, K(1UL) << order
);
1508 spin_unlock_irqrestore(&zone
->lock
, flags
);
1509 printk("= %lukB\n", K(total
));
1512 show_swap_cache_info();
1516 * Builds allocation fallback zone lists.
1518 * Add all populated zones of a node to the zonelist.
1520 static int __meminit
build_zonelists_node(pg_data_t
*pgdat
,
1521 struct zonelist
*zonelist
, int nr_zones
, int zone_type
)
1525 BUG_ON(zone_type
> ZONE_HIGHMEM
);
1528 zone
= pgdat
->node_zones
+ zone_type
;
1529 if (populated_zone(zone
)) {
1530 #ifndef CONFIG_HIGHMEM
1531 BUG_ON(zone_type
> ZONE_NORMAL
);
1533 zonelist
->zones
[nr_zones
++] = zone
;
1534 check_highest_zone(zone_type
);
1538 } while (zone_type
>= 0);
1542 static inline int highest_zone(int zone_bits
)
1544 int res
= ZONE_NORMAL
;
1545 if (zone_bits
& (__force
int)__GFP_HIGHMEM
)
1547 if (zone_bits
& (__force
int)__GFP_DMA32
)
1549 if (zone_bits
& (__force
int)__GFP_DMA
)
1555 #define MAX_NODE_LOAD (num_online_nodes())
1556 static int __meminitdata node_load
[MAX_NUMNODES
];
1558 * find_next_best_node - find the next node that should appear in a given node's fallback list
1559 * @node: node whose fallback list we're appending
1560 * @used_node_mask: nodemask_t of already used nodes
1562 * We use a number of factors to determine which is the next node that should
1563 * appear on a given node's fallback list. The node should not have appeared
1564 * already in @node's fallback list, and it should be the next closest node
1565 * according to the distance array (which contains arbitrary distance values
1566 * from each node to each node in the system), and should also prefer nodes
1567 * with no CPUs, since presumably they'll have very little allocation pressure
1568 * on them otherwise.
1569 * It returns -1 if no node is found.
1571 static int __meminit
find_next_best_node(int node
, nodemask_t
*used_node_mask
)
1574 int min_val
= INT_MAX
;
1577 /* Use the local node if we haven't already */
1578 if (!node_isset(node
, *used_node_mask
)) {
1579 node_set(node
, *used_node_mask
);
1583 for_each_online_node(n
) {
1586 /* Don't want a node to appear more than once */
1587 if (node_isset(n
, *used_node_mask
))
1590 /* Use the distance array to find the distance */
1591 val
= node_distance(node
, n
);
1593 /* Penalize nodes under us ("prefer the next node") */
1596 /* Give preference to headless and unused nodes */
1597 tmp
= node_to_cpumask(n
);
1598 if (!cpus_empty(tmp
))
1599 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
1601 /* Slight preference for less loaded node */
1602 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
1603 val
+= node_load
[n
];
1605 if (val
< min_val
) {
1612 node_set(best_node
, *used_node_mask
);
1617 static void __meminit
build_zonelists(pg_data_t
*pgdat
)
1619 int i
, j
, k
, node
, local_node
;
1620 int prev_node
, load
;
1621 struct zonelist
*zonelist
;
1622 nodemask_t used_mask
;
1624 /* initialize zonelists */
1625 for (i
= 0; i
< GFP_ZONETYPES
; i
++) {
1626 zonelist
= pgdat
->node_zonelists
+ i
;
1627 zonelist
->zones
[0] = NULL
;
1630 /* NUMA-aware ordering of nodes */
1631 local_node
= pgdat
->node_id
;
1632 load
= num_online_nodes();
1633 prev_node
= local_node
;
1634 nodes_clear(used_mask
);
1635 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
1636 int distance
= node_distance(local_node
, node
);
1639 * If another node is sufficiently far away then it is better
1640 * to reclaim pages in a zone before going off node.
1642 if (distance
> RECLAIM_DISTANCE
)
1643 zone_reclaim_mode
= 1;
1646 * We don't want to pressure a particular node.
1647 * So adding penalty to the first node in same
1648 * distance group to make it round-robin.
1651 if (distance
!= node_distance(local_node
, prev_node
))
1652 node_load
[node
] += load
;
1655 for (i
= 0; i
< GFP_ZONETYPES
; i
++) {
1656 zonelist
= pgdat
->node_zonelists
+ i
;
1657 for (j
= 0; zonelist
->zones
[j
] != NULL
; j
++);
1659 k
= highest_zone(i
);
1661 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
, k
);
1662 zonelist
->zones
[j
] = NULL
;
1667 #else /* CONFIG_NUMA */
1669 static void __meminit
build_zonelists(pg_data_t
*pgdat
)
1671 int i
, j
, k
, node
, local_node
;
1673 local_node
= pgdat
->node_id
;
1674 for (i
= 0; i
< GFP_ZONETYPES
; i
++) {
1675 struct zonelist
*zonelist
;
1677 zonelist
= pgdat
->node_zonelists
+ i
;
1680 k
= highest_zone(i
);
1681 j
= build_zonelists_node(pgdat
, zonelist
, j
, k
);
1683 * Now we build the zonelist so that it contains the zones
1684 * of all the other nodes.
1685 * We don't want to pressure a particular node, so when
1686 * building the zones for node N, we make sure that the
1687 * zones coming right after the local ones are those from
1688 * node N+1 (modulo N)
1690 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
1691 if (!node_online(node
))
1693 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
, k
);
1695 for (node
= 0; node
< local_node
; node
++) {
1696 if (!node_online(node
))
1698 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
, k
);
1701 zonelist
->zones
[j
] = NULL
;
1705 #endif /* CONFIG_NUMA */
1707 void __init
build_all_zonelists(void)
1711 for_each_online_node(i
)
1712 build_zonelists(NODE_DATA(i
));
1713 printk("Built %i zonelists\n", num_online_nodes());
1714 cpuset_init_current_mems_allowed();
1718 * Helper functions to size the waitqueue hash table.
1719 * Essentially these want to choose hash table sizes sufficiently
1720 * large so that collisions trying to wait on pages are rare.
1721 * But in fact, the number of active page waitqueues on typical
1722 * systems is ridiculously low, less than 200. So this is even
1723 * conservative, even though it seems large.
1725 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
1726 * waitqueues, i.e. the size of the waitq table given the number of pages.
1728 #define PAGES_PER_WAITQUEUE 256
1730 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
1732 unsigned long size
= 1;
1734 pages
/= PAGES_PER_WAITQUEUE
;
1736 while (size
< pages
)
1740 * Once we have dozens or even hundreds of threads sleeping
1741 * on IO we've got bigger problems than wait queue collision.
1742 * Limit the size of the wait table to a reasonable size.
1744 size
= min(size
, 4096UL);
1746 return max(size
, 4UL);
1750 * This is an integer logarithm so that shifts can be used later
1751 * to extract the more random high bits from the multiplicative
1752 * hash function before the remainder is taken.
1754 static inline unsigned long wait_table_bits(unsigned long size
)
1759 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
1761 static void __init
calculate_zone_totalpages(struct pglist_data
*pgdat
,
1762 unsigned long *zones_size
, unsigned long *zholes_size
)
1764 unsigned long realtotalpages
, totalpages
= 0;
1767 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
1768 totalpages
+= zones_size
[i
];
1769 pgdat
->node_spanned_pages
= totalpages
;
1771 realtotalpages
= totalpages
;
1773 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
1774 realtotalpages
-= zholes_size
[i
];
1775 pgdat
->node_present_pages
= realtotalpages
;
1776 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
, realtotalpages
);
1781 * Initially all pages are reserved - free ones are freed
1782 * up by free_all_bootmem() once the early boot process is
1783 * done. Non-atomic initialization, single-pass.
1785 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
1786 unsigned long start_pfn
)
1789 unsigned long end_pfn
= start_pfn
+ size
;
1792 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
1793 if (!early_pfn_valid(pfn
))
1795 page
= pfn_to_page(pfn
);
1796 set_page_links(page
, zone
, nid
, pfn
);
1797 init_page_count(page
);
1798 reset_page_mapcount(page
);
1799 SetPageReserved(page
);
1800 INIT_LIST_HEAD(&page
->lru
);
1801 #ifdef WANT_PAGE_VIRTUAL
1802 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1803 if (!is_highmem_idx(zone
))
1804 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
1809 void zone_init_free_lists(struct pglist_data
*pgdat
, struct zone
*zone
,
1813 for (order
= 0; order
< MAX_ORDER
; order
++) {
1814 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
);
1815 zone
->free_area
[order
].nr_free
= 0;
1819 #define ZONETABLE_INDEX(x, zone_nr) ((x << ZONES_SHIFT) | zone_nr)
1820 void zonetable_add(struct zone
*zone
, int nid
, int zid
, unsigned long pfn
,
1823 unsigned long snum
= pfn_to_section_nr(pfn
);
1824 unsigned long end
= pfn_to_section_nr(pfn
+ size
);
1827 zone_table
[ZONETABLE_INDEX(nid
, zid
)] = zone
;
1829 for (; snum
<= end
; snum
++)
1830 zone_table
[ZONETABLE_INDEX(snum
, zid
)] = zone
;
1833 #ifndef __HAVE_ARCH_MEMMAP_INIT
1834 #define memmap_init(size, nid, zone, start_pfn) \
1835 memmap_init_zone((size), (nid), (zone), (start_pfn))
1838 static int __cpuinit
zone_batchsize(struct zone
*zone
)
1843 * The per-cpu-pages pools are set to around 1000th of the
1844 * size of the zone. But no more than 1/2 of a meg.
1846 * OK, so we don't know how big the cache is. So guess.
1848 batch
= zone
->present_pages
/ 1024;
1849 if (batch
* PAGE_SIZE
> 512 * 1024)
1850 batch
= (512 * 1024) / PAGE_SIZE
;
1851 batch
/= 4; /* We effectively *= 4 below */
1856 * Clamp the batch to a 2^n - 1 value. Having a power
1857 * of 2 value was found to be more likely to have
1858 * suboptimal cache aliasing properties in some cases.
1860 * For example if 2 tasks are alternately allocating
1861 * batches of pages, one task can end up with a lot
1862 * of pages of one half of the possible page colors
1863 * and the other with pages of the other colors.
1865 batch
= (1 << (fls(batch
+ batch
/2)-1)) - 1;
1870 inline void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
1872 struct per_cpu_pages
*pcp
;
1874 memset(p
, 0, sizeof(*p
));
1876 pcp
= &p
->pcp
[0]; /* hot */
1878 pcp
->high
= 6 * batch
;
1879 pcp
->batch
= max(1UL, 1 * batch
);
1880 INIT_LIST_HEAD(&pcp
->list
);
1882 pcp
= &p
->pcp
[1]; /* cold*/
1884 pcp
->high
= 2 * batch
;
1885 pcp
->batch
= max(1UL, batch
/2);
1886 INIT_LIST_HEAD(&pcp
->list
);
1890 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
1891 * to the value high for the pageset p.
1894 static void setup_pagelist_highmark(struct per_cpu_pageset
*p
,
1897 struct per_cpu_pages
*pcp
;
1899 pcp
= &p
->pcp
[0]; /* hot list */
1901 pcp
->batch
= max(1UL, high
/4);
1902 if ((high
/4) > (PAGE_SHIFT
* 8))
1903 pcp
->batch
= PAGE_SHIFT
* 8;
1909 * Boot pageset table. One per cpu which is going to be used for all
1910 * zones and all nodes. The parameters will be set in such a way
1911 * that an item put on a list will immediately be handed over to
1912 * the buddy list. This is safe since pageset manipulation is done
1913 * with interrupts disabled.
1915 * Some NUMA counter updates may also be caught by the boot pagesets.
1917 * The boot_pagesets must be kept even after bootup is complete for
1918 * unused processors and/or zones. They do play a role for bootstrapping
1919 * hotplugged processors.
1921 * zoneinfo_show() and maybe other functions do
1922 * not check if the processor is online before following the pageset pointer.
1923 * Other parts of the kernel may not check if the zone is available.
1925 static struct per_cpu_pageset boot_pageset
[NR_CPUS
];
1928 * Dynamically allocate memory for the
1929 * per cpu pageset array in struct zone.
1931 static int __cpuinit
process_zones(int cpu
)
1933 struct zone
*zone
, *dzone
;
1935 for_each_zone(zone
) {
1937 zone_pcp(zone
, cpu
) = kmalloc_node(sizeof(struct per_cpu_pageset
),
1938 GFP_KERNEL
, cpu_to_node(cpu
));
1939 if (!zone_pcp(zone
, cpu
))
1942 setup_pageset(zone_pcp(zone
, cpu
), zone_batchsize(zone
));
1944 if (percpu_pagelist_fraction
)
1945 setup_pagelist_highmark(zone_pcp(zone
, cpu
),
1946 (zone
->present_pages
/ percpu_pagelist_fraction
));
1951 for_each_zone(dzone
) {
1954 kfree(zone_pcp(dzone
, cpu
));
1955 zone_pcp(dzone
, cpu
) = NULL
;
1960 static inline void free_zone_pagesets(int cpu
)
1964 for_each_zone(zone
) {
1965 struct per_cpu_pageset
*pset
= zone_pcp(zone
, cpu
);
1967 zone_pcp(zone
, cpu
) = NULL
;
1972 static int pageset_cpuup_callback(struct notifier_block
*nfb
,
1973 unsigned long action
,
1976 int cpu
= (long)hcpu
;
1977 int ret
= NOTIFY_OK
;
1980 case CPU_UP_PREPARE
:
1981 if (process_zones(cpu
))
1984 case CPU_UP_CANCELED
:
1986 free_zone_pagesets(cpu
);
1994 static struct notifier_block pageset_notifier
=
1995 { &pageset_cpuup_callback
, NULL
, 0 };
1997 void __init
setup_per_cpu_pageset(void)
2001 /* Initialize per_cpu_pageset for cpu 0.
2002 * A cpuup callback will do this for every cpu
2003 * as it comes online
2005 err
= process_zones(smp_processor_id());
2007 register_cpu_notifier(&pageset_notifier
);
2013 void zone_wait_table_init(struct zone
*zone
, unsigned long zone_size_pages
)
2016 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
2019 * The per-page waitqueue mechanism uses hashed waitqueues
2022 zone
->wait_table_hash_nr_entries
=
2023 wait_table_hash_nr_entries(zone_size_pages
);
2024 zone
->wait_table_bits
=
2025 wait_table_bits(zone
->wait_table_hash_nr_entries
);
2026 zone
->wait_table
= (wait_queue_head_t
*)
2027 alloc_bootmem_node(pgdat
, zone
->wait_table_hash_nr_entries
2028 * sizeof(wait_queue_head_t
));
2030 for(i
= 0; i
< zone
->wait_table_hash_nr_entries
; ++i
)
2031 init_waitqueue_head(zone
->wait_table
+ i
);
2034 static __meminit
void zone_pcp_init(struct zone
*zone
)
2037 unsigned long batch
= zone_batchsize(zone
);
2039 for (cpu
= 0; cpu
< NR_CPUS
; cpu
++) {
2041 /* Early boot. Slab allocator not functional yet */
2042 zone_pcp(zone
, cpu
) = &boot_pageset
[cpu
];
2043 setup_pageset(&boot_pageset
[cpu
],0);
2045 setup_pageset(zone_pcp(zone
,cpu
), batch
);
2048 if (zone
->present_pages
)
2049 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%lu\n",
2050 zone
->name
, zone
->present_pages
, batch
);
2053 __meminit
int init_currently_empty_zone(struct zone
*zone
,
2054 unsigned long zone_start_pfn
,
2057 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
2059 zone_wait_table_init(zone
, size
);
2060 pgdat
->nr_zones
= zone_idx(zone
) + 1;
2062 zone
->zone_start_pfn
= zone_start_pfn
;
2064 memmap_init(size
, pgdat
->node_id
, zone_idx(zone
), zone_start_pfn
);
2066 zone_init_free_lists(pgdat
, zone
, zone
->spanned_pages
);
2072 * Set up the zone data structures:
2073 * - mark all pages reserved
2074 * - mark all memory queues empty
2075 * - clear the memory bitmaps
2077 static void __meminit
free_area_init_core(struct pglist_data
*pgdat
,
2078 unsigned long *zones_size
, unsigned long *zholes_size
)
2081 int nid
= pgdat
->node_id
;
2082 unsigned long zone_start_pfn
= pgdat
->node_start_pfn
;
2085 pgdat_resize_init(pgdat
);
2086 pgdat
->nr_zones
= 0;
2087 init_waitqueue_head(&pgdat
->kswapd_wait
);
2088 pgdat
->kswapd_max_order
= 0;
2090 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
2091 struct zone
*zone
= pgdat
->node_zones
+ j
;
2092 unsigned long size
, realsize
;
2094 realsize
= size
= zones_size
[j
];
2096 realsize
-= zholes_size
[j
];
2098 if (j
< ZONE_HIGHMEM
)
2099 nr_kernel_pages
+= realsize
;
2100 nr_all_pages
+= realsize
;
2102 zone
->spanned_pages
= size
;
2103 zone
->present_pages
= realsize
;
2104 zone
->name
= zone_names
[j
];
2105 spin_lock_init(&zone
->lock
);
2106 spin_lock_init(&zone
->lru_lock
);
2107 zone_seqlock_init(zone
);
2108 zone
->zone_pgdat
= pgdat
;
2109 zone
->free_pages
= 0;
2111 zone
->temp_priority
= zone
->prev_priority
= DEF_PRIORITY
;
2113 zone_pcp_init(zone
);
2114 INIT_LIST_HEAD(&zone
->active_list
);
2115 INIT_LIST_HEAD(&zone
->inactive_list
);
2116 zone
->nr_scan_active
= 0;
2117 zone
->nr_scan_inactive
= 0;
2118 zone
->nr_active
= 0;
2119 zone
->nr_inactive
= 0;
2120 atomic_set(&zone
->reclaim_in_progress
, 0);
2124 zonetable_add(zone
, nid
, j
, zone_start_pfn
, size
);
2125 ret
= init_currently_empty_zone(zone
, zone_start_pfn
, size
);
2127 zone_start_pfn
+= size
;
2131 static void __init
alloc_node_mem_map(struct pglist_data
*pgdat
)
2133 /* Skip empty nodes */
2134 if (!pgdat
->node_spanned_pages
)
2137 #ifdef CONFIG_FLAT_NODE_MEM_MAP
2138 /* ia64 gets its own node_mem_map, before this, without bootmem */
2139 if (!pgdat
->node_mem_map
) {
2140 unsigned long size
, start
, end
;
2144 * The zone's endpoints aren't required to be MAX_ORDER
2145 * aligned but the node_mem_map endpoints must be in order
2146 * for the buddy allocator to function correctly.
2148 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
2149 end
= pgdat
->node_start_pfn
+ pgdat
->node_spanned_pages
;
2150 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
2151 size
= (end
- start
) * sizeof(struct page
);
2152 map
= alloc_remap(pgdat
->node_id
, size
);
2154 map
= alloc_bootmem_node(pgdat
, size
);
2155 pgdat
->node_mem_map
= map
+ (pgdat
->node_start_pfn
- start
);
2157 #ifdef CONFIG_FLATMEM
2159 * With no DISCONTIG, the global mem_map is just set as node 0's
2161 if (pgdat
== NODE_DATA(0))
2162 mem_map
= NODE_DATA(0)->node_mem_map
;
2164 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
2167 void __meminit
free_area_init_node(int nid
, struct pglist_data
*pgdat
,
2168 unsigned long *zones_size
, unsigned long node_start_pfn
,
2169 unsigned long *zholes_size
)
2171 pgdat
->node_id
= nid
;
2172 pgdat
->node_start_pfn
= node_start_pfn
;
2173 calculate_zone_totalpages(pgdat
, zones_size
, zholes_size
);
2175 alloc_node_mem_map(pgdat
);
2177 free_area_init_core(pgdat
, zones_size
, zholes_size
);
2180 #ifndef CONFIG_NEED_MULTIPLE_NODES
2181 static bootmem_data_t contig_bootmem_data
;
2182 struct pglist_data contig_page_data
= { .bdata
= &contig_bootmem_data
};
2184 EXPORT_SYMBOL(contig_page_data
);
2187 void __init
free_area_init(unsigned long *zones_size
)
2189 free_area_init_node(0, NODE_DATA(0), zones_size
,
2190 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
2193 #ifdef CONFIG_PROC_FS
2195 #include <linux/seq_file.h>
2197 static void *frag_start(struct seq_file
*m
, loff_t
*pos
)
2201 for (pgdat
= first_online_pgdat();
2203 pgdat
= next_online_pgdat(pgdat
))
2209 static void *frag_next(struct seq_file
*m
, void *arg
, loff_t
*pos
)
2211 pg_data_t
*pgdat
= (pg_data_t
*)arg
;
2214 return next_online_pgdat(pgdat
);
2217 static void frag_stop(struct seq_file
*m
, void *arg
)
2222 * This walks the free areas for each zone.
2224 static int frag_show(struct seq_file
*m
, void *arg
)
2226 pg_data_t
*pgdat
= (pg_data_t
*)arg
;
2228 struct zone
*node_zones
= pgdat
->node_zones
;
2229 unsigned long flags
;
2232 for (zone
= node_zones
; zone
- node_zones
< MAX_NR_ZONES
; ++zone
) {
2233 if (!populated_zone(zone
))
2236 spin_lock_irqsave(&zone
->lock
, flags
);
2237 seq_printf(m
, "Node %d, zone %8s ", pgdat
->node_id
, zone
->name
);
2238 for (order
= 0; order
< MAX_ORDER
; ++order
)
2239 seq_printf(m
, "%6lu ", zone
->free_area
[order
].nr_free
);
2240 spin_unlock_irqrestore(&zone
->lock
, flags
);
2246 struct seq_operations fragmentation_op
= {
2247 .start
= frag_start
,
2254 * Output information about zones in @pgdat.
2256 static int zoneinfo_show(struct seq_file
*m
, void *arg
)
2258 pg_data_t
*pgdat
= arg
;
2260 struct zone
*node_zones
= pgdat
->node_zones
;
2261 unsigned long flags
;
2263 for (zone
= node_zones
; zone
- node_zones
< MAX_NR_ZONES
; zone
++) {
2266 if (!populated_zone(zone
))
2269 spin_lock_irqsave(&zone
->lock
, flags
);
2270 seq_printf(m
, "Node %d, zone %8s", pgdat
->node_id
, zone
->name
);
2278 "\n scanned %lu (a: %lu i: %lu)"
2287 zone
->pages_scanned
,
2288 zone
->nr_scan_active
, zone
->nr_scan_inactive
,
2289 zone
->spanned_pages
,
2290 zone
->present_pages
);
2292 "\n protection: (%lu",
2293 zone
->lowmem_reserve
[0]);
2294 for (i
= 1; i
< ARRAY_SIZE(zone
->lowmem_reserve
); i
++)
2295 seq_printf(m
, ", %lu", zone
->lowmem_reserve
[i
]);
2299 for_each_online_cpu(i
) {
2300 struct per_cpu_pageset
*pageset
;
2303 pageset
= zone_pcp(zone
, i
);
2304 for (j
= 0; j
< ARRAY_SIZE(pageset
->pcp
); j
++) {
2305 if (pageset
->pcp
[j
].count
)
2308 if (j
== ARRAY_SIZE(pageset
->pcp
))
2310 for (j
= 0; j
< ARRAY_SIZE(pageset
->pcp
); j
++) {
2312 "\n cpu: %i pcp: %i"
2317 pageset
->pcp
[j
].count
,
2318 pageset
->pcp
[j
].high
,
2319 pageset
->pcp
[j
].batch
);
2325 "\n numa_foreign: %lu"
2326 "\n interleave_hit: %lu"
2327 "\n local_node: %lu"
2328 "\n other_node: %lu",
2331 pageset
->numa_foreign
,
2332 pageset
->interleave_hit
,
2333 pageset
->local_node
,
2334 pageset
->other_node
);
2338 "\n all_unreclaimable: %u"
2339 "\n prev_priority: %i"
2340 "\n temp_priority: %i"
2341 "\n start_pfn: %lu",
2342 zone
->all_unreclaimable
,
2343 zone
->prev_priority
,
2344 zone
->temp_priority
,
2345 zone
->zone_start_pfn
);
2346 spin_unlock_irqrestore(&zone
->lock
, flags
);
2352 struct seq_operations zoneinfo_op
= {
2353 .start
= frag_start
, /* iterate over all zones. The same as in
2357 .show
= zoneinfo_show
,
2360 static char *vmstat_text
[] = {
2364 "nr_page_table_pages",
2395 "pgscan_kswapd_high",
2396 "pgscan_kswapd_normal",
2397 "pgscan_kswapd_dma32",
2398 "pgscan_kswapd_dma",
2400 "pgscan_direct_high",
2401 "pgscan_direct_normal",
2402 "pgscan_direct_dma32",
2403 "pgscan_direct_dma",
2408 "kswapd_inodesteal",
2416 static void *vmstat_start(struct seq_file
*m
, loff_t
*pos
)
2418 struct page_state
*ps
;
2420 if (*pos
>= ARRAY_SIZE(vmstat_text
))
2423 ps
= kmalloc(sizeof(*ps
), GFP_KERNEL
);
2426 return ERR_PTR(-ENOMEM
);
2427 get_full_page_state(ps
);
2428 ps
->pgpgin
/= 2; /* sectors -> kbytes */
2430 return (unsigned long *)ps
+ *pos
;
2433 static void *vmstat_next(struct seq_file
*m
, void *arg
, loff_t
*pos
)
2436 if (*pos
>= ARRAY_SIZE(vmstat_text
))
2438 return (unsigned long *)m
->private + *pos
;
2441 static int vmstat_show(struct seq_file
*m
, void *arg
)
2443 unsigned long *l
= arg
;
2444 unsigned long off
= l
- (unsigned long *)m
->private;
2446 seq_printf(m
, "%s %lu\n", vmstat_text
[off
], *l
);
2450 static void vmstat_stop(struct seq_file
*m
, void *arg
)
2456 struct seq_operations vmstat_op
= {
2457 .start
= vmstat_start
,
2458 .next
= vmstat_next
,
2459 .stop
= vmstat_stop
,
2460 .show
= vmstat_show
,
2463 #endif /* CONFIG_PROC_FS */
2465 #ifdef CONFIG_HOTPLUG_CPU
2466 static int page_alloc_cpu_notify(struct notifier_block
*self
,
2467 unsigned long action
, void *hcpu
)
2469 int cpu
= (unsigned long)hcpu
;
2471 unsigned long *src
, *dest
;
2473 if (action
== CPU_DEAD
) {
2476 /* Drain local pagecache count. */
2477 count
= &per_cpu(nr_pagecache_local
, cpu
);
2478 atomic_add(*count
, &nr_pagecache
);
2480 local_irq_disable();
2483 /* Add dead cpu's page_states to our own. */
2484 dest
= (unsigned long *)&__get_cpu_var(page_states
);
2485 src
= (unsigned long *)&per_cpu(page_states
, cpu
);
2487 for (i
= 0; i
< sizeof(struct page_state
)/sizeof(unsigned long);
2497 #endif /* CONFIG_HOTPLUG_CPU */
2499 void __init
page_alloc_init(void)
2501 hotcpu_notifier(page_alloc_cpu_notify
, 0);
2505 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
2506 * or min_free_kbytes changes.
2508 static void calculate_totalreserve_pages(void)
2510 struct pglist_data
*pgdat
;
2511 unsigned long reserve_pages
= 0;
2514 for_each_online_pgdat(pgdat
) {
2515 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
2516 struct zone
*zone
= pgdat
->node_zones
+ i
;
2517 unsigned long max
= 0;
2519 /* Find valid and maximum lowmem_reserve in the zone */
2520 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
2521 if (zone
->lowmem_reserve
[j
] > max
)
2522 max
= zone
->lowmem_reserve
[j
];
2525 /* we treat pages_high as reserved pages. */
2526 max
+= zone
->pages_high
;
2528 if (max
> zone
->present_pages
)
2529 max
= zone
->present_pages
;
2530 reserve_pages
+= max
;
2533 totalreserve_pages
= reserve_pages
;
2537 * setup_per_zone_lowmem_reserve - called whenever
2538 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
2539 * has a correct pages reserved value, so an adequate number of
2540 * pages are left in the zone after a successful __alloc_pages().
2542 static void setup_per_zone_lowmem_reserve(void)
2544 struct pglist_data
*pgdat
;
2547 for_each_online_pgdat(pgdat
) {
2548 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
2549 struct zone
*zone
= pgdat
->node_zones
+ j
;
2550 unsigned long present_pages
= zone
->present_pages
;
2552 zone
->lowmem_reserve
[j
] = 0;
2554 for (idx
= j
-1; idx
>= 0; idx
--) {
2555 struct zone
*lower_zone
;
2557 if (sysctl_lowmem_reserve_ratio
[idx
] < 1)
2558 sysctl_lowmem_reserve_ratio
[idx
] = 1;
2560 lower_zone
= pgdat
->node_zones
+ idx
;
2561 lower_zone
->lowmem_reserve
[j
] = present_pages
/
2562 sysctl_lowmem_reserve_ratio
[idx
];
2563 present_pages
+= lower_zone
->present_pages
;
2568 /* update totalreserve_pages */
2569 calculate_totalreserve_pages();
2573 * setup_per_zone_pages_min - called when min_free_kbytes changes. Ensures
2574 * that the pages_{min,low,high} values for each zone are set correctly
2575 * with respect to min_free_kbytes.
2577 void setup_per_zone_pages_min(void)
2579 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
2580 unsigned long lowmem_pages
= 0;
2582 unsigned long flags
;
2584 /* Calculate total number of !ZONE_HIGHMEM pages */
2585 for_each_zone(zone
) {
2586 if (!is_highmem(zone
))
2587 lowmem_pages
+= zone
->present_pages
;
2590 for_each_zone(zone
) {
2593 spin_lock_irqsave(&zone
->lru_lock
, flags
);
2594 tmp
= (u64
)pages_min
* zone
->present_pages
;
2595 do_div(tmp
, lowmem_pages
);
2596 if (is_highmem(zone
)) {
2598 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
2599 * need highmem pages, so cap pages_min to a small
2602 * The (pages_high-pages_low) and (pages_low-pages_min)
2603 * deltas controls asynch page reclaim, and so should
2604 * not be capped for highmem.
2608 min_pages
= zone
->present_pages
/ 1024;
2609 if (min_pages
< SWAP_CLUSTER_MAX
)
2610 min_pages
= SWAP_CLUSTER_MAX
;
2611 if (min_pages
> 128)
2613 zone
->pages_min
= min_pages
;
2616 * If it's a lowmem zone, reserve a number of pages
2617 * proportionate to the zone's size.
2619 zone
->pages_min
= tmp
;
2622 zone
->pages_low
= zone
->pages_min
+ (tmp
>> 2);
2623 zone
->pages_high
= zone
->pages_min
+ (tmp
>> 1);
2624 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
2627 /* update totalreserve_pages */
2628 calculate_totalreserve_pages();
2632 * Initialise min_free_kbytes.
2634 * For small machines we want it small (128k min). For large machines
2635 * we want it large (64MB max). But it is not linear, because network
2636 * bandwidth does not increase linearly with machine size. We use
2638 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
2639 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
2655 static int __init
init_per_zone_pages_min(void)
2657 unsigned long lowmem_kbytes
;
2659 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
2661 min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
2662 if (min_free_kbytes
< 128)
2663 min_free_kbytes
= 128;
2664 if (min_free_kbytes
> 65536)
2665 min_free_kbytes
= 65536;
2666 setup_per_zone_pages_min();
2667 setup_per_zone_lowmem_reserve();
2670 module_init(init_per_zone_pages_min
)
2673 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
2674 * that we can call two helper functions whenever min_free_kbytes
2677 int min_free_kbytes_sysctl_handler(ctl_table
*table
, int write
,
2678 struct file
*file
, void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2680 proc_dointvec(table
, write
, file
, buffer
, length
, ppos
);
2681 setup_per_zone_pages_min();
2686 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
2687 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
2688 * whenever sysctl_lowmem_reserve_ratio changes.
2690 * The reserve ratio obviously has absolutely no relation with the
2691 * pages_min watermarks. The lowmem reserve ratio can only make sense
2692 * if in function of the boot time zone sizes.
2694 int lowmem_reserve_ratio_sysctl_handler(ctl_table
*table
, int write
,
2695 struct file
*file
, void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2697 proc_dointvec_minmax(table
, write
, file
, buffer
, length
, ppos
);
2698 setup_per_zone_lowmem_reserve();
2703 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
2704 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
2705 * can have before it gets flushed back to buddy allocator.
2708 int percpu_pagelist_fraction_sysctl_handler(ctl_table
*table
, int write
,
2709 struct file
*file
, void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2715 ret
= proc_dointvec_minmax(table
, write
, file
, buffer
, length
, ppos
);
2716 if (!write
|| (ret
== -EINVAL
))
2718 for_each_zone(zone
) {
2719 for_each_online_cpu(cpu
) {
2721 high
= zone
->present_pages
/ percpu_pagelist_fraction
;
2722 setup_pagelist_highmark(zone_pcp(zone
, cpu
), high
);
2728 __initdata
int hashdist
= HASHDIST_DEFAULT
;
2731 static int __init
set_hashdist(char *str
)
2735 hashdist
= simple_strtoul(str
, &str
, 0);
2738 __setup("hashdist=", set_hashdist
);
2742 * allocate a large system hash table from bootmem
2743 * - it is assumed that the hash table must contain an exact power-of-2
2744 * quantity of entries
2745 * - limit is the number of hash buckets, not the total allocation size
2747 void *__init
alloc_large_system_hash(const char *tablename
,
2748 unsigned long bucketsize
,
2749 unsigned long numentries
,
2752 unsigned int *_hash_shift
,
2753 unsigned int *_hash_mask
,
2754 unsigned long limit
)
2756 unsigned long long max
= limit
;
2757 unsigned long log2qty
, size
;
2760 /* allow the kernel cmdline to have a say */
2762 /* round applicable memory size up to nearest megabyte */
2763 numentries
= (flags
& HASH_HIGHMEM
) ? nr_all_pages
: nr_kernel_pages
;
2764 numentries
+= (1UL << (20 - PAGE_SHIFT
)) - 1;
2765 numentries
>>= 20 - PAGE_SHIFT
;
2766 numentries
<<= 20 - PAGE_SHIFT
;
2768 /* limit to 1 bucket per 2^scale bytes of low memory */
2769 if (scale
> PAGE_SHIFT
)
2770 numentries
>>= (scale
- PAGE_SHIFT
);
2772 numentries
<<= (PAGE_SHIFT
- scale
);
2774 numentries
= roundup_pow_of_two(numentries
);
2776 /* limit allocation size to 1/16 total memory by default */
2778 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
2779 do_div(max
, bucketsize
);
2782 if (numentries
> max
)
2785 log2qty
= long_log2(numentries
);
2788 size
= bucketsize
<< log2qty
;
2789 if (flags
& HASH_EARLY
)
2790 table
= alloc_bootmem(size
);
2792 table
= __vmalloc(size
, GFP_ATOMIC
, PAGE_KERNEL
);
2794 unsigned long order
;
2795 for (order
= 0; ((1UL << order
) << PAGE_SHIFT
) < size
; order
++)
2797 table
= (void*) __get_free_pages(GFP_ATOMIC
, order
);
2799 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
2802 panic("Failed to allocate %s hash table\n", tablename
);
2804 printk("%s hash table entries: %d (order: %d, %lu bytes)\n",
2807 long_log2(size
) - PAGE_SHIFT
,
2811 *_hash_shift
= log2qty
;
2813 *_hash_mask
= (1 << log2qty
) - 1;
2818 #ifdef CONFIG_OUT_OF_LINE_PFN_TO_PAGE
2820 * pfn <-> page translation. out-of-line version.
2821 * (see asm-generic/memory_model.h)
2823 #if defined(CONFIG_FLATMEM)
2824 struct page
*pfn_to_page(unsigned long pfn
)
2826 return mem_map
+ (pfn
- ARCH_PFN_OFFSET
);
2828 unsigned long page_to_pfn(struct page
*page
)
2830 return (page
- mem_map
) + ARCH_PFN_OFFSET
;
2832 #elif defined(CONFIG_DISCONTIGMEM)
2833 struct page
*pfn_to_page(unsigned long pfn
)
2835 int nid
= arch_pfn_to_nid(pfn
);
2836 return NODE_DATA(nid
)->node_mem_map
+ arch_local_page_offset(pfn
,nid
);
2838 unsigned long page_to_pfn(struct page
*page
)
2840 struct pglist_data
*pgdat
= NODE_DATA(page_to_nid(page
));
2841 return (page
- pgdat
->node_mem_map
) + pgdat
->node_start_pfn
;
2843 #elif defined(CONFIG_SPARSEMEM)
2844 struct page
*pfn_to_page(unsigned long pfn
)
2846 return __section_mem_map_addr(__pfn_to_section(pfn
)) + pfn
;
2849 unsigned long page_to_pfn(struct page
*page
)
2851 long section_id
= page_to_section(page
);
2852 return page
- __section_mem_map_addr(__nr_to_section(section_id
));
2854 #endif /* CONFIG_FLATMEM/DISCONTIGMME/SPARSEMEM */
2855 EXPORT_SYMBOL(pfn_to_page
);
2856 EXPORT_SYMBOL(page_to_pfn
);
2857 #endif /* CONFIG_OUT_OF_LINE_PFN_TO_PAGE */