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/nodemask.h>
37 #include <linux/vmalloc.h>
39 #include <asm/tlbflush.h>
43 * MCD - HACK: Find somewhere to initialize this EARLY, or make this
46 nodemask_t node_online_map __read_mostly
= { { [0] = 1UL } };
47 EXPORT_SYMBOL(node_online_map
);
48 nodemask_t node_possible_map __read_mostly
= NODE_MASK_ALL
;
49 EXPORT_SYMBOL(node_possible_map
);
50 struct pglist_data
*pgdat_list __read_mostly
;
51 unsigned long totalram_pages __read_mostly
;
52 unsigned long totalhigh_pages __read_mostly
;
56 * results with 256, 32 in the lowmem_reserve sysctl:
57 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
58 * 1G machine -> (16M dma, 784M normal, 224M high)
59 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
60 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
61 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
63 int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
-1] = { 256, 32 };
65 EXPORT_SYMBOL(totalram_pages
);
66 EXPORT_SYMBOL(nr_swap_pages
);
69 * Used by page_zone() to look up the address of the struct zone whose
70 * id is encoded in the upper bits of page->flags
72 struct zone
*zone_table
[1 << ZONETABLE_SHIFT
] __read_mostly
;
73 EXPORT_SYMBOL(zone_table
);
75 static char *zone_names
[MAX_NR_ZONES
] = { "DMA", "Normal", "HighMem" };
76 int min_free_kbytes
= 1024;
78 unsigned long __initdata nr_kernel_pages
;
79 unsigned long __initdata nr_all_pages
;
82 * Temporary debugging check for pages not lying within a given zone.
84 static int bad_range(struct zone
*zone
, struct page
*page
)
86 if (page_to_pfn(page
) >= zone
->zone_start_pfn
+ zone
->spanned_pages
)
88 if (page_to_pfn(page
) < zone
->zone_start_pfn
)
90 #ifdef CONFIG_HOLES_IN_ZONE
91 if (!pfn_valid(page_to_pfn(page
)))
94 if (zone
!= page_zone(page
))
99 static void bad_page(const char *function
, struct page
*page
)
101 printk(KERN_EMERG
"Bad page state at %s (in process '%s', page %p)\n",
102 function
, current
->comm
, page
);
103 printk(KERN_EMERG
"flags:0x%0*lx mapping:%p mapcount:%d count:%d\n",
104 (int)(2*sizeof(page_flags_t
)), (unsigned long)page
->flags
,
105 page
->mapping
, page_mapcount(page
), page_count(page
));
106 printk(KERN_EMERG
"Backtrace:\n");
108 printk(KERN_EMERG
"Trying to fix it up, but a reboot is needed\n");
109 page
->flags
&= ~(1 << PG_lru
|
119 set_page_count(page
, 0);
120 reset_page_mapcount(page
);
121 page
->mapping
= NULL
;
122 add_taint(TAINT_BAD_PAGE
);
125 #ifndef CONFIG_HUGETLB_PAGE
126 #define prep_compound_page(page, order) do { } while (0)
127 #define destroy_compound_page(page, order) do { } while (0)
130 * Higher-order pages are called "compound pages". They are structured thusly:
132 * The first PAGE_SIZE page is called the "head page".
134 * The remaining PAGE_SIZE pages are called "tail pages".
136 * All pages have PG_compound set. All pages have their ->private pointing at
137 * the head page (even the head page has this).
139 * The first tail page's ->mapping, if non-zero, holds the address of the
140 * compound page's put_page() function.
142 * The order of the allocation is stored in the first tail page's ->index
143 * This is only for debug at present. This usage means that zero-order pages
144 * may not be compound.
146 static void prep_compound_page(struct page
*page
, unsigned long order
)
149 int nr_pages
= 1 << order
;
151 page
[1].mapping
= NULL
;
152 page
[1].index
= order
;
153 for (i
= 0; i
< nr_pages
; i
++) {
154 struct page
*p
= page
+ i
;
157 p
->private = (unsigned long)page
;
161 static void destroy_compound_page(struct page
*page
, unsigned long order
)
164 int nr_pages
= 1 << order
;
166 if (!PageCompound(page
))
169 if (page
[1].index
!= order
)
170 bad_page(__FUNCTION__
, page
);
172 for (i
= 0; i
< nr_pages
; i
++) {
173 struct page
*p
= page
+ i
;
175 if (!PageCompound(p
))
176 bad_page(__FUNCTION__
, page
);
177 if (p
->private != (unsigned long)page
)
178 bad_page(__FUNCTION__
, page
);
179 ClearPageCompound(p
);
182 #endif /* CONFIG_HUGETLB_PAGE */
185 * function for dealing with page's order in buddy system.
186 * zone->lock is already acquired when we use these.
187 * So, we don't need atomic page->flags operations here.
189 static inline unsigned long page_order(struct page
*page
) {
190 return page
->private;
193 static inline void set_page_order(struct page
*page
, int order
) {
194 page
->private = order
;
195 __SetPagePrivate(page
);
198 static inline void rmv_page_order(struct page
*page
)
200 __ClearPagePrivate(page
);
205 * Locate the struct page for both the matching buddy in our
206 * pair (buddy1) and the combined O(n+1) page they form (page).
208 * 1) Any buddy B1 will have an order O twin B2 which satisfies
209 * the following equation:
211 * For example, if the starting buddy (buddy2) is #8 its order
213 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
215 * 2) Any buddy B will have an order O+1 parent P which
216 * satisfies the following equation:
219 * Assumption: *_mem_map is contigious at least up to MAX_ORDER
221 static inline struct page
*
222 __page_find_buddy(struct page
*page
, unsigned long page_idx
, unsigned int order
)
224 unsigned long buddy_idx
= page_idx
^ (1 << order
);
226 return page
+ (buddy_idx
- page_idx
);
229 static inline unsigned long
230 __find_combined_index(unsigned long page_idx
, unsigned int order
)
232 return (page_idx
& ~(1 << order
));
236 * This function checks whether a page is free && is the buddy
237 * we can do coalesce a page and its buddy if
238 * (a) the buddy is free &&
239 * (b) the buddy is on the buddy system &&
240 * (c) a page and its buddy have the same order.
241 * for recording page's order, we use page->private and PG_private.
244 static inline int page_is_buddy(struct page
*page
, int order
)
246 if (PagePrivate(page
) &&
247 (page_order(page
) == order
) &&
248 page_count(page
) == 0)
254 * Freeing function for a buddy system allocator.
256 * The concept of a buddy system is to maintain direct-mapped table
257 * (containing bit values) for memory blocks of various "orders".
258 * The bottom level table contains the map for the smallest allocatable
259 * units of memory (here, pages), and each level above it describes
260 * pairs of units from the levels below, hence, "buddies".
261 * At a high level, all that happens here is marking the table entry
262 * at the bottom level available, and propagating the changes upward
263 * as necessary, plus some accounting needed to play nicely with other
264 * parts of the VM system.
265 * At each level, we keep a list of pages, which are heads of continuous
266 * free pages of length of (1 << order) and marked with PG_Private.Page's
267 * order is recorded in page->private field.
268 * So when we are allocating or freeing one, we can derive the state of the
269 * other. That is, if we allocate a small block, and both were
270 * free, the remainder of the region must be split into blocks.
271 * If a block is freed, and its buddy is also free, then this
272 * triggers coalescing into a block of larger size.
277 static inline void __free_pages_bulk (struct page
*page
,
278 struct zone
*zone
, unsigned int order
)
280 unsigned long page_idx
;
281 int order_size
= 1 << order
;
284 destroy_compound_page(page
, order
);
286 page_idx
= page_to_pfn(page
) & ((1 << MAX_ORDER
) - 1);
288 BUG_ON(page_idx
& (order_size
- 1));
289 BUG_ON(bad_range(zone
, page
));
291 zone
->free_pages
+= order_size
;
292 while (order
< MAX_ORDER
-1) {
293 unsigned long combined_idx
;
294 struct free_area
*area
;
297 combined_idx
= __find_combined_index(page_idx
, order
);
298 buddy
= __page_find_buddy(page
, page_idx
, order
);
300 if (bad_range(zone
, buddy
))
302 if (!page_is_buddy(buddy
, order
))
303 break; /* Move the buddy up one level. */
304 list_del(&buddy
->lru
);
305 area
= zone
->free_area
+ order
;
307 rmv_page_order(buddy
);
308 page
= page
+ (combined_idx
- page_idx
);
309 page_idx
= combined_idx
;
312 set_page_order(page
, order
);
313 list_add(&page
->lru
, &zone
->free_area
[order
].free_list
);
314 zone
->free_area
[order
].nr_free
++;
317 static inline void free_pages_check(const char *function
, struct page
*page
)
319 if ( page_mapcount(page
) ||
320 page
->mapping
!= NULL
||
321 page_count(page
) != 0 ||
332 bad_page(function
, page
);
334 __ClearPageDirty(page
);
338 * Frees a list of pages.
339 * Assumes all pages on list are in same zone, and of same order.
340 * count is the number of pages to free.
342 * If the zone was previously in an "all pages pinned" state then look to
343 * see if this freeing clears that state.
345 * And clear the zone's pages_scanned counter, to hold off the "all pages are
346 * pinned" detection logic.
349 free_pages_bulk(struct zone
*zone
, int count
,
350 struct list_head
*list
, unsigned int order
)
353 struct page
*page
= NULL
;
356 spin_lock_irqsave(&zone
->lock
, flags
);
357 zone
->all_unreclaimable
= 0;
358 zone
->pages_scanned
= 0;
359 while (!list_empty(list
) && count
--) {
360 page
= list_entry(list
->prev
, struct page
, lru
);
361 /* have to delete it as __free_pages_bulk list manipulates */
362 list_del(&page
->lru
);
363 __free_pages_bulk(page
, zone
, order
);
366 spin_unlock_irqrestore(&zone
->lock
, flags
);
370 void __free_pages_ok(struct page
*page
, unsigned int order
)
375 arch_free_page(page
, order
);
377 mod_page_state(pgfree
, 1 << order
);
381 for (i
= 1 ; i
< (1 << order
) ; ++i
)
382 __put_page(page
+ i
);
385 for (i
= 0 ; i
< (1 << order
) ; ++i
)
386 free_pages_check(__FUNCTION__
, page
+ i
);
387 list_add(&page
->lru
, &list
);
388 kernel_map_pages(page
, 1<<order
, 0);
389 free_pages_bulk(page_zone(page
), 1, &list
, order
);
394 * The order of subdivision here is critical for the IO subsystem.
395 * Please do not alter this order without good reasons and regression
396 * testing. Specifically, as large blocks of memory are subdivided,
397 * the order in which smaller blocks are delivered depends on the order
398 * they're subdivided in this function. This is the primary factor
399 * influencing the order in which pages are delivered to the IO
400 * subsystem according to empirical testing, and this is also justified
401 * by considering the behavior of a buddy system containing a single
402 * large block of memory acted on by a series of small allocations.
403 * This behavior is a critical factor in sglist merging's success.
407 static inline struct page
*
408 expand(struct zone
*zone
, struct page
*page
,
409 int low
, int high
, struct free_area
*area
)
411 unsigned long size
= 1 << high
;
417 BUG_ON(bad_range(zone
, &page
[size
]));
418 list_add(&page
[size
].lru
, &area
->free_list
);
420 set_page_order(&page
[size
], high
);
425 void set_page_refs(struct page
*page
, int order
)
428 set_page_count(page
, 1);
433 * We need to reference all the pages for this order, otherwise if
434 * anyone accesses one of the pages with (get/put) it will be freed.
435 * - eg: access_process_vm()
437 for (i
= 0; i
< (1 << order
); i
++)
438 set_page_count(page
+ i
, 1);
439 #endif /* CONFIG_MMU */
443 * This page is about to be returned from the page allocator
445 static void prep_new_page(struct page
*page
, int order
)
447 if ( page_mapcount(page
) ||
448 page
->mapping
!= NULL
||
449 page_count(page
) != 0 ||
461 bad_page(__FUNCTION__
, page
);
463 page
->flags
&= ~(1 << PG_uptodate
| 1 << PG_error
|
464 1 << PG_referenced
| 1 << PG_arch_1
|
465 1 << PG_checked
| 1 << PG_mappedtodisk
);
467 set_page_refs(page
, order
);
468 kernel_map_pages(page
, 1 << order
, 1);
472 * Do the hard work of removing an element from the buddy allocator.
473 * Call me with the zone->lock already held.
475 static struct page
*__rmqueue(struct zone
*zone
, unsigned int order
)
477 struct free_area
* area
;
478 unsigned int current_order
;
481 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
482 area
= zone
->free_area
+ current_order
;
483 if (list_empty(&area
->free_list
))
486 page
= list_entry(area
->free_list
.next
, struct page
, lru
);
487 list_del(&page
->lru
);
488 rmv_page_order(page
);
490 zone
->free_pages
-= 1UL << order
;
491 return expand(zone
, page
, order
, current_order
, area
);
498 * Obtain a specified number of elements from the buddy allocator, all under
499 * a single hold of the lock, for efficiency. Add them to the supplied list.
500 * Returns the number of new pages which were placed at *list.
502 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
503 unsigned long count
, struct list_head
*list
)
510 spin_lock_irqsave(&zone
->lock
, flags
);
511 for (i
= 0; i
< count
; ++i
) {
512 page
= __rmqueue(zone
, order
);
516 list_add_tail(&page
->lru
, list
);
518 spin_unlock_irqrestore(&zone
->lock
, flags
);
523 /* Called from the slab reaper to drain remote pagesets */
524 void drain_remote_pages(void)
530 local_irq_save(flags
);
531 for_each_zone(zone
) {
532 struct per_cpu_pageset
*pset
;
534 /* Do not drain local pagesets */
535 if (zone
->zone_pgdat
->node_id
== numa_node_id())
538 pset
= zone
->pageset
[smp_processor_id()];
539 for (i
= 0; i
< ARRAY_SIZE(pset
->pcp
); i
++) {
540 struct per_cpu_pages
*pcp
;
544 pcp
->count
-= free_pages_bulk(zone
, pcp
->count
,
548 local_irq_restore(flags
);
552 #if defined(CONFIG_PM) || defined(CONFIG_HOTPLUG_CPU)
553 static void __drain_pages(unsigned int cpu
)
558 for_each_zone(zone
) {
559 struct per_cpu_pageset
*pset
;
561 pset
= zone_pcp(zone
, cpu
);
562 for (i
= 0; i
< ARRAY_SIZE(pset
->pcp
); i
++) {
563 struct per_cpu_pages
*pcp
;
566 pcp
->count
-= free_pages_bulk(zone
, pcp
->count
,
571 #endif /* CONFIG_PM || CONFIG_HOTPLUG_CPU */
575 void mark_free_pages(struct zone
*zone
)
577 unsigned long zone_pfn
, flags
;
579 struct list_head
*curr
;
581 if (!zone
->spanned_pages
)
584 spin_lock_irqsave(&zone
->lock
, flags
);
585 for (zone_pfn
= 0; zone_pfn
< zone
->spanned_pages
; ++zone_pfn
)
586 ClearPageNosaveFree(pfn_to_page(zone_pfn
+ zone
->zone_start_pfn
));
588 for (order
= MAX_ORDER
- 1; order
>= 0; --order
)
589 list_for_each(curr
, &zone
->free_area
[order
].free_list
) {
590 unsigned long start_pfn
, i
;
592 start_pfn
= page_to_pfn(list_entry(curr
, struct page
, lru
));
594 for (i
=0; i
< (1<<order
); i
++)
595 SetPageNosaveFree(pfn_to_page(start_pfn
+i
));
597 spin_unlock_irqrestore(&zone
->lock
, flags
);
601 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
603 void drain_local_pages(void)
607 local_irq_save(flags
);
608 __drain_pages(smp_processor_id());
609 local_irq_restore(flags
);
611 #endif /* CONFIG_PM */
613 static void zone_statistics(struct zonelist
*zonelist
, struct zone
*z
)
618 pg_data_t
*pg
= z
->zone_pgdat
;
619 pg_data_t
*orig
= zonelist
->zones
[0]->zone_pgdat
;
620 struct per_cpu_pageset
*p
;
622 local_irq_save(flags
);
623 cpu
= smp_processor_id();
629 zone_pcp(zonelist
->zones
[0], cpu
)->numa_foreign
++;
631 if (pg
== NODE_DATA(numa_node_id()))
635 local_irq_restore(flags
);
640 * Free a 0-order page
642 static void FASTCALL(free_hot_cold_page(struct page
*page
, int cold
));
643 static void fastcall
free_hot_cold_page(struct page
*page
, int cold
)
645 struct zone
*zone
= page_zone(page
);
646 struct per_cpu_pages
*pcp
;
649 arch_free_page(page
, 0);
651 kernel_map_pages(page
, 1, 0);
652 inc_page_state(pgfree
);
654 page
->mapping
= NULL
;
655 free_pages_check(__FUNCTION__
, page
);
656 pcp
= &zone_pcp(zone
, get_cpu())->pcp
[cold
];
657 local_irq_save(flags
);
658 list_add(&page
->lru
, &pcp
->list
);
660 if (pcp
->count
>= pcp
->high
)
661 pcp
->count
-= free_pages_bulk(zone
, pcp
->batch
, &pcp
->list
, 0);
662 local_irq_restore(flags
);
666 void fastcall
free_hot_page(struct page
*page
)
668 free_hot_cold_page(page
, 0);
671 void fastcall
free_cold_page(struct page
*page
)
673 free_hot_cold_page(page
, 1);
676 static inline void prep_zero_page(struct page
*page
, int order
, gfp_t gfp_flags
)
680 BUG_ON((gfp_flags
& (__GFP_WAIT
| __GFP_HIGHMEM
)) == __GFP_HIGHMEM
);
681 for(i
= 0; i
< (1 << order
); i
++)
682 clear_highpage(page
+ i
);
686 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
687 * we cheat by calling it from here, in the order > 0 path. Saves a branch
691 buffered_rmqueue(struct zone
*zone
, int order
, gfp_t gfp_flags
)
694 struct page
*page
= NULL
;
695 int cold
= !!(gfp_flags
& __GFP_COLD
);
698 struct per_cpu_pages
*pcp
;
700 pcp
= &zone_pcp(zone
, get_cpu())->pcp
[cold
];
701 local_irq_save(flags
);
702 if (pcp
->count
<= pcp
->low
)
703 pcp
->count
+= rmqueue_bulk(zone
, 0,
704 pcp
->batch
, &pcp
->list
);
706 page
= list_entry(pcp
->list
.next
, struct page
, lru
);
707 list_del(&page
->lru
);
710 local_irq_restore(flags
);
715 spin_lock_irqsave(&zone
->lock
, flags
);
716 page
= __rmqueue(zone
, order
);
717 spin_unlock_irqrestore(&zone
->lock
, flags
);
721 BUG_ON(bad_range(zone
, page
));
722 mod_page_state_zone(zone
, pgalloc
, 1 << order
);
723 prep_new_page(page
, order
);
725 if (gfp_flags
& __GFP_ZERO
)
726 prep_zero_page(page
, order
, gfp_flags
);
728 if (order
&& (gfp_flags
& __GFP_COMP
))
729 prep_compound_page(page
, order
);
735 * Return 1 if free pages are above 'mark'. This takes into account the order
738 int zone_watermark_ok(struct zone
*z
, int order
, unsigned long mark
,
739 int classzone_idx
, int can_try_harder
, gfp_t gfp_high
)
741 /* free_pages my go negative - that's OK */
742 long min
= mark
, free_pages
= z
->free_pages
- (1 << order
) + 1;
750 if (free_pages
<= min
+ z
->lowmem_reserve
[classzone_idx
])
752 for (o
= 0; o
< order
; o
++) {
753 /* At the next order, this order's pages become unavailable */
754 free_pages
-= z
->free_area
[o
].nr_free
<< o
;
756 /* Require fewer higher order pages to be free */
759 if (free_pages
<= min
)
766 should_reclaim_zone(struct zone
*z
, gfp_t gfp_mask
)
768 if (!z
->reclaim_pages
)
770 if (gfp_mask
& __GFP_NORECLAIM
)
776 * This is the 'heart' of the zoned buddy allocator.
778 struct page
* fastcall
779 __alloc_pages(gfp_t gfp_mask
, unsigned int order
,
780 struct zonelist
*zonelist
)
782 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
783 struct zone
**zones
, *z
;
785 struct reclaim_state reclaim_state
;
786 struct task_struct
*p
= current
;
791 int did_some_progress
;
793 might_sleep_if(wait
);
796 * The caller may dip into page reserves a bit more if the caller
797 * cannot run direct reclaim, or is the caller has realtime scheduling
800 can_try_harder
= (unlikely(rt_task(p
)) && !in_interrupt()) || !wait
;
802 zones
= zonelist
->zones
; /* the list of zones suitable for gfp_mask */
804 if (unlikely(zones
[0] == NULL
)) {
805 /* Should this ever happen?? */
809 classzone_idx
= zone_idx(zones
[0]);
813 * Go through the zonelist once, looking for a zone with enough free.
814 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
816 for (i
= 0; (z
= zones
[i
]) != NULL
; i
++) {
817 int do_reclaim
= should_reclaim_zone(z
, gfp_mask
);
819 if (!cpuset_zone_allowed(z
, __GFP_HARDWALL
))
823 * If the zone is to attempt early page reclaim then this loop
824 * will try to reclaim pages and check the watermark a second
825 * time before giving up and falling back to the next zone.
828 if (!zone_watermark_ok(z
, order
, z
->pages_low
,
829 classzone_idx
, 0, 0)) {
833 zone_reclaim(z
, gfp_mask
, order
);
834 /* Only try reclaim once */
836 goto zone_reclaim_retry
;
840 page
= buffered_rmqueue(z
, order
, gfp_mask
);
845 for (i
= 0; (z
= zones
[i
]) != NULL
; i
++)
846 wakeup_kswapd(z
, order
);
849 * Go through the zonelist again. Let __GFP_HIGH and allocations
850 * coming from realtime tasks to go deeper into reserves
852 * This is the last chance, in general, before the goto nopage.
853 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
854 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
856 for (i
= 0; (z
= zones
[i
]) != NULL
; i
++) {
857 if (!zone_watermark_ok(z
, order
, z
->pages_min
,
858 classzone_idx
, can_try_harder
,
859 gfp_mask
& __GFP_HIGH
))
862 if (wait
&& !cpuset_zone_allowed(z
, gfp_mask
))
865 page
= buffered_rmqueue(z
, order
, gfp_mask
);
870 /* This allocation should allow future memory freeing. */
872 if (((p
->flags
& PF_MEMALLOC
) || unlikely(test_thread_flag(TIF_MEMDIE
)))
873 && !in_interrupt()) {
874 if (!(gfp_mask
& __GFP_NOMEMALLOC
)) {
875 /* go through the zonelist yet again, ignoring mins */
876 for (i
= 0; (z
= zones
[i
]) != NULL
; i
++) {
877 if (!cpuset_zone_allowed(z
, gfp_mask
))
879 page
= buffered_rmqueue(z
, order
, gfp_mask
);
887 /* Atomic allocations - we can't balance anything */
894 /* We now go into synchronous reclaim */
895 p
->flags
|= PF_MEMALLOC
;
896 reclaim_state
.reclaimed_slab
= 0;
897 p
->reclaim_state
= &reclaim_state
;
899 did_some_progress
= try_to_free_pages(zones
, gfp_mask
);
901 p
->reclaim_state
= NULL
;
902 p
->flags
&= ~PF_MEMALLOC
;
906 if (likely(did_some_progress
)) {
907 for (i
= 0; (z
= zones
[i
]) != NULL
; i
++) {
908 if (!zone_watermark_ok(z
, order
, z
->pages_min
,
909 classzone_idx
, can_try_harder
,
910 gfp_mask
& __GFP_HIGH
))
913 if (!cpuset_zone_allowed(z
, gfp_mask
))
916 page
= buffered_rmqueue(z
, order
, gfp_mask
);
920 } else if ((gfp_mask
& __GFP_FS
) && !(gfp_mask
& __GFP_NORETRY
)) {
922 * Go through the zonelist yet one more time, keep
923 * very high watermark here, this is only to catch
924 * a parallel oom killing, we must fail if we're still
925 * under heavy pressure.
927 for (i
= 0; (z
= zones
[i
]) != NULL
; i
++) {
928 if (!zone_watermark_ok(z
, order
, z
->pages_high
,
929 classzone_idx
, 0, 0))
932 if (!cpuset_zone_allowed(z
, __GFP_HARDWALL
))
935 page
= buffered_rmqueue(z
, order
, gfp_mask
);
940 out_of_memory(gfp_mask
, order
);
945 * Don't let big-order allocations loop unless the caller explicitly
946 * requests that. Wait for some write requests to complete then retry.
948 * In this implementation, __GFP_REPEAT means __GFP_NOFAIL for order
949 * <= 3, but that may not be true in other implementations.
952 if (!(gfp_mask
& __GFP_NORETRY
)) {
953 if ((order
<= 3) || (gfp_mask
& __GFP_REPEAT
))
955 if (gfp_mask
& __GFP_NOFAIL
)
959 blk_congestion_wait(WRITE
, HZ
/50);
964 if (!(gfp_mask
& __GFP_NOWARN
) && printk_ratelimit()) {
965 printk(KERN_WARNING
"%s: page allocation failure."
966 " order:%d, mode:0x%x\n",
967 p
->comm
, order
, gfp_mask
);
973 zone_statistics(zonelist
, z
);
977 EXPORT_SYMBOL(__alloc_pages
);
980 * Common helper functions.
982 fastcall
unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
985 page
= alloc_pages(gfp_mask
, order
);
988 return (unsigned long) page_address(page
);
991 EXPORT_SYMBOL(__get_free_pages
);
993 fastcall
unsigned long get_zeroed_page(gfp_t gfp_mask
)
998 * get_zeroed_page() returns a 32-bit address, which cannot represent
1001 BUG_ON((gfp_mask
& __GFP_HIGHMEM
) != 0);
1003 page
= alloc_pages(gfp_mask
| __GFP_ZERO
, 0);
1005 return (unsigned long) page_address(page
);
1009 EXPORT_SYMBOL(get_zeroed_page
);
1011 void __pagevec_free(struct pagevec
*pvec
)
1013 int i
= pagevec_count(pvec
);
1016 free_hot_cold_page(pvec
->pages
[i
], pvec
->cold
);
1019 fastcall
void __free_pages(struct page
*page
, unsigned int order
)
1021 if (put_page_testzero(page
)) {
1023 free_hot_page(page
);
1025 __free_pages_ok(page
, order
);
1029 EXPORT_SYMBOL(__free_pages
);
1031 fastcall
void free_pages(unsigned long addr
, unsigned int order
)
1034 BUG_ON(!virt_addr_valid((void *)addr
));
1035 __free_pages(virt_to_page((void *)addr
), order
);
1039 EXPORT_SYMBOL(free_pages
);
1042 * Total amount of free (allocatable) RAM:
1044 unsigned int nr_free_pages(void)
1046 unsigned int sum
= 0;
1050 sum
+= zone
->free_pages
;
1055 EXPORT_SYMBOL(nr_free_pages
);
1058 unsigned int nr_free_pages_pgdat(pg_data_t
*pgdat
)
1060 unsigned int i
, sum
= 0;
1062 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
1063 sum
+= pgdat
->node_zones
[i
].free_pages
;
1069 static unsigned int nr_free_zone_pages(int offset
)
1071 /* Just pick one node, since fallback list is circular */
1072 pg_data_t
*pgdat
= NODE_DATA(numa_node_id());
1073 unsigned int sum
= 0;
1075 struct zonelist
*zonelist
= pgdat
->node_zonelists
+ offset
;
1076 struct zone
**zonep
= zonelist
->zones
;
1079 for (zone
= *zonep
++; zone
; zone
= *zonep
++) {
1080 unsigned long size
= zone
->present_pages
;
1081 unsigned long high
= zone
->pages_high
;
1090 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
1092 unsigned int nr_free_buffer_pages(void)
1094 return nr_free_zone_pages(gfp_zone(GFP_USER
));
1098 * Amount of free RAM allocatable within all zones
1100 unsigned int nr_free_pagecache_pages(void)
1102 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER
));
1105 #ifdef CONFIG_HIGHMEM
1106 unsigned int nr_free_highpages (void)
1109 unsigned int pages
= 0;
1111 for_each_pgdat(pgdat
)
1112 pages
+= pgdat
->node_zones
[ZONE_HIGHMEM
].free_pages
;
1119 static void show_node(struct zone
*zone
)
1121 printk("Node %d ", zone
->zone_pgdat
->node_id
);
1124 #define show_node(zone) do { } while (0)
1128 * Accumulate the page_state information across all CPUs.
1129 * The result is unavoidably approximate - it can change
1130 * during and after execution of this function.
1132 static DEFINE_PER_CPU(struct page_state
, page_states
) = {0};
1134 atomic_t nr_pagecache
= ATOMIC_INIT(0);
1135 EXPORT_SYMBOL(nr_pagecache
);
1137 DEFINE_PER_CPU(long, nr_pagecache_local
) = 0;
1140 void __get_page_state(struct page_state
*ret
, int nr
, cpumask_t
*cpumask
)
1144 memset(ret
, 0, sizeof(*ret
));
1145 cpus_and(*cpumask
, *cpumask
, cpu_online_map
);
1147 cpu
= first_cpu(*cpumask
);
1148 while (cpu
< NR_CPUS
) {
1149 unsigned long *in
, *out
, off
;
1151 in
= (unsigned long *)&per_cpu(page_states
, cpu
);
1153 cpu
= next_cpu(cpu
, *cpumask
);
1156 prefetch(&per_cpu(page_states
, cpu
));
1158 out
= (unsigned long *)ret
;
1159 for (off
= 0; off
< nr
; off
++)
1164 void get_page_state_node(struct page_state
*ret
, int node
)
1167 cpumask_t mask
= node_to_cpumask(node
);
1169 nr
= offsetof(struct page_state
, GET_PAGE_STATE_LAST
);
1170 nr
/= sizeof(unsigned long);
1172 __get_page_state(ret
, nr
+1, &mask
);
1175 void get_page_state(struct page_state
*ret
)
1178 cpumask_t mask
= CPU_MASK_ALL
;
1180 nr
= offsetof(struct page_state
, GET_PAGE_STATE_LAST
);
1181 nr
/= sizeof(unsigned long);
1183 __get_page_state(ret
, nr
+ 1, &mask
);
1186 void get_full_page_state(struct page_state
*ret
)
1188 cpumask_t mask
= CPU_MASK_ALL
;
1190 __get_page_state(ret
, sizeof(*ret
) / sizeof(unsigned long), &mask
);
1193 unsigned long __read_page_state(unsigned long offset
)
1195 unsigned long ret
= 0;
1198 for_each_online_cpu(cpu
) {
1201 in
= (unsigned long)&per_cpu(page_states
, cpu
) + offset
;
1202 ret
+= *((unsigned long *)in
);
1207 void __mod_page_state(unsigned long offset
, unsigned long delta
)
1209 unsigned long flags
;
1212 local_irq_save(flags
);
1213 ptr
= &__get_cpu_var(page_states
);
1214 *(unsigned long*)(ptr
+ offset
) += delta
;
1215 local_irq_restore(flags
);
1218 EXPORT_SYMBOL(__mod_page_state
);
1220 void __get_zone_counts(unsigned long *active
, unsigned long *inactive
,
1221 unsigned long *free
, struct pglist_data
*pgdat
)
1223 struct zone
*zones
= pgdat
->node_zones
;
1229 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
1230 *active
+= zones
[i
].nr_active
;
1231 *inactive
+= zones
[i
].nr_inactive
;
1232 *free
+= zones
[i
].free_pages
;
1236 void get_zone_counts(unsigned long *active
,
1237 unsigned long *inactive
, unsigned long *free
)
1239 struct pglist_data
*pgdat
;
1244 for_each_pgdat(pgdat
) {
1245 unsigned long l
, m
, n
;
1246 __get_zone_counts(&l
, &m
, &n
, pgdat
);
1253 void si_meminfo(struct sysinfo
*val
)
1255 val
->totalram
= totalram_pages
;
1257 val
->freeram
= nr_free_pages();
1258 val
->bufferram
= nr_blockdev_pages();
1259 #ifdef CONFIG_HIGHMEM
1260 val
->totalhigh
= totalhigh_pages
;
1261 val
->freehigh
= nr_free_highpages();
1266 val
->mem_unit
= PAGE_SIZE
;
1269 EXPORT_SYMBOL(si_meminfo
);
1272 void si_meminfo_node(struct sysinfo
*val
, int nid
)
1274 pg_data_t
*pgdat
= NODE_DATA(nid
);
1276 val
->totalram
= pgdat
->node_present_pages
;
1277 val
->freeram
= nr_free_pages_pgdat(pgdat
);
1278 val
->totalhigh
= pgdat
->node_zones
[ZONE_HIGHMEM
].present_pages
;
1279 val
->freehigh
= pgdat
->node_zones
[ZONE_HIGHMEM
].free_pages
;
1280 val
->mem_unit
= PAGE_SIZE
;
1284 #define K(x) ((x) << (PAGE_SHIFT-10))
1287 * Show free area list (used inside shift_scroll-lock stuff)
1288 * We also calculate the percentage fragmentation. We do this by counting the
1289 * memory on each free list with the exception of the first item on the list.
1291 void show_free_areas(void)
1293 struct page_state ps
;
1294 int cpu
, temperature
;
1295 unsigned long active
;
1296 unsigned long inactive
;
1300 for_each_zone(zone
) {
1302 printk("%s per-cpu:", zone
->name
);
1304 if (!zone
->present_pages
) {
1310 for (cpu
= 0; cpu
< NR_CPUS
; ++cpu
) {
1311 struct per_cpu_pageset
*pageset
;
1313 if (!cpu_possible(cpu
))
1316 pageset
= zone_pcp(zone
, cpu
);
1318 for (temperature
= 0; temperature
< 2; temperature
++)
1319 printk("cpu %d %s: low %d, high %d, batch %d used:%d\n",
1321 temperature
? "cold" : "hot",
1322 pageset
->pcp
[temperature
].low
,
1323 pageset
->pcp
[temperature
].high
,
1324 pageset
->pcp
[temperature
].batch
,
1325 pageset
->pcp
[temperature
].count
);
1329 get_page_state(&ps
);
1330 get_zone_counts(&active
, &inactive
, &free
);
1332 printk("Free pages: %11ukB (%ukB HighMem)\n",
1334 K(nr_free_highpages()));
1336 printk("Active:%lu inactive:%lu dirty:%lu writeback:%lu "
1337 "unstable:%lu free:%u slab:%lu mapped:%lu pagetables:%lu\n",
1346 ps
.nr_page_table_pages
);
1348 for_each_zone(zone
) {
1360 " pages_scanned:%lu"
1361 " all_unreclaimable? %s"
1364 K(zone
->free_pages
),
1367 K(zone
->pages_high
),
1369 K(zone
->nr_inactive
),
1370 K(zone
->present_pages
),
1371 zone
->pages_scanned
,
1372 (zone
->all_unreclaimable
? "yes" : "no")
1374 printk("lowmem_reserve[]:");
1375 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
1376 printk(" %lu", zone
->lowmem_reserve
[i
]);
1380 for_each_zone(zone
) {
1381 unsigned long nr
, flags
, order
, total
= 0;
1384 printk("%s: ", zone
->name
);
1385 if (!zone
->present_pages
) {
1390 spin_lock_irqsave(&zone
->lock
, flags
);
1391 for (order
= 0; order
< MAX_ORDER
; order
++) {
1392 nr
= zone
->free_area
[order
].nr_free
;
1393 total
+= nr
<< order
;
1394 printk("%lu*%lukB ", nr
, K(1UL) << order
);
1396 spin_unlock_irqrestore(&zone
->lock
, flags
);
1397 printk("= %lukB\n", K(total
));
1400 show_swap_cache_info();
1404 * Builds allocation fallback zone lists.
1406 static int __init
build_zonelists_node(pg_data_t
*pgdat
, struct zonelist
*zonelist
, int j
, int k
)
1413 zone
= pgdat
->node_zones
+ ZONE_HIGHMEM
;
1414 if (zone
->present_pages
) {
1415 #ifndef CONFIG_HIGHMEM
1418 zonelist
->zones
[j
++] = zone
;
1421 zone
= pgdat
->node_zones
+ ZONE_NORMAL
;
1422 if (zone
->present_pages
)
1423 zonelist
->zones
[j
++] = zone
;
1425 zone
= pgdat
->node_zones
+ ZONE_DMA
;
1426 if (zone
->present_pages
)
1427 zonelist
->zones
[j
++] = zone
;
1433 static inline int highest_zone(int zone_bits
)
1435 int res
= ZONE_NORMAL
;
1436 if (zone_bits
& (__force
int)__GFP_HIGHMEM
)
1438 if (zone_bits
& (__force
int)__GFP_DMA
)
1444 #define MAX_NODE_LOAD (num_online_nodes())
1445 static int __initdata node_load
[MAX_NUMNODES
];
1447 * find_next_best_node - find the next node that should appear in a given node's fallback list
1448 * @node: node whose fallback list we're appending
1449 * @used_node_mask: nodemask_t of already used nodes
1451 * We use a number of factors to determine which is the next node that should
1452 * appear on a given node's fallback list. The node should not have appeared
1453 * already in @node's fallback list, and it should be the next closest node
1454 * according to the distance array (which contains arbitrary distance values
1455 * from each node to each node in the system), and should also prefer nodes
1456 * with no CPUs, since presumably they'll have very little allocation pressure
1457 * on them otherwise.
1458 * It returns -1 if no node is found.
1460 static int __init
find_next_best_node(int node
, nodemask_t
*used_node_mask
)
1463 int min_val
= INT_MAX
;
1466 for_each_online_node(i
) {
1469 /* Start from local node */
1470 n
= (node
+i
) % num_online_nodes();
1472 /* Don't want a node to appear more than once */
1473 if (node_isset(n
, *used_node_mask
))
1476 /* Use the local node if we haven't already */
1477 if (!node_isset(node
, *used_node_mask
)) {
1482 /* Use the distance array to find the distance */
1483 val
= node_distance(node
, n
);
1485 /* Give preference to headless and unused nodes */
1486 tmp
= node_to_cpumask(n
);
1487 if (!cpus_empty(tmp
))
1488 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
1490 /* Slight preference for less loaded node */
1491 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
1492 val
+= node_load
[n
];
1494 if (val
< min_val
) {
1501 node_set(best_node
, *used_node_mask
);
1506 static void __init
build_zonelists(pg_data_t
*pgdat
)
1508 int i
, j
, k
, node
, local_node
;
1509 int prev_node
, load
;
1510 struct zonelist
*zonelist
;
1511 nodemask_t used_mask
;
1513 /* initialize zonelists */
1514 for (i
= 0; i
< GFP_ZONETYPES
; i
++) {
1515 zonelist
= pgdat
->node_zonelists
+ i
;
1516 zonelist
->zones
[0] = NULL
;
1519 /* NUMA-aware ordering of nodes */
1520 local_node
= pgdat
->node_id
;
1521 load
= num_online_nodes();
1522 prev_node
= local_node
;
1523 nodes_clear(used_mask
);
1524 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
1526 * We don't want to pressure a particular node.
1527 * So adding penalty to the first node in same
1528 * distance group to make it round-robin.
1530 if (node_distance(local_node
, node
) !=
1531 node_distance(local_node
, prev_node
))
1532 node_load
[node
] += load
;
1535 for (i
= 0; i
< GFP_ZONETYPES
; i
++) {
1536 zonelist
= pgdat
->node_zonelists
+ i
;
1537 for (j
= 0; zonelist
->zones
[j
] != NULL
; j
++);
1539 k
= highest_zone(i
);
1541 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
, k
);
1542 zonelist
->zones
[j
] = NULL
;
1547 #else /* CONFIG_NUMA */
1549 static void __init
build_zonelists(pg_data_t
*pgdat
)
1551 int i
, j
, k
, node
, local_node
;
1553 local_node
= pgdat
->node_id
;
1554 for (i
= 0; i
< GFP_ZONETYPES
; i
++) {
1555 struct zonelist
*zonelist
;
1557 zonelist
= pgdat
->node_zonelists
+ i
;
1560 k
= highest_zone(i
);
1561 j
= build_zonelists_node(pgdat
, zonelist
, j
, k
);
1563 * Now we build the zonelist so that it contains the zones
1564 * of all the other nodes.
1565 * We don't want to pressure a particular node, so when
1566 * building the zones for node N, we make sure that the
1567 * zones coming right after the local ones are those from
1568 * node N+1 (modulo N)
1570 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
1571 if (!node_online(node
))
1573 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
, k
);
1575 for (node
= 0; node
< local_node
; node
++) {
1576 if (!node_online(node
))
1578 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
, k
);
1581 zonelist
->zones
[j
] = NULL
;
1585 #endif /* CONFIG_NUMA */
1587 void __init
build_all_zonelists(void)
1591 for_each_online_node(i
)
1592 build_zonelists(NODE_DATA(i
));
1593 printk("Built %i zonelists\n", num_online_nodes());
1594 cpuset_init_current_mems_allowed();
1598 * Helper functions to size the waitqueue hash table.
1599 * Essentially these want to choose hash table sizes sufficiently
1600 * large so that collisions trying to wait on pages are rare.
1601 * But in fact, the number of active page waitqueues on typical
1602 * systems is ridiculously low, less than 200. So this is even
1603 * conservative, even though it seems large.
1605 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
1606 * waitqueues, i.e. the size of the waitq table given the number of pages.
1608 #define PAGES_PER_WAITQUEUE 256
1610 static inline unsigned long wait_table_size(unsigned long pages
)
1612 unsigned long size
= 1;
1614 pages
/= PAGES_PER_WAITQUEUE
;
1616 while (size
< pages
)
1620 * Once we have dozens or even hundreds of threads sleeping
1621 * on IO we've got bigger problems than wait queue collision.
1622 * Limit the size of the wait table to a reasonable size.
1624 size
= min(size
, 4096UL);
1626 return max(size
, 4UL);
1630 * This is an integer logarithm so that shifts can be used later
1631 * to extract the more random high bits from the multiplicative
1632 * hash function before the remainder is taken.
1634 static inline unsigned long wait_table_bits(unsigned long size
)
1639 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
1641 static void __init
calculate_zone_totalpages(struct pglist_data
*pgdat
,
1642 unsigned long *zones_size
, unsigned long *zholes_size
)
1644 unsigned long realtotalpages
, totalpages
= 0;
1647 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
1648 totalpages
+= zones_size
[i
];
1649 pgdat
->node_spanned_pages
= totalpages
;
1651 realtotalpages
= totalpages
;
1653 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
1654 realtotalpages
-= zholes_size
[i
];
1655 pgdat
->node_present_pages
= realtotalpages
;
1656 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
, realtotalpages
);
1661 * Initially all pages are reserved - free ones are freed
1662 * up by free_all_bootmem() once the early boot process is
1663 * done. Non-atomic initialization, single-pass.
1665 void __init
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
1666 unsigned long start_pfn
)
1669 unsigned long end_pfn
= start_pfn
+ size
;
1672 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++, page
++) {
1673 if (!early_pfn_valid(pfn
))
1675 if (!early_pfn_in_nid(pfn
, nid
))
1677 page
= pfn_to_page(pfn
);
1678 set_page_links(page
, zone
, nid
, pfn
);
1679 set_page_count(page
, 1);
1680 reset_page_mapcount(page
);
1681 SetPageReserved(page
);
1682 INIT_LIST_HEAD(&page
->lru
);
1683 #ifdef WANT_PAGE_VIRTUAL
1684 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1685 if (!is_highmem_idx(zone
))
1686 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
1691 void zone_init_free_lists(struct pglist_data
*pgdat
, struct zone
*zone
,
1695 for (order
= 0; order
< MAX_ORDER
; order
++) {
1696 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
);
1697 zone
->free_area
[order
].nr_free
= 0;
1701 #define ZONETABLE_INDEX(x, zone_nr) ((x << ZONES_SHIFT) | zone_nr)
1702 void zonetable_add(struct zone
*zone
, int nid
, int zid
, unsigned long pfn
,
1705 unsigned long snum
= pfn_to_section_nr(pfn
);
1706 unsigned long end
= pfn_to_section_nr(pfn
+ size
);
1709 zone_table
[ZONETABLE_INDEX(nid
, zid
)] = zone
;
1711 for (; snum
<= end
; snum
++)
1712 zone_table
[ZONETABLE_INDEX(snum
, zid
)] = zone
;
1715 #ifndef __HAVE_ARCH_MEMMAP_INIT
1716 #define memmap_init(size, nid, zone, start_pfn) \
1717 memmap_init_zone((size), (nid), (zone), (start_pfn))
1720 static int __devinit
zone_batchsize(struct zone
*zone
)
1725 * The per-cpu-pages pools are set to around 1000th of the
1726 * size of the zone. But no more than 1/2 of a meg.
1728 * OK, so we don't know how big the cache is. So guess.
1730 batch
= zone
->present_pages
/ 1024;
1731 if (batch
* PAGE_SIZE
> 512 * 1024)
1732 batch
= (512 * 1024) / PAGE_SIZE
;
1733 batch
/= 4; /* We effectively *= 4 below */
1738 * We will be trying to allcoate bigger chunks of contiguous
1739 * memory of the order of fls(batch). This should result in
1740 * better cache coloring.
1742 * A sanity check also to ensure that batch is still in limits.
1744 batch
= (1 << fls(batch
+ batch
/2));
1746 if (fls(batch
) >= (PAGE_SHIFT
+ MAX_ORDER
- 2))
1747 batch
= PAGE_SHIFT
+ ((MAX_ORDER
- 1 - PAGE_SHIFT
)/2);
1752 inline void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
1754 struct per_cpu_pages
*pcp
;
1756 memset(p
, 0, sizeof(*p
));
1758 pcp
= &p
->pcp
[0]; /* hot */
1761 pcp
->high
= 6 * batch
;
1762 pcp
->batch
= max(1UL, 1 * batch
);
1763 INIT_LIST_HEAD(&pcp
->list
);
1765 pcp
= &p
->pcp
[1]; /* cold*/
1768 pcp
->high
= 2 * batch
;
1769 pcp
->batch
= max(1UL, batch
/2);
1770 INIT_LIST_HEAD(&pcp
->list
);
1775 * Boot pageset table. One per cpu which is going to be used for all
1776 * zones and all nodes. The parameters will be set in such a way
1777 * that an item put on a list will immediately be handed over to
1778 * the buddy list. This is safe since pageset manipulation is done
1779 * with interrupts disabled.
1781 * Some NUMA counter updates may also be caught by the boot pagesets.
1783 * The boot_pagesets must be kept even after bootup is complete for
1784 * unused processors and/or zones. They do play a role for bootstrapping
1785 * hotplugged processors.
1787 * zoneinfo_show() and maybe other functions do
1788 * not check if the processor is online before following the pageset pointer.
1789 * Other parts of the kernel may not check if the zone is available.
1791 static struct per_cpu_pageset
1792 boot_pageset
[NR_CPUS
];
1795 * Dynamically allocate memory for the
1796 * per cpu pageset array in struct zone.
1798 static int __devinit
process_zones(int cpu
)
1800 struct zone
*zone
, *dzone
;
1802 for_each_zone(zone
) {
1804 zone
->pageset
[cpu
] = kmalloc_node(sizeof(struct per_cpu_pageset
),
1805 GFP_KERNEL
, cpu_to_node(cpu
));
1806 if (!zone
->pageset
[cpu
])
1809 setup_pageset(zone
->pageset
[cpu
], zone_batchsize(zone
));
1814 for_each_zone(dzone
) {
1817 kfree(dzone
->pageset
[cpu
]);
1818 dzone
->pageset
[cpu
] = NULL
;
1823 static inline void free_zone_pagesets(int cpu
)
1828 for_each_zone(zone
) {
1829 struct per_cpu_pageset
*pset
= zone_pcp(zone
, cpu
);
1831 zone_pcp(zone
, cpu
) = NULL
;
1837 static int __devinit
pageset_cpuup_callback(struct notifier_block
*nfb
,
1838 unsigned long action
,
1841 int cpu
= (long)hcpu
;
1842 int ret
= NOTIFY_OK
;
1845 case CPU_UP_PREPARE
:
1846 if (process_zones(cpu
))
1849 #ifdef CONFIG_HOTPLUG_CPU
1851 free_zone_pagesets(cpu
);
1860 static struct notifier_block pageset_notifier
=
1861 { &pageset_cpuup_callback
, NULL
, 0 };
1863 void __init
setup_per_cpu_pageset()
1867 /* Initialize per_cpu_pageset for cpu 0.
1868 * A cpuup callback will do this for every cpu
1869 * as it comes online
1871 err
= process_zones(smp_processor_id());
1873 register_cpu_notifier(&pageset_notifier
);
1879 * Set up the zone data structures:
1880 * - mark all pages reserved
1881 * - mark all memory queues empty
1882 * - clear the memory bitmaps
1884 static void __init
free_area_init_core(struct pglist_data
*pgdat
,
1885 unsigned long *zones_size
, unsigned long *zholes_size
)
1888 int cpu
, nid
= pgdat
->node_id
;
1889 unsigned long zone_start_pfn
= pgdat
->node_start_pfn
;
1891 pgdat
->nr_zones
= 0;
1892 init_waitqueue_head(&pgdat
->kswapd_wait
);
1893 pgdat
->kswapd_max_order
= 0;
1895 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
1896 struct zone
*zone
= pgdat
->node_zones
+ j
;
1897 unsigned long size
, realsize
;
1898 unsigned long batch
;
1900 realsize
= size
= zones_size
[j
];
1902 realsize
-= zholes_size
[j
];
1904 if (j
== ZONE_DMA
|| j
== ZONE_NORMAL
)
1905 nr_kernel_pages
+= realsize
;
1906 nr_all_pages
+= realsize
;
1908 zone
->spanned_pages
= size
;
1909 zone
->present_pages
= realsize
;
1910 zone
->name
= zone_names
[j
];
1911 spin_lock_init(&zone
->lock
);
1912 spin_lock_init(&zone
->lru_lock
);
1913 zone
->zone_pgdat
= pgdat
;
1914 zone
->free_pages
= 0;
1916 zone
->temp_priority
= zone
->prev_priority
= DEF_PRIORITY
;
1918 batch
= zone_batchsize(zone
);
1920 for (cpu
= 0; cpu
< NR_CPUS
; cpu
++) {
1922 /* Early boot. Slab allocator not functional yet */
1923 zone
->pageset
[cpu
] = &boot_pageset
[cpu
];
1924 setup_pageset(&boot_pageset
[cpu
],0);
1926 setup_pageset(zone_pcp(zone
,cpu
), batch
);
1929 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%lu\n",
1930 zone_names
[j
], realsize
, batch
);
1931 INIT_LIST_HEAD(&zone
->active_list
);
1932 INIT_LIST_HEAD(&zone
->inactive_list
);
1933 zone
->nr_scan_active
= 0;
1934 zone
->nr_scan_inactive
= 0;
1935 zone
->nr_active
= 0;
1936 zone
->nr_inactive
= 0;
1937 atomic_set(&zone
->reclaim_in_progress
, 0);
1942 * The per-page waitqueue mechanism uses hashed waitqueues
1945 zone
->wait_table_size
= wait_table_size(size
);
1946 zone
->wait_table_bits
=
1947 wait_table_bits(zone
->wait_table_size
);
1948 zone
->wait_table
= (wait_queue_head_t
*)
1949 alloc_bootmem_node(pgdat
, zone
->wait_table_size
1950 * sizeof(wait_queue_head_t
));
1952 for(i
= 0; i
< zone
->wait_table_size
; ++i
)
1953 init_waitqueue_head(zone
->wait_table
+ i
);
1955 pgdat
->nr_zones
= j
+1;
1957 zone
->zone_mem_map
= pfn_to_page(zone_start_pfn
);
1958 zone
->zone_start_pfn
= zone_start_pfn
;
1960 memmap_init(size
, nid
, j
, zone_start_pfn
);
1962 zonetable_add(zone
, nid
, j
, zone_start_pfn
, size
);
1964 zone_start_pfn
+= size
;
1966 zone_init_free_lists(pgdat
, zone
, zone
->spanned_pages
);
1970 static void __init
alloc_node_mem_map(struct pglist_data
*pgdat
)
1972 /* Skip empty nodes */
1973 if (!pgdat
->node_spanned_pages
)
1976 #ifdef CONFIG_FLAT_NODE_MEM_MAP
1977 /* ia64 gets its own node_mem_map, before this, without bootmem */
1978 if (!pgdat
->node_mem_map
) {
1982 size
= (pgdat
->node_spanned_pages
+ 1) * sizeof(struct page
);
1983 map
= alloc_remap(pgdat
->node_id
, size
);
1985 map
= alloc_bootmem_node(pgdat
, size
);
1986 pgdat
->node_mem_map
= map
;
1988 #ifdef CONFIG_FLATMEM
1990 * With no DISCONTIG, the global mem_map is just set as node 0's
1992 if (pgdat
== NODE_DATA(0))
1993 mem_map
= NODE_DATA(0)->node_mem_map
;
1995 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
1998 void __init
free_area_init_node(int nid
, struct pglist_data
*pgdat
,
1999 unsigned long *zones_size
, unsigned long node_start_pfn
,
2000 unsigned long *zholes_size
)
2002 pgdat
->node_id
= nid
;
2003 pgdat
->node_start_pfn
= node_start_pfn
;
2004 calculate_zone_totalpages(pgdat
, zones_size
, zholes_size
);
2006 alloc_node_mem_map(pgdat
);
2008 free_area_init_core(pgdat
, zones_size
, zholes_size
);
2011 #ifndef CONFIG_NEED_MULTIPLE_NODES
2012 static bootmem_data_t contig_bootmem_data
;
2013 struct pglist_data contig_page_data
= { .bdata
= &contig_bootmem_data
};
2015 EXPORT_SYMBOL(contig_page_data
);
2018 void __init
free_area_init(unsigned long *zones_size
)
2020 free_area_init_node(0, NODE_DATA(0), zones_size
,
2021 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
2024 #ifdef CONFIG_PROC_FS
2026 #include <linux/seq_file.h>
2028 static void *frag_start(struct seq_file
*m
, loff_t
*pos
)
2033 for (pgdat
= pgdat_list
; pgdat
&& node
; pgdat
= pgdat
->pgdat_next
)
2039 static void *frag_next(struct seq_file
*m
, void *arg
, loff_t
*pos
)
2041 pg_data_t
*pgdat
= (pg_data_t
*)arg
;
2044 return pgdat
->pgdat_next
;
2047 static void frag_stop(struct seq_file
*m
, void *arg
)
2052 * This walks the free areas for each zone.
2054 static int frag_show(struct seq_file
*m
, void *arg
)
2056 pg_data_t
*pgdat
= (pg_data_t
*)arg
;
2058 struct zone
*node_zones
= pgdat
->node_zones
;
2059 unsigned long flags
;
2062 for (zone
= node_zones
; zone
- node_zones
< MAX_NR_ZONES
; ++zone
) {
2063 if (!zone
->present_pages
)
2066 spin_lock_irqsave(&zone
->lock
, flags
);
2067 seq_printf(m
, "Node %d, zone %8s ", pgdat
->node_id
, zone
->name
);
2068 for (order
= 0; order
< MAX_ORDER
; ++order
)
2069 seq_printf(m
, "%6lu ", zone
->free_area
[order
].nr_free
);
2070 spin_unlock_irqrestore(&zone
->lock
, flags
);
2076 struct seq_operations fragmentation_op
= {
2077 .start
= frag_start
,
2084 * Output information about zones in @pgdat.
2086 static int zoneinfo_show(struct seq_file
*m
, void *arg
)
2088 pg_data_t
*pgdat
= arg
;
2090 struct zone
*node_zones
= pgdat
->node_zones
;
2091 unsigned long flags
;
2093 for (zone
= node_zones
; zone
- node_zones
< MAX_NR_ZONES
; zone
++) {
2096 if (!zone
->present_pages
)
2099 spin_lock_irqsave(&zone
->lock
, flags
);
2100 seq_printf(m
, "Node %d, zone %8s", pgdat
->node_id
, zone
->name
);
2108 "\n scanned %lu (a: %lu i: %lu)"
2117 zone
->pages_scanned
,
2118 zone
->nr_scan_active
, zone
->nr_scan_inactive
,
2119 zone
->spanned_pages
,
2120 zone
->present_pages
);
2122 "\n protection: (%lu",
2123 zone
->lowmem_reserve
[0]);
2124 for (i
= 1; i
< ARRAY_SIZE(zone
->lowmem_reserve
); i
++)
2125 seq_printf(m
, ", %lu", zone
->lowmem_reserve
[i
]);
2129 for (i
= 0; i
< ARRAY_SIZE(zone
->pageset
); i
++) {
2130 struct per_cpu_pageset
*pageset
;
2133 pageset
= zone_pcp(zone
, i
);
2134 for (j
= 0; j
< ARRAY_SIZE(pageset
->pcp
); j
++) {
2135 if (pageset
->pcp
[j
].count
)
2138 if (j
== ARRAY_SIZE(pageset
->pcp
))
2140 for (j
= 0; j
< ARRAY_SIZE(pageset
->pcp
); j
++) {
2142 "\n cpu: %i pcp: %i"
2148 pageset
->pcp
[j
].count
,
2149 pageset
->pcp
[j
].low
,
2150 pageset
->pcp
[j
].high
,
2151 pageset
->pcp
[j
].batch
);
2157 "\n numa_foreign: %lu"
2158 "\n interleave_hit: %lu"
2159 "\n local_node: %lu"
2160 "\n other_node: %lu",
2163 pageset
->numa_foreign
,
2164 pageset
->interleave_hit
,
2165 pageset
->local_node
,
2166 pageset
->other_node
);
2170 "\n all_unreclaimable: %u"
2171 "\n prev_priority: %i"
2172 "\n temp_priority: %i"
2173 "\n start_pfn: %lu",
2174 zone
->all_unreclaimable
,
2175 zone
->prev_priority
,
2176 zone
->temp_priority
,
2177 zone
->zone_start_pfn
);
2178 spin_unlock_irqrestore(&zone
->lock
, flags
);
2184 struct seq_operations zoneinfo_op
= {
2185 .start
= frag_start
, /* iterate over all zones. The same as in
2189 .show
= zoneinfo_show
,
2192 static char *vmstat_text
[] = {
2196 "nr_page_table_pages",
2221 "pgscan_kswapd_high",
2222 "pgscan_kswapd_normal",
2224 "pgscan_kswapd_dma",
2225 "pgscan_direct_high",
2226 "pgscan_direct_normal",
2227 "pgscan_direct_dma",
2232 "kswapd_inodesteal",
2240 static void *vmstat_start(struct seq_file
*m
, loff_t
*pos
)
2242 struct page_state
*ps
;
2244 if (*pos
>= ARRAY_SIZE(vmstat_text
))
2247 ps
= kmalloc(sizeof(*ps
), GFP_KERNEL
);
2250 return ERR_PTR(-ENOMEM
);
2251 get_full_page_state(ps
);
2252 ps
->pgpgin
/= 2; /* sectors -> kbytes */
2254 return (unsigned long *)ps
+ *pos
;
2257 static void *vmstat_next(struct seq_file
*m
, void *arg
, loff_t
*pos
)
2260 if (*pos
>= ARRAY_SIZE(vmstat_text
))
2262 return (unsigned long *)m
->private + *pos
;
2265 static int vmstat_show(struct seq_file
*m
, void *arg
)
2267 unsigned long *l
= arg
;
2268 unsigned long off
= l
- (unsigned long *)m
->private;
2270 seq_printf(m
, "%s %lu\n", vmstat_text
[off
], *l
);
2274 static void vmstat_stop(struct seq_file
*m
, void *arg
)
2280 struct seq_operations vmstat_op
= {
2281 .start
= vmstat_start
,
2282 .next
= vmstat_next
,
2283 .stop
= vmstat_stop
,
2284 .show
= vmstat_show
,
2287 #endif /* CONFIG_PROC_FS */
2289 #ifdef CONFIG_HOTPLUG_CPU
2290 static int page_alloc_cpu_notify(struct notifier_block
*self
,
2291 unsigned long action
, void *hcpu
)
2293 int cpu
= (unsigned long)hcpu
;
2295 unsigned long *src
, *dest
;
2297 if (action
== CPU_DEAD
) {
2300 /* Drain local pagecache count. */
2301 count
= &per_cpu(nr_pagecache_local
, cpu
);
2302 atomic_add(*count
, &nr_pagecache
);
2304 local_irq_disable();
2307 /* Add dead cpu's page_states to our own. */
2308 dest
= (unsigned long *)&__get_cpu_var(page_states
);
2309 src
= (unsigned long *)&per_cpu(page_states
, cpu
);
2311 for (i
= 0; i
< sizeof(struct page_state
)/sizeof(unsigned long);
2321 #endif /* CONFIG_HOTPLUG_CPU */
2323 void __init
page_alloc_init(void)
2325 hotcpu_notifier(page_alloc_cpu_notify
, 0);
2329 * setup_per_zone_lowmem_reserve - called whenever
2330 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
2331 * has a correct pages reserved value, so an adequate number of
2332 * pages are left in the zone after a successful __alloc_pages().
2334 static void setup_per_zone_lowmem_reserve(void)
2336 struct pglist_data
*pgdat
;
2339 for_each_pgdat(pgdat
) {
2340 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
2341 struct zone
*zone
= pgdat
->node_zones
+ j
;
2342 unsigned long present_pages
= zone
->present_pages
;
2344 zone
->lowmem_reserve
[j
] = 0;
2346 for (idx
= j
-1; idx
>= 0; idx
--) {
2347 struct zone
*lower_zone
;
2349 if (sysctl_lowmem_reserve_ratio
[idx
] < 1)
2350 sysctl_lowmem_reserve_ratio
[idx
] = 1;
2352 lower_zone
= pgdat
->node_zones
+ idx
;
2353 lower_zone
->lowmem_reserve
[j
] = present_pages
/
2354 sysctl_lowmem_reserve_ratio
[idx
];
2355 present_pages
+= lower_zone
->present_pages
;
2362 * setup_per_zone_pages_min - called when min_free_kbytes changes. Ensures
2363 * that the pages_{min,low,high} values for each zone are set correctly
2364 * with respect to min_free_kbytes.
2366 static void setup_per_zone_pages_min(void)
2368 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
2369 unsigned long lowmem_pages
= 0;
2371 unsigned long flags
;
2373 /* Calculate total number of !ZONE_HIGHMEM pages */
2374 for_each_zone(zone
) {
2375 if (!is_highmem(zone
))
2376 lowmem_pages
+= zone
->present_pages
;
2379 for_each_zone(zone
) {
2380 spin_lock_irqsave(&zone
->lru_lock
, flags
);
2381 if (is_highmem(zone
)) {
2383 * Often, highmem doesn't need to reserve any pages.
2384 * But the pages_min/low/high values are also used for
2385 * batching up page reclaim activity so we need a
2386 * decent value here.
2390 min_pages
= zone
->present_pages
/ 1024;
2391 if (min_pages
< SWAP_CLUSTER_MAX
)
2392 min_pages
= SWAP_CLUSTER_MAX
;
2393 if (min_pages
> 128)
2395 zone
->pages_min
= min_pages
;
2397 /* if it's a lowmem zone, reserve a number of pages
2398 * proportionate to the zone's size.
2400 zone
->pages_min
= (pages_min
* zone
->present_pages
) /
2405 * When interpreting these watermarks, just keep in mind that:
2406 * zone->pages_min == (zone->pages_min * 4) / 4;
2408 zone
->pages_low
= (zone
->pages_min
* 5) / 4;
2409 zone
->pages_high
= (zone
->pages_min
* 6) / 4;
2410 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
2415 * Initialise min_free_kbytes.
2417 * For small machines we want it small (128k min). For large machines
2418 * we want it large (64MB max). But it is not linear, because network
2419 * bandwidth does not increase linearly with machine size. We use
2421 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
2422 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
2438 static int __init
init_per_zone_pages_min(void)
2440 unsigned long lowmem_kbytes
;
2442 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
2444 min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
2445 if (min_free_kbytes
< 128)
2446 min_free_kbytes
= 128;
2447 if (min_free_kbytes
> 65536)
2448 min_free_kbytes
= 65536;
2449 setup_per_zone_pages_min();
2450 setup_per_zone_lowmem_reserve();
2453 module_init(init_per_zone_pages_min
)
2456 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
2457 * that we can call two helper functions whenever min_free_kbytes
2460 int min_free_kbytes_sysctl_handler(ctl_table
*table
, int write
,
2461 struct file
*file
, void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2463 proc_dointvec(table
, write
, file
, buffer
, length
, ppos
);
2464 setup_per_zone_pages_min();
2469 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
2470 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
2471 * whenever sysctl_lowmem_reserve_ratio changes.
2473 * The reserve ratio obviously has absolutely no relation with the
2474 * pages_min watermarks. The lowmem reserve ratio can only make sense
2475 * if in function of the boot time zone sizes.
2477 int lowmem_reserve_ratio_sysctl_handler(ctl_table
*table
, int write
,
2478 struct file
*file
, void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2480 proc_dointvec_minmax(table
, write
, file
, buffer
, length
, ppos
);
2481 setup_per_zone_lowmem_reserve();
2485 __initdata
int hashdist
= HASHDIST_DEFAULT
;
2488 static int __init
set_hashdist(char *str
)
2492 hashdist
= simple_strtoul(str
, &str
, 0);
2495 __setup("hashdist=", set_hashdist
);
2499 * allocate a large system hash table from bootmem
2500 * - it is assumed that the hash table must contain an exact power-of-2
2501 * quantity of entries
2502 * - limit is the number of hash buckets, not the total allocation size
2504 void *__init
alloc_large_system_hash(const char *tablename
,
2505 unsigned long bucketsize
,
2506 unsigned long numentries
,
2509 unsigned int *_hash_shift
,
2510 unsigned int *_hash_mask
,
2511 unsigned long limit
)
2513 unsigned long long max
= limit
;
2514 unsigned long log2qty
, size
;
2517 /* allow the kernel cmdline to have a say */
2519 /* round applicable memory size up to nearest megabyte */
2520 numentries
= (flags
& HASH_HIGHMEM
) ? nr_all_pages
: nr_kernel_pages
;
2521 numentries
+= (1UL << (20 - PAGE_SHIFT
)) - 1;
2522 numentries
>>= 20 - PAGE_SHIFT
;
2523 numentries
<<= 20 - PAGE_SHIFT
;
2525 /* limit to 1 bucket per 2^scale bytes of low memory */
2526 if (scale
> PAGE_SHIFT
)
2527 numentries
>>= (scale
- PAGE_SHIFT
);
2529 numentries
<<= (PAGE_SHIFT
- scale
);
2531 /* rounded up to nearest power of 2 in size */
2532 numentries
= 1UL << (long_log2(numentries
) + 1);
2534 /* limit allocation size to 1/16 total memory by default */
2536 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
2537 do_div(max
, bucketsize
);
2540 if (numentries
> max
)
2543 log2qty
= long_log2(numentries
);
2546 size
= bucketsize
<< log2qty
;
2547 if (flags
& HASH_EARLY
)
2548 table
= alloc_bootmem(size
);
2550 table
= __vmalloc(size
, GFP_ATOMIC
, PAGE_KERNEL
);
2552 unsigned long order
;
2553 for (order
= 0; ((1UL << order
) << PAGE_SHIFT
) < size
; order
++)
2555 table
= (void*) __get_free_pages(GFP_ATOMIC
, order
);
2557 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
2560 panic("Failed to allocate %s hash table\n", tablename
);
2562 printk("%s hash table entries: %d (order: %d, %lu bytes)\n",
2565 long_log2(size
) - PAGE_SHIFT
,
2569 *_hash_shift
= log2qty
;
2571 *_hash_mask
= (1 << log2qty
) - 1;