2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/compiler.h>
25 #include <linux/kernel.h>
26 #include <linux/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/oom.h>
32 #include <linux/notifier.h>
33 #include <linux/topology.h>
34 #include <linux/sysctl.h>
35 #include <linux/cpu.h>
36 #include <linux/cpuset.h>
37 #include <linux/memory_hotplug.h>
38 #include <linux/nodemask.h>
39 #include <linux/vmalloc.h>
40 #include <linux/mempolicy.h>
41 #include <linux/stop_machine.h>
42 #include <linux/sort.h>
43 #include <linux/pfn.h>
44 #include <linux/backing-dev.h>
45 #include <linux/fault-inject.h>
46 #include <linux/page-isolation.h>
47 #include <linux/page_cgroup.h>
48 #include <linux/debugobjects.h>
49 #include <linux/kmemleak.h>
51 #include <asm/tlbflush.h>
52 #include <asm/div64.h>
56 * Array of node states.
58 nodemask_t node_states
[NR_NODE_STATES
] __read_mostly
= {
59 [N_POSSIBLE
] = NODE_MASK_ALL
,
60 [N_ONLINE
] = { { [0] = 1UL } },
62 [N_NORMAL_MEMORY
] = { { [0] = 1UL } },
64 [N_HIGH_MEMORY
] = { { [0] = 1UL } },
66 [N_CPU
] = { { [0] = 1UL } },
69 EXPORT_SYMBOL(node_states
);
71 unsigned long totalram_pages __read_mostly
;
72 unsigned long totalreserve_pages __read_mostly
;
73 unsigned long highest_memmap_pfn __read_mostly
;
74 int percpu_pagelist_fraction
;
76 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
77 int pageblock_order __read_mostly
;
80 static void __free_pages_ok(struct page
*page
, unsigned int order
);
83 * results with 256, 32 in the lowmem_reserve sysctl:
84 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
85 * 1G machine -> (16M dma, 784M normal, 224M high)
86 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
87 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
88 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
90 * TBD: should special case ZONE_DMA32 machines here - in those we normally
91 * don't need any ZONE_NORMAL reservation
93 int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
-1] = {
94 #ifdef CONFIG_ZONE_DMA
97 #ifdef CONFIG_ZONE_DMA32
100 #ifdef CONFIG_HIGHMEM
106 EXPORT_SYMBOL(totalram_pages
);
108 static char * const zone_names
[MAX_NR_ZONES
] = {
109 #ifdef CONFIG_ZONE_DMA
112 #ifdef CONFIG_ZONE_DMA32
116 #ifdef CONFIG_HIGHMEM
122 int min_free_kbytes
= 1024;
124 unsigned long __meminitdata nr_kernel_pages
;
125 unsigned long __meminitdata nr_all_pages
;
126 static unsigned long __meminitdata dma_reserve
;
128 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
130 * MAX_ACTIVE_REGIONS determines the maximum number of distinct
131 * ranges of memory (RAM) that may be registered with add_active_range().
132 * Ranges passed to add_active_range() will be merged if possible
133 * so the number of times add_active_range() can be called is
134 * related to the number of nodes and the number of holes
136 #ifdef CONFIG_MAX_ACTIVE_REGIONS
137 /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */
138 #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS
140 #if MAX_NUMNODES >= 32
141 /* If there can be many nodes, allow up to 50 holes per node */
142 #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50)
144 /* By default, allow up to 256 distinct regions */
145 #define MAX_ACTIVE_REGIONS 256
149 static struct node_active_region __meminitdata early_node_map
[MAX_ACTIVE_REGIONS
];
150 static int __meminitdata nr_nodemap_entries
;
151 static unsigned long __meminitdata arch_zone_lowest_possible_pfn
[MAX_NR_ZONES
];
152 static unsigned long __meminitdata arch_zone_highest_possible_pfn
[MAX_NR_ZONES
];
153 static unsigned long __initdata required_kernelcore
;
154 static unsigned long __initdata required_movablecore
;
155 static unsigned long __meminitdata zone_movable_pfn
[MAX_NUMNODES
];
157 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
159 EXPORT_SYMBOL(movable_zone
);
160 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
163 int nr_node_ids __read_mostly
= MAX_NUMNODES
;
164 int nr_online_nodes __read_mostly
= 1;
165 EXPORT_SYMBOL(nr_node_ids
);
166 EXPORT_SYMBOL(nr_online_nodes
);
169 int page_group_by_mobility_disabled __read_mostly
;
171 static void set_pageblock_migratetype(struct page
*page
, int migratetype
)
174 if (unlikely(page_group_by_mobility_disabled
))
175 migratetype
= MIGRATE_UNMOVABLE
;
177 set_pageblock_flags_group(page
, (unsigned long)migratetype
,
178 PB_migrate
, PB_migrate_end
);
181 #ifdef CONFIG_DEBUG_VM
182 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
186 unsigned long pfn
= page_to_pfn(page
);
189 seq
= zone_span_seqbegin(zone
);
190 if (pfn
>= zone
->zone_start_pfn
+ zone
->spanned_pages
)
192 else if (pfn
< zone
->zone_start_pfn
)
194 } while (zone_span_seqretry(zone
, seq
));
199 static int page_is_consistent(struct zone
*zone
, struct page
*page
)
201 if (!pfn_valid_within(page_to_pfn(page
)))
203 if (zone
!= page_zone(page
))
209 * Temporary debugging check for pages not lying within a given zone.
211 static int bad_range(struct zone
*zone
, struct page
*page
)
213 if (page_outside_zone_boundaries(zone
, page
))
215 if (!page_is_consistent(zone
, page
))
221 static inline int bad_range(struct zone
*zone
, struct page
*page
)
227 static void bad_page(struct page
*page
)
229 static unsigned long resume
;
230 static unsigned long nr_shown
;
231 static unsigned long nr_unshown
;
234 * Allow a burst of 60 reports, then keep quiet for that minute;
235 * or allow a steady drip of one report per second.
237 if (nr_shown
== 60) {
238 if (time_before(jiffies
, resume
)) {
244 "BUG: Bad page state: %lu messages suppressed\n",
251 resume
= jiffies
+ 60 * HZ
;
253 printk(KERN_ALERT
"BUG: Bad page state in process %s pfn:%05lx\n",
254 current
->comm
, page_to_pfn(page
));
256 "page:%p flags:%p count:%d mapcount:%d mapping:%p index:%lx\n",
257 page
, (void *)page
->flags
, page_count(page
),
258 page_mapcount(page
), page
->mapping
, page
->index
);
262 /* Leave bad fields for debug, except PageBuddy could make trouble */
263 __ClearPageBuddy(page
);
264 add_taint(TAINT_BAD_PAGE
);
268 * Higher-order pages are called "compound pages". They are structured thusly:
270 * The first PAGE_SIZE page is called the "head page".
272 * The remaining PAGE_SIZE pages are called "tail pages".
274 * All pages have PG_compound set. All pages have their ->private pointing at
275 * the head page (even the head page has this).
277 * The first tail page's ->lru.next holds the address of the compound page's
278 * put_page() function. Its ->lru.prev holds the order of allocation.
279 * This usage means that zero-order pages may not be compound.
282 static void free_compound_page(struct page
*page
)
284 __free_pages_ok(page
, compound_order(page
));
287 void prep_compound_page(struct page
*page
, unsigned long order
)
290 int nr_pages
= 1 << order
;
292 set_compound_page_dtor(page
, free_compound_page
);
293 set_compound_order(page
, order
);
295 for (i
= 1; i
< nr_pages
; i
++) {
296 struct page
*p
= page
+ i
;
299 p
->first_page
= page
;
303 static int destroy_compound_page(struct page
*page
, unsigned long order
)
306 int nr_pages
= 1 << order
;
309 if (unlikely(compound_order(page
) != order
) ||
310 unlikely(!PageHead(page
))) {
315 __ClearPageHead(page
);
317 for (i
= 1; i
< nr_pages
; i
++) {
318 struct page
*p
= page
+ i
;
320 if (unlikely(!PageTail(p
) || (p
->first_page
!= page
))) {
330 static inline void prep_zero_page(struct page
*page
, int order
, gfp_t gfp_flags
)
335 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
336 * and __GFP_HIGHMEM from hard or soft interrupt context.
338 VM_BUG_ON((gfp_flags
& __GFP_HIGHMEM
) && in_interrupt());
339 for (i
= 0; i
< (1 << order
); i
++)
340 clear_highpage(page
+ i
);
343 static inline void set_page_order(struct page
*page
, int order
)
345 set_page_private(page
, order
);
346 __SetPageBuddy(page
);
349 static inline void rmv_page_order(struct page
*page
)
351 __ClearPageBuddy(page
);
352 set_page_private(page
, 0);
356 * Locate the struct page for both the matching buddy in our
357 * pair (buddy1) and the combined O(n+1) page they form (page).
359 * 1) Any buddy B1 will have an order O twin B2 which satisfies
360 * the following equation:
362 * For example, if the starting buddy (buddy2) is #8 its order
364 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
366 * 2) Any buddy B will have an order O+1 parent P which
367 * satisfies the following equation:
370 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
372 static inline struct page
*
373 __page_find_buddy(struct page
*page
, unsigned long page_idx
, unsigned int order
)
375 unsigned long buddy_idx
= page_idx
^ (1 << order
);
377 return page
+ (buddy_idx
- page_idx
);
380 static inline unsigned long
381 __find_combined_index(unsigned long page_idx
, unsigned int order
)
383 return (page_idx
& ~(1 << order
));
387 * This function checks whether a page is free && is the buddy
388 * we can do coalesce a page and its buddy if
389 * (a) the buddy is not in a hole &&
390 * (b) the buddy is in the buddy system &&
391 * (c) a page and its buddy have the same order &&
392 * (d) a page and its buddy are in the same zone.
394 * For recording whether a page is in the buddy system, we use PG_buddy.
395 * Setting, clearing, and testing PG_buddy is serialized by zone->lock.
397 * For recording page's order, we use page_private(page).
399 static inline int page_is_buddy(struct page
*page
, struct page
*buddy
,
402 if (!pfn_valid_within(page_to_pfn(buddy
)))
405 if (page_zone_id(page
) != page_zone_id(buddy
))
408 if (PageBuddy(buddy
) && page_order(buddy
) == order
) {
409 VM_BUG_ON(page_count(buddy
) != 0);
416 * Freeing function for a buddy system allocator.
418 * The concept of a buddy system is to maintain direct-mapped table
419 * (containing bit values) for memory blocks of various "orders".
420 * The bottom level table contains the map for the smallest allocatable
421 * units of memory (here, pages), and each level above it describes
422 * pairs of units from the levels below, hence, "buddies".
423 * At a high level, all that happens here is marking the table entry
424 * at the bottom level available, and propagating the changes upward
425 * as necessary, plus some accounting needed to play nicely with other
426 * parts of the VM system.
427 * At each level, we keep a list of pages, which are heads of continuous
428 * free pages of length of (1 << order) and marked with PG_buddy. Page's
429 * order is recorded in page_private(page) field.
430 * So when we are allocating or freeing one, we can derive the state of the
431 * other. That is, if we allocate a small block, and both were
432 * free, the remainder of the region must be split into blocks.
433 * If a block is freed, and its buddy is also free, then this
434 * triggers coalescing into a block of larger size.
439 static inline void __free_one_page(struct page
*page
,
440 struct zone
*zone
, unsigned int order
,
443 unsigned long page_idx
;
445 if (unlikely(PageCompound(page
)))
446 if (unlikely(destroy_compound_page(page
, order
)))
449 VM_BUG_ON(migratetype
== -1);
451 page_idx
= page_to_pfn(page
) & ((1 << MAX_ORDER
) - 1);
453 VM_BUG_ON(page_idx
& ((1 << order
) - 1));
454 VM_BUG_ON(bad_range(zone
, page
));
456 while (order
< MAX_ORDER
-1) {
457 unsigned long combined_idx
;
460 buddy
= __page_find_buddy(page
, page_idx
, order
);
461 if (!page_is_buddy(page
, buddy
, order
))
464 /* Our buddy is free, merge with it and move up one order. */
465 list_del(&buddy
->lru
);
466 zone
->free_area
[order
].nr_free
--;
467 rmv_page_order(buddy
);
468 combined_idx
= __find_combined_index(page_idx
, order
);
469 page
= page
+ (combined_idx
- page_idx
);
470 page_idx
= combined_idx
;
473 set_page_order(page
, order
);
475 &zone
->free_area
[order
].free_list
[migratetype
]);
476 zone
->free_area
[order
].nr_free
++;
479 #ifdef CONFIG_HAVE_MLOCKED_PAGE_BIT
481 * free_page_mlock() -- clean up attempts to free and mlocked() page.
482 * Page should not be on lru, so no need to fix that up.
483 * free_pages_check() will verify...
485 static inline void free_page_mlock(struct page
*page
)
487 __ClearPageMlocked(page
);
488 __dec_zone_page_state(page
, NR_MLOCK
);
489 __count_vm_event(UNEVICTABLE_MLOCKFREED
);
492 static void free_page_mlock(struct page
*page
) { }
495 static inline int free_pages_check(struct page
*page
)
497 if (unlikely(page_mapcount(page
) |
498 (page
->mapping
!= NULL
) |
499 (atomic_read(&page
->_count
) != 0) |
500 (page
->flags
& PAGE_FLAGS_CHECK_AT_FREE
))) {
504 if (page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)
505 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
510 * Frees a list of pages.
511 * Assumes all pages on list are in same zone, and of same order.
512 * count is the number of pages to free.
514 * If the zone was previously in an "all pages pinned" state then look to
515 * see if this freeing clears that state.
517 * And clear the zone's pages_scanned counter, to hold off the "all pages are
518 * pinned" detection logic.
520 static void free_pages_bulk(struct zone
*zone
, int count
,
521 struct list_head
*list
, int order
)
523 spin_lock(&zone
->lock
);
524 zone_clear_flag(zone
, ZONE_ALL_UNRECLAIMABLE
);
525 zone
->pages_scanned
= 0;
527 __mod_zone_page_state(zone
, NR_FREE_PAGES
, count
<< order
);
531 VM_BUG_ON(list_empty(list
));
532 page
= list_entry(list
->prev
, struct page
, lru
);
533 /* have to delete it as __free_one_page list manipulates */
534 list_del(&page
->lru
);
535 __free_one_page(page
, zone
, order
, page_private(page
));
537 spin_unlock(&zone
->lock
);
540 static void free_one_page(struct zone
*zone
, struct page
*page
, int order
,
543 spin_lock(&zone
->lock
);
544 zone_clear_flag(zone
, ZONE_ALL_UNRECLAIMABLE
);
545 zone
->pages_scanned
= 0;
547 __mod_zone_page_state(zone
, NR_FREE_PAGES
, 1 << order
);
548 __free_one_page(page
, zone
, order
, migratetype
);
549 spin_unlock(&zone
->lock
);
552 static void __free_pages_ok(struct page
*page
, unsigned int order
)
557 int clearMlocked
= PageMlocked(page
);
559 for (i
= 0 ; i
< (1 << order
) ; ++i
)
560 bad
+= free_pages_check(page
+ i
);
564 if (!PageHighMem(page
)) {
565 debug_check_no_locks_freed(page_address(page
),PAGE_SIZE
<<order
);
566 debug_check_no_obj_freed(page_address(page
),
569 arch_free_page(page
, order
);
570 kernel_map_pages(page
, 1 << order
, 0);
572 local_irq_save(flags
);
573 if (unlikely(clearMlocked
))
574 free_page_mlock(page
);
575 __count_vm_events(PGFREE
, 1 << order
);
576 free_one_page(page_zone(page
), page
, order
,
577 get_pageblock_migratetype(page
));
578 local_irq_restore(flags
);
582 * permit the bootmem allocator to evade page validation on high-order frees
584 void __meminit
__free_pages_bootmem(struct page
*page
, unsigned int order
)
587 __ClearPageReserved(page
);
588 set_page_count(page
, 0);
589 set_page_refcounted(page
);
595 for (loop
= 0; loop
< BITS_PER_LONG
; loop
++) {
596 struct page
*p
= &page
[loop
];
598 if (loop
+ 1 < BITS_PER_LONG
)
600 __ClearPageReserved(p
);
601 set_page_count(p
, 0);
604 set_page_refcounted(page
);
605 __free_pages(page
, order
);
611 * The order of subdivision here is critical for the IO subsystem.
612 * Please do not alter this order without good reasons and regression
613 * testing. Specifically, as large blocks of memory are subdivided,
614 * the order in which smaller blocks are delivered depends on the order
615 * they're subdivided in this function. This is the primary factor
616 * influencing the order in which pages are delivered to the IO
617 * subsystem according to empirical testing, and this is also justified
618 * by considering the behavior of a buddy system containing a single
619 * large block of memory acted on by a series of small allocations.
620 * This behavior is a critical factor in sglist merging's success.
624 static inline void expand(struct zone
*zone
, struct page
*page
,
625 int low
, int high
, struct free_area
*area
,
628 unsigned long size
= 1 << high
;
634 VM_BUG_ON(bad_range(zone
, &page
[size
]));
635 list_add(&page
[size
].lru
, &area
->free_list
[migratetype
]);
637 set_page_order(&page
[size
], high
);
642 * This page is about to be returned from the page allocator
644 static int prep_new_page(struct page
*page
, int order
, gfp_t gfp_flags
)
646 if (unlikely(page_mapcount(page
) |
647 (page
->mapping
!= NULL
) |
648 (atomic_read(&page
->_count
) != 0) |
649 (page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
))) {
654 set_page_private(page
, 0);
655 set_page_refcounted(page
);
657 arch_alloc_page(page
, order
);
658 kernel_map_pages(page
, 1 << order
, 1);
660 if (gfp_flags
& __GFP_ZERO
)
661 prep_zero_page(page
, order
, gfp_flags
);
663 if (order
&& (gfp_flags
& __GFP_COMP
))
664 prep_compound_page(page
, order
);
670 * Go through the free lists for the given migratetype and remove
671 * the smallest available page from the freelists
674 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
677 unsigned int current_order
;
678 struct free_area
* area
;
681 /* Find a page of the appropriate size in the preferred list */
682 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
683 area
= &(zone
->free_area
[current_order
]);
684 if (list_empty(&area
->free_list
[migratetype
]))
687 page
= list_entry(area
->free_list
[migratetype
].next
,
689 list_del(&page
->lru
);
690 rmv_page_order(page
);
692 expand(zone
, page
, order
, current_order
, area
, migratetype
);
701 * This array describes the order lists are fallen back to when
702 * the free lists for the desirable migrate type are depleted
704 static int fallbacks
[MIGRATE_TYPES
][MIGRATE_TYPES
-1] = {
705 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
706 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
707 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_RESERVE
},
708 [MIGRATE_RESERVE
] = { MIGRATE_RESERVE
, MIGRATE_RESERVE
, MIGRATE_RESERVE
}, /* Never used */
712 * Move the free pages in a range to the free lists of the requested type.
713 * Note that start_page and end_pages are not aligned on a pageblock
714 * boundary. If alignment is required, use move_freepages_block()
716 static int move_freepages(struct zone
*zone
,
717 struct page
*start_page
, struct page
*end_page
,
724 #ifndef CONFIG_HOLES_IN_ZONE
726 * page_zone is not safe to call in this context when
727 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
728 * anyway as we check zone boundaries in move_freepages_block().
729 * Remove at a later date when no bug reports exist related to
730 * grouping pages by mobility
732 BUG_ON(page_zone(start_page
) != page_zone(end_page
));
735 for (page
= start_page
; page
<= end_page
;) {
736 /* Make sure we are not inadvertently changing nodes */
737 VM_BUG_ON(page_to_nid(page
) != zone_to_nid(zone
));
739 if (!pfn_valid_within(page_to_pfn(page
))) {
744 if (!PageBuddy(page
)) {
749 order
= page_order(page
);
750 list_del(&page
->lru
);
752 &zone
->free_area
[order
].free_list
[migratetype
]);
754 pages_moved
+= 1 << order
;
760 static int move_freepages_block(struct zone
*zone
, struct page
*page
,
763 unsigned long start_pfn
, end_pfn
;
764 struct page
*start_page
, *end_page
;
766 start_pfn
= page_to_pfn(page
);
767 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
768 start_page
= pfn_to_page(start_pfn
);
769 end_page
= start_page
+ pageblock_nr_pages
- 1;
770 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
772 /* Do not cross zone boundaries */
773 if (start_pfn
< zone
->zone_start_pfn
)
775 if (end_pfn
>= zone
->zone_start_pfn
+ zone
->spanned_pages
)
778 return move_freepages(zone
, start_page
, end_page
, migratetype
);
781 /* Remove an element from the buddy allocator from the fallback list */
782 static inline struct page
*
783 __rmqueue_fallback(struct zone
*zone
, int order
, int start_migratetype
)
785 struct free_area
* area
;
790 /* Find the largest possible block of pages in the other list */
791 for (current_order
= MAX_ORDER
-1; current_order
>= order
;
793 for (i
= 0; i
< MIGRATE_TYPES
- 1; i
++) {
794 migratetype
= fallbacks
[start_migratetype
][i
];
796 /* MIGRATE_RESERVE handled later if necessary */
797 if (migratetype
== MIGRATE_RESERVE
)
800 area
= &(zone
->free_area
[current_order
]);
801 if (list_empty(&area
->free_list
[migratetype
]))
804 page
= list_entry(area
->free_list
[migratetype
].next
,
809 * If breaking a large block of pages, move all free
810 * pages to the preferred allocation list. If falling
811 * back for a reclaimable kernel allocation, be more
812 * agressive about taking ownership of free pages
814 if (unlikely(current_order
>= (pageblock_order
>> 1)) ||
815 start_migratetype
== MIGRATE_RECLAIMABLE
) {
817 pages
= move_freepages_block(zone
, page
,
820 /* Claim the whole block if over half of it is free */
821 if (pages
>= (1 << (pageblock_order
-1)))
822 set_pageblock_migratetype(page
,
825 migratetype
= start_migratetype
;
828 /* Remove the page from the freelists */
829 list_del(&page
->lru
);
830 rmv_page_order(page
);
832 if (current_order
== pageblock_order
)
833 set_pageblock_migratetype(page
,
836 expand(zone
, page
, order
, current_order
, area
, migratetype
);
845 * Do the hard work of removing an element from the buddy allocator.
846 * Call me with the zone->lock already held.
848 static struct page
*__rmqueue(struct zone
*zone
, unsigned int order
,
854 page
= __rmqueue_smallest(zone
, order
, migratetype
);
856 if (unlikely(!page
) && migratetype
!= MIGRATE_RESERVE
) {
857 page
= __rmqueue_fallback(zone
, order
, migratetype
);
860 * Use MIGRATE_RESERVE rather than fail an allocation. goto
861 * is used because __rmqueue_smallest is an inline function
862 * and we want just one call site
865 migratetype
= MIGRATE_RESERVE
;
874 * Obtain a specified number of elements from the buddy allocator, all under
875 * a single hold of the lock, for efficiency. Add them to the supplied list.
876 * Returns the number of new pages which were placed at *list.
878 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
879 unsigned long count
, struct list_head
*list
,
884 spin_lock(&zone
->lock
);
885 for (i
= 0; i
< count
; ++i
) {
886 struct page
*page
= __rmqueue(zone
, order
, migratetype
);
887 if (unlikely(page
== NULL
))
891 * Split buddy pages returned by expand() are received here
892 * in physical page order. The page is added to the callers and
893 * list and the list head then moves forward. From the callers
894 * perspective, the linked list is ordered by page number in
895 * some conditions. This is useful for IO devices that can
896 * merge IO requests if the physical pages are ordered
899 list_add(&page
->lru
, list
);
900 set_page_private(page
, migratetype
);
903 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
904 spin_unlock(&zone
->lock
);
910 * Called from the vmstat counter updater to drain pagesets of this
911 * currently executing processor on remote nodes after they have
914 * Note that this function must be called with the thread pinned to
915 * a single processor.
917 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
922 local_irq_save(flags
);
923 if (pcp
->count
>= pcp
->batch
)
924 to_drain
= pcp
->batch
;
926 to_drain
= pcp
->count
;
927 free_pages_bulk(zone
, to_drain
, &pcp
->list
, 0);
928 pcp
->count
-= to_drain
;
929 local_irq_restore(flags
);
934 * Drain pages of the indicated processor.
936 * The processor must either be the current processor and the
937 * thread pinned to the current processor or a processor that
940 static void drain_pages(unsigned int cpu
)
945 for_each_populated_zone(zone
) {
946 struct per_cpu_pageset
*pset
;
947 struct per_cpu_pages
*pcp
;
949 pset
= zone_pcp(zone
, cpu
);
952 local_irq_save(flags
);
953 free_pages_bulk(zone
, pcp
->count
, &pcp
->list
, 0);
955 local_irq_restore(flags
);
960 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
962 void drain_local_pages(void *arg
)
964 drain_pages(smp_processor_id());
968 * Spill all the per-cpu pages from all CPUs back into the buddy allocator
970 void drain_all_pages(void)
972 on_each_cpu(drain_local_pages
, NULL
, 1);
975 #ifdef CONFIG_HIBERNATION
977 void mark_free_pages(struct zone
*zone
)
979 unsigned long pfn
, max_zone_pfn
;
982 struct list_head
*curr
;
984 if (!zone
->spanned_pages
)
987 spin_lock_irqsave(&zone
->lock
, flags
);
989 max_zone_pfn
= zone
->zone_start_pfn
+ zone
->spanned_pages
;
990 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
991 if (pfn_valid(pfn
)) {
992 struct page
*page
= pfn_to_page(pfn
);
994 if (!swsusp_page_is_forbidden(page
))
995 swsusp_unset_page_free(page
);
998 for_each_migratetype_order(order
, t
) {
999 list_for_each(curr
, &zone
->free_area
[order
].free_list
[t
]) {
1002 pfn
= page_to_pfn(list_entry(curr
, struct page
, lru
));
1003 for (i
= 0; i
< (1UL << order
); i
++)
1004 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
1007 spin_unlock_irqrestore(&zone
->lock
, flags
);
1009 #endif /* CONFIG_PM */
1012 * Free a 0-order page
1014 static void free_hot_cold_page(struct page
*page
, int cold
)
1016 struct zone
*zone
= page_zone(page
);
1017 struct per_cpu_pages
*pcp
;
1018 unsigned long flags
;
1019 int clearMlocked
= PageMlocked(page
);
1022 page
->mapping
= NULL
;
1023 if (free_pages_check(page
))
1026 if (!PageHighMem(page
)) {
1027 debug_check_no_locks_freed(page_address(page
), PAGE_SIZE
);
1028 debug_check_no_obj_freed(page_address(page
), PAGE_SIZE
);
1030 arch_free_page(page
, 0);
1031 kernel_map_pages(page
, 1, 0);
1033 pcp
= &zone_pcp(zone
, get_cpu())->pcp
;
1034 set_page_private(page
, get_pageblock_migratetype(page
));
1035 local_irq_save(flags
);
1036 if (unlikely(clearMlocked
))
1037 free_page_mlock(page
);
1038 __count_vm_event(PGFREE
);
1041 list_add_tail(&page
->lru
, &pcp
->list
);
1043 list_add(&page
->lru
, &pcp
->list
);
1045 if (pcp
->count
>= pcp
->high
) {
1046 free_pages_bulk(zone
, pcp
->batch
, &pcp
->list
, 0);
1047 pcp
->count
-= pcp
->batch
;
1049 local_irq_restore(flags
);
1053 void free_hot_page(struct page
*page
)
1055 free_hot_cold_page(page
, 0);
1058 void free_cold_page(struct page
*page
)
1060 free_hot_cold_page(page
, 1);
1064 * split_page takes a non-compound higher-order page, and splits it into
1065 * n (1<<order) sub-pages: page[0..n]
1066 * Each sub-page must be freed individually.
1068 * Note: this is probably too low level an operation for use in drivers.
1069 * Please consult with lkml before using this in your driver.
1071 void split_page(struct page
*page
, unsigned int order
)
1075 VM_BUG_ON(PageCompound(page
));
1076 VM_BUG_ON(!page_count(page
));
1077 for (i
= 1; i
< (1 << order
); i
++)
1078 set_page_refcounted(page
+ i
);
1082 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1083 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1087 struct page
*buffered_rmqueue(struct zone
*preferred_zone
,
1088 struct zone
*zone
, int order
, gfp_t gfp_flags
,
1091 unsigned long flags
;
1093 int cold
= !!(gfp_flags
& __GFP_COLD
);
1098 if (likely(order
== 0)) {
1099 struct per_cpu_pages
*pcp
;
1101 pcp
= &zone_pcp(zone
, cpu
)->pcp
;
1102 local_irq_save(flags
);
1104 pcp
->count
= rmqueue_bulk(zone
, 0,
1105 pcp
->batch
, &pcp
->list
, migratetype
);
1106 if (unlikely(!pcp
->count
))
1110 /* Find a page of the appropriate migrate type */
1112 list_for_each_entry_reverse(page
, &pcp
->list
, lru
)
1113 if (page_private(page
) == migratetype
)
1116 list_for_each_entry(page
, &pcp
->list
, lru
)
1117 if (page_private(page
) == migratetype
)
1121 /* Allocate more to the pcp list if necessary */
1122 if (unlikely(&page
->lru
== &pcp
->list
)) {
1123 pcp
->count
+= rmqueue_bulk(zone
, 0,
1124 pcp
->batch
, &pcp
->list
, migratetype
);
1125 page
= list_entry(pcp
->list
.next
, struct page
, lru
);
1128 list_del(&page
->lru
);
1131 if (unlikely(gfp_flags
& __GFP_NOFAIL
)) {
1133 * __GFP_NOFAIL is not to be used in new code.
1135 * All __GFP_NOFAIL callers should be fixed so that they
1136 * properly detect and handle allocation failures.
1138 * We most definitely don't want callers attempting to
1139 * allocate greater than single-page units with
1142 WARN_ON_ONCE(order
> 0);
1144 spin_lock_irqsave(&zone
->lock
, flags
);
1145 page
= __rmqueue(zone
, order
, migratetype
);
1146 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(1 << order
));
1147 spin_unlock(&zone
->lock
);
1152 __count_zone_vm_events(PGALLOC
, zone
, 1 << order
);
1153 zone_statistics(preferred_zone
, zone
);
1154 local_irq_restore(flags
);
1157 VM_BUG_ON(bad_range(zone
, page
));
1158 if (prep_new_page(page
, order
, gfp_flags
))
1163 local_irq_restore(flags
);
1168 /* The ALLOC_WMARK bits are used as an index to zone->watermark */
1169 #define ALLOC_WMARK_MIN WMARK_MIN
1170 #define ALLOC_WMARK_LOW WMARK_LOW
1171 #define ALLOC_WMARK_HIGH WMARK_HIGH
1172 #define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */
1174 /* Mask to get the watermark bits */
1175 #define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1)
1177 #define ALLOC_HARDER 0x10 /* try to alloc harder */
1178 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
1179 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
1181 #ifdef CONFIG_FAIL_PAGE_ALLOC
1183 static struct fail_page_alloc_attr
{
1184 struct fault_attr attr
;
1186 u32 ignore_gfp_highmem
;
1187 u32 ignore_gfp_wait
;
1190 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1192 struct dentry
*ignore_gfp_highmem_file
;
1193 struct dentry
*ignore_gfp_wait_file
;
1194 struct dentry
*min_order_file
;
1196 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1198 } fail_page_alloc
= {
1199 .attr
= FAULT_ATTR_INITIALIZER
,
1200 .ignore_gfp_wait
= 1,
1201 .ignore_gfp_highmem
= 1,
1205 static int __init
setup_fail_page_alloc(char *str
)
1207 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
1209 __setup("fail_page_alloc=", setup_fail_page_alloc
);
1211 static int should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1213 if (order
< fail_page_alloc
.min_order
)
1215 if (gfp_mask
& __GFP_NOFAIL
)
1217 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
1219 if (fail_page_alloc
.ignore_gfp_wait
&& (gfp_mask
& __GFP_WAIT
))
1222 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
1225 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1227 static int __init
fail_page_alloc_debugfs(void)
1229 mode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
1233 err
= init_fault_attr_dentries(&fail_page_alloc
.attr
,
1237 dir
= fail_page_alloc
.attr
.dentries
.dir
;
1239 fail_page_alloc
.ignore_gfp_wait_file
=
1240 debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
1241 &fail_page_alloc
.ignore_gfp_wait
);
1243 fail_page_alloc
.ignore_gfp_highmem_file
=
1244 debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
1245 &fail_page_alloc
.ignore_gfp_highmem
);
1246 fail_page_alloc
.min_order_file
=
1247 debugfs_create_u32("min-order", mode
, dir
,
1248 &fail_page_alloc
.min_order
);
1250 if (!fail_page_alloc
.ignore_gfp_wait_file
||
1251 !fail_page_alloc
.ignore_gfp_highmem_file
||
1252 !fail_page_alloc
.min_order_file
) {
1254 debugfs_remove(fail_page_alloc
.ignore_gfp_wait_file
);
1255 debugfs_remove(fail_page_alloc
.ignore_gfp_highmem_file
);
1256 debugfs_remove(fail_page_alloc
.min_order_file
);
1257 cleanup_fault_attr_dentries(&fail_page_alloc
.attr
);
1263 late_initcall(fail_page_alloc_debugfs
);
1265 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1267 #else /* CONFIG_FAIL_PAGE_ALLOC */
1269 static inline int should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1274 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1277 * Return 1 if free pages are above 'mark'. This takes into account the order
1278 * of the allocation.
1280 int zone_watermark_ok(struct zone
*z
, int order
, unsigned long mark
,
1281 int classzone_idx
, int alloc_flags
)
1283 /* free_pages my go negative - that's OK */
1285 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
) - (1 << order
) + 1;
1288 if (alloc_flags
& ALLOC_HIGH
)
1290 if (alloc_flags
& ALLOC_HARDER
)
1293 if (free_pages
<= min
+ z
->lowmem_reserve
[classzone_idx
])
1295 for (o
= 0; o
< order
; o
++) {
1296 /* At the next order, this order's pages become unavailable */
1297 free_pages
-= z
->free_area
[o
].nr_free
<< o
;
1299 /* Require fewer higher order pages to be free */
1302 if (free_pages
<= min
)
1310 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1311 * skip over zones that are not allowed by the cpuset, or that have
1312 * been recently (in last second) found to be nearly full. See further
1313 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1314 * that have to skip over a lot of full or unallowed zones.
1316 * If the zonelist cache is present in the passed in zonelist, then
1317 * returns a pointer to the allowed node mask (either the current
1318 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1320 * If the zonelist cache is not available for this zonelist, does
1321 * nothing and returns NULL.
1323 * If the fullzones BITMAP in the zonelist cache is stale (more than
1324 * a second since last zap'd) then we zap it out (clear its bits.)
1326 * We hold off even calling zlc_setup, until after we've checked the
1327 * first zone in the zonelist, on the theory that most allocations will
1328 * be satisfied from that first zone, so best to examine that zone as
1329 * quickly as we can.
1331 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1333 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1334 nodemask_t
*allowednodes
; /* zonelist_cache approximation */
1336 zlc
= zonelist
->zlcache_ptr
;
1340 if (time_after(jiffies
, zlc
->last_full_zap
+ HZ
)) {
1341 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
1342 zlc
->last_full_zap
= jiffies
;
1345 allowednodes
= !in_interrupt() && (alloc_flags
& ALLOC_CPUSET
) ?
1346 &cpuset_current_mems_allowed
:
1347 &node_states
[N_HIGH_MEMORY
];
1348 return allowednodes
;
1352 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1353 * if it is worth looking at further for free memory:
1354 * 1) Check that the zone isn't thought to be full (doesn't have its
1355 * bit set in the zonelist_cache fullzones BITMAP).
1356 * 2) Check that the zones node (obtained from the zonelist_cache
1357 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1358 * Return true (non-zero) if zone is worth looking at further, or
1359 * else return false (zero) if it is not.
1361 * This check -ignores- the distinction between various watermarks,
1362 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1363 * found to be full for any variation of these watermarks, it will
1364 * be considered full for up to one second by all requests, unless
1365 * we are so low on memory on all allowed nodes that we are forced
1366 * into the second scan of the zonelist.
1368 * In the second scan we ignore this zonelist cache and exactly
1369 * apply the watermarks to all zones, even it is slower to do so.
1370 * We are low on memory in the second scan, and should leave no stone
1371 * unturned looking for a free page.
1373 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1374 nodemask_t
*allowednodes
)
1376 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1377 int i
; /* index of *z in zonelist zones */
1378 int n
; /* node that zone *z is on */
1380 zlc
= zonelist
->zlcache_ptr
;
1384 i
= z
- zonelist
->_zonerefs
;
1387 /* This zone is worth trying if it is allowed but not full */
1388 return node_isset(n
, *allowednodes
) && !test_bit(i
, zlc
->fullzones
);
1392 * Given 'z' scanning a zonelist, set the corresponding bit in
1393 * zlc->fullzones, so that subsequent attempts to allocate a page
1394 * from that zone don't waste time re-examining it.
1396 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1398 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1399 int i
; /* index of *z in zonelist zones */
1401 zlc
= zonelist
->zlcache_ptr
;
1405 i
= z
- zonelist
->_zonerefs
;
1407 set_bit(i
, zlc
->fullzones
);
1410 #else /* CONFIG_NUMA */
1412 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1417 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1418 nodemask_t
*allowednodes
)
1423 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1426 #endif /* CONFIG_NUMA */
1429 * get_page_from_freelist goes through the zonelist trying to allocate
1432 static struct page
*
1433 get_page_from_freelist(gfp_t gfp_mask
, nodemask_t
*nodemask
, unsigned int order
,
1434 struct zonelist
*zonelist
, int high_zoneidx
, int alloc_flags
,
1435 struct zone
*preferred_zone
, int migratetype
)
1438 struct page
*page
= NULL
;
1441 nodemask_t
*allowednodes
= NULL
;/* zonelist_cache approximation */
1442 int zlc_active
= 0; /* set if using zonelist_cache */
1443 int did_zlc_setup
= 0; /* just call zlc_setup() one time */
1445 classzone_idx
= zone_idx(preferred_zone
);
1448 * Scan zonelist, looking for a zone with enough free.
1449 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1451 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
1452 high_zoneidx
, nodemask
) {
1453 if (NUMA_BUILD
&& zlc_active
&&
1454 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
1456 if ((alloc_flags
& ALLOC_CPUSET
) &&
1457 !cpuset_zone_allowed_softwall(zone
, gfp_mask
))
1460 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
1461 if (!(alloc_flags
& ALLOC_NO_WATERMARKS
)) {
1463 mark
= zone
->watermark
[alloc_flags
& ALLOC_WMARK_MASK
];
1464 if (!zone_watermark_ok(zone
, order
, mark
,
1465 classzone_idx
, alloc_flags
)) {
1466 if (!zone_reclaim_mode
||
1467 !zone_reclaim(zone
, gfp_mask
, order
))
1468 goto this_zone_full
;
1472 page
= buffered_rmqueue(preferred_zone
, zone
, order
,
1473 gfp_mask
, migratetype
);
1478 zlc_mark_zone_full(zonelist
, z
);
1480 if (NUMA_BUILD
&& !did_zlc_setup
&& nr_online_nodes
> 1) {
1482 * we do zlc_setup after the first zone is tried but only
1483 * if there are multiple nodes make it worthwhile
1485 allowednodes
= zlc_setup(zonelist
, alloc_flags
);
1491 if (unlikely(NUMA_BUILD
&& page
== NULL
&& zlc_active
)) {
1492 /* Disable zlc cache for second zonelist scan */
1500 should_alloc_retry(gfp_t gfp_mask
, unsigned int order
,
1501 unsigned long pages_reclaimed
)
1503 /* Do not loop if specifically requested */
1504 if (gfp_mask
& __GFP_NORETRY
)
1508 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
1509 * means __GFP_NOFAIL, but that may not be true in other
1512 if (order
<= PAGE_ALLOC_COSTLY_ORDER
)
1516 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
1517 * specified, then we retry until we no longer reclaim any pages
1518 * (above), or we've reclaimed an order of pages at least as
1519 * large as the allocation's order. In both cases, if the
1520 * allocation still fails, we stop retrying.
1522 if (gfp_mask
& __GFP_REPEAT
&& pages_reclaimed
< (1 << order
))
1526 * Don't let big-order allocations loop unless the caller
1527 * explicitly requests that.
1529 if (gfp_mask
& __GFP_NOFAIL
)
1535 static inline struct page
*
1536 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
1537 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1538 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
1543 /* Acquire the OOM killer lock for the zones in zonelist */
1544 if (!try_set_zone_oom(zonelist
, gfp_mask
)) {
1545 schedule_timeout_uninterruptible(1);
1550 * Go through the zonelist yet one more time, keep very high watermark
1551 * here, this is only to catch a parallel oom killing, we must fail if
1552 * we're still under heavy pressure.
1554 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
,
1555 order
, zonelist
, high_zoneidx
,
1556 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
,
1557 preferred_zone
, migratetype
);
1561 /* The OOM killer will not help higher order allocs */
1562 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
1565 /* Exhausted what can be done so it's blamo time */
1566 out_of_memory(zonelist
, gfp_mask
, order
);
1569 clear_zonelist_oom(zonelist
, gfp_mask
);
1573 /* The really slow allocator path where we enter direct reclaim */
1574 static inline struct page
*
1575 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
1576 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1577 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
1578 int migratetype
, unsigned long *did_some_progress
)
1580 struct page
*page
= NULL
;
1581 struct reclaim_state reclaim_state
;
1582 struct task_struct
*p
= current
;
1586 /* We now go into synchronous reclaim */
1587 cpuset_memory_pressure_bump();
1590 * The task's cpuset might have expanded its set of allowable nodes
1592 p
->flags
|= PF_MEMALLOC
;
1593 lockdep_set_current_reclaim_state(gfp_mask
);
1594 reclaim_state
.reclaimed_slab
= 0;
1595 p
->reclaim_state
= &reclaim_state
;
1597 *did_some_progress
= try_to_free_pages(zonelist
, order
, gfp_mask
, nodemask
);
1599 p
->reclaim_state
= NULL
;
1600 lockdep_clear_current_reclaim_state();
1601 p
->flags
&= ~PF_MEMALLOC
;
1608 if (likely(*did_some_progress
))
1609 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
1610 zonelist
, high_zoneidx
,
1611 alloc_flags
, preferred_zone
,
1617 * This is called in the allocator slow-path if the allocation request is of
1618 * sufficient urgency to ignore watermarks and take other desperate measures
1620 static inline struct page
*
1621 __alloc_pages_high_priority(gfp_t gfp_mask
, unsigned int order
,
1622 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1623 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
1629 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
1630 zonelist
, high_zoneidx
, ALLOC_NO_WATERMARKS
,
1631 preferred_zone
, migratetype
);
1633 if (!page
&& gfp_mask
& __GFP_NOFAIL
)
1634 congestion_wait(WRITE
, HZ
/50);
1635 } while (!page
&& (gfp_mask
& __GFP_NOFAIL
));
1641 void wake_all_kswapd(unsigned int order
, struct zonelist
*zonelist
,
1642 enum zone_type high_zoneidx
)
1647 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
)
1648 wakeup_kswapd(zone
, order
);
1652 gfp_to_alloc_flags(gfp_t gfp_mask
)
1654 struct task_struct
*p
= current
;
1655 int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
1656 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
1658 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
1659 BUILD_BUG_ON(__GFP_HIGH
!= ALLOC_HIGH
);
1662 * The caller may dip into page reserves a bit more if the caller
1663 * cannot run direct reclaim, or if the caller has realtime scheduling
1664 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
1665 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
1667 alloc_flags
|= (gfp_mask
& __GFP_HIGH
);
1670 alloc_flags
|= ALLOC_HARDER
;
1672 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
1673 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1675 alloc_flags
&= ~ALLOC_CPUSET
;
1676 } else if (unlikely(rt_task(p
)))
1677 alloc_flags
|= ALLOC_HARDER
;
1679 if (likely(!(gfp_mask
& __GFP_NOMEMALLOC
))) {
1680 if (!in_interrupt() &&
1681 ((p
->flags
& PF_MEMALLOC
) ||
1682 unlikely(test_thread_flag(TIF_MEMDIE
))))
1683 alloc_flags
|= ALLOC_NO_WATERMARKS
;
1689 static inline struct page
*
1690 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
1691 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1692 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
1695 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
1696 struct page
*page
= NULL
;
1698 unsigned long pages_reclaimed
= 0;
1699 unsigned long did_some_progress
;
1700 struct task_struct
*p
= current
;
1703 * In the slowpath, we sanity check order to avoid ever trying to
1704 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
1705 * be using allocators in order of preference for an area that is
1708 if (WARN_ON_ONCE(order
>= MAX_ORDER
))
1712 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
1713 * __GFP_NOWARN set) should not cause reclaim since the subsystem
1714 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
1715 * using a larger set of nodes after it has established that the
1716 * allowed per node queues are empty and that nodes are
1719 if (NUMA_BUILD
&& (gfp_mask
& GFP_THISNODE
) == GFP_THISNODE
)
1722 wake_all_kswapd(order
, zonelist
, high_zoneidx
);
1725 * OK, we're below the kswapd watermark and have kicked background
1726 * reclaim. Now things get more complex, so set up alloc_flags according
1727 * to how we want to proceed.
1729 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
1732 /* This is the last chance, in general, before the goto nopage. */
1733 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
, zonelist
,
1734 high_zoneidx
, alloc_flags
& ~ALLOC_NO_WATERMARKS
,
1735 preferred_zone
, migratetype
);
1740 /* Allocate without watermarks if the context allows */
1741 if (alloc_flags
& ALLOC_NO_WATERMARKS
) {
1742 page
= __alloc_pages_high_priority(gfp_mask
, order
,
1743 zonelist
, high_zoneidx
, nodemask
,
1744 preferred_zone
, migratetype
);
1749 /* Atomic allocations - we can't balance anything */
1753 /* Avoid recursion of direct reclaim */
1754 if (p
->flags
& PF_MEMALLOC
)
1757 /* Try direct reclaim and then allocating */
1758 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
,
1759 zonelist
, high_zoneidx
,
1761 alloc_flags
, preferred_zone
,
1762 migratetype
, &did_some_progress
);
1767 * If we failed to make any progress reclaiming, then we are
1768 * running out of options and have to consider going OOM
1770 if (!did_some_progress
) {
1771 if ((gfp_mask
& __GFP_FS
) && !(gfp_mask
& __GFP_NORETRY
)) {
1772 page
= __alloc_pages_may_oom(gfp_mask
, order
,
1773 zonelist
, high_zoneidx
,
1774 nodemask
, preferred_zone
,
1780 * The OOM killer does not trigger for high-order allocations
1781 * but if no progress is being made, there are no other
1782 * options and retrying is unlikely to help
1784 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
1791 /* Check if we should retry the allocation */
1792 pages_reclaimed
+= did_some_progress
;
1793 if (should_alloc_retry(gfp_mask
, order
, pages_reclaimed
)) {
1794 /* Wait for some write requests to complete then retry */
1795 congestion_wait(WRITE
, HZ
/50);
1800 if (!(gfp_mask
& __GFP_NOWARN
) && printk_ratelimit()) {
1801 printk(KERN_WARNING
"%s: page allocation failure."
1802 " order:%d, mode:0x%x\n",
1803 p
->comm
, order
, gfp_mask
);
1813 * This is the 'heart' of the zoned buddy allocator.
1816 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
,
1817 struct zonelist
*zonelist
, nodemask_t
*nodemask
)
1819 enum zone_type high_zoneidx
= gfp_zone(gfp_mask
);
1820 struct zone
*preferred_zone
;
1822 int migratetype
= allocflags_to_migratetype(gfp_mask
);
1824 lockdep_trace_alloc(gfp_mask
);
1826 might_sleep_if(gfp_mask
& __GFP_WAIT
);
1828 if (should_fail_alloc_page(gfp_mask
, order
))
1832 * Check the zones suitable for the gfp_mask contain at least one
1833 * valid zone. It's possible to have an empty zonelist as a result
1834 * of GFP_THISNODE and a memoryless node
1836 if (unlikely(!zonelist
->_zonerefs
->zone
))
1839 /* The preferred zone is used for statistics later */
1840 first_zones_zonelist(zonelist
, high_zoneidx
, nodemask
, &preferred_zone
);
1841 if (!preferred_zone
)
1844 /* First allocation attempt */
1845 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
, order
,
1846 zonelist
, high_zoneidx
, ALLOC_WMARK_LOW
|ALLOC_CPUSET
,
1847 preferred_zone
, migratetype
);
1848 if (unlikely(!page
))
1849 page
= __alloc_pages_slowpath(gfp_mask
, order
,
1850 zonelist
, high_zoneidx
, nodemask
,
1851 preferred_zone
, migratetype
);
1855 EXPORT_SYMBOL(__alloc_pages_nodemask
);
1858 * Common helper functions.
1860 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
1863 page
= alloc_pages(gfp_mask
, order
);
1866 return (unsigned long) page_address(page
);
1869 EXPORT_SYMBOL(__get_free_pages
);
1871 unsigned long get_zeroed_page(gfp_t gfp_mask
)
1876 * get_zeroed_page() returns a 32-bit address, which cannot represent
1879 VM_BUG_ON((gfp_mask
& __GFP_HIGHMEM
) != 0);
1881 page
= alloc_pages(gfp_mask
| __GFP_ZERO
, 0);
1883 return (unsigned long) page_address(page
);
1887 EXPORT_SYMBOL(get_zeroed_page
);
1889 void __pagevec_free(struct pagevec
*pvec
)
1891 int i
= pagevec_count(pvec
);
1894 free_hot_cold_page(pvec
->pages
[i
], pvec
->cold
);
1897 void __free_pages(struct page
*page
, unsigned int order
)
1899 if (put_page_testzero(page
)) {
1901 free_hot_page(page
);
1903 __free_pages_ok(page
, order
);
1907 EXPORT_SYMBOL(__free_pages
);
1909 void free_pages(unsigned long addr
, unsigned int order
)
1912 VM_BUG_ON(!virt_addr_valid((void *)addr
));
1913 __free_pages(virt_to_page((void *)addr
), order
);
1917 EXPORT_SYMBOL(free_pages
);
1920 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
1921 * @size: the number of bytes to allocate
1922 * @gfp_mask: GFP flags for the allocation
1924 * This function is similar to alloc_pages(), except that it allocates the
1925 * minimum number of pages to satisfy the request. alloc_pages() can only
1926 * allocate memory in power-of-two pages.
1928 * This function is also limited by MAX_ORDER.
1930 * Memory allocated by this function must be released by free_pages_exact().
1932 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
1934 unsigned int order
= get_order(size
);
1937 addr
= __get_free_pages(gfp_mask
, order
);
1939 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
1940 unsigned long used
= addr
+ PAGE_ALIGN(size
);
1942 split_page(virt_to_page(addr
), order
);
1943 while (used
< alloc_end
) {
1949 return (void *)addr
;
1951 EXPORT_SYMBOL(alloc_pages_exact
);
1954 * free_pages_exact - release memory allocated via alloc_pages_exact()
1955 * @virt: the value returned by alloc_pages_exact.
1956 * @size: size of allocation, same value as passed to alloc_pages_exact().
1958 * Release the memory allocated by a previous call to alloc_pages_exact.
1960 void free_pages_exact(void *virt
, size_t size
)
1962 unsigned long addr
= (unsigned long)virt
;
1963 unsigned long end
= addr
+ PAGE_ALIGN(size
);
1965 while (addr
< end
) {
1970 EXPORT_SYMBOL(free_pages_exact
);
1972 static unsigned int nr_free_zone_pages(int offset
)
1977 /* Just pick one node, since fallback list is circular */
1978 unsigned int sum
= 0;
1980 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
1982 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
1983 unsigned long size
= zone
->present_pages
;
1984 unsigned long high
= high_wmark_pages(zone
);
1993 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
1995 unsigned int nr_free_buffer_pages(void)
1997 return nr_free_zone_pages(gfp_zone(GFP_USER
));
1999 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
2002 * Amount of free RAM allocatable within all zones
2004 unsigned int nr_free_pagecache_pages(void)
2006 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
2009 static inline void show_node(struct zone
*zone
)
2012 printk("Node %d ", zone_to_nid(zone
));
2015 void si_meminfo(struct sysinfo
*val
)
2017 val
->totalram
= totalram_pages
;
2019 val
->freeram
= global_page_state(NR_FREE_PAGES
);
2020 val
->bufferram
= nr_blockdev_pages();
2021 val
->totalhigh
= totalhigh_pages
;
2022 val
->freehigh
= nr_free_highpages();
2023 val
->mem_unit
= PAGE_SIZE
;
2026 EXPORT_SYMBOL(si_meminfo
);
2029 void si_meminfo_node(struct sysinfo
*val
, int nid
)
2031 pg_data_t
*pgdat
= NODE_DATA(nid
);
2033 val
->totalram
= pgdat
->node_present_pages
;
2034 val
->freeram
= node_page_state(nid
, NR_FREE_PAGES
);
2035 #ifdef CONFIG_HIGHMEM
2036 val
->totalhigh
= pgdat
->node_zones
[ZONE_HIGHMEM
].present_pages
;
2037 val
->freehigh
= zone_page_state(&pgdat
->node_zones
[ZONE_HIGHMEM
],
2043 val
->mem_unit
= PAGE_SIZE
;
2047 #define K(x) ((x) << (PAGE_SHIFT-10))
2050 * Show free area list (used inside shift_scroll-lock stuff)
2051 * We also calculate the percentage fragmentation. We do this by counting the
2052 * memory on each free list with the exception of the first item on the list.
2054 void show_free_areas(void)
2059 for_each_populated_zone(zone
) {
2061 printk("%s per-cpu:\n", zone
->name
);
2063 for_each_online_cpu(cpu
) {
2064 struct per_cpu_pageset
*pageset
;
2066 pageset
= zone_pcp(zone
, cpu
);
2068 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2069 cpu
, pageset
->pcp
.high
,
2070 pageset
->pcp
.batch
, pageset
->pcp
.count
);
2074 printk("Active_anon:%lu active_file:%lu inactive_anon:%lu\n"
2075 " inactive_file:%lu"
2076 //TODO: check/adjust line lengths
2077 #ifdef CONFIG_UNEVICTABLE_LRU
2080 " dirty:%lu writeback:%lu unstable:%lu\n"
2081 " free:%lu slab:%lu mapped:%lu pagetables:%lu bounce:%lu\n",
2082 global_page_state(NR_ACTIVE_ANON
),
2083 global_page_state(NR_ACTIVE_FILE
),
2084 global_page_state(NR_INACTIVE_ANON
),
2085 global_page_state(NR_INACTIVE_FILE
),
2086 #ifdef CONFIG_UNEVICTABLE_LRU
2087 global_page_state(NR_UNEVICTABLE
),
2089 global_page_state(NR_FILE_DIRTY
),
2090 global_page_state(NR_WRITEBACK
),
2091 global_page_state(NR_UNSTABLE_NFS
),
2092 global_page_state(NR_FREE_PAGES
),
2093 global_page_state(NR_SLAB_RECLAIMABLE
) +
2094 global_page_state(NR_SLAB_UNRECLAIMABLE
),
2095 global_page_state(NR_FILE_MAPPED
),
2096 global_page_state(NR_PAGETABLE
),
2097 global_page_state(NR_BOUNCE
));
2099 for_each_populated_zone(zone
) {
2108 " active_anon:%lukB"
2109 " inactive_anon:%lukB"
2110 " active_file:%lukB"
2111 " inactive_file:%lukB"
2112 #ifdef CONFIG_UNEVICTABLE_LRU
2113 " unevictable:%lukB"
2116 " pages_scanned:%lu"
2117 " all_unreclaimable? %s"
2120 K(zone_page_state(zone
, NR_FREE_PAGES
)),
2121 K(min_wmark_pages(zone
)),
2122 K(low_wmark_pages(zone
)),
2123 K(high_wmark_pages(zone
)),
2124 K(zone_page_state(zone
, NR_ACTIVE_ANON
)),
2125 K(zone_page_state(zone
, NR_INACTIVE_ANON
)),
2126 K(zone_page_state(zone
, NR_ACTIVE_FILE
)),
2127 K(zone_page_state(zone
, NR_INACTIVE_FILE
)),
2128 #ifdef CONFIG_UNEVICTABLE_LRU
2129 K(zone_page_state(zone
, NR_UNEVICTABLE
)),
2131 K(zone
->present_pages
),
2132 zone
->pages_scanned
,
2133 (zone_is_all_unreclaimable(zone
) ? "yes" : "no")
2135 printk("lowmem_reserve[]:");
2136 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
2137 printk(" %lu", zone
->lowmem_reserve
[i
]);
2141 for_each_populated_zone(zone
) {
2142 unsigned long nr
[MAX_ORDER
], flags
, order
, total
= 0;
2145 printk("%s: ", zone
->name
);
2147 spin_lock_irqsave(&zone
->lock
, flags
);
2148 for (order
= 0; order
< MAX_ORDER
; order
++) {
2149 nr
[order
] = zone
->free_area
[order
].nr_free
;
2150 total
+= nr
[order
] << order
;
2152 spin_unlock_irqrestore(&zone
->lock
, flags
);
2153 for (order
= 0; order
< MAX_ORDER
; order
++)
2154 printk("%lu*%lukB ", nr
[order
], K(1UL) << order
);
2155 printk("= %lukB\n", K(total
));
2158 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES
));
2160 show_swap_cache_info();
2163 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
2165 zoneref
->zone
= zone
;
2166 zoneref
->zone_idx
= zone_idx(zone
);
2170 * Builds allocation fallback zone lists.
2172 * Add all populated zones of a node to the zonelist.
2174 static int build_zonelists_node(pg_data_t
*pgdat
, struct zonelist
*zonelist
,
2175 int nr_zones
, enum zone_type zone_type
)
2179 BUG_ON(zone_type
>= MAX_NR_ZONES
);
2184 zone
= pgdat
->node_zones
+ zone_type
;
2185 if (populated_zone(zone
)) {
2186 zoneref_set_zone(zone
,
2187 &zonelist
->_zonerefs
[nr_zones
++]);
2188 check_highest_zone(zone_type
);
2191 } while (zone_type
);
2198 * 0 = automatic detection of better ordering.
2199 * 1 = order by ([node] distance, -zonetype)
2200 * 2 = order by (-zonetype, [node] distance)
2202 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
2203 * the same zonelist. So only NUMA can configure this param.
2205 #define ZONELIST_ORDER_DEFAULT 0
2206 #define ZONELIST_ORDER_NODE 1
2207 #define ZONELIST_ORDER_ZONE 2
2209 /* zonelist order in the kernel.
2210 * set_zonelist_order() will set this to NODE or ZONE.
2212 static int current_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
2213 static char zonelist_order_name
[3][8] = {"Default", "Node", "Zone"};
2217 /* The value user specified ....changed by config */
2218 static int user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
2219 /* string for sysctl */
2220 #define NUMA_ZONELIST_ORDER_LEN 16
2221 char numa_zonelist_order
[16] = "default";
2224 * interface for configure zonelist ordering.
2225 * command line option "numa_zonelist_order"
2226 * = "[dD]efault - default, automatic configuration.
2227 * = "[nN]ode - order by node locality, then by zone within node
2228 * = "[zZ]one - order by zone, then by locality within zone
2231 static int __parse_numa_zonelist_order(char *s
)
2233 if (*s
== 'd' || *s
== 'D') {
2234 user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
2235 } else if (*s
== 'n' || *s
== 'N') {
2236 user_zonelist_order
= ZONELIST_ORDER_NODE
;
2237 } else if (*s
== 'z' || *s
== 'Z') {
2238 user_zonelist_order
= ZONELIST_ORDER_ZONE
;
2241 "Ignoring invalid numa_zonelist_order value: "
2248 static __init
int setup_numa_zonelist_order(char *s
)
2251 return __parse_numa_zonelist_order(s
);
2254 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
2257 * sysctl handler for numa_zonelist_order
2259 int numa_zonelist_order_handler(ctl_table
*table
, int write
,
2260 struct file
*file
, void __user
*buffer
, size_t *length
,
2263 char saved_string
[NUMA_ZONELIST_ORDER_LEN
];
2267 strncpy(saved_string
, (char*)table
->data
,
2268 NUMA_ZONELIST_ORDER_LEN
);
2269 ret
= proc_dostring(table
, write
, file
, buffer
, length
, ppos
);
2273 int oldval
= user_zonelist_order
;
2274 if (__parse_numa_zonelist_order((char*)table
->data
)) {
2276 * bogus value. restore saved string
2278 strncpy((char*)table
->data
, saved_string
,
2279 NUMA_ZONELIST_ORDER_LEN
);
2280 user_zonelist_order
= oldval
;
2281 } else if (oldval
!= user_zonelist_order
)
2282 build_all_zonelists();
2288 #define MAX_NODE_LOAD (nr_online_nodes)
2289 static int node_load
[MAX_NUMNODES
];
2292 * find_next_best_node - find the next node that should appear in a given node's fallback list
2293 * @node: node whose fallback list we're appending
2294 * @used_node_mask: nodemask_t of already used nodes
2296 * We use a number of factors to determine which is the next node that should
2297 * appear on a given node's fallback list. The node should not have appeared
2298 * already in @node's fallback list, and it should be the next closest node
2299 * according to the distance array (which contains arbitrary distance values
2300 * from each node to each node in the system), and should also prefer nodes
2301 * with no CPUs, since presumably they'll have very little allocation pressure
2302 * on them otherwise.
2303 * It returns -1 if no node is found.
2305 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
2308 int min_val
= INT_MAX
;
2310 const struct cpumask
*tmp
= cpumask_of_node(0);
2312 /* Use the local node if we haven't already */
2313 if (!node_isset(node
, *used_node_mask
)) {
2314 node_set(node
, *used_node_mask
);
2318 for_each_node_state(n
, N_HIGH_MEMORY
) {
2320 /* Don't want a node to appear more than once */
2321 if (node_isset(n
, *used_node_mask
))
2324 /* Use the distance array to find the distance */
2325 val
= node_distance(node
, n
);
2327 /* Penalize nodes under us ("prefer the next node") */
2330 /* Give preference to headless and unused nodes */
2331 tmp
= cpumask_of_node(n
);
2332 if (!cpumask_empty(tmp
))
2333 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
2335 /* Slight preference for less loaded node */
2336 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
2337 val
+= node_load
[n
];
2339 if (val
< min_val
) {
2346 node_set(best_node
, *used_node_mask
);
2353 * Build zonelists ordered by node and zones within node.
2354 * This results in maximum locality--normal zone overflows into local
2355 * DMA zone, if any--but risks exhausting DMA zone.
2357 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int node
)
2360 struct zonelist
*zonelist
;
2362 zonelist
= &pgdat
->node_zonelists
[0];
2363 for (j
= 0; zonelist
->_zonerefs
[j
].zone
!= NULL
; j
++)
2365 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
2367 zonelist
->_zonerefs
[j
].zone
= NULL
;
2368 zonelist
->_zonerefs
[j
].zone_idx
= 0;
2372 * Build gfp_thisnode zonelists
2374 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
2377 struct zonelist
*zonelist
;
2379 zonelist
= &pgdat
->node_zonelists
[1];
2380 j
= build_zonelists_node(pgdat
, zonelist
, 0, MAX_NR_ZONES
- 1);
2381 zonelist
->_zonerefs
[j
].zone
= NULL
;
2382 zonelist
->_zonerefs
[j
].zone_idx
= 0;
2386 * Build zonelists ordered by zone and nodes within zones.
2387 * This results in conserving DMA zone[s] until all Normal memory is
2388 * exhausted, but results in overflowing to remote node while memory
2389 * may still exist in local DMA zone.
2391 static int node_order
[MAX_NUMNODES
];
2393 static void build_zonelists_in_zone_order(pg_data_t
*pgdat
, int nr_nodes
)
2396 int zone_type
; /* needs to be signed */
2398 struct zonelist
*zonelist
;
2400 zonelist
= &pgdat
->node_zonelists
[0];
2402 for (zone_type
= MAX_NR_ZONES
- 1; zone_type
>= 0; zone_type
--) {
2403 for (j
= 0; j
< nr_nodes
; j
++) {
2404 node
= node_order
[j
];
2405 z
= &NODE_DATA(node
)->node_zones
[zone_type
];
2406 if (populated_zone(z
)) {
2408 &zonelist
->_zonerefs
[pos
++]);
2409 check_highest_zone(zone_type
);
2413 zonelist
->_zonerefs
[pos
].zone
= NULL
;
2414 zonelist
->_zonerefs
[pos
].zone_idx
= 0;
2417 static int default_zonelist_order(void)
2420 unsigned long low_kmem_size
,total_size
;
2424 * ZONE_DMA and ZONE_DMA32 can be very small area in the sytem.
2425 * If they are really small and used heavily, the system can fall
2426 * into OOM very easily.
2427 * This function detect ZONE_DMA/DMA32 size and confgigures zone order.
2429 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
2432 for_each_online_node(nid
) {
2433 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
2434 z
= &NODE_DATA(nid
)->node_zones
[zone_type
];
2435 if (populated_zone(z
)) {
2436 if (zone_type
< ZONE_NORMAL
)
2437 low_kmem_size
+= z
->present_pages
;
2438 total_size
+= z
->present_pages
;
2442 if (!low_kmem_size
|| /* there are no DMA area. */
2443 low_kmem_size
> total_size
/2) /* DMA/DMA32 is big. */
2444 return ZONELIST_ORDER_NODE
;
2446 * look into each node's config.
2447 * If there is a node whose DMA/DMA32 memory is very big area on
2448 * local memory, NODE_ORDER may be suitable.
2450 average_size
= total_size
/
2451 (nodes_weight(node_states
[N_HIGH_MEMORY
]) + 1);
2452 for_each_online_node(nid
) {
2455 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
2456 z
= &NODE_DATA(nid
)->node_zones
[zone_type
];
2457 if (populated_zone(z
)) {
2458 if (zone_type
< ZONE_NORMAL
)
2459 low_kmem_size
+= z
->present_pages
;
2460 total_size
+= z
->present_pages
;
2463 if (low_kmem_size
&&
2464 total_size
> average_size
&& /* ignore small node */
2465 low_kmem_size
> total_size
* 70/100)
2466 return ZONELIST_ORDER_NODE
;
2468 return ZONELIST_ORDER_ZONE
;
2471 static void set_zonelist_order(void)
2473 if (user_zonelist_order
== ZONELIST_ORDER_DEFAULT
)
2474 current_zonelist_order
= default_zonelist_order();
2476 current_zonelist_order
= user_zonelist_order
;
2479 static void build_zonelists(pg_data_t
*pgdat
)
2483 nodemask_t used_mask
;
2484 int local_node
, prev_node
;
2485 struct zonelist
*zonelist
;
2486 int order
= current_zonelist_order
;
2488 /* initialize zonelists */
2489 for (i
= 0; i
< MAX_ZONELISTS
; i
++) {
2490 zonelist
= pgdat
->node_zonelists
+ i
;
2491 zonelist
->_zonerefs
[0].zone
= NULL
;
2492 zonelist
->_zonerefs
[0].zone_idx
= 0;
2495 /* NUMA-aware ordering of nodes */
2496 local_node
= pgdat
->node_id
;
2497 load
= nr_online_nodes
;
2498 prev_node
= local_node
;
2499 nodes_clear(used_mask
);
2501 memset(node_load
, 0, sizeof(node_load
));
2502 memset(node_order
, 0, sizeof(node_order
));
2505 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
2506 int distance
= node_distance(local_node
, node
);
2509 * If another node is sufficiently far away then it is better
2510 * to reclaim pages in a zone before going off node.
2512 if (distance
> RECLAIM_DISTANCE
)
2513 zone_reclaim_mode
= 1;
2516 * We don't want to pressure a particular node.
2517 * So adding penalty to the first node in same
2518 * distance group to make it round-robin.
2520 if (distance
!= node_distance(local_node
, prev_node
))
2521 node_load
[node
] = load
;
2525 if (order
== ZONELIST_ORDER_NODE
)
2526 build_zonelists_in_node_order(pgdat
, node
);
2528 node_order
[j
++] = node
; /* remember order */
2531 if (order
== ZONELIST_ORDER_ZONE
) {
2532 /* calculate node order -- i.e., DMA last! */
2533 build_zonelists_in_zone_order(pgdat
, j
);
2536 build_thisnode_zonelists(pgdat
);
2539 /* Construct the zonelist performance cache - see further mmzone.h */
2540 static void build_zonelist_cache(pg_data_t
*pgdat
)
2542 struct zonelist
*zonelist
;
2543 struct zonelist_cache
*zlc
;
2546 zonelist
= &pgdat
->node_zonelists
[0];
2547 zonelist
->zlcache_ptr
= zlc
= &zonelist
->zlcache
;
2548 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
2549 for (z
= zonelist
->_zonerefs
; z
->zone
; z
++)
2550 zlc
->z_to_n
[z
- zonelist
->_zonerefs
] = zonelist_node_idx(z
);
2554 #else /* CONFIG_NUMA */
2556 static void set_zonelist_order(void)
2558 current_zonelist_order
= ZONELIST_ORDER_ZONE
;
2561 static void build_zonelists(pg_data_t
*pgdat
)
2563 int node
, local_node
;
2565 struct zonelist
*zonelist
;
2567 local_node
= pgdat
->node_id
;
2569 zonelist
= &pgdat
->node_zonelists
[0];
2570 j
= build_zonelists_node(pgdat
, zonelist
, 0, MAX_NR_ZONES
- 1);
2573 * Now we build the zonelist so that it contains the zones
2574 * of all the other nodes.
2575 * We don't want to pressure a particular node, so when
2576 * building the zones for node N, we make sure that the
2577 * zones coming right after the local ones are those from
2578 * node N+1 (modulo N)
2580 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
2581 if (!node_online(node
))
2583 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
2586 for (node
= 0; node
< local_node
; node
++) {
2587 if (!node_online(node
))
2589 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
2593 zonelist
->_zonerefs
[j
].zone
= NULL
;
2594 zonelist
->_zonerefs
[j
].zone_idx
= 0;
2597 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
2598 static void build_zonelist_cache(pg_data_t
*pgdat
)
2600 pgdat
->node_zonelists
[0].zlcache_ptr
= NULL
;
2603 #endif /* CONFIG_NUMA */
2605 /* return values int ....just for stop_machine() */
2606 static int __build_all_zonelists(void *dummy
)
2610 for_each_online_node(nid
) {
2611 pg_data_t
*pgdat
= NODE_DATA(nid
);
2613 build_zonelists(pgdat
);
2614 build_zonelist_cache(pgdat
);
2619 void build_all_zonelists(void)
2621 set_zonelist_order();
2623 if (system_state
== SYSTEM_BOOTING
) {
2624 __build_all_zonelists(NULL
);
2625 mminit_verify_zonelist();
2626 cpuset_init_current_mems_allowed();
2628 /* we have to stop all cpus to guarantee there is no user
2630 stop_machine(__build_all_zonelists
, NULL
, NULL
);
2631 /* cpuset refresh routine should be here */
2633 vm_total_pages
= nr_free_pagecache_pages();
2635 * Disable grouping by mobility if the number of pages in the
2636 * system is too low to allow the mechanism to work. It would be
2637 * more accurate, but expensive to check per-zone. This check is
2638 * made on memory-hotadd so a system can start with mobility
2639 * disabled and enable it later
2641 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
2642 page_group_by_mobility_disabled
= 1;
2644 page_group_by_mobility_disabled
= 0;
2646 printk("Built %i zonelists in %s order, mobility grouping %s. "
2647 "Total pages: %ld\n",
2649 zonelist_order_name
[current_zonelist_order
],
2650 page_group_by_mobility_disabled
? "off" : "on",
2653 printk("Policy zone: %s\n", zone_names
[policy_zone
]);
2658 * Helper functions to size the waitqueue hash table.
2659 * Essentially these want to choose hash table sizes sufficiently
2660 * large so that collisions trying to wait on pages are rare.
2661 * But in fact, the number of active page waitqueues on typical
2662 * systems is ridiculously low, less than 200. So this is even
2663 * conservative, even though it seems large.
2665 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
2666 * waitqueues, i.e. the size of the waitq table given the number of pages.
2668 #define PAGES_PER_WAITQUEUE 256
2670 #ifndef CONFIG_MEMORY_HOTPLUG
2671 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
2673 unsigned long size
= 1;
2675 pages
/= PAGES_PER_WAITQUEUE
;
2677 while (size
< pages
)
2681 * Once we have dozens or even hundreds of threads sleeping
2682 * on IO we've got bigger problems than wait queue collision.
2683 * Limit the size of the wait table to a reasonable size.
2685 size
= min(size
, 4096UL);
2687 return max(size
, 4UL);
2691 * A zone's size might be changed by hot-add, so it is not possible to determine
2692 * a suitable size for its wait_table. So we use the maximum size now.
2694 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
2696 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
2697 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
2698 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
2700 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
2701 * or more by the traditional way. (See above). It equals:
2703 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
2704 * ia64(16K page size) : = ( 8G + 4M)byte.
2705 * powerpc (64K page size) : = (32G +16M)byte.
2707 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
2714 * This is an integer logarithm so that shifts can be used later
2715 * to extract the more random high bits from the multiplicative
2716 * hash function before the remainder is taken.
2718 static inline unsigned long wait_table_bits(unsigned long size
)
2723 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
2726 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
2727 * of blocks reserved is based on min_wmark_pages(zone). The memory within
2728 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
2729 * higher will lead to a bigger reserve which will get freed as contiguous
2730 * blocks as reclaim kicks in
2732 static void setup_zone_migrate_reserve(struct zone
*zone
)
2734 unsigned long start_pfn
, pfn
, end_pfn
;
2736 unsigned long reserve
, block_migratetype
;
2738 /* Get the start pfn, end pfn and the number of blocks to reserve */
2739 start_pfn
= zone
->zone_start_pfn
;
2740 end_pfn
= start_pfn
+ zone
->spanned_pages
;
2741 reserve
= roundup(min_wmark_pages(zone
), pageblock_nr_pages
) >>
2744 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= pageblock_nr_pages
) {
2745 if (!pfn_valid(pfn
))
2747 page
= pfn_to_page(pfn
);
2749 /* Watch out for overlapping nodes */
2750 if (page_to_nid(page
) != zone_to_nid(zone
))
2753 /* Blocks with reserved pages will never free, skip them. */
2754 if (PageReserved(page
))
2757 block_migratetype
= get_pageblock_migratetype(page
);
2759 /* If this block is reserved, account for it */
2760 if (reserve
> 0 && block_migratetype
== MIGRATE_RESERVE
) {
2765 /* Suitable for reserving if this block is movable */
2766 if (reserve
> 0 && block_migratetype
== MIGRATE_MOVABLE
) {
2767 set_pageblock_migratetype(page
, MIGRATE_RESERVE
);
2768 move_freepages_block(zone
, page
, MIGRATE_RESERVE
);
2774 * If the reserve is met and this is a previous reserved block,
2777 if (block_migratetype
== MIGRATE_RESERVE
) {
2778 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
2779 move_freepages_block(zone
, page
, MIGRATE_MOVABLE
);
2785 * Initially all pages are reserved - free ones are freed
2786 * up by free_all_bootmem() once the early boot process is
2787 * done. Non-atomic initialization, single-pass.
2789 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
2790 unsigned long start_pfn
, enum memmap_context context
)
2793 unsigned long end_pfn
= start_pfn
+ size
;
2797 if (highest_memmap_pfn
< end_pfn
- 1)
2798 highest_memmap_pfn
= end_pfn
- 1;
2800 z
= &NODE_DATA(nid
)->node_zones
[zone
];
2801 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
2803 * There can be holes in boot-time mem_map[]s
2804 * handed to this function. They do not
2805 * exist on hotplugged memory.
2807 if (context
== MEMMAP_EARLY
) {
2808 if (!early_pfn_valid(pfn
))
2810 if (!early_pfn_in_nid(pfn
, nid
))
2813 page
= pfn_to_page(pfn
);
2814 set_page_links(page
, zone
, nid
, pfn
);
2815 mminit_verify_page_links(page
, zone
, nid
, pfn
);
2816 init_page_count(page
);
2817 reset_page_mapcount(page
);
2818 SetPageReserved(page
);
2820 * Mark the block movable so that blocks are reserved for
2821 * movable at startup. This will force kernel allocations
2822 * to reserve their blocks rather than leaking throughout
2823 * the address space during boot when many long-lived
2824 * kernel allocations are made. Later some blocks near
2825 * the start are marked MIGRATE_RESERVE by
2826 * setup_zone_migrate_reserve()
2828 * bitmap is created for zone's valid pfn range. but memmap
2829 * can be created for invalid pages (for alignment)
2830 * check here not to call set_pageblock_migratetype() against
2833 if ((z
->zone_start_pfn
<= pfn
)
2834 && (pfn
< z
->zone_start_pfn
+ z
->spanned_pages
)
2835 && !(pfn
& (pageblock_nr_pages
- 1)))
2836 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
2838 INIT_LIST_HEAD(&page
->lru
);
2839 #ifdef WANT_PAGE_VIRTUAL
2840 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
2841 if (!is_highmem_idx(zone
))
2842 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
2847 static void __meminit
zone_init_free_lists(struct zone
*zone
)
2850 for_each_migratetype_order(order
, t
) {
2851 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
2852 zone
->free_area
[order
].nr_free
= 0;
2856 #ifndef __HAVE_ARCH_MEMMAP_INIT
2857 #define memmap_init(size, nid, zone, start_pfn) \
2858 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
2861 static int zone_batchsize(struct zone
*zone
)
2867 * The per-cpu-pages pools are set to around 1000th of the
2868 * size of the zone. But no more than 1/2 of a meg.
2870 * OK, so we don't know how big the cache is. So guess.
2872 batch
= zone
->present_pages
/ 1024;
2873 if (batch
* PAGE_SIZE
> 512 * 1024)
2874 batch
= (512 * 1024) / PAGE_SIZE
;
2875 batch
/= 4; /* We effectively *= 4 below */
2880 * Clamp the batch to a 2^n - 1 value. Having a power
2881 * of 2 value was found to be more likely to have
2882 * suboptimal cache aliasing properties in some cases.
2884 * For example if 2 tasks are alternately allocating
2885 * batches of pages, one task can end up with a lot
2886 * of pages of one half of the possible page colors
2887 * and the other with pages of the other colors.
2889 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
2894 /* The deferral and batching of frees should be suppressed under NOMMU
2897 * The problem is that NOMMU needs to be able to allocate large chunks
2898 * of contiguous memory as there's no hardware page translation to
2899 * assemble apparent contiguous memory from discontiguous pages.
2901 * Queueing large contiguous runs of pages for batching, however,
2902 * causes the pages to actually be freed in smaller chunks. As there
2903 * can be a significant delay between the individual batches being
2904 * recycled, this leads to the once large chunks of space being
2905 * fragmented and becoming unavailable for high-order allocations.
2911 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
2913 struct per_cpu_pages
*pcp
;
2915 memset(p
, 0, sizeof(*p
));
2919 pcp
->high
= 6 * batch
;
2920 pcp
->batch
= max(1UL, 1 * batch
);
2921 INIT_LIST_HEAD(&pcp
->list
);
2925 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
2926 * to the value high for the pageset p.
2929 static void setup_pagelist_highmark(struct per_cpu_pageset
*p
,
2932 struct per_cpu_pages
*pcp
;
2936 pcp
->batch
= max(1UL, high
/4);
2937 if ((high
/4) > (PAGE_SHIFT
* 8))
2938 pcp
->batch
= PAGE_SHIFT
* 8;
2944 * Boot pageset table. One per cpu which is going to be used for all
2945 * zones and all nodes. The parameters will be set in such a way
2946 * that an item put on a list will immediately be handed over to
2947 * the buddy list. This is safe since pageset manipulation is done
2948 * with interrupts disabled.
2950 * Some NUMA counter updates may also be caught by the boot pagesets.
2952 * The boot_pagesets must be kept even after bootup is complete for
2953 * unused processors and/or zones. They do play a role for bootstrapping
2954 * hotplugged processors.
2956 * zoneinfo_show() and maybe other functions do
2957 * not check if the processor is online before following the pageset pointer.
2958 * Other parts of the kernel may not check if the zone is available.
2960 static struct per_cpu_pageset boot_pageset
[NR_CPUS
];
2963 * Dynamically allocate memory for the
2964 * per cpu pageset array in struct zone.
2966 static int __cpuinit
process_zones(int cpu
)
2968 struct zone
*zone
, *dzone
;
2969 int node
= cpu_to_node(cpu
);
2971 node_set_state(node
, N_CPU
); /* this node has a cpu */
2973 for_each_populated_zone(zone
) {
2974 zone_pcp(zone
, cpu
) = kmalloc_node(sizeof(struct per_cpu_pageset
),
2976 if (!zone_pcp(zone
, cpu
))
2979 setup_pageset(zone_pcp(zone
, cpu
), zone_batchsize(zone
));
2981 if (percpu_pagelist_fraction
)
2982 setup_pagelist_highmark(zone_pcp(zone
, cpu
),
2983 (zone
->present_pages
/ percpu_pagelist_fraction
));
2988 for_each_zone(dzone
) {
2989 if (!populated_zone(dzone
))
2993 kfree(zone_pcp(dzone
, cpu
));
2994 zone_pcp(dzone
, cpu
) = NULL
;
2999 static inline void free_zone_pagesets(int cpu
)
3003 for_each_zone(zone
) {
3004 struct per_cpu_pageset
*pset
= zone_pcp(zone
, cpu
);
3006 /* Free per_cpu_pageset if it is slab allocated */
3007 if (pset
!= &boot_pageset
[cpu
])
3009 zone_pcp(zone
, cpu
) = NULL
;
3013 static int __cpuinit
pageset_cpuup_callback(struct notifier_block
*nfb
,
3014 unsigned long action
,
3017 int cpu
= (long)hcpu
;
3018 int ret
= NOTIFY_OK
;
3021 case CPU_UP_PREPARE
:
3022 case CPU_UP_PREPARE_FROZEN
:
3023 if (process_zones(cpu
))
3026 case CPU_UP_CANCELED
:
3027 case CPU_UP_CANCELED_FROZEN
:
3029 case CPU_DEAD_FROZEN
:
3030 free_zone_pagesets(cpu
);
3038 static struct notifier_block __cpuinitdata pageset_notifier
=
3039 { &pageset_cpuup_callback
, NULL
, 0 };
3041 void __init
setup_per_cpu_pageset(void)
3045 /* Initialize per_cpu_pageset for cpu 0.
3046 * A cpuup callback will do this for every cpu
3047 * as it comes online
3049 err
= process_zones(smp_processor_id());
3051 register_cpu_notifier(&pageset_notifier
);
3056 static noinline __init_refok
3057 int zone_wait_table_init(struct zone
*zone
, unsigned long zone_size_pages
)
3060 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
3064 * The per-page waitqueue mechanism uses hashed waitqueues
3067 zone
->wait_table_hash_nr_entries
=
3068 wait_table_hash_nr_entries(zone_size_pages
);
3069 zone
->wait_table_bits
=
3070 wait_table_bits(zone
->wait_table_hash_nr_entries
);
3071 alloc_size
= zone
->wait_table_hash_nr_entries
3072 * sizeof(wait_queue_head_t
);
3074 if (!slab_is_available()) {
3075 zone
->wait_table
= (wait_queue_head_t
*)
3076 alloc_bootmem_node(pgdat
, alloc_size
);
3079 * This case means that a zone whose size was 0 gets new memory
3080 * via memory hot-add.
3081 * But it may be the case that a new node was hot-added. In
3082 * this case vmalloc() will not be able to use this new node's
3083 * memory - this wait_table must be initialized to use this new
3084 * node itself as well.
3085 * To use this new node's memory, further consideration will be
3088 zone
->wait_table
= vmalloc(alloc_size
);
3090 if (!zone
->wait_table
)
3093 for(i
= 0; i
< zone
->wait_table_hash_nr_entries
; ++i
)
3094 init_waitqueue_head(zone
->wait_table
+ i
);
3099 static __meminit
void zone_pcp_init(struct zone
*zone
)
3102 unsigned long batch
= zone_batchsize(zone
);
3104 for (cpu
= 0; cpu
< NR_CPUS
; cpu
++) {
3106 /* Early boot. Slab allocator not functional yet */
3107 zone_pcp(zone
, cpu
) = &boot_pageset
[cpu
];
3108 setup_pageset(&boot_pageset
[cpu
],0);
3110 setup_pageset(zone_pcp(zone
,cpu
), batch
);
3113 if (zone
->present_pages
)
3114 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%lu\n",
3115 zone
->name
, zone
->present_pages
, batch
);
3118 __meminit
int init_currently_empty_zone(struct zone
*zone
,
3119 unsigned long zone_start_pfn
,
3121 enum memmap_context context
)
3123 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
3125 ret
= zone_wait_table_init(zone
, size
);
3128 pgdat
->nr_zones
= zone_idx(zone
) + 1;
3130 zone
->zone_start_pfn
= zone_start_pfn
;
3132 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
3133 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
3135 (unsigned long)zone_idx(zone
),
3136 zone_start_pfn
, (zone_start_pfn
+ size
));
3138 zone_init_free_lists(zone
);
3143 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3145 * Basic iterator support. Return the first range of PFNs for a node
3146 * Note: nid == MAX_NUMNODES returns first region regardless of node
3148 static int __meminit
first_active_region_index_in_nid(int nid
)
3152 for (i
= 0; i
< nr_nodemap_entries
; i
++)
3153 if (nid
== MAX_NUMNODES
|| early_node_map
[i
].nid
== nid
)
3160 * Basic iterator support. Return the next active range of PFNs for a node
3161 * Note: nid == MAX_NUMNODES returns next region regardless of node
3163 static int __meminit
next_active_region_index_in_nid(int index
, int nid
)
3165 for (index
= index
+ 1; index
< nr_nodemap_entries
; index
++)
3166 if (nid
== MAX_NUMNODES
|| early_node_map
[index
].nid
== nid
)
3172 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
3174 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
3175 * Architectures may implement their own version but if add_active_range()
3176 * was used and there are no special requirements, this is a convenient
3179 int __meminit
__early_pfn_to_nid(unsigned long pfn
)
3183 for (i
= 0; i
< nr_nodemap_entries
; i
++) {
3184 unsigned long start_pfn
= early_node_map
[i
].start_pfn
;
3185 unsigned long end_pfn
= early_node_map
[i
].end_pfn
;
3187 if (start_pfn
<= pfn
&& pfn
< end_pfn
)
3188 return early_node_map
[i
].nid
;
3190 /* This is a memory hole */
3193 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
3195 int __meminit
early_pfn_to_nid(unsigned long pfn
)
3199 nid
= __early_pfn_to_nid(pfn
);
3202 /* just returns 0 */
3206 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
3207 bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
3211 nid
= __early_pfn_to_nid(pfn
);
3212 if (nid
>= 0 && nid
!= node
)
3218 /* Basic iterator support to walk early_node_map[] */
3219 #define for_each_active_range_index_in_nid(i, nid) \
3220 for (i = first_active_region_index_in_nid(nid); i != -1; \
3221 i = next_active_region_index_in_nid(i, nid))
3224 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
3225 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
3226 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
3228 * If an architecture guarantees that all ranges registered with
3229 * add_active_ranges() contain no holes and may be freed, this
3230 * this function may be used instead of calling free_bootmem() manually.
3232 void __init
free_bootmem_with_active_regions(int nid
,
3233 unsigned long max_low_pfn
)
3237 for_each_active_range_index_in_nid(i
, nid
) {
3238 unsigned long size_pages
= 0;
3239 unsigned long end_pfn
= early_node_map
[i
].end_pfn
;
3241 if (early_node_map
[i
].start_pfn
>= max_low_pfn
)
3244 if (end_pfn
> max_low_pfn
)
3245 end_pfn
= max_low_pfn
;
3247 size_pages
= end_pfn
- early_node_map
[i
].start_pfn
;
3248 free_bootmem_node(NODE_DATA(early_node_map
[i
].nid
),
3249 PFN_PHYS(early_node_map
[i
].start_pfn
),
3250 size_pages
<< PAGE_SHIFT
);
3254 void __init
work_with_active_regions(int nid
, work_fn_t work_fn
, void *data
)
3259 for_each_active_range_index_in_nid(i
, nid
) {
3260 ret
= work_fn(early_node_map
[i
].start_pfn
,
3261 early_node_map
[i
].end_pfn
, data
);
3267 * sparse_memory_present_with_active_regions - Call memory_present for each active range
3268 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
3270 * If an architecture guarantees that all ranges registered with
3271 * add_active_ranges() contain no holes and may be freed, this
3272 * function may be used instead of calling memory_present() manually.
3274 void __init
sparse_memory_present_with_active_regions(int nid
)
3278 for_each_active_range_index_in_nid(i
, nid
)
3279 memory_present(early_node_map
[i
].nid
,
3280 early_node_map
[i
].start_pfn
,
3281 early_node_map
[i
].end_pfn
);
3285 * get_pfn_range_for_nid - Return the start and end page frames for a node
3286 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
3287 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
3288 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
3290 * It returns the start and end page frame of a node based on information
3291 * provided by an arch calling add_active_range(). If called for a node
3292 * with no available memory, a warning is printed and the start and end
3295 void __meminit
get_pfn_range_for_nid(unsigned int nid
,
3296 unsigned long *start_pfn
, unsigned long *end_pfn
)
3302 for_each_active_range_index_in_nid(i
, nid
) {
3303 *start_pfn
= min(*start_pfn
, early_node_map
[i
].start_pfn
);
3304 *end_pfn
= max(*end_pfn
, early_node_map
[i
].end_pfn
);
3307 if (*start_pfn
== -1UL)
3312 * This finds a zone that can be used for ZONE_MOVABLE pages. The
3313 * assumption is made that zones within a node are ordered in monotonic
3314 * increasing memory addresses so that the "highest" populated zone is used
3316 static void __init
find_usable_zone_for_movable(void)
3319 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
3320 if (zone_index
== ZONE_MOVABLE
)
3323 if (arch_zone_highest_possible_pfn
[zone_index
] >
3324 arch_zone_lowest_possible_pfn
[zone_index
])
3328 VM_BUG_ON(zone_index
== -1);
3329 movable_zone
= zone_index
;
3333 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
3334 * because it is sized independant of architecture. Unlike the other zones,
3335 * the starting point for ZONE_MOVABLE is not fixed. It may be different
3336 * in each node depending on the size of each node and how evenly kernelcore
3337 * is distributed. This helper function adjusts the zone ranges
3338 * provided by the architecture for a given node by using the end of the
3339 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
3340 * zones within a node are in order of monotonic increases memory addresses
3342 static void __meminit
adjust_zone_range_for_zone_movable(int nid
,
3343 unsigned long zone_type
,
3344 unsigned long node_start_pfn
,
3345 unsigned long node_end_pfn
,
3346 unsigned long *zone_start_pfn
,
3347 unsigned long *zone_end_pfn
)
3349 /* Only adjust if ZONE_MOVABLE is on this node */
3350 if (zone_movable_pfn
[nid
]) {
3351 /* Size ZONE_MOVABLE */
3352 if (zone_type
== ZONE_MOVABLE
) {
3353 *zone_start_pfn
= zone_movable_pfn
[nid
];
3354 *zone_end_pfn
= min(node_end_pfn
,
3355 arch_zone_highest_possible_pfn
[movable_zone
]);
3357 /* Adjust for ZONE_MOVABLE starting within this range */
3358 } else if (*zone_start_pfn
< zone_movable_pfn
[nid
] &&
3359 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
3360 *zone_end_pfn
= zone_movable_pfn
[nid
];
3362 /* Check if this whole range is within ZONE_MOVABLE */
3363 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
3364 *zone_start_pfn
= *zone_end_pfn
;
3369 * Return the number of pages a zone spans in a node, including holes
3370 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
3372 static unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
3373 unsigned long zone_type
,
3374 unsigned long *ignored
)
3376 unsigned long node_start_pfn
, node_end_pfn
;
3377 unsigned long zone_start_pfn
, zone_end_pfn
;
3379 /* Get the start and end of the node and zone */
3380 get_pfn_range_for_nid(nid
, &node_start_pfn
, &node_end_pfn
);
3381 zone_start_pfn
= arch_zone_lowest_possible_pfn
[zone_type
];
3382 zone_end_pfn
= arch_zone_highest_possible_pfn
[zone_type
];
3383 adjust_zone_range_for_zone_movable(nid
, zone_type
,
3384 node_start_pfn
, node_end_pfn
,
3385 &zone_start_pfn
, &zone_end_pfn
);
3387 /* Check that this node has pages within the zone's required range */
3388 if (zone_end_pfn
< node_start_pfn
|| zone_start_pfn
> node_end_pfn
)
3391 /* Move the zone boundaries inside the node if necessary */
3392 zone_end_pfn
= min(zone_end_pfn
, node_end_pfn
);
3393 zone_start_pfn
= max(zone_start_pfn
, node_start_pfn
);
3395 /* Return the spanned pages */
3396 return zone_end_pfn
- zone_start_pfn
;
3400 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
3401 * then all holes in the requested range will be accounted for.
3403 static unsigned long __meminit
__absent_pages_in_range(int nid
,
3404 unsigned long range_start_pfn
,
3405 unsigned long range_end_pfn
)
3408 unsigned long prev_end_pfn
= 0, hole_pages
= 0;
3409 unsigned long start_pfn
;
3411 /* Find the end_pfn of the first active range of pfns in the node */
3412 i
= first_active_region_index_in_nid(nid
);
3416 prev_end_pfn
= min(early_node_map
[i
].start_pfn
, range_end_pfn
);
3418 /* Account for ranges before physical memory on this node */
3419 if (early_node_map
[i
].start_pfn
> range_start_pfn
)
3420 hole_pages
= prev_end_pfn
- range_start_pfn
;
3422 /* Find all holes for the zone within the node */
3423 for (; i
!= -1; i
= next_active_region_index_in_nid(i
, nid
)) {
3425 /* No need to continue if prev_end_pfn is outside the zone */
3426 if (prev_end_pfn
>= range_end_pfn
)
3429 /* Make sure the end of the zone is not within the hole */
3430 start_pfn
= min(early_node_map
[i
].start_pfn
, range_end_pfn
);
3431 prev_end_pfn
= max(prev_end_pfn
, range_start_pfn
);
3433 /* Update the hole size cound and move on */
3434 if (start_pfn
> range_start_pfn
) {
3435 BUG_ON(prev_end_pfn
> start_pfn
);
3436 hole_pages
+= start_pfn
- prev_end_pfn
;
3438 prev_end_pfn
= early_node_map
[i
].end_pfn
;
3441 /* Account for ranges past physical memory on this node */
3442 if (range_end_pfn
> prev_end_pfn
)
3443 hole_pages
+= range_end_pfn
-
3444 max(range_start_pfn
, prev_end_pfn
);
3450 * absent_pages_in_range - Return number of page frames in holes within a range
3451 * @start_pfn: The start PFN to start searching for holes
3452 * @end_pfn: The end PFN to stop searching for holes
3454 * It returns the number of pages frames in memory holes within a range.
3456 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
3457 unsigned long end_pfn
)
3459 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
3462 /* Return the number of page frames in holes in a zone on a node */
3463 static unsigned long __meminit
zone_absent_pages_in_node(int nid
,
3464 unsigned long zone_type
,
3465 unsigned long *ignored
)
3467 unsigned long node_start_pfn
, node_end_pfn
;
3468 unsigned long zone_start_pfn
, zone_end_pfn
;
3470 get_pfn_range_for_nid(nid
, &node_start_pfn
, &node_end_pfn
);
3471 zone_start_pfn
= max(arch_zone_lowest_possible_pfn
[zone_type
],
3473 zone_end_pfn
= min(arch_zone_highest_possible_pfn
[zone_type
],
3476 adjust_zone_range_for_zone_movable(nid
, zone_type
,
3477 node_start_pfn
, node_end_pfn
,
3478 &zone_start_pfn
, &zone_end_pfn
);
3479 return __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
3483 static inline unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
3484 unsigned long zone_type
,
3485 unsigned long *zones_size
)
3487 return zones_size
[zone_type
];
3490 static inline unsigned long __meminit
zone_absent_pages_in_node(int nid
,
3491 unsigned long zone_type
,
3492 unsigned long *zholes_size
)
3497 return zholes_size
[zone_type
];
3502 static void __meminit
calculate_node_totalpages(struct pglist_data
*pgdat
,
3503 unsigned long *zones_size
, unsigned long *zholes_size
)
3505 unsigned long realtotalpages
, totalpages
= 0;
3508 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
3509 totalpages
+= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
3511 pgdat
->node_spanned_pages
= totalpages
;
3513 realtotalpages
= totalpages
;
3514 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
3516 zone_absent_pages_in_node(pgdat
->node_id
, i
,
3518 pgdat
->node_present_pages
= realtotalpages
;
3519 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
3523 #ifndef CONFIG_SPARSEMEM
3525 * Calculate the size of the zone->blockflags rounded to an unsigned long
3526 * Start by making sure zonesize is a multiple of pageblock_order by rounding
3527 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
3528 * round what is now in bits to nearest long in bits, then return it in
3531 static unsigned long __init
usemap_size(unsigned long zonesize
)
3533 unsigned long usemapsize
;
3535 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
3536 usemapsize
= usemapsize
>> pageblock_order
;
3537 usemapsize
*= NR_PAGEBLOCK_BITS
;
3538 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
3540 return usemapsize
/ 8;
3543 static void __init
setup_usemap(struct pglist_data
*pgdat
,
3544 struct zone
*zone
, unsigned long zonesize
)
3546 unsigned long usemapsize
= usemap_size(zonesize
);
3547 zone
->pageblock_flags
= NULL
;
3549 zone
->pageblock_flags
= alloc_bootmem_node(pgdat
, usemapsize
);
3552 static void inline setup_usemap(struct pglist_data
*pgdat
,
3553 struct zone
*zone
, unsigned long zonesize
) {}
3554 #endif /* CONFIG_SPARSEMEM */
3556 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
3558 /* Return a sensible default order for the pageblock size. */
3559 static inline int pageblock_default_order(void)
3561 if (HPAGE_SHIFT
> PAGE_SHIFT
)
3562 return HUGETLB_PAGE_ORDER
;
3567 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
3568 static inline void __init
set_pageblock_order(unsigned int order
)
3570 /* Check that pageblock_nr_pages has not already been setup */
3571 if (pageblock_order
)
3575 * Assume the largest contiguous order of interest is a huge page.
3576 * This value may be variable depending on boot parameters on IA64
3578 pageblock_order
= order
;
3580 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
3583 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
3584 * and pageblock_default_order() are unused as pageblock_order is set
3585 * at compile-time. See include/linux/pageblock-flags.h for the values of
3586 * pageblock_order based on the kernel config
3588 static inline int pageblock_default_order(unsigned int order
)
3592 #define set_pageblock_order(x) do {} while (0)
3594 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
3597 * Set up the zone data structures:
3598 * - mark all pages reserved
3599 * - mark all memory queues empty
3600 * - clear the memory bitmaps
3602 static void __paginginit
free_area_init_core(struct pglist_data
*pgdat
,
3603 unsigned long *zones_size
, unsigned long *zholes_size
)
3606 int nid
= pgdat
->node_id
;
3607 unsigned long zone_start_pfn
= pgdat
->node_start_pfn
;
3610 pgdat_resize_init(pgdat
);
3611 pgdat
->nr_zones
= 0;
3612 init_waitqueue_head(&pgdat
->kswapd_wait
);
3613 pgdat
->kswapd_max_order
= 0;
3614 pgdat_page_cgroup_init(pgdat
);
3616 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
3617 struct zone
*zone
= pgdat
->node_zones
+ j
;
3618 unsigned long size
, realsize
, memmap_pages
;
3621 size
= zone_spanned_pages_in_node(nid
, j
, zones_size
);
3622 realsize
= size
- zone_absent_pages_in_node(nid
, j
,
3626 * Adjust realsize so that it accounts for how much memory
3627 * is used by this zone for memmap. This affects the watermark
3628 * and per-cpu initialisations
3631 PAGE_ALIGN(size
* sizeof(struct page
)) >> PAGE_SHIFT
;
3632 if (realsize
>= memmap_pages
) {
3633 realsize
-= memmap_pages
;
3636 " %s zone: %lu pages used for memmap\n",
3637 zone_names
[j
], memmap_pages
);
3640 " %s zone: %lu pages exceeds realsize %lu\n",
3641 zone_names
[j
], memmap_pages
, realsize
);
3643 /* Account for reserved pages */
3644 if (j
== 0 && realsize
> dma_reserve
) {
3645 realsize
-= dma_reserve
;
3646 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
3647 zone_names
[0], dma_reserve
);
3650 if (!is_highmem_idx(j
))
3651 nr_kernel_pages
+= realsize
;
3652 nr_all_pages
+= realsize
;
3654 zone
->spanned_pages
= size
;
3655 zone
->present_pages
= realsize
;
3658 zone
->min_unmapped_pages
= (realsize
*sysctl_min_unmapped_ratio
)
3660 zone
->min_slab_pages
= (realsize
* sysctl_min_slab_ratio
) / 100;
3662 zone
->name
= zone_names
[j
];
3663 spin_lock_init(&zone
->lock
);
3664 spin_lock_init(&zone
->lru_lock
);
3665 zone_seqlock_init(zone
);
3666 zone
->zone_pgdat
= pgdat
;
3668 zone
->prev_priority
= DEF_PRIORITY
;
3670 zone_pcp_init(zone
);
3672 INIT_LIST_HEAD(&zone
->lru
[l
].list
);
3673 zone
->lru
[l
].nr_saved_scan
= 0;
3675 zone
->reclaim_stat
.recent_rotated
[0] = 0;
3676 zone
->reclaim_stat
.recent_rotated
[1] = 0;
3677 zone
->reclaim_stat
.recent_scanned
[0] = 0;
3678 zone
->reclaim_stat
.recent_scanned
[1] = 0;
3679 zap_zone_vm_stats(zone
);
3684 set_pageblock_order(pageblock_default_order());
3685 setup_usemap(pgdat
, zone
, size
);
3686 ret
= init_currently_empty_zone(zone
, zone_start_pfn
,
3687 size
, MEMMAP_EARLY
);
3689 memmap_init(size
, nid
, j
, zone_start_pfn
);
3690 zone_start_pfn
+= size
;
3694 static void __init_refok
alloc_node_mem_map(struct pglist_data
*pgdat
)
3696 /* Skip empty nodes */
3697 if (!pgdat
->node_spanned_pages
)
3700 #ifdef CONFIG_FLAT_NODE_MEM_MAP
3701 /* ia64 gets its own node_mem_map, before this, without bootmem */
3702 if (!pgdat
->node_mem_map
) {
3703 unsigned long size
, start
, end
;
3707 * The zone's endpoints aren't required to be MAX_ORDER
3708 * aligned but the node_mem_map endpoints must be in order
3709 * for the buddy allocator to function correctly.
3711 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
3712 end
= pgdat
->node_start_pfn
+ pgdat
->node_spanned_pages
;
3713 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
3714 size
= (end
- start
) * sizeof(struct page
);
3715 map
= alloc_remap(pgdat
->node_id
, size
);
3717 map
= alloc_bootmem_node(pgdat
, size
);
3718 pgdat
->node_mem_map
= map
+ (pgdat
->node_start_pfn
- start
);
3720 #ifndef CONFIG_NEED_MULTIPLE_NODES
3722 * With no DISCONTIG, the global mem_map is just set as node 0's
3724 if (pgdat
== NODE_DATA(0)) {
3725 mem_map
= NODE_DATA(0)->node_mem_map
;
3726 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3727 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
3728 mem_map
-= (pgdat
->node_start_pfn
- ARCH_PFN_OFFSET
);
3729 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
3732 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
3735 void __paginginit
free_area_init_node(int nid
, unsigned long *zones_size
,
3736 unsigned long node_start_pfn
, unsigned long *zholes_size
)
3738 pg_data_t
*pgdat
= NODE_DATA(nid
);
3740 pgdat
->node_id
= nid
;
3741 pgdat
->node_start_pfn
= node_start_pfn
;
3742 calculate_node_totalpages(pgdat
, zones_size
, zholes_size
);
3744 alloc_node_mem_map(pgdat
);
3745 #ifdef CONFIG_FLAT_NODE_MEM_MAP
3746 printk(KERN_DEBUG
"free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
3747 nid
, (unsigned long)pgdat
,
3748 (unsigned long)pgdat
->node_mem_map
);
3751 free_area_init_core(pgdat
, zones_size
, zholes_size
);
3754 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3756 #if MAX_NUMNODES > 1
3758 * Figure out the number of possible node ids.
3760 static void __init
setup_nr_node_ids(void)
3763 unsigned int highest
= 0;
3765 for_each_node_mask(node
, node_possible_map
)
3767 nr_node_ids
= highest
+ 1;
3770 static inline void setup_nr_node_ids(void)
3776 * add_active_range - Register a range of PFNs backed by physical memory
3777 * @nid: The node ID the range resides on
3778 * @start_pfn: The start PFN of the available physical memory
3779 * @end_pfn: The end PFN of the available physical memory
3781 * These ranges are stored in an early_node_map[] and later used by
3782 * free_area_init_nodes() to calculate zone sizes and holes. If the
3783 * range spans a memory hole, it is up to the architecture to ensure
3784 * the memory is not freed by the bootmem allocator. If possible
3785 * the range being registered will be merged with existing ranges.
3787 void __init
add_active_range(unsigned int nid
, unsigned long start_pfn
,
3788 unsigned long end_pfn
)
3792 mminit_dprintk(MMINIT_TRACE
, "memory_register",
3793 "Entering add_active_range(%d, %#lx, %#lx) "
3794 "%d entries of %d used\n",
3795 nid
, start_pfn
, end_pfn
,
3796 nr_nodemap_entries
, MAX_ACTIVE_REGIONS
);
3798 mminit_validate_memmodel_limits(&start_pfn
, &end_pfn
);
3800 /* Merge with existing active regions if possible */
3801 for (i
= 0; i
< nr_nodemap_entries
; i
++) {
3802 if (early_node_map
[i
].nid
!= nid
)
3805 /* Skip if an existing region covers this new one */
3806 if (start_pfn
>= early_node_map
[i
].start_pfn
&&
3807 end_pfn
<= early_node_map
[i
].end_pfn
)
3810 /* Merge forward if suitable */
3811 if (start_pfn
<= early_node_map
[i
].end_pfn
&&
3812 end_pfn
> early_node_map
[i
].end_pfn
) {
3813 early_node_map
[i
].end_pfn
= end_pfn
;
3817 /* Merge backward if suitable */
3818 if (start_pfn
< early_node_map
[i
].end_pfn
&&
3819 end_pfn
>= early_node_map
[i
].start_pfn
) {
3820 early_node_map
[i
].start_pfn
= start_pfn
;
3825 /* Check that early_node_map is large enough */
3826 if (i
>= MAX_ACTIVE_REGIONS
) {
3827 printk(KERN_CRIT
"More than %d memory regions, truncating\n",
3828 MAX_ACTIVE_REGIONS
);
3832 early_node_map
[i
].nid
= nid
;
3833 early_node_map
[i
].start_pfn
= start_pfn
;
3834 early_node_map
[i
].end_pfn
= end_pfn
;
3835 nr_nodemap_entries
= i
+ 1;
3839 * remove_active_range - Shrink an existing registered range of PFNs
3840 * @nid: The node id the range is on that should be shrunk
3841 * @start_pfn: The new PFN of the range
3842 * @end_pfn: The new PFN of the range
3844 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
3845 * The map is kept near the end physical page range that has already been
3846 * registered. This function allows an arch to shrink an existing registered
3849 void __init
remove_active_range(unsigned int nid
, unsigned long start_pfn
,
3850 unsigned long end_pfn
)
3855 printk(KERN_DEBUG
"remove_active_range (%d, %lu, %lu)\n",
3856 nid
, start_pfn
, end_pfn
);
3858 /* Find the old active region end and shrink */
3859 for_each_active_range_index_in_nid(i
, nid
) {
3860 if (early_node_map
[i
].start_pfn
>= start_pfn
&&
3861 early_node_map
[i
].end_pfn
<= end_pfn
) {
3863 early_node_map
[i
].start_pfn
= 0;
3864 early_node_map
[i
].end_pfn
= 0;
3868 if (early_node_map
[i
].start_pfn
< start_pfn
&&
3869 early_node_map
[i
].end_pfn
> start_pfn
) {
3870 unsigned long temp_end_pfn
= early_node_map
[i
].end_pfn
;
3871 early_node_map
[i
].end_pfn
= start_pfn
;
3872 if (temp_end_pfn
> end_pfn
)
3873 add_active_range(nid
, end_pfn
, temp_end_pfn
);
3876 if (early_node_map
[i
].start_pfn
>= start_pfn
&&
3877 early_node_map
[i
].end_pfn
> end_pfn
&&
3878 early_node_map
[i
].start_pfn
< end_pfn
) {
3879 early_node_map
[i
].start_pfn
= end_pfn
;
3887 /* remove the blank ones */
3888 for (i
= nr_nodemap_entries
- 1; i
> 0; i
--) {
3889 if (early_node_map
[i
].nid
!= nid
)
3891 if (early_node_map
[i
].end_pfn
)
3893 /* we found it, get rid of it */
3894 for (j
= i
; j
< nr_nodemap_entries
- 1; j
++)
3895 memcpy(&early_node_map
[j
], &early_node_map
[j
+1],
3896 sizeof(early_node_map
[j
]));
3897 j
= nr_nodemap_entries
- 1;
3898 memset(&early_node_map
[j
], 0, sizeof(early_node_map
[j
]));
3899 nr_nodemap_entries
--;
3904 * remove_all_active_ranges - Remove all currently registered regions
3906 * During discovery, it may be found that a table like SRAT is invalid
3907 * and an alternative discovery method must be used. This function removes
3908 * all currently registered regions.
3910 void __init
remove_all_active_ranges(void)
3912 memset(early_node_map
, 0, sizeof(early_node_map
));
3913 nr_nodemap_entries
= 0;
3916 /* Compare two active node_active_regions */
3917 static int __init
cmp_node_active_region(const void *a
, const void *b
)
3919 struct node_active_region
*arange
= (struct node_active_region
*)a
;
3920 struct node_active_region
*brange
= (struct node_active_region
*)b
;
3922 /* Done this way to avoid overflows */
3923 if (arange
->start_pfn
> brange
->start_pfn
)
3925 if (arange
->start_pfn
< brange
->start_pfn
)
3931 /* sort the node_map by start_pfn */
3932 static void __init
sort_node_map(void)
3934 sort(early_node_map
, (size_t)nr_nodemap_entries
,
3935 sizeof(struct node_active_region
),
3936 cmp_node_active_region
, NULL
);
3939 /* Find the lowest pfn for a node */
3940 static unsigned long __init
find_min_pfn_for_node(int nid
)
3943 unsigned long min_pfn
= ULONG_MAX
;
3945 /* Assuming a sorted map, the first range found has the starting pfn */
3946 for_each_active_range_index_in_nid(i
, nid
)
3947 min_pfn
= min(min_pfn
, early_node_map
[i
].start_pfn
);
3949 if (min_pfn
== ULONG_MAX
) {
3951 "Could not find start_pfn for node %d\n", nid
);
3959 * find_min_pfn_with_active_regions - Find the minimum PFN registered
3961 * It returns the minimum PFN based on information provided via
3962 * add_active_range().
3964 unsigned long __init
find_min_pfn_with_active_regions(void)
3966 return find_min_pfn_for_node(MAX_NUMNODES
);
3970 * early_calculate_totalpages()
3971 * Sum pages in active regions for movable zone.
3972 * Populate N_HIGH_MEMORY for calculating usable_nodes.
3974 static unsigned long __init
early_calculate_totalpages(void)
3977 unsigned long totalpages
= 0;
3979 for (i
= 0; i
< nr_nodemap_entries
; i
++) {
3980 unsigned long pages
= early_node_map
[i
].end_pfn
-
3981 early_node_map
[i
].start_pfn
;
3982 totalpages
+= pages
;
3984 node_set_state(early_node_map
[i
].nid
, N_HIGH_MEMORY
);
3990 * Find the PFN the Movable zone begins in each node. Kernel memory
3991 * is spread evenly between nodes as long as the nodes have enough
3992 * memory. When they don't, some nodes will have more kernelcore than
3995 static void __init
find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn
)
3998 unsigned long usable_startpfn
;
3999 unsigned long kernelcore_node
, kernelcore_remaining
;
4000 unsigned long totalpages
= early_calculate_totalpages();
4001 int usable_nodes
= nodes_weight(node_states
[N_HIGH_MEMORY
]);
4004 * If movablecore was specified, calculate what size of
4005 * kernelcore that corresponds so that memory usable for
4006 * any allocation type is evenly spread. If both kernelcore
4007 * and movablecore are specified, then the value of kernelcore
4008 * will be used for required_kernelcore if it's greater than
4009 * what movablecore would have allowed.
4011 if (required_movablecore
) {
4012 unsigned long corepages
;
4015 * Round-up so that ZONE_MOVABLE is at least as large as what
4016 * was requested by the user
4018 required_movablecore
=
4019 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
4020 corepages
= totalpages
- required_movablecore
;
4022 required_kernelcore
= max(required_kernelcore
, corepages
);
4025 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4026 if (!required_kernelcore
)
4029 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4030 find_usable_zone_for_movable();
4031 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
4034 /* Spread kernelcore memory as evenly as possible throughout nodes */
4035 kernelcore_node
= required_kernelcore
/ usable_nodes
;
4036 for_each_node_state(nid
, N_HIGH_MEMORY
) {
4038 * Recalculate kernelcore_node if the division per node
4039 * now exceeds what is necessary to satisfy the requested
4040 * amount of memory for the kernel
4042 if (required_kernelcore
< kernelcore_node
)
4043 kernelcore_node
= required_kernelcore
/ usable_nodes
;
4046 * As the map is walked, we track how much memory is usable
4047 * by the kernel using kernelcore_remaining. When it is
4048 * 0, the rest of the node is usable by ZONE_MOVABLE
4050 kernelcore_remaining
= kernelcore_node
;
4052 /* Go through each range of PFNs within this node */
4053 for_each_active_range_index_in_nid(i
, nid
) {
4054 unsigned long start_pfn
, end_pfn
;
4055 unsigned long size_pages
;
4057 start_pfn
= max(early_node_map
[i
].start_pfn
,
4058 zone_movable_pfn
[nid
]);
4059 end_pfn
= early_node_map
[i
].end_pfn
;
4060 if (start_pfn
>= end_pfn
)
4063 /* Account for what is only usable for kernelcore */
4064 if (start_pfn
< usable_startpfn
) {
4065 unsigned long kernel_pages
;
4066 kernel_pages
= min(end_pfn
, usable_startpfn
)
4069 kernelcore_remaining
-= min(kernel_pages
,
4070 kernelcore_remaining
);
4071 required_kernelcore
-= min(kernel_pages
,
4072 required_kernelcore
);
4074 /* Continue if range is now fully accounted */
4075 if (end_pfn
<= usable_startpfn
) {
4078 * Push zone_movable_pfn to the end so
4079 * that if we have to rebalance
4080 * kernelcore across nodes, we will
4081 * not double account here
4083 zone_movable_pfn
[nid
] = end_pfn
;
4086 start_pfn
= usable_startpfn
;
4090 * The usable PFN range for ZONE_MOVABLE is from
4091 * start_pfn->end_pfn. Calculate size_pages as the
4092 * number of pages used as kernelcore
4094 size_pages
= end_pfn
- start_pfn
;
4095 if (size_pages
> kernelcore_remaining
)
4096 size_pages
= kernelcore_remaining
;
4097 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
4100 * Some kernelcore has been met, update counts and
4101 * break if the kernelcore for this node has been
4104 required_kernelcore
-= min(required_kernelcore
,
4106 kernelcore_remaining
-= size_pages
;
4107 if (!kernelcore_remaining
)
4113 * If there is still required_kernelcore, we do another pass with one
4114 * less node in the count. This will push zone_movable_pfn[nid] further
4115 * along on the nodes that still have memory until kernelcore is
4119 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
4122 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4123 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
4124 zone_movable_pfn
[nid
] =
4125 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
4128 /* Any regular memory on that node ? */
4129 static void check_for_regular_memory(pg_data_t
*pgdat
)
4131 #ifdef CONFIG_HIGHMEM
4132 enum zone_type zone_type
;
4134 for (zone_type
= 0; zone_type
<= ZONE_NORMAL
; zone_type
++) {
4135 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
4136 if (zone
->present_pages
)
4137 node_set_state(zone_to_nid(zone
), N_NORMAL_MEMORY
);
4143 * free_area_init_nodes - Initialise all pg_data_t and zone data
4144 * @max_zone_pfn: an array of max PFNs for each zone
4146 * This will call free_area_init_node() for each active node in the system.
4147 * Using the page ranges provided by add_active_range(), the size of each
4148 * zone in each node and their holes is calculated. If the maximum PFN
4149 * between two adjacent zones match, it is assumed that the zone is empty.
4150 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4151 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4152 * starts where the previous one ended. For example, ZONE_DMA32 starts
4153 * at arch_max_dma_pfn.
4155 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
4160 /* Sort early_node_map as initialisation assumes it is sorted */
4163 /* Record where the zone boundaries are */
4164 memset(arch_zone_lowest_possible_pfn
, 0,
4165 sizeof(arch_zone_lowest_possible_pfn
));
4166 memset(arch_zone_highest_possible_pfn
, 0,
4167 sizeof(arch_zone_highest_possible_pfn
));
4168 arch_zone_lowest_possible_pfn
[0] = find_min_pfn_with_active_regions();
4169 arch_zone_highest_possible_pfn
[0] = max_zone_pfn
[0];
4170 for (i
= 1; i
< MAX_NR_ZONES
; i
++) {
4171 if (i
== ZONE_MOVABLE
)
4173 arch_zone_lowest_possible_pfn
[i
] =
4174 arch_zone_highest_possible_pfn
[i
-1];
4175 arch_zone_highest_possible_pfn
[i
] =
4176 max(max_zone_pfn
[i
], arch_zone_lowest_possible_pfn
[i
]);
4178 arch_zone_lowest_possible_pfn
[ZONE_MOVABLE
] = 0;
4179 arch_zone_highest_possible_pfn
[ZONE_MOVABLE
] = 0;
4181 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4182 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
4183 find_zone_movable_pfns_for_nodes(zone_movable_pfn
);
4185 /* Print out the zone ranges */
4186 printk("Zone PFN ranges:\n");
4187 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
4188 if (i
== ZONE_MOVABLE
)
4190 printk(" %-8s %0#10lx -> %0#10lx\n",
4192 arch_zone_lowest_possible_pfn
[i
],
4193 arch_zone_highest_possible_pfn
[i
]);
4196 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4197 printk("Movable zone start PFN for each node\n");
4198 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
4199 if (zone_movable_pfn
[i
])
4200 printk(" Node %d: %lu\n", i
, zone_movable_pfn
[i
]);
4203 /* Print out the early_node_map[] */
4204 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries
);
4205 for (i
= 0; i
< nr_nodemap_entries
; i
++)
4206 printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map
[i
].nid
,
4207 early_node_map
[i
].start_pfn
,
4208 early_node_map
[i
].end_pfn
);
4210 /* Initialise every node */
4211 mminit_verify_pageflags_layout();
4212 setup_nr_node_ids();
4213 for_each_online_node(nid
) {
4214 pg_data_t
*pgdat
= NODE_DATA(nid
);
4215 free_area_init_node(nid
, NULL
,
4216 find_min_pfn_for_node(nid
), NULL
);
4218 /* Any memory on that node */
4219 if (pgdat
->node_present_pages
)
4220 node_set_state(nid
, N_HIGH_MEMORY
);
4221 check_for_regular_memory(pgdat
);
4225 static int __init
cmdline_parse_core(char *p
, unsigned long *core
)
4227 unsigned long long coremem
;
4231 coremem
= memparse(p
, &p
);
4232 *core
= coremem
>> PAGE_SHIFT
;
4234 /* Paranoid check that UL is enough for the coremem value */
4235 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
4241 * kernelcore=size sets the amount of memory for use for allocations that
4242 * cannot be reclaimed or migrated.
4244 static int __init
cmdline_parse_kernelcore(char *p
)
4246 return cmdline_parse_core(p
, &required_kernelcore
);
4250 * movablecore=size sets the amount of memory for use for allocations that
4251 * can be reclaimed or migrated.
4253 static int __init
cmdline_parse_movablecore(char *p
)
4255 return cmdline_parse_core(p
, &required_movablecore
);
4258 early_param("kernelcore", cmdline_parse_kernelcore
);
4259 early_param("movablecore", cmdline_parse_movablecore
);
4261 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4264 * set_dma_reserve - set the specified number of pages reserved in the first zone
4265 * @new_dma_reserve: The number of pages to mark reserved
4267 * The per-cpu batchsize and zone watermarks are determined by present_pages.
4268 * In the DMA zone, a significant percentage may be consumed by kernel image
4269 * and other unfreeable allocations which can skew the watermarks badly. This
4270 * function may optionally be used to account for unfreeable pages in the
4271 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
4272 * smaller per-cpu batchsize.
4274 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
4276 dma_reserve
= new_dma_reserve
;
4279 #ifndef CONFIG_NEED_MULTIPLE_NODES
4280 struct pglist_data __refdata contig_page_data
= { .bdata
= &bootmem_node_data
[0] };
4281 EXPORT_SYMBOL(contig_page_data
);
4284 void __init
free_area_init(unsigned long *zones_size
)
4286 free_area_init_node(0, zones_size
,
4287 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
4290 static int page_alloc_cpu_notify(struct notifier_block
*self
,
4291 unsigned long action
, void *hcpu
)
4293 int cpu
= (unsigned long)hcpu
;
4295 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
) {
4299 * Spill the event counters of the dead processor
4300 * into the current processors event counters.
4301 * This artificially elevates the count of the current
4304 vm_events_fold_cpu(cpu
);
4307 * Zero the differential counters of the dead processor
4308 * so that the vm statistics are consistent.
4310 * This is only okay since the processor is dead and cannot
4311 * race with what we are doing.
4313 refresh_cpu_vm_stats(cpu
);
4318 void __init
page_alloc_init(void)
4320 hotcpu_notifier(page_alloc_cpu_notify
, 0);
4324 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
4325 * or min_free_kbytes changes.
4327 static void calculate_totalreserve_pages(void)
4329 struct pglist_data
*pgdat
;
4330 unsigned long reserve_pages
= 0;
4331 enum zone_type i
, j
;
4333 for_each_online_pgdat(pgdat
) {
4334 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
4335 struct zone
*zone
= pgdat
->node_zones
+ i
;
4336 unsigned long max
= 0;
4338 /* Find valid and maximum lowmem_reserve in the zone */
4339 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
4340 if (zone
->lowmem_reserve
[j
] > max
)
4341 max
= zone
->lowmem_reserve
[j
];
4344 /* we treat the high watermark as reserved pages. */
4345 max
+= high_wmark_pages(zone
);
4347 if (max
> zone
->present_pages
)
4348 max
= zone
->present_pages
;
4349 reserve_pages
+= max
;
4352 totalreserve_pages
= reserve_pages
;
4356 * setup_per_zone_lowmem_reserve - called whenever
4357 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
4358 * has a correct pages reserved value, so an adequate number of
4359 * pages are left in the zone after a successful __alloc_pages().
4361 static void setup_per_zone_lowmem_reserve(void)
4363 struct pglist_data
*pgdat
;
4364 enum zone_type j
, idx
;
4366 for_each_online_pgdat(pgdat
) {
4367 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
4368 struct zone
*zone
= pgdat
->node_zones
+ j
;
4369 unsigned long present_pages
= zone
->present_pages
;
4371 zone
->lowmem_reserve
[j
] = 0;
4375 struct zone
*lower_zone
;
4379 if (sysctl_lowmem_reserve_ratio
[idx
] < 1)
4380 sysctl_lowmem_reserve_ratio
[idx
] = 1;
4382 lower_zone
= pgdat
->node_zones
+ idx
;
4383 lower_zone
->lowmem_reserve
[j
] = present_pages
/
4384 sysctl_lowmem_reserve_ratio
[idx
];
4385 present_pages
+= lower_zone
->present_pages
;
4390 /* update totalreserve_pages */
4391 calculate_totalreserve_pages();
4395 * setup_per_zone_pages_min - called when min_free_kbytes changes.
4397 * Ensures that the pages_{min,low,high} values for each zone are set correctly
4398 * with respect to min_free_kbytes.
4400 void setup_per_zone_pages_min(void)
4402 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
4403 unsigned long lowmem_pages
= 0;
4405 unsigned long flags
;
4407 /* Calculate total number of !ZONE_HIGHMEM pages */
4408 for_each_zone(zone
) {
4409 if (!is_highmem(zone
))
4410 lowmem_pages
+= zone
->present_pages
;
4413 for_each_zone(zone
) {
4416 spin_lock_irqsave(&zone
->lock
, flags
);
4417 tmp
= (u64
)pages_min
* zone
->present_pages
;
4418 do_div(tmp
, lowmem_pages
);
4419 if (is_highmem(zone
)) {
4421 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
4422 * need highmem pages, so cap pages_min to a small
4425 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
4426 * deltas controls asynch page reclaim, and so should
4427 * not be capped for highmem.
4431 min_pages
= zone
->present_pages
/ 1024;
4432 if (min_pages
< SWAP_CLUSTER_MAX
)
4433 min_pages
= SWAP_CLUSTER_MAX
;
4434 if (min_pages
> 128)
4436 zone
->watermark
[WMARK_MIN
] = min_pages
;
4439 * If it's a lowmem zone, reserve a number of pages
4440 * proportionate to the zone's size.
4442 zone
->watermark
[WMARK_MIN
] = tmp
;
4445 zone
->watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + (tmp
>> 2);
4446 zone
->watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + (tmp
>> 1);
4447 setup_zone_migrate_reserve(zone
);
4448 spin_unlock_irqrestore(&zone
->lock
, flags
);
4451 /* update totalreserve_pages */
4452 calculate_totalreserve_pages();
4456 * The inactive anon list should be small enough that the VM never has to
4457 * do too much work, but large enough that each inactive page has a chance
4458 * to be referenced again before it is swapped out.
4460 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
4461 * INACTIVE_ANON pages on this zone's LRU, maintained by the
4462 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
4463 * the anonymous pages are kept on the inactive list.
4466 * memory ratio inactive anon
4467 * -------------------------------------
4476 static void __init
setup_per_zone_inactive_ratio(void)
4480 for_each_zone(zone
) {
4481 unsigned int gb
, ratio
;
4483 /* Zone size in gigabytes */
4484 gb
= zone
->present_pages
>> (30 - PAGE_SHIFT
);
4486 ratio
= int_sqrt(10 * gb
);
4490 zone
->inactive_ratio
= ratio
;
4495 * Initialise min_free_kbytes.
4497 * For small machines we want it small (128k min). For large machines
4498 * we want it large (64MB max). But it is not linear, because network
4499 * bandwidth does not increase linearly with machine size. We use
4501 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
4502 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
4518 static int __init
init_per_zone_pages_min(void)
4520 unsigned long lowmem_kbytes
;
4522 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
4524 min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
4525 if (min_free_kbytes
< 128)
4526 min_free_kbytes
= 128;
4527 if (min_free_kbytes
> 65536)
4528 min_free_kbytes
= 65536;
4529 setup_per_zone_pages_min();
4530 setup_per_zone_lowmem_reserve();
4531 setup_per_zone_inactive_ratio();
4534 module_init(init_per_zone_pages_min
)
4537 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
4538 * that we can call two helper functions whenever min_free_kbytes
4541 int min_free_kbytes_sysctl_handler(ctl_table
*table
, int write
,
4542 struct file
*file
, void __user
*buffer
, size_t *length
, loff_t
*ppos
)
4544 proc_dointvec(table
, write
, file
, buffer
, length
, ppos
);
4546 setup_per_zone_pages_min();
4551 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table
*table
, int write
,
4552 struct file
*file
, void __user
*buffer
, size_t *length
, loff_t
*ppos
)
4557 rc
= proc_dointvec_minmax(table
, write
, file
, buffer
, length
, ppos
);
4562 zone
->min_unmapped_pages
= (zone
->present_pages
*
4563 sysctl_min_unmapped_ratio
) / 100;
4567 int sysctl_min_slab_ratio_sysctl_handler(ctl_table
*table
, int write
,
4568 struct file
*file
, void __user
*buffer
, size_t *length
, loff_t
*ppos
)
4573 rc
= proc_dointvec_minmax(table
, write
, file
, buffer
, length
, ppos
);
4578 zone
->min_slab_pages
= (zone
->present_pages
*
4579 sysctl_min_slab_ratio
) / 100;
4585 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
4586 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
4587 * whenever sysctl_lowmem_reserve_ratio changes.
4589 * The reserve ratio obviously has absolutely no relation with the
4590 * minimum watermarks. The lowmem reserve ratio can only make sense
4591 * if in function of the boot time zone sizes.
4593 int lowmem_reserve_ratio_sysctl_handler(ctl_table
*table
, int write
,
4594 struct file
*file
, void __user
*buffer
, size_t *length
, loff_t
*ppos
)
4596 proc_dointvec_minmax(table
, write
, file
, buffer
, length
, ppos
);
4597 setup_per_zone_lowmem_reserve();
4602 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
4603 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
4604 * can have before it gets flushed back to buddy allocator.
4607 int percpu_pagelist_fraction_sysctl_handler(ctl_table
*table
, int write
,
4608 struct file
*file
, void __user
*buffer
, size_t *length
, loff_t
*ppos
)
4614 ret
= proc_dointvec_minmax(table
, write
, file
, buffer
, length
, ppos
);
4615 if (!write
|| (ret
== -EINVAL
))
4617 for_each_zone(zone
) {
4618 for_each_online_cpu(cpu
) {
4620 high
= zone
->present_pages
/ percpu_pagelist_fraction
;
4621 setup_pagelist_highmark(zone_pcp(zone
, cpu
), high
);
4627 int hashdist
= HASHDIST_DEFAULT
;
4630 static int __init
set_hashdist(char *str
)
4634 hashdist
= simple_strtoul(str
, &str
, 0);
4637 __setup("hashdist=", set_hashdist
);
4641 * allocate a large system hash table from bootmem
4642 * - it is assumed that the hash table must contain an exact power-of-2
4643 * quantity of entries
4644 * - limit is the number of hash buckets, not the total allocation size
4646 void *__init
alloc_large_system_hash(const char *tablename
,
4647 unsigned long bucketsize
,
4648 unsigned long numentries
,
4651 unsigned int *_hash_shift
,
4652 unsigned int *_hash_mask
,
4653 unsigned long limit
)
4655 unsigned long long max
= limit
;
4656 unsigned long log2qty
, size
;
4659 /* allow the kernel cmdline to have a say */
4661 /* round applicable memory size up to nearest megabyte */
4662 numentries
= nr_kernel_pages
;
4663 numentries
+= (1UL << (20 - PAGE_SHIFT
)) - 1;
4664 numentries
>>= 20 - PAGE_SHIFT
;
4665 numentries
<<= 20 - PAGE_SHIFT
;
4667 /* limit to 1 bucket per 2^scale bytes of low memory */
4668 if (scale
> PAGE_SHIFT
)
4669 numentries
>>= (scale
- PAGE_SHIFT
);
4671 numentries
<<= (PAGE_SHIFT
- scale
);
4673 /* Make sure we've got at least a 0-order allocation.. */
4674 if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
4675 numentries
= PAGE_SIZE
/ bucketsize
;
4677 numentries
= roundup_pow_of_two(numentries
);
4679 /* limit allocation size to 1/16 total memory by default */
4681 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
4682 do_div(max
, bucketsize
);
4685 if (numentries
> max
)
4688 log2qty
= ilog2(numentries
);
4691 size
= bucketsize
<< log2qty
;
4692 if (flags
& HASH_EARLY
)
4693 table
= alloc_bootmem_nopanic(size
);
4695 table
= __vmalloc(size
, GFP_ATOMIC
, PAGE_KERNEL
);
4698 * If bucketsize is not a power-of-two, we may free
4699 * some pages at the end of hash table which
4700 * alloc_pages_exact() automatically does
4702 if (get_order(size
) < MAX_ORDER
)
4703 table
= alloc_pages_exact(size
, GFP_ATOMIC
);
4705 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
4708 panic("Failed to allocate %s hash table\n", tablename
);
4710 printk(KERN_INFO
"%s hash table entries: %d (order: %d, %lu bytes)\n",
4713 ilog2(size
) - PAGE_SHIFT
,
4717 *_hash_shift
= log2qty
;
4719 *_hash_mask
= (1 << log2qty
) - 1;
4722 * If hashdist is set, the table allocation is done with __vmalloc()
4723 * which invokes the kmemleak_alloc() callback. This function may also
4724 * be called before the slab and kmemleak are initialised when
4725 * kmemleak simply buffers the request to be executed later
4726 * (GFP_ATOMIC flag ignored in this case).
4729 kmemleak_alloc(table
, size
, 1, GFP_ATOMIC
);
4734 /* Return a pointer to the bitmap storing bits affecting a block of pages */
4735 static inline unsigned long *get_pageblock_bitmap(struct zone
*zone
,
4738 #ifdef CONFIG_SPARSEMEM
4739 return __pfn_to_section(pfn
)->pageblock_flags
;
4741 return zone
->pageblock_flags
;
4742 #endif /* CONFIG_SPARSEMEM */
4745 static inline int pfn_to_bitidx(struct zone
*zone
, unsigned long pfn
)
4747 #ifdef CONFIG_SPARSEMEM
4748 pfn
&= (PAGES_PER_SECTION
-1);
4749 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
4751 pfn
= pfn
- zone
->zone_start_pfn
;
4752 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
4753 #endif /* CONFIG_SPARSEMEM */
4757 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
4758 * @page: The page within the block of interest
4759 * @start_bitidx: The first bit of interest to retrieve
4760 * @end_bitidx: The last bit of interest
4761 * returns pageblock_bits flags
4763 unsigned long get_pageblock_flags_group(struct page
*page
,
4764 int start_bitidx
, int end_bitidx
)
4767 unsigned long *bitmap
;
4768 unsigned long pfn
, bitidx
;
4769 unsigned long flags
= 0;
4770 unsigned long value
= 1;
4772 zone
= page_zone(page
);
4773 pfn
= page_to_pfn(page
);
4774 bitmap
= get_pageblock_bitmap(zone
, pfn
);
4775 bitidx
= pfn_to_bitidx(zone
, pfn
);
4777 for (; start_bitidx
<= end_bitidx
; start_bitidx
++, value
<<= 1)
4778 if (test_bit(bitidx
+ start_bitidx
, bitmap
))
4785 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
4786 * @page: The page within the block of interest
4787 * @start_bitidx: The first bit of interest
4788 * @end_bitidx: The last bit of interest
4789 * @flags: The flags to set
4791 void set_pageblock_flags_group(struct page
*page
, unsigned long flags
,
4792 int start_bitidx
, int end_bitidx
)
4795 unsigned long *bitmap
;
4796 unsigned long pfn
, bitidx
;
4797 unsigned long value
= 1;
4799 zone
= page_zone(page
);
4800 pfn
= page_to_pfn(page
);
4801 bitmap
= get_pageblock_bitmap(zone
, pfn
);
4802 bitidx
= pfn_to_bitidx(zone
, pfn
);
4803 VM_BUG_ON(pfn
< zone
->zone_start_pfn
);
4804 VM_BUG_ON(pfn
>= zone
->zone_start_pfn
+ zone
->spanned_pages
);
4806 for (; start_bitidx
<= end_bitidx
; start_bitidx
++, value
<<= 1)
4808 __set_bit(bitidx
+ start_bitidx
, bitmap
);
4810 __clear_bit(bitidx
+ start_bitidx
, bitmap
);
4814 * This is designed as sub function...plz see page_isolation.c also.
4815 * set/clear page block's type to be ISOLATE.
4816 * page allocater never alloc memory from ISOLATE block.
4819 int set_migratetype_isolate(struct page
*page
)
4822 unsigned long flags
;
4825 zone
= page_zone(page
);
4826 spin_lock_irqsave(&zone
->lock
, flags
);
4828 * In future, more migrate types will be able to be isolation target.
4830 if (get_pageblock_migratetype(page
) != MIGRATE_MOVABLE
)
4832 set_pageblock_migratetype(page
, MIGRATE_ISOLATE
);
4833 move_freepages_block(zone
, page
, MIGRATE_ISOLATE
);
4836 spin_unlock_irqrestore(&zone
->lock
, flags
);
4842 void unset_migratetype_isolate(struct page
*page
)
4845 unsigned long flags
;
4846 zone
= page_zone(page
);
4847 spin_lock_irqsave(&zone
->lock
, flags
);
4848 if (get_pageblock_migratetype(page
) != MIGRATE_ISOLATE
)
4850 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
4851 move_freepages_block(zone
, page
, MIGRATE_MOVABLE
);
4853 spin_unlock_irqrestore(&zone
->lock
, flags
);
4856 #ifdef CONFIG_MEMORY_HOTREMOVE
4858 * All pages in the range must be isolated before calling this.
4861 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
4867 unsigned long flags
;
4868 /* find the first valid pfn */
4869 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
4874 zone
= page_zone(pfn_to_page(pfn
));
4875 spin_lock_irqsave(&zone
->lock
, flags
);
4877 while (pfn
< end_pfn
) {
4878 if (!pfn_valid(pfn
)) {
4882 page
= pfn_to_page(pfn
);
4883 BUG_ON(page_count(page
));
4884 BUG_ON(!PageBuddy(page
));
4885 order
= page_order(page
);
4886 #ifdef CONFIG_DEBUG_VM
4887 printk(KERN_INFO
"remove from free list %lx %d %lx\n",
4888 pfn
, 1 << order
, end_pfn
);
4890 list_del(&page
->lru
);
4891 rmv_page_order(page
);
4892 zone
->free_area
[order
].nr_free
--;
4893 __mod_zone_page_state(zone
, NR_FREE_PAGES
,
4895 for (i
= 0; i
< (1 << order
); i
++)
4896 SetPageReserved((page
+i
));
4897 pfn
+= (1 << order
);
4899 spin_unlock_irqrestore(&zone
->lock
, flags
);