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/kmemcheck.h>
27 #include <linux/module.h>
28 #include <linux/suspend.h>
29 #include <linux/pagevec.h>
30 #include <linux/blkdev.h>
31 #include <linux/slab.h>
32 #include <linux/oom.h>
33 #include <linux/notifier.h>
34 #include <linux/topology.h>
35 #include <linux/sysctl.h>
36 #include <linux/cpu.h>
37 #include <linux/cpuset.h>
38 #include <linux/memory_hotplug.h>
39 #include <linux/nodemask.h>
40 #include <linux/vmalloc.h>
41 #include <linux/mempolicy.h>
42 #include <linux/stop_machine.h>
43 #include <linux/sort.h>
44 #include <linux/pfn.h>
45 #include <linux/backing-dev.h>
46 #include <linux/fault-inject.h>
47 #include <linux/page-isolation.h>
48 #include <linux/page_cgroup.h>
49 #include <linux/debugobjects.h>
50 #include <linux/kmemleak.h>
51 #include <linux/memory.h>
52 #include <trace/events/kmem.h>
54 #include <asm/tlbflush.h>
55 #include <asm/div64.h>
59 * Array of node states.
61 nodemask_t node_states
[NR_NODE_STATES
] __read_mostly
= {
62 [N_POSSIBLE
] = NODE_MASK_ALL
,
63 [N_ONLINE
] = { { [0] = 1UL } },
65 [N_NORMAL_MEMORY
] = { { [0] = 1UL } },
67 [N_HIGH_MEMORY
] = { { [0] = 1UL } },
69 [N_CPU
] = { { [0] = 1UL } },
72 EXPORT_SYMBOL(node_states
);
74 unsigned long totalram_pages __read_mostly
;
75 unsigned long totalreserve_pages __read_mostly
;
76 int percpu_pagelist_fraction
;
77 gfp_t gfp_allowed_mask __read_mostly
= GFP_BOOT_MASK
;
79 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
80 int pageblock_order __read_mostly
;
83 static void __free_pages_ok(struct page
*page
, unsigned int order
);
86 * results with 256, 32 in the lowmem_reserve sysctl:
87 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
88 * 1G machine -> (16M dma, 784M normal, 224M high)
89 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
90 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
91 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
93 * TBD: should special case ZONE_DMA32 machines here - in those we normally
94 * don't need any ZONE_NORMAL reservation
96 int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
-1] = {
97 #ifdef CONFIG_ZONE_DMA
100 #ifdef CONFIG_ZONE_DMA32
103 #ifdef CONFIG_HIGHMEM
109 EXPORT_SYMBOL(totalram_pages
);
111 static char * const zone_names
[MAX_NR_ZONES
] = {
112 #ifdef CONFIG_ZONE_DMA
115 #ifdef CONFIG_ZONE_DMA32
119 #ifdef CONFIG_HIGHMEM
125 int min_free_kbytes
= 1024;
127 static unsigned long __meminitdata nr_kernel_pages
;
128 static unsigned long __meminitdata nr_all_pages
;
129 static unsigned long __meminitdata dma_reserve
;
131 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
133 * MAX_ACTIVE_REGIONS determines the maximum number of distinct
134 * ranges of memory (RAM) that may be registered with add_active_range().
135 * Ranges passed to add_active_range() will be merged if possible
136 * so the number of times add_active_range() can be called is
137 * related to the number of nodes and the number of holes
139 #ifdef CONFIG_MAX_ACTIVE_REGIONS
140 /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */
141 #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS
143 #if MAX_NUMNODES >= 32
144 /* If there can be many nodes, allow up to 50 holes per node */
145 #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50)
147 /* By default, allow up to 256 distinct regions */
148 #define MAX_ACTIVE_REGIONS 256
152 static struct node_active_region __meminitdata early_node_map
[MAX_ACTIVE_REGIONS
];
153 static int __meminitdata nr_nodemap_entries
;
154 static unsigned long __meminitdata arch_zone_lowest_possible_pfn
[MAX_NR_ZONES
];
155 static unsigned long __meminitdata arch_zone_highest_possible_pfn
[MAX_NR_ZONES
];
156 static unsigned long __initdata required_kernelcore
;
157 static unsigned long __initdata required_movablecore
;
158 static unsigned long __meminitdata zone_movable_pfn
[MAX_NUMNODES
];
160 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
162 EXPORT_SYMBOL(movable_zone
);
163 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
166 int nr_node_ids __read_mostly
= MAX_NUMNODES
;
167 int nr_online_nodes __read_mostly
= 1;
168 EXPORT_SYMBOL(nr_node_ids
);
169 EXPORT_SYMBOL(nr_online_nodes
);
172 int page_group_by_mobility_disabled __read_mostly
;
174 static void set_pageblock_migratetype(struct page
*page
, int migratetype
)
177 if (unlikely(page_group_by_mobility_disabled
))
178 migratetype
= MIGRATE_UNMOVABLE
;
180 set_pageblock_flags_group(page
, (unsigned long)migratetype
,
181 PB_migrate
, PB_migrate_end
);
184 bool oom_killer_disabled __read_mostly
;
186 #ifdef CONFIG_DEBUG_VM
187 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
191 unsigned long pfn
= page_to_pfn(page
);
194 seq
= zone_span_seqbegin(zone
);
195 if (pfn
>= zone
->zone_start_pfn
+ zone
->spanned_pages
)
197 else if (pfn
< zone
->zone_start_pfn
)
199 } while (zone_span_seqretry(zone
, seq
));
204 static int page_is_consistent(struct zone
*zone
, struct page
*page
)
206 if (!pfn_valid_within(page_to_pfn(page
)))
208 if (zone
!= page_zone(page
))
214 * Temporary debugging check for pages not lying within a given zone.
216 static int bad_range(struct zone
*zone
, struct page
*page
)
218 if (page_outside_zone_boundaries(zone
, page
))
220 if (!page_is_consistent(zone
, page
))
226 static inline int bad_range(struct zone
*zone
, struct page
*page
)
232 static void bad_page(struct page
*page
)
234 static unsigned long resume
;
235 static unsigned long nr_shown
;
236 static unsigned long nr_unshown
;
238 /* Don't complain about poisoned pages */
239 if (PageHWPoison(page
)) {
240 __ClearPageBuddy(page
);
245 * Allow a burst of 60 reports, then keep quiet for that minute;
246 * or allow a steady drip of one report per second.
248 if (nr_shown
== 60) {
249 if (time_before(jiffies
, resume
)) {
255 "BUG: Bad page state: %lu messages suppressed\n",
262 resume
= jiffies
+ 60 * HZ
;
264 printk(KERN_ALERT
"BUG: Bad page state in process %s pfn:%05lx\n",
265 current
->comm
, page_to_pfn(page
));
267 "page:%p flags:%p count:%d mapcount:%d mapping:%p index:%lx\n",
268 page
, (void *)page
->flags
, page_count(page
),
269 page_mapcount(page
), page
->mapping
, page
->index
);
273 /* Leave bad fields for debug, except PageBuddy could make trouble */
274 __ClearPageBuddy(page
);
275 add_taint(TAINT_BAD_PAGE
);
279 * Higher-order pages are called "compound pages". They are structured thusly:
281 * The first PAGE_SIZE page is called the "head page".
283 * The remaining PAGE_SIZE pages are called "tail pages".
285 * All pages have PG_compound set. All pages have their ->private pointing at
286 * the head page (even the head page has this).
288 * The first tail page's ->lru.next holds the address of the compound page's
289 * put_page() function. Its ->lru.prev holds the order of allocation.
290 * This usage means that zero-order pages may not be compound.
293 static void free_compound_page(struct page
*page
)
295 __free_pages_ok(page
, compound_order(page
));
298 void prep_compound_page(struct page
*page
, unsigned long order
)
301 int nr_pages
= 1 << order
;
303 set_compound_page_dtor(page
, free_compound_page
);
304 set_compound_order(page
, order
);
306 for (i
= 1; i
< nr_pages
; i
++) {
307 struct page
*p
= page
+ i
;
310 p
->first_page
= page
;
314 static int destroy_compound_page(struct page
*page
, unsigned long order
)
317 int nr_pages
= 1 << order
;
320 if (unlikely(compound_order(page
) != order
) ||
321 unlikely(!PageHead(page
))) {
326 __ClearPageHead(page
);
328 for (i
= 1; i
< nr_pages
; i
++) {
329 struct page
*p
= page
+ i
;
331 if (unlikely(!PageTail(p
) || (p
->first_page
!= page
))) {
341 static inline void prep_zero_page(struct page
*page
, int order
, gfp_t gfp_flags
)
346 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
347 * and __GFP_HIGHMEM from hard or soft interrupt context.
349 VM_BUG_ON((gfp_flags
& __GFP_HIGHMEM
) && in_interrupt());
350 for (i
= 0; i
< (1 << order
); i
++)
351 clear_highpage(page
+ i
);
354 static inline void set_page_order(struct page
*page
, int order
)
356 set_page_private(page
, order
);
357 __SetPageBuddy(page
);
360 static inline void rmv_page_order(struct page
*page
)
362 __ClearPageBuddy(page
);
363 set_page_private(page
, 0);
367 * Locate the struct page for both the matching buddy in our
368 * pair (buddy1) and the combined O(n+1) page they form (page).
370 * 1) Any buddy B1 will have an order O twin B2 which satisfies
371 * the following equation:
373 * For example, if the starting buddy (buddy2) is #8 its order
375 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
377 * 2) Any buddy B will have an order O+1 parent P which
378 * satisfies the following equation:
381 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
383 static inline struct page
*
384 __page_find_buddy(struct page
*page
, unsigned long page_idx
, unsigned int order
)
386 unsigned long buddy_idx
= page_idx
^ (1 << order
);
388 return page
+ (buddy_idx
- page_idx
);
391 static inline unsigned long
392 __find_combined_index(unsigned long page_idx
, unsigned int order
)
394 return (page_idx
& ~(1 << order
));
398 * This function checks whether a page is free && is the buddy
399 * we can do coalesce a page and its buddy if
400 * (a) the buddy is not in a hole &&
401 * (b) the buddy is in the buddy system &&
402 * (c) a page and its buddy have the same order &&
403 * (d) a page and its buddy are in the same zone.
405 * For recording whether a page is in the buddy system, we use PG_buddy.
406 * Setting, clearing, and testing PG_buddy is serialized by zone->lock.
408 * For recording page's order, we use page_private(page).
410 static inline int page_is_buddy(struct page
*page
, struct page
*buddy
,
413 if (!pfn_valid_within(page_to_pfn(buddy
)))
416 if (page_zone_id(page
) != page_zone_id(buddy
))
419 if (PageBuddy(buddy
) && page_order(buddy
) == order
) {
420 VM_BUG_ON(page_count(buddy
) != 0);
427 * Freeing function for a buddy system allocator.
429 * The concept of a buddy system is to maintain direct-mapped table
430 * (containing bit values) for memory blocks of various "orders".
431 * The bottom level table contains the map for the smallest allocatable
432 * units of memory (here, pages), and each level above it describes
433 * pairs of units from the levels below, hence, "buddies".
434 * At a high level, all that happens here is marking the table entry
435 * at the bottom level available, and propagating the changes upward
436 * as necessary, plus some accounting needed to play nicely with other
437 * parts of the VM system.
438 * At each level, we keep a list of pages, which are heads of continuous
439 * free pages of length of (1 << order) and marked with PG_buddy. Page's
440 * order is recorded in page_private(page) field.
441 * So when we are allocating or freeing one, we can derive the state of the
442 * other. That is, if we allocate a small block, and both were
443 * free, the remainder of the region must be split into blocks.
444 * If a block is freed, and its buddy is also free, then this
445 * triggers coalescing into a block of larger size.
450 static inline void __free_one_page(struct page
*page
,
451 struct zone
*zone
, unsigned int order
,
454 unsigned long page_idx
;
456 if (unlikely(PageCompound(page
)))
457 if (unlikely(destroy_compound_page(page
, order
)))
460 VM_BUG_ON(migratetype
== -1);
462 page_idx
= page_to_pfn(page
) & ((1 << MAX_ORDER
) - 1);
464 VM_BUG_ON(page_idx
& ((1 << order
) - 1));
465 VM_BUG_ON(bad_range(zone
, page
));
467 while (order
< MAX_ORDER
-1) {
468 unsigned long combined_idx
;
471 buddy
= __page_find_buddy(page
, page_idx
, order
);
472 if (!page_is_buddy(page
, buddy
, order
))
475 /* Our buddy is free, merge with it and move up one order. */
476 list_del(&buddy
->lru
);
477 zone
->free_area
[order
].nr_free
--;
478 rmv_page_order(buddy
);
479 combined_idx
= __find_combined_index(page_idx
, order
);
480 page
= page
+ (combined_idx
- page_idx
);
481 page_idx
= combined_idx
;
484 set_page_order(page
, order
);
486 &zone
->free_area
[order
].free_list
[migratetype
]);
487 zone
->free_area
[order
].nr_free
++;
491 * free_page_mlock() -- clean up attempts to free and mlocked() page.
492 * Page should not be on lru, so no need to fix that up.
493 * free_pages_check() will verify...
495 static inline void free_page_mlock(struct page
*page
)
497 __dec_zone_page_state(page
, NR_MLOCK
);
498 __count_vm_event(UNEVICTABLE_MLOCKFREED
);
501 static inline int free_pages_check(struct page
*page
)
503 if (unlikely(page_mapcount(page
) |
504 (page
->mapping
!= NULL
) |
505 (atomic_read(&page
->_count
) != 0) |
506 (page
->flags
& PAGE_FLAGS_CHECK_AT_FREE
))) {
510 if (page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)
511 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
516 * Frees a number of pages from the PCP lists
517 * Assumes all pages on list are in same zone, and of same order.
518 * count is the number of pages to free.
520 * If the zone was previously in an "all pages pinned" state then look to
521 * see if this freeing clears that state.
523 * And clear the zone's pages_scanned counter, to hold off the "all pages are
524 * pinned" detection logic.
526 static void free_pcppages_bulk(struct zone
*zone
, int count
,
527 struct per_cpu_pages
*pcp
)
532 spin_lock(&zone
->lock
);
533 zone_clear_flag(zone
, ZONE_ALL_UNRECLAIMABLE
);
534 zone
->pages_scanned
= 0;
536 __mod_zone_page_state(zone
, NR_FREE_PAGES
, count
);
539 struct list_head
*list
;
542 * Remove pages from lists in a round-robin fashion. A
543 * batch_free count is maintained that is incremented when an
544 * empty list is encountered. This is so more pages are freed
545 * off fuller lists instead of spinning excessively around empty
550 if (++migratetype
== MIGRATE_PCPTYPES
)
552 list
= &pcp
->lists
[migratetype
];
553 } while (list_empty(list
));
556 page
= list_entry(list
->prev
, struct page
, lru
);
557 /* must delete as __free_one_page list manipulates */
558 list_del(&page
->lru
);
559 __free_one_page(page
, zone
, 0, migratetype
);
560 trace_mm_page_pcpu_drain(page
, 0, migratetype
);
561 } while (--count
&& --batch_free
&& !list_empty(list
));
563 spin_unlock(&zone
->lock
);
566 static void free_one_page(struct zone
*zone
, struct page
*page
, int order
,
569 spin_lock(&zone
->lock
);
570 zone_clear_flag(zone
, ZONE_ALL_UNRECLAIMABLE
);
571 zone
->pages_scanned
= 0;
573 __mod_zone_page_state(zone
, NR_FREE_PAGES
, 1 << order
);
574 __free_one_page(page
, zone
, order
, migratetype
);
575 spin_unlock(&zone
->lock
);
578 static void __free_pages_ok(struct page
*page
, unsigned int order
)
583 int wasMlocked
= __TestClearPageMlocked(page
);
585 kmemcheck_free_shadow(page
, order
);
587 for (i
= 0 ; i
< (1 << order
) ; ++i
)
588 bad
+= free_pages_check(page
+ i
);
592 if (!PageHighMem(page
)) {
593 debug_check_no_locks_freed(page_address(page
),PAGE_SIZE
<<order
);
594 debug_check_no_obj_freed(page_address(page
),
597 arch_free_page(page
, order
);
598 kernel_map_pages(page
, 1 << order
, 0);
600 local_irq_save(flags
);
601 if (unlikely(wasMlocked
))
602 free_page_mlock(page
);
603 __count_vm_events(PGFREE
, 1 << order
);
604 free_one_page(page_zone(page
), page
, order
,
605 get_pageblock_migratetype(page
));
606 local_irq_restore(flags
);
610 * permit the bootmem allocator to evade page validation on high-order frees
612 void __meminit
__free_pages_bootmem(struct page
*page
, unsigned int order
)
615 __ClearPageReserved(page
);
616 set_page_count(page
, 0);
617 set_page_refcounted(page
);
623 for (loop
= 0; loop
< BITS_PER_LONG
; loop
++) {
624 struct page
*p
= &page
[loop
];
626 if (loop
+ 1 < BITS_PER_LONG
)
628 __ClearPageReserved(p
);
629 set_page_count(p
, 0);
632 set_page_refcounted(page
);
633 __free_pages(page
, order
);
639 * The order of subdivision here is critical for the IO subsystem.
640 * Please do not alter this order without good reasons and regression
641 * testing. Specifically, as large blocks of memory are subdivided,
642 * the order in which smaller blocks are delivered depends on the order
643 * they're subdivided in this function. This is the primary factor
644 * influencing the order in which pages are delivered to the IO
645 * subsystem according to empirical testing, and this is also justified
646 * by considering the behavior of a buddy system containing a single
647 * large block of memory acted on by a series of small allocations.
648 * This behavior is a critical factor in sglist merging's success.
652 static inline void expand(struct zone
*zone
, struct page
*page
,
653 int low
, int high
, struct free_area
*area
,
656 unsigned long size
= 1 << high
;
662 VM_BUG_ON(bad_range(zone
, &page
[size
]));
663 list_add(&page
[size
].lru
, &area
->free_list
[migratetype
]);
665 set_page_order(&page
[size
], high
);
670 * This page is about to be returned from the page allocator
672 static inline int check_new_page(struct page
*page
)
674 if (unlikely(page_mapcount(page
) |
675 (page
->mapping
!= NULL
) |
676 (atomic_read(&page
->_count
) != 0) |
677 (page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
))) {
684 static int prep_new_page(struct page
*page
, int order
, gfp_t gfp_flags
)
688 for (i
= 0; i
< (1 << order
); i
++) {
689 struct page
*p
= page
+ i
;
690 if (unlikely(check_new_page(p
)))
694 set_page_private(page
, 0);
695 set_page_refcounted(page
);
697 arch_alloc_page(page
, order
);
698 kernel_map_pages(page
, 1 << order
, 1);
700 if (gfp_flags
& __GFP_ZERO
)
701 prep_zero_page(page
, order
, gfp_flags
);
703 if (order
&& (gfp_flags
& __GFP_COMP
))
704 prep_compound_page(page
, order
);
710 * Go through the free lists for the given migratetype and remove
711 * the smallest available page from the freelists
714 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
717 unsigned int current_order
;
718 struct free_area
* area
;
721 /* Find a page of the appropriate size in the preferred list */
722 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
723 area
= &(zone
->free_area
[current_order
]);
724 if (list_empty(&area
->free_list
[migratetype
]))
727 page
= list_entry(area
->free_list
[migratetype
].next
,
729 list_del(&page
->lru
);
730 rmv_page_order(page
);
732 expand(zone
, page
, order
, current_order
, area
, migratetype
);
741 * This array describes the order lists are fallen back to when
742 * the free lists for the desirable migrate type are depleted
744 static int fallbacks
[MIGRATE_TYPES
][MIGRATE_TYPES
-1] = {
745 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
746 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
747 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_RESERVE
},
748 [MIGRATE_RESERVE
] = { MIGRATE_RESERVE
, MIGRATE_RESERVE
, MIGRATE_RESERVE
}, /* Never used */
752 * Move the free pages in a range to the free lists of the requested type.
753 * Note that start_page and end_pages are not aligned on a pageblock
754 * boundary. If alignment is required, use move_freepages_block()
756 static int move_freepages(struct zone
*zone
,
757 struct page
*start_page
, struct page
*end_page
,
764 #ifndef CONFIG_HOLES_IN_ZONE
766 * page_zone is not safe to call in this context when
767 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
768 * anyway as we check zone boundaries in move_freepages_block().
769 * Remove at a later date when no bug reports exist related to
770 * grouping pages by mobility
772 BUG_ON(page_zone(start_page
) != page_zone(end_page
));
775 for (page
= start_page
; page
<= end_page
;) {
776 /* Make sure we are not inadvertently changing nodes */
777 VM_BUG_ON(page_to_nid(page
) != zone_to_nid(zone
));
779 if (!pfn_valid_within(page_to_pfn(page
))) {
784 if (!PageBuddy(page
)) {
789 order
= page_order(page
);
790 list_del(&page
->lru
);
792 &zone
->free_area
[order
].free_list
[migratetype
]);
794 pages_moved
+= 1 << order
;
800 static int move_freepages_block(struct zone
*zone
, struct page
*page
,
803 unsigned long start_pfn
, end_pfn
;
804 struct page
*start_page
, *end_page
;
806 start_pfn
= page_to_pfn(page
);
807 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
808 start_page
= pfn_to_page(start_pfn
);
809 end_page
= start_page
+ pageblock_nr_pages
- 1;
810 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
812 /* Do not cross zone boundaries */
813 if (start_pfn
< zone
->zone_start_pfn
)
815 if (end_pfn
>= zone
->zone_start_pfn
+ zone
->spanned_pages
)
818 return move_freepages(zone
, start_page
, end_page
, migratetype
);
821 static void change_pageblock_range(struct page
*pageblock_page
,
822 int start_order
, int migratetype
)
824 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
826 while (nr_pageblocks
--) {
827 set_pageblock_migratetype(pageblock_page
, migratetype
);
828 pageblock_page
+= pageblock_nr_pages
;
832 /* Remove an element from the buddy allocator from the fallback list */
833 static inline struct page
*
834 __rmqueue_fallback(struct zone
*zone
, int order
, int start_migratetype
)
836 struct free_area
* area
;
841 /* Find the largest possible block of pages in the other list */
842 for (current_order
= MAX_ORDER
-1; current_order
>= order
;
844 for (i
= 0; i
< MIGRATE_TYPES
- 1; i
++) {
845 migratetype
= fallbacks
[start_migratetype
][i
];
847 /* MIGRATE_RESERVE handled later if necessary */
848 if (migratetype
== MIGRATE_RESERVE
)
851 area
= &(zone
->free_area
[current_order
]);
852 if (list_empty(&area
->free_list
[migratetype
]))
855 page
= list_entry(area
->free_list
[migratetype
].next
,
860 * If breaking a large block of pages, move all free
861 * pages to the preferred allocation list. If falling
862 * back for a reclaimable kernel allocation, be more
863 * agressive about taking ownership of free pages
865 if (unlikely(current_order
>= (pageblock_order
>> 1)) ||
866 start_migratetype
== MIGRATE_RECLAIMABLE
||
867 page_group_by_mobility_disabled
) {
869 pages
= move_freepages_block(zone
, page
,
872 /* Claim the whole block if over half of it is free */
873 if (pages
>= (1 << (pageblock_order
-1)) ||
874 page_group_by_mobility_disabled
)
875 set_pageblock_migratetype(page
,
878 migratetype
= start_migratetype
;
881 /* Remove the page from the freelists */
882 list_del(&page
->lru
);
883 rmv_page_order(page
);
885 /* Take ownership for orders >= pageblock_order */
886 if (current_order
>= pageblock_order
)
887 change_pageblock_range(page
, current_order
,
890 expand(zone
, page
, order
, current_order
, area
, migratetype
);
892 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
893 start_migratetype
, migratetype
);
903 * Do the hard work of removing an element from the buddy allocator.
904 * Call me with the zone->lock already held.
906 static struct page
*__rmqueue(struct zone
*zone
, unsigned int order
,
912 page
= __rmqueue_smallest(zone
, order
, migratetype
);
914 if (unlikely(!page
) && migratetype
!= MIGRATE_RESERVE
) {
915 page
= __rmqueue_fallback(zone
, order
, migratetype
);
918 * Use MIGRATE_RESERVE rather than fail an allocation. goto
919 * is used because __rmqueue_smallest is an inline function
920 * and we want just one call site
923 migratetype
= MIGRATE_RESERVE
;
928 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
933 * Obtain a specified number of elements from the buddy allocator, all under
934 * a single hold of the lock, for efficiency. Add them to the supplied list.
935 * Returns the number of new pages which were placed at *list.
937 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
938 unsigned long count
, struct list_head
*list
,
939 int migratetype
, int cold
)
943 spin_lock(&zone
->lock
);
944 for (i
= 0; i
< count
; ++i
) {
945 struct page
*page
= __rmqueue(zone
, order
, migratetype
);
946 if (unlikely(page
== NULL
))
950 * Split buddy pages returned by expand() are received here
951 * in physical page order. The page is added to the callers and
952 * list and the list head then moves forward. From the callers
953 * perspective, the linked list is ordered by page number in
954 * some conditions. This is useful for IO devices that can
955 * merge IO requests if the physical pages are ordered
958 if (likely(cold
== 0))
959 list_add(&page
->lru
, list
);
961 list_add_tail(&page
->lru
, list
);
962 set_page_private(page
, migratetype
);
965 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
966 spin_unlock(&zone
->lock
);
972 * Called from the vmstat counter updater to drain pagesets of this
973 * currently executing processor on remote nodes after they have
976 * Note that this function must be called with the thread pinned to
977 * a single processor.
979 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
984 local_irq_save(flags
);
985 if (pcp
->count
>= pcp
->batch
)
986 to_drain
= pcp
->batch
;
988 to_drain
= pcp
->count
;
989 free_pcppages_bulk(zone
, to_drain
, pcp
);
990 pcp
->count
-= to_drain
;
991 local_irq_restore(flags
);
996 * Drain pages of the indicated processor.
998 * The processor must either be the current processor and the
999 * thread pinned to the current processor or a processor that
1002 static void drain_pages(unsigned int cpu
)
1004 unsigned long flags
;
1007 for_each_populated_zone(zone
) {
1008 struct per_cpu_pageset
*pset
;
1009 struct per_cpu_pages
*pcp
;
1011 pset
= zone_pcp(zone
, cpu
);
1014 local_irq_save(flags
);
1015 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
1017 local_irq_restore(flags
);
1022 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1024 void drain_local_pages(void *arg
)
1026 drain_pages(smp_processor_id());
1030 * Spill all the per-cpu pages from all CPUs back into the buddy allocator
1032 void drain_all_pages(void)
1034 on_each_cpu(drain_local_pages
, NULL
, 1);
1037 #ifdef CONFIG_HIBERNATION
1039 void mark_free_pages(struct zone
*zone
)
1041 unsigned long pfn
, max_zone_pfn
;
1042 unsigned long flags
;
1044 struct list_head
*curr
;
1046 if (!zone
->spanned_pages
)
1049 spin_lock_irqsave(&zone
->lock
, flags
);
1051 max_zone_pfn
= zone
->zone_start_pfn
+ zone
->spanned_pages
;
1052 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
1053 if (pfn_valid(pfn
)) {
1054 struct page
*page
= pfn_to_page(pfn
);
1056 if (!swsusp_page_is_forbidden(page
))
1057 swsusp_unset_page_free(page
);
1060 for_each_migratetype_order(order
, t
) {
1061 list_for_each(curr
, &zone
->free_area
[order
].free_list
[t
]) {
1064 pfn
= page_to_pfn(list_entry(curr
, struct page
, lru
));
1065 for (i
= 0; i
< (1UL << order
); i
++)
1066 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
1069 spin_unlock_irqrestore(&zone
->lock
, flags
);
1071 #endif /* CONFIG_PM */
1074 * Free a 0-order page
1076 static void free_hot_cold_page(struct page
*page
, int cold
)
1078 struct zone
*zone
= page_zone(page
);
1079 struct per_cpu_pages
*pcp
;
1080 unsigned long flags
;
1082 int wasMlocked
= __TestClearPageMlocked(page
);
1084 kmemcheck_free_shadow(page
, 0);
1087 page
->mapping
= NULL
;
1088 if (free_pages_check(page
))
1091 if (!PageHighMem(page
)) {
1092 debug_check_no_locks_freed(page_address(page
), PAGE_SIZE
);
1093 debug_check_no_obj_freed(page_address(page
), PAGE_SIZE
);
1095 arch_free_page(page
, 0);
1096 kernel_map_pages(page
, 1, 0);
1098 pcp
= &zone_pcp(zone
, get_cpu())->pcp
;
1099 migratetype
= get_pageblock_migratetype(page
);
1100 set_page_private(page
, migratetype
);
1101 local_irq_save(flags
);
1102 if (unlikely(wasMlocked
))
1103 free_page_mlock(page
);
1104 __count_vm_event(PGFREE
);
1107 * We only track unmovable, reclaimable and movable on pcp lists.
1108 * Free ISOLATE pages back to the allocator because they are being
1109 * offlined but treat RESERVE as movable pages so we can get those
1110 * areas back if necessary. Otherwise, we may have to free
1111 * excessively into the page allocator
1113 if (migratetype
>= MIGRATE_PCPTYPES
) {
1114 if (unlikely(migratetype
== MIGRATE_ISOLATE
)) {
1115 free_one_page(zone
, page
, 0, migratetype
);
1118 migratetype
= MIGRATE_MOVABLE
;
1122 list_add_tail(&page
->lru
, &pcp
->lists
[migratetype
]);
1124 list_add(&page
->lru
, &pcp
->lists
[migratetype
]);
1126 if (pcp
->count
>= pcp
->high
) {
1127 free_pcppages_bulk(zone
, pcp
->batch
, pcp
);
1128 pcp
->count
-= pcp
->batch
;
1132 local_irq_restore(flags
);
1136 void free_hot_page(struct page
*page
)
1138 trace_mm_page_free_direct(page
, 0);
1139 free_hot_cold_page(page
, 0);
1143 * split_page takes a non-compound higher-order page, and splits it into
1144 * n (1<<order) sub-pages: page[0..n]
1145 * Each sub-page must be freed individually.
1147 * Note: this is probably too low level an operation for use in drivers.
1148 * Please consult with lkml before using this in your driver.
1150 void split_page(struct page
*page
, unsigned int order
)
1154 VM_BUG_ON(PageCompound(page
));
1155 VM_BUG_ON(!page_count(page
));
1157 #ifdef CONFIG_KMEMCHECK
1159 * Split shadow pages too, because free(page[0]) would
1160 * otherwise free the whole shadow.
1162 if (kmemcheck_page_is_tracked(page
))
1163 split_page(virt_to_page(page
[0].shadow
), order
);
1166 for (i
= 1; i
< (1 << order
); i
++)
1167 set_page_refcounted(page
+ i
);
1171 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1172 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1176 struct page
*buffered_rmqueue(struct zone
*preferred_zone
,
1177 struct zone
*zone
, int order
, gfp_t gfp_flags
,
1180 unsigned long flags
;
1182 int cold
= !!(gfp_flags
& __GFP_COLD
);
1187 if (likely(order
== 0)) {
1188 struct per_cpu_pages
*pcp
;
1189 struct list_head
*list
;
1191 pcp
= &zone_pcp(zone
, cpu
)->pcp
;
1192 list
= &pcp
->lists
[migratetype
];
1193 local_irq_save(flags
);
1194 if (list_empty(list
)) {
1195 pcp
->count
+= rmqueue_bulk(zone
, 0,
1198 if (unlikely(list_empty(list
)))
1203 page
= list_entry(list
->prev
, struct page
, lru
);
1205 page
= list_entry(list
->next
, struct page
, lru
);
1207 list_del(&page
->lru
);
1210 if (unlikely(gfp_flags
& __GFP_NOFAIL
)) {
1212 * __GFP_NOFAIL is not to be used in new code.
1214 * All __GFP_NOFAIL callers should be fixed so that they
1215 * properly detect and handle allocation failures.
1217 * We most definitely don't want callers attempting to
1218 * allocate greater than order-1 page units with
1221 WARN_ON_ONCE(order
> 1);
1223 spin_lock_irqsave(&zone
->lock
, flags
);
1224 page
= __rmqueue(zone
, order
, migratetype
);
1225 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(1 << order
));
1226 spin_unlock(&zone
->lock
);
1231 __count_zone_vm_events(PGALLOC
, zone
, 1 << order
);
1232 zone_statistics(preferred_zone
, zone
);
1233 local_irq_restore(flags
);
1236 VM_BUG_ON(bad_range(zone
, page
));
1237 if (prep_new_page(page
, order
, gfp_flags
))
1242 local_irq_restore(flags
);
1247 /* The ALLOC_WMARK bits are used as an index to zone->watermark */
1248 #define ALLOC_WMARK_MIN WMARK_MIN
1249 #define ALLOC_WMARK_LOW WMARK_LOW
1250 #define ALLOC_WMARK_HIGH WMARK_HIGH
1251 #define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */
1253 /* Mask to get the watermark bits */
1254 #define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1)
1256 #define ALLOC_HARDER 0x10 /* try to alloc harder */
1257 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
1258 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
1260 #ifdef CONFIG_FAIL_PAGE_ALLOC
1262 static struct fail_page_alloc_attr
{
1263 struct fault_attr attr
;
1265 u32 ignore_gfp_highmem
;
1266 u32 ignore_gfp_wait
;
1269 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1271 struct dentry
*ignore_gfp_highmem_file
;
1272 struct dentry
*ignore_gfp_wait_file
;
1273 struct dentry
*min_order_file
;
1275 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1277 } fail_page_alloc
= {
1278 .attr
= FAULT_ATTR_INITIALIZER
,
1279 .ignore_gfp_wait
= 1,
1280 .ignore_gfp_highmem
= 1,
1284 static int __init
setup_fail_page_alloc(char *str
)
1286 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
1288 __setup("fail_page_alloc=", setup_fail_page_alloc
);
1290 static int should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1292 if (order
< fail_page_alloc
.min_order
)
1294 if (gfp_mask
& __GFP_NOFAIL
)
1296 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
1298 if (fail_page_alloc
.ignore_gfp_wait
&& (gfp_mask
& __GFP_WAIT
))
1301 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
1304 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1306 static int __init
fail_page_alloc_debugfs(void)
1308 mode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
1312 err
= init_fault_attr_dentries(&fail_page_alloc
.attr
,
1316 dir
= fail_page_alloc
.attr
.dentries
.dir
;
1318 fail_page_alloc
.ignore_gfp_wait_file
=
1319 debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
1320 &fail_page_alloc
.ignore_gfp_wait
);
1322 fail_page_alloc
.ignore_gfp_highmem_file
=
1323 debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
1324 &fail_page_alloc
.ignore_gfp_highmem
);
1325 fail_page_alloc
.min_order_file
=
1326 debugfs_create_u32("min-order", mode
, dir
,
1327 &fail_page_alloc
.min_order
);
1329 if (!fail_page_alloc
.ignore_gfp_wait_file
||
1330 !fail_page_alloc
.ignore_gfp_highmem_file
||
1331 !fail_page_alloc
.min_order_file
) {
1333 debugfs_remove(fail_page_alloc
.ignore_gfp_wait_file
);
1334 debugfs_remove(fail_page_alloc
.ignore_gfp_highmem_file
);
1335 debugfs_remove(fail_page_alloc
.min_order_file
);
1336 cleanup_fault_attr_dentries(&fail_page_alloc
.attr
);
1342 late_initcall(fail_page_alloc_debugfs
);
1344 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1346 #else /* CONFIG_FAIL_PAGE_ALLOC */
1348 static inline int should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1353 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1356 * Return 1 if free pages are above 'mark'. This takes into account the order
1357 * of the allocation.
1359 int zone_watermark_ok(struct zone
*z
, int order
, unsigned long mark
,
1360 int classzone_idx
, int alloc_flags
)
1362 /* free_pages my go negative - that's OK */
1364 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
) - (1 << order
) + 1;
1367 if (alloc_flags
& ALLOC_HIGH
)
1369 if (alloc_flags
& ALLOC_HARDER
)
1372 if (free_pages
<= min
+ z
->lowmem_reserve
[classzone_idx
])
1374 for (o
= 0; o
< order
; o
++) {
1375 /* At the next order, this order's pages become unavailable */
1376 free_pages
-= z
->free_area
[o
].nr_free
<< o
;
1378 /* Require fewer higher order pages to be free */
1381 if (free_pages
<= min
)
1389 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1390 * skip over zones that are not allowed by the cpuset, or that have
1391 * been recently (in last second) found to be nearly full. See further
1392 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1393 * that have to skip over a lot of full or unallowed zones.
1395 * If the zonelist cache is present in the passed in zonelist, then
1396 * returns a pointer to the allowed node mask (either the current
1397 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1399 * If the zonelist cache is not available for this zonelist, does
1400 * nothing and returns NULL.
1402 * If the fullzones BITMAP in the zonelist cache is stale (more than
1403 * a second since last zap'd) then we zap it out (clear its bits.)
1405 * We hold off even calling zlc_setup, until after we've checked the
1406 * first zone in the zonelist, on the theory that most allocations will
1407 * be satisfied from that first zone, so best to examine that zone as
1408 * quickly as we can.
1410 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1412 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1413 nodemask_t
*allowednodes
; /* zonelist_cache approximation */
1415 zlc
= zonelist
->zlcache_ptr
;
1419 if (time_after(jiffies
, zlc
->last_full_zap
+ HZ
)) {
1420 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
1421 zlc
->last_full_zap
= jiffies
;
1424 allowednodes
= !in_interrupt() && (alloc_flags
& ALLOC_CPUSET
) ?
1425 &cpuset_current_mems_allowed
:
1426 &node_states
[N_HIGH_MEMORY
];
1427 return allowednodes
;
1431 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1432 * if it is worth looking at further for free memory:
1433 * 1) Check that the zone isn't thought to be full (doesn't have its
1434 * bit set in the zonelist_cache fullzones BITMAP).
1435 * 2) Check that the zones node (obtained from the zonelist_cache
1436 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1437 * Return true (non-zero) if zone is worth looking at further, or
1438 * else return false (zero) if it is not.
1440 * This check -ignores- the distinction between various watermarks,
1441 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1442 * found to be full for any variation of these watermarks, it will
1443 * be considered full for up to one second by all requests, unless
1444 * we are so low on memory on all allowed nodes that we are forced
1445 * into the second scan of the zonelist.
1447 * In the second scan we ignore this zonelist cache and exactly
1448 * apply the watermarks to all zones, even it is slower to do so.
1449 * We are low on memory in the second scan, and should leave no stone
1450 * unturned looking for a free page.
1452 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1453 nodemask_t
*allowednodes
)
1455 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1456 int i
; /* index of *z in zonelist zones */
1457 int n
; /* node that zone *z is on */
1459 zlc
= zonelist
->zlcache_ptr
;
1463 i
= z
- zonelist
->_zonerefs
;
1466 /* This zone is worth trying if it is allowed but not full */
1467 return node_isset(n
, *allowednodes
) && !test_bit(i
, zlc
->fullzones
);
1471 * Given 'z' scanning a zonelist, set the corresponding bit in
1472 * zlc->fullzones, so that subsequent attempts to allocate a page
1473 * from that zone don't waste time re-examining it.
1475 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1477 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1478 int i
; /* index of *z in zonelist zones */
1480 zlc
= zonelist
->zlcache_ptr
;
1484 i
= z
- zonelist
->_zonerefs
;
1486 set_bit(i
, zlc
->fullzones
);
1489 #else /* CONFIG_NUMA */
1491 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1496 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1497 nodemask_t
*allowednodes
)
1502 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1505 #endif /* CONFIG_NUMA */
1508 * get_page_from_freelist goes through the zonelist trying to allocate
1511 static struct page
*
1512 get_page_from_freelist(gfp_t gfp_mask
, nodemask_t
*nodemask
, unsigned int order
,
1513 struct zonelist
*zonelist
, int high_zoneidx
, int alloc_flags
,
1514 struct zone
*preferred_zone
, int migratetype
)
1517 struct page
*page
= NULL
;
1520 nodemask_t
*allowednodes
= NULL
;/* zonelist_cache approximation */
1521 int zlc_active
= 0; /* set if using zonelist_cache */
1522 int did_zlc_setup
= 0; /* just call zlc_setup() one time */
1524 classzone_idx
= zone_idx(preferred_zone
);
1527 * Scan zonelist, looking for a zone with enough free.
1528 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1530 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
1531 high_zoneidx
, nodemask
) {
1532 if (NUMA_BUILD
&& zlc_active
&&
1533 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
1535 if ((alloc_flags
& ALLOC_CPUSET
) &&
1536 !cpuset_zone_allowed_softwall(zone
, gfp_mask
))
1539 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
1540 if (!(alloc_flags
& ALLOC_NO_WATERMARKS
)) {
1544 mark
= zone
->watermark
[alloc_flags
& ALLOC_WMARK_MASK
];
1545 if (zone_watermark_ok(zone
, order
, mark
,
1546 classzone_idx
, alloc_flags
))
1549 if (zone_reclaim_mode
== 0)
1550 goto this_zone_full
;
1552 ret
= zone_reclaim(zone
, gfp_mask
, order
);
1554 case ZONE_RECLAIM_NOSCAN
:
1557 case ZONE_RECLAIM_FULL
:
1558 /* scanned but unreclaimable */
1559 goto this_zone_full
;
1561 /* did we reclaim enough */
1562 if (!zone_watermark_ok(zone
, order
, mark
,
1563 classzone_idx
, alloc_flags
))
1564 goto this_zone_full
;
1569 page
= buffered_rmqueue(preferred_zone
, zone
, order
,
1570 gfp_mask
, migratetype
);
1575 zlc_mark_zone_full(zonelist
, z
);
1577 if (NUMA_BUILD
&& !did_zlc_setup
&& nr_online_nodes
> 1) {
1579 * we do zlc_setup after the first zone is tried but only
1580 * if there are multiple nodes make it worthwhile
1582 allowednodes
= zlc_setup(zonelist
, alloc_flags
);
1588 if (unlikely(NUMA_BUILD
&& page
== NULL
&& zlc_active
)) {
1589 /* Disable zlc cache for second zonelist scan */
1597 should_alloc_retry(gfp_t gfp_mask
, unsigned int order
,
1598 unsigned long pages_reclaimed
)
1600 /* Do not loop if specifically requested */
1601 if (gfp_mask
& __GFP_NORETRY
)
1605 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
1606 * means __GFP_NOFAIL, but that may not be true in other
1609 if (order
<= PAGE_ALLOC_COSTLY_ORDER
)
1613 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
1614 * specified, then we retry until we no longer reclaim any pages
1615 * (above), or we've reclaimed an order of pages at least as
1616 * large as the allocation's order. In both cases, if the
1617 * allocation still fails, we stop retrying.
1619 if (gfp_mask
& __GFP_REPEAT
&& pages_reclaimed
< (1 << order
))
1623 * Don't let big-order allocations loop unless the caller
1624 * explicitly requests that.
1626 if (gfp_mask
& __GFP_NOFAIL
)
1632 static inline struct page
*
1633 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
1634 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1635 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
1640 /* Acquire the OOM killer lock for the zones in zonelist */
1641 if (!try_set_zone_oom(zonelist
, gfp_mask
)) {
1642 schedule_timeout_uninterruptible(1);
1647 * Go through the zonelist yet one more time, keep very high watermark
1648 * here, this is only to catch a parallel oom killing, we must fail if
1649 * we're still under heavy pressure.
1651 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
,
1652 order
, zonelist
, high_zoneidx
,
1653 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
,
1654 preferred_zone
, migratetype
);
1658 if (!(gfp_mask
& __GFP_NOFAIL
)) {
1659 /* The OOM killer will not help higher order allocs */
1660 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
1663 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
1664 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
1665 * The caller should handle page allocation failure by itself if
1666 * it specifies __GFP_THISNODE.
1667 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
1669 if (gfp_mask
& __GFP_THISNODE
)
1672 /* Exhausted what can be done so it's blamo time */
1673 out_of_memory(zonelist
, gfp_mask
, order
, nodemask
);
1676 clear_zonelist_oom(zonelist
, gfp_mask
);
1680 /* The really slow allocator path where we enter direct reclaim */
1681 static inline struct page
*
1682 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
1683 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1684 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
1685 int migratetype
, unsigned long *did_some_progress
)
1687 struct page
*page
= NULL
;
1688 struct reclaim_state reclaim_state
;
1689 struct task_struct
*p
= current
;
1693 /* We now go into synchronous reclaim */
1694 cpuset_memory_pressure_bump();
1695 p
->flags
|= PF_MEMALLOC
;
1696 lockdep_set_current_reclaim_state(gfp_mask
);
1697 reclaim_state
.reclaimed_slab
= 0;
1698 p
->reclaim_state
= &reclaim_state
;
1700 *did_some_progress
= try_to_free_pages(zonelist
, order
, gfp_mask
, nodemask
);
1702 p
->reclaim_state
= NULL
;
1703 lockdep_clear_current_reclaim_state();
1704 p
->flags
&= ~PF_MEMALLOC
;
1711 if (likely(*did_some_progress
))
1712 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
1713 zonelist
, high_zoneidx
,
1714 alloc_flags
, preferred_zone
,
1720 * This is called in the allocator slow-path if the allocation request is of
1721 * sufficient urgency to ignore watermarks and take other desperate measures
1723 static inline struct page
*
1724 __alloc_pages_high_priority(gfp_t gfp_mask
, unsigned int order
,
1725 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1726 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
1732 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
1733 zonelist
, high_zoneidx
, ALLOC_NO_WATERMARKS
,
1734 preferred_zone
, migratetype
);
1736 if (!page
&& gfp_mask
& __GFP_NOFAIL
)
1737 congestion_wait(BLK_RW_ASYNC
, HZ
/50);
1738 } while (!page
&& (gfp_mask
& __GFP_NOFAIL
));
1744 void wake_all_kswapd(unsigned int order
, struct zonelist
*zonelist
,
1745 enum zone_type high_zoneidx
)
1750 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
)
1751 wakeup_kswapd(zone
, order
);
1755 gfp_to_alloc_flags(gfp_t gfp_mask
)
1757 struct task_struct
*p
= current
;
1758 int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
1759 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
1761 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
1762 BUILD_BUG_ON(__GFP_HIGH
!= ALLOC_HIGH
);
1765 * The caller may dip into page reserves a bit more if the caller
1766 * cannot run direct reclaim, or if the caller has realtime scheduling
1767 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
1768 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
1770 alloc_flags
|= (gfp_mask
& __GFP_HIGH
);
1773 alloc_flags
|= ALLOC_HARDER
;
1775 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
1776 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1778 alloc_flags
&= ~ALLOC_CPUSET
;
1779 } else if (unlikely(rt_task(p
)) && !in_interrupt())
1780 alloc_flags
|= ALLOC_HARDER
;
1782 if (likely(!(gfp_mask
& __GFP_NOMEMALLOC
))) {
1783 if (!in_interrupt() &&
1784 ((p
->flags
& PF_MEMALLOC
) ||
1785 unlikely(test_thread_flag(TIF_MEMDIE
))))
1786 alloc_flags
|= ALLOC_NO_WATERMARKS
;
1792 static inline struct page
*
1793 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
1794 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1795 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
1798 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
1799 struct page
*page
= NULL
;
1801 unsigned long pages_reclaimed
= 0;
1802 unsigned long did_some_progress
;
1803 struct task_struct
*p
= current
;
1806 * In the slowpath, we sanity check order to avoid ever trying to
1807 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
1808 * be using allocators in order of preference for an area that is
1811 if (order
>= MAX_ORDER
) {
1812 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
1817 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
1818 * __GFP_NOWARN set) should not cause reclaim since the subsystem
1819 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
1820 * using a larger set of nodes after it has established that the
1821 * allowed per node queues are empty and that nodes are
1824 if (NUMA_BUILD
&& (gfp_mask
& GFP_THISNODE
) == GFP_THISNODE
)
1828 wake_all_kswapd(order
, zonelist
, high_zoneidx
);
1831 * OK, we're below the kswapd watermark and have kicked background
1832 * reclaim. Now things get more complex, so set up alloc_flags according
1833 * to how we want to proceed.
1835 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
1837 /* This is the last chance, in general, before the goto nopage. */
1838 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
, zonelist
,
1839 high_zoneidx
, alloc_flags
& ~ALLOC_NO_WATERMARKS
,
1840 preferred_zone
, migratetype
);
1845 /* Allocate without watermarks if the context allows */
1846 if (alloc_flags
& ALLOC_NO_WATERMARKS
) {
1847 page
= __alloc_pages_high_priority(gfp_mask
, order
,
1848 zonelist
, high_zoneidx
, nodemask
,
1849 preferred_zone
, migratetype
);
1854 /* Atomic allocations - we can't balance anything */
1858 /* Avoid recursion of direct reclaim */
1859 if (p
->flags
& PF_MEMALLOC
)
1862 /* Avoid allocations with no watermarks from looping endlessly */
1863 if (test_thread_flag(TIF_MEMDIE
) && !(gfp_mask
& __GFP_NOFAIL
))
1866 /* Try direct reclaim and then allocating */
1867 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
,
1868 zonelist
, high_zoneidx
,
1870 alloc_flags
, preferred_zone
,
1871 migratetype
, &did_some_progress
);
1876 * If we failed to make any progress reclaiming, then we are
1877 * running out of options and have to consider going OOM
1879 if (!did_some_progress
) {
1880 if ((gfp_mask
& __GFP_FS
) && !(gfp_mask
& __GFP_NORETRY
)) {
1881 if (oom_killer_disabled
)
1883 page
= __alloc_pages_may_oom(gfp_mask
, order
,
1884 zonelist
, high_zoneidx
,
1885 nodemask
, preferred_zone
,
1891 * The OOM killer does not trigger for high-order
1892 * ~__GFP_NOFAIL allocations so if no progress is being
1893 * made, there are no other options and retrying is
1896 if (order
> PAGE_ALLOC_COSTLY_ORDER
&&
1897 !(gfp_mask
& __GFP_NOFAIL
))
1904 /* Check if we should retry the allocation */
1905 pages_reclaimed
+= did_some_progress
;
1906 if (should_alloc_retry(gfp_mask
, order
, pages_reclaimed
)) {
1907 /* Wait for some write requests to complete then retry */
1908 congestion_wait(BLK_RW_ASYNC
, HZ
/50);
1913 if (!(gfp_mask
& __GFP_NOWARN
) && printk_ratelimit()) {
1914 printk(KERN_WARNING
"%s: page allocation failure."
1915 " order:%d, mode:0x%x\n",
1916 p
->comm
, order
, gfp_mask
);
1922 if (kmemcheck_enabled
)
1923 kmemcheck_pagealloc_alloc(page
, order
, gfp_mask
);
1929 * This is the 'heart' of the zoned buddy allocator.
1932 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
,
1933 struct zonelist
*zonelist
, nodemask_t
*nodemask
)
1935 enum zone_type high_zoneidx
= gfp_zone(gfp_mask
);
1936 struct zone
*preferred_zone
;
1938 int migratetype
= allocflags_to_migratetype(gfp_mask
);
1940 gfp_mask
&= gfp_allowed_mask
;
1942 lockdep_trace_alloc(gfp_mask
);
1944 might_sleep_if(gfp_mask
& __GFP_WAIT
);
1946 if (should_fail_alloc_page(gfp_mask
, order
))
1950 * Check the zones suitable for the gfp_mask contain at least one
1951 * valid zone. It's possible to have an empty zonelist as a result
1952 * of GFP_THISNODE and a memoryless node
1954 if (unlikely(!zonelist
->_zonerefs
->zone
))
1957 /* The preferred zone is used for statistics later */
1958 first_zones_zonelist(zonelist
, high_zoneidx
, nodemask
, &preferred_zone
);
1959 if (!preferred_zone
)
1962 /* First allocation attempt */
1963 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
, order
,
1964 zonelist
, high_zoneidx
, ALLOC_WMARK_LOW
|ALLOC_CPUSET
,
1965 preferred_zone
, migratetype
);
1966 if (unlikely(!page
))
1967 page
= __alloc_pages_slowpath(gfp_mask
, order
,
1968 zonelist
, high_zoneidx
, nodemask
,
1969 preferred_zone
, migratetype
);
1971 trace_mm_page_alloc(page
, order
, gfp_mask
, migratetype
);
1974 EXPORT_SYMBOL(__alloc_pages_nodemask
);
1977 * Common helper functions.
1979 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
1984 * __get_free_pages() returns a 32-bit address, which cannot represent
1987 VM_BUG_ON((gfp_mask
& __GFP_HIGHMEM
) != 0);
1989 page
= alloc_pages(gfp_mask
, order
);
1992 return (unsigned long) page_address(page
);
1994 EXPORT_SYMBOL(__get_free_pages
);
1996 unsigned long get_zeroed_page(gfp_t gfp_mask
)
1998 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
2000 EXPORT_SYMBOL(get_zeroed_page
);
2002 void __pagevec_free(struct pagevec
*pvec
)
2004 int i
= pagevec_count(pvec
);
2007 trace_mm_pagevec_free(pvec
->pages
[i
], pvec
->cold
);
2008 free_hot_cold_page(pvec
->pages
[i
], pvec
->cold
);
2012 void __free_pages(struct page
*page
, unsigned int order
)
2014 if (put_page_testzero(page
)) {
2015 trace_mm_page_free_direct(page
, order
);
2017 free_hot_page(page
);
2019 __free_pages_ok(page
, order
);
2023 EXPORT_SYMBOL(__free_pages
);
2025 void free_pages(unsigned long addr
, unsigned int order
)
2028 VM_BUG_ON(!virt_addr_valid((void *)addr
));
2029 __free_pages(virt_to_page((void *)addr
), order
);
2033 EXPORT_SYMBOL(free_pages
);
2036 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2037 * @size: the number of bytes to allocate
2038 * @gfp_mask: GFP flags for the allocation
2040 * This function is similar to alloc_pages(), except that it allocates the
2041 * minimum number of pages to satisfy the request. alloc_pages() can only
2042 * allocate memory in power-of-two pages.
2044 * This function is also limited by MAX_ORDER.
2046 * Memory allocated by this function must be released by free_pages_exact().
2048 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
2050 unsigned int order
= get_order(size
);
2053 addr
= __get_free_pages(gfp_mask
, order
);
2055 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
2056 unsigned long used
= addr
+ PAGE_ALIGN(size
);
2058 split_page(virt_to_page((void *)addr
), order
);
2059 while (used
< alloc_end
) {
2065 return (void *)addr
;
2067 EXPORT_SYMBOL(alloc_pages_exact
);
2070 * free_pages_exact - release memory allocated via alloc_pages_exact()
2071 * @virt: the value returned by alloc_pages_exact.
2072 * @size: size of allocation, same value as passed to alloc_pages_exact().
2074 * Release the memory allocated by a previous call to alloc_pages_exact.
2076 void free_pages_exact(void *virt
, size_t size
)
2078 unsigned long addr
= (unsigned long)virt
;
2079 unsigned long end
= addr
+ PAGE_ALIGN(size
);
2081 while (addr
< end
) {
2086 EXPORT_SYMBOL(free_pages_exact
);
2088 static unsigned int nr_free_zone_pages(int offset
)
2093 /* Just pick one node, since fallback list is circular */
2094 unsigned int sum
= 0;
2096 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
2098 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
2099 unsigned long size
= zone
->present_pages
;
2100 unsigned long high
= high_wmark_pages(zone
);
2109 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2111 unsigned int nr_free_buffer_pages(void)
2113 return nr_free_zone_pages(gfp_zone(GFP_USER
));
2115 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
2118 * Amount of free RAM allocatable within all zones
2120 unsigned int nr_free_pagecache_pages(void)
2122 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
2125 static inline void show_node(struct zone
*zone
)
2128 printk("Node %d ", zone_to_nid(zone
));
2131 void si_meminfo(struct sysinfo
*val
)
2133 val
->totalram
= totalram_pages
;
2135 val
->freeram
= global_page_state(NR_FREE_PAGES
);
2136 val
->bufferram
= nr_blockdev_pages();
2137 val
->totalhigh
= totalhigh_pages
;
2138 val
->freehigh
= nr_free_highpages();
2139 val
->mem_unit
= PAGE_SIZE
;
2142 EXPORT_SYMBOL(si_meminfo
);
2145 void si_meminfo_node(struct sysinfo
*val
, int nid
)
2147 pg_data_t
*pgdat
= NODE_DATA(nid
);
2149 val
->totalram
= pgdat
->node_present_pages
;
2150 val
->freeram
= node_page_state(nid
, NR_FREE_PAGES
);
2151 #ifdef CONFIG_HIGHMEM
2152 val
->totalhigh
= pgdat
->node_zones
[ZONE_HIGHMEM
].present_pages
;
2153 val
->freehigh
= zone_page_state(&pgdat
->node_zones
[ZONE_HIGHMEM
],
2159 val
->mem_unit
= PAGE_SIZE
;
2163 #define K(x) ((x) << (PAGE_SHIFT-10))
2166 * Show free area list (used inside shift_scroll-lock stuff)
2167 * We also calculate the percentage fragmentation. We do this by counting the
2168 * memory on each free list with the exception of the first item on the list.
2170 void show_free_areas(void)
2175 for_each_populated_zone(zone
) {
2177 printk("%s per-cpu:\n", zone
->name
);
2179 for_each_online_cpu(cpu
) {
2180 struct per_cpu_pageset
*pageset
;
2182 pageset
= zone_pcp(zone
, cpu
);
2184 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2185 cpu
, pageset
->pcp
.high
,
2186 pageset
->pcp
.batch
, pageset
->pcp
.count
);
2190 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2191 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2193 " dirty:%lu writeback:%lu unstable:%lu\n"
2194 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2195 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n",
2196 global_page_state(NR_ACTIVE_ANON
),
2197 global_page_state(NR_INACTIVE_ANON
),
2198 global_page_state(NR_ISOLATED_ANON
),
2199 global_page_state(NR_ACTIVE_FILE
),
2200 global_page_state(NR_INACTIVE_FILE
),
2201 global_page_state(NR_ISOLATED_FILE
),
2202 global_page_state(NR_UNEVICTABLE
),
2203 global_page_state(NR_FILE_DIRTY
),
2204 global_page_state(NR_WRITEBACK
),
2205 global_page_state(NR_UNSTABLE_NFS
),
2206 global_page_state(NR_FREE_PAGES
),
2207 global_page_state(NR_SLAB_RECLAIMABLE
),
2208 global_page_state(NR_SLAB_UNRECLAIMABLE
),
2209 global_page_state(NR_FILE_MAPPED
),
2210 global_page_state(NR_SHMEM
),
2211 global_page_state(NR_PAGETABLE
),
2212 global_page_state(NR_BOUNCE
));
2214 for_each_populated_zone(zone
) {
2223 " active_anon:%lukB"
2224 " inactive_anon:%lukB"
2225 " active_file:%lukB"
2226 " inactive_file:%lukB"
2227 " unevictable:%lukB"
2228 " isolated(anon):%lukB"
2229 " isolated(file):%lukB"
2236 " slab_reclaimable:%lukB"
2237 " slab_unreclaimable:%lukB"
2238 " kernel_stack:%lukB"
2242 " writeback_tmp:%lukB"
2243 " pages_scanned:%lu"
2244 " all_unreclaimable? %s"
2247 K(zone_page_state(zone
, NR_FREE_PAGES
)),
2248 K(min_wmark_pages(zone
)),
2249 K(low_wmark_pages(zone
)),
2250 K(high_wmark_pages(zone
)),
2251 K(zone_page_state(zone
, NR_ACTIVE_ANON
)),
2252 K(zone_page_state(zone
, NR_INACTIVE_ANON
)),
2253 K(zone_page_state(zone
, NR_ACTIVE_FILE
)),
2254 K(zone_page_state(zone
, NR_INACTIVE_FILE
)),
2255 K(zone_page_state(zone
, NR_UNEVICTABLE
)),
2256 K(zone_page_state(zone
, NR_ISOLATED_ANON
)),
2257 K(zone_page_state(zone
, NR_ISOLATED_FILE
)),
2258 K(zone
->present_pages
),
2259 K(zone_page_state(zone
, NR_MLOCK
)),
2260 K(zone_page_state(zone
, NR_FILE_DIRTY
)),
2261 K(zone_page_state(zone
, NR_WRITEBACK
)),
2262 K(zone_page_state(zone
, NR_FILE_MAPPED
)),
2263 K(zone_page_state(zone
, NR_SHMEM
)),
2264 K(zone_page_state(zone
, NR_SLAB_RECLAIMABLE
)),
2265 K(zone_page_state(zone
, NR_SLAB_UNRECLAIMABLE
)),
2266 zone_page_state(zone
, NR_KERNEL_STACK
) *
2268 K(zone_page_state(zone
, NR_PAGETABLE
)),
2269 K(zone_page_state(zone
, NR_UNSTABLE_NFS
)),
2270 K(zone_page_state(zone
, NR_BOUNCE
)),
2271 K(zone_page_state(zone
, NR_WRITEBACK_TEMP
)),
2272 zone
->pages_scanned
,
2273 (zone_is_all_unreclaimable(zone
) ? "yes" : "no")
2275 printk("lowmem_reserve[]:");
2276 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
2277 printk(" %lu", zone
->lowmem_reserve
[i
]);
2281 for_each_populated_zone(zone
) {
2282 unsigned long nr
[MAX_ORDER
], flags
, order
, total
= 0;
2285 printk("%s: ", zone
->name
);
2287 spin_lock_irqsave(&zone
->lock
, flags
);
2288 for (order
= 0; order
< MAX_ORDER
; order
++) {
2289 nr
[order
] = zone
->free_area
[order
].nr_free
;
2290 total
+= nr
[order
] << order
;
2292 spin_unlock_irqrestore(&zone
->lock
, flags
);
2293 for (order
= 0; order
< MAX_ORDER
; order
++)
2294 printk("%lu*%lukB ", nr
[order
], K(1UL) << order
);
2295 printk("= %lukB\n", K(total
));
2298 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES
));
2300 show_swap_cache_info();
2303 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
2305 zoneref
->zone
= zone
;
2306 zoneref
->zone_idx
= zone_idx(zone
);
2310 * Builds allocation fallback zone lists.
2312 * Add all populated zones of a node to the zonelist.
2314 static int build_zonelists_node(pg_data_t
*pgdat
, struct zonelist
*zonelist
,
2315 int nr_zones
, enum zone_type zone_type
)
2319 BUG_ON(zone_type
>= MAX_NR_ZONES
);
2324 zone
= pgdat
->node_zones
+ zone_type
;
2325 if (populated_zone(zone
)) {
2326 zoneref_set_zone(zone
,
2327 &zonelist
->_zonerefs
[nr_zones
++]);
2328 check_highest_zone(zone_type
);
2331 } while (zone_type
);
2338 * 0 = automatic detection of better ordering.
2339 * 1 = order by ([node] distance, -zonetype)
2340 * 2 = order by (-zonetype, [node] distance)
2342 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
2343 * the same zonelist. So only NUMA can configure this param.
2345 #define ZONELIST_ORDER_DEFAULT 0
2346 #define ZONELIST_ORDER_NODE 1
2347 #define ZONELIST_ORDER_ZONE 2
2349 /* zonelist order in the kernel.
2350 * set_zonelist_order() will set this to NODE or ZONE.
2352 static int current_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
2353 static char zonelist_order_name
[3][8] = {"Default", "Node", "Zone"};
2357 /* The value user specified ....changed by config */
2358 static int user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
2359 /* string for sysctl */
2360 #define NUMA_ZONELIST_ORDER_LEN 16
2361 char numa_zonelist_order
[16] = "default";
2364 * interface for configure zonelist ordering.
2365 * command line option "numa_zonelist_order"
2366 * = "[dD]efault - default, automatic configuration.
2367 * = "[nN]ode - order by node locality, then by zone within node
2368 * = "[zZ]one - order by zone, then by locality within zone
2371 static int __parse_numa_zonelist_order(char *s
)
2373 if (*s
== 'd' || *s
== 'D') {
2374 user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
2375 } else if (*s
== 'n' || *s
== 'N') {
2376 user_zonelist_order
= ZONELIST_ORDER_NODE
;
2377 } else if (*s
== 'z' || *s
== 'Z') {
2378 user_zonelist_order
= ZONELIST_ORDER_ZONE
;
2381 "Ignoring invalid numa_zonelist_order value: "
2388 static __init
int setup_numa_zonelist_order(char *s
)
2391 return __parse_numa_zonelist_order(s
);
2394 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
2397 * sysctl handler for numa_zonelist_order
2399 int numa_zonelist_order_handler(ctl_table
*table
, int write
,
2400 void __user
*buffer
, size_t *length
,
2403 char saved_string
[NUMA_ZONELIST_ORDER_LEN
];
2407 strncpy(saved_string
, (char*)table
->data
,
2408 NUMA_ZONELIST_ORDER_LEN
);
2409 ret
= proc_dostring(table
, write
, buffer
, length
, ppos
);
2413 int oldval
= user_zonelist_order
;
2414 if (__parse_numa_zonelist_order((char*)table
->data
)) {
2416 * bogus value. restore saved string
2418 strncpy((char*)table
->data
, saved_string
,
2419 NUMA_ZONELIST_ORDER_LEN
);
2420 user_zonelist_order
= oldval
;
2421 } else if (oldval
!= user_zonelist_order
)
2422 build_all_zonelists();
2428 #define MAX_NODE_LOAD (nr_online_nodes)
2429 static int node_load
[MAX_NUMNODES
];
2432 * find_next_best_node - find the next node that should appear in a given node's fallback list
2433 * @node: node whose fallback list we're appending
2434 * @used_node_mask: nodemask_t of already used nodes
2436 * We use a number of factors to determine which is the next node that should
2437 * appear on a given node's fallback list. The node should not have appeared
2438 * already in @node's fallback list, and it should be the next closest node
2439 * according to the distance array (which contains arbitrary distance values
2440 * from each node to each node in the system), and should also prefer nodes
2441 * with no CPUs, since presumably they'll have very little allocation pressure
2442 * on them otherwise.
2443 * It returns -1 if no node is found.
2445 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
2448 int min_val
= INT_MAX
;
2450 const struct cpumask
*tmp
= cpumask_of_node(0);
2452 /* Use the local node if we haven't already */
2453 if (!node_isset(node
, *used_node_mask
)) {
2454 node_set(node
, *used_node_mask
);
2458 for_each_node_state(n
, N_HIGH_MEMORY
) {
2460 /* Don't want a node to appear more than once */
2461 if (node_isset(n
, *used_node_mask
))
2464 /* Use the distance array to find the distance */
2465 val
= node_distance(node
, n
);
2467 /* Penalize nodes under us ("prefer the next node") */
2470 /* Give preference to headless and unused nodes */
2471 tmp
= cpumask_of_node(n
);
2472 if (!cpumask_empty(tmp
))
2473 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
2475 /* Slight preference for less loaded node */
2476 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
2477 val
+= node_load
[n
];
2479 if (val
< min_val
) {
2486 node_set(best_node
, *used_node_mask
);
2493 * Build zonelists ordered by node and zones within node.
2494 * This results in maximum locality--normal zone overflows into local
2495 * DMA zone, if any--but risks exhausting DMA zone.
2497 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int node
)
2500 struct zonelist
*zonelist
;
2502 zonelist
= &pgdat
->node_zonelists
[0];
2503 for (j
= 0; zonelist
->_zonerefs
[j
].zone
!= NULL
; j
++)
2505 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
2507 zonelist
->_zonerefs
[j
].zone
= NULL
;
2508 zonelist
->_zonerefs
[j
].zone_idx
= 0;
2512 * Build gfp_thisnode zonelists
2514 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
2517 struct zonelist
*zonelist
;
2519 zonelist
= &pgdat
->node_zonelists
[1];
2520 j
= build_zonelists_node(pgdat
, zonelist
, 0, MAX_NR_ZONES
- 1);
2521 zonelist
->_zonerefs
[j
].zone
= NULL
;
2522 zonelist
->_zonerefs
[j
].zone_idx
= 0;
2526 * Build zonelists ordered by zone and nodes within zones.
2527 * This results in conserving DMA zone[s] until all Normal memory is
2528 * exhausted, but results in overflowing to remote node while memory
2529 * may still exist in local DMA zone.
2531 static int node_order
[MAX_NUMNODES
];
2533 static void build_zonelists_in_zone_order(pg_data_t
*pgdat
, int nr_nodes
)
2536 int zone_type
; /* needs to be signed */
2538 struct zonelist
*zonelist
;
2540 zonelist
= &pgdat
->node_zonelists
[0];
2542 for (zone_type
= MAX_NR_ZONES
- 1; zone_type
>= 0; zone_type
--) {
2543 for (j
= 0; j
< nr_nodes
; j
++) {
2544 node
= node_order
[j
];
2545 z
= &NODE_DATA(node
)->node_zones
[zone_type
];
2546 if (populated_zone(z
)) {
2548 &zonelist
->_zonerefs
[pos
++]);
2549 check_highest_zone(zone_type
);
2553 zonelist
->_zonerefs
[pos
].zone
= NULL
;
2554 zonelist
->_zonerefs
[pos
].zone_idx
= 0;
2557 static int default_zonelist_order(void)
2560 unsigned long low_kmem_size
,total_size
;
2564 * ZONE_DMA and ZONE_DMA32 can be very small area in the sytem.
2565 * If they are really small and used heavily, the system can fall
2566 * into OOM very easily.
2567 * This function detect ZONE_DMA/DMA32 size and confgigures zone order.
2569 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
2572 for_each_online_node(nid
) {
2573 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
2574 z
= &NODE_DATA(nid
)->node_zones
[zone_type
];
2575 if (populated_zone(z
)) {
2576 if (zone_type
< ZONE_NORMAL
)
2577 low_kmem_size
+= z
->present_pages
;
2578 total_size
+= z
->present_pages
;
2582 if (!low_kmem_size
|| /* there are no DMA area. */
2583 low_kmem_size
> total_size
/2) /* DMA/DMA32 is big. */
2584 return ZONELIST_ORDER_NODE
;
2586 * look into each node's config.
2587 * If there is a node whose DMA/DMA32 memory is very big area on
2588 * local memory, NODE_ORDER may be suitable.
2590 average_size
= total_size
/
2591 (nodes_weight(node_states
[N_HIGH_MEMORY
]) + 1);
2592 for_each_online_node(nid
) {
2595 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
2596 z
= &NODE_DATA(nid
)->node_zones
[zone_type
];
2597 if (populated_zone(z
)) {
2598 if (zone_type
< ZONE_NORMAL
)
2599 low_kmem_size
+= z
->present_pages
;
2600 total_size
+= z
->present_pages
;
2603 if (low_kmem_size
&&
2604 total_size
> average_size
&& /* ignore small node */
2605 low_kmem_size
> total_size
* 70/100)
2606 return ZONELIST_ORDER_NODE
;
2608 return ZONELIST_ORDER_ZONE
;
2611 static void set_zonelist_order(void)
2613 if (user_zonelist_order
== ZONELIST_ORDER_DEFAULT
)
2614 current_zonelist_order
= default_zonelist_order();
2616 current_zonelist_order
= user_zonelist_order
;
2619 static void build_zonelists(pg_data_t
*pgdat
)
2623 nodemask_t used_mask
;
2624 int local_node
, prev_node
;
2625 struct zonelist
*zonelist
;
2626 int order
= current_zonelist_order
;
2628 /* initialize zonelists */
2629 for (i
= 0; i
< MAX_ZONELISTS
; i
++) {
2630 zonelist
= pgdat
->node_zonelists
+ i
;
2631 zonelist
->_zonerefs
[0].zone
= NULL
;
2632 zonelist
->_zonerefs
[0].zone_idx
= 0;
2635 /* NUMA-aware ordering of nodes */
2636 local_node
= pgdat
->node_id
;
2637 load
= nr_online_nodes
;
2638 prev_node
= local_node
;
2639 nodes_clear(used_mask
);
2641 memset(node_order
, 0, sizeof(node_order
));
2644 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
2645 int distance
= node_distance(local_node
, node
);
2648 * If another node is sufficiently far away then it is better
2649 * to reclaim pages in a zone before going off node.
2651 if (distance
> RECLAIM_DISTANCE
)
2652 zone_reclaim_mode
= 1;
2655 * We don't want to pressure a particular node.
2656 * So adding penalty to the first node in same
2657 * distance group to make it round-robin.
2659 if (distance
!= node_distance(local_node
, prev_node
))
2660 node_load
[node
] = load
;
2664 if (order
== ZONELIST_ORDER_NODE
)
2665 build_zonelists_in_node_order(pgdat
, node
);
2667 node_order
[j
++] = node
; /* remember order */
2670 if (order
== ZONELIST_ORDER_ZONE
) {
2671 /* calculate node order -- i.e., DMA last! */
2672 build_zonelists_in_zone_order(pgdat
, j
);
2675 build_thisnode_zonelists(pgdat
);
2678 /* Construct the zonelist performance cache - see further mmzone.h */
2679 static void build_zonelist_cache(pg_data_t
*pgdat
)
2681 struct zonelist
*zonelist
;
2682 struct zonelist_cache
*zlc
;
2685 zonelist
= &pgdat
->node_zonelists
[0];
2686 zonelist
->zlcache_ptr
= zlc
= &zonelist
->zlcache
;
2687 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
2688 for (z
= zonelist
->_zonerefs
; z
->zone
; z
++)
2689 zlc
->z_to_n
[z
- zonelist
->_zonerefs
] = zonelist_node_idx(z
);
2693 #else /* CONFIG_NUMA */
2695 static void set_zonelist_order(void)
2697 current_zonelist_order
= ZONELIST_ORDER_ZONE
;
2700 static void build_zonelists(pg_data_t
*pgdat
)
2702 int node
, local_node
;
2704 struct zonelist
*zonelist
;
2706 local_node
= pgdat
->node_id
;
2708 zonelist
= &pgdat
->node_zonelists
[0];
2709 j
= build_zonelists_node(pgdat
, zonelist
, 0, MAX_NR_ZONES
- 1);
2712 * Now we build the zonelist so that it contains the zones
2713 * of all the other nodes.
2714 * We don't want to pressure a particular node, so when
2715 * building the zones for node N, we make sure that the
2716 * zones coming right after the local ones are those from
2717 * node N+1 (modulo N)
2719 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
2720 if (!node_online(node
))
2722 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
2725 for (node
= 0; node
< local_node
; node
++) {
2726 if (!node_online(node
))
2728 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
2732 zonelist
->_zonerefs
[j
].zone
= NULL
;
2733 zonelist
->_zonerefs
[j
].zone_idx
= 0;
2736 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
2737 static void build_zonelist_cache(pg_data_t
*pgdat
)
2739 pgdat
->node_zonelists
[0].zlcache_ptr
= NULL
;
2742 #endif /* CONFIG_NUMA */
2744 /* return values int ....just for stop_machine() */
2745 static int __build_all_zonelists(void *dummy
)
2750 memset(node_load
, 0, sizeof(node_load
));
2752 for_each_online_node(nid
) {
2753 pg_data_t
*pgdat
= NODE_DATA(nid
);
2755 build_zonelists(pgdat
);
2756 build_zonelist_cache(pgdat
);
2761 void build_all_zonelists(void)
2763 set_zonelist_order();
2765 if (system_state
== SYSTEM_BOOTING
) {
2766 __build_all_zonelists(NULL
);
2767 mminit_verify_zonelist();
2768 cpuset_init_current_mems_allowed();
2770 /* we have to stop all cpus to guarantee there is no user
2772 stop_machine(__build_all_zonelists
, NULL
, NULL
);
2773 /* cpuset refresh routine should be here */
2775 vm_total_pages
= nr_free_pagecache_pages();
2777 * Disable grouping by mobility if the number of pages in the
2778 * system is too low to allow the mechanism to work. It would be
2779 * more accurate, but expensive to check per-zone. This check is
2780 * made on memory-hotadd so a system can start with mobility
2781 * disabled and enable it later
2783 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
2784 page_group_by_mobility_disabled
= 1;
2786 page_group_by_mobility_disabled
= 0;
2788 printk("Built %i zonelists in %s order, mobility grouping %s. "
2789 "Total pages: %ld\n",
2791 zonelist_order_name
[current_zonelist_order
],
2792 page_group_by_mobility_disabled
? "off" : "on",
2795 printk("Policy zone: %s\n", zone_names
[policy_zone
]);
2800 * Helper functions to size the waitqueue hash table.
2801 * Essentially these want to choose hash table sizes sufficiently
2802 * large so that collisions trying to wait on pages are rare.
2803 * But in fact, the number of active page waitqueues on typical
2804 * systems is ridiculously low, less than 200. So this is even
2805 * conservative, even though it seems large.
2807 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
2808 * waitqueues, i.e. the size of the waitq table given the number of pages.
2810 #define PAGES_PER_WAITQUEUE 256
2812 #ifndef CONFIG_MEMORY_HOTPLUG
2813 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
2815 unsigned long size
= 1;
2817 pages
/= PAGES_PER_WAITQUEUE
;
2819 while (size
< pages
)
2823 * Once we have dozens or even hundreds of threads sleeping
2824 * on IO we've got bigger problems than wait queue collision.
2825 * Limit the size of the wait table to a reasonable size.
2827 size
= min(size
, 4096UL);
2829 return max(size
, 4UL);
2833 * A zone's size might be changed by hot-add, so it is not possible to determine
2834 * a suitable size for its wait_table. So we use the maximum size now.
2836 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
2838 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
2839 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
2840 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
2842 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
2843 * or more by the traditional way. (See above). It equals:
2845 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
2846 * ia64(16K page size) : = ( 8G + 4M)byte.
2847 * powerpc (64K page size) : = (32G +16M)byte.
2849 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
2856 * This is an integer logarithm so that shifts can be used later
2857 * to extract the more random high bits from the multiplicative
2858 * hash function before the remainder is taken.
2860 static inline unsigned long wait_table_bits(unsigned long size
)
2865 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
2868 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
2869 * of blocks reserved is based on min_wmark_pages(zone). The memory within
2870 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
2871 * higher will lead to a bigger reserve which will get freed as contiguous
2872 * blocks as reclaim kicks in
2874 static void setup_zone_migrate_reserve(struct zone
*zone
)
2876 unsigned long start_pfn
, pfn
, end_pfn
;
2878 unsigned long block_migratetype
;
2881 /* Get the start pfn, end pfn and the number of blocks to reserve */
2882 start_pfn
= zone
->zone_start_pfn
;
2883 end_pfn
= start_pfn
+ zone
->spanned_pages
;
2884 reserve
= roundup(min_wmark_pages(zone
), pageblock_nr_pages
) >>
2888 * Reserve blocks are generally in place to help high-order atomic
2889 * allocations that are short-lived. A min_free_kbytes value that
2890 * would result in more than 2 reserve blocks for atomic allocations
2891 * is assumed to be in place to help anti-fragmentation for the
2892 * future allocation of hugepages at runtime.
2894 reserve
= min(2, reserve
);
2896 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= pageblock_nr_pages
) {
2897 if (!pfn_valid(pfn
))
2899 page
= pfn_to_page(pfn
);
2901 /* Watch out for overlapping nodes */
2902 if (page_to_nid(page
) != zone_to_nid(zone
))
2905 /* Blocks with reserved pages will never free, skip them. */
2906 if (PageReserved(page
))
2909 block_migratetype
= get_pageblock_migratetype(page
);
2911 /* If this block is reserved, account for it */
2912 if (reserve
> 0 && block_migratetype
== MIGRATE_RESERVE
) {
2917 /* Suitable for reserving if this block is movable */
2918 if (reserve
> 0 && block_migratetype
== MIGRATE_MOVABLE
) {
2919 set_pageblock_migratetype(page
, MIGRATE_RESERVE
);
2920 move_freepages_block(zone
, page
, MIGRATE_RESERVE
);
2926 * If the reserve is met and this is a previous reserved block,
2929 if (block_migratetype
== MIGRATE_RESERVE
) {
2930 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
2931 move_freepages_block(zone
, page
, MIGRATE_MOVABLE
);
2937 * Initially all pages are reserved - free ones are freed
2938 * up by free_all_bootmem() once the early boot process is
2939 * done. Non-atomic initialization, single-pass.
2941 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
2942 unsigned long start_pfn
, enum memmap_context context
)
2945 unsigned long end_pfn
= start_pfn
+ size
;
2949 if (highest_memmap_pfn
< end_pfn
- 1)
2950 highest_memmap_pfn
= end_pfn
- 1;
2952 z
= &NODE_DATA(nid
)->node_zones
[zone
];
2953 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
2955 * There can be holes in boot-time mem_map[]s
2956 * handed to this function. They do not
2957 * exist on hotplugged memory.
2959 if (context
== MEMMAP_EARLY
) {
2960 if (!early_pfn_valid(pfn
))
2962 if (!early_pfn_in_nid(pfn
, nid
))
2965 page
= pfn_to_page(pfn
);
2966 set_page_links(page
, zone
, nid
, pfn
);
2967 mminit_verify_page_links(page
, zone
, nid
, pfn
);
2968 init_page_count(page
);
2969 reset_page_mapcount(page
);
2970 SetPageReserved(page
);
2972 * Mark the block movable so that blocks are reserved for
2973 * movable at startup. This will force kernel allocations
2974 * to reserve their blocks rather than leaking throughout
2975 * the address space during boot when many long-lived
2976 * kernel allocations are made. Later some blocks near
2977 * the start are marked MIGRATE_RESERVE by
2978 * setup_zone_migrate_reserve()
2980 * bitmap is created for zone's valid pfn range. but memmap
2981 * can be created for invalid pages (for alignment)
2982 * check here not to call set_pageblock_migratetype() against
2985 if ((z
->zone_start_pfn
<= pfn
)
2986 && (pfn
< z
->zone_start_pfn
+ z
->spanned_pages
)
2987 && !(pfn
& (pageblock_nr_pages
- 1)))
2988 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
2990 INIT_LIST_HEAD(&page
->lru
);
2991 #ifdef WANT_PAGE_VIRTUAL
2992 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
2993 if (!is_highmem_idx(zone
))
2994 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
2999 static void __meminit
zone_init_free_lists(struct zone
*zone
)
3002 for_each_migratetype_order(order
, t
) {
3003 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
3004 zone
->free_area
[order
].nr_free
= 0;
3008 #ifndef __HAVE_ARCH_MEMMAP_INIT
3009 #define memmap_init(size, nid, zone, start_pfn) \
3010 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3013 static int zone_batchsize(struct zone
*zone
)
3019 * The per-cpu-pages pools are set to around 1000th of the
3020 * size of the zone. But no more than 1/2 of a meg.
3022 * OK, so we don't know how big the cache is. So guess.
3024 batch
= zone
->present_pages
/ 1024;
3025 if (batch
* PAGE_SIZE
> 512 * 1024)
3026 batch
= (512 * 1024) / PAGE_SIZE
;
3027 batch
/= 4; /* We effectively *= 4 below */
3032 * Clamp the batch to a 2^n - 1 value. Having a power
3033 * of 2 value was found to be more likely to have
3034 * suboptimal cache aliasing properties in some cases.
3036 * For example if 2 tasks are alternately allocating
3037 * batches of pages, one task can end up with a lot
3038 * of pages of one half of the possible page colors
3039 * and the other with pages of the other colors.
3041 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
3046 /* The deferral and batching of frees should be suppressed under NOMMU
3049 * The problem is that NOMMU needs to be able to allocate large chunks
3050 * of contiguous memory as there's no hardware page translation to
3051 * assemble apparent contiguous memory from discontiguous pages.
3053 * Queueing large contiguous runs of pages for batching, however,
3054 * causes the pages to actually be freed in smaller chunks. As there
3055 * can be a significant delay between the individual batches being
3056 * recycled, this leads to the once large chunks of space being
3057 * fragmented and becoming unavailable for high-order allocations.
3063 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
3065 struct per_cpu_pages
*pcp
;
3068 memset(p
, 0, sizeof(*p
));
3072 pcp
->high
= 6 * batch
;
3073 pcp
->batch
= max(1UL, 1 * batch
);
3074 for (migratetype
= 0; migratetype
< MIGRATE_PCPTYPES
; migratetype
++)
3075 INIT_LIST_HEAD(&pcp
->lists
[migratetype
]);
3079 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3080 * to the value high for the pageset p.
3083 static void setup_pagelist_highmark(struct per_cpu_pageset
*p
,
3086 struct per_cpu_pages
*pcp
;
3090 pcp
->batch
= max(1UL, high
/4);
3091 if ((high
/4) > (PAGE_SHIFT
* 8))
3092 pcp
->batch
= PAGE_SHIFT
* 8;
3098 * Boot pageset table. One per cpu which is going to be used for all
3099 * zones and all nodes. The parameters will be set in such a way
3100 * that an item put on a list will immediately be handed over to
3101 * the buddy list. This is safe since pageset manipulation is done
3102 * with interrupts disabled.
3104 * Some NUMA counter updates may also be caught by the boot pagesets.
3106 * The boot_pagesets must be kept even after bootup is complete for
3107 * unused processors and/or zones. They do play a role for bootstrapping
3108 * hotplugged processors.
3110 * zoneinfo_show() and maybe other functions do
3111 * not check if the processor is online before following the pageset pointer.
3112 * Other parts of the kernel may not check if the zone is available.
3114 static struct per_cpu_pageset boot_pageset
[NR_CPUS
];
3117 * Dynamically allocate memory for the
3118 * per cpu pageset array in struct zone.
3120 static int __cpuinit
process_zones(int cpu
)
3122 struct zone
*zone
, *dzone
;
3123 int node
= cpu_to_node(cpu
);
3125 node_set_state(node
, N_CPU
); /* this node has a cpu */
3127 for_each_populated_zone(zone
) {
3128 zone_pcp(zone
, cpu
) = kmalloc_node(sizeof(struct per_cpu_pageset
),
3130 if (!zone_pcp(zone
, cpu
))
3133 setup_pageset(zone_pcp(zone
, cpu
), zone_batchsize(zone
));
3135 if (percpu_pagelist_fraction
)
3136 setup_pagelist_highmark(zone_pcp(zone
, cpu
),
3137 (zone
->present_pages
/ percpu_pagelist_fraction
));
3142 for_each_zone(dzone
) {
3143 if (!populated_zone(dzone
))
3147 kfree(zone_pcp(dzone
, cpu
));
3148 zone_pcp(dzone
, cpu
) = &boot_pageset
[cpu
];
3153 static inline void free_zone_pagesets(int cpu
)
3157 for_each_zone(zone
) {
3158 struct per_cpu_pageset
*pset
= zone_pcp(zone
, cpu
);
3160 /* Free per_cpu_pageset if it is slab allocated */
3161 if (pset
!= &boot_pageset
[cpu
])
3163 zone_pcp(zone
, cpu
) = &boot_pageset
[cpu
];
3167 static int __cpuinit
pageset_cpuup_callback(struct notifier_block
*nfb
,
3168 unsigned long action
,
3171 int cpu
= (long)hcpu
;
3172 int ret
= NOTIFY_OK
;
3175 case CPU_UP_PREPARE
:
3176 case CPU_UP_PREPARE_FROZEN
:
3177 if (process_zones(cpu
))
3180 case CPU_UP_CANCELED
:
3181 case CPU_UP_CANCELED_FROZEN
:
3183 case CPU_DEAD_FROZEN
:
3184 free_zone_pagesets(cpu
);
3192 static struct notifier_block __cpuinitdata pageset_notifier
=
3193 { &pageset_cpuup_callback
, NULL
, 0 };
3195 void __init
setup_per_cpu_pageset(void)
3199 /* Initialize per_cpu_pageset for cpu 0.
3200 * A cpuup callback will do this for every cpu
3201 * as it comes online
3203 err
= process_zones(smp_processor_id());
3205 register_cpu_notifier(&pageset_notifier
);
3210 static noinline __init_refok
3211 int zone_wait_table_init(struct zone
*zone
, unsigned long zone_size_pages
)
3214 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
3218 * The per-page waitqueue mechanism uses hashed waitqueues
3221 zone
->wait_table_hash_nr_entries
=
3222 wait_table_hash_nr_entries(zone_size_pages
);
3223 zone
->wait_table_bits
=
3224 wait_table_bits(zone
->wait_table_hash_nr_entries
);
3225 alloc_size
= zone
->wait_table_hash_nr_entries
3226 * sizeof(wait_queue_head_t
);
3228 if (!slab_is_available()) {
3229 zone
->wait_table
= (wait_queue_head_t
*)
3230 alloc_bootmem_node(pgdat
, alloc_size
);
3233 * This case means that a zone whose size was 0 gets new memory
3234 * via memory hot-add.
3235 * But it may be the case that a new node was hot-added. In
3236 * this case vmalloc() will not be able to use this new node's
3237 * memory - this wait_table must be initialized to use this new
3238 * node itself as well.
3239 * To use this new node's memory, further consideration will be
3242 zone
->wait_table
= vmalloc(alloc_size
);
3244 if (!zone
->wait_table
)
3247 for(i
= 0; i
< zone
->wait_table_hash_nr_entries
; ++i
)
3248 init_waitqueue_head(zone
->wait_table
+ i
);
3253 static int __zone_pcp_update(void *data
)
3255 struct zone
*zone
= data
;
3257 unsigned long batch
= zone_batchsize(zone
), flags
;
3259 for (cpu
= 0; cpu
< NR_CPUS
; cpu
++) {
3260 struct per_cpu_pageset
*pset
;
3261 struct per_cpu_pages
*pcp
;
3263 pset
= zone_pcp(zone
, cpu
);
3266 local_irq_save(flags
);
3267 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
3268 setup_pageset(pset
, batch
);
3269 local_irq_restore(flags
);
3274 void zone_pcp_update(struct zone
*zone
)
3276 stop_machine(__zone_pcp_update
, zone
, NULL
);
3279 static __meminit
void zone_pcp_init(struct zone
*zone
)
3282 unsigned long batch
= zone_batchsize(zone
);
3284 for (cpu
= 0; cpu
< NR_CPUS
; cpu
++) {
3286 /* Early boot. Slab allocator not functional yet */
3287 zone_pcp(zone
, cpu
) = &boot_pageset
[cpu
];
3288 setup_pageset(&boot_pageset
[cpu
],0);
3290 setup_pageset(zone_pcp(zone
,cpu
), batch
);
3293 if (zone
->present_pages
)
3294 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%lu\n",
3295 zone
->name
, zone
->present_pages
, batch
);
3298 __meminit
int init_currently_empty_zone(struct zone
*zone
,
3299 unsigned long zone_start_pfn
,
3301 enum memmap_context context
)
3303 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
3305 ret
= zone_wait_table_init(zone
, size
);
3308 pgdat
->nr_zones
= zone_idx(zone
) + 1;
3310 zone
->zone_start_pfn
= zone_start_pfn
;
3312 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
3313 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
3315 (unsigned long)zone_idx(zone
),
3316 zone_start_pfn
, (zone_start_pfn
+ size
));
3318 zone_init_free_lists(zone
);
3323 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3325 * Basic iterator support. Return the first range of PFNs for a node
3326 * Note: nid == MAX_NUMNODES returns first region regardless of node
3328 static int __meminit
first_active_region_index_in_nid(int nid
)
3332 for (i
= 0; i
< nr_nodemap_entries
; i
++)
3333 if (nid
== MAX_NUMNODES
|| early_node_map
[i
].nid
== nid
)
3340 * Basic iterator support. Return the next active range of PFNs for a node
3341 * Note: nid == MAX_NUMNODES returns next region regardless of node
3343 static int __meminit
next_active_region_index_in_nid(int index
, int nid
)
3345 for (index
= index
+ 1; index
< nr_nodemap_entries
; index
++)
3346 if (nid
== MAX_NUMNODES
|| early_node_map
[index
].nid
== nid
)
3352 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
3354 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
3355 * Architectures may implement their own version but if add_active_range()
3356 * was used and there are no special requirements, this is a convenient
3359 int __meminit
__early_pfn_to_nid(unsigned long pfn
)
3363 for (i
= 0; i
< nr_nodemap_entries
; i
++) {
3364 unsigned long start_pfn
= early_node_map
[i
].start_pfn
;
3365 unsigned long end_pfn
= early_node_map
[i
].end_pfn
;
3367 if (start_pfn
<= pfn
&& pfn
< end_pfn
)
3368 return early_node_map
[i
].nid
;
3370 /* This is a memory hole */
3373 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
3375 int __meminit
early_pfn_to_nid(unsigned long pfn
)
3379 nid
= __early_pfn_to_nid(pfn
);
3382 /* just returns 0 */
3386 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
3387 bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
3391 nid
= __early_pfn_to_nid(pfn
);
3392 if (nid
>= 0 && nid
!= node
)
3398 /* Basic iterator support to walk early_node_map[] */
3399 #define for_each_active_range_index_in_nid(i, nid) \
3400 for (i = first_active_region_index_in_nid(nid); i != -1; \
3401 i = next_active_region_index_in_nid(i, nid))
3404 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
3405 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
3406 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
3408 * If an architecture guarantees that all ranges registered with
3409 * add_active_ranges() contain no holes and may be freed, this
3410 * this function may be used instead of calling free_bootmem() manually.
3412 void __init
free_bootmem_with_active_regions(int nid
,
3413 unsigned long max_low_pfn
)
3417 for_each_active_range_index_in_nid(i
, nid
) {
3418 unsigned long size_pages
= 0;
3419 unsigned long end_pfn
= early_node_map
[i
].end_pfn
;
3421 if (early_node_map
[i
].start_pfn
>= max_low_pfn
)
3424 if (end_pfn
> max_low_pfn
)
3425 end_pfn
= max_low_pfn
;
3427 size_pages
= end_pfn
- early_node_map
[i
].start_pfn
;
3428 free_bootmem_node(NODE_DATA(early_node_map
[i
].nid
),
3429 PFN_PHYS(early_node_map
[i
].start_pfn
),
3430 size_pages
<< PAGE_SHIFT
);
3434 void __init
work_with_active_regions(int nid
, work_fn_t work_fn
, void *data
)
3439 for_each_active_range_index_in_nid(i
, nid
) {
3440 ret
= work_fn(early_node_map
[i
].start_pfn
,
3441 early_node_map
[i
].end_pfn
, data
);
3447 * sparse_memory_present_with_active_regions - Call memory_present for each active range
3448 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
3450 * If an architecture guarantees that all ranges registered with
3451 * add_active_ranges() contain no holes and may be freed, this
3452 * function may be used instead of calling memory_present() manually.
3454 void __init
sparse_memory_present_with_active_regions(int nid
)
3458 for_each_active_range_index_in_nid(i
, nid
)
3459 memory_present(early_node_map
[i
].nid
,
3460 early_node_map
[i
].start_pfn
,
3461 early_node_map
[i
].end_pfn
);
3465 * get_pfn_range_for_nid - Return the start and end page frames for a node
3466 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
3467 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
3468 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
3470 * It returns the start and end page frame of a node based on information
3471 * provided by an arch calling add_active_range(). If called for a node
3472 * with no available memory, a warning is printed and the start and end
3475 void __meminit
get_pfn_range_for_nid(unsigned int nid
,
3476 unsigned long *start_pfn
, unsigned long *end_pfn
)
3482 for_each_active_range_index_in_nid(i
, nid
) {
3483 *start_pfn
= min(*start_pfn
, early_node_map
[i
].start_pfn
);
3484 *end_pfn
= max(*end_pfn
, early_node_map
[i
].end_pfn
);
3487 if (*start_pfn
== -1UL)
3492 * This finds a zone that can be used for ZONE_MOVABLE pages. The
3493 * assumption is made that zones within a node are ordered in monotonic
3494 * increasing memory addresses so that the "highest" populated zone is used
3496 static void __init
find_usable_zone_for_movable(void)
3499 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
3500 if (zone_index
== ZONE_MOVABLE
)
3503 if (arch_zone_highest_possible_pfn
[zone_index
] >
3504 arch_zone_lowest_possible_pfn
[zone_index
])
3508 VM_BUG_ON(zone_index
== -1);
3509 movable_zone
= zone_index
;
3513 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
3514 * because it is sized independant of architecture. Unlike the other zones,
3515 * the starting point for ZONE_MOVABLE is not fixed. It may be different
3516 * in each node depending on the size of each node and how evenly kernelcore
3517 * is distributed. This helper function adjusts the zone ranges
3518 * provided by the architecture for a given node by using the end of the
3519 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
3520 * zones within a node are in order of monotonic increases memory addresses
3522 static void __meminit
adjust_zone_range_for_zone_movable(int nid
,
3523 unsigned long zone_type
,
3524 unsigned long node_start_pfn
,
3525 unsigned long node_end_pfn
,
3526 unsigned long *zone_start_pfn
,
3527 unsigned long *zone_end_pfn
)
3529 /* Only adjust if ZONE_MOVABLE is on this node */
3530 if (zone_movable_pfn
[nid
]) {
3531 /* Size ZONE_MOVABLE */
3532 if (zone_type
== ZONE_MOVABLE
) {
3533 *zone_start_pfn
= zone_movable_pfn
[nid
];
3534 *zone_end_pfn
= min(node_end_pfn
,
3535 arch_zone_highest_possible_pfn
[movable_zone
]);
3537 /* Adjust for ZONE_MOVABLE starting within this range */
3538 } else if (*zone_start_pfn
< zone_movable_pfn
[nid
] &&
3539 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
3540 *zone_end_pfn
= zone_movable_pfn
[nid
];
3542 /* Check if this whole range is within ZONE_MOVABLE */
3543 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
3544 *zone_start_pfn
= *zone_end_pfn
;
3549 * Return the number of pages a zone spans in a node, including holes
3550 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
3552 static unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
3553 unsigned long zone_type
,
3554 unsigned long *ignored
)
3556 unsigned long node_start_pfn
, node_end_pfn
;
3557 unsigned long zone_start_pfn
, zone_end_pfn
;
3559 /* Get the start and end of the node and zone */
3560 get_pfn_range_for_nid(nid
, &node_start_pfn
, &node_end_pfn
);
3561 zone_start_pfn
= arch_zone_lowest_possible_pfn
[zone_type
];
3562 zone_end_pfn
= arch_zone_highest_possible_pfn
[zone_type
];
3563 adjust_zone_range_for_zone_movable(nid
, zone_type
,
3564 node_start_pfn
, node_end_pfn
,
3565 &zone_start_pfn
, &zone_end_pfn
);
3567 /* Check that this node has pages within the zone's required range */
3568 if (zone_end_pfn
< node_start_pfn
|| zone_start_pfn
> node_end_pfn
)
3571 /* Move the zone boundaries inside the node if necessary */
3572 zone_end_pfn
= min(zone_end_pfn
, node_end_pfn
);
3573 zone_start_pfn
= max(zone_start_pfn
, node_start_pfn
);
3575 /* Return the spanned pages */
3576 return zone_end_pfn
- zone_start_pfn
;
3580 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
3581 * then all holes in the requested range will be accounted for.
3583 unsigned long __meminit
__absent_pages_in_range(int nid
,
3584 unsigned long range_start_pfn
,
3585 unsigned long range_end_pfn
)
3588 unsigned long prev_end_pfn
= 0, hole_pages
= 0;
3589 unsigned long start_pfn
;
3591 /* Find the end_pfn of the first active range of pfns in the node */
3592 i
= first_active_region_index_in_nid(nid
);
3596 prev_end_pfn
= min(early_node_map
[i
].start_pfn
, range_end_pfn
);
3598 /* Account for ranges before physical memory on this node */
3599 if (early_node_map
[i
].start_pfn
> range_start_pfn
)
3600 hole_pages
= prev_end_pfn
- range_start_pfn
;
3602 /* Find all holes for the zone within the node */
3603 for (; i
!= -1; i
= next_active_region_index_in_nid(i
, nid
)) {
3605 /* No need to continue if prev_end_pfn is outside the zone */
3606 if (prev_end_pfn
>= range_end_pfn
)
3609 /* Make sure the end of the zone is not within the hole */
3610 start_pfn
= min(early_node_map
[i
].start_pfn
, range_end_pfn
);
3611 prev_end_pfn
= max(prev_end_pfn
, range_start_pfn
);
3613 /* Update the hole size cound and move on */
3614 if (start_pfn
> range_start_pfn
) {
3615 BUG_ON(prev_end_pfn
> start_pfn
);
3616 hole_pages
+= start_pfn
- prev_end_pfn
;
3618 prev_end_pfn
= early_node_map
[i
].end_pfn
;
3621 /* Account for ranges past physical memory on this node */
3622 if (range_end_pfn
> prev_end_pfn
)
3623 hole_pages
+= range_end_pfn
-
3624 max(range_start_pfn
, prev_end_pfn
);
3630 * absent_pages_in_range - Return number of page frames in holes within a range
3631 * @start_pfn: The start PFN to start searching for holes
3632 * @end_pfn: The end PFN to stop searching for holes
3634 * It returns the number of pages frames in memory holes within a range.
3636 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
3637 unsigned long end_pfn
)
3639 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
3642 /* Return the number of page frames in holes in a zone on a node */
3643 static unsigned long __meminit
zone_absent_pages_in_node(int nid
,
3644 unsigned long zone_type
,
3645 unsigned long *ignored
)
3647 unsigned long node_start_pfn
, node_end_pfn
;
3648 unsigned long zone_start_pfn
, zone_end_pfn
;
3650 get_pfn_range_for_nid(nid
, &node_start_pfn
, &node_end_pfn
);
3651 zone_start_pfn
= max(arch_zone_lowest_possible_pfn
[zone_type
],
3653 zone_end_pfn
= min(arch_zone_highest_possible_pfn
[zone_type
],
3656 adjust_zone_range_for_zone_movable(nid
, zone_type
,
3657 node_start_pfn
, node_end_pfn
,
3658 &zone_start_pfn
, &zone_end_pfn
);
3659 return __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
3663 static inline unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
3664 unsigned long zone_type
,
3665 unsigned long *zones_size
)
3667 return zones_size
[zone_type
];
3670 static inline unsigned long __meminit
zone_absent_pages_in_node(int nid
,
3671 unsigned long zone_type
,
3672 unsigned long *zholes_size
)
3677 return zholes_size
[zone_type
];
3682 static void __meminit
calculate_node_totalpages(struct pglist_data
*pgdat
,
3683 unsigned long *zones_size
, unsigned long *zholes_size
)
3685 unsigned long realtotalpages
, totalpages
= 0;
3688 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
3689 totalpages
+= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
3691 pgdat
->node_spanned_pages
= totalpages
;
3693 realtotalpages
= totalpages
;
3694 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
3696 zone_absent_pages_in_node(pgdat
->node_id
, i
,
3698 pgdat
->node_present_pages
= realtotalpages
;
3699 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
3703 #ifndef CONFIG_SPARSEMEM
3705 * Calculate the size of the zone->blockflags rounded to an unsigned long
3706 * Start by making sure zonesize is a multiple of pageblock_order by rounding
3707 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
3708 * round what is now in bits to nearest long in bits, then return it in
3711 static unsigned long __init
usemap_size(unsigned long zonesize
)
3713 unsigned long usemapsize
;
3715 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
3716 usemapsize
= usemapsize
>> pageblock_order
;
3717 usemapsize
*= NR_PAGEBLOCK_BITS
;
3718 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
3720 return usemapsize
/ 8;
3723 static void __init
setup_usemap(struct pglist_data
*pgdat
,
3724 struct zone
*zone
, unsigned long zonesize
)
3726 unsigned long usemapsize
= usemap_size(zonesize
);
3727 zone
->pageblock_flags
= NULL
;
3729 zone
->pageblock_flags
= alloc_bootmem_node(pgdat
, usemapsize
);
3732 static void inline setup_usemap(struct pglist_data
*pgdat
,
3733 struct zone
*zone
, unsigned long zonesize
) {}
3734 #endif /* CONFIG_SPARSEMEM */
3736 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
3738 /* Return a sensible default order for the pageblock size. */
3739 static inline int pageblock_default_order(void)
3741 if (HPAGE_SHIFT
> PAGE_SHIFT
)
3742 return HUGETLB_PAGE_ORDER
;
3747 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
3748 static inline void __init
set_pageblock_order(unsigned int order
)
3750 /* Check that pageblock_nr_pages has not already been setup */
3751 if (pageblock_order
)
3755 * Assume the largest contiguous order of interest is a huge page.
3756 * This value may be variable depending on boot parameters on IA64
3758 pageblock_order
= order
;
3760 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
3763 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
3764 * and pageblock_default_order() are unused as pageblock_order is set
3765 * at compile-time. See include/linux/pageblock-flags.h for the values of
3766 * pageblock_order based on the kernel config
3768 static inline int pageblock_default_order(unsigned int order
)
3772 #define set_pageblock_order(x) do {} while (0)
3774 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
3777 * Set up the zone data structures:
3778 * - mark all pages reserved
3779 * - mark all memory queues empty
3780 * - clear the memory bitmaps
3782 static void __paginginit
free_area_init_core(struct pglist_data
*pgdat
,
3783 unsigned long *zones_size
, unsigned long *zholes_size
)
3786 int nid
= pgdat
->node_id
;
3787 unsigned long zone_start_pfn
= pgdat
->node_start_pfn
;
3790 pgdat_resize_init(pgdat
);
3791 pgdat
->nr_zones
= 0;
3792 init_waitqueue_head(&pgdat
->kswapd_wait
);
3793 pgdat
->kswapd_max_order
= 0;
3794 pgdat_page_cgroup_init(pgdat
);
3796 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
3797 struct zone
*zone
= pgdat
->node_zones
+ j
;
3798 unsigned long size
, realsize
, memmap_pages
;
3801 size
= zone_spanned_pages_in_node(nid
, j
, zones_size
);
3802 realsize
= size
- zone_absent_pages_in_node(nid
, j
,
3806 * Adjust realsize so that it accounts for how much memory
3807 * is used by this zone for memmap. This affects the watermark
3808 * and per-cpu initialisations
3811 PAGE_ALIGN(size
* sizeof(struct page
)) >> PAGE_SHIFT
;
3812 if (realsize
>= memmap_pages
) {
3813 realsize
-= memmap_pages
;
3816 " %s zone: %lu pages used for memmap\n",
3817 zone_names
[j
], memmap_pages
);
3820 " %s zone: %lu pages exceeds realsize %lu\n",
3821 zone_names
[j
], memmap_pages
, realsize
);
3823 /* Account for reserved pages */
3824 if (j
== 0 && realsize
> dma_reserve
) {
3825 realsize
-= dma_reserve
;
3826 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
3827 zone_names
[0], dma_reserve
);
3830 if (!is_highmem_idx(j
))
3831 nr_kernel_pages
+= realsize
;
3832 nr_all_pages
+= realsize
;
3834 zone
->spanned_pages
= size
;
3835 zone
->present_pages
= realsize
;
3838 zone
->min_unmapped_pages
= (realsize
*sysctl_min_unmapped_ratio
)
3840 zone
->min_slab_pages
= (realsize
* sysctl_min_slab_ratio
) / 100;
3842 zone
->name
= zone_names
[j
];
3843 spin_lock_init(&zone
->lock
);
3844 spin_lock_init(&zone
->lru_lock
);
3845 zone_seqlock_init(zone
);
3846 zone
->zone_pgdat
= pgdat
;
3848 zone
->prev_priority
= DEF_PRIORITY
;
3850 zone_pcp_init(zone
);
3852 INIT_LIST_HEAD(&zone
->lru
[l
].list
);
3853 zone
->reclaim_stat
.nr_saved_scan
[l
] = 0;
3855 zone
->reclaim_stat
.recent_rotated
[0] = 0;
3856 zone
->reclaim_stat
.recent_rotated
[1] = 0;
3857 zone
->reclaim_stat
.recent_scanned
[0] = 0;
3858 zone
->reclaim_stat
.recent_scanned
[1] = 0;
3859 zap_zone_vm_stats(zone
);
3864 set_pageblock_order(pageblock_default_order());
3865 setup_usemap(pgdat
, zone
, size
);
3866 ret
= init_currently_empty_zone(zone
, zone_start_pfn
,
3867 size
, MEMMAP_EARLY
);
3869 memmap_init(size
, nid
, j
, zone_start_pfn
);
3870 zone_start_pfn
+= size
;
3874 static void __init_refok
alloc_node_mem_map(struct pglist_data
*pgdat
)
3876 /* Skip empty nodes */
3877 if (!pgdat
->node_spanned_pages
)
3880 #ifdef CONFIG_FLAT_NODE_MEM_MAP
3881 /* ia64 gets its own node_mem_map, before this, without bootmem */
3882 if (!pgdat
->node_mem_map
) {
3883 unsigned long size
, start
, end
;
3887 * The zone's endpoints aren't required to be MAX_ORDER
3888 * aligned but the node_mem_map endpoints must be in order
3889 * for the buddy allocator to function correctly.
3891 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
3892 end
= pgdat
->node_start_pfn
+ pgdat
->node_spanned_pages
;
3893 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
3894 size
= (end
- start
) * sizeof(struct page
);
3895 map
= alloc_remap(pgdat
->node_id
, size
);
3897 map
= alloc_bootmem_node(pgdat
, size
);
3898 pgdat
->node_mem_map
= map
+ (pgdat
->node_start_pfn
- start
);
3900 #ifndef CONFIG_NEED_MULTIPLE_NODES
3902 * With no DISCONTIG, the global mem_map is just set as node 0's
3904 if (pgdat
== NODE_DATA(0)) {
3905 mem_map
= NODE_DATA(0)->node_mem_map
;
3906 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3907 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
3908 mem_map
-= (pgdat
->node_start_pfn
- ARCH_PFN_OFFSET
);
3909 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
3912 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
3915 void __paginginit
free_area_init_node(int nid
, unsigned long *zones_size
,
3916 unsigned long node_start_pfn
, unsigned long *zholes_size
)
3918 pg_data_t
*pgdat
= NODE_DATA(nid
);
3920 pgdat
->node_id
= nid
;
3921 pgdat
->node_start_pfn
= node_start_pfn
;
3922 calculate_node_totalpages(pgdat
, zones_size
, zholes_size
);
3924 alloc_node_mem_map(pgdat
);
3925 #ifdef CONFIG_FLAT_NODE_MEM_MAP
3926 printk(KERN_DEBUG
"free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
3927 nid
, (unsigned long)pgdat
,
3928 (unsigned long)pgdat
->node_mem_map
);
3931 free_area_init_core(pgdat
, zones_size
, zholes_size
);
3934 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3936 #if MAX_NUMNODES > 1
3938 * Figure out the number of possible node ids.
3940 static void __init
setup_nr_node_ids(void)
3943 unsigned int highest
= 0;
3945 for_each_node_mask(node
, node_possible_map
)
3947 nr_node_ids
= highest
+ 1;
3950 static inline void setup_nr_node_ids(void)
3956 * add_active_range - Register a range of PFNs backed by physical memory
3957 * @nid: The node ID the range resides on
3958 * @start_pfn: The start PFN of the available physical memory
3959 * @end_pfn: The end PFN of the available physical memory
3961 * These ranges are stored in an early_node_map[] and later used by
3962 * free_area_init_nodes() to calculate zone sizes and holes. If the
3963 * range spans a memory hole, it is up to the architecture to ensure
3964 * the memory is not freed by the bootmem allocator. If possible
3965 * the range being registered will be merged with existing ranges.
3967 void __init
add_active_range(unsigned int nid
, unsigned long start_pfn
,
3968 unsigned long end_pfn
)
3972 mminit_dprintk(MMINIT_TRACE
, "memory_register",
3973 "Entering add_active_range(%d, %#lx, %#lx) "
3974 "%d entries of %d used\n",
3975 nid
, start_pfn
, end_pfn
,
3976 nr_nodemap_entries
, MAX_ACTIVE_REGIONS
);
3978 mminit_validate_memmodel_limits(&start_pfn
, &end_pfn
);
3980 /* Merge with existing active regions if possible */
3981 for (i
= 0; i
< nr_nodemap_entries
; i
++) {
3982 if (early_node_map
[i
].nid
!= nid
)
3985 /* Skip if an existing region covers this new one */
3986 if (start_pfn
>= early_node_map
[i
].start_pfn
&&
3987 end_pfn
<= early_node_map
[i
].end_pfn
)
3990 /* Merge forward if suitable */
3991 if (start_pfn
<= early_node_map
[i
].end_pfn
&&
3992 end_pfn
> early_node_map
[i
].end_pfn
) {
3993 early_node_map
[i
].end_pfn
= end_pfn
;
3997 /* Merge backward if suitable */
3998 if (start_pfn
< early_node_map
[i
].end_pfn
&&
3999 end_pfn
>= early_node_map
[i
].start_pfn
) {
4000 early_node_map
[i
].start_pfn
= start_pfn
;
4005 /* Check that early_node_map is large enough */
4006 if (i
>= MAX_ACTIVE_REGIONS
) {
4007 printk(KERN_CRIT
"More than %d memory regions, truncating\n",
4008 MAX_ACTIVE_REGIONS
);
4012 early_node_map
[i
].nid
= nid
;
4013 early_node_map
[i
].start_pfn
= start_pfn
;
4014 early_node_map
[i
].end_pfn
= end_pfn
;
4015 nr_nodemap_entries
= i
+ 1;
4019 * remove_active_range - Shrink an existing registered range of PFNs
4020 * @nid: The node id the range is on that should be shrunk
4021 * @start_pfn: The new PFN of the range
4022 * @end_pfn: The new PFN of the range
4024 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
4025 * The map is kept near the end physical page range that has already been
4026 * registered. This function allows an arch to shrink an existing registered
4029 void __init
remove_active_range(unsigned int nid
, unsigned long start_pfn
,
4030 unsigned long end_pfn
)
4035 printk(KERN_DEBUG
"remove_active_range (%d, %lu, %lu)\n",
4036 nid
, start_pfn
, end_pfn
);
4038 /* Find the old active region end and shrink */
4039 for_each_active_range_index_in_nid(i
, nid
) {
4040 if (early_node_map
[i
].start_pfn
>= start_pfn
&&
4041 early_node_map
[i
].end_pfn
<= end_pfn
) {
4043 early_node_map
[i
].start_pfn
= 0;
4044 early_node_map
[i
].end_pfn
= 0;
4048 if (early_node_map
[i
].start_pfn
< start_pfn
&&
4049 early_node_map
[i
].end_pfn
> start_pfn
) {
4050 unsigned long temp_end_pfn
= early_node_map
[i
].end_pfn
;
4051 early_node_map
[i
].end_pfn
= start_pfn
;
4052 if (temp_end_pfn
> end_pfn
)
4053 add_active_range(nid
, end_pfn
, temp_end_pfn
);
4056 if (early_node_map
[i
].start_pfn
>= start_pfn
&&
4057 early_node_map
[i
].end_pfn
> end_pfn
&&
4058 early_node_map
[i
].start_pfn
< end_pfn
) {
4059 early_node_map
[i
].start_pfn
= end_pfn
;
4067 /* remove the blank ones */
4068 for (i
= nr_nodemap_entries
- 1; i
> 0; i
--) {
4069 if (early_node_map
[i
].nid
!= nid
)
4071 if (early_node_map
[i
].end_pfn
)
4073 /* we found it, get rid of it */
4074 for (j
= i
; j
< nr_nodemap_entries
- 1; j
++)
4075 memcpy(&early_node_map
[j
], &early_node_map
[j
+1],
4076 sizeof(early_node_map
[j
]));
4077 j
= nr_nodemap_entries
- 1;
4078 memset(&early_node_map
[j
], 0, sizeof(early_node_map
[j
]));
4079 nr_nodemap_entries
--;
4084 * remove_all_active_ranges - Remove all currently registered regions
4086 * During discovery, it may be found that a table like SRAT is invalid
4087 * and an alternative discovery method must be used. This function removes
4088 * all currently registered regions.
4090 void __init
remove_all_active_ranges(void)
4092 memset(early_node_map
, 0, sizeof(early_node_map
));
4093 nr_nodemap_entries
= 0;
4096 /* Compare two active node_active_regions */
4097 static int __init
cmp_node_active_region(const void *a
, const void *b
)
4099 struct node_active_region
*arange
= (struct node_active_region
*)a
;
4100 struct node_active_region
*brange
= (struct node_active_region
*)b
;
4102 /* Done this way to avoid overflows */
4103 if (arange
->start_pfn
> brange
->start_pfn
)
4105 if (arange
->start_pfn
< brange
->start_pfn
)
4111 /* sort the node_map by start_pfn */
4112 void __init
sort_node_map(void)
4114 sort(early_node_map
, (size_t)nr_nodemap_entries
,
4115 sizeof(struct node_active_region
),
4116 cmp_node_active_region
, NULL
);
4119 /* Find the lowest pfn for a node */
4120 static unsigned long __init
find_min_pfn_for_node(int nid
)
4123 unsigned long min_pfn
= ULONG_MAX
;
4125 /* Assuming a sorted map, the first range found has the starting pfn */
4126 for_each_active_range_index_in_nid(i
, nid
)
4127 min_pfn
= min(min_pfn
, early_node_map
[i
].start_pfn
);
4129 if (min_pfn
== ULONG_MAX
) {
4131 "Could not find start_pfn for node %d\n", nid
);
4139 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4141 * It returns the minimum PFN based on information provided via
4142 * add_active_range().
4144 unsigned long __init
find_min_pfn_with_active_regions(void)
4146 return find_min_pfn_for_node(MAX_NUMNODES
);
4150 * early_calculate_totalpages()
4151 * Sum pages in active regions for movable zone.
4152 * Populate N_HIGH_MEMORY for calculating usable_nodes.
4154 static unsigned long __init
early_calculate_totalpages(void)
4157 unsigned long totalpages
= 0;
4159 for (i
= 0; i
< nr_nodemap_entries
; i
++) {
4160 unsigned long pages
= early_node_map
[i
].end_pfn
-
4161 early_node_map
[i
].start_pfn
;
4162 totalpages
+= pages
;
4164 node_set_state(early_node_map
[i
].nid
, N_HIGH_MEMORY
);
4170 * Find the PFN the Movable zone begins in each node. Kernel memory
4171 * is spread evenly between nodes as long as the nodes have enough
4172 * memory. When they don't, some nodes will have more kernelcore than
4175 static void __init
find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn
)
4178 unsigned long usable_startpfn
;
4179 unsigned long kernelcore_node
, kernelcore_remaining
;
4180 /* save the state before borrow the nodemask */
4181 nodemask_t saved_node_state
= node_states
[N_HIGH_MEMORY
];
4182 unsigned long totalpages
= early_calculate_totalpages();
4183 int usable_nodes
= nodes_weight(node_states
[N_HIGH_MEMORY
]);
4186 * If movablecore was specified, calculate what size of
4187 * kernelcore that corresponds so that memory usable for
4188 * any allocation type is evenly spread. If both kernelcore
4189 * and movablecore are specified, then the value of kernelcore
4190 * will be used for required_kernelcore if it's greater than
4191 * what movablecore would have allowed.
4193 if (required_movablecore
) {
4194 unsigned long corepages
;
4197 * Round-up so that ZONE_MOVABLE is at least as large as what
4198 * was requested by the user
4200 required_movablecore
=
4201 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
4202 corepages
= totalpages
- required_movablecore
;
4204 required_kernelcore
= max(required_kernelcore
, corepages
);
4207 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4208 if (!required_kernelcore
)
4211 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4212 find_usable_zone_for_movable();
4213 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
4216 /* Spread kernelcore memory as evenly as possible throughout nodes */
4217 kernelcore_node
= required_kernelcore
/ usable_nodes
;
4218 for_each_node_state(nid
, N_HIGH_MEMORY
) {
4220 * Recalculate kernelcore_node if the division per node
4221 * now exceeds what is necessary to satisfy the requested
4222 * amount of memory for the kernel
4224 if (required_kernelcore
< kernelcore_node
)
4225 kernelcore_node
= required_kernelcore
/ usable_nodes
;
4228 * As the map is walked, we track how much memory is usable
4229 * by the kernel using kernelcore_remaining. When it is
4230 * 0, the rest of the node is usable by ZONE_MOVABLE
4232 kernelcore_remaining
= kernelcore_node
;
4234 /* Go through each range of PFNs within this node */
4235 for_each_active_range_index_in_nid(i
, nid
) {
4236 unsigned long start_pfn
, end_pfn
;
4237 unsigned long size_pages
;
4239 start_pfn
= max(early_node_map
[i
].start_pfn
,
4240 zone_movable_pfn
[nid
]);
4241 end_pfn
= early_node_map
[i
].end_pfn
;
4242 if (start_pfn
>= end_pfn
)
4245 /* Account for what is only usable for kernelcore */
4246 if (start_pfn
< usable_startpfn
) {
4247 unsigned long kernel_pages
;
4248 kernel_pages
= min(end_pfn
, usable_startpfn
)
4251 kernelcore_remaining
-= min(kernel_pages
,
4252 kernelcore_remaining
);
4253 required_kernelcore
-= min(kernel_pages
,
4254 required_kernelcore
);
4256 /* Continue if range is now fully accounted */
4257 if (end_pfn
<= usable_startpfn
) {
4260 * Push zone_movable_pfn to the end so
4261 * that if we have to rebalance
4262 * kernelcore across nodes, we will
4263 * not double account here
4265 zone_movable_pfn
[nid
] = end_pfn
;
4268 start_pfn
= usable_startpfn
;
4272 * The usable PFN range for ZONE_MOVABLE is from
4273 * start_pfn->end_pfn. Calculate size_pages as the
4274 * number of pages used as kernelcore
4276 size_pages
= end_pfn
- start_pfn
;
4277 if (size_pages
> kernelcore_remaining
)
4278 size_pages
= kernelcore_remaining
;
4279 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
4282 * Some kernelcore has been met, update counts and
4283 * break if the kernelcore for this node has been
4286 required_kernelcore
-= min(required_kernelcore
,
4288 kernelcore_remaining
-= size_pages
;
4289 if (!kernelcore_remaining
)
4295 * If there is still required_kernelcore, we do another pass with one
4296 * less node in the count. This will push zone_movable_pfn[nid] further
4297 * along on the nodes that still have memory until kernelcore is
4301 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
4304 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4305 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
4306 zone_movable_pfn
[nid
] =
4307 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
4310 /* restore the node_state */
4311 node_states
[N_HIGH_MEMORY
] = saved_node_state
;
4314 /* Any regular memory on that node ? */
4315 static void check_for_regular_memory(pg_data_t
*pgdat
)
4317 #ifdef CONFIG_HIGHMEM
4318 enum zone_type zone_type
;
4320 for (zone_type
= 0; zone_type
<= ZONE_NORMAL
; zone_type
++) {
4321 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
4322 if (zone
->present_pages
)
4323 node_set_state(zone_to_nid(zone
), N_NORMAL_MEMORY
);
4329 * free_area_init_nodes - Initialise all pg_data_t and zone data
4330 * @max_zone_pfn: an array of max PFNs for each zone
4332 * This will call free_area_init_node() for each active node in the system.
4333 * Using the page ranges provided by add_active_range(), the size of each
4334 * zone in each node and their holes is calculated. If the maximum PFN
4335 * between two adjacent zones match, it is assumed that the zone is empty.
4336 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4337 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4338 * starts where the previous one ended. For example, ZONE_DMA32 starts
4339 * at arch_max_dma_pfn.
4341 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
4346 /* Sort early_node_map as initialisation assumes it is sorted */
4349 /* Record where the zone boundaries are */
4350 memset(arch_zone_lowest_possible_pfn
, 0,
4351 sizeof(arch_zone_lowest_possible_pfn
));
4352 memset(arch_zone_highest_possible_pfn
, 0,
4353 sizeof(arch_zone_highest_possible_pfn
));
4354 arch_zone_lowest_possible_pfn
[0] = find_min_pfn_with_active_regions();
4355 arch_zone_highest_possible_pfn
[0] = max_zone_pfn
[0];
4356 for (i
= 1; i
< MAX_NR_ZONES
; i
++) {
4357 if (i
== ZONE_MOVABLE
)
4359 arch_zone_lowest_possible_pfn
[i
] =
4360 arch_zone_highest_possible_pfn
[i
-1];
4361 arch_zone_highest_possible_pfn
[i
] =
4362 max(max_zone_pfn
[i
], arch_zone_lowest_possible_pfn
[i
]);
4364 arch_zone_lowest_possible_pfn
[ZONE_MOVABLE
] = 0;
4365 arch_zone_highest_possible_pfn
[ZONE_MOVABLE
] = 0;
4367 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4368 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
4369 find_zone_movable_pfns_for_nodes(zone_movable_pfn
);
4371 /* Print out the zone ranges */
4372 printk("Zone PFN ranges:\n");
4373 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
4374 if (i
== ZONE_MOVABLE
)
4376 printk(" %-8s %0#10lx -> %0#10lx\n",
4378 arch_zone_lowest_possible_pfn
[i
],
4379 arch_zone_highest_possible_pfn
[i
]);
4382 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4383 printk("Movable zone start PFN for each node\n");
4384 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
4385 if (zone_movable_pfn
[i
])
4386 printk(" Node %d: %lu\n", i
, zone_movable_pfn
[i
]);
4389 /* Print out the early_node_map[] */
4390 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries
);
4391 for (i
= 0; i
< nr_nodemap_entries
; i
++)
4392 printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map
[i
].nid
,
4393 early_node_map
[i
].start_pfn
,
4394 early_node_map
[i
].end_pfn
);
4396 /* Initialise every node */
4397 mminit_verify_pageflags_layout();
4398 setup_nr_node_ids();
4399 for_each_online_node(nid
) {
4400 pg_data_t
*pgdat
= NODE_DATA(nid
);
4401 free_area_init_node(nid
, NULL
,
4402 find_min_pfn_for_node(nid
), NULL
);
4404 /* Any memory on that node */
4405 if (pgdat
->node_present_pages
)
4406 node_set_state(nid
, N_HIGH_MEMORY
);
4407 check_for_regular_memory(pgdat
);
4411 static int __init
cmdline_parse_core(char *p
, unsigned long *core
)
4413 unsigned long long coremem
;
4417 coremem
= memparse(p
, &p
);
4418 *core
= coremem
>> PAGE_SHIFT
;
4420 /* Paranoid check that UL is enough for the coremem value */
4421 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
4427 * kernelcore=size sets the amount of memory for use for allocations that
4428 * cannot be reclaimed or migrated.
4430 static int __init
cmdline_parse_kernelcore(char *p
)
4432 return cmdline_parse_core(p
, &required_kernelcore
);
4436 * movablecore=size sets the amount of memory for use for allocations that
4437 * can be reclaimed or migrated.
4439 static int __init
cmdline_parse_movablecore(char *p
)
4441 return cmdline_parse_core(p
, &required_movablecore
);
4444 early_param("kernelcore", cmdline_parse_kernelcore
);
4445 early_param("movablecore", cmdline_parse_movablecore
);
4447 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4450 * set_dma_reserve - set the specified number of pages reserved in the first zone
4451 * @new_dma_reserve: The number of pages to mark reserved
4453 * The per-cpu batchsize and zone watermarks are determined by present_pages.
4454 * In the DMA zone, a significant percentage may be consumed by kernel image
4455 * and other unfreeable allocations which can skew the watermarks badly. This
4456 * function may optionally be used to account for unfreeable pages in the
4457 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
4458 * smaller per-cpu batchsize.
4460 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
4462 dma_reserve
= new_dma_reserve
;
4465 #ifndef CONFIG_NEED_MULTIPLE_NODES
4466 struct pglist_data __refdata contig_page_data
= { .bdata
= &bootmem_node_data
[0] };
4467 EXPORT_SYMBOL(contig_page_data
);
4470 void __init
free_area_init(unsigned long *zones_size
)
4472 free_area_init_node(0, zones_size
,
4473 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
4476 static int page_alloc_cpu_notify(struct notifier_block
*self
,
4477 unsigned long action
, void *hcpu
)
4479 int cpu
= (unsigned long)hcpu
;
4481 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
) {
4485 * Spill the event counters of the dead processor
4486 * into the current processors event counters.
4487 * This artificially elevates the count of the current
4490 vm_events_fold_cpu(cpu
);
4493 * Zero the differential counters of the dead processor
4494 * so that the vm statistics are consistent.
4496 * This is only okay since the processor is dead and cannot
4497 * race with what we are doing.
4499 refresh_cpu_vm_stats(cpu
);
4504 void __init
page_alloc_init(void)
4506 hotcpu_notifier(page_alloc_cpu_notify
, 0);
4510 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
4511 * or min_free_kbytes changes.
4513 static void calculate_totalreserve_pages(void)
4515 struct pglist_data
*pgdat
;
4516 unsigned long reserve_pages
= 0;
4517 enum zone_type i
, j
;
4519 for_each_online_pgdat(pgdat
) {
4520 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
4521 struct zone
*zone
= pgdat
->node_zones
+ i
;
4522 unsigned long max
= 0;
4524 /* Find valid and maximum lowmem_reserve in the zone */
4525 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
4526 if (zone
->lowmem_reserve
[j
] > max
)
4527 max
= zone
->lowmem_reserve
[j
];
4530 /* we treat the high watermark as reserved pages. */
4531 max
+= high_wmark_pages(zone
);
4533 if (max
> zone
->present_pages
)
4534 max
= zone
->present_pages
;
4535 reserve_pages
+= max
;
4538 totalreserve_pages
= reserve_pages
;
4542 * setup_per_zone_lowmem_reserve - called whenever
4543 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
4544 * has a correct pages reserved value, so an adequate number of
4545 * pages are left in the zone after a successful __alloc_pages().
4547 static void setup_per_zone_lowmem_reserve(void)
4549 struct pglist_data
*pgdat
;
4550 enum zone_type j
, idx
;
4552 for_each_online_pgdat(pgdat
) {
4553 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
4554 struct zone
*zone
= pgdat
->node_zones
+ j
;
4555 unsigned long present_pages
= zone
->present_pages
;
4557 zone
->lowmem_reserve
[j
] = 0;
4561 struct zone
*lower_zone
;
4565 if (sysctl_lowmem_reserve_ratio
[idx
] < 1)
4566 sysctl_lowmem_reserve_ratio
[idx
] = 1;
4568 lower_zone
= pgdat
->node_zones
+ idx
;
4569 lower_zone
->lowmem_reserve
[j
] = present_pages
/
4570 sysctl_lowmem_reserve_ratio
[idx
];
4571 present_pages
+= lower_zone
->present_pages
;
4576 /* update totalreserve_pages */
4577 calculate_totalreserve_pages();
4581 * setup_per_zone_wmarks - called when min_free_kbytes changes
4582 * or when memory is hot-{added|removed}
4584 * Ensures that the watermark[min,low,high] values for each zone are set
4585 * correctly with respect to min_free_kbytes.
4587 void setup_per_zone_wmarks(void)
4589 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
4590 unsigned long lowmem_pages
= 0;
4592 unsigned long flags
;
4594 /* Calculate total number of !ZONE_HIGHMEM pages */
4595 for_each_zone(zone
) {
4596 if (!is_highmem(zone
))
4597 lowmem_pages
+= zone
->present_pages
;
4600 for_each_zone(zone
) {
4603 spin_lock_irqsave(&zone
->lock
, flags
);
4604 tmp
= (u64
)pages_min
* zone
->present_pages
;
4605 do_div(tmp
, lowmem_pages
);
4606 if (is_highmem(zone
)) {
4608 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
4609 * need highmem pages, so cap pages_min to a small
4612 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
4613 * deltas controls asynch page reclaim, and so should
4614 * not be capped for highmem.
4618 min_pages
= zone
->present_pages
/ 1024;
4619 if (min_pages
< SWAP_CLUSTER_MAX
)
4620 min_pages
= SWAP_CLUSTER_MAX
;
4621 if (min_pages
> 128)
4623 zone
->watermark
[WMARK_MIN
] = min_pages
;
4626 * If it's a lowmem zone, reserve a number of pages
4627 * proportionate to the zone's size.
4629 zone
->watermark
[WMARK_MIN
] = tmp
;
4632 zone
->watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + (tmp
>> 2);
4633 zone
->watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + (tmp
>> 1);
4634 setup_zone_migrate_reserve(zone
);
4635 spin_unlock_irqrestore(&zone
->lock
, flags
);
4638 /* update totalreserve_pages */
4639 calculate_totalreserve_pages();
4643 * The inactive anon list should be small enough that the VM never has to
4644 * do too much work, but large enough that each inactive page has a chance
4645 * to be referenced again before it is swapped out.
4647 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
4648 * INACTIVE_ANON pages on this zone's LRU, maintained by the
4649 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
4650 * the anonymous pages are kept on the inactive list.
4653 * memory ratio inactive anon
4654 * -------------------------------------
4663 void calculate_zone_inactive_ratio(struct zone
*zone
)
4665 unsigned int gb
, ratio
;
4667 /* Zone size in gigabytes */
4668 gb
= zone
->present_pages
>> (30 - PAGE_SHIFT
);
4670 ratio
= int_sqrt(10 * gb
);
4674 zone
->inactive_ratio
= ratio
;
4677 static void __init
setup_per_zone_inactive_ratio(void)
4682 calculate_zone_inactive_ratio(zone
);
4686 * Initialise min_free_kbytes.
4688 * For small machines we want it small (128k min). For large machines
4689 * we want it large (64MB max). But it is not linear, because network
4690 * bandwidth does not increase linearly with machine size. We use
4692 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
4693 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
4709 static int __init
init_per_zone_wmark_min(void)
4711 unsigned long lowmem_kbytes
;
4713 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
4715 min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
4716 if (min_free_kbytes
< 128)
4717 min_free_kbytes
= 128;
4718 if (min_free_kbytes
> 65536)
4719 min_free_kbytes
= 65536;
4720 setup_per_zone_wmarks();
4721 setup_per_zone_lowmem_reserve();
4722 setup_per_zone_inactive_ratio();
4725 module_init(init_per_zone_wmark_min
)
4728 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
4729 * that we can call two helper functions whenever min_free_kbytes
4732 int min_free_kbytes_sysctl_handler(ctl_table
*table
, int write
,
4733 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
4735 proc_dointvec(table
, write
, buffer
, length
, ppos
);
4737 setup_per_zone_wmarks();
4742 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table
*table
, int write
,
4743 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
4748 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
4753 zone
->min_unmapped_pages
= (zone
->present_pages
*
4754 sysctl_min_unmapped_ratio
) / 100;
4758 int sysctl_min_slab_ratio_sysctl_handler(ctl_table
*table
, int write
,
4759 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
4764 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
4769 zone
->min_slab_pages
= (zone
->present_pages
*
4770 sysctl_min_slab_ratio
) / 100;
4776 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
4777 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
4778 * whenever sysctl_lowmem_reserve_ratio changes.
4780 * The reserve ratio obviously has absolutely no relation with the
4781 * minimum watermarks. The lowmem reserve ratio can only make sense
4782 * if in function of the boot time zone sizes.
4784 int lowmem_reserve_ratio_sysctl_handler(ctl_table
*table
, int write
,
4785 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
4787 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
4788 setup_per_zone_lowmem_reserve();
4793 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
4794 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
4795 * can have before it gets flushed back to buddy allocator.
4798 int percpu_pagelist_fraction_sysctl_handler(ctl_table
*table
, int write
,
4799 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
4805 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
4806 if (!write
|| (ret
== -EINVAL
))
4808 for_each_populated_zone(zone
) {
4809 for_each_online_cpu(cpu
) {
4811 high
= zone
->present_pages
/ percpu_pagelist_fraction
;
4812 setup_pagelist_highmark(zone_pcp(zone
, cpu
), high
);
4818 int hashdist
= HASHDIST_DEFAULT
;
4821 static int __init
set_hashdist(char *str
)
4825 hashdist
= simple_strtoul(str
, &str
, 0);
4828 __setup("hashdist=", set_hashdist
);
4832 * allocate a large system hash table from bootmem
4833 * - it is assumed that the hash table must contain an exact power-of-2
4834 * quantity of entries
4835 * - limit is the number of hash buckets, not the total allocation size
4837 void *__init
alloc_large_system_hash(const char *tablename
,
4838 unsigned long bucketsize
,
4839 unsigned long numentries
,
4842 unsigned int *_hash_shift
,
4843 unsigned int *_hash_mask
,
4844 unsigned long limit
)
4846 unsigned long long max
= limit
;
4847 unsigned long log2qty
, size
;
4850 /* allow the kernel cmdline to have a say */
4852 /* round applicable memory size up to nearest megabyte */
4853 numentries
= nr_kernel_pages
;
4854 numentries
+= (1UL << (20 - PAGE_SHIFT
)) - 1;
4855 numentries
>>= 20 - PAGE_SHIFT
;
4856 numentries
<<= 20 - PAGE_SHIFT
;
4858 /* limit to 1 bucket per 2^scale bytes of low memory */
4859 if (scale
> PAGE_SHIFT
)
4860 numentries
>>= (scale
- PAGE_SHIFT
);
4862 numentries
<<= (PAGE_SHIFT
- scale
);
4864 /* Make sure we've got at least a 0-order allocation.. */
4865 if (unlikely(flags
& HASH_SMALL
)) {
4866 /* Makes no sense without HASH_EARLY */
4867 WARN_ON(!(flags
& HASH_EARLY
));
4868 if (!(numentries
>> *_hash_shift
)) {
4869 numentries
= 1UL << *_hash_shift
;
4870 BUG_ON(!numentries
);
4872 } else if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
4873 numentries
= PAGE_SIZE
/ bucketsize
;
4875 numentries
= roundup_pow_of_two(numentries
);
4877 /* limit allocation size to 1/16 total memory by default */
4879 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
4880 do_div(max
, bucketsize
);
4883 if (numentries
> max
)
4886 log2qty
= ilog2(numentries
);
4889 size
= bucketsize
<< log2qty
;
4890 if (flags
& HASH_EARLY
)
4891 table
= alloc_bootmem_nopanic(size
);
4893 table
= __vmalloc(size
, GFP_ATOMIC
, PAGE_KERNEL
);
4896 * If bucketsize is not a power-of-two, we may free
4897 * some pages at the end of hash table which
4898 * alloc_pages_exact() automatically does
4900 if (get_order(size
) < MAX_ORDER
) {
4901 table
= alloc_pages_exact(size
, GFP_ATOMIC
);
4902 kmemleak_alloc(table
, size
, 1, GFP_ATOMIC
);
4905 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
4908 panic("Failed to allocate %s hash table\n", tablename
);
4910 printk(KERN_INFO
"%s hash table entries: %d (order: %d, %lu bytes)\n",
4913 ilog2(size
) - PAGE_SHIFT
,
4917 *_hash_shift
= log2qty
;
4919 *_hash_mask
= (1 << log2qty
) - 1;
4924 /* Return a pointer to the bitmap storing bits affecting a block of pages */
4925 static inline unsigned long *get_pageblock_bitmap(struct zone
*zone
,
4928 #ifdef CONFIG_SPARSEMEM
4929 return __pfn_to_section(pfn
)->pageblock_flags
;
4931 return zone
->pageblock_flags
;
4932 #endif /* CONFIG_SPARSEMEM */
4935 static inline int pfn_to_bitidx(struct zone
*zone
, unsigned long pfn
)
4937 #ifdef CONFIG_SPARSEMEM
4938 pfn
&= (PAGES_PER_SECTION
-1);
4939 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
4941 pfn
= pfn
- zone
->zone_start_pfn
;
4942 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
4943 #endif /* CONFIG_SPARSEMEM */
4947 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
4948 * @page: The page within the block of interest
4949 * @start_bitidx: The first bit of interest to retrieve
4950 * @end_bitidx: The last bit of interest
4951 * returns pageblock_bits flags
4953 unsigned long get_pageblock_flags_group(struct page
*page
,
4954 int start_bitidx
, int end_bitidx
)
4957 unsigned long *bitmap
;
4958 unsigned long pfn
, bitidx
;
4959 unsigned long flags
= 0;
4960 unsigned long value
= 1;
4962 zone
= page_zone(page
);
4963 pfn
= page_to_pfn(page
);
4964 bitmap
= get_pageblock_bitmap(zone
, pfn
);
4965 bitidx
= pfn_to_bitidx(zone
, pfn
);
4967 for (; start_bitidx
<= end_bitidx
; start_bitidx
++, value
<<= 1)
4968 if (test_bit(bitidx
+ start_bitidx
, bitmap
))
4975 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
4976 * @page: The page within the block of interest
4977 * @start_bitidx: The first bit of interest
4978 * @end_bitidx: The last bit of interest
4979 * @flags: The flags to set
4981 void set_pageblock_flags_group(struct page
*page
, unsigned long flags
,
4982 int start_bitidx
, int end_bitidx
)
4985 unsigned long *bitmap
;
4986 unsigned long pfn
, bitidx
;
4987 unsigned long value
= 1;
4989 zone
= page_zone(page
);
4990 pfn
= page_to_pfn(page
);
4991 bitmap
= get_pageblock_bitmap(zone
, pfn
);
4992 bitidx
= pfn_to_bitidx(zone
, pfn
);
4993 VM_BUG_ON(pfn
< zone
->zone_start_pfn
);
4994 VM_BUG_ON(pfn
>= zone
->zone_start_pfn
+ zone
->spanned_pages
);
4996 for (; start_bitidx
<= end_bitidx
; start_bitidx
++, value
<<= 1)
4998 __set_bit(bitidx
+ start_bitidx
, bitmap
);
5000 __clear_bit(bitidx
+ start_bitidx
, bitmap
);
5004 * This is designed as sub function...plz see page_isolation.c also.
5005 * set/clear page block's type to be ISOLATE.
5006 * page allocater never alloc memory from ISOLATE block.
5009 int set_migratetype_isolate(struct page
*page
)
5012 struct page
*curr_page
;
5013 unsigned long flags
, pfn
, iter
;
5014 unsigned long immobile
= 0;
5015 struct memory_isolate_notify arg
;
5020 zone
= page_zone(page
);
5021 zone_idx
= zone_idx(zone
);
5023 spin_lock_irqsave(&zone
->lock
, flags
);
5024 if (get_pageblock_migratetype(page
) == MIGRATE_MOVABLE
||
5025 zone_idx
== ZONE_MOVABLE
) {
5030 pfn
= page_to_pfn(page
);
5031 arg
.start_pfn
= pfn
;
5032 arg
.nr_pages
= pageblock_nr_pages
;
5033 arg
.pages_found
= 0;
5036 * It may be possible to isolate a pageblock even if the
5037 * migratetype is not MIGRATE_MOVABLE. The memory isolation
5038 * notifier chain is used by balloon drivers to return the
5039 * number of pages in a range that are held by the balloon
5040 * driver to shrink memory. If all the pages are accounted for
5041 * by balloons, are free, or on the LRU, isolation can continue.
5042 * Later, for example, when memory hotplug notifier runs, these
5043 * pages reported as "can be isolated" should be isolated(freed)
5044 * by the balloon driver through the memory notifier chain.
5046 notifier_ret
= memory_isolate_notify(MEM_ISOLATE_COUNT
, &arg
);
5047 notifier_ret
= notifier_to_errno(notifier_ret
);
5048 if (notifier_ret
|| !arg
.pages_found
)
5051 for (iter
= pfn
; iter
< (pfn
+ pageblock_nr_pages
); iter
++) {
5052 if (!pfn_valid_within(pfn
))
5055 curr_page
= pfn_to_page(iter
);
5056 if (!page_count(curr_page
) || PageLRU(curr_page
))
5062 if (arg
.pages_found
== immobile
)
5067 set_pageblock_migratetype(page
, MIGRATE_ISOLATE
);
5068 move_freepages_block(zone
, page
, MIGRATE_ISOLATE
);
5071 spin_unlock_irqrestore(&zone
->lock
, flags
);
5077 void unset_migratetype_isolate(struct page
*page
)
5080 unsigned long flags
;
5081 zone
= page_zone(page
);
5082 spin_lock_irqsave(&zone
->lock
, flags
);
5083 if (get_pageblock_migratetype(page
) != MIGRATE_ISOLATE
)
5085 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
5086 move_freepages_block(zone
, page
, MIGRATE_MOVABLE
);
5088 spin_unlock_irqrestore(&zone
->lock
, flags
);
5091 #ifdef CONFIG_MEMORY_HOTREMOVE
5093 * All pages in the range must be isolated before calling this.
5096 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
5102 unsigned long flags
;
5103 /* find the first valid pfn */
5104 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
5109 zone
= page_zone(pfn_to_page(pfn
));
5110 spin_lock_irqsave(&zone
->lock
, flags
);
5112 while (pfn
< end_pfn
) {
5113 if (!pfn_valid(pfn
)) {
5117 page
= pfn_to_page(pfn
);
5118 BUG_ON(page_count(page
));
5119 BUG_ON(!PageBuddy(page
));
5120 order
= page_order(page
);
5121 #ifdef CONFIG_DEBUG_VM
5122 printk(KERN_INFO
"remove from free list %lx %d %lx\n",
5123 pfn
, 1 << order
, end_pfn
);
5125 list_del(&page
->lru
);
5126 rmv_page_order(page
);
5127 zone
->free_area
[order
].nr_free
--;
5128 __mod_zone_page_state(zone
, NR_FREE_PAGES
,
5130 for (i
= 0; i
< (1 << order
); i
++)
5131 SetPageReserved((page
+i
));
5132 pfn
+= (1 << order
);
5134 spin_unlock_irqrestore(&zone
->lock
, flags
);
5138 #ifdef CONFIG_MEMORY_FAILURE
5139 bool is_free_buddy_page(struct page
*page
)
5141 struct zone
*zone
= page_zone(page
);
5142 unsigned long pfn
= page_to_pfn(page
);
5143 unsigned long flags
;
5146 spin_lock_irqsave(&zone
->lock
, flags
);
5147 for (order
= 0; order
< MAX_ORDER
; order
++) {
5148 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
5150 if (PageBuddy(page_head
) && page_order(page_head
) >= order
)
5153 spin_unlock_irqrestore(&zone
->lock
, flags
);
5155 return order
< MAX_ORDER
;