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>
52 #include <asm/tlbflush.h>
53 #include <asm/div64.h>
57 * Array of node states.
59 nodemask_t node_states
[NR_NODE_STATES
] __read_mostly
= {
60 [N_POSSIBLE
] = NODE_MASK_ALL
,
61 [N_ONLINE
] = { { [0] = 1UL } },
63 [N_NORMAL_MEMORY
] = { { [0] = 1UL } },
65 [N_HIGH_MEMORY
] = { { [0] = 1UL } },
67 [N_CPU
] = { { [0] = 1UL } },
70 EXPORT_SYMBOL(node_states
);
72 unsigned long totalram_pages __read_mostly
;
73 unsigned long totalreserve_pages __read_mostly
;
74 unsigned long highest_memmap_pfn __read_mostly
;
75 int percpu_pagelist_fraction
;
76 gfp_t gfp_allowed_mask __read_mostly
= GFP_BOOT_MASK
;
78 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
79 int pageblock_order __read_mostly
;
82 static void __free_pages_ok(struct page
*page
, unsigned int order
);
85 * results with 256, 32 in the lowmem_reserve sysctl:
86 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
87 * 1G machine -> (16M dma, 784M normal, 224M high)
88 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
89 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
90 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
92 * TBD: should special case ZONE_DMA32 machines here - in those we normally
93 * don't need any ZONE_NORMAL reservation
95 int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
-1] = {
96 #ifdef CONFIG_ZONE_DMA
99 #ifdef CONFIG_ZONE_DMA32
102 #ifdef CONFIG_HIGHMEM
108 EXPORT_SYMBOL(totalram_pages
);
110 static char * const zone_names
[MAX_NR_ZONES
] = {
111 #ifdef CONFIG_ZONE_DMA
114 #ifdef CONFIG_ZONE_DMA32
118 #ifdef CONFIG_HIGHMEM
124 int min_free_kbytes
= 1024;
126 unsigned long __meminitdata nr_kernel_pages
;
127 unsigned long __meminitdata nr_all_pages
;
128 static unsigned long __meminitdata dma_reserve
;
130 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
132 * MAX_ACTIVE_REGIONS determines the maximum number of distinct
133 * ranges of memory (RAM) that may be registered with add_active_range().
134 * Ranges passed to add_active_range() will be merged if possible
135 * so the number of times add_active_range() can be called is
136 * related to the number of nodes and the number of holes
138 #ifdef CONFIG_MAX_ACTIVE_REGIONS
139 /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */
140 #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS
142 #if MAX_NUMNODES >= 32
143 /* If there can be many nodes, allow up to 50 holes per node */
144 #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50)
146 /* By default, allow up to 256 distinct regions */
147 #define MAX_ACTIVE_REGIONS 256
151 static struct node_active_region __meminitdata early_node_map
[MAX_ACTIVE_REGIONS
];
152 static int __meminitdata nr_nodemap_entries
;
153 static unsigned long __meminitdata arch_zone_lowest_possible_pfn
[MAX_NR_ZONES
];
154 static unsigned long __meminitdata arch_zone_highest_possible_pfn
[MAX_NR_ZONES
];
155 static unsigned long __initdata required_kernelcore
;
156 static unsigned long __initdata required_movablecore
;
157 static unsigned long __meminitdata zone_movable_pfn
[MAX_NUMNODES
];
159 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
161 EXPORT_SYMBOL(movable_zone
);
162 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
165 int nr_node_ids __read_mostly
= MAX_NUMNODES
;
166 int nr_online_nodes __read_mostly
= 1;
167 EXPORT_SYMBOL(nr_node_ids
);
168 EXPORT_SYMBOL(nr_online_nodes
);
171 int page_group_by_mobility_disabled __read_mostly
;
173 static void set_pageblock_migratetype(struct page
*page
, int migratetype
)
176 if (unlikely(page_group_by_mobility_disabled
))
177 migratetype
= MIGRATE_UNMOVABLE
;
179 set_pageblock_flags_group(page
, (unsigned long)migratetype
,
180 PB_migrate
, PB_migrate_end
);
183 bool oom_killer_disabled __read_mostly
;
185 #ifdef CONFIG_DEBUG_VM
186 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
190 unsigned long pfn
= page_to_pfn(page
);
193 seq
= zone_span_seqbegin(zone
);
194 if (pfn
>= zone
->zone_start_pfn
+ zone
->spanned_pages
)
196 else if (pfn
< zone
->zone_start_pfn
)
198 } while (zone_span_seqretry(zone
, seq
));
203 static int page_is_consistent(struct zone
*zone
, struct page
*page
)
205 if (!pfn_valid_within(page_to_pfn(page
)))
207 if (zone
!= page_zone(page
))
213 * Temporary debugging check for pages not lying within a given zone.
215 static int bad_range(struct zone
*zone
, struct page
*page
)
217 if (page_outside_zone_boundaries(zone
, page
))
219 if (!page_is_consistent(zone
, page
))
225 static inline int bad_range(struct zone
*zone
, struct page
*page
)
231 static void bad_page(struct page
*page
)
233 static unsigned long resume
;
234 static unsigned long nr_shown
;
235 static unsigned long nr_unshown
;
238 * Allow a burst of 60 reports, then keep quiet for that minute;
239 * or allow a steady drip of one report per second.
241 if (nr_shown
== 60) {
242 if (time_before(jiffies
, resume
)) {
248 "BUG: Bad page state: %lu messages suppressed\n",
255 resume
= jiffies
+ 60 * HZ
;
257 printk(KERN_ALERT
"BUG: Bad page state in process %s pfn:%05lx\n",
258 current
->comm
, page_to_pfn(page
));
260 "page:%p flags:%p count:%d mapcount:%d mapping:%p index:%lx\n",
261 page
, (void *)page
->flags
, page_count(page
),
262 page_mapcount(page
), page
->mapping
, page
->index
);
266 /* Leave bad fields for debug, except PageBuddy could make trouble */
267 __ClearPageBuddy(page
);
268 add_taint(TAINT_BAD_PAGE
);
272 * Higher-order pages are called "compound pages". They are structured thusly:
274 * The first PAGE_SIZE page is called the "head page".
276 * The remaining PAGE_SIZE pages are called "tail pages".
278 * All pages have PG_compound set. All pages have their ->private pointing at
279 * the head page (even the head page has this).
281 * The first tail page's ->lru.next holds the address of the compound page's
282 * put_page() function. Its ->lru.prev holds the order of allocation.
283 * This usage means that zero-order pages may not be compound.
286 static void free_compound_page(struct page
*page
)
288 __free_pages_ok(page
, compound_order(page
));
291 void prep_compound_page(struct page
*page
, unsigned long order
)
294 int nr_pages
= 1 << order
;
296 set_compound_page_dtor(page
, free_compound_page
);
297 set_compound_order(page
, order
);
299 for (i
= 1; i
< nr_pages
; i
++) {
300 struct page
*p
= page
+ i
;
303 p
->first_page
= page
;
307 static int destroy_compound_page(struct page
*page
, unsigned long order
)
310 int nr_pages
= 1 << order
;
313 if (unlikely(compound_order(page
) != order
) ||
314 unlikely(!PageHead(page
))) {
319 __ClearPageHead(page
);
321 for (i
= 1; i
< nr_pages
; i
++) {
322 struct page
*p
= page
+ i
;
324 if (unlikely(!PageTail(p
) || (p
->first_page
!= page
))) {
334 static inline void prep_zero_page(struct page
*page
, int order
, gfp_t gfp_flags
)
339 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
340 * and __GFP_HIGHMEM from hard or soft interrupt context.
342 VM_BUG_ON((gfp_flags
& __GFP_HIGHMEM
) && in_interrupt());
343 for (i
= 0; i
< (1 << order
); i
++)
344 clear_highpage(page
+ i
);
347 static inline void set_page_order(struct page
*page
, int order
)
349 set_page_private(page
, order
);
350 __SetPageBuddy(page
);
353 static inline void rmv_page_order(struct page
*page
)
355 __ClearPageBuddy(page
);
356 set_page_private(page
, 0);
360 * Locate the struct page for both the matching buddy in our
361 * pair (buddy1) and the combined O(n+1) page they form (page).
363 * 1) Any buddy B1 will have an order O twin B2 which satisfies
364 * the following equation:
366 * For example, if the starting buddy (buddy2) is #8 its order
368 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
370 * 2) Any buddy B will have an order O+1 parent P which
371 * satisfies the following equation:
374 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
376 static inline struct page
*
377 __page_find_buddy(struct page
*page
, unsigned long page_idx
, unsigned int order
)
379 unsigned long buddy_idx
= page_idx
^ (1 << order
);
381 return page
+ (buddy_idx
- page_idx
);
384 static inline unsigned long
385 __find_combined_index(unsigned long page_idx
, unsigned int order
)
387 return (page_idx
& ~(1 << order
));
391 * This function checks whether a page is free && is the buddy
392 * we can do coalesce a page and its buddy if
393 * (a) the buddy is not in a hole &&
394 * (b) the buddy is in the buddy system &&
395 * (c) a page and its buddy have the same order &&
396 * (d) a page and its buddy are in the same zone.
398 * For recording whether a page is in the buddy system, we use PG_buddy.
399 * Setting, clearing, and testing PG_buddy is serialized by zone->lock.
401 * For recording page's order, we use page_private(page).
403 static inline int page_is_buddy(struct page
*page
, struct page
*buddy
,
406 if (!pfn_valid_within(page_to_pfn(buddy
)))
409 if (page_zone_id(page
) != page_zone_id(buddy
))
412 if (PageBuddy(buddy
) && page_order(buddy
) == order
) {
413 VM_BUG_ON(page_count(buddy
) != 0);
420 * Freeing function for a buddy system allocator.
422 * The concept of a buddy system is to maintain direct-mapped table
423 * (containing bit values) for memory blocks of various "orders".
424 * The bottom level table contains the map for the smallest allocatable
425 * units of memory (here, pages), and each level above it describes
426 * pairs of units from the levels below, hence, "buddies".
427 * At a high level, all that happens here is marking the table entry
428 * at the bottom level available, and propagating the changes upward
429 * as necessary, plus some accounting needed to play nicely with other
430 * parts of the VM system.
431 * At each level, we keep a list of pages, which are heads of continuous
432 * free pages of length of (1 << order) and marked with PG_buddy. Page's
433 * order is recorded in page_private(page) field.
434 * So when we are allocating or freeing one, we can derive the state of the
435 * other. That is, if we allocate a small block, and both were
436 * free, the remainder of the region must be split into blocks.
437 * If a block is freed, and its buddy is also free, then this
438 * triggers coalescing into a block of larger size.
443 static inline void __free_one_page(struct page
*page
,
444 struct zone
*zone
, unsigned int order
,
447 unsigned long page_idx
;
449 if (unlikely(PageCompound(page
)))
450 if (unlikely(destroy_compound_page(page
, order
)))
453 VM_BUG_ON(migratetype
== -1);
455 page_idx
= page_to_pfn(page
) & ((1 << MAX_ORDER
) - 1);
457 VM_BUG_ON(page_idx
& ((1 << order
) - 1));
458 VM_BUG_ON(bad_range(zone
, page
));
460 while (order
< MAX_ORDER
-1) {
461 unsigned long combined_idx
;
464 buddy
= __page_find_buddy(page
, page_idx
, order
);
465 if (!page_is_buddy(page
, buddy
, order
))
468 /* Our buddy is free, merge with it and move up one order. */
469 list_del(&buddy
->lru
);
470 zone
->free_area
[order
].nr_free
--;
471 rmv_page_order(buddy
);
472 combined_idx
= __find_combined_index(page_idx
, order
);
473 page
= page
+ (combined_idx
- page_idx
);
474 page_idx
= combined_idx
;
477 set_page_order(page
, order
);
479 &zone
->free_area
[order
].free_list
[migratetype
]);
480 zone
->free_area
[order
].nr_free
++;
483 #ifdef CONFIG_HAVE_MLOCKED_PAGE_BIT
485 * free_page_mlock() -- clean up attempts to free and mlocked() page.
486 * Page should not be on lru, so no need to fix that up.
487 * free_pages_check() will verify...
489 static inline void free_page_mlock(struct page
*page
)
491 __dec_zone_page_state(page
, NR_MLOCK
);
492 __count_vm_event(UNEVICTABLE_MLOCKFREED
);
495 static void free_page_mlock(struct page
*page
) { }
498 static inline int free_pages_check(struct page
*page
)
500 if (unlikely(page_mapcount(page
) |
501 (page
->mapping
!= NULL
) |
502 (atomic_read(&page
->_count
) != 0) |
503 (page
->flags
& PAGE_FLAGS_CHECK_AT_FREE
))) {
507 if (page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)
508 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
513 * Frees a list of pages.
514 * Assumes all pages on list are in same zone, and of same order.
515 * count is the number of pages to free.
517 * If the zone was previously in an "all pages pinned" state then look to
518 * see if this freeing clears that state.
520 * And clear the zone's pages_scanned counter, to hold off the "all pages are
521 * pinned" detection logic.
523 static void free_pages_bulk(struct zone
*zone
, int count
,
524 struct list_head
*list
, int order
)
526 spin_lock(&zone
->lock
);
527 zone_clear_flag(zone
, ZONE_ALL_UNRECLAIMABLE
);
528 zone
->pages_scanned
= 0;
530 __mod_zone_page_state(zone
, NR_FREE_PAGES
, count
<< order
);
534 VM_BUG_ON(list_empty(list
));
535 page
= list_entry(list
->prev
, struct page
, lru
);
536 /* have to delete it as __free_one_page list manipulates */
537 list_del(&page
->lru
);
538 __free_one_page(page
, zone
, order
, page_private(page
));
540 spin_unlock(&zone
->lock
);
543 static void free_one_page(struct zone
*zone
, struct page
*page
, int order
,
546 spin_lock(&zone
->lock
);
547 zone_clear_flag(zone
, ZONE_ALL_UNRECLAIMABLE
);
548 zone
->pages_scanned
= 0;
550 __mod_zone_page_state(zone
, NR_FREE_PAGES
, 1 << order
);
551 __free_one_page(page
, zone
, order
, migratetype
);
552 spin_unlock(&zone
->lock
);
555 static void __free_pages_ok(struct page
*page
, unsigned int order
)
560 int wasMlocked
= __TestClearPageMlocked(page
);
562 kmemcheck_free_shadow(page
, order
);
564 for (i
= 0 ; i
< (1 << order
) ; ++i
)
565 bad
+= free_pages_check(page
+ i
);
569 if (!PageHighMem(page
)) {
570 debug_check_no_locks_freed(page_address(page
),PAGE_SIZE
<<order
);
571 debug_check_no_obj_freed(page_address(page
),
574 arch_free_page(page
, order
);
575 kernel_map_pages(page
, 1 << order
, 0);
577 local_irq_save(flags
);
578 if (unlikely(wasMlocked
))
579 free_page_mlock(page
);
580 __count_vm_events(PGFREE
, 1 << order
);
581 free_one_page(page_zone(page
), page
, order
,
582 get_pageblock_migratetype(page
));
583 local_irq_restore(flags
);
587 * permit the bootmem allocator to evade page validation on high-order frees
589 void __meminit
__free_pages_bootmem(struct page
*page
, unsigned int order
)
592 __ClearPageReserved(page
);
593 set_page_count(page
, 0);
594 set_page_refcounted(page
);
600 for (loop
= 0; loop
< BITS_PER_LONG
; loop
++) {
601 struct page
*p
= &page
[loop
];
603 if (loop
+ 1 < BITS_PER_LONG
)
605 __ClearPageReserved(p
);
606 set_page_count(p
, 0);
609 set_page_refcounted(page
);
610 __free_pages(page
, order
);
616 * The order of subdivision here is critical for the IO subsystem.
617 * Please do not alter this order without good reasons and regression
618 * testing. Specifically, as large blocks of memory are subdivided,
619 * the order in which smaller blocks are delivered depends on the order
620 * they're subdivided in this function. This is the primary factor
621 * influencing the order in which pages are delivered to the IO
622 * subsystem according to empirical testing, and this is also justified
623 * by considering the behavior of a buddy system containing a single
624 * large block of memory acted on by a series of small allocations.
625 * This behavior is a critical factor in sglist merging's success.
629 static inline void expand(struct zone
*zone
, struct page
*page
,
630 int low
, int high
, struct free_area
*area
,
633 unsigned long size
= 1 << high
;
639 VM_BUG_ON(bad_range(zone
, &page
[size
]));
640 list_add(&page
[size
].lru
, &area
->free_list
[migratetype
]);
642 set_page_order(&page
[size
], high
);
647 * This page is about to be returned from the page allocator
649 static int prep_new_page(struct page
*page
, int order
, gfp_t gfp_flags
)
651 if (unlikely(page_mapcount(page
) |
652 (page
->mapping
!= NULL
) |
653 (atomic_read(&page
->_count
) != 0) |
654 (page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
))) {
659 set_page_private(page
, 0);
660 set_page_refcounted(page
);
662 arch_alloc_page(page
, order
);
663 kernel_map_pages(page
, 1 << order
, 1);
665 if (gfp_flags
& __GFP_ZERO
)
666 prep_zero_page(page
, order
, gfp_flags
);
668 if (order
&& (gfp_flags
& __GFP_COMP
))
669 prep_compound_page(page
, order
);
675 * Go through the free lists for the given migratetype and remove
676 * the smallest available page from the freelists
679 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
682 unsigned int current_order
;
683 struct free_area
* area
;
686 /* Find a page of the appropriate size in the preferred list */
687 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
688 area
= &(zone
->free_area
[current_order
]);
689 if (list_empty(&area
->free_list
[migratetype
]))
692 page
= list_entry(area
->free_list
[migratetype
].next
,
694 list_del(&page
->lru
);
695 rmv_page_order(page
);
697 expand(zone
, page
, order
, current_order
, area
, migratetype
);
706 * This array describes the order lists are fallen back to when
707 * the free lists for the desirable migrate type are depleted
709 static int fallbacks
[MIGRATE_TYPES
][MIGRATE_TYPES
-1] = {
710 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
711 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
712 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_RESERVE
},
713 [MIGRATE_RESERVE
] = { MIGRATE_RESERVE
, MIGRATE_RESERVE
, MIGRATE_RESERVE
}, /* Never used */
717 * Move the free pages in a range to the free lists of the requested type.
718 * Note that start_page and end_pages are not aligned on a pageblock
719 * boundary. If alignment is required, use move_freepages_block()
721 static int move_freepages(struct zone
*zone
,
722 struct page
*start_page
, struct page
*end_page
,
729 #ifndef CONFIG_HOLES_IN_ZONE
731 * page_zone is not safe to call in this context when
732 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
733 * anyway as we check zone boundaries in move_freepages_block().
734 * Remove at a later date when no bug reports exist related to
735 * grouping pages by mobility
737 BUG_ON(page_zone(start_page
) != page_zone(end_page
));
740 for (page
= start_page
; page
<= end_page
;) {
741 /* Make sure we are not inadvertently changing nodes */
742 VM_BUG_ON(page_to_nid(page
) != zone_to_nid(zone
));
744 if (!pfn_valid_within(page_to_pfn(page
))) {
749 if (!PageBuddy(page
)) {
754 order
= page_order(page
);
755 list_del(&page
->lru
);
757 &zone
->free_area
[order
].free_list
[migratetype
]);
759 pages_moved
+= 1 << order
;
765 static int move_freepages_block(struct zone
*zone
, struct page
*page
,
768 unsigned long start_pfn
, end_pfn
;
769 struct page
*start_page
, *end_page
;
771 start_pfn
= page_to_pfn(page
);
772 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
773 start_page
= pfn_to_page(start_pfn
);
774 end_page
= start_page
+ pageblock_nr_pages
- 1;
775 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
777 /* Do not cross zone boundaries */
778 if (start_pfn
< zone
->zone_start_pfn
)
780 if (end_pfn
>= zone
->zone_start_pfn
+ zone
->spanned_pages
)
783 return move_freepages(zone
, start_page
, end_page
, migratetype
);
786 static void change_pageblock_range(struct page
*pageblock_page
,
787 int start_order
, int migratetype
)
789 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
791 while (nr_pageblocks
--) {
792 set_pageblock_migratetype(pageblock_page
, migratetype
);
793 pageblock_page
+= pageblock_nr_pages
;
797 /* Remove an element from the buddy allocator from the fallback list */
798 static inline struct page
*
799 __rmqueue_fallback(struct zone
*zone
, int order
, int start_migratetype
)
801 struct free_area
* area
;
806 /* Find the largest possible block of pages in the other list */
807 for (current_order
= MAX_ORDER
-1; current_order
>= order
;
809 for (i
= 0; i
< MIGRATE_TYPES
- 1; i
++) {
810 migratetype
= fallbacks
[start_migratetype
][i
];
812 /* MIGRATE_RESERVE handled later if necessary */
813 if (migratetype
== MIGRATE_RESERVE
)
816 area
= &(zone
->free_area
[current_order
]);
817 if (list_empty(&area
->free_list
[migratetype
]))
820 page
= list_entry(area
->free_list
[migratetype
].next
,
825 * If breaking a large block of pages, move all free
826 * pages to the preferred allocation list. If falling
827 * back for a reclaimable kernel allocation, be more
828 * agressive about taking ownership of free pages
830 if (unlikely(current_order
>= (pageblock_order
>> 1)) ||
831 start_migratetype
== MIGRATE_RECLAIMABLE
||
832 page_group_by_mobility_disabled
) {
834 pages
= move_freepages_block(zone
, page
,
837 /* Claim the whole block if over half of it is free */
838 if (pages
>= (1 << (pageblock_order
-1)) ||
839 page_group_by_mobility_disabled
)
840 set_pageblock_migratetype(page
,
843 migratetype
= start_migratetype
;
846 /* Remove the page from the freelists */
847 list_del(&page
->lru
);
848 rmv_page_order(page
);
850 /* Take ownership for orders >= pageblock_order */
851 if (current_order
>= pageblock_order
)
852 change_pageblock_range(page
, current_order
,
855 expand(zone
, page
, order
, current_order
, area
, migratetype
);
864 * Do the hard work of removing an element from the buddy allocator.
865 * Call me with the zone->lock already held.
867 static struct page
*__rmqueue(struct zone
*zone
, unsigned int order
,
873 page
= __rmqueue_smallest(zone
, order
, migratetype
);
875 if (unlikely(!page
) && migratetype
!= MIGRATE_RESERVE
) {
876 page
= __rmqueue_fallback(zone
, order
, migratetype
);
879 * Use MIGRATE_RESERVE rather than fail an allocation. goto
880 * is used because __rmqueue_smallest is an inline function
881 * and we want just one call site
884 migratetype
= MIGRATE_RESERVE
;
893 * Obtain a specified number of elements from the buddy allocator, all under
894 * a single hold of the lock, for efficiency. Add them to the supplied list.
895 * Returns the number of new pages which were placed at *list.
897 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
898 unsigned long count
, struct list_head
*list
,
899 int migratetype
, int cold
)
903 spin_lock(&zone
->lock
);
904 for (i
= 0; i
< count
; ++i
) {
905 struct page
*page
= __rmqueue(zone
, order
, migratetype
);
906 if (unlikely(page
== NULL
))
910 * Split buddy pages returned by expand() are received here
911 * in physical page order. The page is added to the callers and
912 * list and the list head then moves forward. From the callers
913 * perspective, the linked list is ordered by page number in
914 * some conditions. This is useful for IO devices that can
915 * merge IO requests if the physical pages are ordered
918 if (likely(cold
== 0))
919 list_add(&page
->lru
, list
);
921 list_add_tail(&page
->lru
, list
);
922 set_page_private(page
, migratetype
);
925 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
926 spin_unlock(&zone
->lock
);
932 * Called from the vmstat counter updater to drain pagesets of this
933 * currently executing processor on remote nodes after they have
936 * Note that this function must be called with the thread pinned to
937 * a single processor.
939 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
944 local_irq_save(flags
);
945 if (pcp
->count
>= pcp
->batch
)
946 to_drain
= pcp
->batch
;
948 to_drain
= pcp
->count
;
949 free_pages_bulk(zone
, to_drain
, &pcp
->list
, 0);
950 pcp
->count
-= to_drain
;
951 local_irq_restore(flags
);
956 * Drain pages of the indicated processor.
958 * The processor must either be the current processor and the
959 * thread pinned to the current processor or a processor that
962 static void drain_pages(unsigned int cpu
)
967 for_each_populated_zone(zone
) {
968 struct per_cpu_pageset
*pset
;
969 struct per_cpu_pages
*pcp
;
971 pset
= zone_pcp(zone
, cpu
);
974 local_irq_save(flags
);
975 free_pages_bulk(zone
, pcp
->count
, &pcp
->list
, 0);
977 local_irq_restore(flags
);
982 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
984 void drain_local_pages(void *arg
)
986 drain_pages(smp_processor_id());
990 * Spill all the per-cpu pages from all CPUs back into the buddy allocator
992 void drain_all_pages(void)
994 on_each_cpu(drain_local_pages
, NULL
, 1);
997 #ifdef CONFIG_HIBERNATION
999 void mark_free_pages(struct zone
*zone
)
1001 unsigned long pfn
, max_zone_pfn
;
1002 unsigned long flags
;
1004 struct list_head
*curr
;
1006 if (!zone
->spanned_pages
)
1009 spin_lock_irqsave(&zone
->lock
, flags
);
1011 max_zone_pfn
= zone
->zone_start_pfn
+ zone
->spanned_pages
;
1012 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
1013 if (pfn_valid(pfn
)) {
1014 struct page
*page
= pfn_to_page(pfn
);
1016 if (!swsusp_page_is_forbidden(page
))
1017 swsusp_unset_page_free(page
);
1020 for_each_migratetype_order(order
, t
) {
1021 list_for_each(curr
, &zone
->free_area
[order
].free_list
[t
]) {
1024 pfn
= page_to_pfn(list_entry(curr
, struct page
, lru
));
1025 for (i
= 0; i
< (1UL << order
); i
++)
1026 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
1029 spin_unlock_irqrestore(&zone
->lock
, flags
);
1031 #endif /* CONFIG_PM */
1034 * Free a 0-order page
1036 static void free_hot_cold_page(struct page
*page
, int cold
)
1038 struct zone
*zone
= page_zone(page
);
1039 struct per_cpu_pages
*pcp
;
1040 unsigned long flags
;
1041 int wasMlocked
= __TestClearPageMlocked(page
);
1043 kmemcheck_free_shadow(page
, 0);
1046 page
->mapping
= NULL
;
1047 if (free_pages_check(page
))
1050 if (!PageHighMem(page
)) {
1051 debug_check_no_locks_freed(page_address(page
), PAGE_SIZE
);
1052 debug_check_no_obj_freed(page_address(page
), PAGE_SIZE
);
1054 arch_free_page(page
, 0);
1055 kernel_map_pages(page
, 1, 0);
1057 pcp
= &zone_pcp(zone
, get_cpu())->pcp
;
1058 set_page_private(page
, get_pageblock_migratetype(page
));
1059 local_irq_save(flags
);
1060 if (unlikely(wasMlocked
))
1061 free_page_mlock(page
);
1062 __count_vm_event(PGFREE
);
1065 list_add_tail(&page
->lru
, &pcp
->list
);
1067 list_add(&page
->lru
, &pcp
->list
);
1069 if (pcp
->count
>= pcp
->high
) {
1070 free_pages_bulk(zone
, pcp
->batch
, &pcp
->list
, 0);
1071 pcp
->count
-= pcp
->batch
;
1073 local_irq_restore(flags
);
1077 void free_hot_page(struct page
*page
)
1079 free_hot_cold_page(page
, 0);
1083 * split_page takes a non-compound higher-order page, and splits it into
1084 * n (1<<order) sub-pages: page[0..n]
1085 * Each sub-page must be freed individually.
1087 * Note: this is probably too low level an operation for use in drivers.
1088 * Please consult with lkml before using this in your driver.
1090 void split_page(struct page
*page
, unsigned int order
)
1094 VM_BUG_ON(PageCompound(page
));
1095 VM_BUG_ON(!page_count(page
));
1097 #ifdef CONFIG_KMEMCHECK
1099 * Split shadow pages too, because free(page[0]) would
1100 * otherwise free the whole shadow.
1102 if (kmemcheck_page_is_tracked(page
))
1103 split_page(virt_to_page(page
[0].shadow
), order
);
1106 for (i
= 1; i
< (1 << order
); i
++)
1107 set_page_refcounted(page
+ i
);
1111 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1112 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1116 struct page
*buffered_rmqueue(struct zone
*preferred_zone
,
1117 struct zone
*zone
, int order
, gfp_t gfp_flags
,
1120 unsigned long flags
;
1122 int cold
= !!(gfp_flags
& __GFP_COLD
);
1127 if (likely(order
== 0)) {
1128 struct per_cpu_pages
*pcp
;
1130 pcp
= &zone_pcp(zone
, cpu
)->pcp
;
1131 local_irq_save(flags
);
1133 pcp
->count
= rmqueue_bulk(zone
, 0,
1134 pcp
->batch
, &pcp
->list
,
1136 if (unlikely(!pcp
->count
))
1140 /* Find a page of the appropriate migrate type */
1142 list_for_each_entry_reverse(page
, &pcp
->list
, lru
)
1143 if (page_private(page
) == migratetype
)
1146 list_for_each_entry(page
, &pcp
->list
, lru
)
1147 if (page_private(page
) == migratetype
)
1151 /* Allocate more to the pcp list if necessary */
1152 if (unlikely(&page
->lru
== &pcp
->list
)) {
1153 int get_one_page
= 0;
1155 pcp
->count
+= rmqueue_bulk(zone
, 0,
1156 pcp
->batch
, &pcp
->list
,
1158 list_for_each_entry(page
, &pcp
->list
, lru
) {
1159 if (get_pageblock_migratetype(page
) !=
1169 list_del(&page
->lru
);
1172 if (unlikely(gfp_flags
& __GFP_NOFAIL
)) {
1174 * __GFP_NOFAIL is not to be used in new code.
1176 * All __GFP_NOFAIL callers should be fixed so that they
1177 * properly detect and handle allocation failures.
1179 * We most definitely don't want callers attempting to
1180 * allocate greater than order-1 page units with
1183 WARN_ON_ONCE(order
> 1);
1185 spin_lock_irqsave(&zone
->lock
, flags
);
1186 page
= __rmqueue(zone
, order
, migratetype
);
1187 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(1 << order
));
1188 spin_unlock(&zone
->lock
);
1193 __count_zone_vm_events(PGALLOC
, zone
, 1 << order
);
1194 zone_statistics(preferred_zone
, zone
);
1195 local_irq_restore(flags
);
1198 VM_BUG_ON(bad_range(zone
, page
));
1199 if (prep_new_page(page
, order
, gfp_flags
))
1204 local_irq_restore(flags
);
1209 /* The ALLOC_WMARK bits are used as an index to zone->watermark */
1210 #define ALLOC_WMARK_MIN WMARK_MIN
1211 #define ALLOC_WMARK_LOW WMARK_LOW
1212 #define ALLOC_WMARK_HIGH WMARK_HIGH
1213 #define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */
1215 /* Mask to get the watermark bits */
1216 #define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1)
1218 #define ALLOC_HARDER 0x10 /* try to alloc harder */
1219 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
1220 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
1222 #ifdef CONFIG_FAIL_PAGE_ALLOC
1224 static struct fail_page_alloc_attr
{
1225 struct fault_attr attr
;
1227 u32 ignore_gfp_highmem
;
1228 u32 ignore_gfp_wait
;
1231 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1233 struct dentry
*ignore_gfp_highmem_file
;
1234 struct dentry
*ignore_gfp_wait_file
;
1235 struct dentry
*min_order_file
;
1237 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1239 } fail_page_alloc
= {
1240 .attr
= FAULT_ATTR_INITIALIZER
,
1241 .ignore_gfp_wait
= 1,
1242 .ignore_gfp_highmem
= 1,
1246 static int __init
setup_fail_page_alloc(char *str
)
1248 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
1250 __setup("fail_page_alloc=", setup_fail_page_alloc
);
1252 static int should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1254 if (order
< fail_page_alloc
.min_order
)
1256 if (gfp_mask
& __GFP_NOFAIL
)
1258 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
1260 if (fail_page_alloc
.ignore_gfp_wait
&& (gfp_mask
& __GFP_WAIT
))
1263 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
1266 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1268 static int __init
fail_page_alloc_debugfs(void)
1270 mode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
1274 err
= init_fault_attr_dentries(&fail_page_alloc
.attr
,
1278 dir
= fail_page_alloc
.attr
.dentries
.dir
;
1280 fail_page_alloc
.ignore_gfp_wait_file
=
1281 debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
1282 &fail_page_alloc
.ignore_gfp_wait
);
1284 fail_page_alloc
.ignore_gfp_highmem_file
=
1285 debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
1286 &fail_page_alloc
.ignore_gfp_highmem
);
1287 fail_page_alloc
.min_order_file
=
1288 debugfs_create_u32("min-order", mode
, dir
,
1289 &fail_page_alloc
.min_order
);
1291 if (!fail_page_alloc
.ignore_gfp_wait_file
||
1292 !fail_page_alloc
.ignore_gfp_highmem_file
||
1293 !fail_page_alloc
.min_order_file
) {
1295 debugfs_remove(fail_page_alloc
.ignore_gfp_wait_file
);
1296 debugfs_remove(fail_page_alloc
.ignore_gfp_highmem_file
);
1297 debugfs_remove(fail_page_alloc
.min_order_file
);
1298 cleanup_fault_attr_dentries(&fail_page_alloc
.attr
);
1304 late_initcall(fail_page_alloc_debugfs
);
1306 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1308 #else /* CONFIG_FAIL_PAGE_ALLOC */
1310 static inline int should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1315 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1318 * Return 1 if free pages are above 'mark'. This takes into account the order
1319 * of the allocation.
1321 int zone_watermark_ok(struct zone
*z
, int order
, unsigned long mark
,
1322 int classzone_idx
, int alloc_flags
)
1324 /* free_pages my go negative - that's OK */
1326 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
) - (1 << order
) + 1;
1329 if (alloc_flags
& ALLOC_HIGH
)
1331 if (alloc_flags
& ALLOC_HARDER
)
1334 if (free_pages
<= min
+ z
->lowmem_reserve
[classzone_idx
])
1336 for (o
= 0; o
< order
; o
++) {
1337 /* At the next order, this order's pages become unavailable */
1338 free_pages
-= z
->free_area
[o
].nr_free
<< o
;
1340 /* Require fewer higher order pages to be free */
1343 if (free_pages
<= min
)
1351 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1352 * skip over zones that are not allowed by the cpuset, or that have
1353 * been recently (in last second) found to be nearly full. See further
1354 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1355 * that have to skip over a lot of full or unallowed zones.
1357 * If the zonelist cache is present in the passed in zonelist, then
1358 * returns a pointer to the allowed node mask (either the current
1359 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1361 * If the zonelist cache is not available for this zonelist, does
1362 * nothing and returns NULL.
1364 * If the fullzones BITMAP in the zonelist cache is stale (more than
1365 * a second since last zap'd) then we zap it out (clear its bits.)
1367 * We hold off even calling zlc_setup, until after we've checked the
1368 * first zone in the zonelist, on the theory that most allocations will
1369 * be satisfied from that first zone, so best to examine that zone as
1370 * quickly as we can.
1372 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1374 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1375 nodemask_t
*allowednodes
; /* zonelist_cache approximation */
1377 zlc
= zonelist
->zlcache_ptr
;
1381 if (time_after(jiffies
, zlc
->last_full_zap
+ HZ
)) {
1382 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
1383 zlc
->last_full_zap
= jiffies
;
1386 allowednodes
= !in_interrupt() && (alloc_flags
& ALLOC_CPUSET
) ?
1387 &cpuset_current_mems_allowed
:
1388 &node_states
[N_HIGH_MEMORY
];
1389 return allowednodes
;
1393 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1394 * if it is worth looking at further for free memory:
1395 * 1) Check that the zone isn't thought to be full (doesn't have its
1396 * bit set in the zonelist_cache fullzones BITMAP).
1397 * 2) Check that the zones node (obtained from the zonelist_cache
1398 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1399 * Return true (non-zero) if zone is worth looking at further, or
1400 * else return false (zero) if it is not.
1402 * This check -ignores- the distinction between various watermarks,
1403 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1404 * found to be full for any variation of these watermarks, it will
1405 * be considered full for up to one second by all requests, unless
1406 * we are so low on memory on all allowed nodes that we are forced
1407 * into the second scan of the zonelist.
1409 * In the second scan we ignore this zonelist cache and exactly
1410 * apply the watermarks to all zones, even it is slower to do so.
1411 * We are low on memory in the second scan, and should leave no stone
1412 * unturned looking for a free page.
1414 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1415 nodemask_t
*allowednodes
)
1417 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1418 int i
; /* index of *z in zonelist zones */
1419 int n
; /* node that zone *z is on */
1421 zlc
= zonelist
->zlcache_ptr
;
1425 i
= z
- zonelist
->_zonerefs
;
1428 /* This zone is worth trying if it is allowed but not full */
1429 return node_isset(n
, *allowednodes
) && !test_bit(i
, zlc
->fullzones
);
1433 * Given 'z' scanning a zonelist, set the corresponding bit in
1434 * zlc->fullzones, so that subsequent attempts to allocate a page
1435 * from that zone don't waste time re-examining it.
1437 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1439 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1440 int i
; /* index of *z in zonelist zones */
1442 zlc
= zonelist
->zlcache_ptr
;
1446 i
= z
- zonelist
->_zonerefs
;
1448 set_bit(i
, zlc
->fullzones
);
1451 #else /* CONFIG_NUMA */
1453 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1458 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1459 nodemask_t
*allowednodes
)
1464 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1467 #endif /* CONFIG_NUMA */
1470 * get_page_from_freelist goes through the zonelist trying to allocate
1473 static struct page
*
1474 get_page_from_freelist(gfp_t gfp_mask
, nodemask_t
*nodemask
, unsigned int order
,
1475 struct zonelist
*zonelist
, int high_zoneidx
, int alloc_flags
,
1476 struct zone
*preferred_zone
, int migratetype
)
1479 struct page
*page
= NULL
;
1482 nodemask_t
*allowednodes
= NULL
;/* zonelist_cache approximation */
1483 int zlc_active
= 0; /* set if using zonelist_cache */
1484 int did_zlc_setup
= 0; /* just call zlc_setup() one time */
1486 classzone_idx
= zone_idx(preferred_zone
);
1489 * Scan zonelist, looking for a zone with enough free.
1490 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1492 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
1493 high_zoneidx
, nodemask
) {
1494 if (NUMA_BUILD
&& zlc_active
&&
1495 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
1497 if ((alloc_flags
& ALLOC_CPUSET
) &&
1498 !cpuset_zone_allowed_softwall(zone
, gfp_mask
))
1501 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
1502 if (!(alloc_flags
& ALLOC_NO_WATERMARKS
)) {
1506 mark
= zone
->watermark
[alloc_flags
& ALLOC_WMARK_MASK
];
1507 if (zone_watermark_ok(zone
, order
, mark
,
1508 classzone_idx
, alloc_flags
))
1511 if (zone_reclaim_mode
== 0)
1512 goto this_zone_full
;
1514 ret
= zone_reclaim(zone
, gfp_mask
, order
);
1516 case ZONE_RECLAIM_NOSCAN
:
1519 case ZONE_RECLAIM_FULL
:
1520 /* scanned but unreclaimable */
1521 goto this_zone_full
;
1523 /* did we reclaim enough */
1524 if (!zone_watermark_ok(zone
, order
, mark
,
1525 classzone_idx
, alloc_flags
))
1526 goto this_zone_full
;
1531 page
= buffered_rmqueue(preferred_zone
, zone
, order
,
1532 gfp_mask
, migratetype
);
1537 zlc_mark_zone_full(zonelist
, z
);
1539 if (NUMA_BUILD
&& !did_zlc_setup
&& nr_online_nodes
> 1) {
1541 * we do zlc_setup after the first zone is tried but only
1542 * if there are multiple nodes make it worthwhile
1544 allowednodes
= zlc_setup(zonelist
, alloc_flags
);
1550 if (unlikely(NUMA_BUILD
&& page
== NULL
&& zlc_active
)) {
1551 /* Disable zlc cache for second zonelist scan */
1559 should_alloc_retry(gfp_t gfp_mask
, unsigned int order
,
1560 unsigned long pages_reclaimed
)
1562 /* Do not loop if specifically requested */
1563 if (gfp_mask
& __GFP_NORETRY
)
1567 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
1568 * means __GFP_NOFAIL, but that may not be true in other
1571 if (order
<= PAGE_ALLOC_COSTLY_ORDER
)
1575 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
1576 * specified, then we retry until we no longer reclaim any pages
1577 * (above), or we've reclaimed an order of pages at least as
1578 * large as the allocation's order. In both cases, if the
1579 * allocation still fails, we stop retrying.
1581 if (gfp_mask
& __GFP_REPEAT
&& pages_reclaimed
< (1 << order
))
1585 * Don't let big-order allocations loop unless the caller
1586 * explicitly requests that.
1588 if (gfp_mask
& __GFP_NOFAIL
)
1594 static inline struct page
*
1595 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
1596 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1597 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
1602 /* Acquire the OOM killer lock for the zones in zonelist */
1603 if (!try_set_zone_oom(zonelist
, gfp_mask
)) {
1604 schedule_timeout_uninterruptible(1);
1609 * Go through the zonelist yet one more time, keep very high watermark
1610 * here, this is only to catch a parallel oom killing, we must fail if
1611 * we're still under heavy pressure.
1613 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
,
1614 order
, zonelist
, high_zoneidx
,
1615 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
,
1616 preferred_zone
, migratetype
);
1620 /* The OOM killer will not help higher order allocs */
1621 if (order
> PAGE_ALLOC_COSTLY_ORDER
&& !(gfp_mask
& __GFP_NOFAIL
))
1624 /* Exhausted what can be done so it's blamo time */
1625 out_of_memory(zonelist
, gfp_mask
, order
);
1628 clear_zonelist_oom(zonelist
, gfp_mask
);
1632 /* The really slow allocator path where we enter direct reclaim */
1633 static inline struct page
*
1634 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
1635 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1636 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
1637 int migratetype
, unsigned long *did_some_progress
)
1639 struct page
*page
= NULL
;
1640 struct reclaim_state reclaim_state
;
1641 struct task_struct
*p
= current
;
1645 /* We now go into synchronous reclaim */
1646 cpuset_memory_pressure_bump();
1647 p
->flags
|= PF_MEMALLOC
;
1648 lockdep_set_current_reclaim_state(gfp_mask
);
1649 reclaim_state
.reclaimed_slab
= 0;
1650 p
->reclaim_state
= &reclaim_state
;
1652 *did_some_progress
= try_to_free_pages(zonelist
, order
, gfp_mask
, nodemask
);
1654 p
->reclaim_state
= NULL
;
1655 lockdep_clear_current_reclaim_state();
1656 p
->flags
&= ~PF_MEMALLOC
;
1663 if (likely(*did_some_progress
))
1664 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
1665 zonelist
, high_zoneidx
,
1666 alloc_flags
, preferred_zone
,
1672 * This is called in the allocator slow-path if the allocation request is of
1673 * sufficient urgency to ignore watermarks and take other desperate measures
1675 static inline struct page
*
1676 __alloc_pages_high_priority(gfp_t gfp_mask
, unsigned int order
,
1677 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1678 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
1684 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
1685 zonelist
, high_zoneidx
, ALLOC_NO_WATERMARKS
,
1686 preferred_zone
, migratetype
);
1688 if (!page
&& gfp_mask
& __GFP_NOFAIL
)
1689 congestion_wait(BLK_RW_ASYNC
, HZ
/50);
1690 } while (!page
&& (gfp_mask
& __GFP_NOFAIL
));
1696 void wake_all_kswapd(unsigned int order
, struct zonelist
*zonelist
,
1697 enum zone_type high_zoneidx
)
1702 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
)
1703 wakeup_kswapd(zone
, order
);
1707 gfp_to_alloc_flags(gfp_t gfp_mask
)
1709 struct task_struct
*p
= current
;
1710 int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
1711 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
1713 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
1714 BUILD_BUG_ON(__GFP_HIGH
!= ALLOC_HIGH
);
1717 * The caller may dip into page reserves a bit more if the caller
1718 * cannot run direct reclaim, or if the caller has realtime scheduling
1719 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
1720 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
1722 alloc_flags
|= (gfp_mask
& __GFP_HIGH
);
1725 alloc_flags
|= ALLOC_HARDER
;
1727 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
1728 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1730 alloc_flags
&= ~ALLOC_CPUSET
;
1731 } else if (unlikely(rt_task(p
)))
1732 alloc_flags
|= ALLOC_HARDER
;
1734 if (likely(!(gfp_mask
& __GFP_NOMEMALLOC
))) {
1735 if (!in_interrupt() &&
1736 ((p
->flags
& PF_MEMALLOC
) ||
1737 unlikely(test_thread_flag(TIF_MEMDIE
))))
1738 alloc_flags
|= ALLOC_NO_WATERMARKS
;
1744 static inline struct page
*
1745 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
1746 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1747 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
1750 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
1751 struct page
*page
= NULL
;
1753 unsigned long pages_reclaimed
= 0;
1754 unsigned long did_some_progress
;
1755 struct task_struct
*p
= current
;
1758 * In the slowpath, we sanity check order to avoid ever trying to
1759 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
1760 * be using allocators in order of preference for an area that is
1763 if (order
>= MAX_ORDER
) {
1764 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
1769 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
1770 * __GFP_NOWARN set) should not cause reclaim since the subsystem
1771 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
1772 * using a larger set of nodes after it has established that the
1773 * allowed per node queues are empty and that nodes are
1776 if (NUMA_BUILD
&& (gfp_mask
& GFP_THISNODE
) == GFP_THISNODE
)
1779 wake_all_kswapd(order
, zonelist
, high_zoneidx
);
1783 * OK, we're below the kswapd watermark and have kicked background
1784 * reclaim. Now things get more complex, so set up alloc_flags according
1785 * to how we want to proceed.
1787 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
1789 /* This is the last chance, in general, before the goto nopage. */
1790 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
, zonelist
,
1791 high_zoneidx
, alloc_flags
& ~ALLOC_NO_WATERMARKS
,
1792 preferred_zone
, migratetype
);
1797 /* Allocate without watermarks if the context allows */
1798 if (alloc_flags
& ALLOC_NO_WATERMARKS
) {
1799 page
= __alloc_pages_high_priority(gfp_mask
, order
,
1800 zonelist
, high_zoneidx
, nodemask
,
1801 preferred_zone
, migratetype
);
1806 /* Atomic allocations - we can't balance anything */
1810 /* Avoid recursion of direct reclaim */
1811 if (p
->flags
& PF_MEMALLOC
)
1814 /* Avoid allocations with no watermarks from looping endlessly */
1815 if (test_thread_flag(TIF_MEMDIE
) && !(gfp_mask
& __GFP_NOFAIL
))
1818 /* Try direct reclaim and then allocating */
1819 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
,
1820 zonelist
, high_zoneidx
,
1822 alloc_flags
, preferred_zone
,
1823 migratetype
, &did_some_progress
);
1828 * If we failed to make any progress reclaiming, then we are
1829 * running out of options and have to consider going OOM
1831 if (!did_some_progress
) {
1832 if ((gfp_mask
& __GFP_FS
) && !(gfp_mask
& __GFP_NORETRY
)) {
1833 if (oom_killer_disabled
)
1835 page
= __alloc_pages_may_oom(gfp_mask
, order
,
1836 zonelist
, high_zoneidx
,
1837 nodemask
, preferred_zone
,
1843 * The OOM killer does not trigger for high-order
1844 * ~__GFP_NOFAIL allocations so if no progress is being
1845 * made, there are no other options and retrying is
1848 if (order
> PAGE_ALLOC_COSTLY_ORDER
&&
1849 !(gfp_mask
& __GFP_NOFAIL
))
1856 /* Check if we should retry the allocation */
1857 pages_reclaimed
+= did_some_progress
;
1858 if (should_alloc_retry(gfp_mask
, order
, pages_reclaimed
)) {
1859 /* Wait for some write requests to complete then retry */
1860 congestion_wait(BLK_RW_ASYNC
, HZ
/50);
1865 if (!(gfp_mask
& __GFP_NOWARN
) && printk_ratelimit()) {
1866 printk(KERN_WARNING
"%s: page allocation failure."
1867 " order:%d, mode:0x%x\n",
1868 p
->comm
, order
, gfp_mask
);
1874 if (kmemcheck_enabled
)
1875 kmemcheck_pagealloc_alloc(page
, order
, gfp_mask
);
1881 * This is the 'heart' of the zoned buddy allocator.
1884 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
,
1885 struct zonelist
*zonelist
, nodemask_t
*nodemask
)
1887 enum zone_type high_zoneidx
= gfp_zone(gfp_mask
);
1888 struct zone
*preferred_zone
;
1890 int migratetype
= allocflags_to_migratetype(gfp_mask
);
1892 gfp_mask
&= gfp_allowed_mask
;
1894 lockdep_trace_alloc(gfp_mask
);
1896 might_sleep_if(gfp_mask
& __GFP_WAIT
);
1898 if (should_fail_alloc_page(gfp_mask
, order
))
1902 * Check the zones suitable for the gfp_mask contain at least one
1903 * valid zone. It's possible to have an empty zonelist as a result
1904 * of GFP_THISNODE and a memoryless node
1906 if (unlikely(!zonelist
->_zonerefs
->zone
))
1909 /* The preferred zone is used for statistics later */
1910 first_zones_zonelist(zonelist
, high_zoneidx
, nodemask
, &preferred_zone
);
1911 if (!preferred_zone
)
1914 /* First allocation attempt */
1915 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
, order
,
1916 zonelist
, high_zoneidx
, ALLOC_WMARK_LOW
|ALLOC_CPUSET
,
1917 preferred_zone
, migratetype
);
1918 if (unlikely(!page
))
1919 page
= __alloc_pages_slowpath(gfp_mask
, order
,
1920 zonelist
, high_zoneidx
, nodemask
,
1921 preferred_zone
, migratetype
);
1925 EXPORT_SYMBOL(__alloc_pages_nodemask
);
1928 * Common helper functions.
1930 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
1935 * __get_free_pages() returns a 32-bit address, which cannot represent
1938 VM_BUG_ON((gfp_mask
& __GFP_HIGHMEM
) != 0);
1940 page
= alloc_pages(gfp_mask
, order
);
1943 return (unsigned long) page_address(page
);
1945 EXPORT_SYMBOL(__get_free_pages
);
1947 unsigned long get_zeroed_page(gfp_t gfp_mask
)
1949 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
1951 EXPORT_SYMBOL(get_zeroed_page
);
1953 void __pagevec_free(struct pagevec
*pvec
)
1955 int i
= pagevec_count(pvec
);
1958 free_hot_cold_page(pvec
->pages
[i
], pvec
->cold
);
1961 void __free_pages(struct page
*page
, unsigned int order
)
1963 if (put_page_testzero(page
)) {
1965 free_hot_page(page
);
1967 __free_pages_ok(page
, order
);
1971 EXPORT_SYMBOL(__free_pages
);
1973 void free_pages(unsigned long addr
, unsigned int order
)
1976 VM_BUG_ON(!virt_addr_valid((void *)addr
));
1977 __free_pages(virt_to_page((void *)addr
), order
);
1981 EXPORT_SYMBOL(free_pages
);
1984 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
1985 * @size: the number of bytes to allocate
1986 * @gfp_mask: GFP flags for the allocation
1988 * This function is similar to alloc_pages(), except that it allocates the
1989 * minimum number of pages to satisfy the request. alloc_pages() can only
1990 * allocate memory in power-of-two pages.
1992 * This function is also limited by MAX_ORDER.
1994 * Memory allocated by this function must be released by free_pages_exact().
1996 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
1998 unsigned int order
= get_order(size
);
2001 addr
= __get_free_pages(gfp_mask
, order
);
2003 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
2004 unsigned long used
= addr
+ PAGE_ALIGN(size
);
2006 split_page(virt_to_page((void *)addr
), order
);
2007 while (used
< alloc_end
) {
2013 return (void *)addr
;
2015 EXPORT_SYMBOL(alloc_pages_exact
);
2018 * free_pages_exact - release memory allocated via alloc_pages_exact()
2019 * @virt: the value returned by alloc_pages_exact.
2020 * @size: size of allocation, same value as passed to alloc_pages_exact().
2022 * Release the memory allocated by a previous call to alloc_pages_exact.
2024 void free_pages_exact(void *virt
, size_t size
)
2026 unsigned long addr
= (unsigned long)virt
;
2027 unsigned long end
= addr
+ PAGE_ALIGN(size
);
2029 while (addr
< end
) {
2034 EXPORT_SYMBOL(free_pages_exact
);
2036 static unsigned int nr_free_zone_pages(int offset
)
2041 /* Just pick one node, since fallback list is circular */
2042 unsigned int sum
= 0;
2044 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
2046 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
2047 unsigned long size
= zone
->present_pages
;
2048 unsigned long high
= high_wmark_pages(zone
);
2057 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2059 unsigned int nr_free_buffer_pages(void)
2061 return nr_free_zone_pages(gfp_zone(GFP_USER
));
2063 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
2066 * Amount of free RAM allocatable within all zones
2068 unsigned int nr_free_pagecache_pages(void)
2070 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
2073 static inline void show_node(struct zone
*zone
)
2076 printk("Node %d ", zone_to_nid(zone
));
2079 void si_meminfo(struct sysinfo
*val
)
2081 val
->totalram
= totalram_pages
;
2083 val
->freeram
= global_page_state(NR_FREE_PAGES
);
2084 val
->bufferram
= nr_blockdev_pages();
2085 val
->totalhigh
= totalhigh_pages
;
2086 val
->freehigh
= nr_free_highpages();
2087 val
->mem_unit
= PAGE_SIZE
;
2090 EXPORT_SYMBOL(si_meminfo
);
2093 void si_meminfo_node(struct sysinfo
*val
, int nid
)
2095 pg_data_t
*pgdat
= NODE_DATA(nid
);
2097 val
->totalram
= pgdat
->node_present_pages
;
2098 val
->freeram
= node_page_state(nid
, NR_FREE_PAGES
);
2099 #ifdef CONFIG_HIGHMEM
2100 val
->totalhigh
= pgdat
->node_zones
[ZONE_HIGHMEM
].present_pages
;
2101 val
->freehigh
= zone_page_state(&pgdat
->node_zones
[ZONE_HIGHMEM
],
2107 val
->mem_unit
= PAGE_SIZE
;
2111 #define K(x) ((x) << (PAGE_SHIFT-10))
2114 * Show free area list (used inside shift_scroll-lock stuff)
2115 * We also calculate the percentage fragmentation. We do this by counting the
2116 * memory on each free list with the exception of the first item on the list.
2118 void show_free_areas(void)
2123 for_each_populated_zone(zone
) {
2125 printk("%s per-cpu:\n", zone
->name
);
2127 for_each_online_cpu(cpu
) {
2128 struct per_cpu_pageset
*pageset
;
2130 pageset
= zone_pcp(zone
, cpu
);
2132 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2133 cpu
, pageset
->pcp
.high
,
2134 pageset
->pcp
.batch
, pageset
->pcp
.count
);
2138 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2139 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2141 " dirty:%lu writeback:%lu unstable:%lu buffer:%lu\n"
2142 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2143 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n",
2144 global_page_state(NR_ACTIVE_ANON
),
2145 global_page_state(NR_INACTIVE_ANON
),
2146 global_page_state(NR_ISOLATED_ANON
),
2147 global_page_state(NR_ACTIVE_FILE
),
2148 global_page_state(NR_INACTIVE_FILE
),
2149 global_page_state(NR_ISOLATED_FILE
),
2150 global_page_state(NR_UNEVICTABLE
),
2151 global_page_state(NR_FILE_DIRTY
),
2152 global_page_state(NR_WRITEBACK
),
2153 global_page_state(NR_UNSTABLE_NFS
),
2154 nr_blockdev_pages(),
2155 global_page_state(NR_FREE_PAGES
),
2156 global_page_state(NR_SLAB_RECLAIMABLE
),
2157 global_page_state(NR_SLAB_UNRECLAIMABLE
),
2158 global_page_state(NR_FILE_MAPPED
),
2159 global_page_state(NR_SHMEM
),
2160 global_page_state(NR_PAGETABLE
),
2161 global_page_state(NR_BOUNCE
));
2163 for_each_populated_zone(zone
) {
2172 " active_anon:%lukB"
2173 " inactive_anon:%lukB"
2174 " active_file:%lukB"
2175 " inactive_file:%lukB"
2176 " unevictable:%lukB"
2177 " isolated(anon):%lukB"
2178 " isolated(file):%lukB"
2185 " slab_reclaimable:%lukB"
2186 " slab_unreclaimable:%lukB"
2187 " kernel_stack:%lukB"
2191 " writeback_tmp:%lukB"
2192 " pages_scanned:%lu"
2193 " all_unreclaimable? %s"
2196 K(zone_page_state(zone
, NR_FREE_PAGES
)),
2197 K(min_wmark_pages(zone
)),
2198 K(low_wmark_pages(zone
)),
2199 K(high_wmark_pages(zone
)),
2200 K(zone_page_state(zone
, NR_ACTIVE_ANON
)),
2201 K(zone_page_state(zone
, NR_INACTIVE_ANON
)),
2202 K(zone_page_state(zone
, NR_ACTIVE_FILE
)),
2203 K(zone_page_state(zone
, NR_INACTIVE_FILE
)),
2204 K(zone_page_state(zone
, NR_UNEVICTABLE
)),
2205 K(zone_page_state(zone
, NR_ISOLATED_ANON
)),
2206 K(zone_page_state(zone
, NR_ISOLATED_FILE
)),
2207 K(zone
->present_pages
),
2208 K(zone_page_state(zone
, NR_MLOCK
)),
2209 K(zone_page_state(zone
, NR_FILE_DIRTY
)),
2210 K(zone_page_state(zone
, NR_WRITEBACK
)),
2211 K(zone_page_state(zone
, NR_FILE_MAPPED
)),
2212 K(zone_page_state(zone
, NR_SHMEM
)),
2213 K(zone_page_state(zone
, NR_SLAB_RECLAIMABLE
)),
2214 K(zone_page_state(zone
, NR_SLAB_UNRECLAIMABLE
)),
2215 zone_page_state(zone
, NR_KERNEL_STACK
) *
2217 K(zone_page_state(zone
, NR_PAGETABLE
)),
2218 K(zone_page_state(zone
, NR_UNSTABLE_NFS
)),
2219 K(zone_page_state(zone
, NR_BOUNCE
)),
2220 K(zone_page_state(zone
, NR_WRITEBACK_TEMP
)),
2221 zone
->pages_scanned
,
2222 (zone_is_all_unreclaimable(zone
) ? "yes" : "no")
2224 printk("lowmem_reserve[]:");
2225 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
2226 printk(" %lu", zone
->lowmem_reserve
[i
]);
2230 for_each_populated_zone(zone
) {
2231 unsigned long nr
[MAX_ORDER
], flags
, order
, total
= 0;
2234 printk("%s: ", zone
->name
);
2236 spin_lock_irqsave(&zone
->lock
, flags
);
2237 for (order
= 0; order
< MAX_ORDER
; order
++) {
2238 nr
[order
] = zone
->free_area
[order
].nr_free
;
2239 total
+= nr
[order
] << order
;
2241 spin_unlock_irqrestore(&zone
->lock
, flags
);
2242 for (order
= 0; order
< MAX_ORDER
; order
++)
2243 printk("%lu*%lukB ", nr
[order
], K(1UL) << order
);
2244 printk("= %lukB\n", K(total
));
2247 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES
));
2249 show_swap_cache_info();
2252 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
2254 zoneref
->zone
= zone
;
2255 zoneref
->zone_idx
= zone_idx(zone
);
2259 * Builds allocation fallback zone lists.
2261 * Add all populated zones of a node to the zonelist.
2263 static int build_zonelists_node(pg_data_t
*pgdat
, struct zonelist
*zonelist
,
2264 int nr_zones
, enum zone_type zone_type
)
2268 BUG_ON(zone_type
>= MAX_NR_ZONES
);
2273 zone
= pgdat
->node_zones
+ zone_type
;
2274 if (populated_zone(zone
)) {
2275 zoneref_set_zone(zone
,
2276 &zonelist
->_zonerefs
[nr_zones
++]);
2277 check_highest_zone(zone_type
);
2280 } while (zone_type
);
2287 * 0 = automatic detection of better ordering.
2288 * 1 = order by ([node] distance, -zonetype)
2289 * 2 = order by (-zonetype, [node] distance)
2291 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
2292 * the same zonelist. So only NUMA can configure this param.
2294 #define ZONELIST_ORDER_DEFAULT 0
2295 #define ZONELIST_ORDER_NODE 1
2296 #define ZONELIST_ORDER_ZONE 2
2298 /* zonelist order in the kernel.
2299 * set_zonelist_order() will set this to NODE or ZONE.
2301 static int current_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
2302 static char zonelist_order_name
[3][8] = {"Default", "Node", "Zone"};
2306 /* The value user specified ....changed by config */
2307 static int user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
2308 /* string for sysctl */
2309 #define NUMA_ZONELIST_ORDER_LEN 16
2310 char numa_zonelist_order
[16] = "default";
2313 * interface for configure zonelist ordering.
2314 * command line option "numa_zonelist_order"
2315 * = "[dD]efault - default, automatic configuration.
2316 * = "[nN]ode - order by node locality, then by zone within node
2317 * = "[zZ]one - order by zone, then by locality within zone
2320 static int __parse_numa_zonelist_order(char *s
)
2322 if (*s
== 'd' || *s
== 'D') {
2323 user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
2324 } else if (*s
== 'n' || *s
== 'N') {
2325 user_zonelist_order
= ZONELIST_ORDER_NODE
;
2326 } else if (*s
== 'z' || *s
== 'Z') {
2327 user_zonelist_order
= ZONELIST_ORDER_ZONE
;
2330 "Ignoring invalid numa_zonelist_order value: "
2337 static __init
int setup_numa_zonelist_order(char *s
)
2340 return __parse_numa_zonelist_order(s
);
2343 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
2346 * sysctl handler for numa_zonelist_order
2348 int numa_zonelist_order_handler(ctl_table
*table
, int write
,
2349 struct file
*file
, void __user
*buffer
, size_t *length
,
2352 char saved_string
[NUMA_ZONELIST_ORDER_LEN
];
2356 strncpy(saved_string
, (char*)table
->data
,
2357 NUMA_ZONELIST_ORDER_LEN
);
2358 ret
= proc_dostring(table
, write
, file
, buffer
, length
, ppos
);
2362 int oldval
= user_zonelist_order
;
2363 if (__parse_numa_zonelist_order((char*)table
->data
)) {
2365 * bogus value. restore saved string
2367 strncpy((char*)table
->data
, saved_string
,
2368 NUMA_ZONELIST_ORDER_LEN
);
2369 user_zonelist_order
= oldval
;
2370 } else if (oldval
!= user_zonelist_order
)
2371 build_all_zonelists();
2377 #define MAX_NODE_LOAD (nr_online_nodes)
2378 static int node_load
[MAX_NUMNODES
];
2381 * find_next_best_node - find the next node that should appear in a given node's fallback list
2382 * @node: node whose fallback list we're appending
2383 * @used_node_mask: nodemask_t of already used nodes
2385 * We use a number of factors to determine which is the next node that should
2386 * appear on a given node's fallback list. The node should not have appeared
2387 * already in @node's fallback list, and it should be the next closest node
2388 * according to the distance array (which contains arbitrary distance values
2389 * from each node to each node in the system), and should also prefer nodes
2390 * with no CPUs, since presumably they'll have very little allocation pressure
2391 * on them otherwise.
2392 * It returns -1 if no node is found.
2394 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
2397 int min_val
= INT_MAX
;
2399 const struct cpumask
*tmp
= cpumask_of_node(0);
2401 /* Use the local node if we haven't already */
2402 if (!node_isset(node
, *used_node_mask
)) {
2403 node_set(node
, *used_node_mask
);
2407 for_each_node_state(n
, N_HIGH_MEMORY
) {
2409 /* Don't want a node to appear more than once */
2410 if (node_isset(n
, *used_node_mask
))
2413 /* Use the distance array to find the distance */
2414 val
= node_distance(node
, n
);
2416 /* Penalize nodes under us ("prefer the next node") */
2419 /* Give preference to headless and unused nodes */
2420 tmp
= cpumask_of_node(n
);
2421 if (!cpumask_empty(tmp
))
2422 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
2424 /* Slight preference for less loaded node */
2425 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
2426 val
+= node_load
[n
];
2428 if (val
< min_val
) {
2435 node_set(best_node
, *used_node_mask
);
2442 * Build zonelists ordered by node and zones within node.
2443 * This results in maximum locality--normal zone overflows into local
2444 * DMA zone, if any--but risks exhausting DMA zone.
2446 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int node
)
2449 struct zonelist
*zonelist
;
2451 zonelist
= &pgdat
->node_zonelists
[0];
2452 for (j
= 0; zonelist
->_zonerefs
[j
].zone
!= NULL
; j
++)
2454 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
2456 zonelist
->_zonerefs
[j
].zone
= NULL
;
2457 zonelist
->_zonerefs
[j
].zone_idx
= 0;
2461 * Build gfp_thisnode zonelists
2463 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
2466 struct zonelist
*zonelist
;
2468 zonelist
= &pgdat
->node_zonelists
[1];
2469 j
= build_zonelists_node(pgdat
, zonelist
, 0, MAX_NR_ZONES
- 1);
2470 zonelist
->_zonerefs
[j
].zone
= NULL
;
2471 zonelist
->_zonerefs
[j
].zone_idx
= 0;
2475 * Build zonelists ordered by zone and nodes within zones.
2476 * This results in conserving DMA zone[s] until all Normal memory is
2477 * exhausted, but results in overflowing to remote node while memory
2478 * may still exist in local DMA zone.
2480 static int node_order
[MAX_NUMNODES
];
2482 static void build_zonelists_in_zone_order(pg_data_t
*pgdat
, int nr_nodes
)
2485 int zone_type
; /* needs to be signed */
2487 struct zonelist
*zonelist
;
2489 zonelist
= &pgdat
->node_zonelists
[0];
2491 for (zone_type
= MAX_NR_ZONES
- 1; zone_type
>= 0; zone_type
--) {
2492 for (j
= 0; j
< nr_nodes
; j
++) {
2493 node
= node_order
[j
];
2494 z
= &NODE_DATA(node
)->node_zones
[zone_type
];
2495 if (populated_zone(z
)) {
2497 &zonelist
->_zonerefs
[pos
++]);
2498 check_highest_zone(zone_type
);
2502 zonelist
->_zonerefs
[pos
].zone
= NULL
;
2503 zonelist
->_zonerefs
[pos
].zone_idx
= 0;
2506 static int default_zonelist_order(void)
2509 unsigned long low_kmem_size
,total_size
;
2513 * ZONE_DMA and ZONE_DMA32 can be very small area in the sytem.
2514 * If they are really small and used heavily, the system can fall
2515 * into OOM very easily.
2516 * This function detect ZONE_DMA/DMA32 size and confgigures zone order.
2518 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
2521 for_each_online_node(nid
) {
2522 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
2523 z
= &NODE_DATA(nid
)->node_zones
[zone_type
];
2524 if (populated_zone(z
)) {
2525 if (zone_type
< ZONE_NORMAL
)
2526 low_kmem_size
+= z
->present_pages
;
2527 total_size
+= z
->present_pages
;
2531 if (!low_kmem_size
|| /* there are no DMA area. */
2532 low_kmem_size
> total_size
/2) /* DMA/DMA32 is big. */
2533 return ZONELIST_ORDER_NODE
;
2535 * look into each node's config.
2536 * If there is a node whose DMA/DMA32 memory is very big area on
2537 * local memory, NODE_ORDER may be suitable.
2539 average_size
= total_size
/
2540 (nodes_weight(node_states
[N_HIGH_MEMORY
]) + 1);
2541 for_each_online_node(nid
) {
2544 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
2545 z
= &NODE_DATA(nid
)->node_zones
[zone_type
];
2546 if (populated_zone(z
)) {
2547 if (zone_type
< ZONE_NORMAL
)
2548 low_kmem_size
+= z
->present_pages
;
2549 total_size
+= z
->present_pages
;
2552 if (low_kmem_size
&&
2553 total_size
> average_size
&& /* ignore small node */
2554 low_kmem_size
> total_size
* 70/100)
2555 return ZONELIST_ORDER_NODE
;
2557 return ZONELIST_ORDER_ZONE
;
2560 static void set_zonelist_order(void)
2562 if (user_zonelist_order
== ZONELIST_ORDER_DEFAULT
)
2563 current_zonelist_order
= default_zonelist_order();
2565 current_zonelist_order
= user_zonelist_order
;
2568 static void build_zonelists(pg_data_t
*pgdat
)
2572 nodemask_t used_mask
;
2573 int local_node
, prev_node
;
2574 struct zonelist
*zonelist
;
2575 int order
= current_zonelist_order
;
2577 /* initialize zonelists */
2578 for (i
= 0; i
< MAX_ZONELISTS
; i
++) {
2579 zonelist
= pgdat
->node_zonelists
+ i
;
2580 zonelist
->_zonerefs
[0].zone
= NULL
;
2581 zonelist
->_zonerefs
[0].zone_idx
= 0;
2584 /* NUMA-aware ordering of nodes */
2585 local_node
= pgdat
->node_id
;
2586 load
= nr_online_nodes
;
2587 prev_node
= local_node
;
2588 nodes_clear(used_mask
);
2590 memset(node_order
, 0, sizeof(node_order
));
2593 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
2594 int distance
= node_distance(local_node
, node
);
2597 * If another node is sufficiently far away then it is better
2598 * to reclaim pages in a zone before going off node.
2600 if (distance
> RECLAIM_DISTANCE
)
2601 zone_reclaim_mode
= 1;
2604 * We don't want to pressure a particular node.
2605 * So adding penalty to the first node in same
2606 * distance group to make it round-robin.
2608 if (distance
!= node_distance(local_node
, prev_node
))
2609 node_load
[node
] = load
;
2613 if (order
== ZONELIST_ORDER_NODE
)
2614 build_zonelists_in_node_order(pgdat
, node
);
2616 node_order
[j
++] = node
; /* remember order */
2619 if (order
== ZONELIST_ORDER_ZONE
) {
2620 /* calculate node order -- i.e., DMA last! */
2621 build_zonelists_in_zone_order(pgdat
, j
);
2624 build_thisnode_zonelists(pgdat
);
2627 /* Construct the zonelist performance cache - see further mmzone.h */
2628 static void build_zonelist_cache(pg_data_t
*pgdat
)
2630 struct zonelist
*zonelist
;
2631 struct zonelist_cache
*zlc
;
2634 zonelist
= &pgdat
->node_zonelists
[0];
2635 zonelist
->zlcache_ptr
= zlc
= &zonelist
->zlcache
;
2636 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
2637 for (z
= zonelist
->_zonerefs
; z
->zone
; z
++)
2638 zlc
->z_to_n
[z
- zonelist
->_zonerefs
] = zonelist_node_idx(z
);
2642 #else /* CONFIG_NUMA */
2644 static void set_zonelist_order(void)
2646 current_zonelist_order
= ZONELIST_ORDER_ZONE
;
2649 static void build_zonelists(pg_data_t
*pgdat
)
2651 int node
, local_node
;
2653 struct zonelist
*zonelist
;
2655 local_node
= pgdat
->node_id
;
2657 zonelist
= &pgdat
->node_zonelists
[0];
2658 j
= build_zonelists_node(pgdat
, zonelist
, 0, MAX_NR_ZONES
- 1);
2661 * Now we build the zonelist so that it contains the zones
2662 * of all the other nodes.
2663 * We don't want to pressure a particular node, so when
2664 * building the zones for node N, we make sure that the
2665 * zones coming right after the local ones are those from
2666 * node N+1 (modulo N)
2668 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
2669 if (!node_online(node
))
2671 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
2674 for (node
= 0; node
< local_node
; node
++) {
2675 if (!node_online(node
))
2677 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
2681 zonelist
->_zonerefs
[j
].zone
= NULL
;
2682 zonelist
->_zonerefs
[j
].zone_idx
= 0;
2685 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
2686 static void build_zonelist_cache(pg_data_t
*pgdat
)
2688 pgdat
->node_zonelists
[0].zlcache_ptr
= NULL
;
2691 #endif /* CONFIG_NUMA */
2693 /* return values int ....just for stop_machine() */
2694 static int __build_all_zonelists(void *dummy
)
2699 memset(node_load
, 0, sizeof(node_load
));
2701 for_each_online_node(nid
) {
2702 pg_data_t
*pgdat
= NODE_DATA(nid
);
2704 build_zonelists(pgdat
);
2705 build_zonelist_cache(pgdat
);
2710 void build_all_zonelists(void)
2712 set_zonelist_order();
2714 if (system_state
== SYSTEM_BOOTING
) {
2715 __build_all_zonelists(NULL
);
2716 mminit_verify_zonelist();
2717 cpuset_init_current_mems_allowed();
2719 /* we have to stop all cpus to guarantee there is no user
2721 stop_machine(__build_all_zonelists
, NULL
, NULL
);
2722 /* cpuset refresh routine should be here */
2724 vm_total_pages
= nr_free_pagecache_pages();
2726 * Disable grouping by mobility if the number of pages in the
2727 * system is too low to allow the mechanism to work. It would be
2728 * more accurate, but expensive to check per-zone. This check is
2729 * made on memory-hotadd so a system can start with mobility
2730 * disabled and enable it later
2732 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
2733 page_group_by_mobility_disabled
= 1;
2735 page_group_by_mobility_disabled
= 0;
2737 printk("Built %i zonelists in %s order, mobility grouping %s. "
2738 "Total pages: %ld\n",
2740 zonelist_order_name
[current_zonelist_order
],
2741 page_group_by_mobility_disabled
? "off" : "on",
2744 printk("Policy zone: %s\n", zone_names
[policy_zone
]);
2749 * Helper functions to size the waitqueue hash table.
2750 * Essentially these want to choose hash table sizes sufficiently
2751 * large so that collisions trying to wait on pages are rare.
2752 * But in fact, the number of active page waitqueues on typical
2753 * systems is ridiculously low, less than 200. So this is even
2754 * conservative, even though it seems large.
2756 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
2757 * waitqueues, i.e. the size of the waitq table given the number of pages.
2759 #define PAGES_PER_WAITQUEUE 256
2761 #ifndef CONFIG_MEMORY_HOTPLUG
2762 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
2764 unsigned long size
= 1;
2766 pages
/= PAGES_PER_WAITQUEUE
;
2768 while (size
< pages
)
2772 * Once we have dozens or even hundreds of threads sleeping
2773 * on IO we've got bigger problems than wait queue collision.
2774 * Limit the size of the wait table to a reasonable size.
2776 size
= min(size
, 4096UL);
2778 return max(size
, 4UL);
2782 * A zone's size might be changed by hot-add, so it is not possible to determine
2783 * a suitable size for its wait_table. So we use the maximum size now.
2785 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
2787 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
2788 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
2789 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
2791 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
2792 * or more by the traditional way. (See above). It equals:
2794 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
2795 * ia64(16K page size) : = ( 8G + 4M)byte.
2796 * powerpc (64K page size) : = (32G +16M)byte.
2798 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
2805 * This is an integer logarithm so that shifts can be used later
2806 * to extract the more random high bits from the multiplicative
2807 * hash function before the remainder is taken.
2809 static inline unsigned long wait_table_bits(unsigned long size
)
2814 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
2817 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
2818 * of blocks reserved is based on min_wmark_pages(zone). The memory within
2819 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
2820 * higher will lead to a bigger reserve which will get freed as contiguous
2821 * blocks as reclaim kicks in
2823 static void setup_zone_migrate_reserve(struct zone
*zone
)
2825 unsigned long start_pfn
, pfn
, end_pfn
;
2827 unsigned long reserve
, block_migratetype
;
2829 /* Get the start pfn, end pfn and the number of blocks to reserve */
2830 start_pfn
= zone
->zone_start_pfn
;
2831 end_pfn
= start_pfn
+ zone
->spanned_pages
;
2832 reserve
= roundup(min_wmark_pages(zone
), pageblock_nr_pages
) >>
2835 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= pageblock_nr_pages
) {
2836 if (!pfn_valid(pfn
))
2838 page
= pfn_to_page(pfn
);
2840 /* Watch out for overlapping nodes */
2841 if (page_to_nid(page
) != zone_to_nid(zone
))
2844 /* Blocks with reserved pages will never free, skip them. */
2845 if (PageReserved(page
))
2848 block_migratetype
= get_pageblock_migratetype(page
);
2850 /* If this block is reserved, account for it */
2851 if (reserve
> 0 && block_migratetype
== MIGRATE_RESERVE
) {
2856 /* Suitable for reserving if this block is movable */
2857 if (reserve
> 0 && block_migratetype
== MIGRATE_MOVABLE
) {
2858 set_pageblock_migratetype(page
, MIGRATE_RESERVE
);
2859 move_freepages_block(zone
, page
, MIGRATE_RESERVE
);
2865 * If the reserve is met and this is a previous reserved block,
2868 if (block_migratetype
== MIGRATE_RESERVE
) {
2869 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
2870 move_freepages_block(zone
, page
, MIGRATE_MOVABLE
);
2876 * Initially all pages are reserved - free ones are freed
2877 * up by free_all_bootmem() once the early boot process is
2878 * done. Non-atomic initialization, single-pass.
2880 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
2881 unsigned long start_pfn
, enum memmap_context context
)
2884 unsigned long end_pfn
= start_pfn
+ size
;
2888 if (highest_memmap_pfn
< end_pfn
- 1)
2889 highest_memmap_pfn
= end_pfn
- 1;
2891 z
= &NODE_DATA(nid
)->node_zones
[zone
];
2892 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
2894 * There can be holes in boot-time mem_map[]s
2895 * handed to this function. They do not
2896 * exist on hotplugged memory.
2898 if (context
== MEMMAP_EARLY
) {
2899 if (!early_pfn_valid(pfn
))
2901 if (!early_pfn_in_nid(pfn
, nid
))
2904 page
= pfn_to_page(pfn
);
2905 set_page_links(page
, zone
, nid
, pfn
);
2906 mminit_verify_page_links(page
, zone
, nid
, pfn
);
2907 init_page_count(page
);
2908 reset_page_mapcount(page
);
2909 SetPageReserved(page
);
2911 * Mark the block movable so that blocks are reserved for
2912 * movable at startup. This will force kernel allocations
2913 * to reserve their blocks rather than leaking throughout
2914 * the address space during boot when many long-lived
2915 * kernel allocations are made. Later some blocks near
2916 * the start are marked MIGRATE_RESERVE by
2917 * setup_zone_migrate_reserve()
2919 * bitmap is created for zone's valid pfn range. but memmap
2920 * can be created for invalid pages (for alignment)
2921 * check here not to call set_pageblock_migratetype() against
2924 if ((z
->zone_start_pfn
<= pfn
)
2925 && (pfn
< z
->zone_start_pfn
+ z
->spanned_pages
)
2926 && !(pfn
& (pageblock_nr_pages
- 1)))
2927 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
2929 INIT_LIST_HEAD(&page
->lru
);
2930 #ifdef WANT_PAGE_VIRTUAL
2931 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
2932 if (!is_highmem_idx(zone
))
2933 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
2938 static void __meminit
zone_init_free_lists(struct zone
*zone
)
2941 for_each_migratetype_order(order
, t
) {
2942 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
2943 zone
->free_area
[order
].nr_free
= 0;
2947 #ifndef __HAVE_ARCH_MEMMAP_INIT
2948 #define memmap_init(size, nid, zone, start_pfn) \
2949 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
2952 static int zone_batchsize(struct zone
*zone
)
2958 * The per-cpu-pages pools are set to around 1000th of the
2959 * size of the zone. But no more than 1/2 of a meg.
2961 * OK, so we don't know how big the cache is. So guess.
2963 batch
= zone
->present_pages
/ 1024;
2964 if (batch
* PAGE_SIZE
> 512 * 1024)
2965 batch
= (512 * 1024) / PAGE_SIZE
;
2966 batch
/= 4; /* We effectively *= 4 below */
2971 * Clamp the batch to a 2^n - 1 value. Having a power
2972 * of 2 value was found to be more likely to have
2973 * suboptimal cache aliasing properties in some cases.
2975 * For example if 2 tasks are alternately allocating
2976 * batches of pages, one task can end up with a lot
2977 * of pages of one half of the possible page colors
2978 * and the other with pages of the other colors.
2980 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
2985 /* The deferral and batching of frees should be suppressed under NOMMU
2988 * The problem is that NOMMU needs to be able to allocate large chunks
2989 * of contiguous memory as there's no hardware page translation to
2990 * assemble apparent contiguous memory from discontiguous pages.
2992 * Queueing large contiguous runs of pages for batching, however,
2993 * causes the pages to actually be freed in smaller chunks. As there
2994 * can be a significant delay between the individual batches being
2995 * recycled, this leads to the once large chunks of space being
2996 * fragmented and becoming unavailable for high-order allocations.
3002 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
3004 struct per_cpu_pages
*pcp
;
3006 memset(p
, 0, sizeof(*p
));
3010 pcp
->high
= 6 * batch
;
3011 pcp
->batch
= max(1UL, 1 * batch
);
3012 INIT_LIST_HEAD(&pcp
->list
);
3016 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3017 * to the value high for the pageset p.
3020 static void setup_pagelist_highmark(struct per_cpu_pageset
*p
,
3023 struct per_cpu_pages
*pcp
;
3027 pcp
->batch
= max(1UL, high
/4);
3028 if ((high
/4) > (PAGE_SHIFT
* 8))
3029 pcp
->batch
= PAGE_SHIFT
* 8;
3035 * Boot pageset table. One per cpu which is going to be used for all
3036 * zones and all nodes. The parameters will be set in such a way
3037 * that an item put on a list will immediately be handed over to
3038 * the buddy list. This is safe since pageset manipulation is done
3039 * with interrupts disabled.
3041 * Some NUMA counter updates may also be caught by the boot pagesets.
3043 * The boot_pagesets must be kept even after bootup is complete for
3044 * unused processors and/or zones. They do play a role for bootstrapping
3045 * hotplugged processors.
3047 * zoneinfo_show() and maybe other functions do
3048 * not check if the processor is online before following the pageset pointer.
3049 * Other parts of the kernel may not check if the zone is available.
3051 static struct per_cpu_pageset boot_pageset
[NR_CPUS
];
3054 * Dynamically allocate memory for the
3055 * per cpu pageset array in struct zone.
3057 static int __cpuinit
process_zones(int cpu
)
3059 struct zone
*zone
, *dzone
;
3060 int node
= cpu_to_node(cpu
);
3062 node_set_state(node
, N_CPU
); /* this node has a cpu */
3064 for_each_populated_zone(zone
) {
3065 zone_pcp(zone
, cpu
) = kmalloc_node(sizeof(struct per_cpu_pageset
),
3067 if (!zone_pcp(zone
, cpu
))
3070 setup_pageset(zone_pcp(zone
, cpu
), zone_batchsize(zone
));
3072 if (percpu_pagelist_fraction
)
3073 setup_pagelist_highmark(zone_pcp(zone
, cpu
),
3074 (zone
->present_pages
/ percpu_pagelist_fraction
));
3079 for_each_zone(dzone
) {
3080 if (!populated_zone(dzone
))
3084 kfree(zone_pcp(dzone
, cpu
));
3085 zone_pcp(dzone
, cpu
) = &boot_pageset
[cpu
];
3090 static inline void free_zone_pagesets(int cpu
)
3094 for_each_zone(zone
) {
3095 struct per_cpu_pageset
*pset
= zone_pcp(zone
, cpu
);
3097 /* Free per_cpu_pageset if it is slab allocated */
3098 if (pset
!= &boot_pageset
[cpu
])
3100 zone_pcp(zone
, cpu
) = &boot_pageset
[cpu
];
3104 static int __cpuinit
pageset_cpuup_callback(struct notifier_block
*nfb
,
3105 unsigned long action
,
3108 int cpu
= (long)hcpu
;
3109 int ret
= NOTIFY_OK
;
3112 case CPU_UP_PREPARE
:
3113 case CPU_UP_PREPARE_FROZEN
:
3114 if (process_zones(cpu
))
3117 case CPU_UP_CANCELED
:
3118 case CPU_UP_CANCELED_FROZEN
:
3120 case CPU_DEAD_FROZEN
:
3121 free_zone_pagesets(cpu
);
3129 static struct notifier_block __cpuinitdata pageset_notifier
=
3130 { &pageset_cpuup_callback
, NULL
, 0 };
3132 void __init
setup_per_cpu_pageset(void)
3136 /* Initialize per_cpu_pageset for cpu 0.
3137 * A cpuup callback will do this for every cpu
3138 * as it comes online
3140 err
= process_zones(smp_processor_id());
3142 register_cpu_notifier(&pageset_notifier
);
3147 static noinline __init_refok
3148 int zone_wait_table_init(struct zone
*zone
, unsigned long zone_size_pages
)
3151 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
3155 * The per-page waitqueue mechanism uses hashed waitqueues
3158 zone
->wait_table_hash_nr_entries
=
3159 wait_table_hash_nr_entries(zone_size_pages
);
3160 zone
->wait_table_bits
=
3161 wait_table_bits(zone
->wait_table_hash_nr_entries
);
3162 alloc_size
= zone
->wait_table_hash_nr_entries
3163 * sizeof(wait_queue_head_t
);
3165 if (!slab_is_available()) {
3166 zone
->wait_table
= (wait_queue_head_t
*)
3167 alloc_bootmem_node(pgdat
, alloc_size
);
3170 * This case means that a zone whose size was 0 gets new memory
3171 * via memory hot-add.
3172 * But it may be the case that a new node was hot-added. In
3173 * this case vmalloc() will not be able to use this new node's
3174 * memory - this wait_table must be initialized to use this new
3175 * node itself as well.
3176 * To use this new node's memory, further consideration will be
3179 zone
->wait_table
= vmalloc(alloc_size
);
3181 if (!zone
->wait_table
)
3184 for(i
= 0; i
< zone
->wait_table_hash_nr_entries
; ++i
)
3185 init_waitqueue_head(zone
->wait_table
+ i
);
3190 static int __zone_pcp_update(void *data
)
3192 struct zone
*zone
= data
;
3194 unsigned long batch
= zone_batchsize(zone
), flags
;
3196 for (cpu
= 0; cpu
< NR_CPUS
; cpu
++) {
3197 struct per_cpu_pageset
*pset
;
3198 struct per_cpu_pages
*pcp
;
3200 pset
= zone_pcp(zone
, cpu
);
3203 local_irq_save(flags
);
3204 free_pages_bulk(zone
, pcp
->count
, &pcp
->list
, 0);
3205 setup_pageset(pset
, batch
);
3206 local_irq_restore(flags
);
3211 void zone_pcp_update(struct zone
*zone
)
3213 stop_machine(__zone_pcp_update
, zone
, NULL
);
3216 static __meminit
void zone_pcp_init(struct zone
*zone
)
3219 unsigned long batch
= zone_batchsize(zone
);
3221 for (cpu
= 0; cpu
< NR_CPUS
; cpu
++) {
3223 /* Early boot. Slab allocator not functional yet */
3224 zone_pcp(zone
, cpu
) = &boot_pageset
[cpu
];
3225 setup_pageset(&boot_pageset
[cpu
],0);
3227 setup_pageset(zone_pcp(zone
,cpu
), batch
);
3230 if (zone
->present_pages
)
3231 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%lu\n",
3232 zone
->name
, zone
->present_pages
, batch
);
3235 __meminit
int init_currently_empty_zone(struct zone
*zone
,
3236 unsigned long zone_start_pfn
,
3238 enum memmap_context context
)
3240 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
3242 ret
= zone_wait_table_init(zone
, size
);
3245 pgdat
->nr_zones
= zone_idx(zone
) + 1;
3247 zone
->zone_start_pfn
= zone_start_pfn
;
3249 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
3250 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
3252 (unsigned long)zone_idx(zone
),
3253 zone_start_pfn
, (zone_start_pfn
+ size
));
3255 zone_init_free_lists(zone
);
3260 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3262 * Basic iterator support. Return the first range of PFNs for a node
3263 * Note: nid == MAX_NUMNODES returns first region regardless of node
3265 static int __meminit
first_active_region_index_in_nid(int nid
)
3269 for (i
= 0; i
< nr_nodemap_entries
; i
++)
3270 if (nid
== MAX_NUMNODES
|| early_node_map
[i
].nid
== nid
)
3277 * Basic iterator support. Return the next active range of PFNs for a node
3278 * Note: nid == MAX_NUMNODES returns next region regardless of node
3280 static int __meminit
next_active_region_index_in_nid(int index
, int nid
)
3282 for (index
= index
+ 1; index
< nr_nodemap_entries
; index
++)
3283 if (nid
== MAX_NUMNODES
|| early_node_map
[index
].nid
== nid
)
3289 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
3291 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
3292 * Architectures may implement their own version but if add_active_range()
3293 * was used and there are no special requirements, this is a convenient
3296 int __meminit
__early_pfn_to_nid(unsigned long pfn
)
3300 for (i
= 0; i
< nr_nodemap_entries
; i
++) {
3301 unsigned long start_pfn
= early_node_map
[i
].start_pfn
;
3302 unsigned long end_pfn
= early_node_map
[i
].end_pfn
;
3304 if (start_pfn
<= pfn
&& pfn
< end_pfn
)
3305 return early_node_map
[i
].nid
;
3307 /* This is a memory hole */
3310 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
3312 int __meminit
early_pfn_to_nid(unsigned long pfn
)
3316 nid
= __early_pfn_to_nid(pfn
);
3319 /* just returns 0 */
3323 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
3324 bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
3328 nid
= __early_pfn_to_nid(pfn
);
3329 if (nid
>= 0 && nid
!= node
)
3335 /* Basic iterator support to walk early_node_map[] */
3336 #define for_each_active_range_index_in_nid(i, nid) \
3337 for (i = first_active_region_index_in_nid(nid); i != -1; \
3338 i = next_active_region_index_in_nid(i, nid))
3341 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
3342 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
3343 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
3345 * If an architecture guarantees that all ranges registered with
3346 * add_active_ranges() contain no holes and may be freed, this
3347 * this function may be used instead of calling free_bootmem() manually.
3349 void __init
free_bootmem_with_active_regions(int nid
,
3350 unsigned long max_low_pfn
)
3354 for_each_active_range_index_in_nid(i
, nid
) {
3355 unsigned long size_pages
= 0;
3356 unsigned long end_pfn
= early_node_map
[i
].end_pfn
;
3358 if (early_node_map
[i
].start_pfn
>= max_low_pfn
)
3361 if (end_pfn
> max_low_pfn
)
3362 end_pfn
= max_low_pfn
;
3364 size_pages
= end_pfn
- early_node_map
[i
].start_pfn
;
3365 free_bootmem_node(NODE_DATA(early_node_map
[i
].nid
),
3366 PFN_PHYS(early_node_map
[i
].start_pfn
),
3367 size_pages
<< PAGE_SHIFT
);
3371 void __init
work_with_active_regions(int nid
, work_fn_t work_fn
, void *data
)
3376 for_each_active_range_index_in_nid(i
, nid
) {
3377 ret
= work_fn(early_node_map
[i
].start_pfn
,
3378 early_node_map
[i
].end_pfn
, data
);
3384 * sparse_memory_present_with_active_regions - Call memory_present for each active range
3385 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
3387 * If an architecture guarantees that all ranges registered with
3388 * add_active_ranges() contain no holes and may be freed, this
3389 * function may be used instead of calling memory_present() manually.
3391 void __init
sparse_memory_present_with_active_regions(int nid
)
3395 for_each_active_range_index_in_nid(i
, nid
)
3396 memory_present(early_node_map
[i
].nid
,
3397 early_node_map
[i
].start_pfn
,
3398 early_node_map
[i
].end_pfn
);
3402 * get_pfn_range_for_nid - Return the start and end page frames for a node
3403 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
3404 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
3405 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
3407 * It returns the start and end page frame of a node based on information
3408 * provided by an arch calling add_active_range(). If called for a node
3409 * with no available memory, a warning is printed and the start and end
3412 void __meminit
get_pfn_range_for_nid(unsigned int nid
,
3413 unsigned long *start_pfn
, unsigned long *end_pfn
)
3419 for_each_active_range_index_in_nid(i
, nid
) {
3420 *start_pfn
= min(*start_pfn
, early_node_map
[i
].start_pfn
);
3421 *end_pfn
= max(*end_pfn
, early_node_map
[i
].end_pfn
);
3424 if (*start_pfn
== -1UL)
3429 * This finds a zone that can be used for ZONE_MOVABLE pages. The
3430 * assumption is made that zones within a node are ordered in monotonic
3431 * increasing memory addresses so that the "highest" populated zone is used
3433 static void __init
find_usable_zone_for_movable(void)
3436 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
3437 if (zone_index
== ZONE_MOVABLE
)
3440 if (arch_zone_highest_possible_pfn
[zone_index
] >
3441 arch_zone_lowest_possible_pfn
[zone_index
])
3445 VM_BUG_ON(zone_index
== -1);
3446 movable_zone
= zone_index
;
3450 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
3451 * because it is sized independant of architecture. Unlike the other zones,
3452 * the starting point for ZONE_MOVABLE is not fixed. It may be different
3453 * in each node depending on the size of each node and how evenly kernelcore
3454 * is distributed. This helper function adjusts the zone ranges
3455 * provided by the architecture for a given node by using the end of the
3456 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
3457 * zones within a node are in order of monotonic increases memory addresses
3459 static void __meminit
adjust_zone_range_for_zone_movable(int nid
,
3460 unsigned long zone_type
,
3461 unsigned long node_start_pfn
,
3462 unsigned long node_end_pfn
,
3463 unsigned long *zone_start_pfn
,
3464 unsigned long *zone_end_pfn
)
3466 /* Only adjust if ZONE_MOVABLE is on this node */
3467 if (zone_movable_pfn
[nid
]) {
3468 /* Size ZONE_MOVABLE */
3469 if (zone_type
== ZONE_MOVABLE
) {
3470 *zone_start_pfn
= zone_movable_pfn
[nid
];
3471 *zone_end_pfn
= min(node_end_pfn
,
3472 arch_zone_highest_possible_pfn
[movable_zone
]);
3474 /* Adjust for ZONE_MOVABLE starting within this range */
3475 } else if (*zone_start_pfn
< zone_movable_pfn
[nid
] &&
3476 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
3477 *zone_end_pfn
= zone_movable_pfn
[nid
];
3479 /* Check if this whole range is within ZONE_MOVABLE */
3480 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
3481 *zone_start_pfn
= *zone_end_pfn
;
3486 * Return the number of pages a zone spans in a node, including holes
3487 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
3489 static unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
3490 unsigned long zone_type
,
3491 unsigned long *ignored
)
3493 unsigned long node_start_pfn
, node_end_pfn
;
3494 unsigned long zone_start_pfn
, zone_end_pfn
;
3496 /* Get the start and end of the node and zone */
3497 get_pfn_range_for_nid(nid
, &node_start_pfn
, &node_end_pfn
);
3498 zone_start_pfn
= arch_zone_lowest_possible_pfn
[zone_type
];
3499 zone_end_pfn
= arch_zone_highest_possible_pfn
[zone_type
];
3500 adjust_zone_range_for_zone_movable(nid
, zone_type
,
3501 node_start_pfn
, node_end_pfn
,
3502 &zone_start_pfn
, &zone_end_pfn
);
3504 /* Check that this node has pages within the zone's required range */
3505 if (zone_end_pfn
< node_start_pfn
|| zone_start_pfn
> node_end_pfn
)
3508 /* Move the zone boundaries inside the node if necessary */
3509 zone_end_pfn
= min(zone_end_pfn
, node_end_pfn
);
3510 zone_start_pfn
= max(zone_start_pfn
, node_start_pfn
);
3512 /* Return the spanned pages */
3513 return zone_end_pfn
- zone_start_pfn
;
3517 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
3518 * then all holes in the requested range will be accounted for.
3520 static unsigned long __meminit
__absent_pages_in_range(int nid
,
3521 unsigned long range_start_pfn
,
3522 unsigned long range_end_pfn
)
3525 unsigned long prev_end_pfn
= 0, hole_pages
= 0;
3526 unsigned long start_pfn
;
3528 /* Find the end_pfn of the first active range of pfns in the node */
3529 i
= first_active_region_index_in_nid(nid
);
3533 prev_end_pfn
= min(early_node_map
[i
].start_pfn
, range_end_pfn
);
3535 /* Account for ranges before physical memory on this node */
3536 if (early_node_map
[i
].start_pfn
> range_start_pfn
)
3537 hole_pages
= prev_end_pfn
- range_start_pfn
;
3539 /* Find all holes for the zone within the node */
3540 for (; i
!= -1; i
= next_active_region_index_in_nid(i
, nid
)) {
3542 /* No need to continue if prev_end_pfn is outside the zone */
3543 if (prev_end_pfn
>= range_end_pfn
)
3546 /* Make sure the end of the zone is not within the hole */
3547 start_pfn
= min(early_node_map
[i
].start_pfn
, range_end_pfn
);
3548 prev_end_pfn
= max(prev_end_pfn
, range_start_pfn
);
3550 /* Update the hole size cound and move on */
3551 if (start_pfn
> range_start_pfn
) {
3552 BUG_ON(prev_end_pfn
> start_pfn
);
3553 hole_pages
+= start_pfn
- prev_end_pfn
;
3555 prev_end_pfn
= early_node_map
[i
].end_pfn
;
3558 /* Account for ranges past physical memory on this node */
3559 if (range_end_pfn
> prev_end_pfn
)
3560 hole_pages
+= range_end_pfn
-
3561 max(range_start_pfn
, prev_end_pfn
);
3567 * absent_pages_in_range - Return number of page frames in holes within a range
3568 * @start_pfn: The start PFN to start searching for holes
3569 * @end_pfn: The end PFN to stop searching for holes
3571 * It returns the number of pages frames in memory holes within a range.
3573 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
3574 unsigned long end_pfn
)
3576 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
3579 /* Return the number of page frames in holes in a zone on a node */
3580 static unsigned long __meminit
zone_absent_pages_in_node(int nid
,
3581 unsigned long zone_type
,
3582 unsigned long *ignored
)
3584 unsigned long node_start_pfn
, node_end_pfn
;
3585 unsigned long zone_start_pfn
, zone_end_pfn
;
3587 get_pfn_range_for_nid(nid
, &node_start_pfn
, &node_end_pfn
);
3588 zone_start_pfn
= max(arch_zone_lowest_possible_pfn
[zone_type
],
3590 zone_end_pfn
= min(arch_zone_highest_possible_pfn
[zone_type
],
3593 adjust_zone_range_for_zone_movable(nid
, zone_type
,
3594 node_start_pfn
, node_end_pfn
,
3595 &zone_start_pfn
, &zone_end_pfn
);
3596 return __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
3600 static inline unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
3601 unsigned long zone_type
,
3602 unsigned long *zones_size
)
3604 return zones_size
[zone_type
];
3607 static inline unsigned long __meminit
zone_absent_pages_in_node(int nid
,
3608 unsigned long zone_type
,
3609 unsigned long *zholes_size
)
3614 return zholes_size
[zone_type
];
3619 static void __meminit
calculate_node_totalpages(struct pglist_data
*pgdat
,
3620 unsigned long *zones_size
, unsigned long *zholes_size
)
3622 unsigned long realtotalpages
, totalpages
= 0;
3625 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
3626 totalpages
+= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
3628 pgdat
->node_spanned_pages
= totalpages
;
3630 realtotalpages
= totalpages
;
3631 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
3633 zone_absent_pages_in_node(pgdat
->node_id
, i
,
3635 pgdat
->node_present_pages
= realtotalpages
;
3636 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
3640 #ifndef CONFIG_SPARSEMEM
3642 * Calculate the size of the zone->blockflags rounded to an unsigned long
3643 * Start by making sure zonesize is a multiple of pageblock_order by rounding
3644 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
3645 * round what is now in bits to nearest long in bits, then return it in
3648 static unsigned long __init
usemap_size(unsigned long zonesize
)
3650 unsigned long usemapsize
;
3652 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
3653 usemapsize
= usemapsize
>> pageblock_order
;
3654 usemapsize
*= NR_PAGEBLOCK_BITS
;
3655 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
3657 return usemapsize
/ 8;
3660 static void __init
setup_usemap(struct pglist_data
*pgdat
,
3661 struct zone
*zone
, unsigned long zonesize
)
3663 unsigned long usemapsize
= usemap_size(zonesize
);
3664 zone
->pageblock_flags
= NULL
;
3666 zone
->pageblock_flags
= alloc_bootmem_node(pgdat
, usemapsize
);
3669 static void inline setup_usemap(struct pglist_data
*pgdat
,
3670 struct zone
*zone
, unsigned long zonesize
) {}
3671 #endif /* CONFIG_SPARSEMEM */
3673 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
3675 /* Return a sensible default order for the pageblock size. */
3676 static inline int pageblock_default_order(void)
3678 if (HPAGE_SHIFT
> PAGE_SHIFT
)
3679 return HUGETLB_PAGE_ORDER
;
3684 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
3685 static inline void __init
set_pageblock_order(unsigned int order
)
3687 /* Check that pageblock_nr_pages has not already been setup */
3688 if (pageblock_order
)
3692 * Assume the largest contiguous order of interest is a huge page.
3693 * This value may be variable depending on boot parameters on IA64
3695 pageblock_order
= order
;
3697 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
3700 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
3701 * and pageblock_default_order() are unused as pageblock_order is set
3702 * at compile-time. See include/linux/pageblock-flags.h for the values of
3703 * pageblock_order based on the kernel config
3705 static inline int pageblock_default_order(unsigned int order
)
3709 #define set_pageblock_order(x) do {} while (0)
3711 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
3714 * Set up the zone data structures:
3715 * - mark all pages reserved
3716 * - mark all memory queues empty
3717 * - clear the memory bitmaps
3719 static void __paginginit
free_area_init_core(struct pglist_data
*pgdat
,
3720 unsigned long *zones_size
, unsigned long *zholes_size
)
3723 int nid
= pgdat
->node_id
;
3724 unsigned long zone_start_pfn
= pgdat
->node_start_pfn
;
3727 pgdat_resize_init(pgdat
);
3728 pgdat
->nr_zones
= 0;
3729 init_waitqueue_head(&pgdat
->kswapd_wait
);
3730 pgdat
->kswapd_max_order
= 0;
3731 pgdat_page_cgroup_init(pgdat
);
3733 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
3734 struct zone
*zone
= pgdat
->node_zones
+ j
;
3735 unsigned long size
, realsize
, memmap_pages
;
3738 size
= zone_spanned_pages_in_node(nid
, j
, zones_size
);
3739 realsize
= size
- zone_absent_pages_in_node(nid
, j
,
3743 * Adjust realsize so that it accounts for how much memory
3744 * is used by this zone for memmap. This affects the watermark
3745 * and per-cpu initialisations
3748 PAGE_ALIGN(size
* sizeof(struct page
)) >> PAGE_SHIFT
;
3749 if (realsize
>= memmap_pages
) {
3750 realsize
-= memmap_pages
;
3753 " %s zone: %lu pages used for memmap\n",
3754 zone_names
[j
], memmap_pages
);
3757 " %s zone: %lu pages exceeds realsize %lu\n",
3758 zone_names
[j
], memmap_pages
, realsize
);
3760 /* Account for reserved pages */
3761 if (j
== 0 && realsize
> dma_reserve
) {
3762 realsize
-= dma_reserve
;
3763 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
3764 zone_names
[0], dma_reserve
);
3767 if (!is_highmem_idx(j
))
3768 nr_kernel_pages
+= realsize
;
3769 nr_all_pages
+= realsize
;
3771 zone
->spanned_pages
= size
;
3772 zone
->present_pages
= realsize
;
3775 zone
->min_unmapped_pages
= (realsize
*sysctl_min_unmapped_ratio
)
3777 zone
->min_slab_pages
= (realsize
* sysctl_min_slab_ratio
) / 100;
3779 zone
->name
= zone_names
[j
];
3780 spin_lock_init(&zone
->lock
);
3781 spin_lock_init(&zone
->lru_lock
);
3782 zone_seqlock_init(zone
);
3783 zone
->zone_pgdat
= pgdat
;
3785 zone
->prev_priority
= DEF_PRIORITY
;
3787 zone_pcp_init(zone
);
3789 INIT_LIST_HEAD(&zone
->lru
[l
].list
);
3790 zone
->lru
[l
].nr_saved_scan
= 0;
3792 zone
->reclaim_stat
.recent_rotated
[0] = 0;
3793 zone
->reclaim_stat
.recent_rotated
[1] = 0;
3794 zone
->reclaim_stat
.recent_scanned
[0] = 0;
3795 zone
->reclaim_stat
.recent_scanned
[1] = 0;
3796 zap_zone_vm_stats(zone
);
3801 set_pageblock_order(pageblock_default_order());
3802 setup_usemap(pgdat
, zone
, size
);
3803 ret
= init_currently_empty_zone(zone
, zone_start_pfn
,
3804 size
, MEMMAP_EARLY
);
3806 memmap_init(size
, nid
, j
, zone_start_pfn
);
3807 zone_start_pfn
+= size
;
3811 static void __init_refok
alloc_node_mem_map(struct pglist_data
*pgdat
)
3813 /* Skip empty nodes */
3814 if (!pgdat
->node_spanned_pages
)
3817 #ifdef CONFIG_FLAT_NODE_MEM_MAP
3818 /* ia64 gets its own node_mem_map, before this, without bootmem */
3819 if (!pgdat
->node_mem_map
) {
3820 unsigned long size
, start
, end
;
3824 * The zone's endpoints aren't required to be MAX_ORDER
3825 * aligned but the node_mem_map endpoints must be in order
3826 * for the buddy allocator to function correctly.
3828 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
3829 end
= pgdat
->node_start_pfn
+ pgdat
->node_spanned_pages
;
3830 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
3831 size
= (end
- start
) * sizeof(struct page
);
3832 map
= alloc_remap(pgdat
->node_id
, size
);
3834 map
= alloc_bootmem_node(pgdat
, size
);
3835 pgdat
->node_mem_map
= map
+ (pgdat
->node_start_pfn
- start
);
3837 #ifndef CONFIG_NEED_MULTIPLE_NODES
3839 * With no DISCONTIG, the global mem_map is just set as node 0's
3841 if (pgdat
== NODE_DATA(0)) {
3842 mem_map
= NODE_DATA(0)->node_mem_map
;
3843 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3844 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
3845 mem_map
-= (pgdat
->node_start_pfn
- ARCH_PFN_OFFSET
);
3846 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
3849 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
3852 void __paginginit
free_area_init_node(int nid
, unsigned long *zones_size
,
3853 unsigned long node_start_pfn
, unsigned long *zholes_size
)
3855 pg_data_t
*pgdat
= NODE_DATA(nid
);
3857 pgdat
->node_id
= nid
;
3858 pgdat
->node_start_pfn
= node_start_pfn
;
3859 calculate_node_totalpages(pgdat
, zones_size
, zholes_size
);
3861 alloc_node_mem_map(pgdat
);
3862 #ifdef CONFIG_FLAT_NODE_MEM_MAP
3863 printk(KERN_DEBUG
"free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
3864 nid
, (unsigned long)pgdat
,
3865 (unsigned long)pgdat
->node_mem_map
);
3868 free_area_init_core(pgdat
, zones_size
, zholes_size
);
3871 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3873 #if MAX_NUMNODES > 1
3875 * Figure out the number of possible node ids.
3877 static void __init
setup_nr_node_ids(void)
3880 unsigned int highest
= 0;
3882 for_each_node_mask(node
, node_possible_map
)
3884 nr_node_ids
= highest
+ 1;
3887 static inline void setup_nr_node_ids(void)
3893 * add_active_range - Register a range of PFNs backed by physical memory
3894 * @nid: The node ID the range resides on
3895 * @start_pfn: The start PFN of the available physical memory
3896 * @end_pfn: The end PFN of the available physical memory
3898 * These ranges are stored in an early_node_map[] and later used by
3899 * free_area_init_nodes() to calculate zone sizes and holes. If the
3900 * range spans a memory hole, it is up to the architecture to ensure
3901 * the memory is not freed by the bootmem allocator. If possible
3902 * the range being registered will be merged with existing ranges.
3904 void __init
add_active_range(unsigned int nid
, unsigned long start_pfn
,
3905 unsigned long end_pfn
)
3909 mminit_dprintk(MMINIT_TRACE
, "memory_register",
3910 "Entering add_active_range(%d, %#lx, %#lx) "
3911 "%d entries of %d used\n",
3912 nid
, start_pfn
, end_pfn
,
3913 nr_nodemap_entries
, MAX_ACTIVE_REGIONS
);
3915 mminit_validate_memmodel_limits(&start_pfn
, &end_pfn
);
3917 /* Merge with existing active regions if possible */
3918 for (i
= 0; i
< nr_nodemap_entries
; i
++) {
3919 if (early_node_map
[i
].nid
!= nid
)
3922 /* Skip if an existing region covers this new one */
3923 if (start_pfn
>= early_node_map
[i
].start_pfn
&&
3924 end_pfn
<= early_node_map
[i
].end_pfn
)
3927 /* Merge forward if suitable */
3928 if (start_pfn
<= early_node_map
[i
].end_pfn
&&
3929 end_pfn
> early_node_map
[i
].end_pfn
) {
3930 early_node_map
[i
].end_pfn
= end_pfn
;
3934 /* Merge backward if suitable */
3935 if (start_pfn
< early_node_map
[i
].end_pfn
&&
3936 end_pfn
>= early_node_map
[i
].start_pfn
) {
3937 early_node_map
[i
].start_pfn
= start_pfn
;
3942 /* Check that early_node_map is large enough */
3943 if (i
>= MAX_ACTIVE_REGIONS
) {
3944 printk(KERN_CRIT
"More than %d memory regions, truncating\n",
3945 MAX_ACTIVE_REGIONS
);
3949 early_node_map
[i
].nid
= nid
;
3950 early_node_map
[i
].start_pfn
= start_pfn
;
3951 early_node_map
[i
].end_pfn
= end_pfn
;
3952 nr_nodemap_entries
= i
+ 1;
3956 * remove_active_range - Shrink an existing registered range of PFNs
3957 * @nid: The node id the range is on that should be shrunk
3958 * @start_pfn: The new PFN of the range
3959 * @end_pfn: The new PFN of the range
3961 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
3962 * The map is kept near the end physical page range that has already been
3963 * registered. This function allows an arch to shrink an existing registered
3966 void __init
remove_active_range(unsigned int nid
, unsigned long start_pfn
,
3967 unsigned long end_pfn
)
3972 printk(KERN_DEBUG
"remove_active_range (%d, %lu, %lu)\n",
3973 nid
, start_pfn
, end_pfn
);
3975 /* Find the old active region end and shrink */
3976 for_each_active_range_index_in_nid(i
, nid
) {
3977 if (early_node_map
[i
].start_pfn
>= start_pfn
&&
3978 early_node_map
[i
].end_pfn
<= end_pfn
) {
3980 early_node_map
[i
].start_pfn
= 0;
3981 early_node_map
[i
].end_pfn
= 0;
3985 if (early_node_map
[i
].start_pfn
< start_pfn
&&
3986 early_node_map
[i
].end_pfn
> start_pfn
) {
3987 unsigned long temp_end_pfn
= early_node_map
[i
].end_pfn
;
3988 early_node_map
[i
].end_pfn
= start_pfn
;
3989 if (temp_end_pfn
> end_pfn
)
3990 add_active_range(nid
, end_pfn
, temp_end_pfn
);
3993 if (early_node_map
[i
].start_pfn
>= start_pfn
&&
3994 early_node_map
[i
].end_pfn
> end_pfn
&&
3995 early_node_map
[i
].start_pfn
< end_pfn
) {
3996 early_node_map
[i
].start_pfn
= end_pfn
;
4004 /* remove the blank ones */
4005 for (i
= nr_nodemap_entries
- 1; i
> 0; i
--) {
4006 if (early_node_map
[i
].nid
!= nid
)
4008 if (early_node_map
[i
].end_pfn
)
4010 /* we found it, get rid of it */
4011 for (j
= i
; j
< nr_nodemap_entries
- 1; j
++)
4012 memcpy(&early_node_map
[j
], &early_node_map
[j
+1],
4013 sizeof(early_node_map
[j
]));
4014 j
= nr_nodemap_entries
- 1;
4015 memset(&early_node_map
[j
], 0, sizeof(early_node_map
[j
]));
4016 nr_nodemap_entries
--;
4021 * remove_all_active_ranges - Remove all currently registered regions
4023 * During discovery, it may be found that a table like SRAT is invalid
4024 * and an alternative discovery method must be used. This function removes
4025 * all currently registered regions.
4027 void __init
remove_all_active_ranges(void)
4029 memset(early_node_map
, 0, sizeof(early_node_map
));
4030 nr_nodemap_entries
= 0;
4033 /* Compare two active node_active_regions */
4034 static int __init
cmp_node_active_region(const void *a
, const void *b
)
4036 struct node_active_region
*arange
= (struct node_active_region
*)a
;
4037 struct node_active_region
*brange
= (struct node_active_region
*)b
;
4039 /* Done this way to avoid overflows */
4040 if (arange
->start_pfn
> brange
->start_pfn
)
4042 if (arange
->start_pfn
< brange
->start_pfn
)
4048 /* sort the node_map by start_pfn */
4049 static void __init
sort_node_map(void)
4051 sort(early_node_map
, (size_t)nr_nodemap_entries
,
4052 sizeof(struct node_active_region
),
4053 cmp_node_active_region
, NULL
);
4056 /* Find the lowest pfn for a node */
4057 static unsigned long __init
find_min_pfn_for_node(int nid
)
4060 unsigned long min_pfn
= ULONG_MAX
;
4062 /* Assuming a sorted map, the first range found has the starting pfn */
4063 for_each_active_range_index_in_nid(i
, nid
)
4064 min_pfn
= min(min_pfn
, early_node_map
[i
].start_pfn
);
4066 if (min_pfn
== ULONG_MAX
) {
4068 "Could not find start_pfn for node %d\n", nid
);
4076 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4078 * It returns the minimum PFN based on information provided via
4079 * add_active_range().
4081 unsigned long __init
find_min_pfn_with_active_regions(void)
4083 return find_min_pfn_for_node(MAX_NUMNODES
);
4087 * early_calculate_totalpages()
4088 * Sum pages in active regions for movable zone.
4089 * Populate N_HIGH_MEMORY for calculating usable_nodes.
4091 static unsigned long __init
early_calculate_totalpages(void)
4094 unsigned long totalpages
= 0;
4096 for (i
= 0; i
< nr_nodemap_entries
; i
++) {
4097 unsigned long pages
= early_node_map
[i
].end_pfn
-
4098 early_node_map
[i
].start_pfn
;
4099 totalpages
+= pages
;
4101 node_set_state(early_node_map
[i
].nid
, N_HIGH_MEMORY
);
4107 * Find the PFN the Movable zone begins in each node. Kernel memory
4108 * is spread evenly between nodes as long as the nodes have enough
4109 * memory. When they don't, some nodes will have more kernelcore than
4112 static void __init
find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn
)
4115 unsigned long usable_startpfn
;
4116 unsigned long kernelcore_node
, kernelcore_remaining
;
4117 /* save the state before borrow the nodemask */
4118 nodemask_t saved_node_state
= node_states
[N_HIGH_MEMORY
];
4119 unsigned long totalpages
= early_calculate_totalpages();
4120 int usable_nodes
= nodes_weight(node_states
[N_HIGH_MEMORY
]);
4123 * If movablecore was specified, calculate what size of
4124 * kernelcore that corresponds so that memory usable for
4125 * any allocation type is evenly spread. If both kernelcore
4126 * and movablecore are specified, then the value of kernelcore
4127 * will be used for required_kernelcore if it's greater than
4128 * what movablecore would have allowed.
4130 if (required_movablecore
) {
4131 unsigned long corepages
;
4134 * Round-up so that ZONE_MOVABLE is at least as large as what
4135 * was requested by the user
4137 required_movablecore
=
4138 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
4139 corepages
= totalpages
- required_movablecore
;
4141 required_kernelcore
= max(required_kernelcore
, corepages
);
4144 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4145 if (!required_kernelcore
)
4148 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4149 find_usable_zone_for_movable();
4150 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
4153 /* Spread kernelcore memory as evenly as possible throughout nodes */
4154 kernelcore_node
= required_kernelcore
/ usable_nodes
;
4155 for_each_node_state(nid
, N_HIGH_MEMORY
) {
4157 * Recalculate kernelcore_node if the division per node
4158 * now exceeds what is necessary to satisfy the requested
4159 * amount of memory for the kernel
4161 if (required_kernelcore
< kernelcore_node
)
4162 kernelcore_node
= required_kernelcore
/ usable_nodes
;
4165 * As the map is walked, we track how much memory is usable
4166 * by the kernel using kernelcore_remaining. When it is
4167 * 0, the rest of the node is usable by ZONE_MOVABLE
4169 kernelcore_remaining
= kernelcore_node
;
4171 /* Go through each range of PFNs within this node */
4172 for_each_active_range_index_in_nid(i
, nid
) {
4173 unsigned long start_pfn
, end_pfn
;
4174 unsigned long size_pages
;
4176 start_pfn
= max(early_node_map
[i
].start_pfn
,
4177 zone_movable_pfn
[nid
]);
4178 end_pfn
= early_node_map
[i
].end_pfn
;
4179 if (start_pfn
>= end_pfn
)
4182 /* Account for what is only usable for kernelcore */
4183 if (start_pfn
< usable_startpfn
) {
4184 unsigned long kernel_pages
;
4185 kernel_pages
= min(end_pfn
, usable_startpfn
)
4188 kernelcore_remaining
-= min(kernel_pages
,
4189 kernelcore_remaining
);
4190 required_kernelcore
-= min(kernel_pages
,
4191 required_kernelcore
);
4193 /* Continue if range is now fully accounted */
4194 if (end_pfn
<= usable_startpfn
) {
4197 * Push zone_movable_pfn to the end so
4198 * that if we have to rebalance
4199 * kernelcore across nodes, we will
4200 * not double account here
4202 zone_movable_pfn
[nid
] = end_pfn
;
4205 start_pfn
= usable_startpfn
;
4209 * The usable PFN range for ZONE_MOVABLE is from
4210 * start_pfn->end_pfn. Calculate size_pages as the
4211 * number of pages used as kernelcore
4213 size_pages
= end_pfn
- start_pfn
;
4214 if (size_pages
> kernelcore_remaining
)
4215 size_pages
= kernelcore_remaining
;
4216 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
4219 * Some kernelcore has been met, update counts and
4220 * break if the kernelcore for this node has been
4223 required_kernelcore
-= min(required_kernelcore
,
4225 kernelcore_remaining
-= size_pages
;
4226 if (!kernelcore_remaining
)
4232 * If there is still required_kernelcore, we do another pass with one
4233 * less node in the count. This will push zone_movable_pfn[nid] further
4234 * along on the nodes that still have memory until kernelcore is
4238 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
4241 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4242 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
4243 zone_movable_pfn
[nid
] =
4244 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
4247 /* restore the node_state */
4248 node_states
[N_HIGH_MEMORY
] = saved_node_state
;
4251 /* Any regular memory on that node ? */
4252 static void check_for_regular_memory(pg_data_t
*pgdat
)
4254 #ifdef CONFIG_HIGHMEM
4255 enum zone_type zone_type
;
4257 for (zone_type
= 0; zone_type
<= ZONE_NORMAL
; zone_type
++) {
4258 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
4259 if (zone
->present_pages
)
4260 node_set_state(zone_to_nid(zone
), N_NORMAL_MEMORY
);
4266 * free_area_init_nodes - Initialise all pg_data_t and zone data
4267 * @max_zone_pfn: an array of max PFNs for each zone
4269 * This will call free_area_init_node() for each active node in the system.
4270 * Using the page ranges provided by add_active_range(), the size of each
4271 * zone in each node and their holes is calculated. If the maximum PFN
4272 * between two adjacent zones match, it is assumed that the zone is empty.
4273 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4274 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4275 * starts where the previous one ended. For example, ZONE_DMA32 starts
4276 * at arch_max_dma_pfn.
4278 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
4283 /* Sort early_node_map as initialisation assumes it is sorted */
4286 /* Record where the zone boundaries are */
4287 memset(arch_zone_lowest_possible_pfn
, 0,
4288 sizeof(arch_zone_lowest_possible_pfn
));
4289 memset(arch_zone_highest_possible_pfn
, 0,
4290 sizeof(arch_zone_highest_possible_pfn
));
4291 arch_zone_lowest_possible_pfn
[0] = find_min_pfn_with_active_regions();
4292 arch_zone_highest_possible_pfn
[0] = max_zone_pfn
[0];
4293 for (i
= 1; i
< MAX_NR_ZONES
; i
++) {
4294 if (i
== ZONE_MOVABLE
)
4296 arch_zone_lowest_possible_pfn
[i
] =
4297 arch_zone_highest_possible_pfn
[i
-1];
4298 arch_zone_highest_possible_pfn
[i
] =
4299 max(max_zone_pfn
[i
], arch_zone_lowest_possible_pfn
[i
]);
4301 arch_zone_lowest_possible_pfn
[ZONE_MOVABLE
] = 0;
4302 arch_zone_highest_possible_pfn
[ZONE_MOVABLE
] = 0;
4304 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4305 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
4306 find_zone_movable_pfns_for_nodes(zone_movable_pfn
);
4308 /* Print out the zone ranges */
4309 printk("Zone PFN ranges:\n");
4310 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
4311 if (i
== ZONE_MOVABLE
)
4313 printk(" %-8s %0#10lx -> %0#10lx\n",
4315 arch_zone_lowest_possible_pfn
[i
],
4316 arch_zone_highest_possible_pfn
[i
]);
4319 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4320 printk("Movable zone start PFN for each node\n");
4321 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
4322 if (zone_movable_pfn
[i
])
4323 printk(" Node %d: %lu\n", i
, zone_movable_pfn
[i
]);
4326 /* Print out the early_node_map[] */
4327 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries
);
4328 for (i
= 0; i
< nr_nodemap_entries
; i
++)
4329 printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map
[i
].nid
,
4330 early_node_map
[i
].start_pfn
,
4331 early_node_map
[i
].end_pfn
);
4333 /* Initialise every node */
4334 mminit_verify_pageflags_layout();
4335 setup_nr_node_ids();
4336 for_each_online_node(nid
) {
4337 pg_data_t
*pgdat
= NODE_DATA(nid
);
4338 free_area_init_node(nid
, NULL
,
4339 find_min_pfn_for_node(nid
), NULL
);
4341 /* Any memory on that node */
4342 if (pgdat
->node_present_pages
)
4343 node_set_state(nid
, N_HIGH_MEMORY
);
4344 check_for_regular_memory(pgdat
);
4348 static int __init
cmdline_parse_core(char *p
, unsigned long *core
)
4350 unsigned long long coremem
;
4354 coremem
= memparse(p
, &p
);
4355 *core
= coremem
>> PAGE_SHIFT
;
4357 /* Paranoid check that UL is enough for the coremem value */
4358 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
4364 * kernelcore=size sets the amount of memory for use for allocations that
4365 * cannot be reclaimed or migrated.
4367 static int __init
cmdline_parse_kernelcore(char *p
)
4369 return cmdline_parse_core(p
, &required_kernelcore
);
4373 * movablecore=size sets the amount of memory for use for allocations that
4374 * can be reclaimed or migrated.
4376 static int __init
cmdline_parse_movablecore(char *p
)
4378 return cmdline_parse_core(p
, &required_movablecore
);
4381 early_param("kernelcore", cmdline_parse_kernelcore
);
4382 early_param("movablecore", cmdline_parse_movablecore
);
4384 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4387 * set_dma_reserve - set the specified number of pages reserved in the first zone
4388 * @new_dma_reserve: The number of pages to mark reserved
4390 * The per-cpu batchsize and zone watermarks are determined by present_pages.
4391 * In the DMA zone, a significant percentage may be consumed by kernel image
4392 * and other unfreeable allocations which can skew the watermarks badly. This
4393 * function may optionally be used to account for unfreeable pages in the
4394 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
4395 * smaller per-cpu batchsize.
4397 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
4399 dma_reserve
= new_dma_reserve
;
4402 #ifndef CONFIG_NEED_MULTIPLE_NODES
4403 struct pglist_data __refdata contig_page_data
= { .bdata
= &bootmem_node_data
[0] };
4404 EXPORT_SYMBOL(contig_page_data
);
4407 void __init
free_area_init(unsigned long *zones_size
)
4409 free_area_init_node(0, zones_size
,
4410 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
4413 static int page_alloc_cpu_notify(struct notifier_block
*self
,
4414 unsigned long action
, void *hcpu
)
4416 int cpu
= (unsigned long)hcpu
;
4418 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
) {
4422 * Spill the event counters of the dead processor
4423 * into the current processors event counters.
4424 * This artificially elevates the count of the current
4427 vm_events_fold_cpu(cpu
);
4430 * Zero the differential counters of the dead processor
4431 * so that the vm statistics are consistent.
4433 * This is only okay since the processor is dead and cannot
4434 * race with what we are doing.
4436 refresh_cpu_vm_stats(cpu
);
4441 void __init
page_alloc_init(void)
4443 hotcpu_notifier(page_alloc_cpu_notify
, 0);
4447 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
4448 * or min_free_kbytes changes.
4450 static void calculate_totalreserve_pages(void)
4452 struct pglist_data
*pgdat
;
4453 unsigned long reserve_pages
= 0;
4454 enum zone_type i
, j
;
4456 for_each_online_pgdat(pgdat
) {
4457 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
4458 struct zone
*zone
= pgdat
->node_zones
+ i
;
4459 unsigned long max
= 0;
4461 /* Find valid and maximum lowmem_reserve in the zone */
4462 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
4463 if (zone
->lowmem_reserve
[j
] > max
)
4464 max
= zone
->lowmem_reserve
[j
];
4467 /* we treat the high watermark as reserved pages. */
4468 max
+= high_wmark_pages(zone
);
4470 if (max
> zone
->present_pages
)
4471 max
= zone
->present_pages
;
4472 reserve_pages
+= max
;
4475 totalreserve_pages
= reserve_pages
;
4479 * setup_per_zone_lowmem_reserve - called whenever
4480 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
4481 * has a correct pages reserved value, so an adequate number of
4482 * pages are left in the zone after a successful __alloc_pages().
4484 static void setup_per_zone_lowmem_reserve(void)
4486 struct pglist_data
*pgdat
;
4487 enum zone_type j
, idx
;
4489 for_each_online_pgdat(pgdat
) {
4490 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
4491 struct zone
*zone
= pgdat
->node_zones
+ j
;
4492 unsigned long present_pages
= zone
->present_pages
;
4494 zone
->lowmem_reserve
[j
] = 0;
4498 struct zone
*lower_zone
;
4502 if (sysctl_lowmem_reserve_ratio
[idx
] < 1)
4503 sysctl_lowmem_reserve_ratio
[idx
] = 1;
4505 lower_zone
= pgdat
->node_zones
+ idx
;
4506 lower_zone
->lowmem_reserve
[j
] = present_pages
/
4507 sysctl_lowmem_reserve_ratio
[idx
];
4508 present_pages
+= lower_zone
->present_pages
;
4513 /* update totalreserve_pages */
4514 calculate_totalreserve_pages();
4518 * setup_per_zone_wmarks - called when min_free_kbytes changes
4519 * or when memory is hot-{added|removed}
4521 * Ensures that the watermark[min,low,high] values for each zone are set
4522 * correctly with respect to min_free_kbytes.
4524 void setup_per_zone_wmarks(void)
4526 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
4527 unsigned long lowmem_pages
= 0;
4529 unsigned long flags
;
4531 /* Calculate total number of !ZONE_HIGHMEM pages */
4532 for_each_zone(zone
) {
4533 if (!is_highmem(zone
))
4534 lowmem_pages
+= zone
->present_pages
;
4537 for_each_zone(zone
) {
4540 spin_lock_irqsave(&zone
->lock
, flags
);
4541 tmp
= (u64
)pages_min
* zone
->present_pages
;
4542 do_div(tmp
, lowmem_pages
);
4543 if (is_highmem(zone
)) {
4545 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
4546 * need highmem pages, so cap pages_min to a small
4549 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
4550 * deltas controls asynch page reclaim, and so should
4551 * not be capped for highmem.
4555 min_pages
= zone
->present_pages
/ 1024;
4556 if (min_pages
< SWAP_CLUSTER_MAX
)
4557 min_pages
= SWAP_CLUSTER_MAX
;
4558 if (min_pages
> 128)
4560 zone
->watermark
[WMARK_MIN
] = min_pages
;
4563 * If it's a lowmem zone, reserve a number of pages
4564 * proportionate to the zone's size.
4566 zone
->watermark
[WMARK_MIN
] = tmp
;
4569 zone
->watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + (tmp
>> 2);
4570 zone
->watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + (tmp
>> 1);
4571 setup_zone_migrate_reserve(zone
);
4572 spin_unlock_irqrestore(&zone
->lock
, flags
);
4575 /* update totalreserve_pages */
4576 calculate_totalreserve_pages();
4580 * The inactive anon list should be small enough that the VM never has to
4581 * do too much work, but large enough that each inactive page has a chance
4582 * to be referenced again before it is swapped out.
4584 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
4585 * INACTIVE_ANON pages on this zone's LRU, maintained by the
4586 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
4587 * the anonymous pages are kept on the inactive list.
4590 * memory ratio inactive anon
4591 * -------------------------------------
4600 void calculate_zone_inactive_ratio(struct zone
*zone
)
4602 unsigned int gb
, ratio
;
4604 /* Zone size in gigabytes */
4605 gb
= zone
->present_pages
>> (30 - PAGE_SHIFT
);
4607 ratio
= int_sqrt(10 * gb
);
4611 zone
->inactive_ratio
= ratio
;
4614 static void __init
setup_per_zone_inactive_ratio(void)
4619 calculate_zone_inactive_ratio(zone
);
4623 * Initialise min_free_kbytes.
4625 * For small machines we want it small (128k min). For large machines
4626 * we want it large (64MB max). But it is not linear, because network
4627 * bandwidth does not increase linearly with machine size. We use
4629 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
4630 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
4646 static int __init
init_per_zone_wmark_min(void)
4648 unsigned long lowmem_kbytes
;
4650 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
4652 min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
4653 if (min_free_kbytes
< 128)
4654 min_free_kbytes
= 128;
4655 if (min_free_kbytes
> 65536)
4656 min_free_kbytes
= 65536;
4657 setup_per_zone_wmarks();
4658 setup_per_zone_lowmem_reserve();
4659 setup_per_zone_inactive_ratio();
4662 module_init(init_per_zone_wmark_min
)
4665 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
4666 * that we can call two helper functions whenever min_free_kbytes
4669 int min_free_kbytes_sysctl_handler(ctl_table
*table
, int write
,
4670 struct file
*file
, void __user
*buffer
, size_t *length
, loff_t
*ppos
)
4672 proc_dointvec(table
, write
, file
, buffer
, length
, ppos
);
4674 setup_per_zone_wmarks();
4679 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table
*table
, int write
,
4680 struct file
*file
, void __user
*buffer
, size_t *length
, loff_t
*ppos
)
4685 rc
= proc_dointvec_minmax(table
, write
, file
, buffer
, length
, ppos
);
4690 zone
->min_unmapped_pages
= (zone
->present_pages
*
4691 sysctl_min_unmapped_ratio
) / 100;
4695 int sysctl_min_slab_ratio_sysctl_handler(ctl_table
*table
, int write
,
4696 struct file
*file
, void __user
*buffer
, size_t *length
, loff_t
*ppos
)
4701 rc
= proc_dointvec_minmax(table
, write
, file
, buffer
, length
, ppos
);
4706 zone
->min_slab_pages
= (zone
->present_pages
*
4707 sysctl_min_slab_ratio
) / 100;
4713 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
4714 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
4715 * whenever sysctl_lowmem_reserve_ratio changes.
4717 * The reserve ratio obviously has absolutely no relation with the
4718 * minimum watermarks. The lowmem reserve ratio can only make sense
4719 * if in function of the boot time zone sizes.
4721 int lowmem_reserve_ratio_sysctl_handler(ctl_table
*table
, int write
,
4722 struct file
*file
, void __user
*buffer
, size_t *length
, loff_t
*ppos
)
4724 proc_dointvec_minmax(table
, write
, file
, buffer
, length
, ppos
);
4725 setup_per_zone_lowmem_reserve();
4730 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
4731 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
4732 * can have before it gets flushed back to buddy allocator.
4735 int percpu_pagelist_fraction_sysctl_handler(ctl_table
*table
, int write
,
4736 struct file
*file
, void __user
*buffer
, size_t *length
, loff_t
*ppos
)
4742 ret
= proc_dointvec_minmax(table
, write
, file
, buffer
, length
, ppos
);
4743 if (!write
|| (ret
== -EINVAL
))
4745 for_each_populated_zone(zone
) {
4746 for_each_online_cpu(cpu
) {
4748 high
= zone
->present_pages
/ percpu_pagelist_fraction
;
4749 setup_pagelist_highmark(zone_pcp(zone
, cpu
), high
);
4755 int hashdist
= HASHDIST_DEFAULT
;
4758 static int __init
set_hashdist(char *str
)
4762 hashdist
= simple_strtoul(str
, &str
, 0);
4765 __setup("hashdist=", set_hashdist
);
4769 * allocate a large system hash table from bootmem
4770 * - it is assumed that the hash table must contain an exact power-of-2
4771 * quantity of entries
4772 * - limit is the number of hash buckets, not the total allocation size
4774 void *__init
alloc_large_system_hash(const char *tablename
,
4775 unsigned long bucketsize
,
4776 unsigned long numentries
,
4779 unsigned int *_hash_shift
,
4780 unsigned int *_hash_mask
,
4781 unsigned long limit
)
4783 unsigned long long max
= limit
;
4784 unsigned long log2qty
, size
;
4787 /* allow the kernel cmdline to have a say */
4789 /* round applicable memory size up to nearest megabyte */
4790 numentries
= nr_kernel_pages
;
4791 numentries
+= (1UL << (20 - PAGE_SHIFT
)) - 1;
4792 numentries
>>= 20 - PAGE_SHIFT
;
4793 numentries
<<= 20 - PAGE_SHIFT
;
4795 /* limit to 1 bucket per 2^scale bytes of low memory */
4796 if (scale
> PAGE_SHIFT
)
4797 numentries
>>= (scale
- PAGE_SHIFT
);
4799 numentries
<<= (PAGE_SHIFT
- scale
);
4801 /* Make sure we've got at least a 0-order allocation.. */
4802 if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
4803 numentries
= PAGE_SIZE
/ bucketsize
;
4805 numentries
= roundup_pow_of_two(numentries
);
4807 /* limit allocation size to 1/16 total memory by default */
4809 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
4810 do_div(max
, bucketsize
);
4813 if (numentries
> max
)
4816 log2qty
= ilog2(numentries
);
4819 size
= bucketsize
<< log2qty
;
4820 if (flags
& HASH_EARLY
)
4821 table
= alloc_bootmem_nopanic(size
);
4823 table
= __vmalloc(size
, GFP_ATOMIC
, PAGE_KERNEL
);
4826 * If bucketsize is not a power-of-two, we may free
4827 * some pages at the end of hash table which
4828 * alloc_pages_exact() automatically does
4830 if (get_order(size
) < MAX_ORDER
) {
4831 table
= alloc_pages_exact(size
, GFP_ATOMIC
);
4832 kmemleak_alloc(table
, size
, 1, GFP_ATOMIC
);
4835 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
4838 panic("Failed to allocate %s hash table\n", tablename
);
4840 printk(KERN_INFO
"%s hash table entries: %d (order: %d, %lu bytes)\n",
4843 ilog2(size
) - PAGE_SHIFT
,
4847 *_hash_shift
= log2qty
;
4849 *_hash_mask
= (1 << log2qty
) - 1;
4854 /* Return a pointer to the bitmap storing bits affecting a block of pages */
4855 static inline unsigned long *get_pageblock_bitmap(struct zone
*zone
,
4858 #ifdef CONFIG_SPARSEMEM
4859 return __pfn_to_section(pfn
)->pageblock_flags
;
4861 return zone
->pageblock_flags
;
4862 #endif /* CONFIG_SPARSEMEM */
4865 static inline int pfn_to_bitidx(struct zone
*zone
, unsigned long pfn
)
4867 #ifdef CONFIG_SPARSEMEM
4868 pfn
&= (PAGES_PER_SECTION
-1);
4869 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
4871 pfn
= pfn
- zone
->zone_start_pfn
;
4872 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
4873 #endif /* CONFIG_SPARSEMEM */
4877 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
4878 * @page: The page within the block of interest
4879 * @start_bitidx: The first bit of interest to retrieve
4880 * @end_bitidx: The last bit of interest
4881 * returns pageblock_bits flags
4883 unsigned long get_pageblock_flags_group(struct page
*page
,
4884 int start_bitidx
, int end_bitidx
)
4887 unsigned long *bitmap
;
4888 unsigned long pfn
, bitidx
;
4889 unsigned long flags
= 0;
4890 unsigned long value
= 1;
4892 zone
= page_zone(page
);
4893 pfn
= page_to_pfn(page
);
4894 bitmap
= get_pageblock_bitmap(zone
, pfn
);
4895 bitidx
= pfn_to_bitidx(zone
, pfn
);
4897 for (; start_bitidx
<= end_bitidx
; start_bitidx
++, value
<<= 1)
4898 if (test_bit(bitidx
+ start_bitidx
, bitmap
))
4905 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
4906 * @page: The page within the block of interest
4907 * @start_bitidx: The first bit of interest
4908 * @end_bitidx: The last bit of interest
4909 * @flags: The flags to set
4911 void set_pageblock_flags_group(struct page
*page
, unsigned long flags
,
4912 int start_bitidx
, int end_bitidx
)
4915 unsigned long *bitmap
;
4916 unsigned long pfn
, bitidx
;
4917 unsigned long value
= 1;
4919 zone
= page_zone(page
);
4920 pfn
= page_to_pfn(page
);
4921 bitmap
= get_pageblock_bitmap(zone
, pfn
);
4922 bitidx
= pfn_to_bitidx(zone
, pfn
);
4923 VM_BUG_ON(pfn
< zone
->zone_start_pfn
);
4924 VM_BUG_ON(pfn
>= zone
->zone_start_pfn
+ zone
->spanned_pages
);
4926 for (; start_bitidx
<= end_bitidx
; start_bitidx
++, value
<<= 1)
4928 __set_bit(bitidx
+ start_bitidx
, bitmap
);
4930 __clear_bit(bitidx
+ start_bitidx
, bitmap
);
4934 * This is designed as sub function...plz see page_isolation.c also.
4935 * set/clear page block's type to be ISOLATE.
4936 * page allocater never alloc memory from ISOLATE block.
4939 int set_migratetype_isolate(struct page
*page
)
4942 unsigned long flags
;
4946 zone
= page_zone(page
);
4947 zone_idx
= zone_idx(zone
);
4948 spin_lock_irqsave(&zone
->lock
, flags
);
4950 * In future, more migrate types will be able to be isolation target.
4952 if (get_pageblock_migratetype(page
) != MIGRATE_MOVABLE
&&
4953 zone_idx
!= ZONE_MOVABLE
)
4955 set_pageblock_migratetype(page
, MIGRATE_ISOLATE
);
4956 move_freepages_block(zone
, page
, MIGRATE_ISOLATE
);
4959 spin_unlock_irqrestore(&zone
->lock
, flags
);
4965 void unset_migratetype_isolate(struct page
*page
)
4968 unsigned long flags
;
4969 zone
= page_zone(page
);
4970 spin_lock_irqsave(&zone
->lock
, flags
);
4971 if (get_pageblock_migratetype(page
) != MIGRATE_ISOLATE
)
4973 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
4974 move_freepages_block(zone
, page
, MIGRATE_MOVABLE
);
4976 spin_unlock_irqrestore(&zone
->lock
, flags
);
4979 #ifdef CONFIG_MEMORY_HOTREMOVE
4981 * All pages in the range must be isolated before calling this.
4984 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
4990 unsigned long flags
;
4991 /* find the first valid pfn */
4992 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
4997 zone
= page_zone(pfn_to_page(pfn
));
4998 spin_lock_irqsave(&zone
->lock
, flags
);
5000 while (pfn
< end_pfn
) {
5001 if (!pfn_valid(pfn
)) {
5005 page
= pfn_to_page(pfn
);
5006 BUG_ON(page_count(page
));
5007 BUG_ON(!PageBuddy(page
));
5008 order
= page_order(page
);
5009 #ifdef CONFIG_DEBUG_VM
5010 printk(KERN_INFO
"remove from free list %lx %d %lx\n",
5011 pfn
, 1 << order
, end_pfn
);
5013 list_del(&page
->lru
);
5014 rmv_page_order(page
);
5015 zone
->free_area
[order
].nr_free
--;
5016 __mod_zone_page_state(zone
, NR_FREE_PAGES
,
5018 for (i
= 0; i
< (1 << order
); i
++)
5019 SetPageReserved((page
+i
));
5020 pfn
+= (1 << order
);
5022 spin_unlock_irqrestore(&zone
->lock
, flags
);