2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/compiler.h>
25 #include <linux/kernel.h>
26 #include <linux/kmemcheck.h>
27 #include <linux/module.h>
28 #include <linux/suspend.h>
29 #include <linux/pagevec.h>
30 #include <linux/blkdev.h>
31 #include <linux/slab.h>
32 #include <linux/oom.h>
33 #include <linux/notifier.h>
34 #include <linux/topology.h>
35 #include <linux/sysctl.h>
36 #include <linux/cpu.h>
37 #include <linux/cpuset.h>
38 #include <linux/memory_hotplug.h>
39 #include <linux/nodemask.h>
40 #include <linux/vmalloc.h>
41 #include <linux/mempolicy.h>
42 #include <linux/stop_machine.h>
43 #include <linux/sort.h>
44 #include <linux/pfn.h>
45 #include <linux/backing-dev.h>
46 #include <linux/fault-inject.h>
47 #include <linux/page-isolation.h>
48 #include <linux/page_cgroup.h>
49 #include <linux/debugobjects.h>
50 #include <linux/kmemleak.h>
51 #include <linux/memory.h>
52 #include <trace/events/kmem.h>
54 #include <asm/tlbflush.h>
55 #include <asm/div64.h>
59 * Array of node states.
61 nodemask_t node_states
[NR_NODE_STATES
] __read_mostly
= {
62 [N_POSSIBLE
] = NODE_MASK_ALL
,
63 [N_ONLINE
] = { { [0] = 1UL } },
65 [N_NORMAL_MEMORY
] = { { [0] = 1UL } },
67 [N_HIGH_MEMORY
] = { { [0] = 1UL } },
69 [N_CPU
] = { { [0] = 1UL } },
72 EXPORT_SYMBOL(node_states
);
74 unsigned long totalram_pages __read_mostly
;
75 unsigned long totalreserve_pages __read_mostly
;
76 int percpu_pagelist_fraction
;
77 gfp_t gfp_allowed_mask __read_mostly
= GFP_BOOT_MASK
;
79 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
80 int pageblock_order __read_mostly
;
83 static void __free_pages_ok(struct page
*page
, unsigned int order
);
86 * results with 256, 32 in the lowmem_reserve sysctl:
87 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
88 * 1G machine -> (16M dma, 784M normal, 224M high)
89 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
90 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
91 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
93 * TBD: should special case ZONE_DMA32 machines here - in those we normally
94 * don't need any ZONE_NORMAL reservation
96 int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
-1] = {
97 #ifdef CONFIG_ZONE_DMA
100 #ifdef CONFIG_ZONE_DMA32
103 #ifdef CONFIG_HIGHMEM
109 EXPORT_SYMBOL(totalram_pages
);
111 static char * const zone_names
[MAX_NR_ZONES
] = {
112 #ifdef CONFIG_ZONE_DMA
115 #ifdef CONFIG_ZONE_DMA32
119 #ifdef CONFIG_HIGHMEM
125 int min_free_kbytes
= 1024;
127 static unsigned long __meminitdata nr_kernel_pages
;
128 static unsigned long __meminitdata nr_all_pages
;
129 static unsigned long __meminitdata dma_reserve
;
131 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
133 * MAX_ACTIVE_REGIONS determines the maximum number of distinct
134 * ranges of memory (RAM) that may be registered with add_active_range().
135 * Ranges passed to add_active_range() will be merged if possible
136 * so the number of times add_active_range() can be called is
137 * related to the number of nodes and the number of holes
139 #ifdef CONFIG_MAX_ACTIVE_REGIONS
140 /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */
141 #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS
143 #if MAX_NUMNODES >= 32
144 /* If there can be many nodes, allow up to 50 holes per node */
145 #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50)
147 /* By default, allow up to 256 distinct regions */
148 #define MAX_ACTIVE_REGIONS 256
152 static struct node_active_region __meminitdata early_node_map
[MAX_ACTIVE_REGIONS
];
153 static int __meminitdata nr_nodemap_entries
;
154 static unsigned long __meminitdata arch_zone_lowest_possible_pfn
[MAX_NR_ZONES
];
155 static unsigned long __meminitdata arch_zone_highest_possible_pfn
[MAX_NR_ZONES
];
156 static unsigned long __initdata required_kernelcore
;
157 static unsigned long __initdata required_movablecore
;
158 static unsigned long __meminitdata zone_movable_pfn
[MAX_NUMNODES
];
160 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
162 EXPORT_SYMBOL(movable_zone
);
163 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
166 int nr_node_ids __read_mostly
= MAX_NUMNODES
;
167 int nr_online_nodes __read_mostly
= 1;
168 EXPORT_SYMBOL(nr_node_ids
);
169 EXPORT_SYMBOL(nr_online_nodes
);
172 int page_group_by_mobility_disabled __read_mostly
;
174 static void set_pageblock_migratetype(struct page
*page
, int migratetype
)
177 if (unlikely(page_group_by_mobility_disabled
))
178 migratetype
= MIGRATE_UNMOVABLE
;
180 set_pageblock_flags_group(page
, (unsigned long)migratetype
,
181 PB_migrate
, PB_migrate_end
);
184 bool oom_killer_disabled __read_mostly
;
186 #ifdef CONFIG_DEBUG_VM
187 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
191 unsigned long pfn
= page_to_pfn(page
);
194 seq
= zone_span_seqbegin(zone
);
195 if (pfn
>= zone
->zone_start_pfn
+ zone
->spanned_pages
)
197 else if (pfn
< zone
->zone_start_pfn
)
199 } while (zone_span_seqretry(zone
, seq
));
204 static int page_is_consistent(struct zone
*zone
, struct page
*page
)
206 if (!pfn_valid_within(page_to_pfn(page
)))
208 if (zone
!= page_zone(page
))
214 * Temporary debugging check for pages not lying within a given zone.
216 static int bad_range(struct zone
*zone
, struct page
*page
)
218 if (page_outside_zone_boundaries(zone
, page
))
220 if (!page_is_consistent(zone
, page
))
226 static inline int bad_range(struct zone
*zone
, struct page
*page
)
232 static void bad_page(struct page
*page
)
234 static unsigned long resume
;
235 static unsigned long nr_shown
;
236 static unsigned long nr_unshown
;
238 /* Don't complain about poisoned pages */
239 if (PageHWPoison(page
)) {
240 __ClearPageBuddy(page
);
245 * Allow a burst of 60 reports, then keep quiet for that minute;
246 * or allow a steady drip of one report per second.
248 if (nr_shown
== 60) {
249 if (time_before(jiffies
, resume
)) {
255 "BUG: Bad page state: %lu messages suppressed\n",
262 resume
= jiffies
+ 60 * HZ
;
264 printk(KERN_ALERT
"BUG: Bad page state in process %s pfn:%05lx\n",
265 current
->comm
, page_to_pfn(page
));
267 "page:%p flags:%p count:%d mapcount:%d mapping:%p index:%lx\n",
268 page
, (void *)page
->flags
, page_count(page
),
269 page_mapcount(page
), page
->mapping
, page
->index
);
273 /* Leave bad fields for debug, except PageBuddy could make trouble */
274 __ClearPageBuddy(page
);
275 add_taint(TAINT_BAD_PAGE
);
279 * Higher-order pages are called "compound pages". They are structured thusly:
281 * The first PAGE_SIZE page is called the "head page".
283 * The remaining PAGE_SIZE pages are called "tail pages".
285 * All pages have PG_compound set. All pages have their ->private pointing at
286 * the head page (even the head page has this).
288 * The first tail page's ->lru.next holds the address of the compound page's
289 * put_page() function. Its ->lru.prev holds the order of allocation.
290 * This usage means that zero-order pages may not be compound.
293 static void free_compound_page(struct page
*page
)
295 __free_pages_ok(page
, compound_order(page
));
298 void prep_compound_page(struct page
*page
, unsigned long order
)
301 int nr_pages
= 1 << order
;
303 set_compound_page_dtor(page
, free_compound_page
);
304 set_compound_order(page
, order
);
306 for (i
= 1; i
< nr_pages
; i
++) {
307 struct page
*p
= page
+ i
;
310 p
->first_page
= page
;
314 static int destroy_compound_page(struct page
*page
, unsigned long order
)
317 int nr_pages
= 1 << order
;
320 if (unlikely(compound_order(page
) != order
) ||
321 unlikely(!PageHead(page
))) {
326 __ClearPageHead(page
);
328 for (i
= 1; i
< nr_pages
; i
++) {
329 struct page
*p
= page
+ i
;
331 if (unlikely(!PageTail(p
) || (p
->first_page
!= page
))) {
341 static inline void prep_zero_page(struct page
*page
, int order
, gfp_t gfp_flags
)
346 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
347 * and __GFP_HIGHMEM from hard or soft interrupt context.
349 VM_BUG_ON((gfp_flags
& __GFP_HIGHMEM
) && in_interrupt());
350 for (i
= 0; i
< (1 << order
); i
++)
351 clear_highpage(page
+ i
);
354 static inline void set_page_order(struct page
*page
, int order
)
356 set_page_private(page
, order
);
357 __SetPageBuddy(page
);
360 static inline void rmv_page_order(struct page
*page
)
362 __ClearPageBuddy(page
);
363 set_page_private(page
, 0);
367 * Locate the struct page for both the matching buddy in our
368 * pair (buddy1) and the combined O(n+1) page they form (page).
370 * 1) Any buddy B1 will have an order O twin B2 which satisfies
371 * the following equation:
373 * For example, if the starting buddy (buddy2) is #8 its order
375 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
377 * 2) Any buddy B will have an order O+1 parent P which
378 * satisfies the following equation:
381 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
383 static inline struct page
*
384 __page_find_buddy(struct page
*page
, unsigned long page_idx
, unsigned int order
)
386 unsigned long buddy_idx
= page_idx
^ (1 << order
);
388 return page
+ (buddy_idx
- page_idx
);
391 static inline unsigned long
392 __find_combined_index(unsigned long page_idx
, unsigned int order
)
394 return (page_idx
& ~(1 << order
));
398 * This function checks whether a page is free && is the buddy
399 * we can do coalesce a page and its buddy if
400 * (a) the buddy is not in a hole &&
401 * (b) the buddy is in the buddy system &&
402 * (c) a page and its buddy have the same order &&
403 * (d) a page and its buddy are in the same zone.
405 * For recording whether a page is in the buddy system, we use PG_buddy.
406 * Setting, clearing, and testing PG_buddy is serialized by zone->lock.
408 * For recording page's order, we use page_private(page).
410 static inline int page_is_buddy(struct page
*page
, struct page
*buddy
,
413 if (!pfn_valid_within(page_to_pfn(buddy
)))
416 if (page_zone_id(page
) != page_zone_id(buddy
))
419 if (PageBuddy(buddy
) && page_order(buddy
) == order
) {
420 VM_BUG_ON(page_count(buddy
) != 0);
427 * Freeing function for a buddy system allocator.
429 * The concept of a buddy system is to maintain direct-mapped table
430 * (containing bit values) for memory blocks of various "orders".
431 * The bottom level table contains the map for the smallest allocatable
432 * units of memory (here, pages), and each level above it describes
433 * pairs of units from the levels below, hence, "buddies".
434 * At a high level, all that happens here is marking the table entry
435 * at the bottom level available, and propagating the changes upward
436 * as necessary, plus some accounting needed to play nicely with other
437 * parts of the VM system.
438 * At each level, we keep a list of pages, which are heads of continuous
439 * free pages of length of (1 << order) and marked with PG_buddy. Page's
440 * order is recorded in page_private(page) field.
441 * So when we are allocating or freeing one, we can derive the state of the
442 * other. That is, if we allocate a small block, and both were
443 * free, the remainder of the region must be split into blocks.
444 * If a block is freed, and its buddy is also free, then this
445 * triggers coalescing into a block of larger size.
450 static inline void __free_one_page(struct page
*page
,
451 struct zone
*zone
, unsigned int order
,
454 unsigned long page_idx
;
456 if (unlikely(PageCompound(page
)))
457 if (unlikely(destroy_compound_page(page
, order
)))
460 VM_BUG_ON(migratetype
== -1);
462 page_idx
= page_to_pfn(page
) & ((1 << MAX_ORDER
) - 1);
464 VM_BUG_ON(page_idx
& ((1 << order
) - 1));
465 VM_BUG_ON(bad_range(zone
, page
));
467 while (order
< MAX_ORDER
-1) {
468 unsigned long combined_idx
;
471 buddy
= __page_find_buddy(page
, page_idx
, order
);
472 if (!page_is_buddy(page
, buddy
, order
))
475 /* Our buddy is free, merge with it and move up one order. */
476 list_del(&buddy
->lru
);
477 zone
->free_area
[order
].nr_free
--;
478 rmv_page_order(buddy
);
479 combined_idx
= __find_combined_index(page_idx
, order
);
480 page
= page
+ (combined_idx
- page_idx
);
481 page_idx
= combined_idx
;
484 set_page_order(page
, order
);
486 &zone
->free_area
[order
].free_list
[migratetype
]);
487 zone
->free_area
[order
].nr_free
++;
491 * free_page_mlock() -- clean up attempts to free and mlocked() page.
492 * Page should not be on lru, so no need to fix that up.
493 * free_pages_check() will verify...
495 static inline void free_page_mlock(struct page
*page
)
497 __dec_zone_page_state(page
, NR_MLOCK
);
498 __count_vm_event(UNEVICTABLE_MLOCKFREED
);
501 static inline int free_pages_check(struct page
*page
)
503 if (unlikely(page_mapcount(page
) |
504 (page
->mapping
!= NULL
) |
505 (atomic_read(&page
->_count
) != 0) |
506 (page
->flags
& PAGE_FLAGS_CHECK_AT_FREE
))) {
510 if (page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)
511 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
516 * Frees a number of pages from the PCP lists
517 * Assumes all pages on list are in same zone, and of same order.
518 * count is the number of pages to free.
520 * If the zone was previously in an "all pages pinned" state then look to
521 * see if this freeing clears that state.
523 * And clear the zone's pages_scanned counter, to hold off the "all pages are
524 * pinned" detection logic.
526 static void free_pcppages_bulk(struct zone
*zone
, int count
,
527 struct per_cpu_pages
*pcp
)
532 spin_lock(&zone
->lock
);
533 zone_clear_flag(zone
, ZONE_ALL_UNRECLAIMABLE
);
534 zone
->pages_scanned
= 0;
536 __mod_zone_page_state(zone
, NR_FREE_PAGES
, count
);
539 struct list_head
*list
;
542 * Remove pages from lists in a round-robin fashion. A
543 * batch_free count is maintained that is incremented when an
544 * empty list is encountered. This is so more pages are freed
545 * off fuller lists instead of spinning excessively around empty
550 if (++migratetype
== MIGRATE_PCPTYPES
)
552 list
= &pcp
->lists
[migratetype
];
553 } while (list_empty(list
));
556 page
= list_entry(list
->prev
, struct page
, lru
);
557 /* must delete as __free_one_page list manipulates */
558 list_del(&page
->lru
);
559 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
560 __free_one_page(page
, zone
, 0, page_private(page
));
561 trace_mm_page_pcpu_drain(page
, 0, page_private(page
));
562 } while (--count
&& --batch_free
&& !list_empty(list
));
564 spin_unlock(&zone
->lock
);
567 static void free_one_page(struct zone
*zone
, struct page
*page
, int order
,
570 spin_lock(&zone
->lock
);
571 zone_clear_flag(zone
, ZONE_ALL_UNRECLAIMABLE
);
572 zone
->pages_scanned
= 0;
574 __mod_zone_page_state(zone
, NR_FREE_PAGES
, 1 << order
);
575 __free_one_page(page
, zone
, order
, migratetype
);
576 spin_unlock(&zone
->lock
);
579 static void __free_pages_ok(struct page
*page
, unsigned int order
)
584 int wasMlocked
= __TestClearPageMlocked(page
);
586 kmemcheck_free_shadow(page
, order
);
588 for (i
= 0 ; i
< (1 << order
) ; ++i
)
589 bad
+= free_pages_check(page
+ i
);
593 if (!PageHighMem(page
)) {
594 debug_check_no_locks_freed(page_address(page
),PAGE_SIZE
<<order
);
595 debug_check_no_obj_freed(page_address(page
),
598 arch_free_page(page
, order
);
599 kernel_map_pages(page
, 1 << order
, 0);
601 local_irq_save(flags
);
602 if (unlikely(wasMlocked
))
603 free_page_mlock(page
);
604 __count_vm_events(PGFREE
, 1 << order
);
605 free_one_page(page_zone(page
), page
, order
,
606 get_pageblock_migratetype(page
));
607 local_irq_restore(flags
);
611 * permit the bootmem allocator to evade page validation on high-order frees
613 void __meminit
__free_pages_bootmem(struct page
*page
, unsigned int order
)
616 __ClearPageReserved(page
);
617 set_page_count(page
, 0);
618 set_page_refcounted(page
);
624 for (loop
= 0; loop
< BITS_PER_LONG
; loop
++) {
625 struct page
*p
= &page
[loop
];
627 if (loop
+ 1 < BITS_PER_LONG
)
629 __ClearPageReserved(p
);
630 set_page_count(p
, 0);
633 set_page_refcounted(page
);
634 __free_pages(page
, order
);
640 * The order of subdivision here is critical for the IO subsystem.
641 * Please do not alter this order without good reasons and regression
642 * testing. Specifically, as large blocks of memory are subdivided,
643 * the order in which smaller blocks are delivered depends on the order
644 * they're subdivided in this function. This is the primary factor
645 * influencing the order in which pages are delivered to the IO
646 * subsystem according to empirical testing, and this is also justified
647 * by considering the behavior of a buddy system containing a single
648 * large block of memory acted on by a series of small allocations.
649 * This behavior is a critical factor in sglist merging's success.
653 static inline void expand(struct zone
*zone
, struct page
*page
,
654 int low
, int high
, struct free_area
*area
,
657 unsigned long size
= 1 << high
;
663 VM_BUG_ON(bad_range(zone
, &page
[size
]));
664 list_add(&page
[size
].lru
, &area
->free_list
[migratetype
]);
666 set_page_order(&page
[size
], high
);
671 * This page is about to be returned from the page allocator
673 static inline int check_new_page(struct page
*page
)
675 if (unlikely(page_mapcount(page
) |
676 (page
->mapping
!= NULL
) |
677 (atomic_read(&page
->_count
) != 0) |
678 (page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
))) {
685 static int prep_new_page(struct page
*page
, int order
, gfp_t gfp_flags
)
689 for (i
= 0; i
< (1 << order
); i
++) {
690 struct page
*p
= page
+ i
;
691 if (unlikely(check_new_page(p
)))
695 set_page_private(page
, 0);
696 set_page_refcounted(page
);
698 arch_alloc_page(page
, order
);
699 kernel_map_pages(page
, 1 << order
, 1);
701 if (gfp_flags
& __GFP_ZERO
)
702 prep_zero_page(page
, order
, gfp_flags
);
704 if (order
&& (gfp_flags
& __GFP_COMP
))
705 prep_compound_page(page
, order
);
711 * Go through the free lists for the given migratetype and remove
712 * the smallest available page from the freelists
715 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
718 unsigned int current_order
;
719 struct free_area
* area
;
722 /* Find a page of the appropriate size in the preferred list */
723 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
724 area
= &(zone
->free_area
[current_order
]);
725 if (list_empty(&area
->free_list
[migratetype
]))
728 page
= list_entry(area
->free_list
[migratetype
].next
,
730 list_del(&page
->lru
);
731 rmv_page_order(page
);
733 expand(zone
, page
, order
, current_order
, area
, migratetype
);
742 * This array describes the order lists are fallen back to when
743 * the free lists for the desirable migrate type are depleted
745 static int fallbacks
[MIGRATE_TYPES
][MIGRATE_TYPES
-1] = {
746 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
747 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
748 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_RESERVE
},
749 [MIGRATE_RESERVE
] = { MIGRATE_RESERVE
, MIGRATE_RESERVE
, MIGRATE_RESERVE
}, /* Never used */
753 * Move the free pages in a range to the free lists of the requested type.
754 * Note that start_page and end_pages are not aligned on a pageblock
755 * boundary. If alignment is required, use move_freepages_block()
757 static int move_freepages(struct zone
*zone
,
758 struct page
*start_page
, struct page
*end_page
,
765 #ifndef CONFIG_HOLES_IN_ZONE
767 * page_zone is not safe to call in this context when
768 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
769 * anyway as we check zone boundaries in move_freepages_block().
770 * Remove at a later date when no bug reports exist related to
771 * grouping pages by mobility
773 BUG_ON(page_zone(start_page
) != page_zone(end_page
));
776 for (page
= start_page
; page
<= end_page
;) {
777 /* Make sure we are not inadvertently changing nodes */
778 VM_BUG_ON(page_to_nid(page
) != zone_to_nid(zone
));
780 if (!pfn_valid_within(page_to_pfn(page
))) {
785 if (!PageBuddy(page
)) {
790 order
= page_order(page
);
791 list_del(&page
->lru
);
793 &zone
->free_area
[order
].free_list
[migratetype
]);
795 pages_moved
+= 1 << order
;
801 static int move_freepages_block(struct zone
*zone
, struct page
*page
,
804 unsigned long start_pfn
, end_pfn
;
805 struct page
*start_page
, *end_page
;
807 start_pfn
= page_to_pfn(page
);
808 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
809 start_page
= pfn_to_page(start_pfn
);
810 end_page
= start_page
+ pageblock_nr_pages
- 1;
811 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
813 /* Do not cross zone boundaries */
814 if (start_pfn
< zone
->zone_start_pfn
)
816 if (end_pfn
>= zone
->zone_start_pfn
+ zone
->spanned_pages
)
819 return move_freepages(zone
, start_page
, end_page
, migratetype
);
822 static void change_pageblock_range(struct page
*pageblock_page
,
823 int start_order
, int migratetype
)
825 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
827 while (nr_pageblocks
--) {
828 set_pageblock_migratetype(pageblock_page
, migratetype
);
829 pageblock_page
+= pageblock_nr_pages
;
833 /* Remove an element from the buddy allocator from the fallback list */
834 static inline struct page
*
835 __rmqueue_fallback(struct zone
*zone
, int order
, int start_migratetype
)
837 struct free_area
* area
;
842 /* Find the largest possible block of pages in the other list */
843 for (current_order
= MAX_ORDER
-1; current_order
>= order
;
845 for (i
= 0; i
< MIGRATE_TYPES
- 1; i
++) {
846 migratetype
= fallbacks
[start_migratetype
][i
];
848 /* MIGRATE_RESERVE handled later if necessary */
849 if (migratetype
== MIGRATE_RESERVE
)
852 area
= &(zone
->free_area
[current_order
]);
853 if (list_empty(&area
->free_list
[migratetype
]))
856 page
= list_entry(area
->free_list
[migratetype
].next
,
861 * If breaking a large block of pages, move all free
862 * pages to the preferred allocation list. If falling
863 * back for a reclaimable kernel allocation, be more
864 * agressive about taking ownership of free pages
866 if (unlikely(current_order
>= (pageblock_order
>> 1)) ||
867 start_migratetype
== MIGRATE_RECLAIMABLE
||
868 page_group_by_mobility_disabled
) {
870 pages
= move_freepages_block(zone
, page
,
873 /* Claim the whole block if over half of it is free */
874 if (pages
>= (1 << (pageblock_order
-1)) ||
875 page_group_by_mobility_disabled
)
876 set_pageblock_migratetype(page
,
879 migratetype
= start_migratetype
;
882 /* Remove the page from the freelists */
883 list_del(&page
->lru
);
884 rmv_page_order(page
);
886 /* Take ownership for orders >= pageblock_order */
887 if (current_order
>= pageblock_order
)
888 change_pageblock_range(page
, current_order
,
891 expand(zone
, page
, order
, current_order
, area
, migratetype
);
893 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
894 start_migratetype
, migratetype
);
904 * Do the hard work of removing an element from the buddy allocator.
905 * Call me with the zone->lock already held.
907 static struct page
*__rmqueue(struct zone
*zone
, unsigned int order
,
913 page
= __rmqueue_smallest(zone
, order
, migratetype
);
915 if (unlikely(!page
) && migratetype
!= MIGRATE_RESERVE
) {
916 page
= __rmqueue_fallback(zone
, order
, migratetype
);
919 * Use MIGRATE_RESERVE rather than fail an allocation. goto
920 * is used because __rmqueue_smallest is an inline function
921 * and we want just one call site
924 migratetype
= MIGRATE_RESERVE
;
929 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
934 * Obtain a specified number of elements from the buddy allocator, all under
935 * a single hold of the lock, for efficiency. Add them to the supplied list.
936 * Returns the number of new pages which were placed at *list.
938 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
939 unsigned long count
, struct list_head
*list
,
940 int migratetype
, int cold
)
944 spin_lock(&zone
->lock
);
945 for (i
= 0; i
< count
; ++i
) {
946 struct page
*page
= __rmqueue(zone
, order
, migratetype
);
947 if (unlikely(page
== NULL
))
951 * Split buddy pages returned by expand() are received here
952 * in physical page order. The page is added to the callers and
953 * list and the list head then moves forward. From the callers
954 * perspective, the linked list is ordered by page number in
955 * some conditions. This is useful for IO devices that can
956 * merge IO requests if the physical pages are ordered
959 if (likely(cold
== 0))
960 list_add(&page
->lru
, list
);
962 list_add_tail(&page
->lru
, list
);
963 set_page_private(page
, migratetype
);
966 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
967 spin_unlock(&zone
->lock
);
973 * Called from the vmstat counter updater to drain pagesets of this
974 * currently executing processor on remote nodes after they have
977 * Note that this function must be called with the thread pinned to
978 * a single processor.
980 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
985 local_irq_save(flags
);
986 if (pcp
->count
>= pcp
->batch
)
987 to_drain
= pcp
->batch
;
989 to_drain
= pcp
->count
;
990 free_pcppages_bulk(zone
, to_drain
, pcp
);
991 pcp
->count
-= to_drain
;
992 local_irq_restore(flags
);
997 * Drain pages of the indicated processor.
999 * The processor must either be the current processor and the
1000 * thread pinned to the current processor or a processor that
1003 static void drain_pages(unsigned int cpu
)
1005 unsigned long flags
;
1008 for_each_populated_zone(zone
) {
1009 struct per_cpu_pageset
*pset
;
1010 struct per_cpu_pages
*pcp
;
1012 local_irq_save(flags
);
1013 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
1016 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
1018 local_irq_restore(flags
);
1023 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1025 void drain_local_pages(void *arg
)
1027 drain_pages(smp_processor_id());
1031 * Spill all the per-cpu pages from all CPUs back into the buddy allocator
1033 void drain_all_pages(void)
1035 on_each_cpu(drain_local_pages
, NULL
, 1);
1038 #ifdef CONFIG_HIBERNATION
1040 void mark_free_pages(struct zone
*zone
)
1042 unsigned long pfn
, max_zone_pfn
;
1043 unsigned long flags
;
1045 struct list_head
*curr
;
1047 if (!zone
->spanned_pages
)
1050 spin_lock_irqsave(&zone
->lock
, flags
);
1052 max_zone_pfn
= zone
->zone_start_pfn
+ zone
->spanned_pages
;
1053 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
1054 if (pfn_valid(pfn
)) {
1055 struct page
*page
= pfn_to_page(pfn
);
1057 if (!swsusp_page_is_forbidden(page
))
1058 swsusp_unset_page_free(page
);
1061 for_each_migratetype_order(order
, t
) {
1062 list_for_each(curr
, &zone
->free_area
[order
].free_list
[t
]) {
1065 pfn
= page_to_pfn(list_entry(curr
, struct page
, lru
));
1066 for (i
= 0; i
< (1UL << order
); i
++)
1067 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
1070 spin_unlock_irqrestore(&zone
->lock
, flags
);
1072 #endif /* CONFIG_PM */
1075 * Free a 0-order page
1077 static void free_hot_cold_page(struct page
*page
, int cold
)
1079 struct zone
*zone
= page_zone(page
);
1080 struct per_cpu_pages
*pcp
;
1081 unsigned long flags
;
1083 int wasMlocked
= __TestClearPageMlocked(page
);
1085 kmemcheck_free_shadow(page
, 0);
1088 page
->mapping
= NULL
;
1089 if (free_pages_check(page
))
1092 if (!PageHighMem(page
)) {
1093 debug_check_no_locks_freed(page_address(page
), PAGE_SIZE
);
1094 debug_check_no_obj_freed(page_address(page
), PAGE_SIZE
);
1096 arch_free_page(page
, 0);
1097 kernel_map_pages(page
, 1, 0);
1099 migratetype
= get_pageblock_migratetype(page
);
1100 set_page_private(page
, migratetype
);
1101 local_irq_save(flags
);
1102 if (unlikely(wasMlocked
))
1103 free_page_mlock(page
);
1104 __count_vm_event(PGFREE
);
1107 * We only track unmovable, reclaimable and movable on pcp lists.
1108 * Free ISOLATE pages back to the allocator because they are being
1109 * offlined but treat RESERVE as movable pages so we can get those
1110 * areas back if necessary. Otherwise, we may have to free
1111 * excessively into the page allocator
1113 if (migratetype
>= MIGRATE_PCPTYPES
) {
1114 if (unlikely(migratetype
== MIGRATE_ISOLATE
)) {
1115 free_one_page(zone
, page
, 0, migratetype
);
1118 migratetype
= MIGRATE_MOVABLE
;
1121 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
1123 list_add_tail(&page
->lru
, &pcp
->lists
[migratetype
]);
1125 list_add(&page
->lru
, &pcp
->lists
[migratetype
]);
1127 if (pcp
->count
>= pcp
->high
) {
1128 free_pcppages_bulk(zone
, pcp
->batch
, pcp
);
1129 pcp
->count
-= pcp
->batch
;
1133 local_irq_restore(flags
);
1136 void free_hot_page(struct page
*page
)
1138 trace_mm_page_free_direct(page
, 0);
1139 free_hot_cold_page(page
, 0);
1143 * split_page takes a non-compound higher-order page, and splits it into
1144 * n (1<<order) sub-pages: page[0..n]
1145 * Each sub-page must be freed individually.
1147 * Note: this is probably too low level an operation for use in drivers.
1148 * Please consult with lkml before using this in your driver.
1150 void split_page(struct page
*page
, unsigned int order
)
1154 VM_BUG_ON(PageCompound(page
));
1155 VM_BUG_ON(!page_count(page
));
1157 #ifdef CONFIG_KMEMCHECK
1159 * Split shadow pages too, because free(page[0]) would
1160 * otherwise free the whole shadow.
1162 if (kmemcheck_page_is_tracked(page
))
1163 split_page(virt_to_page(page
[0].shadow
), order
);
1166 for (i
= 1; i
< (1 << order
); i
++)
1167 set_page_refcounted(page
+ i
);
1171 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1172 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1176 struct page
*buffered_rmqueue(struct zone
*preferred_zone
,
1177 struct zone
*zone
, int order
, gfp_t gfp_flags
,
1180 unsigned long flags
;
1182 int cold
= !!(gfp_flags
& __GFP_COLD
);
1185 if (likely(order
== 0)) {
1186 struct per_cpu_pages
*pcp
;
1187 struct list_head
*list
;
1189 local_irq_save(flags
);
1190 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
1191 list
= &pcp
->lists
[migratetype
];
1192 if (list_empty(list
)) {
1193 pcp
->count
+= rmqueue_bulk(zone
, 0,
1196 if (unlikely(list_empty(list
)))
1201 page
= list_entry(list
->prev
, struct page
, lru
);
1203 page
= list_entry(list
->next
, struct page
, lru
);
1205 list_del(&page
->lru
);
1208 if (unlikely(gfp_flags
& __GFP_NOFAIL
)) {
1210 * __GFP_NOFAIL is not to be used in new code.
1212 * All __GFP_NOFAIL callers should be fixed so that they
1213 * properly detect and handle allocation failures.
1215 * We most definitely don't want callers attempting to
1216 * allocate greater than order-1 page units with
1219 WARN_ON_ONCE(order
> 1);
1221 spin_lock_irqsave(&zone
->lock
, flags
);
1222 page
= __rmqueue(zone
, order
, migratetype
);
1223 spin_unlock(&zone
->lock
);
1226 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(1 << order
));
1229 __count_zone_vm_events(PGALLOC
, zone
, 1 << order
);
1230 zone_statistics(preferred_zone
, zone
);
1231 local_irq_restore(flags
);
1233 VM_BUG_ON(bad_range(zone
, page
));
1234 if (prep_new_page(page
, order
, gfp_flags
))
1239 local_irq_restore(flags
);
1243 /* The ALLOC_WMARK bits are used as an index to zone->watermark */
1244 #define ALLOC_WMARK_MIN WMARK_MIN
1245 #define ALLOC_WMARK_LOW WMARK_LOW
1246 #define ALLOC_WMARK_HIGH WMARK_HIGH
1247 #define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */
1249 /* Mask to get the watermark bits */
1250 #define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1)
1252 #define ALLOC_HARDER 0x10 /* try to alloc harder */
1253 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
1254 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
1256 #ifdef CONFIG_FAIL_PAGE_ALLOC
1258 static struct fail_page_alloc_attr
{
1259 struct fault_attr attr
;
1261 u32 ignore_gfp_highmem
;
1262 u32 ignore_gfp_wait
;
1265 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1267 struct dentry
*ignore_gfp_highmem_file
;
1268 struct dentry
*ignore_gfp_wait_file
;
1269 struct dentry
*min_order_file
;
1271 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1273 } fail_page_alloc
= {
1274 .attr
= FAULT_ATTR_INITIALIZER
,
1275 .ignore_gfp_wait
= 1,
1276 .ignore_gfp_highmem
= 1,
1280 static int __init
setup_fail_page_alloc(char *str
)
1282 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
1284 __setup("fail_page_alloc=", setup_fail_page_alloc
);
1286 static int should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1288 if (order
< fail_page_alloc
.min_order
)
1290 if (gfp_mask
& __GFP_NOFAIL
)
1292 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
1294 if (fail_page_alloc
.ignore_gfp_wait
&& (gfp_mask
& __GFP_WAIT
))
1297 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
1300 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1302 static int __init
fail_page_alloc_debugfs(void)
1304 mode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
1308 err
= init_fault_attr_dentries(&fail_page_alloc
.attr
,
1312 dir
= fail_page_alloc
.attr
.dentries
.dir
;
1314 fail_page_alloc
.ignore_gfp_wait_file
=
1315 debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
1316 &fail_page_alloc
.ignore_gfp_wait
);
1318 fail_page_alloc
.ignore_gfp_highmem_file
=
1319 debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
1320 &fail_page_alloc
.ignore_gfp_highmem
);
1321 fail_page_alloc
.min_order_file
=
1322 debugfs_create_u32("min-order", mode
, dir
,
1323 &fail_page_alloc
.min_order
);
1325 if (!fail_page_alloc
.ignore_gfp_wait_file
||
1326 !fail_page_alloc
.ignore_gfp_highmem_file
||
1327 !fail_page_alloc
.min_order_file
) {
1329 debugfs_remove(fail_page_alloc
.ignore_gfp_wait_file
);
1330 debugfs_remove(fail_page_alloc
.ignore_gfp_highmem_file
);
1331 debugfs_remove(fail_page_alloc
.min_order_file
);
1332 cleanup_fault_attr_dentries(&fail_page_alloc
.attr
);
1338 late_initcall(fail_page_alloc_debugfs
);
1340 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1342 #else /* CONFIG_FAIL_PAGE_ALLOC */
1344 static inline int should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1349 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1352 * Return 1 if free pages are above 'mark'. This takes into account the order
1353 * of the allocation.
1355 int zone_watermark_ok(struct zone
*z
, int order
, unsigned long mark
,
1356 int classzone_idx
, int alloc_flags
)
1358 /* free_pages my go negative - that's OK */
1360 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
) - (1 << order
) + 1;
1363 if (alloc_flags
& ALLOC_HIGH
)
1365 if (alloc_flags
& ALLOC_HARDER
)
1368 if (free_pages
<= min
+ z
->lowmem_reserve
[classzone_idx
])
1370 for (o
= 0; o
< order
; o
++) {
1371 /* At the next order, this order's pages become unavailable */
1372 free_pages
-= z
->free_area
[o
].nr_free
<< o
;
1374 /* Require fewer higher order pages to be free */
1377 if (free_pages
<= min
)
1385 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1386 * skip over zones that are not allowed by the cpuset, or that have
1387 * been recently (in last second) found to be nearly full. See further
1388 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1389 * that have to skip over a lot of full or unallowed zones.
1391 * If the zonelist cache is present in the passed in zonelist, then
1392 * returns a pointer to the allowed node mask (either the current
1393 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1395 * If the zonelist cache is not available for this zonelist, does
1396 * nothing and returns NULL.
1398 * If the fullzones BITMAP in the zonelist cache is stale (more than
1399 * a second since last zap'd) then we zap it out (clear its bits.)
1401 * We hold off even calling zlc_setup, until after we've checked the
1402 * first zone in the zonelist, on the theory that most allocations will
1403 * be satisfied from that first zone, so best to examine that zone as
1404 * quickly as we can.
1406 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1408 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1409 nodemask_t
*allowednodes
; /* zonelist_cache approximation */
1411 zlc
= zonelist
->zlcache_ptr
;
1415 if (time_after(jiffies
, zlc
->last_full_zap
+ HZ
)) {
1416 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
1417 zlc
->last_full_zap
= jiffies
;
1420 allowednodes
= !in_interrupt() && (alloc_flags
& ALLOC_CPUSET
) ?
1421 &cpuset_current_mems_allowed
:
1422 &node_states
[N_HIGH_MEMORY
];
1423 return allowednodes
;
1427 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1428 * if it is worth looking at further for free memory:
1429 * 1) Check that the zone isn't thought to be full (doesn't have its
1430 * bit set in the zonelist_cache fullzones BITMAP).
1431 * 2) Check that the zones node (obtained from the zonelist_cache
1432 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1433 * Return true (non-zero) if zone is worth looking at further, or
1434 * else return false (zero) if it is not.
1436 * This check -ignores- the distinction between various watermarks,
1437 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1438 * found to be full for any variation of these watermarks, it will
1439 * be considered full for up to one second by all requests, unless
1440 * we are so low on memory on all allowed nodes that we are forced
1441 * into the second scan of the zonelist.
1443 * In the second scan we ignore this zonelist cache and exactly
1444 * apply the watermarks to all zones, even it is slower to do so.
1445 * We are low on memory in the second scan, and should leave no stone
1446 * unturned looking for a free page.
1448 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1449 nodemask_t
*allowednodes
)
1451 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1452 int i
; /* index of *z in zonelist zones */
1453 int n
; /* node that zone *z is on */
1455 zlc
= zonelist
->zlcache_ptr
;
1459 i
= z
- zonelist
->_zonerefs
;
1462 /* This zone is worth trying if it is allowed but not full */
1463 return node_isset(n
, *allowednodes
) && !test_bit(i
, zlc
->fullzones
);
1467 * Given 'z' scanning a zonelist, set the corresponding bit in
1468 * zlc->fullzones, so that subsequent attempts to allocate a page
1469 * from that zone don't waste time re-examining it.
1471 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1473 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1474 int i
; /* index of *z in zonelist zones */
1476 zlc
= zonelist
->zlcache_ptr
;
1480 i
= z
- zonelist
->_zonerefs
;
1482 set_bit(i
, zlc
->fullzones
);
1485 #else /* CONFIG_NUMA */
1487 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1492 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1493 nodemask_t
*allowednodes
)
1498 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1501 #endif /* CONFIG_NUMA */
1504 * get_page_from_freelist goes through the zonelist trying to allocate
1507 static struct page
*
1508 get_page_from_freelist(gfp_t gfp_mask
, nodemask_t
*nodemask
, unsigned int order
,
1509 struct zonelist
*zonelist
, int high_zoneidx
, int alloc_flags
,
1510 struct zone
*preferred_zone
, int migratetype
)
1513 struct page
*page
= NULL
;
1516 nodemask_t
*allowednodes
= NULL
;/* zonelist_cache approximation */
1517 int zlc_active
= 0; /* set if using zonelist_cache */
1518 int did_zlc_setup
= 0; /* just call zlc_setup() one time */
1520 classzone_idx
= zone_idx(preferred_zone
);
1523 * Scan zonelist, looking for a zone with enough free.
1524 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1526 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
1527 high_zoneidx
, nodemask
) {
1528 if (NUMA_BUILD
&& zlc_active
&&
1529 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
1531 if ((alloc_flags
& ALLOC_CPUSET
) &&
1532 !cpuset_zone_allowed_softwall(zone
, gfp_mask
))
1535 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
1536 if (!(alloc_flags
& ALLOC_NO_WATERMARKS
)) {
1540 mark
= zone
->watermark
[alloc_flags
& ALLOC_WMARK_MASK
];
1541 if (zone_watermark_ok(zone
, order
, mark
,
1542 classzone_idx
, alloc_flags
))
1545 if (zone_reclaim_mode
== 0)
1546 goto this_zone_full
;
1548 ret
= zone_reclaim(zone
, gfp_mask
, order
);
1550 case ZONE_RECLAIM_NOSCAN
:
1553 case ZONE_RECLAIM_FULL
:
1554 /* scanned but unreclaimable */
1555 goto this_zone_full
;
1557 /* did we reclaim enough */
1558 if (!zone_watermark_ok(zone
, order
, mark
,
1559 classzone_idx
, alloc_flags
))
1560 goto this_zone_full
;
1565 page
= buffered_rmqueue(preferred_zone
, zone
, order
,
1566 gfp_mask
, migratetype
);
1571 zlc_mark_zone_full(zonelist
, z
);
1573 if (NUMA_BUILD
&& !did_zlc_setup
&& nr_online_nodes
> 1) {
1575 * we do zlc_setup after the first zone is tried but only
1576 * if there are multiple nodes make it worthwhile
1578 allowednodes
= zlc_setup(zonelist
, alloc_flags
);
1584 if (unlikely(NUMA_BUILD
&& page
== NULL
&& zlc_active
)) {
1585 /* Disable zlc cache for second zonelist scan */
1593 should_alloc_retry(gfp_t gfp_mask
, unsigned int order
,
1594 unsigned long pages_reclaimed
)
1596 /* Do not loop if specifically requested */
1597 if (gfp_mask
& __GFP_NORETRY
)
1601 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
1602 * means __GFP_NOFAIL, but that may not be true in other
1605 if (order
<= PAGE_ALLOC_COSTLY_ORDER
)
1609 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
1610 * specified, then we retry until we no longer reclaim any pages
1611 * (above), or we've reclaimed an order of pages at least as
1612 * large as the allocation's order. In both cases, if the
1613 * allocation still fails, we stop retrying.
1615 if (gfp_mask
& __GFP_REPEAT
&& pages_reclaimed
< (1 << order
))
1619 * Don't let big-order allocations loop unless the caller
1620 * explicitly requests that.
1622 if (gfp_mask
& __GFP_NOFAIL
)
1628 static inline struct page
*
1629 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
1630 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1631 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
1636 /* Acquire the OOM killer lock for the zones in zonelist */
1637 if (!try_set_zone_oom(zonelist
, gfp_mask
)) {
1638 schedule_timeout_uninterruptible(1);
1643 * Go through the zonelist yet one more time, keep very high watermark
1644 * here, this is only to catch a parallel oom killing, we must fail if
1645 * we're still under heavy pressure.
1647 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
,
1648 order
, zonelist
, high_zoneidx
,
1649 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
,
1650 preferred_zone
, migratetype
);
1654 if (!(gfp_mask
& __GFP_NOFAIL
)) {
1655 /* The OOM killer will not help higher order allocs */
1656 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
1659 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
1660 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
1661 * The caller should handle page allocation failure by itself if
1662 * it specifies __GFP_THISNODE.
1663 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
1665 if (gfp_mask
& __GFP_THISNODE
)
1668 /* Exhausted what can be done so it's blamo time */
1669 out_of_memory(zonelist
, gfp_mask
, order
, nodemask
);
1672 clear_zonelist_oom(zonelist
, gfp_mask
);
1676 /* The really slow allocator path where we enter direct reclaim */
1677 static inline struct page
*
1678 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
1679 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1680 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
1681 int migratetype
, unsigned long *did_some_progress
)
1683 struct page
*page
= NULL
;
1684 struct reclaim_state reclaim_state
;
1685 struct task_struct
*p
= current
;
1689 /* We now go into synchronous reclaim */
1690 cpuset_memory_pressure_bump();
1691 p
->flags
|= PF_MEMALLOC
;
1692 lockdep_set_current_reclaim_state(gfp_mask
);
1693 reclaim_state
.reclaimed_slab
= 0;
1694 p
->reclaim_state
= &reclaim_state
;
1696 *did_some_progress
= try_to_free_pages(zonelist
, order
, gfp_mask
, nodemask
);
1698 p
->reclaim_state
= NULL
;
1699 lockdep_clear_current_reclaim_state();
1700 p
->flags
&= ~PF_MEMALLOC
;
1707 if (likely(*did_some_progress
))
1708 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
1709 zonelist
, high_zoneidx
,
1710 alloc_flags
, preferred_zone
,
1716 * This is called in the allocator slow-path if the allocation request is of
1717 * sufficient urgency to ignore watermarks and take other desperate measures
1719 static inline struct page
*
1720 __alloc_pages_high_priority(gfp_t gfp_mask
, unsigned int order
,
1721 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1722 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
1728 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
1729 zonelist
, high_zoneidx
, ALLOC_NO_WATERMARKS
,
1730 preferred_zone
, migratetype
);
1732 if (!page
&& gfp_mask
& __GFP_NOFAIL
)
1733 congestion_wait(BLK_RW_ASYNC
, HZ
/50);
1734 } while (!page
&& (gfp_mask
& __GFP_NOFAIL
));
1740 void wake_all_kswapd(unsigned int order
, struct zonelist
*zonelist
,
1741 enum zone_type high_zoneidx
)
1746 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
)
1747 wakeup_kswapd(zone
, order
);
1751 gfp_to_alloc_flags(gfp_t gfp_mask
)
1753 struct task_struct
*p
= current
;
1754 int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
1755 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
1757 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
1758 BUILD_BUG_ON(__GFP_HIGH
!= ALLOC_HIGH
);
1761 * The caller may dip into page reserves a bit more if the caller
1762 * cannot run direct reclaim, or if the caller has realtime scheduling
1763 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
1764 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
1766 alloc_flags
|= (gfp_mask
& __GFP_HIGH
);
1769 alloc_flags
|= ALLOC_HARDER
;
1771 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
1772 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1774 alloc_flags
&= ~ALLOC_CPUSET
;
1775 } else if (unlikely(rt_task(p
)) && !in_interrupt())
1776 alloc_flags
|= ALLOC_HARDER
;
1778 if (likely(!(gfp_mask
& __GFP_NOMEMALLOC
))) {
1779 if (!in_interrupt() &&
1780 ((p
->flags
& PF_MEMALLOC
) ||
1781 unlikely(test_thread_flag(TIF_MEMDIE
))))
1782 alloc_flags
|= ALLOC_NO_WATERMARKS
;
1788 static inline struct page
*
1789 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
1790 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1791 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
1794 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
1795 struct page
*page
= NULL
;
1797 unsigned long pages_reclaimed
= 0;
1798 unsigned long did_some_progress
;
1799 struct task_struct
*p
= current
;
1802 * In the slowpath, we sanity check order to avoid ever trying to
1803 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
1804 * be using allocators in order of preference for an area that is
1807 if (order
>= MAX_ORDER
) {
1808 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
1813 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
1814 * __GFP_NOWARN set) should not cause reclaim since the subsystem
1815 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
1816 * using a larger set of nodes after it has established that the
1817 * allowed per node queues are empty and that nodes are
1820 if (NUMA_BUILD
&& (gfp_mask
& GFP_THISNODE
) == GFP_THISNODE
)
1824 wake_all_kswapd(order
, zonelist
, high_zoneidx
);
1827 * OK, we're below the kswapd watermark and have kicked background
1828 * reclaim. Now things get more complex, so set up alloc_flags according
1829 * to how we want to proceed.
1831 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
1833 /* This is the last chance, in general, before the goto nopage. */
1834 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
, zonelist
,
1835 high_zoneidx
, alloc_flags
& ~ALLOC_NO_WATERMARKS
,
1836 preferred_zone
, migratetype
);
1841 /* Allocate without watermarks if the context allows */
1842 if (alloc_flags
& ALLOC_NO_WATERMARKS
) {
1843 page
= __alloc_pages_high_priority(gfp_mask
, order
,
1844 zonelist
, high_zoneidx
, nodemask
,
1845 preferred_zone
, migratetype
);
1850 /* Atomic allocations - we can't balance anything */
1854 /* Avoid recursion of direct reclaim */
1855 if (p
->flags
& PF_MEMALLOC
)
1858 /* Avoid allocations with no watermarks from looping endlessly */
1859 if (test_thread_flag(TIF_MEMDIE
) && !(gfp_mask
& __GFP_NOFAIL
))
1862 /* Try direct reclaim and then allocating */
1863 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
,
1864 zonelist
, high_zoneidx
,
1866 alloc_flags
, preferred_zone
,
1867 migratetype
, &did_some_progress
);
1872 * If we failed to make any progress reclaiming, then we are
1873 * running out of options and have to consider going OOM
1875 if (!did_some_progress
) {
1876 if ((gfp_mask
& __GFP_FS
) && !(gfp_mask
& __GFP_NORETRY
)) {
1877 if (oom_killer_disabled
)
1879 page
= __alloc_pages_may_oom(gfp_mask
, order
,
1880 zonelist
, high_zoneidx
,
1881 nodemask
, preferred_zone
,
1887 * The OOM killer does not trigger for high-order
1888 * ~__GFP_NOFAIL allocations so if no progress is being
1889 * made, there are no other options and retrying is
1892 if (order
> PAGE_ALLOC_COSTLY_ORDER
&&
1893 !(gfp_mask
& __GFP_NOFAIL
))
1900 /* Check if we should retry the allocation */
1901 pages_reclaimed
+= did_some_progress
;
1902 if (should_alloc_retry(gfp_mask
, order
, pages_reclaimed
)) {
1903 /* Wait for some write requests to complete then retry */
1904 congestion_wait(BLK_RW_ASYNC
, HZ
/50);
1909 if (!(gfp_mask
& __GFP_NOWARN
) && printk_ratelimit()) {
1910 printk(KERN_WARNING
"%s: page allocation failure."
1911 " order:%d, mode:0x%x\n",
1912 p
->comm
, order
, gfp_mask
);
1918 if (kmemcheck_enabled
)
1919 kmemcheck_pagealloc_alloc(page
, order
, gfp_mask
);
1925 * This is the 'heart' of the zoned buddy allocator.
1928 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
,
1929 struct zonelist
*zonelist
, nodemask_t
*nodemask
)
1931 enum zone_type high_zoneidx
= gfp_zone(gfp_mask
);
1932 struct zone
*preferred_zone
;
1934 int migratetype
= allocflags_to_migratetype(gfp_mask
);
1936 gfp_mask
&= gfp_allowed_mask
;
1938 lockdep_trace_alloc(gfp_mask
);
1940 might_sleep_if(gfp_mask
& __GFP_WAIT
);
1942 if (should_fail_alloc_page(gfp_mask
, order
))
1946 * Check the zones suitable for the gfp_mask contain at least one
1947 * valid zone. It's possible to have an empty zonelist as a result
1948 * of GFP_THISNODE and a memoryless node
1950 if (unlikely(!zonelist
->_zonerefs
->zone
))
1953 /* The preferred zone is used for statistics later */
1954 first_zones_zonelist(zonelist
, high_zoneidx
, nodemask
, &preferred_zone
);
1955 if (!preferred_zone
)
1958 /* First allocation attempt */
1959 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
, order
,
1960 zonelist
, high_zoneidx
, ALLOC_WMARK_LOW
|ALLOC_CPUSET
,
1961 preferred_zone
, migratetype
);
1962 if (unlikely(!page
))
1963 page
= __alloc_pages_slowpath(gfp_mask
, order
,
1964 zonelist
, high_zoneidx
, nodemask
,
1965 preferred_zone
, migratetype
);
1967 trace_mm_page_alloc(page
, order
, gfp_mask
, migratetype
);
1970 EXPORT_SYMBOL(__alloc_pages_nodemask
);
1973 * Common helper functions.
1975 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
1980 * __get_free_pages() returns a 32-bit address, which cannot represent
1983 VM_BUG_ON((gfp_mask
& __GFP_HIGHMEM
) != 0);
1985 page
= alloc_pages(gfp_mask
, order
);
1988 return (unsigned long) page_address(page
);
1990 EXPORT_SYMBOL(__get_free_pages
);
1992 unsigned long get_zeroed_page(gfp_t gfp_mask
)
1994 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
1996 EXPORT_SYMBOL(get_zeroed_page
);
1998 void __pagevec_free(struct pagevec
*pvec
)
2000 int i
= pagevec_count(pvec
);
2003 trace_mm_pagevec_free(pvec
->pages
[i
], pvec
->cold
);
2004 free_hot_cold_page(pvec
->pages
[i
], pvec
->cold
);
2008 void __free_pages(struct page
*page
, unsigned int order
)
2010 if (put_page_testzero(page
)) {
2011 trace_mm_page_free_direct(page
, order
);
2013 free_hot_page(page
);
2015 __free_pages_ok(page
, order
);
2019 EXPORT_SYMBOL(__free_pages
);
2021 void free_pages(unsigned long addr
, unsigned int order
)
2024 VM_BUG_ON(!virt_addr_valid((void *)addr
));
2025 __free_pages(virt_to_page((void *)addr
), order
);
2029 EXPORT_SYMBOL(free_pages
);
2032 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2033 * @size: the number of bytes to allocate
2034 * @gfp_mask: GFP flags for the allocation
2036 * This function is similar to alloc_pages(), except that it allocates the
2037 * minimum number of pages to satisfy the request. alloc_pages() can only
2038 * allocate memory in power-of-two pages.
2040 * This function is also limited by MAX_ORDER.
2042 * Memory allocated by this function must be released by free_pages_exact().
2044 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
2046 unsigned int order
= get_order(size
);
2049 addr
= __get_free_pages(gfp_mask
, order
);
2051 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
2052 unsigned long used
= addr
+ PAGE_ALIGN(size
);
2054 split_page(virt_to_page((void *)addr
), order
);
2055 while (used
< alloc_end
) {
2061 return (void *)addr
;
2063 EXPORT_SYMBOL(alloc_pages_exact
);
2066 * free_pages_exact - release memory allocated via alloc_pages_exact()
2067 * @virt: the value returned by alloc_pages_exact.
2068 * @size: size of allocation, same value as passed to alloc_pages_exact().
2070 * Release the memory allocated by a previous call to alloc_pages_exact.
2072 void free_pages_exact(void *virt
, size_t size
)
2074 unsigned long addr
= (unsigned long)virt
;
2075 unsigned long end
= addr
+ PAGE_ALIGN(size
);
2077 while (addr
< end
) {
2082 EXPORT_SYMBOL(free_pages_exact
);
2084 static unsigned int nr_free_zone_pages(int offset
)
2089 /* Just pick one node, since fallback list is circular */
2090 unsigned int sum
= 0;
2092 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
2094 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
2095 unsigned long size
= zone
->present_pages
;
2096 unsigned long high
= high_wmark_pages(zone
);
2105 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2107 unsigned int nr_free_buffer_pages(void)
2109 return nr_free_zone_pages(gfp_zone(GFP_USER
));
2111 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
2114 * Amount of free RAM allocatable within all zones
2116 unsigned int nr_free_pagecache_pages(void)
2118 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
2121 static inline void show_node(struct zone
*zone
)
2124 printk("Node %d ", zone_to_nid(zone
));
2127 void si_meminfo(struct sysinfo
*val
)
2129 val
->totalram
= totalram_pages
;
2131 val
->freeram
= global_page_state(NR_FREE_PAGES
);
2132 val
->bufferram
= nr_blockdev_pages();
2133 val
->totalhigh
= totalhigh_pages
;
2134 val
->freehigh
= nr_free_highpages();
2135 val
->mem_unit
= PAGE_SIZE
;
2138 EXPORT_SYMBOL(si_meminfo
);
2141 void si_meminfo_node(struct sysinfo
*val
, int nid
)
2143 pg_data_t
*pgdat
= NODE_DATA(nid
);
2145 val
->totalram
= pgdat
->node_present_pages
;
2146 val
->freeram
= node_page_state(nid
, NR_FREE_PAGES
);
2147 #ifdef CONFIG_HIGHMEM
2148 val
->totalhigh
= pgdat
->node_zones
[ZONE_HIGHMEM
].present_pages
;
2149 val
->freehigh
= zone_page_state(&pgdat
->node_zones
[ZONE_HIGHMEM
],
2155 val
->mem_unit
= PAGE_SIZE
;
2159 #define K(x) ((x) << (PAGE_SHIFT-10))
2162 * Show free area list (used inside shift_scroll-lock stuff)
2163 * We also calculate the percentage fragmentation. We do this by counting the
2164 * memory on each free list with the exception of the first item on the list.
2166 void show_free_areas(void)
2171 for_each_populated_zone(zone
) {
2173 printk("%s per-cpu:\n", zone
->name
);
2175 for_each_online_cpu(cpu
) {
2176 struct per_cpu_pageset
*pageset
;
2178 pageset
= per_cpu_ptr(zone
->pageset
, cpu
);
2180 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2181 cpu
, pageset
->pcp
.high
,
2182 pageset
->pcp
.batch
, pageset
->pcp
.count
);
2186 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2187 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2189 " dirty:%lu writeback:%lu unstable:%lu\n"
2190 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2191 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n",
2192 global_page_state(NR_ACTIVE_ANON
),
2193 global_page_state(NR_INACTIVE_ANON
),
2194 global_page_state(NR_ISOLATED_ANON
),
2195 global_page_state(NR_ACTIVE_FILE
),
2196 global_page_state(NR_INACTIVE_FILE
),
2197 global_page_state(NR_ISOLATED_FILE
),
2198 global_page_state(NR_UNEVICTABLE
),
2199 global_page_state(NR_FILE_DIRTY
),
2200 global_page_state(NR_WRITEBACK
),
2201 global_page_state(NR_UNSTABLE_NFS
),
2202 global_page_state(NR_FREE_PAGES
),
2203 global_page_state(NR_SLAB_RECLAIMABLE
),
2204 global_page_state(NR_SLAB_UNRECLAIMABLE
),
2205 global_page_state(NR_FILE_MAPPED
),
2206 global_page_state(NR_SHMEM
),
2207 global_page_state(NR_PAGETABLE
),
2208 global_page_state(NR_BOUNCE
));
2210 for_each_populated_zone(zone
) {
2219 " active_anon:%lukB"
2220 " inactive_anon:%lukB"
2221 " active_file:%lukB"
2222 " inactive_file:%lukB"
2223 " unevictable:%lukB"
2224 " isolated(anon):%lukB"
2225 " isolated(file):%lukB"
2232 " slab_reclaimable:%lukB"
2233 " slab_unreclaimable:%lukB"
2234 " kernel_stack:%lukB"
2238 " writeback_tmp:%lukB"
2239 " pages_scanned:%lu"
2240 " all_unreclaimable? %s"
2243 K(zone_page_state(zone
, NR_FREE_PAGES
)),
2244 K(min_wmark_pages(zone
)),
2245 K(low_wmark_pages(zone
)),
2246 K(high_wmark_pages(zone
)),
2247 K(zone_page_state(zone
, NR_ACTIVE_ANON
)),
2248 K(zone_page_state(zone
, NR_INACTIVE_ANON
)),
2249 K(zone_page_state(zone
, NR_ACTIVE_FILE
)),
2250 K(zone_page_state(zone
, NR_INACTIVE_FILE
)),
2251 K(zone_page_state(zone
, NR_UNEVICTABLE
)),
2252 K(zone_page_state(zone
, NR_ISOLATED_ANON
)),
2253 K(zone_page_state(zone
, NR_ISOLATED_FILE
)),
2254 K(zone
->present_pages
),
2255 K(zone_page_state(zone
, NR_MLOCK
)),
2256 K(zone_page_state(zone
, NR_FILE_DIRTY
)),
2257 K(zone_page_state(zone
, NR_WRITEBACK
)),
2258 K(zone_page_state(zone
, NR_FILE_MAPPED
)),
2259 K(zone_page_state(zone
, NR_SHMEM
)),
2260 K(zone_page_state(zone
, NR_SLAB_RECLAIMABLE
)),
2261 K(zone_page_state(zone
, NR_SLAB_UNRECLAIMABLE
)),
2262 zone_page_state(zone
, NR_KERNEL_STACK
) *
2264 K(zone_page_state(zone
, NR_PAGETABLE
)),
2265 K(zone_page_state(zone
, NR_UNSTABLE_NFS
)),
2266 K(zone_page_state(zone
, NR_BOUNCE
)),
2267 K(zone_page_state(zone
, NR_WRITEBACK_TEMP
)),
2268 zone
->pages_scanned
,
2269 (zone_is_all_unreclaimable(zone
) ? "yes" : "no")
2271 printk("lowmem_reserve[]:");
2272 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
2273 printk(" %lu", zone
->lowmem_reserve
[i
]);
2277 for_each_populated_zone(zone
) {
2278 unsigned long nr
[MAX_ORDER
], flags
, order
, total
= 0;
2281 printk("%s: ", zone
->name
);
2283 spin_lock_irqsave(&zone
->lock
, flags
);
2284 for (order
= 0; order
< MAX_ORDER
; order
++) {
2285 nr
[order
] = zone
->free_area
[order
].nr_free
;
2286 total
+= nr
[order
] << order
;
2288 spin_unlock_irqrestore(&zone
->lock
, flags
);
2289 for (order
= 0; order
< MAX_ORDER
; order
++)
2290 printk("%lu*%lukB ", nr
[order
], K(1UL) << order
);
2291 printk("= %lukB\n", K(total
));
2294 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES
));
2296 show_swap_cache_info();
2299 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
2301 zoneref
->zone
= zone
;
2302 zoneref
->zone_idx
= zone_idx(zone
);
2306 * Builds allocation fallback zone lists.
2308 * Add all populated zones of a node to the zonelist.
2310 static int build_zonelists_node(pg_data_t
*pgdat
, struct zonelist
*zonelist
,
2311 int nr_zones
, enum zone_type zone_type
)
2315 BUG_ON(zone_type
>= MAX_NR_ZONES
);
2320 zone
= pgdat
->node_zones
+ zone_type
;
2321 if (populated_zone(zone
)) {
2322 zoneref_set_zone(zone
,
2323 &zonelist
->_zonerefs
[nr_zones
++]);
2324 check_highest_zone(zone_type
);
2327 } while (zone_type
);
2334 * 0 = automatic detection of better ordering.
2335 * 1 = order by ([node] distance, -zonetype)
2336 * 2 = order by (-zonetype, [node] distance)
2338 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
2339 * the same zonelist. So only NUMA can configure this param.
2341 #define ZONELIST_ORDER_DEFAULT 0
2342 #define ZONELIST_ORDER_NODE 1
2343 #define ZONELIST_ORDER_ZONE 2
2345 /* zonelist order in the kernel.
2346 * set_zonelist_order() will set this to NODE or ZONE.
2348 static int current_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
2349 static char zonelist_order_name
[3][8] = {"Default", "Node", "Zone"};
2353 /* The value user specified ....changed by config */
2354 static int user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
2355 /* string for sysctl */
2356 #define NUMA_ZONELIST_ORDER_LEN 16
2357 char numa_zonelist_order
[16] = "default";
2360 * interface for configure zonelist ordering.
2361 * command line option "numa_zonelist_order"
2362 * = "[dD]efault - default, automatic configuration.
2363 * = "[nN]ode - order by node locality, then by zone within node
2364 * = "[zZ]one - order by zone, then by locality within zone
2367 static int __parse_numa_zonelist_order(char *s
)
2369 if (*s
== 'd' || *s
== 'D') {
2370 user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
2371 } else if (*s
== 'n' || *s
== 'N') {
2372 user_zonelist_order
= ZONELIST_ORDER_NODE
;
2373 } else if (*s
== 'z' || *s
== 'Z') {
2374 user_zonelist_order
= ZONELIST_ORDER_ZONE
;
2377 "Ignoring invalid numa_zonelist_order value: "
2384 static __init
int setup_numa_zonelist_order(char *s
)
2387 return __parse_numa_zonelist_order(s
);
2390 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
2393 * sysctl handler for numa_zonelist_order
2395 int numa_zonelist_order_handler(ctl_table
*table
, int write
,
2396 void __user
*buffer
, size_t *length
,
2399 char saved_string
[NUMA_ZONELIST_ORDER_LEN
];
2401 static DEFINE_MUTEX(zl_order_mutex
);
2403 mutex_lock(&zl_order_mutex
);
2405 strcpy(saved_string
, (char*)table
->data
);
2406 ret
= proc_dostring(table
, write
, buffer
, length
, ppos
);
2410 int oldval
= user_zonelist_order
;
2411 if (__parse_numa_zonelist_order((char*)table
->data
)) {
2413 * bogus value. restore saved string
2415 strncpy((char*)table
->data
, saved_string
,
2416 NUMA_ZONELIST_ORDER_LEN
);
2417 user_zonelist_order
= oldval
;
2418 } else if (oldval
!= user_zonelist_order
)
2419 build_all_zonelists();
2422 mutex_unlock(&zl_order_mutex
);
2427 #define MAX_NODE_LOAD (nr_online_nodes)
2428 static int node_load
[MAX_NUMNODES
];
2431 * find_next_best_node - find the next node that should appear in a given node's fallback list
2432 * @node: node whose fallback list we're appending
2433 * @used_node_mask: nodemask_t of already used nodes
2435 * We use a number of factors to determine which is the next node that should
2436 * appear on a given node's fallback list. The node should not have appeared
2437 * already in @node's fallback list, and it should be the next closest node
2438 * according to the distance array (which contains arbitrary distance values
2439 * from each node to each node in the system), and should also prefer nodes
2440 * with no CPUs, since presumably they'll have very little allocation pressure
2441 * on them otherwise.
2442 * It returns -1 if no node is found.
2444 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
2447 int min_val
= INT_MAX
;
2449 const struct cpumask
*tmp
= cpumask_of_node(0);
2451 /* Use the local node if we haven't already */
2452 if (!node_isset(node
, *used_node_mask
)) {
2453 node_set(node
, *used_node_mask
);
2457 for_each_node_state(n
, N_HIGH_MEMORY
) {
2459 /* Don't want a node to appear more than once */
2460 if (node_isset(n
, *used_node_mask
))
2463 /* Use the distance array to find the distance */
2464 val
= node_distance(node
, n
);
2466 /* Penalize nodes under us ("prefer the next node") */
2469 /* Give preference to headless and unused nodes */
2470 tmp
= cpumask_of_node(n
);
2471 if (!cpumask_empty(tmp
))
2472 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
2474 /* Slight preference for less loaded node */
2475 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
2476 val
+= node_load
[n
];
2478 if (val
< min_val
) {
2485 node_set(best_node
, *used_node_mask
);
2492 * Build zonelists ordered by node and zones within node.
2493 * This results in maximum locality--normal zone overflows into local
2494 * DMA zone, if any--but risks exhausting DMA zone.
2496 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int node
)
2499 struct zonelist
*zonelist
;
2501 zonelist
= &pgdat
->node_zonelists
[0];
2502 for (j
= 0; zonelist
->_zonerefs
[j
].zone
!= NULL
; j
++)
2504 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
2506 zonelist
->_zonerefs
[j
].zone
= NULL
;
2507 zonelist
->_zonerefs
[j
].zone_idx
= 0;
2511 * Build gfp_thisnode zonelists
2513 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
2516 struct zonelist
*zonelist
;
2518 zonelist
= &pgdat
->node_zonelists
[1];
2519 j
= build_zonelists_node(pgdat
, zonelist
, 0, MAX_NR_ZONES
- 1);
2520 zonelist
->_zonerefs
[j
].zone
= NULL
;
2521 zonelist
->_zonerefs
[j
].zone_idx
= 0;
2525 * Build zonelists ordered by zone and nodes within zones.
2526 * This results in conserving DMA zone[s] until all Normal memory is
2527 * exhausted, but results in overflowing to remote node while memory
2528 * may still exist in local DMA zone.
2530 static int node_order
[MAX_NUMNODES
];
2532 static void build_zonelists_in_zone_order(pg_data_t
*pgdat
, int nr_nodes
)
2535 int zone_type
; /* needs to be signed */
2537 struct zonelist
*zonelist
;
2539 zonelist
= &pgdat
->node_zonelists
[0];
2541 for (zone_type
= MAX_NR_ZONES
- 1; zone_type
>= 0; zone_type
--) {
2542 for (j
= 0; j
< nr_nodes
; j
++) {
2543 node
= node_order
[j
];
2544 z
= &NODE_DATA(node
)->node_zones
[zone_type
];
2545 if (populated_zone(z
)) {
2547 &zonelist
->_zonerefs
[pos
++]);
2548 check_highest_zone(zone_type
);
2552 zonelist
->_zonerefs
[pos
].zone
= NULL
;
2553 zonelist
->_zonerefs
[pos
].zone_idx
= 0;
2556 static int default_zonelist_order(void)
2559 unsigned long low_kmem_size
,total_size
;
2563 * ZONE_DMA and ZONE_DMA32 can be very small area in the sytem.
2564 * If they are really small and used heavily, the system can fall
2565 * into OOM very easily.
2566 * This function detect ZONE_DMA/DMA32 size and confgigures zone order.
2568 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
2571 for_each_online_node(nid
) {
2572 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
2573 z
= &NODE_DATA(nid
)->node_zones
[zone_type
];
2574 if (populated_zone(z
)) {
2575 if (zone_type
< ZONE_NORMAL
)
2576 low_kmem_size
+= z
->present_pages
;
2577 total_size
+= z
->present_pages
;
2581 if (!low_kmem_size
|| /* there are no DMA area. */
2582 low_kmem_size
> total_size
/2) /* DMA/DMA32 is big. */
2583 return ZONELIST_ORDER_NODE
;
2585 * look into each node's config.
2586 * If there is a node whose DMA/DMA32 memory is very big area on
2587 * local memory, NODE_ORDER may be suitable.
2589 average_size
= total_size
/
2590 (nodes_weight(node_states
[N_HIGH_MEMORY
]) + 1);
2591 for_each_online_node(nid
) {
2594 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
2595 z
= &NODE_DATA(nid
)->node_zones
[zone_type
];
2596 if (populated_zone(z
)) {
2597 if (zone_type
< ZONE_NORMAL
)
2598 low_kmem_size
+= z
->present_pages
;
2599 total_size
+= z
->present_pages
;
2602 if (low_kmem_size
&&
2603 total_size
> average_size
&& /* ignore small node */
2604 low_kmem_size
> total_size
* 70/100)
2605 return ZONELIST_ORDER_NODE
;
2607 return ZONELIST_ORDER_ZONE
;
2610 static void set_zonelist_order(void)
2612 if (user_zonelist_order
== ZONELIST_ORDER_DEFAULT
)
2613 current_zonelist_order
= default_zonelist_order();
2615 current_zonelist_order
= user_zonelist_order
;
2618 static void build_zonelists(pg_data_t
*pgdat
)
2622 nodemask_t used_mask
;
2623 int local_node
, prev_node
;
2624 struct zonelist
*zonelist
;
2625 int order
= current_zonelist_order
;
2627 /* initialize zonelists */
2628 for (i
= 0; i
< MAX_ZONELISTS
; i
++) {
2629 zonelist
= pgdat
->node_zonelists
+ i
;
2630 zonelist
->_zonerefs
[0].zone
= NULL
;
2631 zonelist
->_zonerefs
[0].zone_idx
= 0;
2634 /* NUMA-aware ordering of nodes */
2635 local_node
= pgdat
->node_id
;
2636 load
= nr_online_nodes
;
2637 prev_node
= local_node
;
2638 nodes_clear(used_mask
);
2640 memset(node_order
, 0, sizeof(node_order
));
2643 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
2644 int distance
= node_distance(local_node
, node
);
2647 * If another node is sufficiently far away then it is better
2648 * to reclaim pages in a zone before going off node.
2650 if (distance
> RECLAIM_DISTANCE
)
2651 zone_reclaim_mode
= 1;
2654 * We don't want to pressure a particular node.
2655 * So adding penalty to the first node in same
2656 * distance group to make it round-robin.
2658 if (distance
!= node_distance(local_node
, prev_node
))
2659 node_load
[node
] = load
;
2663 if (order
== ZONELIST_ORDER_NODE
)
2664 build_zonelists_in_node_order(pgdat
, node
);
2666 node_order
[j
++] = node
; /* remember order */
2669 if (order
== ZONELIST_ORDER_ZONE
) {
2670 /* calculate node order -- i.e., DMA last! */
2671 build_zonelists_in_zone_order(pgdat
, j
);
2674 build_thisnode_zonelists(pgdat
);
2677 /* Construct the zonelist performance cache - see further mmzone.h */
2678 static void build_zonelist_cache(pg_data_t
*pgdat
)
2680 struct zonelist
*zonelist
;
2681 struct zonelist_cache
*zlc
;
2684 zonelist
= &pgdat
->node_zonelists
[0];
2685 zonelist
->zlcache_ptr
= zlc
= &zonelist
->zlcache
;
2686 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
2687 for (z
= zonelist
->_zonerefs
; z
->zone
; z
++)
2688 zlc
->z_to_n
[z
- zonelist
->_zonerefs
] = zonelist_node_idx(z
);
2692 #else /* CONFIG_NUMA */
2694 static void set_zonelist_order(void)
2696 current_zonelist_order
= ZONELIST_ORDER_ZONE
;
2699 static void build_zonelists(pg_data_t
*pgdat
)
2701 int node
, local_node
;
2703 struct zonelist
*zonelist
;
2705 local_node
= pgdat
->node_id
;
2707 zonelist
= &pgdat
->node_zonelists
[0];
2708 j
= build_zonelists_node(pgdat
, zonelist
, 0, MAX_NR_ZONES
- 1);
2711 * Now we build the zonelist so that it contains the zones
2712 * of all the other nodes.
2713 * We don't want to pressure a particular node, so when
2714 * building the zones for node N, we make sure that the
2715 * zones coming right after the local ones are those from
2716 * node N+1 (modulo N)
2718 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
2719 if (!node_online(node
))
2721 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
2724 for (node
= 0; node
< local_node
; node
++) {
2725 if (!node_online(node
))
2727 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
2731 zonelist
->_zonerefs
[j
].zone
= NULL
;
2732 zonelist
->_zonerefs
[j
].zone_idx
= 0;
2735 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
2736 static void build_zonelist_cache(pg_data_t
*pgdat
)
2738 pgdat
->node_zonelists
[0].zlcache_ptr
= NULL
;
2741 #endif /* CONFIG_NUMA */
2744 * Boot pageset table. One per cpu which is going to be used for all
2745 * zones and all nodes. The parameters will be set in such a way
2746 * that an item put on a list will immediately be handed over to
2747 * the buddy list. This is safe since pageset manipulation is done
2748 * with interrupts disabled.
2750 * The boot_pagesets must be kept even after bootup is complete for
2751 * unused processors and/or zones. They do play a role for bootstrapping
2752 * hotplugged processors.
2754 * zoneinfo_show() and maybe other functions do
2755 * not check if the processor is online before following the pageset pointer.
2756 * Other parts of the kernel may not check if the zone is available.
2758 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
);
2759 static DEFINE_PER_CPU(struct per_cpu_pageset
, boot_pageset
);
2761 /* return values int ....just for stop_machine() */
2762 static int __build_all_zonelists(void *dummy
)
2768 memset(node_load
, 0, sizeof(node_load
));
2770 for_each_online_node(nid
) {
2771 pg_data_t
*pgdat
= NODE_DATA(nid
);
2773 build_zonelists(pgdat
);
2774 build_zonelist_cache(pgdat
);
2778 * Initialize the boot_pagesets that are going to be used
2779 * for bootstrapping processors. The real pagesets for
2780 * each zone will be allocated later when the per cpu
2781 * allocator is available.
2783 * boot_pagesets are used also for bootstrapping offline
2784 * cpus if the system is already booted because the pagesets
2785 * are needed to initialize allocators on a specific cpu too.
2786 * F.e. the percpu allocator needs the page allocator which
2787 * needs the percpu allocator in order to allocate its pagesets
2788 * (a chicken-egg dilemma).
2790 for_each_possible_cpu(cpu
)
2791 setup_pageset(&per_cpu(boot_pageset
, cpu
), 0);
2796 void build_all_zonelists(void)
2798 set_zonelist_order();
2800 if (system_state
== SYSTEM_BOOTING
) {
2801 __build_all_zonelists(NULL
);
2802 mminit_verify_zonelist();
2803 cpuset_init_current_mems_allowed();
2805 /* we have to stop all cpus to guarantee there is no user
2807 stop_machine(__build_all_zonelists
, NULL
, NULL
);
2808 /* cpuset refresh routine should be here */
2810 vm_total_pages
= nr_free_pagecache_pages();
2812 * Disable grouping by mobility if the number of pages in the
2813 * system is too low to allow the mechanism to work. It would be
2814 * more accurate, but expensive to check per-zone. This check is
2815 * made on memory-hotadd so a system can start with mobility
2816 * disabled and enable it later
2818 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
2819 page_group_by_mobility_disabled
= 1;
2821 page_group_by_mobility_disabled
= 0;
2823 printk("Built %i zonelists in %s order, mobility grouping %s. "
2824 "Total pages: %ld\n",
2826 zonelist_order_name
[current_zonelist_order
],
2827 page_group_by_mobility_disabled
? "off" : "on",
2830 printk("Policy zone: %s\n", zone_names
[policy_zone
]);
2835 * Helper functions to size the waitqueue hash table.
2836 * Essentially these want to choose hash table sizes sufficiently
2837 * large so that collisions trying to wait on pages are rare.
2838 * But in fact, the number of active page waitqueues on typical
2839 * systems is ridiculously low, less than 200. So this is even
2840 * conservative, even though it seems large.
2842 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
2843 * waitqueues, i.e. the size of the waitq table given the number of pages.
2845 #define PAGES_PER_WAITQUEUE 256
2847 #ifndef CONFIG_MEMORY_HOTPLUG
2848 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
2850 unsigned long size
= 1;
2852 pages
/= PAGES_PER_WAITQUEUE
;
2854 while (size
< pages
)
2858 * Once we have dozens or even hundreds of threads sleeping
2859 * on IO we've got bigger problems than wait queue collision.
2860 * Limit the size of the wait table to a reasonable size.
2862 size
= min(size
, 4096UL);
2864 return max(size
, 4UL);
2868 * A zone's size might be changed by hot-add, so it is not possible to determine
2869 * a suitable size for its wait_table. So we use the maximum size now.
2871 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
2873 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
2874 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
2875 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
2877 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
2878 * or more by the traditional way. (See above). It equals:
2880 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
2881 * ia64(16K page size) : = ( 8G + 4M)byte.
2882 * powerpc (64K page size) : = (32G +16M)byte.
2884 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
2891 * This is an integer logarithm so that shifts can be used later
2892 * to extract the more random high bits from the multiplicative
2893 * hash function before the remainder is taken.
2895 static inline unsigned long wait_table_bits(unsigned long size
)
2900 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
2903 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
2904 * of blocks reserved is based on min_wmark_pages(zone). The memory within
2905 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
2906 * higher will lead to a bigger reserve which will get freed as contiguous
2907 * blocks as reclaim kicks in
2909 static void setup_zone_migrate_reserve(struct zone
*zone
)
2911 unsigned long start_pfn
, pfn
, end_pfn
;
2913 unsigned long block_migratetype
;
2916 /* Get the start pfn, end pfn and the number of blocks to reserve */
2917 start_pfn
= zone
->zone_start_pfn
;
2918 end_pfn
= start_pfn
+ zone
->spanned_pages
;
2919 reserve
= roundup(min_wmark_pages(zone
), pageblock_nr_pages
) >>
2923 * Reserve blocks are generally in place to help high-order atomic
2924 * allocations that are short-lived. A min_free_kbytes value that
2925 * would result in more than 2 reserve blocks for atomic allocations
2926 * is assumed to be in place to help anti-fragmentation for the
2927 * future allocation of hugepages at runtime.
2929 reserve
= min(2, reserve
);
2931 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= pageblock_nr_pages
) {
2932 if (!pfn_valid(pfn
))
2934 page
= pfn_to_page(pfn
);
2936 /* Watch out for overlapping nodes */
2937 if (page_to_nid(page
) != zone_to_nid(zone
))
2940 /* Blocks with reserved pages will never free, skip them. */
2941 if (PageReserved(page
))
2944 block_migratetype
= get_pageblock_migratetype(page
);
2946 /* If this block is reserved, account for it */
2947 if (reserve
> 0 && block_migratetype
== MIGRATE_RESERVE
) {
2952 /* Suitable for reserving if this block is movable */
2953 if (reserve
> 0 && block_migratetype
== MIGRATE_MOVABLE
) {
2954 set_pageblock_migratetype(page
, MIGRATE_RESERVE
);
2955 move_freepages_block(zone
, page
, MIGRATE_RESERVE
);
2961 * If the reserve is met and this is a previous reserved block,
2964 if (block_migratetype
== MIGRATE_RESERVE
) {
2965 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
2966 move_freepages_block(zone
, page
, MIGRATE_MOVABLE
);
2972 * Initially all pages are reserved - free ones are freed
2973 * up by free_all_bootmem() once the early boot process is
2974 * done. Non-atomic initialization, single-pass.
2976 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
2977 unsigned long start_pfn
, enum memmap_context context
)
2980 unsigned long end_pfn
= start_pfn
+ size
;
2984 if (highest_memmap_pfn
< end_pfn
- 1)
2985 highest_memmap_pfn
= end_pfn
- 1;
2987 z
= &NODE_DATA(nid
)->node_zones
[zone
];
2988 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
2990 * There can be holes in boot-time mem_map[]s
2991 * handed to this function. They do not
2992 * exist on hotplugged memory.
2994 if (context
== MEMMAP_EARLY
) {
2995 if (!early_pfn_valid(pfn
))
2997 if (!early_pfn_in_nid(pfn
, nid
))
3000 page
= pfn_to_page(pfn
);
3001 set_page_links(page
, zone
, nid
, pfn
);
3002 mminit_verify_page_links(page
, zone
, nid
, pfn
);
3003 init_page_count(page
);
3004 reset_page_mapcount(page
);
3005 SetPageReserved(page
);
3007 * Mark the block movable so that blocks are reserved for
3008 * movable at startup. This will force kernel allocations
3009 * to reserve their blocks rather than leaking throughout
3010 * the address space during boot when many long-lived
3011 * kernel allocations are made. Later some blocks near
3012 * the start are marked MIGRATE_RESERVE by
3013 * setup_zone_migrate_reserve()
3015 * bitmap is created for zone's valid pfn range. but memmap
3016 * can be created for invalid pages (for alignment)
3017 * check here not to call set_pageblock_migratetype() against
3020 if ((z
->zone_start_pfn
<= pfn
)
3021 && (pfn
< z
->zone_start_pfn
+ z
->spanned_pages
)
3022 && !(pfn
& (pageblock_nr_pages
- 1)))
3023 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
3025 INIT_LIST_HEAD(&page
->lru
);
3026 #ifdef WANT_PAGE_VIRTUAL
3027 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3028 if (!is_highmem_idx(zone
))
3029 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
3034 static void __meminit
zone_init_free_lists(struct zone
*zone
)
3037 for_each_migratetype_order(order
, t
) {
3038 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
3039 zone
->free_area
[order
].nr_free
= 0;
3043 #ifndef __HAVE_ARCH_MEMMAP_INIT
3044 #define memmap_init(size, nid, zone, start_pfn) \
3045 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3048 static int zone_batchsize(struct zone
*zone
)
3054 * The per-cpu-pages pools are set to around 1000th of the
3055 * size of the zone. But no more than 1/2 of a meg.
3057 * OK, so we don't know how big the cache is. So guess.
3059 batch
= zone
->present_pages
/ 1024;
3060 if (batch
* PAGE_SIZE
> 512 * 1024)
3061 batch
= (512 * 1024) / PAGE_SIZE
;
3062 batch
/= 4; /* We effectively *= 4 below */
3067 * Clamp the batch to a 2^n - 1 value. Having a power
3068 * of 2 value was found to be more likely to have
3069 * suboptimal cache aliasing properties in some cases.
3071 * For example if 2 tasks are alternately allocating
3072 * batches of pages, one task can end up with a lot
3073 * of pages of one half of the possible page colors
3074 * and the other with pages of the other colors.
3076 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
3081 /* The deferral and batching of frees should be suppressed under NOMMU
3084 * The problem is that NOMMU needs to be able to allocate large chunks
3085 * of contiguous memory as there's no hardware page translation to
3086 * assemble apparent contiguous memory from discontiguous pages.
3088 * Queueing large contiguous runs of pages for batching, however,
3089 * causes the pages to actually be freed in smaller chunks. As there
3090 * can be a significant delay between the individual batches being
3091 * recycled, this leads to the once large chunks of space being
3092 * fragmented and becoming unavailable for high-order allocations.
3098 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
3100 struct per_cpu_pages
*pcp
;
3103 memset(p
, 0, sizeof(*p
));
3107 pcp
->high
= 6 * batch
;
3108 pcp
->batch
= max(1UL, 1 * batch
);
3109 for (migratetype
= 0; migratetype
< MIGRATE_PCPTYPES
; migratetype
++)
3110 INIT_LIST_HEAD(&pcp
->lists
[migratetype
]);
3114 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3115 * to the value high for the pageset p.
3118 static void setup_pagelist_highmark(struct per_cpu_pageset
*p
,
3121 struct per_cpu_pages
*pcp
;
3125 pcp
->batch
= max(1UL, high
/4);
3126 if ((high
/4) > (PAGE_SHIFT
* 8))
3127 pcp
->batch
= PAGE_SHIFT
* 8;
3131 * Allocate per cpu pagesets and initialize them.
3132 * Before this call only boot pagesets were available.
3133 * Boot pagesets will no longer be used by this processorr
3134 * after setup_per_cpu_pageset().
3136 void __init
setup_per_cpu_pageset(void)
3141 for_each_populated_zone(zone
) {
3142 zone
->pageset
= alloc_percpu(struct per_cpu_pageset
);
3144 for_each_possible_cpu(cpu
) {
3145 struct per_cpu_pageset
*pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
3147 setup_pageset(pcp
, zone_batchsize(zone
));
3149 if (percpu_pagelist_fraction
)
3150 setup_pagelist_highmark(pcp
,
3151 (zone
->present_pages
/
3152 percpu_pagelist_fraction
));
3157 static noinline __init_refok
3158 int zone_wait_table_init(struct zone
*zone
, unsigned long zone_size_pages
)
3161 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
3165 * The per-page waitqueue mechanism uses hashed waitqueues
3168 zone
->wait_table_hash_nr_entries
=
3169 wait_table_hash_nr_entries(zone_size_pages
);
3170 zone
->wait_table_bits
=
3171 wait_table_bits(zone
->wait_table_hash_nr_entries
);
3172 alloc_size
= zone
->wait_table_hash_nr_entries
3173 * sizeof(wait_queue_head_t
);
3175 if (!slab_is_available()) {
3176 zone
->wait_table
= (wait_queue_head_t
*)
3177 alloc_bootmem_node(pgdat
, alloc_size
);
3180 * This case means that a zone whose size was 0 gets new memory
3181 * via memory hot-add.
3182 * But it may be the case that a new node was hot-added. In
3183 * this case vmalloc() will not be able to use this new node's
3184 * memory - this wait_table must be initialized to use this new
3185 * node itself as well.
3186 * To use this new node's memory, further consideration will be
3189 zone
->wait_table
= vmalloc(alloc_size
);
3191 if (!zone
->wait_table
)
3194 for(i
= 0; i
< zone
->wait_table_hash_nr_entries
; ++i
)
3195 init_waitqueue_head(zone
->wait_table
+ i
);
3200 static int __zone_pcp_update(void *data
)
3202 struct zone
*zone
= data
;
3204 unsigned long batch
= zone_batchsize(zone
), flags
;
3206 for (cpu
= 0; cpu
< NR_CPUS
; cpu
++) {
3207 struct per_cpu_pageset
*pset
;
3208 struct per_cpu_pages
*pcp
;
3210 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
3213 local_irq_save(flags
);
3214 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
3215 setup_pageset(pset
, batch
);
3216 local_irq_restore(flags
);
3221 void zone_pcp_update(struct zone
*zone
)
3223 stop_machine(__zone_pcp_update
, zone
, NULL
);
3226 static __meminit
void zone_pcp_init(struct zone
*zone
)
3229 * per cpu subsystem is not up at this point. The following code
3230 * relies on the ability of the linker to provide the
3231 * offset of a (static) per cpu variable into the per cpu area.
3233 zone
->pageset
= &boot_pageset
;
3235 if (zone
->present_pages
)
3236 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%u\n",
3237 zone
->name
, zone
->present_pages
,
3238 zone_batchsize(zone
));
3241 __meminit
int init_currently_empty_zone(struct zone
*zone
,
3242 unsigned long zone_start_pfn
,
3244 enum memmap_context context
)
3246 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
3248 ret
= zone_wait_table_init(zone
, size
);
3251 pgdat
->nr_zones
= zone_idx(zone
) + 1;
3253 zone
->zone_start_pfn
= zone_start_pfn
;
3255 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
3256 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
3258 (unsigned long)zone_idx(zone
),
3259 zone_start_pfn
, (zone_start_pfn
+ size
));
3261 zone_init_free_lists(zone
);
3266 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3268 * Basic iterator support. Return the first range of PFNs for a node
3269 * Note: nid == MAX_NUMNODES returns first region regardless of node
3271 static int __meminit
first_active_region_index_in_nid(int nid
)
3275 for (i
= 0; i
< nr_nodemap_entries
; i
++)
3276 if (nid
== MAX_NUMNODES
|| early_node_map
[i
].nid
== nid
)
3283 * Basic iterator support. Return the next active range of PFNs for a node
3284 * Note: nid == MAX_NUMNODES returns next region regardless of node
3286 static int __meminit
next_active_region_index_in_nid(int index
, int nid
)
3288 for (index
= index
+ 1; index
< nr_nodemap_entries
; index
++)
3289 if (nid
== MAX_NUMNODES
|| early_node_map
[index
].nid
== nid
)
3295 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
3297 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
3298 * Architectures may implement their own version but if add_active_range()
3299 * was used and there are no special requirements, this is a convenient
3302 int __meminit
__early_pfn_to_nid(unsigned long pfn
)
3306 for (i
= 0; i
< nr_nodemap_entries
; i
++) {
3307 unsigned long start_pfn
= early_node_map
[i
].start_pfn
;
3308 unsigned long end_pfn
= early_node_map
[i
].end_pfn
;
3310 if (start_pfn
<= pfn
&& pfn
< end_pfn
)
3311 return early_node_map
[i
].nid
;
3313 /* This is a memory hole */
3316 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
3318 int __meminit
early_pfn_to_nid(unsigned long pfn
)
3322 nid
= __early_pfn_to_nid(pfn
);
3325 /* just returns 0 */
3329 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
3330 bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
3334 nid
= __early_pfn_to_nid(pfn
);
3335 if (nid
>= 0 && nid
!= node
)
3341 /* Basic iterator support to walk early_node_map[] */
3342 #define for_each_active_range_index_in_nid(i, nid) \
3343 for (i = first_active_region_index_in_nid(nid); i != -1; \
3344 i = next_active_region_index_in_nid(i, nid))
3347 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
3348 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
3349 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
3351 * If an architecture guarantees that all ranges registered with
3352 * add_active_ranges() contain no holes and may be freed, this
3353 * this function may be used instead of calling free_bootmem() manually.
3355 void __init
free_bootmem_with_active_regions(int nid
,
3356 unsigned long max_low_pfn
)
3360 for_each_active_range_index_in_nid(i
, nid
) {
3361 unsigned long size_pages
= 0;
3362 unsigned long end_pfn
= early_node_map
[i
].end_pfn
;
3364 if (early_node_map
[i
].start_pfn
>= max_low_pfn
)
3367 if (end_pfn
> max_low_pfn
)
3368 end_pfn
= max_low_pfn
;
3370 size_pages
= end_pfn
- early_node_map
[i
].start_pfn
;
3371 free_bootmem_node(NODE_DATA(early_node_map
[i
].nid
),
3372 PFN_PHYS(early_node_map
[i
].start_pfn
),
3373 size_pages
<< PAGE_SHIFT
);
3377 int __init
add_from_early_node_map(struct range
*range
, int az
,
3378 int nr_range
, int nid
)
3383 /* need to go over early_node_map to find out good range for node */
3384 for_each_active_range_index_in_nid(i
, nid
) {
3385 start
= early_node_map
[i
].start_pfn
;
3386 end
= early_node_map
[i
].end_pfn
;
3387 nr_range
= add_range(range
, az
, nr_range
, start
, end
);
3392 #ifdef CONFIG_NO_BOOTMEM
3393 void * __init
__alloc_memory_core_early(int nid
, u64 size
, u64 align
,
3394 u64 goal
, u64 limit
)
3399 /* need to go over early_node_map to find out good range for node */
3400 for_each_active_range_index_in_nid(i
, nid
) {
3402 u64 ei_start
, ei_last
;
3404 ei_last
= early_node_map
[i
].end_pfn
;
3405 ei_last
<<= PAGE_SHIFT
;
3406 ei_start
= early_node_map
[i
].start_pfn
;
3407 ei_start
<<= PAGE_SHIFT
;
3408 addr
= find_early_area(ei_start
, ei_last
,
3409 goal
, limit
, size
, align
);
3415 printk(KERN_DEBUG
"alloc (nid=%d %llx - %llx) (%llx - %llx) %llx %llx => %llx\n",
3417 ei_start
, ei_last
, goal
, limit
, size
,
3421 ptr
= phys_to_virt(addr
);
3422 memset(ptr
, 0, size
);
3423 reserve_early_without_check(addr
, addr
+ size
, "BOOTMEM");
3432 void __init
work_with_active_regions(int nid
, work_fn_t work_fn
, void *data
)
3437 for_each_active_range_index_in_nid(i
, nid
) {
3438 ret
= work_fn(early_node_map
[i
].start_pfn
,
3439 early_node_map
[i
].end_pfn
, data
);
3445 * sparse_memory_present_with_active_regions - Call memory_present for each active range
3446 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
3448 * If an architecture guarantees that all ranges registered with
3449 * add_active_ranges() contain no holes and may be freed, this
3450 * function may be used instead of calling memory_present() manually.
3452 void __init
sparse_memory_present_with_active_regions(int nid
)
3456 for_each_active_range_index_in_nid(i
, nid
)
3457 memory_present(early_node_map
[i
].nid
,
3458 early_node_map
[i
].start_pfn
,
3459 early_node_map
[i
].end_pfn
);
3463 * get_pfn_range_for_nid - Return the start and end page frames for a node
3464 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
3465 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
3466 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
3468 * It returns the start and end page frame of a node based on information
3469 * provided by an arch calling add_active_range(). If called for a node
3470 * with no available memory, a warning is printed and the start and end
3473 void __meminit
get_pfn_range_for_nid(unsigned int nid
,
3474 unsigned long *start_pfn
, unsigned long *end_pfn
)
3480 for_each_active_range_index_in_nid(i
, nid
) {
3481 *start_pfn
= min(*start_pfn
, early_node_map
[i
].start_pfn
);
3482 *end_pfn
= max(*end_pfn
, early_node_map
[i
].end_pfn
);
3485 if (*start_pfn
== -1UL)
3490 * This finds a zone that can be used for ZONE_MOVABLE pages. The
3491 * assumption is made that zones within a node are ordered in monotonic
3492 * increasing memory addresses so that the "highest" populated zone is used
3494 static void __init
find_usable_zone_for_movable(void)
3497 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
3498 if (zone_index
== ZONE_MOVABLE
)
3501 if (arch_zone_highest_possible_pfn
[zone_index
] >
3502 arch_zone_lowest_possible_pfn
[zone_index
])
3506 VM_BUG_ON(zone_index
== -1);
3507 movable_zone
= zone_index
;
3511 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
3512 * because it is sized independant of architecture. Unlike the other zones,
3513 * the starting point for ZONE_MOVABLE is not fixed. It may be different
3514 * in each node depending on the size of each node and how evenly kernelcore
3515 * is distributed. This helper function adjusts the zone ranges
3516 * provided by the architecture for a given node by using the end of the
3517 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
3518 * zones within a node are in order of monotonic increases memory addresses
3520 static void __meminit
adjust_zone_range_for_zone_movable(int nid
,
3521 unsigned long zone_type
,
3522 unsigned long node_start_pfn
,
3523 unsigned long node_end_pfn
,
3524 unsigned long *zone_start_pfn
,
3525 unsigned long *zone_end_pfn
)
3527 /* Only adjust if ZONE_MOVABLE is on this node */
3528 if (zone_movable_pfn
[nid
]) {
3529 /* Size ZONE_MOVABLE */
3530 if (zone_type
== ZONE_MOVABLE
) {
3531 *zone_start_pfn
= zone_movable_pfn
[nid
];
3532 *zone_end_pfn
= min(node_end_pfn
,
3533 arch_zone_highest_possible_pfn
[movable_zone
]);
3535 /* Adjust for ZONE_MOVABLE starting within this range */
3536 } else if (*zone_start_pfn
< zone_movable_pfn
[nid
] &&
3537 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
3538 *zone_end_pfn
= zone_movable_pfn
[nid
];
3540 /* Check if this whole range is within ZONE_MOVABLE */
3541 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
3542 *zone_start_pfn
= *zone_end_pfn
;
3547 * Return the number of pages a zone spans in a node, including holes
3548 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
3550 static unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
3551 unsigned long zone_type
,
3552 unsigned long *ignored
)
3554 unsigned long node_start_pfn
, node_end_pfn
;
3555 unsigned long zone_start_pfn
, zone_end_pfn
;
3557 /* Get the start and end of the node and zone */
3558 get_pfn_range_for_nid(nid
, &node_start_pfn
, &node_end_pfn
);
3559 zone_start_pfn
= arch_zone_lowest_possible_pfn
[zone_type
];
3560 zone_end_pfn
= arch_zone_highest_possible_pfn
[zone_type
];
3561 adjust_zone_range_for_zone_movable(nid
, zone_type
,
3562 node_start_pfn
, node_end_pfn
,
3563 &zone_start_pfn
, &zone_end_pfn
);
3565 /* Check that this node has pages within the zone's required range */
3566 if (zone_end_pfn
< node_start_pfn
|| zone_start_pfn
> node_end_pfn
)
3569 /* Move the zone boundaries inside the node if necessary */
3570 zone_end_pfn
= min(zone_end_pfn
, node_end_pfn
);
3571 zone_start_pfn
= max(zone_start_pfn
, node_start_pfn
);
3573 /* Return the spanned pages */
3574 return zone_end_pfn
- zone_start_pfn
;
3578 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
3579 * then all holes in the requested range will be accounted for.
3581 unsigned long __meminit
__absent_pages_in_range(int nid
,
3582 unsigned long range_start_pfn
,
3583 unsigned long range_end_pfn
)
3586 unsigned long prev_end_pfn
= 0, hole_pages
= 0;
3587 unsigned long start_pfn
;
3589 /* Find the end_pfn of the first active range of pfns in the node */
3590 i
= first_active_region_index_in_nid(nid
);
3594 prev_end_pfn
= min(early_node_map
[i
].start_pfn
, range_end_pfn
);
3596 /* Account for ranges before physical memory on this node */
3597 if (early_node_map
[i
].start_pfn
> range_start_pfn
)
3598 hole_pages
= prev_end_pfn
- range_start_pfn
;
3600 /* Find all holes for the zone within the node */
3601 for (; i
!= -1; i
= next_active_region_index_in_nid(i
, nid
)) {
3603 /* No need to continue if prev_end_pfn is outside the zone */
3604 if (prev_end_pfn
>= range_end_pfn
)
3607 /* Make sure the end of the zone is not within the hole */
3608 start_pfn
= min(early_node_map
[i
].start_pfn
, range_end_pfn
);
3609 prev_end_pfn
= max(prev_end_pfn
, range_start_pfn
);
3611 /* Update the hole size cound and move on */
3612 if (start_pfn
> range_start_pfn
) {
3613 BUG_ON(prev_end_pfn
> start_pfn
);
3614 hole_pages
+= start_pfn
- prev_end_pfn
;
3616 prev_end_pfn
= early_node_map
[i
].end_pfn
;
3619 /* Account for ranges past physical memory on this node */
3620 if (range_end_pfn
> prev_end_pfn
)
3621 hole_pages
+= range_end_pfn
-
3622 max(range_start_pfn
, prev_end_pfn
);
3628 * absent_pages_in_range - Return number of page frames in holes within a range
3629 * @start_pfn: The start PFN to start searching for holes
3630 * @end_pfn: The end PFN to stop searching for holes
3632 * It returns the number of pages frames in memory holes within a range.
3634 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
3635 unsigned long end_pfn
)
3637 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
3640 /* Return the number of page frames in holes in a zone on a node */
3641 static unsigned long __meminit
zone_absent_pages_in_node(int nid
,
3642 unsigned long zone_type
,
3643 unsigned long *ignored
)
3645 unsigned long node_start_pfn
, node_end_pfn
;
3646 unsigned long zone_start_pfn
, zone_end_pfn
;
3648 get_pfn_range_for_nid(nid
, &node_start_pfn
, &node_end_pfn
);
3649 zone_start_pfn
= max(arch_zone_lowest_possible_pfn
[zone_type
],
3651 zone_end_pfn
= min(arch_zone_highest_possible_pfn
[zone_type
],
3654 adjust_zone_range_for_zone_movable(nid
, zone_type
,
3655 node_start_pfn
, node_end_pfn
,
3656 &zone_start_pfn
, &zone_end_pfn
);
3657 return __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
3661 static inline unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
3662 unsigned long zone_type
,
3663 unsigned long *zones_size
)
3665 return zones_size
[zone_type
];
3668 static inline unsigned long __meminit
zone_absent_pages_in_node(int nid
,
3669 unsigned long zone_type
,
3670 unsigned long *zholes_size
)
3675 return zholes_size
[zone_type
];
3680 static void __meminit
calculate_node_totalpages(struct pglist_data
*pgdat
,
3681 unsigned long *zones_size
, unsigned long *zholes_size
)
3683 unsigned long realtotalpages
, totalpages
= 0;
3686 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
3687 totalpages
+= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
3689 pgdat
->node_spanned_pages
= totalpages
;
3691 realtotalpages
= totalpages
;
3692 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
3694 zone_absent_pages_in_node(pgdat
->node_id
, i
,
3696 pgdat
->node_present_pages
= realtotalpages
;
3697 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
3701 #ifndef CONFIG_SPARSEMEM
3703 * Calculate the size of the zone->blockflags rounded to an unsigned long
3704 * Start by making sure zonesize is a multiple of pageblock_order by rounding
3705 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
3706 * round what is now in bits to nearest long in bits, then return it in
3709 static unsigned long __init
usemap_size(unsigned long zonesize
)
3711 unsigned long usemapsize
;
3713 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
3714 usemapsize
= usemapsize
>> pageblock_order
;
3715 usemapsize
*= NR_PAGEBLOCK_BITS
;
3716 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
3718 return usemapsize
/ 8;
3721 static void __init
setup_usemap(struct pglist_data
*pgdat
,
3722 struct zone
*zone
, unsigned long zonesize
)
3724 unsigned long usemapsize
= usemap_size(zonesize
);
3725 zone
->pageblock_flags
= NULL
;
3727 zone
->pageblock_flags
= alloc_bootmem_node(pgdat
, usemapsize
);
3730 static void inline setup_usemap(struct pglist_data
*pgdat
,
3731 struct zone
*zone
, unsigned long zonesize
) {}
3732 #endif /* CONFIG_SPARSEMEM */
3734 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
3736 /* Return a sensible default order for the pageblock size. */
3737 static inline int pageblock_default_order(void)
3739 if (HPAGE_SHIFT
> PAGE_SHIFT
)
3740 return HUGETLB_PAGE_ORDER
;
3745 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
3746 static inline void __init
set_pageblock_order(unsigned int order
)
3748 /* Check that pageblock_nr_pages has not already been setup */
3749 if (pageblock_order
)
3753 * Assume the largest contiguous order of interest is a huge page.
3754 * This value may be variable depending on boot parameters on IA64
3756 pageblock_order
= order
;
3758 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
3761 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
3762 * and pageblock_default_order() are unused as pageblock_order is set
3763 * at compile-time. See include/linux/pageblock-flags.h for the values of
3764 * pageblock_order based on the kernel config
3766 static inline int pageblock_default_order(unsigned int order
)
3770 #define set_pageblock_order(x) do {} while (0)
3772 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
3775 * Set up the zone data structures:
3776 * - mark all pages reserved
3777 * - mark all memory queues empty
3778 * - clear the memory bitmaps
3780 static void __paginginit
free_area_init_core(struct pglist_data
*pgdat
,
3781 unsigned long *zones_size
, unsigned long *zholes_size
)
3784 int nid
= pgdat
->node_id
;
3785 unsigned long zone_start_pfn
= pgdat
->node_start_pfn
;
3788 pgdat_resize_init(pgdat
);
3789 pgdat
->nr_zones
= 0;
3790 init_waitqueue_head(&pgdat
->kswapd_wait
);
3791 pgdat
->kswapd_max_order
= 0;
3792 pgdat_page_cgroup_init(pgdat
);
3794 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
3795 struct zone
*zone
= pgdat
->node_zones
+ j
;
3796 unsigned long size
, realsize
, memmap_pages
;
3799 size
= zone_spanned_pages_in_node(nid
, j
, zones_size
);
3800 realsize
= size
- zone_absent_pages_in_node(nid
, j
,
3804 * Adjust realsize so that it accounts for how much memory
3805 * is used by this zone for memmap. This affects the watermark
3806 * and per-cpu initialisations
3809 PAGE_ALIGN(size
* sizeof(struct page
)) >> PAGE_SHIFT
;
3810 if (realsize
>= memmap_pages
) {
3811 realsize
-= memmap_pages
;
3814 " %s zone: %lu pages used for memmap\n",
3815 zone_names
[j
], memmap_pages
);
3818 " %s zone: %lu pages exceeds realsize %lu\n",
3819 zone_names
[j
], memmap_pages
, realsize
);
3821 /* Account for reserved pages */
3822 if (j
== 0 && realsize
> dma_reserve
) {
3823 realsize
-= dma_reserve
;
3824 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
3825 zone_names
[0], dma_reserve
);
3828 if (!is_highmem_idx(j
))
3829 nr_kernel_pages
+= realsize
;
3830 nr_all_pages
+= realsize
;
3832 zone
->spanned_pages
= size
;
3833 zone
->present_pages
= realsize
;
3836 zone
->min_unmapped_pages
= (realsize
*sysctl_min_unmapped_ratio
)
3838 zone
->min_slab_pages
= (realsize
* sysctl_min_slab_ratio
) / 100;
3840 zone
->name
= zone_names
[j
];
3841 spin_lock_init(&zone
->lock
);
3842 spin_lock_init(&zone
->lru_lock
);
3843 zone_seqlock_init(zone
);
3844 zone
->zone_pgdat
= pgdat
;
3846 zone
->prev_priority
= DEF_PRIORITY
;
3848 zone_pcp_init(zone
);
3850 INIT_LIST_HEAD(&zone
->lru
[l
].list
);
3851 zone
->reclaim_stat
.nr_saved_scan
[l
] = 0;
3853 zone
->reclaim_stat
.recent_rotated
[0] = 0;
3854 zone
->reclaim_stat
.recent_rotated
[1] = 0;
3855 zone
->reclaim_stat
.recent_scanned
[0] = 0;
3856 zone
->reclaim_stat
.recent_scanned
[1] = 0;
3857 zap_zone_vm_stats(zone
);
3862 set_pageblock_order(pageblock_default_order());
3863 setup_usemap(pgdat
, zone
, size
);
3864 ret
= init_currently_empty_zone(zone
, zone_start_pfn
,
3865 size
, MEMMAP_EARLY
);
3867 memmap_init(size
, nid
, j
, zone_start_pfn
);
3868 zone_start_pfn
+= size
;
3872 static void __init_refok
alloc_node_mem_map(struct pglist_data
*pgdat
)
3874 /* Skip empty nodes */
3875 if (!pgdat
->node_spanned_pages
)
3878 #ifdef CONFIG_FLAT_NODE_MEM_MAP
3879 /* ia64 gets its own node_mem_map, before this, without bootmem */
3880 if (!pgdat
->node_mem_map
) {
3881 unsigned long size
, start
, end
;
3885 * The zone's endpoints aren't required to be MAX_ORDER
3886 * aligned but the node_mem_map endpoints must be in order
3887 * for the buddy allocator to function correctly.
3889 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
3890 end
= pgdat
->node_start_pfn
+ pgdat
->node_spanned_pages
;
3891 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
3892 size
= (end
- start
) * sizeof(struct page
);
3893 map
= alloc_remap(pgdat
->node_id
, size
);
3895 map
= alloc_bootmem_node(pgdat
, size
);
3896 pgdat
->node_mem_map
= map
+ (pgdat
->node_start_pfn
- start
);
3898 #ifndef CONFIG_NEED_MULTIPLE_NODES
3900 * With no DISCONTIG, the global mem_map is just set as node 0's
3902 if (pgdat
== NODE_DATA(0)) {
3903 mem_map
= NODE_DATA(0)->node_mem_map
;
3904 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3905 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
3906 mem_map
-= (pgdat
->node_start_pfn
- ARCH_PFN_OFFSET
);
3907 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
3910 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
3913 void __paginginit
free_area_init_node(int nid
, unsigned long *zones_size
,
3914 unsigned long node_start_pfn
, unsigned long *zholes_size
)
3916 pg_data_t
*pgdat
= NODE_DATA(nid
);
3918 pgdat
->node_id
= nid
;
3919 pgdat
->node_start_pfn
= node_start_pfn
;
3920 calculate_node_totalpages(pgdat
, zones_size
, zholes_size
);
3922 alloc_node_mem_map(pgdat
);
3923 #ifdef CONFIG_FLAT_NODE_MEM_MAP
3924 printk(KERN_DEBUG
"free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
3925 nid
, (unsigned long)pgdat
,
3926 (unsigned long)pgdat
->node_mem_map
);
3929 free_area_init_core(pgdat
, zones_size
, zholes_size
);
3932 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3934 #if MAX_NUMNODES > 1
3936 * Figure out the number of possible node ids.
3938 static void __init
setup_nr_node_ids(void)
3941 unsigned int highest
= 0;
3943 for_each_node_mask(node
, node_possible_map
)
3945 nr_node_ids
= highest
+ 1;
3948 static inline void setup_nr_node_ids(void)
3954 * add_active_range - Register a range of PFNs backed by physical memory
3955 * @nid: The node ID the range resides on
3956 * @start_pfn: The start PFN of the available physical memory
3957 * @end_pfn: The end PFN of the available physical memory
3959 * These ranges are stored in an early_node_map[] and later used by
3960 * free_area_init_nodes() to calculate zone sizes and holes. If the
3961 * range spans a memory hole, it is up to the architecture to ensure
3962 * the memory is not freed by the bootmem allocator. If possible
3963 * the range being registered will be merged with existing ranges.
3965 void __init
add_active_range(unsigned int nid
, unsigned long start_pfn
,
3966 unsigned long end_pfn
)
3970 mminit_dprintk(MMINIT_TRACE
, "memory_register",
3971 "Entering add_active_range(%d, %#lx, %#lx) "
3972 "%d entries of %d used\n",
3973 nid
, start_pfn
, end_pfn
,
3974 nr_nodemap_entries
, MAX_ACTIVE_REGIONS
);
3976 mminit_validate_memmodel_limits(&start_pfn
, &end_pfn
);
3978 /* Merge with existing active regions if possible */
3979 for (i
= 0; i
< nr_nodemap_entries
; i
++) {
3980 if (early_node_map
[i
].nid
!= nid
)
3983 /* Skip if an existing region covers this new one */
3984 if (start_pfn
>= early_node_map
[i
].start_pfn
&&
3985 end_pfn
<= early_node_map
[i
].end_pfn
)
3988 /* Merge forward if suitable */
3989 if (start_pfn
<= early_node_map
[i
].end_pfn
&&
3990 end_pfn
> early_node_map
[i
].end_pfn
) {
3991 early_node_map
[i
].end_pfn
= end_pfn
;
3995 /* Merge backward if suitable */
3996 if (start_pfn
< early_node_map
[i
].start_pfn
&&
3997 end_pfn
>= early_node_map
[i
].start_pfn
) {
3998 early_node_map
[i
].start_pfn
= start_pfn
;
4003 /* Check that early_node_map is large enough */
4004 if (i
>= MAX_ACTIVE_REGIONS
) {
4005 printk(KERN_CRIT
"More than %d memory regions, truncating\n",
4006 MAX_ACTIVE_REGIONS
);
4010 early_node_map
[i
].nid
= nid
;
4011 early_node_map
[i
].start_pfn
= start_pfn
;
4012 early_node_map
[i
].end_pfn
= end_pfn
;
4013 nr_nodemap_entries
= i
+ 1;
4017 * remove_active_range - Shrink an existing registered range of PFNs
4018 * @nid: The node id the range is on that should be shrunk
4019 * @start_pfn: The new PFN of the range
4020 * @end_pfn: The new PFN of the range
4022 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
4023 * The map is kept near the end physical page range that has already been
4024 * registered. This function allows an arch to shrink an existing registered
4027 void __init
remove_active_range(unsigned int nid
, unsigned long start_pfn
,
4028 unsigned long end_pfn
)
4033 printk(KERN_DEBUG
"remove_active_range (%d, %lu, %lu)\n",
4034 nid
, start_pfn
, end_pfn
);
4036 /* Find the old active region end and shrink */
4037 for_each_active_range_index_in_nid(i
, nid
) {
4038 if (early_node_map
[i
].start_pfn
>= start_pfn
&&
4039 early_node_map
[i
].end_pfn
<= end_pfn
) {
4041 early_node_map
[i
].start_pfn
= 0;
4042 early_node_map
[i
].end_pfn
= 0;
4046 if (early_node_map
[i
].start_pfn
< start_pfn
&&
4047 early_node_map
[i
].end_pfn
> start_pfn
) {
4048 unsigned long temp_end_pfn
= early_node_map
[i
].end_pfn
;
4049 early_node_map
[i
].end_pfn
= start_pfn
;
4050 if (temp_end_pfn
> end_pfn
)
4051 add_active_range(nid
, end_pfn
, temp_end_pfn
);
4054 if (early_node_map
[i
].start_pfn
>= start_pfn
&&
4055 early_node_map
[i
].end_pfn
> end_pfn
&&
4056 early_node_map
[i
].start_pfn
< end_pfn
) {
4057 early_node_map
[i
].start_pfn
= end_pfn
;
4065 /* remove the blank ones */
4066 for (i
= nr_nodemap_entries
- 1; i
> 0; i
--) {
4067 if (early_node_map
[i
].nid
!= nid
)
4069 if (early_node_map
[i
].end_pfn
)
4071 /* we found it, get rid of it */
4072 for (j
= i
; j
< nr_nodemap_entries
- 1; j
++)
4073 memcpy(&early_node_map
[j
], &early_node_map
[j
+1],
4074 sizeof(early_node_map
[j
]));
4075 j
= nr_nodemap_entries
- 1;
4076 memset(&early_node_map
[j
], 0, sizeof(early_node_map
[j
]));
4077 nr_nodemap_entries
--;
4082 * remove_all_active_ranges - Remove all currently registered regions
4084 * During discovery, it may be found that a table like SRAT is invalid
4085 * and an alternative discovery method must be used. This function removes
4086 * all currently registered regions.
4088 void __init
remove_all_active_ranges(void)
4090 memset(early_node_map
, 0, sizeof(early_node_map
));
4091 nr_nodemap_entries
= 0;
4094 /* Compare two active node_active_regions */
4095 static int __init
cmp_node_active_region(const void *a
, const void *b
)
4097 struct node_active_region
*arange
= (struct node_active_region
*)a
;
4098 struct node_active_region
*brange
= (struct node_active_region
*)b
;
4100 /* Done this way to avoid overflows */
4101 if (arange
->start_pfn
> brange
->start_pfn
)
4103 if (arange
->start_pfn
< brange
->start_pfn
)
4109 /* sort the node_map by start_pfn */
4110 void __init
sort_node_map(void)
4112 sort(early_node_map
, (size_t)nr_nodemap_entries
,
4113 sizeof(struct node_active_region
),
4114 cmp_node_active_region
, NULL
);
4117 /* Find the lowest pfn for a node */
4118 static unsigned long __init
find_min_pfn_for_node(int nid
)
4121 unsigned long min_pfn
= ULONG_MAX
;
4123 /* Assuming a sorted map, the first range found has the starting pfn */
4124 for_each_active_range_index_in_nid(i
, nid
)
4125 min_pfn
= min(min_pfn
, early_node_map
[i
].start_pfn
);
4127 if (min_pfn
== ULONG_MAX
) {
4129 "Could not find start_pfn for node %d\n", nid
);
4137 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4139 * It returns the minimum PFN based on information provided via
4140 * add_active_range().
4142 unsigned long __init
find_min_pfn_with_active_regions(void)
4144 return find_min_pfn_for_node(MAX_NUMNODES
);
4148 * early_calculate_totalpages()
4149 * Sum pages in active regions for movable zone.
4150 * Populate N_HIGH_MEMORY for calculating usable_nodes.
4152 static unsigned long __init
early_calculate_totalpages(void)
4155 unsigned long totalpages
= 0;
4157 for (i
= 0; i
< nr_nodemap_entries
; i
++) {
4158 unsigned long pages
= early_node_map
[i
].end_pfn
-
4159 early_node_map
[i
].start_pfn
;
4160 totalpages
+= pages
;
4162 node_set_state(early_node_map
[i
].nid
, N_HIGH_MEMORY
);
4168 * Find the PFN the Movable zone begins in each node. Kernel memory
4169 * is spread evenly between nodes as long as the nodes have enough
4170 * memory. When they don't, some nodes will have more kernelcore than
4173 static void __init
find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn
)
4176 unsigned long usable_startpfn
;
4177 unsigned long kernelcore_node
, kernelcore_remaining
;
4178 /* save the state before borrow the nodemask */
4179 nodemask_t saved_node_state
= node_states
[N_HIGH_MEMORY
];
4180 unsigned long totalpages
= early_calculate_totalpages();
4181 int usable_nodes
= nodes_weight(node_states
[N_HIGH_MEMORY
]);
4184 * If movablecore was specified, calculate what size of
4185 * kernelcore that corresponds so that memory usable for
4186 * any allocation type is evenly spread. If both kernelcore
4187 * and movablecore are specified, then the value of kernelcore
4188 * will be used for required_kernelcore if it's greater than
4189 * what movablecore would have allowed.
4191 if (required_movablecore
) {
4192 unsigned long corepages
;
4195 * Round-up so that ZONE_MOVABLE is at least as large as what
4196 * was requested by the user
4198 required_movablecore
=
4199 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
4200 corepages
= totalpages
- required_movablecore
;
4202 required_kernelcore
= max(required_kernelcore
, corepages
);
4205 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4206 if (!required_kernelcore
)
4209 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4210 find_usable_zone_for_movable();
4211 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
4214 /* Spread kernelcore memory as evenly as possible throughout nodes */
4215 kernelcore_node
= required_kernelcore
/ usable_nodes
;
4216 for_each_node_state(nid
, N_HIGH_MEMORY
) {
4218 * Recalculate kernelcore_node if the division per node
4219 * now exceeds what is necessary to satisfy the requested
4220 * amount of memory for the kernel
4222 if (required_kernelcore
< kernelcore_node
)
4223 kernelcore_node
= required_kernelcore
/ usable_nodes
;
4226 * As the map is walked, we track how much memory is usable
4227 * by the kernel using kernelcore_remaining. When it is
4228 * 0, the rest of the node is usable by ZONE_MOVABLE
4230 kernelcore_remaining
= kernelcore_node
;
4232 /* Go through each range of PFNs within this node */
4233 for_each_active_range_index_in_nid(i
, nid
) {
4234 unsigned long start_pfn
, end_pfn
;
4235 unsigned long size_pages
;
4237 start_pfn
= max(early_node_map
[i
].start_pfn
,
4238 zone_movable_pfn
[nid
]);
4239 end_pfn
= early_node_map
[i
].end_pfn
;
4240 if (start_pfn
>= end_pfn
)
4243 /* Account for what is only usable for kernelcore */
4244 if (start_pfn
< usable_startpfn
) {
4245 unsigned long kernel_pages
;
4246 kernel_pages
= min(end_pfn
, usable_startpfn
)
4249 kernelcore_remaining
-= min(kernel_pages
,
4250 kernelcore_remaining
);
4251 required_kernelcore
-= min(kernel_pages
,
4252 required_kernelcore
);
4254 /* Continue if range is now fully accounted */
4255 if (end_pfn
<= usable_startpfn
) {
4258 * Push zone_movable_pfn to the end so
4259 * that if we have to rebalance
4260 * kernelcore across nodes, we will
4261 * not double account here
4263 zone_movable_pfn
[nid
] = end_pfn
;
4266 start_pfn
= usable_startpfn
;
4270 * The usable PFN range for ZONE_MOVABLE is from
4271 * start_pfn->end_pfn. Calculate size_pages as the
4272 * number of pages used as kernelcore
4274 size_pages
= end_pfn
- start_pfn
;
4275 if (size_pages
> kernelcore_remaining
)
4276 size_pages
= kernelcore_remaining
;
4277 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
4280 * Some kernelcore has been met, update counts and
4281 * break if the kernelcore for this node has been
4284 required_kernelcore
-= min(required_kernelcore
,
4286 kernelcore_remaining
-= size_pages
;
4287 if (!kernelcore_remaining
)
4293 * If there is still required_kernelcore, we do another pass with one
4294 * less node in the count. This will push zone_movable_pfn[nid] further
4295 * along on the nodes that still have memory until kernelcore is
4299 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
4302 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4303 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
4304 zone_movable_pfn
[nid
] =
4305 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
4308 /* restore the node_state */
4309 node_states
[N_HIGH_MEMORY
] = saved_node_state
;
4312 /* Any regular memory on that node ? */
4313 static void check_for_regular_memory(pg_data_t
*pgdat
)
4315 #ifdef CONFIG_HIGHMEM
4316 enum zone_type zone_type
;
4318 for (zone_type
= 0; zone_type
<= ZONE_NORMAL
; zone_type
++) {
4319 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
4320 if (zone
->present_pages
)
4321 node_set_state(zone_to_nid(zone
), N_NORMAL_MEMORY
);
4327 * free_area_init_nodes - Initialise all pg_data_t and zone data
4328 * @max_zone_pfn: an array of max PFNs for each zone
4330 * This will call free_area_init_node() for each active node in the system.
4331 * Using the page ranges provided by add_active_range(), the size of each
4332 * zone in each node and their holes is calculated. If the maximum PFN
4333 * between two adjacent zones match, it is assumed that the zone is empty.
4334 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4335 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4336 * starts where the previous one ended. For example, ZONE_DMA32 starts
4337 * at arch_max_dma_pfn.
4339 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
4344 /* Sort early_node_map as initialisation assumes it is sorted */
4347 /* Record where the zone boundaries are */
4348 memset(arch_zone_lowest_possible_pfn
, 0,
4349 sizeof(arch_zone_lowest_possible_pfn
));
4350 memset(arch_zone_highest_possible_pfn
, 0,
4351 sizeof(arch_zone_highest_possible_pfn
));
4352 arch_zone_lowest_possible_pfn
[0] = find_min_pfn_with_active_regions();
4353 arch_zone_highest_possible_pfn
[0] = max_zone_pfn
[0];
4354 for (i
= 1; i
< MAX_NR_ZONES
; i
++) {
4355 if (i
== ZONE_MOVABLE
)
4357 arch_zone_lowest_possible_pfn
[i
] =
4358 arch_zone_highest_possible_pfn
[i
-1];
4359 arch_zone_highest_possible_pfn
[i
] =
4360 max(max_zone_pfn
[i
], arch_zone_lowest_possible_pfn
[i
]);
4362 arch_zone_lowest_possible_pfn
[ZONE_MOVABLE
] = 0;
4363 arch_zone_highest_possible_pfn
[ZONE_MOVABLE
] = 0;
4365 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4366 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
4367 find_zone_movable_pfns_for_nodes(zone_movable_pfn
);
4369 /* Print out the zone ranges */
4370 printk("Zone PFN ranges:\n");
4371 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
4372 if (i
== ZONE_MOVABLE
)
4374 printk(" %-8s %0#10lx -> %0#10lx\n",
4376 arch_zone_lowest_possible_pfn
[i
],
4377 arch_zone_highest_possible_pfn
[i
]);
4380 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4381 printk("Movable zone start PFN for each node\n");
4382 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
4383 if (zone_movable_pfn
[i
])
4384 printk(" Node %d: %lu\n", i
, zone_movable_pfn
[i
]);
4387 /* Print out the early_node_map[] */
4388 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries
);
4389 for (i
= 0; i
< nr_nodemap_entries
; i
++)
4390 printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map
[i
].nid
,
4391 early_node_map
[i
].start_pfn
,
4392 early_node_map
[i
].end_pfn
);
4394 /* Initialise every node */
4395 mminit_verify_pageflags_layout();
4396 setup_nr_node_ids();
4397 for_each_online_node(nid
) {
4398 pg_data_t
*pgdat
= NODE_DATA(nid
);
4399 free_area_init_node(nid
, NULL
,
4400 find_min_pfn_for_node(nid
), NULL
);
4402 /* Any memory on that node */
4403 if (pgdat
->node_present_pages
)
4404 node_set_state(nid
, N_HIGH_MEMORY
);
4405 check_for_regular_memory(pgdat
);
4409 static int __init
cmdline_parse_core(char *p
, unsigned long *core
)
4411 unsigned long long coremem
;
4415 coremem
= memparse(p
, &p
);
4416 *core
= coremem
>> PAGE_SHIFT
;
4418 /* Paranoid check that UL is enough for the coremem value */
4419 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
4425 * kernelcore=size sets the amount of memory for use for allocations that
4426 * cannot be reclaimed or migrated.
4428 static int __init
cmdline_parse_kernelcore(char *p
)
4430 return cmdline_parse_core(p
, &required_kernelcore
);
4434 * movablecore=size sets the amount of memory for use for allocations that
4435 * can be reclaimed or migrated.
4437 static int __init
cmdline_parse_movablecore(char *p
)
4439 return cmdline_parse_core(p
, &required_movablecore
);
4442 early_param("kernelcore", cmdline_parse_kernelcore
);
4443 early_param("movablecore", cmdline_parse_movablecore
);
4445 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4448 * set_dma_reserve - set the specified number of pages reserved in the first zone
4449 * @new_dma_reserve: The number of pages to mark reserved
4451 * The per-cpu batchsize and zone watermarks are determined by present_pages.
4452 * In the DMA zone, a significant percentage may be consumed by kernel image
4453 * and other unfreeable allocations which can skew the watermarks badly. This
4454 * function may optionally be used to account for unfreeable pages in the
4455 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
4456 * smaller per-cpu batchsize.
4458 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
4460 dma_reserve
= new_dma_reserve
;
4463 #ifndef CONFIG_NEED_MULTIPLE_NODES
4464 struct pglist_data __refdata contig_page_data
= {
4465 #ifndef CONFIG_NO_BOOTMEM
4466 .bdata
= &bootmem_node_data
[0]
4469 EXPORT_SYMBOL(contig_page_data
);
4472 void __init
free_area_init(unsigned long *zones_size
)
4474 free_area_init_node(0, zones_size
,
4475 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
4478 static int page_alloc_cpu_notify(struct notifier_block
*self
,
4479 unsigned long action
, void *hcpu
)
4481 int cpu
= (unsigned long)hcpu
;
4483 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
) {
4487 * Spill the event counters of the dead processor
4488 * into the current processors event counters.
4489 * This artificially elevates the count of the current
4492 vm_events_fold_cpu(cpu
);
4495 * Zero the differential counters of the dead processor
4496 * so that the vm statistics are consistent.
4498 * This is only okay since the processor is dead and cannot
4499 * race with what we are doing.
4501 refresh_cpu_vm_stats(cpu
);
4506 void __init
page_alloc_init(void)
4508 hotcpu_notifier(page_alloc_cpu_notify
, 0);
4512 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
4513 * or min_free_kbytes changes.
4515 static void calculate_totalreserve_pages(void)
4517 struct pglist_data
*pgdat
;
4518 unsigned long reserve_pages
= 0;
4519 enum zone_type i
, j
;
4521 for_each_online_pgdat(pgdat
) {
4522 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
4523 struct zone
*zone
= pgdat
->node_zones
+ i
;
4524 unsigned long max
= 0;
4526 /* Find valid and maximum lowmem_reserve in the zone */
4527 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
4528 if (zone
->lowmem_reserve
[j
] > max
)
4529 max
= zone
->lowmem_reserve
[j
];
4532 /* we treat the high watermark as reserved pages. */
4533 max
+= high_wmark_pages(zone
);
4535 if (max
> zone
->present_pages
)
4536 max
= zone
->present_pages
;
4537 reserve_pages
+= max
;
4540 totalreserve_pages
= reserve_pages
;
4544 * setup_per_zone_lowmem_reserve - called whenever
4545 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
4546 * has a correct pages reserved value, so an adequate number of
4547 * pages are left in the zone after a successful __alloc_pages().
4549 static void setup_per_zone_lowmem_reserve(void)
4551 struct pglist_data
*pgdat
;
4552 enum zone_type j
, idx
;
4554 for_each_online_pgdat(pgdat
) {
4555 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
4556 struct zone
*zone
= pgdat
->node_zones
+ j
;
4557 unsigned long present_pages
= zone
->present_pages
;
4559 zone
->lowmem_reserve
[j
] = 0;
4563 struct zone
*lower_zone
;
4567 if (sysctl_lowmem_reserve_ratio
[idx
] < 1)
4568 sysctl_lowmem_reserve_ratio
[idx
] = 1;
4570 lower_zone
= pgdat
->node_zones
+ idx
;
4571 lower_zone
->lowmem_reserve
[j
] = present_pages
/
4572 sysctl_lowmem_reserve_ratio
[idx
];
4573 present_pages
+= lower_zone
->present_pages
;
4578 /* update totalreserve_pages */
4579 calculate_totalreserve_pages();
4583 * setup_per_zone_wmarks - called when min_free_kbytes changes
4584 * or when memory is hot-{added|removed}
4586 * Ensures that the watermark[min,low,high] values for each zone are set
4587 * correctly with respect to min_free_kbytes.
4589 void setup_per_zone_wmarks(void)
4591 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
4592 unsigned long lowmem_pages
= 0;
4594 unsigned long flags
;
4596 /* Calculate total number of !ZONE_HIGHMEM pages */
4597 for_each_zone(zone
) {
4598 if (!is_highmem(zone
))
4599 lowmem_pages
+= zone
->present_pages
;
4602 for_each_zone(zone
) {
4605 spin_lock_irqsave(&zone
->lock
, flags
);
4606 tmp
= (u64
)pages_min
* zone
->present_pages
;
4607 do_div(tmp
, lowmem_pages
);
4608 if (is_highmem(zone
)) {
4610 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
4611 * need highmem pages, so cap pages_min to a small
4614 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
4615 * deltas controls asynch page reclaim, and so should
4616 * not be capped for highmem.
4620 min_pages
= zone
->present_pages
/ 1024;
4621 if (min_pages
< SWAP_CLUSTER_MAX
)
4622 min_pages
= SWAP_CLUSTER_MAX
;
4623 if (min_pages
> 128)
4625 zone
->watermark
[WMARK_MIN
] = min_pages
;
4628 * If it's a lowmem zone, reserve a number of pages
4629 * proportionate to the zone's size.
4631 zone
->watermark
[WMARK_MIN
] = tmp
;
4634 zone
->watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + (tmp
>> 2);
4635 zone
->watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + (tmp
>> 1);
4636 setup_zone_migrate_reserve(zone
);
4637 spin_unlock_irqrestore(&zone
->lock
, flags
);
4640 /* update totalreserve_pages */
4641 calculate_totalreserve_pages();
4645 * The inactive anon list should be small enough that the VM never has to
4646 * do too much work, but large enough that each inactive page has a chance
4647 * to be referenced again before it is swapped out.
4649 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
4650 * INACTIVE_ANON pages on this zone's LRU, maintained by the
4651 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
4652 * the anonymous pages are kept on the inactive list.
4655 * memory ratio inactive anon
4656 * -------------------------------------
4665 void calculate_zone_inactive_ratio(struct zone
*zone
)
4667 unsigned int gb
, ratio
;
4669 /* Zone size in gigabytes */
4670 gb
= zone
->present_pages
>> (30 - PAGE_SHIFT
);
4672 ratio
= int_sqrt(10 * gb
);
4676 zone
->inactive_ratio
= ratio
;
4679 static void __init
setup_per_zone_inactive_ratio(void)
4684 calculate_zone_inactive_ratio(zone
);
4688 * Initialise min_free_kbytes.
4690 * For small machines we want it small (128k min). For large machines
4691 * we want it large (64MB max). But it is not linear, because network
4692 * bandwidth does not increase linearly with machine size. We use
4694 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
4695 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
4711 static int __init
init_per_zone_wmark_min(void)
4713 unsigned long lowmem_kbytes
;
4715 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
4717 min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
4718 if (min_free_kbytes
< 128)
4719 min_free_kbytes
= 128;
4720 if (min_free_kbytes
> 65536)
4721 min_free_kbytes
= 65536;
4722 setup_per_zone_wmarks();
4723 setup_per_zone_lowmem_reserve();
4724 setup_per_zone_inactive_ratio();
4727 module_init(init_per_zone_wmark_min
)
4730 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
4731 * that we can call two helper functions whenever min_free_kbytes
4734 int min_free_kbytes_sysctl_handler(ctl_table
*table
, int write
,
4735 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
4737 proc_dointvec(table
, write
, buffer
, length
, ppos
);
4739 setup_per_zone_wmarks();
4744 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table
*table
, int write
,
4745 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
4750 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
4755 zone
->min_unmapped_pages
= (zone
->present_pages
*
4756 sysctl_min_unmapped_ratio
) / 100;
4760 int sysctl_min_slab_ratio_sysctl_handler(ctl_table
*table
, int write
,
4761 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
4766 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
4771 zone
->min_slab_pages
= (zone
->present_pages
*
4772 sysctl_min_slab_ratio
) / 100;
4778 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
4779 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
4780 * whenever sysctl_lowmem_reserve_ratio changes.
4782 * The reserve ratio obviously has absolutely no relation with the
4783 * minimum watermarks. The lowmem reserve ratio can only make sense
4784 * if in function of the boot time zone sizes.
4786 int lowmem_reserve_ratio_sysctl_handler(ctl_table
*table
, int write
,
4787 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
4789 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
4790 setup_per_zone_lowmem_reserve();
4795 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
4796 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
4797 * can have before it gets flushed back to buddy allocator.
4800 int percpu_pagelist_fraction_sysctl_handler(ctl_table
*table
, int write
,
4801 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
4807 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
4808 if (!write
|| (ret
== -EINVAL
))
4810 for_each_populated_zone(zone
) {
4811 for_each_possible_cpu(cpu
) {
4813 high
= zone
->present_pages
/ percpu_pagelist_fraction
;
4814 setup_pagelist_highmark(
4815 per_cpu_ptr(zone
->pageset
, cpu
), high
);
4821 int hashdist
= HASHDIST_DEFAULT
;
4824 static int __init
set_hashdist(char *str
)
4828 hashdist
= simple_strtoul(str
, &str
, 0);
4831 __setup("hashdist=", set_hashdist
);
4835 * allocate a large system hash table from bootmem
4836 * - it is assumed that the hash table must contain an exact power-of-2
4837 * quantity of entries
4838 * - limit is the number of hash buckets, not the total allocation size
4840 void *__init
alloc_large_system_hash(const char *tablename
,
4841 unsigned long bucketsize
,
4842 unsigned long numentries
,
4845 unsigned int *_hash_shift
,
4846 unsigned int *_hash_mask
,
4847 unsigned long limit
)
4849 unsigned long long max
= limit
;
4850 unsigned long log2qty
, size
;
4853 /* allow the kernel cmdline to have a say */
4855 /* round applicable memory size up to nearest megabyte */
4856 numentries
= nr_kernel_pages
;
4857 numentries
+= (1UL << (20 - PAGE_SHIFT
)) - 1;
4858 numentries
>>= 20 - PAGE_SHIFT
;
4859 numentries
<<= 20 - PAGE_SHIFT
;
4861 /* limit to 1 bucket per 2^scale bytes of low memory */
4862 if (scale
> PAGE_SHIFT
)
4863 numentries
>>= (scale
- PAGE_SHIFT
);
4865 numentries
<<= (PAGE_SHIFT
- scale
);
4867 /* Make sure we've got at least a 0-order allocation.. */
4868 if (unlikely(flags
& HASH_SMALL
)) {
4869 /* Makes no sense without HASH_EARLY */
4870 WARN_ON(!(flags
& HASH_EARLY
));
4871 if (!(numentries
>> *_hash_shift
)) {
4872 numentries
= 1UL << *_hash_shift
;
4873 BUG_ON(!numentries
);
4875 } else if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
4876 numentries
= PAGE_SIZE
/ bucketsize
;
4878 numentries
= roundup_pow_of_two(numentries
);
4880 /* limit allocation size to 1/16 total memory by default */
4882 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
4883 do_div(max
, bucketsize
);
4886 if (numentries
> max
)
4889 log2qty
= ilog2(numentries
);
4892 size
= bucketsize
<< log2qty
;
4893 if (flags
& HASH_EARLY
)
4894 table
= alloc_bootmem_nopanic(size
);
4896 table
= __vmalloc(size
, GFP_ATOMIC
, PAGE_KERNEL
);
4899 * If bucketsize is not a power-of-two, we may free
4900 * some pages at the end of hash table which
4901 * alloc_pages_exact() automatically does
4903 if (get_order(size
) < MAX_ORDER
) {
4904 table
= alloc_pages_exact(size
, GFP_ATOMIC
);
4905 kmemleak_alloc(table
, size
, 1, GFP_ATOMIC
);
4908 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
4911 panic("Failed to allocate %s hash table\n", tablename
);
4913 printk(KERN_INFO
"%s hash table entries: %d (order: %d, %lu bytes)\n",
4916 ilog2(size
) - PAGE_SHIFT
,
4920 *_hash_shift
= log2qty
;
4922 *_hash_mask
= (1 << log2qty
) - 1;
4927 /* Return a pointer to the bitmap storing bits affecting a block of pages */
4928 static inline unsigned long *get_pageblock_bitmap(struct zone
*zone
,
4931 #ifdef CONFIG_SPARSEMEM
4932 return __pfn_to_section(pfn
)->pageblock_flags
;
4934 return zone
->pageblock_flags
;
4935 #endif /* CONFIG_SPARSEMEM */
4938 static inline int pfn_to_bitidx(struct zone
*zone
, unsigned long pfn
)
4940 #ifdef CONFIG_SPARSEMEM
4941 pfn
&= (PAGES_PER_SECTION
-1);
4942 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
4944 pfn
= pfn
- zone
->zone_start_pfn
;
4945 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
4946 #endif /* CONFIG_SPARSEMEM */
4950 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
4951 * @page: The page within the block of interest
4952 * @start_bitidx: The first bit of interest to retrieve
4953 * @end_bitidx: The last bit of interest
4954 * returns pageblock_bits flags
4956 unsigned long get_pageblock_flags_group(struct page
*page
,
4957 int start_bitidx
, int end_bitidx
)
4960 unsigned long *bitmap
;
4961 unsigned long pfn
, bitidx
;
4962 unsigned long flags
= 0;
4963 unsigned long value
= 1;
4965 zone
= page_zone(page
);
4966 pfn
= page_to_pfn(page
);
4967 bitmap
= get_pageblock_bitmap(zone
, pfn
);
4968 bitidx
= pfn_to_bitidx(zone
, pfn
);
4970 for (; start_bitidx
<= end_bitidx
; start_bitidx
++, value
<<= 1)
4971 if (test_bit(bitidx
+ start_bitidx
, bitmap
))
4978 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
4979 * @page: The page within the block of interest
4980 * @start_bitidx: The first bit of interest
4981 * @end_bitidx: The last bit of interest
4982 * @flags: The flags to set
4984 void set_pageblock_flags_group(struct page
*page
, unsigned long flags
,
4985 int start_bitidx
, int end_bitidx
)
4988 unsigned long *bitmap
;
4989 unsigned long pfn
, bitidx
;
4990 unsigned long value
= 1;
4992 zone
= page_zone(page
);
4993 pfn
= page_to_pfn(page
);
4994 bitmap
= get_pageblock_bitmap(zone
, pfn
);
4995 bitidx
= pfn_to_bitidx(zone
, pfn
);
4996 VM_BUG_ON(pfn
< zone
->zone_start_pfn
);
4997 VM_BUG_ON(pfn
>= zone
->zone_start_pfn
+ zone
->spanned_pages
);
4999 for (; start_bitidx
<= end_bitidx
; start_bitidx
++, value
<<= 1)
5001 __set_bit(bitidx
+ start_bitidx
, bitmap
);
5003 __clear_bit(bitidx
+ start_bitidx
, bitmap
);
5007 * This is designed as sub function...plz see page_isolation.c also.
5008 * set/clear page block's type to be ISOLATE.
5009 * page allocater never alloc memory from ISOLATE block.
5012 int set_migratetype_isolate(struct page
*page
)
5015 struct page
*curr_page
;
5016 unsigned long flags
, pfn
, iter
;
5017 unsigned long immobile
= 0;
5018 struct memory_isolate_notify arg
;
5023 zone
= page_zone(page
);
5024 zone_idx
= zone_idx(zone
);
5026 spin_lock_irqsave(&zone
->lock
, flags
);
5027 if (get_pageblock_migratetype(page
) == MIGRATE_MOVABLE
||
5028 zone_idx
== ZONE_MOVABLE
) {
5033 pfn
= page_to_pfn(page
);
5034 arg
.start_pfn
= pfn
;
5035 arg
.nr_pages
= pageblock_nr_pages
;
5036 arg
.pages_found
= 0;
5039 * It may be possible to isolate a pageblock even if the
5040 * migratetype is not MIGRATE_MOVABLE. The memory isolation
5041 * notifier chain is used by balloon drivers to return the
5042 * number of pages in a range that are held by the balloon
5043 * driver to shrink memory. If all the pages are accounted for
5044 * by balloons, are free, or on the LRU, isolation can continue.
5045 * Later, for example, when memory hotplug notifier runs, these
5046 * pages reported as "can be isolated" should be isolated(freed)
5047 * by the balloon driver through the memory notifier chain.
5049 notifier_ret
= memory_isolate_notify(MEM_ISOLATE_COUNT
, &arg
);
5050 notifier_ret
= notifier_to_errno(notifier_ret
);
5051 if (notifier_ret
|| !arg
.pages_found
)
5054 for (iter
= pfn
; iter
< (pfn
+ pageblock_nr_pages
); iter
++) {
5055 if (!pfn_valid_within(pfn
))
5058 curr_page
= pfn_to_page(iter
);
5059 if (!page_count(curr_page
) || PageLRU(curr_page
))
5065 if (arg
.pages_found
== immobile
)
5070 set_pageblock_migratetype(page
, MIGRATE_ISOLATE
);
5071 move_freepages_block(zone
, page
, MIGRATE_ISOLATE
);
5074 spin_unlock_irqrestore(&zone
->lock
, flags
);
5080 void unset_migratetype_isolate(struct page
*page
)
5083 unsigned long flags
;
5084 zone
= page_zone(page
);
5085 spin_lock_irqsave(&zone
->lock
, flags
);
5086 if (get_pageblock_migratetype(page
) != MIGRATE_ISOLATE
)
5088 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
5089 move_freepages_block(zone
, page
, MIGRATE_MOVABLE
);
5091 spin_unlock_irqrestore(&zone
->lock
, flags
);
5094 #ifdef CONFIG_MEMORY_HOTREMOVE
5096 * All pages in the range must be isolated before calling this.
5099 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
5105 unsigned long flags
;
5106 /* find the first valid pfn */
5107 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
5112 zone
= page_zone(pfn_to_page(pfn
));
5113 spin_lock_irqsave(&zone
->lock
, flags
);
5115 while (pfn
< end_pfn
) {
5116 if (!pfn_valid(pfn
)) {
5120 page
= pfn_to_page(pfn
);
5121 BUG_ON(page_count(page
));
5122 BUG_ON(!PageBuddy(page
));
5123 order
= page_order(page
);
5124 #ifdef CONFIG_DEBUG_VM
5125 printk(KERN_INFO
"remove from free list %lx %d %lx\n",
5126 pfn
, 1 << order
, end_pfn
);
5128 list_del(&page
->lru
);
5129 rmv_page_order(page
);
5130 zone
->free_area
[order
].nr_free
--;
5131 __mod_zone_page_state(zone
, NR_FREE_PAGES
,
5133 for (i
= 0; i
< (1 << order
); i
++)
5134 SetPageReserved((page
+i
));
5135 pfn
+= (1 << order
);
5137 spin_unlock_irqrestore(&zone
->lock
, flags
);
5141 #ifdef CONFIG_MEMORY_FAILURE
5142 bool is_free_buddy_page(struct page
*page
)
5144 struct zone
*zone
= page_zone(page
);
5145 unsigned long pfn
= page_to_pfn(page
);
5146 unsigned long flags
;
5149 spin_lock_irqsave(&zone
->lock
, flags
);
5150 for (order
= 0; order
< MAX_ORDER
; order
++) {
5151 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
5153 if (PageBuddy(page_head
) && page_order(page_head
) >= order
)
5156 spin_unlock_irqrestore(&zone
->lock
, flags
);
5158 return order
< MAX_ORDER
;