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/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kmemcheck.h>
28 #include <linux/module.h>
29 #include <linux/suspend.h>
30 #include <linux/pagevec.h>
31 #include <linux/blkdev.h>
32 #include <linux/slab.h>
33 #include <linux/oom.h>
34 #include <linux/notifier.h>
35 #include <linux/topology.h>
36 #include <linux/sysctl.h>
37 #include <linux/cpu.h>
38 #include <linux/cpuset.h>
39 #include <linux/memory_hotplug.h>
40 #include <linux/nodemask.h>
41 #include <linux/vmalloc.h>
42 #include <linux/mempolicy.h>
43 #include <linux/stop_machine.h>
44 #include <linux/sort.h>
45 #include <linux/pfn.h>
46 #include <linux/backing-dev.h>
47 #include <linux/fault-inject.h>
48 #include <linux/page-isolation.h>
49 #include <linux/page_cgroup.h>
50 #include <linux/debugobjects.h>
51 #include <linux/kmemleak.h>
52 #include <linux/memory.h>
53 #include <linux/compaction.h>
54 #include <trace/events/kmem.h>
55 #include <linux/ftrace_event.h>
56 #include <linux/memcontrol.h>
57 #include <linux/prefetch.h>
59 #include <asm/tlbflush.h>
60 #include <asm/div64.h>
63 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
64 DEFINE_PER_CPU(int, numa_node
);
65 EXPORT_PER_CPU_SYMBOL(numa_node
);
68 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
70 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
71 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
72 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
73 * defined in <linux/topology.h>.
75 DEFINE_PER_CPU(int, _numa_mem_
); /* Kernel "local memory" node */
76 EXPORT_PER_CPU_SYMBOL(_numa_mem_
);
80 * Array of node states.
82 nodemask_t node_states
[NR_NODE_STATES
] __read_mostly
= {
83 [N_POSSIBLE
] = NODE_MASK_ALL
,
84 [N_ONLINE
] = { { [0] = 1UL } },
86 [N_NORMAL_MEMORY
] = { { [0] = 1UL } },
88 [N_HIGH_MEMORY
] = { { [0] = 1UL } },
90 [N_CPU
] = { { [0] = 1UL } },
93 EXPORT_SYMBOL(node_states
);
95 unsigned long totalram_pages __read_mostly
;
96 unsigned long totalreserve_pages __read_mostly
;
97 int percpu_pagelist_fraction
;
98 gfp_t gfp_allowed_mask __read_mostly
= GFP_BOOT_MASK
;
100 #ifdef CONFIG_PM_SLEEP
102 * The following functions are used by the suspend/hibernate code to temporarily
103 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
104 * while devices are suspended. To avoid races with the suspend/hibernate code,
105 * they should always be called with pm_mutex held (gfp_allowed_mask also should
106 * only be modified with pm_mutex held, unless the suspend/hibernate code is
107 * guaranteed not to run in parallel with that modification).
110 static gfp_t saved_gfp_mask
;
112 void pm_restore_gfp_mask(void)
114 WARN_ON(!mutex_is_locked(&pm_mutex
));
115 if (saved_gfp_mask
) {
116 gfp_allowed_mask
= saved_gfp_mask
;
121 void pm_restrict_gfp_mask(void)
123 WARN_ON(!mutex_is_locked(&pm_mutex
));
124 WARN_ON(saved_gfp_mask
);
125 saved_gfp_mask
= gfp_allowed_mask
;
126 gfp_allowed_mask
&= ~GFP_IOFS
;
128 #endif /* CONFIG_PM_SLEEP */
130 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
131 int pageblock_order __read_mostly
;
134 static void __free_pages_ok(struct page
*page
, unsigned int order
);
137 * results with 256, 32 in the lowmem_reserve sysctl:
138 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
139 * 1G machine -> (16M dma, 784M normal, 224M high)
140 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
141 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
142 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
144 * TBD: should special case ZONE_DMA32 machines here - in those we normally
145 * don't need any ZONE_NORMAL reservation
147 int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
-1] = {
148 #ifdef CONFIG_ZONE_DMA
151 #ifdef CONFIG_ZONE_DMA32
154 #ifdef CONFIG_HIGHMEM
160 EXPORT_SYMBOL(totalram_pages
);
162 static char * const zone_names
[MAX_NR_ZONES
] = {
163 #ifdef CONFIG_ZONE_DMA
166 #ifdef CONFIG_ZONE_DMA32
170 #ifdef CONFIG_HIGHMEM
176 int min_free_kbytes
= 1024;
178 static unsigned long __meminitdata nr_kernel_pages
;
179 static unsigned long __meminitdata nr_all_pages
;
180 static unsigned long __meminitdata dma_reserve
;
182 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
184 * MAX_ACTIVE_REGIONS determines the maximum number of distinct
185 * ranges of memory (RAM) that may be registered with add_active_range().
186 * Ranges passed to add_active_range() will be merged if possible
187 * so the number of times add_active_range() can be called is
188 * related to the number of nodes and the number of holes
190 #ifdef CONFIG_MAX_ACTIVE_REGIONS
191 /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */
192 #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS
194 #if MAX_NUMNODES >= 32
195 /* If there can be many nodes, allow up to 50 holes per node */
196 #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50)
198 /* By default, allow up to 256 distinct regions */
199 #define MAX_ACTIVE_REGIONS 256
203 static struct node_active_region __meminitdata early_node_map
[MAX_ACTIVE_REGIONS
];
204 static int __meminitdata nr_nodemap_entries
;
205 static unsigned long __meminitdata arch_zone_lowest_possible_pfn
[MAX_NR_ZONES
];
206 static unsigned long __meminitdata arch_zone_highest_possible_pfn
[MAX_NR_ZONES
];
207 static unsigned long __initdata required_kernelcore
;
208 static unsigned long __initdata required_movablecore
;
209 static unsigned long __meminitdata zone_movable_pfn
[MAX_NUMNODES
];
211 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
213 EXPORT_SYMBOL(movable_zone
);
214 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
217 int nr_node_ids __read_mostly
= MAX_NUMNODES
;
218 int nr_online_nodes __read_mostly
= 1;
219 EXPORT_SYMBOL(nr_node_ids
);
220 EXPORT_SYMBOL(nr_online_nodes
);
223 int page_group_by_mobility_disabled __read_mostly
;
225 static void set_pageblock_migratetype(struct page
*page
, int migratetype
)
228 if (unlikely(page_group_by_mobility_disabled
))
229 migratetype
= MIGRATE_UNMOVABLE
;
231 set_pageblock_flags_group(page
, (unsigned long)migratetype
,
232 PB_migrate
, PB_migrate_end
);
235 bool oom_killer_disabled __read_mostly
;
237 #ifdef CONFIG_DEBUG_VM
238 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
242 unsigned long pfn
= page_to_pfn(page
);
245 seq
= zone_span_seqbegin(zone
);
246 if (pfn
>= zone
->zone_start_pfn
+ zone
->spanned_pages
)
248 else if (pfn
< zone
->zone_start_pfn
)
250 } while (zone_span_seqretry(zone
, seq
));
255 static int page_is_consistent(struct zone
*zone
, struct page
*page
)
257 if (!pfn_valid_within(page_to_pfn(page
)))
259 if (zone
!= page_zone(page
))
265 * Temporary debugging check for pages not lying within a given zone.
267 static int bad_range(struct zone
*zone
, struct page
*page
)
269 if (page_outside_zone_boundaries(zone
, page
))
271 if (!page_is_consistent(zone
, page
))
277 static inline int bad_range(struct zone
*zone
, struct page
*page
)
283 static void bad_page(struct page
*page
)
285 static unsigned long resume
;
286 static unsigned long nr_shown
;
287 static unsigned long nr_unshown
;
289 /* Don't complain about poisoned pages */
290 if (PageHWPoison(page
)) {
291 reset_page_mapcount(page
); /* remove PageBuddy */
296 * Allow a burst of 60 reports, then keep quiet for that minute;
297 * or allow a steady drip of one report per second.
299 if (nr_shown
== 60) {
300 if (time_before(jiffies
, resume
)) {
306 "BUG: Bad page state: %lu messages suppressed\n",
313 resume
= jiffies
+ 60 * HZ
;
315 printk(KERN_ALERT
"BUG: Bad page state in process %s pfn:%05lx\n",
316 current
->comm
, page_to_pfn(page
));
321 /* Leave bad fields for debug, except PageBuddy could make trouble */
322 reset_page_mapcount(page
); /* remove PageBuddy */
323 add_taint(TAINT_BAD_PAGE
);
327 * Higher-order pages are called "compound pages". They are structured thusly:
329 * The first PAGE_SIZE page is called the "head page".
331 * The remaining PAGE_SIZE pages are called "tail pages".
333 * All pages have PG_compound set. All pages have their ->private pointing at
334 * the head page (even the head page has this).
336 * The first tail page's ->lru.next holds the address of the compound page's
337 * put_page() function. Its ->lru.prev holds the order of allocation.
338 * This usage means that zero-order pages may not be compound.
341 static void free_compound_page(struct page
*page
)
343 __free_pages_ok(page
, compound_order(page
));
346 void prep_compound_page(struct page
*page
, unsigned long order
)
349 int nr_pages
= 1 << order
;
351 set_compound_page_dtor(page
, free_compound_page
);
352 set_compound_order(page
, order
);
354 for (i
= 1; i
< nr_pages
; i
++) {
355 struct page
*p
= page
+ i
;
358 p
->first_page
= page
;
362 /* update __split_huge_page_refcount if you change this function */
363 static int destroy_compound_page(struct page
*page
, unsigned long order
)
366 int nr_pages
= 1 << order
;
369 if (unlikely(compound_order(page
) != order
) ||
370 unlikely(!PageHead(page
))) {
375 __ClearPageHead(page
);
377 for (i
= 1; i
< nr_pages
; i
++) {
378 struct page
*p
= page
+ i
;
380 if (unlikely(!PageTail(p
) || (p
->first_page
!= page
))) {
390 static inline void prep_zero_page(struct page
*page
, int order
, gfp_t gfp_flags
)
395 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
396 * and __GFP_HIGHMEM from hard or soft interrupt context.
398 VM_BUG_ON((gfp_flags
& __GFP_HIGHMEM
) && in_interrupt());
399 for (i
= 0; i
< (1 << order
); i
++)
400 clear_highpage(page
+ i
);
403 static inline void set_page_order(struct page
*page
, int order
)
405 set_page_private(page
, order
);
406 __SetPageBuddy(page
);
409 static inline void rmv_page_order(struct page
*page
)
411 __ClearPageBuddy(page
);
412 set_page_private(page
, 0);
416 * Locate the struct page for both the matching buddy in our
417 * pair (buddy1) and the combined O(n+1) page they form (page).
419 * 1) Any buddy B1 will have an order O twin B2 which satisfies
420 * the following equation:
422 * For example, if the starting buddy (buddy2) is #8 its order
424 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
426 * 2) Any buddy B will have an order O+1 parent P which
427 * satisfies the following equation:
430 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
432 static inline unsigned long
433 __find_buddy_index(unsigned long page_idx
, unsigned int order
)
435 return page_idx
^ (1 << order
);
439 * This function checks whether a page is free && is the buddy
440 * we can do coalesce a page and its buddy if
441 * (a) the buddy is not in a hole &&
442 * (b) the buddy is in the buddy system &&
443 * (c) a page and its buddy have the same order &&
444 * (d) a page and its buddy are in the same zone.
446 * For recording whether a page is in the buddy system, we set ->_mapcount -2.
447 * Setting, clearing, and testing _mapcount -2 is serialized by zone->lock.
449 * For recording page's order, we use page_private(page).
451 static inline int page_is_buddy(struct page
*page
, struct page
*buddy
,
454 if (!pfn_valid_within(page_to_pfn(buddy
)))
457 if (page_zone_id(page
) != page_zone_id(buddy
))
460 if (PageBuddy(buddy
) && page_order(buddy
) == order
) {
461 VM_BUG_ON(page_count(buddy
) != 0);
468 * Freeing function for a buddy system allocator.
470 * The concept of a buddy system is to maintain direct-mapped table
471 * (containing bit values) for memory blocks of various "orders".
472 * The bottom level table contains the map for the smallest allocatable
473 * units of memory (here, pages), and each level above it describes
474 * pairs of units from the levels below, hence, "buddies".
475 * At a high level, all that happens here is marking the table entry
476 * at the bottom level available, and propagating the changes upward
477 * as necessary, plus some accounting needed to play nicely with other
478 * parts of the VM system.
479 * At each level, we keep a list of pages, which are heads of continuous
480 * free pages of length of (1 << order) and marked with _mapcount -2. Page's
481 * order is recorded in page_private(page) field.
482 * So when we are allocating or freeing one, we can derive the state of the
483 * other. That is, if we allocate a small block, and both were
484 * free, the remainder of the region must be split into blocks.
485 * If a block is freed, and its buddy is also free, then this
486 * triggers coalescing into a block of larger size.
491 static inline void __free_one_page(struct page
*page
,
492 struct zone
*zone
, unsigned int order
,
495 unsigned long page_idx
;
496 unsigned long combined_idx
;
497 unsigned long uninitialized_var(buddy_idx
);
500 if (unlikely(PageCompound(page
)))
501 if (unlikely(destroy_compound_page(page
, order
)))
504 VM_BUG_ON(migratetype
== -1);
506 page_idx
= page_to_pfn(page
) & ((1 << MAX_ORDER
) - 1);
508 VM_BUG_ON(page_idx
& ((1 << order
) - 1));
509 VM_BUG_ON(bad_range(zone
, page
));
511 while (order
< MAX_ORDER
-1) {
512 buddy_idx
= __find_buddy_index(page_idx
, order
);
513 buddy
= page
+ (buddy_idx
- page_idx
);
514 if (!page_is_buddy(page
, buddy
, order
))
517 /* Our buddy is free, merge with it and move up one order. */
518 list_del(&buddy
->lru
);
519 zone
->free_area
[order
].nr_free
--;
520 rmv_page_order(buddy
);
521 combined_idx
= buddy_idx
& page_idx
;
522 page
= page
+ (combined_idx
- page_idx
);
523 page_idx
= combined_idx
;
526 set_page_order(page
, order
);
529 * If this is not the largest possible page, check if the buddy
530 * of the next-highest order is free. If it is, it's possible
531 * that pages are being freed that will coalesce soon. In case,
532 * that is happening, add the free page to the tail of the list
533 * so it's less likely to be used soon and more likely to be merged
534 * as a higher order page
536 if ((order
< MAX_ORDER
-2) && pfn_valid_within(page_to_pfn(buddy
))) {
537 struct page
*higher_page
, *higher_buddy
;
538 combined_idx
= buddy_idx
& page_idx
;
539 higher_page
= page
+ (combined_idx
- page_idx
);
540 buddy_idx
= __find_buddy_index(combined_idx
, order
+ 1);
541 higher_buddy
= page
+ (buddy_idx
- combined_idx
);
542 if (page_is_buddy(higher_page
, higher_buddy
, order
+ 1)) {
543 list_add_tail(&page
->lru
,
544 &zone
->free_area
[order
].free_list
[migratetype
]);
549 list_add(&page
->lru
, &zone
->free_area
[order
].free_list
[migratetype
]);
551 zone
->free_area
[order
].nr_free
++;
555 * free_page_mlock() -- clean up attempts to free and mlocked() page.
556 * Page should not be on lru, so no need to fix that up.
557 * free_pages_check() will verify...
559 static inline void free_page_mlock(struct page
*page
)
561 __dec_zone_page_state(page
, NR_MLOCK
);
562 __count_vm_event(UNEVICTABLE_MLOCKFREED
);
565 static inline int free_pages_check(struct page
*page
)
567 if (unlikely(page_mapcount(page
) |
568 (page
->mapping
!= NULL
) |
569 (atomic_read(&page
->_count
) != 0) |
570 (page
->flags
& PAGE_FLAGS_CHECK_AT_FREE
) |
571 (mem_cgroup_bad_page_check(page
)))) {
575 if (page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)
576 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
581 * Frees a number of pages from the PCP lists
582 * Assumes all pages on list are in same zone, and of same order.
583 * count is the number of pages to free.
585 * If the zone was previously in an "all pages pinned" state then look to
586 * see if this freeing clears that state.
588 * And clear the zone's pages_scanned counter, to hold off the "all pages are
589 * pinned" detection logic.
591 static void free_pcppages_bulk(struct zone
*zone
, int count
,
592 struct per_cpu_pages
*pcp
)
598 spin_lock(&zone
->lock
);
599 zone
->all_unreclaimable
= 0;
600 zone
->pages_scanned
= 0;
604 struct list_head
*list
;
607 * Remove pages from lists in a round-robin fashion. A
608 * batch_free count is maintained that is incremented when an
609 * empty list is encountered. This is so more pages are freed
610 * off fuller lists instead of spinning excessively around empty
615 if (++migratetype
== MIGRATE_PCPTYPES
)
617 list
= &pcp
->lists
[migratetype
];
618 } while (list_empty(list
));
620 /* This is the only non-empty list. Free them all. */
621 if (batch_free
== MIGRATE_PCPTYPES
)
622 batch_free
= to_free
;
625 page
= list_entry(list
->prev
, struct page
, lru
);
626 /* must delete as __free_one_page list manipulates */
627 list_del(&page
->lru
);
628 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
629 __free_one_page(page
, zone
, 0, page_private(page
));
630 trace_mm_page_pcpu_drain(page
, 0, page_private(page
));
631 } while (--to_free
&& --batch_free
&& !list_empty(list
));
633 __mod_zone_page_state(zone
, NR_FREE_PAGES
, count
);
634 spin_unlock(&zone
->lock
);
637 static void free_one_page(struct zone
*zone
, struct page
*page
, int order
,
640 spin_lock(&zone
->lock
);
641 zone
->all_unreclaimable
= 0;
642 zone
->pages_scanned
= 0;
644 __free_one_page(page
, zone
, order
, migratetype
);
645 __mod_zone_page_state(zone
, NR_FREE_PAGES
, 1 << order
);
646 spin_unlock(&zone
->lock
);
649 static bool free_pages_prepare(struct page
*page
, unsigned int order
)
654 trace_mm_page_free_direct(page
, order
);
655 kmemcheck_free_shadow(page
, order
);
658 page
->mapping
= NULL
;
659 for (i
= 0; i
< (1 << order
); i
++)
660 bad
+= free_pages_check(page
+ i
);
664 if (!PageHighMem(page
)) {
665 debug_check_no_locks_freed(page_address(page
),PAGE_SIZE
<<order
);
666 debug_check_no_obj_freed(page_address(page
),
669 arch_free_page(page
, order
);
670 kernel_map_pages(page
, 1 << order
, 0);
675 static void __free_pages_ok(struct page
*page
, unsigned int order
)
678 int wasMlocked
= __TestClearPageMlocked(page
);
680 if (!free_pages_prepare(page
, order
))
683 local_irq_save(flags
);
684 if (unlikely(wasMlocked
))
685 free_page_mlock(page
);
686 __count_vm_events(PGFREE
, 1 << order
);
687 free_one_page(page_zone(page
), page
, order
,
688 get_pageblock_migratetype(page
));
689 local_irq_restore(flags
);
693 * permit the bootmem allocator to evade page validation on high-order frees
695 void __meminit
__free_pages_bootmem(struct page
*page
, unsigned int order
)
698 __ClearPageReserved(page
);
699 set_page_count(page
, 0);
700 set_page_refcounted(page
);
706 for (loop
= 0; loop
< BITS_PER_LONG
; loop
++) {
707 struct page
*p
= &page
[loop
];
709 if (loop
+ 1 < BITS_PER_LONG
)
711 __ClearPageReserved(p
);
712 set_page_count(p
, 0);
715 set_page_refcounted(page
);
716 __free_pages(page
, order
);
722 * The order of subdivision here is critical for the IO subsystem.
723 * Please do not alter this order without good reasons and regression
724 * testing. Specifically, as large blocks of memory are subdivided,
725 * the order in which smaller blocks are delivered depends on the order
726 * they're subdivided in this function. This is the primary factor
727 * influencing the order in which pages are delivered to the IO
728 * subsystem according to empirical testing, and this is also justified
729 * by considering the behavior of a buddy system containing a single
730 * large block of memory acted on by a series of small allocations.
731 * This behavior is a critical factor in sglist merging's success.
735 static inline void expand(struct zone
*zone
, struct page
*page
,
736 int low
, int high
, struct free_area
*area
,
739 unsigned long size
= 1 << high
;
745 VM_BUG_ON(bad_range(zone
, &page
[size
]));
746 list_add(&page
[size
].lru
, &area
->free_list
[migratetype
]);
748 set_page_order(&page
[size
], high
);
753 * This page is about to be returned from the page allocator
755 static inline int check_new_page(struct page
*page
)
757 if (unlikely(page_mapcount(page
) |
758 (page
->mapping
!= NULL
) |
759 (atomic_read(&page
->_count
) != 0) |
760 (page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
) |
761 (mem_cgroup_bad_page_check(page
)))) {
768 static int prep_new_page(struct page
*page
, int order
, gfp_t gfp_flags
)
772 for (i
= 0; i
< (1 << order
); i
++) {
773 struct page
*p
= page
+ i
;
774 if (unlikely(check_new_page(p
)))
778 set_page_private(page
, 0);
779 set_page_refcounted(page
);
781 arch_alloc_page(page
, order
);
782 kernel_map_pages(page
, 1 << order
, 1);
784 if (gfp_flags
& __GFP_ZERO
)
785 prep_zero_page(page
, order
, gfp_flags
);
787 if (order
&& (gfp_flags
& __GFP_COMP
))
788 prep_compound_page(page
, order
);
794 * Go through the free lists for the given migratetype and remove
795 * the smallest available page from the freelists
798 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
801 unsigned int current_order
;
802 struct free_area
* area
;
805 /* Find a page of the appropriate size in the preferred list */
806 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
807 area
= &(zone
->free_area
[current_order
]);
808 if (list_empty(&area
->free_list
[migratetype
]))
811 page
= list_entry(area
->free_list
[migratetype
].next
,
813 list_del(&page
->lru
);
814 rmv_page_order(page
);
816 expand(zone
, page
, order
, current_order
, area
, migratetype
);
825 * This array describes the order lists are fallen back to when
826 * the free lists for the desirable migrate type are depleted
828 static int fallbacks
[MIGRATE_TYPES
][MIGRATE_TYPES
-1] = {
829 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
830 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
831 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_RESERVE
},
832 [MIGRATE_RESERVE
] = { MIGRATE_RESERVE
, MIGRATE_RESERVE
, MIGRATE_RESERVE
}, /* Never used */
836 * Move the free pages in a range to the free lists of the requested type.
837 * Note that start_page and end_pages are not aligned on a pageblock
838 * boundary. If alignment is required, use move_freepages_block()
840 static int move_freepages(struct zone
*zone
,
841 struct page
*start_page
, struct page
*end_page
,
848 #ifndef CONFIG_HOLES_IN_ZONE
850 * page_zone is not safe to call in this context when
851 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
852 * anyway as we check zone boundaries in move_freepages_block().
853 * Remove at a later date when no bug reports exist related to
854 * grouping pages by mobility
856 BUG_ON(page_zone(start_page
) != page_zone(end_page
));
859 for (page
= start_page
; page
<= end_page
;) {
860 /* Make sure we are not inadvertently changing nodes */
861 VM_BUG_ON(page_to_nid(page
) != zone_to_nid(zone
));
863 if (!pfn_valid_within(page_to_pfn(page
))) {
868 if (!PageBuddy(page
)) {
873 order
= page_order(page
);
874 list_move(&page
->lru
,
875 &zone
->free_area
[order
].free_list
[migratetype
]);
877 pages_moved
+= 1 << order
;
883 static int move_freepages_block(struct zone
*zone
, struct page
*page
,
886 unsigned long start_pfn
, end_pfn
;
887 struct page
*start_page
, *end_page
;
889 start_pfn
= page_to_pfn(page
);
890 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
891 start_page
= pfn_to_page(start_pfn
);
892 end_page
= start_page
+ pageblock_nr_pages
- 1;
893 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
895 /* Do not cross zone boundaries */
896 if (start_pfn
< zone
->zone_start_pfn
)
898 if (end_pfn
>= zone
->zone_start_pfn
+ zone
->spanned_pages
)
901 return move_freepages(zone
, start_page
, end_page
, migratetype
);
904 static void change_pageblock_range(struct page
*pageblock_page
,
905 int start_order
, int migratetype
)
907 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
909 while (nr_pageblocks
--) {
910 set_pageblock_migratetype(pageblock_page
, migratetype
);
911 pageblock_page
+= pageblock_nr_pages
;
915 /* Remove an element from the buddy allocator from the fallback list */
916 static inline struct page
*
917 __rmqueue_fallback(struct zone
*zone
, int order
, int start_migratetype
)
919 struct free_area
* area
;
924 /* Find the largest possible block of pages in the other list */
925 for (current_order
= MAX_ORDER
-1; current_order
>= order
;
927 for (i
= 0; i
< MIGRATE_TYPES
- 1; i
++) {
928 migratetype
= fallbacks
[start_migratetype
][i
];
930 /* MIGRATE_RESERVE handled later if necessary */
931 if (migratetype
== MIGRATE_RESERVE
)
934 area
= &(zone
->free_area
[current_order
]);
935 if (list_empty(&area
->free_list
[migratetype
]))
938 page
= list_entry(area
->free_list
[migratetype
].next
,
943 * If breaking a large block of pages, move all free
944 * pages to the preferred allocation list. If falling
945 * back for a reclaimable kernel allocation, be more
946 * aggressive about taking ownership of free pages
948 if (unlikely(current_order
>= (pageblock_order
>> 1)) ||
949 start_migratetype
== MIGRATE_RECLAIMABLE
||
950 page_group_by_mobility_disabled
) {
952 pages
= move_freepages_block(zone
, page
,
955 /* Claim the whole block if over half of it is free */
956 if (pages
>= (1 << (pageblock_order
-1)) ||
957 page_group_by_mobility_disabled
)
958 set_pageblock_migratetype(page
,
961 migratetype
= start_migratetype
;
964 /* Remove the page from the freelists */
965 list_del(&page
->lru
);
966 rmv_page_order(page
);
968 /* Take ownership for orders >= pageblock_order */
969 if (current_order
>= pageblock_order
)
970 change_pageblock_range(page
, current_order
,
973 expand(zone
, page
, order
, current_order
, area
, migratetype
);
975 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
976 start_migratetype
, migratetype
);
986 * Do the hard work of removing an element from the buddy allocator.
987 * Call me with the zone->lock already held.
989 static struct page
*__rmqueue(struct zone
*zone
, unsigned int order
,
995 page
= __rmqueue_smallest(zone
, order
, migratetype
);
997 if (unlikely(!page
) && migratetype
!= MIGRATE_RESERVE
) {
998 page
= __rmqueue_fallback(zone
, order
, migratetype
);
1001 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1002 * is used because __rmqueue_smallest is an inline function
1003 * and we want just one call site
1006 migratetype
= MIGRATE_RESERVE
;
1011 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
1016 * Obtain a specified number of elements from the buddy allocator, all under
1017 * a single hold of the lock, for efficiency. Add them to the supplied list.
1018 * Returns the number of new pages which were placed at *list.
1020 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
1021 unsigned long count
, struct list_head
*list
,
1022 int migratetype
, int cold
)
1026 spin_lock(&zone
->lock
);
1027 for (i
= 0; i
< count
; ++i
) {
1028 struct page
*page
= __rmqueue(zone
, order
, migratetype
);
1029 if (unlikely(page
== NULL
))
1033 * Split buddy pages returned by expand() are received here
1034 * in physical page order. The page is added to the callers and
1035 * list and the list head then moves forward. From the callers
1036 * perspective, the linked list is ordered by page number in
1037 * some conditions. This is useful for IO devices that can
1038 * merge IO requests if the physical pages are ordered
1041 if (likely(cold
== 0))
1042 list_add(&page
->lru
, list
);
1044 list_add_tail(&page
->lru
, list
);
1045 set_page_private(page
, migratetype
);
1048 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
1049 spin_unlock(&zone
->lock
);
1055 * Called from the vmstat counter updater to drain pagesets of this
1056 * currently executing processor on remote nodes after they have
1059 * Note that this function must be called with the thread pinned to
1060 * a single processor.
1062 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
1064 unsigned long flags
;
1067 local_irq_save(flags
);
1068 if (pcp
->count
>= pcp
->batch
)
1069 to_drain
= pcp
->batch
;
1071 to_drain
= pcp
->count
;
1072 free_pcppages_bulk(zone
, to_drain
, pcp
);
1073 pcp
->count
-= to_drain
;
1074 local_irq_restore(flags
);
1079 * Drain pages of the indicated processor.
1081 * The processor must either be the current processor and the
1082 * thread pinned to the current processor or a processor that
1085 static void drain_pages(unsigned int cpu
)
1087 unsigned long flags
;
1090 for_each_populated_zone(zone
) {
1091 struct per_cpu_pageset
*pset
;
1092 struct per_cpu_pages
*pcp
;
1094 local_irq_save(flags
);
1095 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
1099 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
1102 local_irq_restore(flags
);
1107 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1109 void drain_local_pages(void *arg
)
1111 drain_pages(smp_processor_id());
1115 * Spill all the per-cpu pages from all CPUs back into the buddy allocator
1117 void drain_all_pages(void)
1119 on_each_cpu(drain_local_pages
, NULL
, 1);
1122 #ifdef CONFIG_HIBERNATION
1124 void mark_free_pages(struct zone
*zone
)
1126 unsigned long pfn
, max_zone_pfn
;
1127 unsigned long flags
;
1129 struct list_head
*curr
;
1131 if (!zone
->spanned_pages
)
1134 spin_lock_irqsave(&zone
->lock
, flags
);
1136 max_zone_pfn
= zone
->zone_start_pfn
+ zone
->spanned_pages
;
1137 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
1138 if (pfn_valid(pfn
)) {
1139 struct page
*page
= pfn_to_page(pfn
);
1141 if (!swsusp_page_is_forbidden(page
))
1142 swsusp_unset_page_free(page
);
1145 for_each_migratetype_order(order
, t
) {
1146 list_for_each(curr
, &zone
->free_area
[order
].free_list
[t
]) {
1149 pfn
= page_to_pfn(list_entry(curr
, struct page
, lru
));
1150 for (i
= 0; i
< (1UL << order
); i
++)
1151 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
1154 spin_unlock_irqrestore(&zone
->lock
, flags
);
1156 #endif /* CONFIG_PM */
1159 * Free a 0-order page
1160 * cold == 1 ? free a cold page : free a hot page
1162 void free_hot_cold_page(struct page
*page
, int cold
)
1164 struct zone
*zone
= page_zone(page
);
1165 struct per_cpu_pages
*pcp
;
1166 unsigned long flags
;
1168 int wasMlocked
= __TestClearPageMlocked(page
);
1170 if (!free_pages_prepare(page
, 0))
1173 migratetype
= get_pageblock_migratetype(page
);
1174 set_page_private(page
, migratetype
);
1175 local_irq_save(flags
);
1176 if (unlikely(wasMlocked
))
1177 free_page_mlock(page
);
1178 __count_vm_event(PGFREE
);
1181 * We only track unmovable, reclaimable and movable on pcp lists.
1182 * Free ISOLATE pages back to the allocator because they are being
1183 * offlined but treat RESERVE as movable pages so we can get those
1184 * areas back if necessary. Otherwise, we may have to free
1185 * excessively into the page allocator
1187 if (migratetype
>= MIGRATE_PCPTYPES
) {
1188 if (unlikely(migratetype
== MIGRATE_ISOLATE
)) {
1189 free_one_page(zone
, page
, 0, migratetype
);
1192 migratetype
= MIGRATE_MOVABLE
;
1195 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
1197 list_add_tail(&page
->lru
, &pcp
->lists
[migratetype
]);
1199 list_add(&page
->lru
, &pcp
->lists
[migratetype
]);
1201 if (pcp
->count
>= pcp
->high
) {
1202 free_pcppages_bulk(zone
, pcp
->batch
, pcp
);
1203 pcp
->count
-= pcp
->batch
;
1207 local_irq_restore(flags
);
1211 * split_page takes a non-compound higher-order page, and splits it into
1212 * n (1<<order) sub-pages: page[0..n]
1213 * Each sub-page must be freed individually.
1215 * Note: this is probably too low level an operation for use in drivers.
1216 * Please consult with lkml before using this in your driver.
1218 void split_page(struct page
*page
, unsigned int order
)
1222 VM_BUG_ON(PageCompound(page
));
1223 VM_BUG_ON(!page_count(page
));
1225 #ifdef CONFIG_KMEMCHECK
1227 * Split shadow pages too, because free(page[0]) would
1228 * otherwise free the whole shadow.
1230 if (kmemcheck_page_is_tracked(page
))
1231 split_page(virt_to_page(page
[0].shadow
), order
);
1234 for (i
= 1; i
< (1 << order
); i
++)
1235 set_page_refcounted(page
+ i
);
1239 * Similar to split_page except the page is already free. As this is only
1240 * being used for migration, the migratetype of the block also changes.
1241 * As this is called with interrupts disabled, the caller is responsible
1242 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1245 * Note: this is probably too low level an operation for use in drivers.
1246 * Please consult with lkml before using this in your driver.
1248 int split_free_page(struct page
*page
)
1251 unsigned long watermark
;
1254 BUG_ON(!PageBuddy(page
));
1256 zone
= page_zone(page
);
1257 order
= page_order(page
);
1259 /* Obey watermarks as if the page was being allocated */
1260 watermark
= low_wmark_pages(zone
) + (1 << order
);
1261 if (!zone_watermark_ok(zone
, 0, watermark
, 0, 0))
1264 /* Remove page from free list */
1265 list_del(&page
->lru
);
1266 zone
->free_area
[order
].nr_free
--;
1267 rmv_page_order(page
);
1268 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(1UL << order
));
1270 /* Split into individual pages */
1271 set_page_refcounted(page
);
1272 split_page(page
, order
);
1274 if (order
>= pageblock_order
- 1) {
1275 struct page
*endpage
= page
+ (1 << order
) - 1;
1276 for (; page
< endpage
; page
+= pageblock_nr_pages
)
1277 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
1284 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1285 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1289 struct page
*buffered_rmqueue(struct zone
*preferred_zone
,
1290 struct zone
*zone
, int order
, gfp_t gfp_flags
,
1293 unsigned long flags
;
1295 int cold
= !!(gfp_flags
& __GFP_COLD
);
1298 if (likely(order
== 0)) {
1299 struct per_cpu_pages
*pcp
;
1300 struct list_head
*list
;
1302 local_irq_save(flags
);
1303 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
1304 list
= &pcp
->lists
[migratetype
];
1305 if (list_empty(list
)) {
1306 pcp
->count
+= rmqueue_bulk(zone
, 0,
1309 if (unlikely(list_empty(list
)))
1314 page
= list_entry(list
->prev
, struct page
, lru
);
1316 page
= list_entry(list
->next
, struct page
, lru
);
1318 list_del(&page
->lru
);
1321 if (unlikely(gfp_flags
& __GFP_NOFAIL
)) {
1323 * __GFP_NOFAIL is not to be used in new code.
1325 * All __GFP_NOFAIL callers should be fixed so that they
1326 * properly detect and handle allocation failures.
1328 * We most definitely don't want callers attempting to
1329 * allocate greater than order-1 page units with
1332 WARN_ON_ONCE(order
> 1);
1334 spin_lock_irqsave(&zone
->lock
, flags
);
1335 page
= __rmqueue(zone
, order
, migratetype
);
1336 spin_unlock(&zone
->lock
);
1339 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(1 << order
));
1342 __count_zone_vm_events(PGALLOC
, zone
, 1 << order
);
1343 zone_statistics(preferred_zone
, zone
, gfp_flags
);
1344 local_irq_restore(flags
);
1346 VM_BUG_ON(bad_range(zone
, page
));
1347 if (prep_new_page(page
, order
, gfp_flags
))
1352 local_irq_restore(flags
);
1356 /* The ALLOC_WMARK bits are used as an index to zone->watermark */
1357 #define ALLOC_WMARK_MIN WMARK_MIN
1358 #define ALLOC_WMARK_LOW WMARK_LOW
1359 #define ALLOC_WMARK_HIGH WMARK_HIGH
1360 #define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */
1362 /* Mask to get the watermark bits */
1363 #define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1)
1365 #define ALLOC_HARDER 0x10 /* try to alloc harder */
1366 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
1367 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
1369 #ifdef CONFIG_FAIL_PAGE_ALLOC
1371 static struct fail_page_alloc_attr
{
1372 struct fault_attr attr
;
1374 u32 ignore_gfp_highmem
;
1375 u32 ignore_gfp_wait
;
1378 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1380 struct dentry
*ignore_gfp_highmem_file
;
1381 struct dentry
*ignore_gfp_wait_file
;
1382 struct dentry
*min_order_file
;
1384 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1386 } fail_page_alloc
= {
1387 .attr
= FAULT_ATTR_INITIALIZER
,
1388 .ignore_gfp_wait
= 1,
1389 .ignore_gfp_highmem
= 1,
1393 static int __init
setup_fail_page_alloc(char *str
)
1395 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
1397 __setup("fail_page_alloc=", setup_fail_page_alloc
);
1399 static int should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1401 if (order
< fail_page_alloc
.min_order
)
1403 if (gfp_mask
& __GFP_NOFAIL
)
1405 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
1407 if (fail_page_alloc
.ignore_gfp_wait
&& (gfp_mask
& __GFP_WAIT
))
1410 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
1413 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1415 static int __init
fail_page_alloc_debugfs(void)
1417 mode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
1421 err
= init_fault_attr_dentries(&fail_page_alloc
.attr
,
1425 dir
= fail_page_alloc
.attr
.dentries
.dir
;
1427 fail_page_alloc
.ignore_gfp_wait_file
=
1428 debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
1429 &fail_page_alloc
.ignore_gfp_wait
);
1431 fail_page_alloc
.ignore_gfp_highmem_file
=
1432 debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
1433 &fail_page_alloc
.ignore_gfp_highmem
);
1434 fail_page_alloc
.min_order_file
=
1435 debugfs_create_u32("min-order", mode
, dir
,
1436 &fail_page_alloc
.min_order
);
1438 if (!fail_page_alloc
.ignore_gfp_wait_file
||
1439 !fail_page_alloc
.ignore_gfp_highmem_file
||
1440 !fail_page_alloc
.min_order_file
) {
1442 debugfs_remove(fail_page_alloc
.ignore_gfp_wait_file
);
1443 debugfs_remove(fail_page_alloc
.ignore_gfp_highmem_file
);
1444 debugfs_remove(fail_page_alloc
.min_order_file
);
1445 cleanup_fault_attr_dentries(&fail_page_alloc
.attr
);
1451 late_initcall(fail_page_alloc_debugfs
);
1453 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1455 #else /* CONFIG_FAIL_PAGE_ALLOC */
1457 static inline int should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1462 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1465 * Return true if free pages are above 'mark'. This takes into account the order
1466 * of the allocation.
1468 static bool __zone_watermark_ok(struct zone
*z
, int order
, unsigned long mark
,
1469 int classzone_idx
, int alloc_flags
, long free_pages
)
1471 /* free_pages my go negative - that's OK */
1475 free_pages
-= (1 << order
) + 1;
1476 if (alloc_flags
& ALLOC_HIGH
)
1478 if (alloc_flags
& ALLOC_HARDER
)
1481 if (free_pages
<= min
+ z
->lowmem_reserve
[classzone_idx
])
1483 for (o
= 0; o
< order
; o
++) {
1484 /* At the next order, this order's pages become unavailable */
1485 free_pages
-= z
->free_area
[o
].nr_free
<< o
;
1487 /* Require fewer higher order pages to be free */
1490 if (free_pages
<= min
)
1496 bool zone_watermark_ok(struct zone
*z
, int order
, unsigned long mark
,
1497 int classzone_idx
, int alloc_flags
)
1499 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
1500 zone_page_state(z
, NR_FREE_PAGES
));
1503 bool zone_watermark_ok_safe(struct zone
*z
, int order
, unsigned long mark
,
1504 int classzone_idx
, int alloc_flags
)
1506 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
1508 if (z
->percpu_drift_mark
&& free_pages
< z
->percpu_drift_mark
)
1509 free_pages
= zone_page_state_snapshot(z
, NR_FREE_PAGES
);
1511 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
1517 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1518 * skip over zones that are not allowed by the cpuset, or that have
1519 * been recently (in last second) found to be nearly full. See further
1520 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1521 * that have to skip over a lot of full or unallowed zones.
1523 * If the zonelist cache is present in the passed in zonelist, then
1524 * returns a pointer to the allowed node mask (either the current
1525 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1527 * If the zonelist cache is not available for this zonelist, does
1528 * nothing and returns NULL.
1530 * If the fullzones BITMAP in the zonelist cache is stale (more than
1531 * a second since last zap'd) then we zap it out (clear its bits.)
1533 * We hold off even calling zlc_setup, until after we've checked the
1534 * first zone in the zonelist, on the theory that most allocations will
1535 * be satisfied from that first zone, so best to examine that zone as
1536 * quickly as we can.
1538 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1540 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1541 nodemask_t
*allowednodes
; /* zonelist_cache approximation */
1543 zlc
= zonelist
->zlcache_ptr
;
1547 if (time_after(jiffies
, zlc
->last_full_zap
+ HZ
)) {
1548 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
1549 zlc
->last_full_zap
= jiffies
;
1552 allowednodes
= !in_interrupt() && (alloc_flags
& ALLOC_CPUSET
) ?
1553 &cpuset_current_mems_allowed
:
1554 &node_states
[N_HIGH_MEMORY
];
1555 return allowednodes
;
1559 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1560 * if it is worth looking at further for free memory:
1561 * 1) Check that the zone isn't thought to be full (doesn't have its
1562 * bit set in the zonelist_cache fullzones BITMAP).
1563 * 2) Check that the zones node (obtained from the zonelist_cache
1564 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1565 * Return true (non-zero) if zone is worth looking at further, or
1566 * else return false (zero) if it is not.
1568 * This check -ignores- the distinction between various watermarks,
1569 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1570 * found to be full for any variation of these watermarks, it will
1571 * be considered full for up to one second by all requests, unless
1572 * we are so low on memory on all allowed nodes that we are forced
1573 * into the second scan of the zonelist.
1575 * In the second scan we ignore this zonelist cache and exactly
1576 * apply the watermarks to all zones, even it is slower to do so.
1577 * We are low on memory in the second scan, and should leave no stone
1578 * unturned looking for a free page.
1580 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1581 nodemask_t
*allowednodes
)
1583 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1584 int i
; /* index of *z in zonelist zones */
1585 int n
; /* node that zone *z is on */
1587 zlc
= zonelist
->zlcache_ptr
;
1591 i
= z
- zonelist
->_zonerefs
;
1594 /* This zone is worth trying if it is allowed but not full */
1595 return node_isset(n
, *allowednodes
) && !test_bit(i
, zlc
->fullzones
);
1599 * Given 'z' scanning a zonelist, set the corresponding bit in
1600 * zlc->fullzones, so that subsequent attempts to allocate a page
1601 * from that zone don't waste time re-examining it.
1603 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1605 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1606 int i
; /* index of *z in zonelist zones */
1608 zlc
= zonelist
->zlcache_ptr
;
1612 i
= z
- zonelist
->_zonerefs
;
1614 set_bit(i
, zlc
->fullzones
);
1617 #else /* CONFIG_NUMA */
1619 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1624 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1625 nodemask_t
*allowednodes
)
1630 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1633 #endif /* CONFIG_NUMA */
1636 * get_page_from_freelist goes through the zonelist trying to allocate
1639 static struct page
*
1640 get_page_from_freelist(gfp_t gfp_mask
, nodemask_t
*nodemask
, unsigned int order
,
1641 struct zonelist
*zonelist
, int high_zoneidx
, int alloc_flags
,
1642 struct zone
*preferred_zone
, int migratetype
)
1645 struct page
*page
= NULL
;
1648 nodemask_t
*allowednodes
= NULL
;/* zonelist_cache approximation */
1649 int zlc_active
= 0; /* set if using zonelist_cache */
1650 int did_zlc_setup
= 0; /* just call zlc_setup() one time */
1652 classzone_idx
= zone_idx(preferred_zone
);
1655 * Scan zonelist, looking for a zone with enough free.
1656 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1658 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
1659 high_zoneidx
, nodemask
) {
1660 if (NUMA_BUILD
&& zlc_active
&&
1661 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
1663 if ((alloc_flags
& ALLOC_CPUSET
) &&
1664 !cpuset_zone_allowed_softwall(zone
, gfp_mask
))
1667 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
1668 if (!(alloc_flags
& ALLOC_NO_WATERMARKS
)) {
1672 mark
= zone
->watermark
[alloc_flags
& ALLOC_WMARK_MASK
];
1673 if (zone_watermark_ok(zone
, order
, mark
,
1674 classzone_idx
, alloc_flags
))
1677 if (zone_reclaim_mode
== 0)
1678 goto this_zone_full
;
1680 ret
= zone_reclaim(zone
, gfp_mask
, order
);
1682 case ZONE_RECLAIM_NOSCAN
:
1685 case ZONE_RECLAIM_FULL
:
1686 /* scanned but unreclaimable */
1687 goto this_zone_full
;
1689 /* did we reclaim enough */
1690 if (!zone_watermark_ok(zone
, order
, mark
,
1691 classzone_idx
, alloc_flags
))
1692 goto this_zone_full
;
1697 page
= buffered_rmqueue(preferred_zone
, zone
, order
,
1698 gfp_mask
, migratetype
);
1703 zlc_mark_zone_full(zonelist
, z
);
1705 if (NUMA_BUILD
&& !did_zlc_setup
&& nr_online_nodes
> 1) {
1707 * we do zlc_setup after the first zone is tried but only
1708 * if there are multiple nodes make it worthwhile
1710 allowednodes
= zlc_setup(zonelist
, alloc_flags
);
1716 if (unlikely(NUMA_BUILD
&& page
== NULL
&& zlc_active
)) {
1717 /* Disable zlc cache for second zonelist scan */
1725 * Large machines with many possible nodes should not always dump per-node
1726 * meminfo in irq context.
1728 static inline bool should_suppress_show_mem(void)
1733 ret
= in_interrupt();
1739 should_alloc_retry(gfp_t gfp_mask
, unsigned int order
,
1740 unsigned long pages_reclaimed
)
1742 /* Do not loop if specifically requested */
1743 if (gfp_mask
& __GFP_NORETRY
)
1747 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
1748 * means __GFP_NOFAIL, but that may not be true in other
1751 if (order
<= PAGE_ALLOC_COSTLY_ORDER
)
1755 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
1756 * specified, then we retry until we no longer reclaim any pages
1757 * (above), or we've reclaimed an order of pages at least as
1758 * large as the allocation's order. In both cases, if the
1759 * allocation still fails, we stop retrying.
1761 if (gfp_mask
& __GFP_REPEAT
&& pages_reclaimed
< (1 << order
))
1765 * Don't let big-order allocations loop unless the caller
1766 * explicitly requests that.
1768 if (gfp_mask
& __GFP_NOFAIL
)
1774 static inline struct page
*
1775 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
1776 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1777 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
1782 /* Acquire the OOM killer lock for the zones in zonelist */
1783 if (!try_set_zonelist_oom(zonelist
, gfp_mask
)) {
1784 schedule_timeout_uninterruptible(1);
1789 * Go through the zonelist yet one more time, keep very high watermark
1790 * here, this is only to catch a parallel oom killing, we must fail if
1791 * we're still under heavy pressure.
1793 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
,
1794 order
, zonelist
, high_zoneidx
,
1795 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
,
1796 preferred_zone
, migratetype
);
1800 if (!(gfp_mask
& __GFP_NOFAIL
)) {
1801 /* The OOM killer will not help higher order allocs */
1802 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
1804 /* The OOM killer does not needlessly kill tasks for lowmem */
1805 if (high_zoneidx
< ZONE_NORMAL
)
1808 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
1809 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
1810 * The caller should handle page allocation failure by itself if
1811 * it specifies __GFP_THISNODE.
1812 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
1814 if (gfp_mask
& __GFP_THISNODE
)
1817 /* Exhausted what can be done so it's blamo time */
1818 out_of_memory(zonelist
, gfp_mask
, order
, nodemask
);
1821 clear_zonelist_oom(zonelist
, gfp_mask
);
1825 #ifdef CONFIG_COMPACTION
1826 /* Try memory compaction for high-order allocations before reclaim */
1827 static struct page
*
1828 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
1829 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1830 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
1831 int migratetype
, unsigned long *did_some_progress
,
1832 bool sync_migration
)
1836 if (!order
|| compaction_deferred(preferred_zone
))
1839 current
->flags
|= PF_MEMALLOC
;
1840 *did_some_progress
= try_to_compact_pages(zonelist
, order
, gfp_mask
,
1841 nodemask
, sync_migration
);
1842 current
->flags
&= ~PF_MEMALLOC
;
1843 if (*did_some_progress
!= COMPACT_SKIPPED
) {
1845 /* Page migration frees to the PCP lists but we want merging */
1846 drain_pages(get_cpu());
1849 page
= get_page_from_freelist(gfp_mask
, nodemask
,
1850 order
, zonelist
, high_zoneidx
,
1851 alloc_flags
, preferred_zone
,
1854 preferred_zone
->compact_considered
= 0;
1855 preferred_zone
->compact_defer_shift
= 0;
1856 count_vm_event(COMPACTSUCCESS
);
1861 * It's bad if compaction run occurs and fails.
1862 * The most likely reason is that pages exist,
1863 * but not enough to satisfy watermarks.
1865 count_vm_event(COMPACTFAIL
);
1866 defer_compaction(preferred_zone
);
1874 static inline struct page
*
1875 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
1876 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1877 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
1878 int migratetype
, unsigned long *did_some_progress
,
1879 bool sync_migration
)
1883 #endif /* CONFIG_COMPACTION */
1885 /* The really slow allocator path where we enter direct reclaim */
1886 static inline struct page
*
1887 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
1888 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1889 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
1890 int migratetype
, unsigned long *did_some_progress
)
1892 struct page
*page
= NULL
;
1893 struct reclaim_state reclaim_state
;
1894 bool drained
= false;
1898 /* We now go into synchronous reclaim */
1899 cpuset_memory_pressure_bump();
1900 current
->flags
|= PF_MEMALLOC
;
1901 lockdep_set_current_reclaim_state(gfp_mask
);
1902 reclaim_state
.reclaimed_slab
= 0;
1903 current
->reclaim_state
= &reclaim_state
;
1905 *did_some_progress
= try_to_free_pages(zonelist
, order
, gfp_mask
, nodemask
);
1907 current
->reclaim_state
= NULL
;
1908 lockdep_clear_current_reclaim_state();
1909 current
->flags
&= ~PF_MEMALLOC
;
1913 if (unlikely(!(*did_some_progress
)))
1917 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
1918 zonelist
, high_zoneidx
,
1919 alloc_flags
, preferred_zone
,
1923 * If an allocation failed after direct reclaim, it could be because
1924 * pages are pinned on the per-cpu lists. Drain them and try again
1926 if (!page
&& !drained
) {
1936 * This is called in the allocator slow-path if the allocation request is of
1937 * sufficient urgency to ignore watermarks and take other desperate measures
1939 static inline struct page
*
1940 __alloc_pages_high_priority(gfp_t gfp_mask
, unsigned int order
,
1941 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1942 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
1948 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
1949 zonelist
, high_zoneidx
, ALLOC_NO_WATERMARKS
,
1950 preferred_zone
, migratetype
);
1952 if (!page
&& gfp_mask
& __GFP_NOFAIL
)
1953 wait_iff_congested(preferred_zone
, BLK_RW_ASYNC
, HZ
/50);
1954 } while (!page
&& (gfp_mask
& __GFP_NOFAIL
));
1960 void wake_all_kswapd(unsigned int order
, struct zonelist
*zonelist
,
1961 enum zone_type high_zoneidx
,
1962 enum zone_type classzone_idx
)
1967 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
)
1968 wakeup_kswapd(zone
, order
, classzone_idx
);
1972 gfp_to_alloc_flags(gfp_t gfp_mask
)
1974 int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
1975 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
1977 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
1978 BUILD_BUG_ON(__GFP_HIGH
!= (__force gfp_t
) ALLOC_HIGH
);
1981 * The caller may dip into page reserves a bit more if the caller
1982 * cannot run direct reclaim, or if the caller has realtime scheduling
1983 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
1984 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
1986 alloc_flags
|= (__force
int) (gfp_mask
& __GFP_HIGH
);
1990 * Not worth trying to allocate harder for
1991 * __GFP_NOMEMALLOC even if it can't schedule.
1993 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
1994 alloc_flags
|= ALLOC_HARDER
;
1996 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
1997 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1999 alloc_flags
&= ~ALLOC_CPUSET
;
2000 } else if (unlikely(rt_task(current
)) && !in_interrupt())
2001 alloc_flags
|= ALLOC_HARDER
;
2003 if (likely(!(gfp_mask
& __GFP_NOMEMALLOC
))) {
2004 if (!in_interrupt() &&
2005 ((current
->flags
& PF_MEMALLOC
) ||
2006 unlikely(test_thread_flag(TIF_MEMDIE
))))
2007 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2013 static inline struct page
*
2014 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
2015 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2016 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
2019 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
2020 struct page
*page
= NULL
;
2022 unsigned long pages_reclaimed
= 0;
2023 unsigned long did_some_progress
;
2024 bool sync_migration
= false;
2027 * In the slowpath, we sanity check order to avoid ever trying to
2028 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2029 * be using allocators in order of preference for an area that is
2032 if (order
>= MAX_ORDER
) {
2033 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
2038 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2039 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2040 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2041 * using a larger set of nodes after it has established that the
2042 * allowed per node queues are empty and that nodes are
2045 if (NUMA_BUILD
&& (gfp_mask
& GFP_THISNODE
) == GFP_THISNODE
)
2049 if (!(gfp_mask
& __GFP_NO_KSWAPD
))
2050 wake_all_kswapd(order
, zonelist
, high_zoneidx
,
2051 zone_idx(preferred_zone
));
2054 * OK, we're below the kswapd watermark and have kicked background
2055 * reclaim. Now things get more complex, so set up alloc_flags according
2056 * to how we want to proceed.
2058 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
2061 * Find the true preferred zone if the allocation is unconstrained by
2064 if (!(alloc_flags
& ALLOC_CPUSET
) && !nodemask
)
2065 first_zones_zonelist(zonelist
, high_zoneidx
, NULL
,
2068 /* This is the last chance, in general, before the goto nopage. */
2069 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
, zonelist
,
2070 high_zoneidx
, alloc_flags
& ~ALLOC_NO_WATERMARKS
,
2071 preferred_zone
, migratetype
);
2076 /* Allocate without watermarks if the context allows */
2077 if (alloc_flags
& ALLOC_NO_WATERMARKS
) {
2078 page
= __alloc_pages_high_priority(gfp_mask
, order
,
2079 zonelist
, high_zoneidx
, nodemask
,
2080 preferred_zone
, migratetype
);
2085 /* Atomic allocations - we can't balance anything */
2089 /* Avoid recursion of direct reclaim */
2090 if (current
->flags
& PF_MEMALLOC
)
2093 /* Avoid allocations with no watermarks from looping endlessly */
2094 if (test_thread_flag(TIF_MEMDIE
) && !(gfp_mask
& __GFP_NOFAIL
))
2098 * Try direct compaction. The first pass is asynchronous. Subsequent
2099 * attempts after direct reclaim are synchronous
2101 page
= __alloc_pages_direct_compact(gfp_mask
, order
,
2102 zonelist
, high_zoneidx
,
2104 alloc_flags
, preferred_zone
,
2105 migratetype
, &did_some_progress
,
2109 sync_migration
= !(gfp_mask
& __GFP_NO_KSWAPD
);
2111 /* Try direct reclaim and then allocating */
2112 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
,
2113 zonelist
, high_zoneidx
,
2115 alloc_flags
, preferred_zone
,
2116 migratetype
, &did_some_progress
);
2121 * If we failed to make any progress reclaiming, then we are
2122 * running out of options and have to consider going OOM
2124 if (!did_some_progress
) {
2125 if ((gfp_mask
& __GFP_FS
) && !(gfp_mask
& __GFP_NORETRY
)) {
2126 if (oom_killer_disabled
)
2128 page
= __alloc_pages_may_oom(gfp_mask
, order
,
2129 zonelist
, high_zoneidx
,
2130 nodemask
, preferred_zone
,
2135 if (!(gfp_mask
& __GFP_NOFAIL
)) {
2137 * The oom killer is not called for high-order
2138 * allocations that may fail, so if no progress
2139 * is being made, there are no other options and
2140 * retrying is unlikely to help.
2142 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
2145 * The oom killer is not called for lowmem
2146 * allocations to prevent needlessly killing
2149 if (high_zoneidx
< ZONE_NORMAL
)
2157 /* Check if we should retry the allocation */
2158 pages_reclaimed
+= did_some_progress
;
2159 if (should_alloc_retry(gfp_mask
, order
, pages_reclaimed
)) {
2160 /* Wait for some write requests to complete then retry */
2161 wait_iff_congested(preferred_zone
, BLK_RW_ASYNC
, HZ
/50);
2165 * High-order allocations do not necessarily loop after
2166 * direct reclaim and reclaim/compaction depends on compaction
2167 * being called after reclaim so call directly if necessary
2169 page
= __alloc_pages_direct_compact(gfp_mask
, order
,
2170 zonelist
, high_zoneidx
,
2172 alloc_flags
, preferred_zone
,
2173 migratetype
, &did_some_progress
,
2180 if (!(gfp_mask
& __GFP_NOWARN
) && printk_ratelimit()) {
2181 unsigned int filter
= SHOW_MEM_FILTER_NODES
;
2184 * This documents exceptions given to allocations in certain
2185 * contexts that are allowed to allocate outside current's set
2188 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
2189 if (test_thread_flag(TIF_MEMDIE
) ||
2190 (current
->flags
& (PF_MEMALLOC
| PF_EXITING
)))
2191 filter
&= ~SHOW_MEM_FILTER_NODES
;
2192 if (in_interrupt() || !wait
)
2193 filter
&= ~SHOW_MEM_FILTER_NODES
;
2195 pr_warning("%s: page allocation failure. order:%d, mode:0x%x\n",
2196 current
->comm
, order
, gfp_mask
);
2198 if (!should_suppress_show_mem())
2203 if (kmemcheck_enabled
)
2204 kmemcheck_pagealloc_alloc(page
, order
, gfp_mask
);
2210 * This is the 'heart' of the zoned buddy allocator.
2213 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
,
2214 struct zonelist
*zonelist
, nodemask_t
*nodemask
)
2216 enum zone_type high_zoneidx
= gfp_zone(gfp_mask
);
2217 struct zone
*preferred_zone
;
2219 int migratetype
= allocflags_to_migratetype(gfp_mask
);
2221 gfp_mask
&= gfp_allowed_mask
;
2223 lockdep_trace_alloc(gfp_mask
);
2225 might_sleep_if(gfp_mask
& __GFP_WAIT
);
2227 if (should_fail_alloc_page(gfp_mask
, order
))
2231 * Check the zones suitable for the gfp_mask contain at least one
2232 * valid zone. It's possible to have an empty zonelist as a result
2233 * of GFP_THISNODE and a memoryless node
2235 if (unlikely(!zonelist
->_zonerefs
->zone
))
2239 /* The preferred zone is used for statistics later */
2240 first_zones_zonelist(zonelist
, high_zoneidx
,
2241 nodemask
? : &cpuset_current_mems_allowed
,
2243 if (!preferred_zone
) {
2248 /* First allocation attempt */
2249 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
, order
,
2250 zonelist
, high_zoneidx
, ALLOC_WMARK_LOW
|ALLOC_CPUSET
,
2251 preferred_zone
, migratetype
);
2252 if (unlikely(!page
))
2253 page
= __alloc_pages_slowpath(gfp_mask
, order
,
2254 zonelist
, high_zoneidx
, nodemask
,
2255 preferred_zone
, migratetype
);
2258 trace_mm_page_alloc(page
, order
, gfp_mask
, migratetype
);
2261 EXPORT_SYMBOL(__alloc_pages_nodemask
);
2264 * Common helper functions.
2266 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
2271 * __get_free_pages() returns a 32-bit address, which cannot represent
2274 VM_BUG_ON((gfp_mask
& __GFP_HIGHMEM
) != 0);
2276 page
= alloc_pages(gfp_mask
, order
);
2279 return (unsigned long) page_address(page
);
2281 EXPORT_SYMBOL(__get_free_pages
);
2283 unsigned long get_zeroed_page(gfp_t gfp_mask
)
2285 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
2287 EXPORT_SYMBOL(get_zeroed_page
);
2289 void __pagevec_free(struct pagevec
*pvec
)
2291 int i
= pagevec_count(pvec
);
2294 trace_mm_pagevec_free(pvec
->pages
[i
], pvec
->cold
);
2295 free_hot_cold_page(pvec
->pages
[i
], pvec
->cold
);
2299 void __free_pages(struct page
*page
, unsigned int order
)
2301 if (put_page_testzero(page
)) {
2303 free_hot_cold_page(page
, 0);
2305 __free_pages_ok(page
, order
);
2309 EXPORT_SYMBOL(__free_pages
);
2311 void free_pages(unsigned long addr
, unsigned int order
)
2314 VM_BUG_ON(!virt_addr_valid((void *)addr
));
2315 __free_pages(virt_to_page((void *)addr
), order
);
2319 EXPORT_SYMBOL(free_pages
);
2321 static void *make_alloc_exact(unsigned long addr
, unsigned order
, size_t size
)
2324 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
2325 unsigned long used
= addr
+ PAGE_ALIGN(size
);
2327 split_page(virt_to_page((void *)addr
), order
);
2328 while (used
< alloc_end
) {
2333 return (void *)addr
;
2337 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2338 * @size: the number of bytes to allocate
2339 * @gfp_mask: GFP flags for the allocation
2341 * This function is similar to alloc_pages(), except that it allocates the
2342 * minimum number of pages to satisfy the request. alloc_pages() can only
2343 * allocate memory in power-of-two pages.
2345 * This function is also limited by MAX_ORDER.
2347 * Memory allocated by this function must be released by free_pages_exact().
2349 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
2351 unsigned int order
= get_order(size
);
2354 addr
= __get_free_pages(gfp_mask
, order
);
2355 return make_alloc_exact(addr
, order
, size
);
2357 EXPORT_SYMBOL(alloc_pages_exact
);
2360 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2362 * @nid: the preferred node ID where memory should be allocated
2363 * @size: the number of bytes to allocate
2364 * @gfp_mask: GFP flags for the allocation
2366 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2368 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2371 void *alloc_pages_exact_nid(int nid
, size_t size
, gfp_t gfp_mask
)
2373 unsigned order
= get_order(size
);
2374 struct page
*p
= alloc_pages_node(nid
, gfp_mask
, order
);
2377 return make_alloc_exact((unsigned long)page_address(p
), order
, size
);
2379 EXPORT_SYMBOL(alloc_pages_exact_nid
);
2382 * free_pages_exact - release memory allocated via alloc_pages_exact()
2383 * @virt: the value returned by alloc_pages_exact.
2384 * @size: size of allocation, same value as passed to alloc_pages_exact().
2386 * Release the memory allocated by a previous call to alloc_pages_exact.
2388 void free_pages_exact(void *virt
, size_t size
)
2390 unsigned long addr
= (unsigned long)virt
;
2391 unsigned long end
= addr
+ PAGE_ALIGN(size
);
2393 while (addr
< end
) {
2398 EXPORT_SYMBOL(free_pages_exact
);
2400 static unsigned int nr_free_zone_pages(int offset
)
2405 /* Just pick one node, since fallback list is circular */
2406 unsigned int sum
= 0;
2408 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
2410 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
2411 unsigned long size
= zone
->present_pages
;
2412 unsigned long high
= high_wmark_pages(zone
);
2421 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2423 unsigned int nr_free_buffer_pages(void)
2425 return nr_free_zone_pages(gfp_zone(GFP_USER
));
2427 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
2430 * Amount of free RAM allocatable within all zones
2432 unsigned int nr_free_pagecache_pages(void)
2434 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
2437 static inline void show_node(struct zone
*zone
)
2440 printk("Node %d ", zone_to_nid(zone
));
2443 void si_meminfo(struct sysinfo
*val
)
2445 val
->totalram
= totalram_pages
;
2447 val
->freeram
= global_page_state(NR_FREE_PAGES
);
2448 val
->bufferram
= nr_blockdev_pages();
2449 val
->totalhigh
= totalhigh_pages
;
2450 val
->freehigh
= nr_free_highpages();
2451 val
->mem_unit
= PAGE_SIZE
;
2454 EXPORT_SYMBOL(si_meminfo
);
2457 void si_meminfo_node(struct sysinfo
*val
, int nid
)
2459 pg_data_t
*pgdat
= NODE_DATA(nid
);
2461 val
->totalram
= pgdat
->node_present_pages
;
2462 val
->freeram
= node_page_state(nid
, NR_FREE_PAGES
);
2463 #ifdef CONFIG_HIGHMEM
2464 val
->totalhigh
= pgdat
->node_zones
[ZONE_HIGHMEM
].present_pages
;
2465 val
->freehigh
= zone_page_state(&pgdat
->node_zones
[ZONE_HIGHMEM
],
2471 val
->mem_unit
= PAGE_SIZE
;
2476 * Determine whether the node should be displayed or not, depending on whether
2477 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
2479 bool skip_free_areas_node(unsigned int flags
, int nid
)
2483 if (!(flags
& SHOW_MEM_FILTER_NODES
))
2487 ret
= !node_isset(nid
, cpuset_current_mems_allowed
);
2493 #define K(x) ((x) << (PAGE_SHIFT-10))
2496 * Show free area list (used inside shift_scroll-lock stuff)
2497 * We also calculate the percentage fragmentation. We do this by counting the
2498 * memory on each free list with the exception of the first item on the list.
2499 * Suppresses nodes that are not allowed by current's cpuset if
2500 * SHOW_MEM_FILTER_NODES is passed.
2502 void show_free_areas(unsigned int filter
)
2507 for_each_populated_zone(zone
) {
2508 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
2511 printk("%s per-cpu:\n", zone
->name
);
2513 for_each_online_cpu(cpu
) {
2514 struct per_cpu_pageset
*pageset
;
2516 pageset
= per_cpu_ptr(zone
->pageset
, cpu
);
2518 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2519 cpu
, pageset
->pcp
.high
,
2520 pageset
->pcp
.batch
, pageset
->pcp
.count
);
2524 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2525 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2527 " dirty:%lu writeback:%lu unstable:%lu\n"
2528 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2529 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n",
2530 global_page_state(NR_ACTIVE_ANON
),
2531 global_page_state(NR_INACTIVE_ANON
),
2532 global_page_state(NR_ISOLATED_ANON
),
2533 global_page_state(NR_ACTIVE_FILE
),
2534 global_page_state(NR_INACTIVE_FILE
),
2535 global_page_state(NR_ISOLATED_FILE
),
2536 global_page_state(NR_UNEVICTABLE
),
2537 global_page_state(NR_FILE_DIRTY
),
2538 global_page_state(NR_WRITEBACK
),
2539 global_page_state(NR_UNSTABLE_NFS
),
2540 global_page_state(NR_FREE_PAGES
),
2541 global_page_state(NR_SLAB_RECLAIMABLE
),
2542 global_page_state(NR_SLAB_UNRECLAIMABLE
),
2543 global_page_state(NR_FILE_MAPPED
),
2544 global_page_state(NR_SHMEM
),
2545 global_page_state(NR_PAGETABLE
),
2546 global_page_state(NR_BOUNCE
));
2548 for_each_populated_zone(zone
) {
2551 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
2559 " active_anon:%lukB"
2560 " inactive_anon:%lukB"
2561 " active_file:%lukB"
2562 " inactive_file:%lukB"
2563 " unevictable:%lukB"
2564 " isolated(anon):%lukB"
2565 " isolated(file):%lukB"
2572 " slab_reclaimable:%lukB"
2573 " slab_unreclaimable:%lukB"
2574 " kernel_stack:%lukB"
2578 " writeback_tmp:%lukB"
2579 " pages_scanned:%lu"
2580 " all_unreclaimable? %s"
2583 K(zone_page_state(zone
, NR_FREE_PAGES
)),
2584 K(min_wmark_pages(zone
)),
2585 K(low_wmark_pages(zone
)),
2586 K(high_wmark_pages(zone
)),
2587 K(zone_page_state(zone
, NR_ACTIVE_ANON
)),
2588 K(zone_page_state(zone
, NR_INACTIVE_ANON
)),
2589 K(zone_page_state(zone
, NR_ACTIVE_FILE
)),
2590 K(zone_page_state(zone
, NR_INACTIVE_FILE
)),
2591 K(zone_page_state(zone
, NR_UNEVICTABLE
)),
2592 K(zone_page_state(zone
, NR_ISOLATED_ANON
)),
2593 K(zone_page_state(zone
, NR_ISOLATED_FILE
)),
2594 K(zone
->present_pages
),
2595 K(zone_page_state(zone
, NR_MLOCK
)),
2596 K(zone_page_state(zone
, NR_FILE_DIRTY
)),
2597 K(zone_page_state(zone
, NR_WRITEBACK
)),
2598 K(zone_page_state(zone
, NR_FILE_MAPPED
)),
2599 K(zone_page_state(zone
, NR_SHMEM
)),
2600 K(zone_page_state(zone
, NR_SLAB_RECLAIMABLE
)),
2601 K(zone_page_state(zone
, NR_SLAB_UNRECLAIMABLE
)),
2602 zone_page_state(zone
, NR_KERNEL_STACK
) *
2604 K(zone_page_state(zone
, NR_PAGETABLE
)),
2605 K(zone_page_state(zone
, NR_UNSTABLE_NFS
)),
2606 K(zone_page_state(zone
, NR_BOUNCE
)),
2607 K(zone_page_state(zone
, NR_WRITEBACK_TEMP
)),
2608 zone
->pages_scanned
,
2609 (zone
->all_unreclaimable
? "yes" : "no")
2611 printk("lowmem_reserve[]:");
2612 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
2613 printk(" %lu", zone
->lowmem_reserve
[i
]);
2617 for_each_populated_zone(zone
) {
2618 unsigned long nr
[MAX_ORDER
], flags
, order
, total
= 0;
2620 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
2623 printk("%s: ", zone
->name
);
2625 spin_lock_irqsave(&zone
->lock
, flags
);
2626 for (order
= 0; order
< MAX_ORDER
; order
++) {
2627 nr
[order
] = zone
->free_area
[order
].nr_free
;
2628 total
+= nr
[order
] << order
;
2630 spin_unlock_irqrestore(&zone
->lock
, flags
);
2631 for (order
= 0; order
< MAX_ORDER
; order
++)
2632 printk("%lu*%lukB ", nr
[order
], K(1UL) << order
);
2633 printk("= %lukB\n", K(total
));
2636 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES
));
2638 show_swap_cache_info();
2641 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
2643 zoneref
->zone
= zone
;
2644 zoneref
->zone_idx
= zone_idx(zone
);
2648 * Builds allocation fallback zone lists.
2650 * Add all populated zones of a node to the zonelist.
2652 static int build_zonelists_node(pg_data_t
*pgdat
, struct zonelist
*zonelist
,
2653 int nr_zones
, enum zone_type zone_type
)
2657 BUG_ON(zone_type
>= MAX_NR_ZONES
);
2662 zone
= pgdat
->node_zones
+ zone_type
;
2663 if (populated_zone(zone
)) {
2664 zoneref_set_zone(zone
,
2665 &zonelist
->_zonerefs
[nr_zones
++]);
2666 check_highest_zone(zone_type
);
2669 } while (zone_type
);
2676 * 0 = automatic detection of better ordering.
2677 * 1 = order by ([node] distance, -zonetype)
2678 * 2 = order by (-zonetype, [node] distance)
2680 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
2681 * the same zonelist. So only NUMA can configure this param.
2683 #define ZONELIST_ORDER_DEFAULT 0
2684 #define ZONELIST_ORDER_NODE 1
2685 #define ZONELIST_ORDER_ZONE 2
2687 /* zonelist order in the kernel.
2688 * set_zonelist_order() will set this to NODE or ZONE.
2690 static int current_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
2691 static char zonelist_order_name
[3][8] = {"Default", "Node", "Zone"};
2695 /* The value user specified ....changed by config */
2696 static int user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
2697 /* string for sysctl */
2698 #define NUMA_ZONELIST_ORDER_LEN 16
2699 char numa_zonelist_order
[16] = "default";
2702 * interface for configure zonelist ordering.
2703 * command line option "numa_zonelist_order"
2704 * = "[dD]efault - default, automatic configuration.
2705 * = "[nN]ode - order by node locality, then by zone within node
2706 * = "[zZ]one - order by zone, then by locality within zone
2709 static int __parse_numa_zonelist_order(char *s
)
2711 if (*s
== 'd' || *s
== 'D') {
2712 user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
2713 } else if (*s
== 'n' || *s
== 'N') {
2714 user_zonelist_order
= ZONELIST_ORDER_NODE
;
2715 } else if (*s
== 'z' || *s
== 'Z') {
2716 user_zonelist_order
= ZONELIST_ORDER_ZONE
;
2719 "Ignoring invalid numa_zonelist_order value: "
2726 static __init
int setup_numa_zonelist_order(char *s
)
2733 ret
= __parse_numa_zonelist_order(s
);
2735 strlcpy(numa_zonelist_order
, s
, NUMA_ZONELIST_ORDER_LEN
);
2739 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
2742 * sysctl handler for numa_zonelist_order
2744 int numa_zonelist_order_handler(ctl_table
*table
, int write
,
2745 void __user
*buffer
, size_t *length
,
2748 char saved_string
[NUMA_ZONELIST_ORDER_LEN
];
2750 static DEFINE_MUTEX(zl_order_mutex
);
2752 mutex_lock(&zl_order_mutex
);
2754 strcpy(saved_string
, (char*)table
->data
);
2755 ret
= proc_dostring(table
, write
, buffer
, length
, ppos
);
2759 int oldval
= user_zonelist_order
;
2760 if (__parse_numa_zonelist_order((char*)table
->data
)) {
2762 * bogus value. restore saved string
2764 strncpy((char*)table
->data
, saved_string
,
2765 NUMA_ZONELIST_ORDER_LEN
);
2766 user_zonelist_order
= oldval
;
2767 } else if (oldval
!= user_zonelist_order
) {
2768 mutex_lock(&zonelists_mutex
);
2769 build_all_zonelists(NULL
);
2770 mutex_unlock(&zonelists_mutex
);
2774 mutex_unlock(&zl_order_mutex
);
2779 #define MAX_NODE_LOAD (nr_online_nodes)
2780 static int node_load
[MAX_NUMNODES
];
2783 * find_next_best_node - find the next node that should appear in a given node's fallback list
2784 * @node: node whose fallback list we're appending
2785 * @used_node_mask: nodemask_t of already used nodes
2787 * We use a number of factors to determine which is the next node that should
2788 * appear on a given node's fallback list. The node should not have appeared
2789 * already in @node's fallback list, and it should be the next closest node
2790 * according to the distance array (which contains arbitrary distance values
2791 * from each node to each node in the system), and should also prefer nodes
2792 * with no CPUs, since presumably they'll have very little allocation pressure
2793 * on them otherwise.
2794 * It returns -1 if no node is found.
2796 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
2799 int min_val
= INT_MAX
;
2801 const struct cpumask
*tmp
= cpumask_of_node(0);
2803 /* Use the local node if we haven't already */
2804 if (!node_isset(node
, *used_node_mask
)) {
2805 node_set(node
, *used_node_mask
);
2809 for_each_node_state(n
, N_HIGH_MEMORY
) {
2811 /* Don't want a node to appear more than once */
2812 if (node_isset(n
, *used_node_mask
))
2815 /* Use the distance array to find the distance */
2816 val
= node_distance(node
, n
);
2818 /* Penalize nodes under us ("prefer the next node") */
2821 /* Give preference to headless and unused nodes */
2822 tmp
= cpumask_of_node(n
);
2823 if (!cpumask_empty(tmp
))
2824 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
2826 /* Slight preference for less loaded node */
2827 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
2828 val
+= node_load
[n
];
2830 if (val
< min_val
) {
2837 node_set(best_node
, *used_node_mask
);
2844 * Build zonelists ordered by node and zones within node.
2845 * This results in maximum locality--normal zone overflows into local
2846 * DMA zone, if any--but risks exhausting DMA zone.
2848 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int node
)
2851 struct zonelist
*zonelist
;
2853 zonelist
= &pgdat
->node_zonelists
[0];
2854 for (j
= 0; zonelist
->_zonerefs
[j
].zone
!= NULL
; j
++)
2856 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
2858 zonelist
->_zonerefs
[j
].zone
= NULL
;
2859 zonelist
->_zonerefs
[j
].zone_idx
= 0;
2863 * Build gfp_thisnode zonelists
2865 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
2868 struct zonelist
*zonelist
;
2870 zonelist
= &pgdat
->node_zonelists
[1];
2871 j
= build_zonelists_node(pgdat
, zonelist
, 0, MAX_NR_ZONES
- 1);
2872 zonelist
->_zonerefs
[j
].zone
= NULL
;
2873 zonelist
->_zonerefs
[j
].zone_idx
= 0;
2877 * Build zonelists ordered by zone and nodes within zones.
2878 * This results in conserving DMA zone[s] until all Normal memory is
2879 * exhausted, but results in overflowing to remote node while memory
2880 * may still exist in local DMA zone.
2882 static int node_order
[MAX_NUMNODES
];
2884 static void build_zonelists_in_zone_order(pg_data_t
*pgdat
, int nr_nodes
)
2887 int zone_type
; /* needs to be signed */
2889 struct zonelist
*zonelist
;
2891 zonelist
= &pgdat
->node_zonelists
[0];
2893 for (zone_type
= MAX_NR_ZONES
- 1; zone_type
>= 0; zone_type
--) {
2894 for (j
= 0; j
< nr_nodes
; j
++) {
2895 node
= node_order
[j
];
2896 z
= &NODE_DATA(node
)->node_zones
[zone_type
];
2897 if (populated_zone(z
)) {
2899 &zonelist
->_zonerefs
[pos
++]);
2900 check_highest_zone(zone_type
);
2904 zonelist
->_zonerefs
[pos
].zone
= NULL
;
2905 zonelist
->_zonerefs
[pos
].zone_idx
= 0;
2908 static int default_zonelist_order(void)
2911 unsigned long low_kmem_size
,total_size
;
2915 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
2916 * If they are really small and used heavily, the system can fall
2917 * into OOM very easily.
2918 * This function detect ZONE_DMA/DMA32 size and configures zone order.
2920 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
2923 for_each_online_node(nid
) {
2924 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
2925 z
= &NODE_DATA(nid
)->node_zones
[zone_type
];
2926 if (populated_zone(z
)) {
2927 if (zone_type
< ZONE_NORMAL
)
2928 low_kmem_size
+= z
->present_pages
;
2929 total_size
+= z
->present_pages
;
2930 } else if (zone_type
== ZONE_NORMAL
) {
2932 * If any node has only lowmem, then node order
2933 * is preferred to allow kernel allocations
2934 * locally; otherwise, they can easily infringe
2935 * on other nodes when there is an abundance of
2936 * lowmem available to allocate from.
2938 return ZONELIST_ORDER_NODE
;
2942 if (!low_kmem_size
|| /* there are no DMA area. */
2943 low_kmem_size
> total_size
/2) /* DMA/DMA32 is big. */
2944 return ZONELIST_ORDER_NODE
;
2946 * look into each node's config.
2947 * If there is a node whose DMA/DMA32 memory is very big area on
2948 * local memory, NODE_ORDER may be suitable.
2950 average_size
= total_size
/
2951 (nodes_weight(node_states
[N_HIGH_MEMORY
]) + 1);
2952 for_each_online_node(nid
) {
2955 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
2956 z
= &NODE_DATA(nid
)->node_zones
[zone_type
];
2957 if (populated_zone(z
)) {
2958 if (zone_type
< ZONE_NORMAL
)
2959 low_kmem_size
+= z
->present_pages
;
2960 total_size
+= z
->present_pages
;
2963 if (low_kmem_size
&&
2964 total_size
> average_size
&& /* ignore small node */
2965 low_kmem_size
> total_size
* 70/100)
2966 return ZONELIST_ORDER_NODE
;
2968 return ZONELIST_ORDER_ZONE
;
2971 static void set_zonelist_order(void)
2973 if (user_zonelist_order
== ZONELIST_ORDER_DEFAULT
)
2974 current_zonelist_order
= default_zonelist_order();
2976 current_zonelist_order
= user_zonelist_order
;
2979 static void build_zonelists(pg_data_t
*pgdat
)
2983 nodemask_t used_mask
;
2984 int local_node
, prev_node
;
2985 struct zonelist
*zonelist
;
2986 int order
= current_zonelist_order
;
2988 /* initialize zonelists */
2989 for (i
= 0; i
< MAX_ZONELISTS
; i
++) {
2990 zonelist
= pgdat
->node_zonelists
+ i
;
2991 zonelist
->_zonerefs
[0].zone
= NULL
;
2992 zonelist
->_zonerefs
[0].zone_idx
= 0;
2995 /* NUMA-aware ordering of nodes */
2996 local_node
= pgdat
->node_id
;
2997 load
= nr_online_nodes
;
2998 prev_node
= local_node
;
2999 nodes_clear(used_mask
);
3001 memset(node_order
, 0, sizeof(node_order
));
3004 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
3005 int distance
= node_distance(local_node
, node
);
3008 * If another node is sufficiently far away then it is better
3009 * to reclaim pages in a zone before going off node.
3011 if (distance
> RECLAIM_DISTANCE
)
3012 zone_reclaim_mode
= 1;
3015 * We don't want to pressure a particular node.
3016 * So adding penalty to the first node in same
3017 * distance group to make it round-robin.
3019 if (distance
!= node_distance(local_node
, prev_node
))
3020 node_load
[node
] = load
;
3024 if (order
== ZONELIST_ORDER_NODE
)
3025 build_zonelists_in_node_order(pgdat
, node
);
3027 node_order
[j
++] = node
; /* remember order */
3030 if (order
== ZONELIST_ORDER_ZONE
) {
3031 /* calculate node order -- i.e., DMA last! */
3032 build_zonelists_in_zone_order(pgdat
, j
);
3035 build_thisnode_zonelists(pgdat
);
3038 /* Construct the zonelist performance cache - see further mmzone.h */
3039 static void build_zonelist_cache(pg_data_t
*pgdat
)
3041 struct zonelist
*zonelist
;
3042 struct zonelist_cache
*zlc
;
3045 zonelist
= &pgdat
->node_zonelists
[0];
3046 zonelist
->zlcache_ptr
= zlc
= &zonelist
->zlcache
;
3047 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
3048 for (z
= zonelist
->_zonerefs
; z
->zone
; z
++)
3049 zlc
->z_to_n
[z
- zonelist
->_zonerefs
] = zonelist_node_idx(z
);
3052 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3054 * Return node id of node used for "local" allocations.
3055 * I.e., first node id of first zone in arg node's generic zonelist.
3056 * Used for initializing percpu 'numa_mem', which is used primarily
3057 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3059 int local_memory_node(int node
)
3063 (void)first_zones_zonelist(node_zonelist(node
, GFP_KERNEL
),
3064 gfp_zone(GFP_KERNEL
),
3071 #else /* CONFIG_NUMA */
3073 static void set_zonelist_order(void)
3075 current_zonelist_order
= ZONELIST_ORDER_ZONE
;
3078 static void build_zonelists(pg_data_t
*pgdat
)
3080 int node
, local_node
;
3082 struct zonelist
*zonelist
;
3084 local_node
= pgdat
->node_id
;
3086 zonelist
= &pgdat
->node_zonelists
[0];
3087 j
= build_zonelists_node(pgdat
, zonelist
, 0, MAX_NR_ZONES
- 1);
3090 * Now we build the zonelist so that it contains the zones
3091 * of all the other nodes.
3092 * We don't want to pressure a particular node, so when
3093 * building the zones for node N, we make sure that the
3094 * zones coming right after the local ones are those from
3095 * node N+1 (modulo N)
3097 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
3098 if (!node_online(node
))
3100 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
3103 for (node
= 0; node
< local_node
; node
++) {
3104 if (!node_online(node
))
3106 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
3110 zonelist
->_zonerefs
[j
].zone
= NULL
;
3111 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3114 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3115 static void build_zonelist_cache(pg_data_t
*pgdat
)
3117 pgdat
->node_zonelists
[0].zlcache_ptr
= NULL
;
3120 #endif /* CONFIG_NUMA */
3123 * Boot pageset table. One per cpu which is going to be used for all
3124 * zones and all nodes. The parameters will be set in such a way
3125 * that an item put on a list will immediately be handed over to
3126 * the buddy list. This is safe since pageset manipulation is done
3127 * with interrupts disabled.
3129 * The boot_pagesets must be kept even after bootup is complete for
3130 * unused processors and/or zones. They do play a role for bootstrapping
3131 * hotplugged processors.
3133 * zoneinfo_show() and maybe other functions do
3134 * not check if the processor is online before following the pageset pointer.
3135 * Other parts of the kernel may not check if the zone is available.
3137 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
);
3138 static DEFINE_PER_CPU(struct per_cpu_pageset
, boot_pageset
);
3139 static void setup_zone_pageset(struct zone
*zone
);
3142 * Global mutex to protect against size modification of zonelists
3143 * as well as to serialize pageset setup for the new populated zone.
3145 DEFINE_MUTEX(zonelists_mutex
);
3147 /* return values int ....just for stop_machine() */
3148 static __init_refok
int __build_all_zonelists(void *data
)
3154 memset(node_load
, 0, sizeof(node_load
));
3156 for_each_online_node(nid
) {
3157 pg_data_t
*pgdat
= NODE_DATA(nid
);
3159 build_zonelists(pgdat
);
3160 build_zonelist_cache(pgdat
);
3164 * Initialize the boot_pagesets that are going to be used
3165 * for bootstrapping processors. The real pagesets for
3166 * each zone will be allocated later when the per cpu
3167 * allocator is available.
3169 * boot_pagesets are used also for bootstrapping offline
3170 * cpus if the system is already booted because the pagesets
3171 * are needed to initialize allocators on a specific cpu too.
3172 * F.e. the percpu allocator needs the page allocator which
3173 * needs the percpu allocator in order to allocate its pagesets
3174 * (a chicken-egg dilemma).
3176 for_each_possible_cpu(cpu
) {
3177 setup_pageset(&per_cpu(boot_pageset
, cpu
), 0);
3179 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3181 * We now know the "local memory node" for each node--
3182 * i.e., the node of the first zone in the generic zonelist.
3183 * Set up numa_mem percpu variable for on-line cpus. During
3184 * boot, only the boot cpu should be on-line; we'll init the
3185 * secondary cpus' numa_mem as they come on-line. During
3186 * node/memory hotplug, we'll fixup all on-line cpus.
3188 if (cpu_online(cpu
))
3189 set_cpu_numa_mem(cpu
, local_memory_node(cpu_to_node(cpu
)));
3197 * Called with zonelists_mutex held always
3198 * unless system_state == SYSTEM_BOOTING.
3200 void __ref
build_all_zonelists(void *data
)
3202 set_zonelist_order();
3204 if (system_state
== SYSTEM_BOOTING
) {
3205 __build_all_zonelists(NULL
);
3206 mminit_verify_zonelist();
3207 cpuset_init_current_mems_allowed();
3209 /* we have to stop all cpus to guarantee there is no user
3211 #ifdef CONFIG_MEMORY_HOTPLUG
3213 setup_zone_pageset((struct zone
*)data
);
3215 stop_machine(__build_all_zonelists
, NULL
, NULL
);
3216 /* cpuset refresh routine should be here */
3218 vm_total_pages
= nr_free_pagecache_pages();
3220 * Disable grouping by mobility if the number of pages in the
3221 * system is too low to allow the mechanism to work. It would be
3222 * more accurate, but expensive to check per-zone. This check is
3223 * made on memory-hotadd so a system can start with mobility
3224 * disabled and enable it later
3226 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
3227 page_group_by_mobility_disabled
= 1;
3229 page_group_by_mobility_disabled
= 0;
3231 printk("Built %i zonelists in %s order, mobility grouping %s. "
3232 "Total pages: %ld\n",
3234 zonelist_order_name
[current_zonelist_order
],
3235 page_group_by_mobility_disabled
? "off" : "on",
3238 printk("Policy zone: %s\n", zone_names
[policy_zone
]);
3243 * Helper functions to size the waitqueue hash table.
3244 * Essentially these want to choose hash table sizes sufficiently
3245 * large so that collisions trying to wait on pages are rare.
3246 * But in fact, the number of active page waitqueues on typical
3247 * systems is ridiculously low, less than 200. So this is even
3248 * conservative, even though it seems large.
3250 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3251 * waitqueues, i.e. the size of the waitq table given the number of pages.
3253 #define PAGES_PER_WAITQUEUE 256
3255 #ifndef CONFIG_MEMORY_HOTPLUG
3256 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
3258 unsigned long size
= 1;
3260 pages
/= PAGES_PER_WAITQUEUE
;
3262 while (size
< pages
)
3266 * Once we have dozens or even hundreds of threads sleeping
3267 * on IO we've got bigger problems than wait queue collision.
3268 * Limit the size of the wait table to a reasonable size.
3270 size
= min(size
, 4096UL);
3272 return max(size
, 4UL);
3276 * A zone's size might be changed by hot-add, so it is not possible to determine
3277 * a suitable size for its wait_table. So we use the maximum size now.
3279 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3281 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3282 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3283 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3285 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3286 * or more by the traditional way. (See above). It equals:
3288 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3289 * ia64(16K page size) : = ( 8G + 4M)byte.
3290 * powerpc (64K page size) : = (32G +16M)byte.
3292 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
3299 * This is an integer logarithm so that shifts can be used later
3300 * to extract the more random high bits from the multiplicative
3301 * hash function before the remainder is taken.
3303 static inline unsigned long wait_table_bits(unsigned long size
)
3308 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3311 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3312 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3313 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3314 * higher will lead to a bigger reserve which will get freed as contiguous
3315 * blocks as reclaim kicks in
3317 static void setup_zone_migrate_reserve(struct zone
*zone
)
3319 unsigned long start_pfn
, pfn
, end_pfn
;
3321 unsigned long block_migratetype
;
3324 /* Get the start pfn, end pfn and the number of blocks to reserve */
3325 start_pfn
= zone
->zone_start_pfn
;
3326 end_pfn
= start_pfn
+ zone
->spanned_pages
;
3327 reserve
= roundup(min_wmark_pages(zone
), pageblock_nr_pages
) >>
3331 * Reserve blocks are generally in place to help high-order atomic
3332 * allocations that are short-lived. A min_free_kbytes value that
3333 * would result in more than 2 reserve blocks for atomic allocations
3334 * is assumed to be in place to help anti-fragmentation for the
3335 * future allocation of hugepages at runtime.
3337 reserve
= min(2, reserve
);
3339 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= pageblock_nr_pages
) {
3340 if (!pfn_valid(pfn
))
3342 page
= pfn_to_page(pfn
);
3344 /* Watch out for overlapping nodes */
3345 if (page_to_nid(page
) != zone_to_nid(zone
))
3348 /* Blocks with reserved pages will never free, skip them. */
3349 if (PageReserved(page
))
3352 block_migratetype
= get_pageblock_migratetype(page
);
3354 /* If this block is reserved, account for it */
3355 if (reserve
> 0 && block_migratetype
== MIGRATE_RESERVE
) {
3360 /* Suitable for reserving if this block is movable */
3361 if (reserve
> 0 && block_migratetype
== MIGRATE_MOVABLE
) {
3362 set_pageblock_migratetype(page
, MIGRATE_RESERVE
);
3363 move_freepages_block(zone
, page
, MIGRATE_RESERVE
);
3369 * If the reserve is met and this is a previous reserved block,
3372 if (block_migratetype
== MIGRATE_RESERVE
) {
3373 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
3374 move_freepages_block(zone
, page
, MIGRATE_MOVABLE
);
3380 * Initially all pages are reserved - free ones are freed
3381 * up by free_all_bootmem() once the early boot process is
3382 * done. Non-atomic initialization, single-pass.
3384 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
3385 unsigned long start_pfn
, enum memmap_context context
)
3388 unsigned long end_pfn
= start_pfn
+ size
;
3392 if (highest_memmap_pfn
< end_pfn
- 1)
3393 highest_memmap_pfn
= end_pfn
- 1;
3395 z
= &NODE_DATA(nid
)->node_zones
[zone
];
3396 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
3398 * There can be holes in boot-time mem_map[]s
3399 * handed to this function. They do not
3400 * exist on hotplugged memory.
3402 if (context
== MEMMAP_EARLY
) {
3403 if (!early_pfn_valid(pfn
))
3405 if (!early_pfn_in_nid(pfn
, nid
))
3408 page
= pfn_to_page(pfn
);
3409 set_page_links(page
, zone
, nid
, pfn
);
3410 mminit_verify_page_links(page
, zone
, nid
, pfn
);
3411 init_page_count(page
);
3412 reset_page_mapcount(page
);
3413 SetPageReserved(page
);
3415 * Mark the block movable so that blocks are reserved for
3416 * movable at startup. This will force kernel allocations
3417 * to reserve their blocks rather than leaking throughout
3418 * the address space during boot when many long-lived
3419 * kernel allocations are made. Later some blocks near
3420 * the start are marked MIGRATE_RESERVE by
3421 * setup_zone_migrate_reserve()
3423 * bitmap is created for zone's valid pfn range. but memmap
3424 * can be created for invalid pages (for alignment)
3425 * check here not to call set_pageblock_migratetype() against
3428 if ((z
->zone_start_pfn
<= pfn
)
3429 && (pfn
< z
->zone_start_pfn
+ z
->spanned_pages
)
3430 && !(pfn
& (pageblock_nr_pages
- 1)))
3431 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
3433 INIT_LIST_HEAD(&page
->lru
);
3434 #ifdef WANT_PAGE_VIRTUAL
3435 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3436 if (!is_highmem_idx(zone
))
3437 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
3442 static void __meminit
zone_init_free_lists(struct zone
*zone
)
3445 for_each_migratetype_order(order
, t
) {
3446 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
3447 zone
->free_area
[order
].nr_free
= 0;
3451 #ifndef __HAVE_ARCH_MEMMAP_INIT
3452 #define memmap_init(size, nid, zone, start_pfn) \
3453 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3456 static int zone_batchsize(struct zone
*zone
)
3462 * The per-cpu-pages pools are set to around 1000th of the
3463 * size of the zone. But no more than 1/2 of a meg.
3465 * OK, so we don't know how big the cache is. So guess.
3467 batch
= zone
->present_pages
/ 1024;
3468 if (batch
* PAGE_SIZE
> 512 * 1024)
3469 batch
= (512 * 1024) / PAGE_SIZE
;
3470 batch
/= 4; /* We effectively *= 4 below */
3475 * Clamp the batch to a 2^n - 1 value. Having a power
3476 * of 2 value was found to be more likely to have
3477 * suboptimal cache aliasing properties in some cases.
3479 * For example if 2 tasks are alternately allocating
3480 * batches of pages, one task can end up with a lot
3481 * of pages of one half of the possible page colors
3482 * and the other with pages of the other colors.
3484 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
3489 /* The deferral and batching of frees should be suppressed under NOMMU
3492 * The problem is that NOMMU needs to be able to allocate large chunks
3493 * of contiguous memory as there's no hardware page translation to
3494 * assemble apparent contiguous memory from discontiguous pages.
3496 * Queueing large contiguous runs of pages for batching, however,
3497 * causes the pages to actually be freed in smaller chunks. As there
3498 * can be a significant delay between the individual batches being
3499 * recycled, this leads to the once large chunks of space being
3500 * fragmented and becoming unavailable for high-order allocations.
3506 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
3508 struct per_cpu_pages
*pcp
;
3511 memset(p
, 0, sizeof(*p
));
3515 pcp
->high
= 6 * batch
;
3516 pcp
->batch
= max(1UL, 1 * batch
);
3517 for (migratetype
= 0; migratetype
< MIGRATE_PCPTYPES
; migratetype
++)
3518 INIT_LIST_HEAD(&pcp
->lists
[migratetype
]);
3522 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3523 * to the value high for the pageset p.
3526 static void setup_pagelist_highmark(struct per_cpu_pageset
*p
,
3529 struct per_cpu_pages
*pcp
;
3533 pcp
->batch
= max(1UL, high
/4);
3534 if ((high
/4) > (PAGE_SHIFT
* 8))
3535 pcp
->batch
= PAGE_SHIFT
* 8;
3538 static void setup_zone_pageset(struct zone
*zone
)
3542 zone
->pageset
= alloc_percpu(struct per_cpu_pageset
);
3544 for_each_possible_cpu(cpu
) {
3545 struct per_cpu_pageset
*pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
3547 setup_pageset(pcp
, zone_batchsize(zone
));
3549 if (percpu_pagelist_fraction
)
3550 setup_pagelist_highmark(pcp
,
3551 (zone
->present_pages
/
3552 percpu_pagelist_fraction
));
3557 * Allocate per cpu pagesets and initialize them.
3558 * Before this call only boot pagesets were available.
3560 void __init
setup_per_cpu_pageset(void)
3564 for_each_populated_zone(zone
)
3565 setup_zone_pageset(zone
);
3568 static noinline __init_refok
3569 int zone_wait_table_init(struct zone
*zone
, unsigned long zone_size_pages
)
3572 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
3576 * The per-page waitqueue mechanism uses hashed waitqueues
3579 zone
->wait_table_hash_nr_entries
=
3580 wait_table_hash_nr_entries(zone_size_pages
);
3581 zone
->wait_table_bits
=
3582 wait_table_bits(zone
->wait_table_hash_nr_entries
);
3583 alloc_size
= zone
->wait_table_hash_nr_entries
3584 * sizeof(wait_queue_head_t
);
3586 if (!slab_is_available()) {
3587 zone
->wait_table
= (wait_queue_head_t
*)
3588 alloc_bootmem_node_nopanic(pgdat
, alloc_size
);
3591 * This case means that a zone whose size was 0 gets new memory
3592 * via memory hot-add.
3593 * But it may be the case that a new node was hot-added. In
3594 * this case vmalloc() will not be able to use this new node's
3595 * memory - this wait_table must be initialized to use this new
3596 * node itself as well.
3597 * To use this new node's memory, further consideration will be
3600 zone
->wait_table
= vmalloc(alloc_size
);
3602 if (!zone
->wait_table
)
3605 for(i
= 0; i
< zone
->wait_table_hash_nr_entries
; ++i
)
3606 init_waitqueue_head(zone
->wait_table
+ i
);
3611 static int __zone_pcp_update(void *data
)
3613 struct zone
*zone
= data
;
3615 unsigned long batch
= zone_batchsize(zone
), flags
;
3617 for_each_possible_cpu(cpu
) {
3618 struct per_cpu_pageset
*pset
;
3619 struct per_cpu_pages
*pcp
;
3621 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
3624 local_irq_save(flags
);
3625 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
3626 setup_pageset(pset
, batch
);
3627 local_irq_restore(flags
);
3632 void zone_pcp_update(struct zone
*zone
)
3634 stop_machine(__zone_pcp_update
, zone
, NULL
);
3637 static __meminit
void zone_pcp_init(struct zone
*zone
)
3640 * per cpu subsystem is not up at this point. The following code
3641 * relies on the ability of the linker to provide the
3642 * offset of a (static) per cpu variable into the per cpu area.
3644 zone
->pageset
= &boot_pageset
;
3646 if (zone
->present_pages
)
3647 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%u\n",
3648 zone
->name
, zone
->present_pages
,
3649 zone_batchsize(zone
));
3652 __meminit
int init_currently_empty_zone(struct zone
*zone
,
3653 unsigned long zone_start_pfn
,
3655 enum memmap_context context
)
3657 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
3659 ret
= zone_wait_table_init(zone
, size
);
3662 pgdat
->nr_zones
= zone_idx(zone
) + 1;
3664 zone
->zone_start_pfn
= zone_start_pfn
;
3666 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
3667 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
3669 (unsigned long)zone_idx(zone
),
3670 zone_start_pfn
, (zone_start_pfn
+ size
));
3672 zone_init_free_lists(zone
);
3677 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3679 * Basic iterator support. Return the first range of PFNs for a node
3680 * Note: nid == MAX_NUMNODES returns first region regardless of node
3682 static int __meminit
first_active_region_index_in_nid(int nid
)
3686 for (i
= 0; i
< nr_nodemap_entries
; i
++)
3687 if (nid
== MAX_NUMNODES
|| early_node_map
[i
].nid
== nid
)
3694 * Basic iterator support. Return the next active range of PFNs for a node
3695 * Note: nid == MAX_NUMNODES returns next region regardless of node
3697 static int __meminit
next_active_region_index_in_nid(int index
, int nid
)
3699 for (index
= index
+ 1; index
< nr_nodemap_entries
; index
++)
3700 if (nid
== MAX_NUMNODES
|| early_node_map
[index
].nid
== nid
)
3706 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
3708 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
3709 * Architectures may implement their own version but if add_active_range()
3710 * was used and there are no special requirements, this is a convenient
3713 int __meminit
__early_pfn_to_nid(unsigned long pfn
)
3717 for (i
= 0; i
< nr_nodemap_entries
; i
++) {
3718 unsigned long start_pfn
= early_node_map
[i
].start_pfn
;
3719 unsigned long end_pfn
= early_node_map
[i
].end_pfn
;
3721 if (start_pfn
<= pfn
&& pfn
< end_pfn
)
3722 return early_node_map
[i
].nid
;
3724 /* This is a memory hole */
3727 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
3729 int __meminit
early_pfn_to_nid(unsigned long pfn
)
3733 nid
= __early_pfn_to_nid(pfn
);
3736 /* just returns 0 */
3740 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
3741 bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
3745 nid
= __early_pfn_to_nid(pfn
);
3746 if (nid
>= 0 && nid
!= node
)
3752 /* Basic iterator support to walk early_node_map[] */
3753 #define for_each_active_range_index_in_nid(i, nid) \
3754 for (i = first_active_region_index_in_nid(nid); i != -1; \
3755 i = next_active_region_index_in_nid(i, nid))
3758 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
3759 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
3760 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
3762 * If an architecture guarantees that all ranges registered with
3763 * add_active_ranges() contain no holes and may be freed, this
3764 * this function may be used instead of calling free_bootmem() manually.
3766 void __init
free_bootmem_with_active_regions(int nid
,
3767 unsigned long max_low_pfn
)
3771 for_each_active_range_index_in_nid(i
, nid
) {
3772 unsigned long size_pages
= 0;
3773 unsigned long end_pfn
= early_node_map
[i
].end_pfn
;
3775 if (early_node_map
[i
].start_pfn
>= max_low_pfn
)
3778 if (end_pfn
> max_low_pfn
)
3779 end_pfn
= max_low_pfn
;
3781 size_pages
= end_pfn
- early_node_map
[i
].start_pfn
;
3782 free_bootmem_node(NODE_DATA(early_node_map
[i
].nid
),
3783 PFN_PHYS(early_node_map
[i
].start_pfn
),
3784 size_pages
<< PAGE_SHIFT
);
3788 #ifdef CONFIG_HAVE_MEMBLOCK
3790 * Basic iterator support. Return the last range of PFNs for a node
3791 * Note: nid == MAX_NUMNODES returns last region regardless of node
3793 static int __meminit
last_active_region_index_in_nid(int nid
)
3797 for (i
= nr_nodemap_entries
- 1; i
>= 0; i
--)
3798 if (nid
== MAX_NUMNODES
|| early_node_map
[i
].nid
== nid
)
3805 * Basic iterator support. Return the previous active range of PFNs for a node
3806 * Note: nid == MAX_NUMNODES returns next region regardless of node
3808 static int __meminit
previous_active_region_index_in_nid(int index
, int nid
)
3810 for (index
= index
- 1; index
>= 0; index
--)
3811 if (nid
== MAX_NUMNODES
|| early_node_map
[index
].nid
== nid
)
3817 #define for_each_active_range_index_in_nid_reverse(i, nid) \
3818 for (i = last_active_region_index_in_nid(nid); i != -1; \
3819 i = previous_active_region_index_in_nid(i, nid))
3821 u64 __init
find_memory_core_early(int nid
, u64 size
, u64 align
,
3822 u64 goal
, u64 limit
)
3826 /* Need to go over early_node_map to find out good range for node */
3827 for_each_active_range_index_in_nid_reverse(i
, nid
) {
3829 u64 ei_start
, ei_last
;
3830 u64 final_start
, final_end
;
3832 ei_last
= early_node_map
[i
].end_pfn
;
3833 ei_last
<<= PAGE_SHIFT
;
3834 ei_start
= early_node_map
[i
].start_pfn
;
3835 ei_start
<<= PAGE_SHIFT
;
3837 final_start
= max(ei_start
, goal
);
3838 final_end
= min(ei_last
, limit
);
3840 if (final_start
>= final_end
)
3843 addr
= memblock_find_in_range(final_start
, final_end
, size
, align
);
3845 if (addr
== MEMBLOCK_ERROR
)
3851 return MEMBLOCK_ERROR
;
3855 int __init
add_from_early_node_map(struct range
*range
, int az
,
3856 int nr_range
, int nid
)
3861 /* need to go over early_node_map to find out good range for node */
3862 for_each_active_range_index_in_nid(i
, nid
) {
3863 start
= early_node_map
[i
].start_pfn
;
3864 end
= early_node_map
[i
].end_pfn
;
3865 nr_range
= add_range(range
, az
, nr_range
, start
, end
);
3870 void __init
work_with_active_regions(int nid
, work_fn_t work_fn
, void *data
)
3875 for_each_active_range_index_in_nid(i
, nid
) {
3876 ret
= work_fn(early_node_map
[i
].start_pfn
,
3877 early_node_map
[i
].end_pfn
, data
);
3883 * sparse_memory_present_with_active_regions - Call memory_present for each active range
3884 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
3886 * If an architecture guarantees that all ranges registered with
3887 * add_active_ranges() contain no holes and may be freed, this
3888 * function may be used instead of calling memory_present() manually.
3890 void __init
sparse_memory_present_with_active_regions(int nid
)
3894 for_each_active_range_index_in_nid(i
, nid
)
3895 memory_present(early_node_map
[i
].nid
,
3896 early_node_map
[i
].start_pfn
,
3897 early_node_map
[i
].end_pfn
);
3901 * get_pfn_range_for_nid - Return the start and end page frames for a node
3902 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
3903 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
3904 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
3906 * It returns the start and end page frame of a node based on information
3907 * provided by an arch calling add_active_range(). If called for a node
3908 * with no available memory, a warning is printed and the start and end
3911 void __meminit
get_pfn_range_for_nid(unsigned int nid
,
3912 unsigned long *start_pfn
, unsigned long *end_pfn
)
3918 for_each_active_range_index_in_nid(i
, nid
) {
3919 *start_pfn
= min(*start_pfn
, early_node_map
[i
].start_pfn
);
3920 *end_pfn
= max(*end_pfn
, early_node_map
[i
].end_pfn
);
3923 if (*start_pfn
== -1UL)
3928 * This finds a zone that can be used for ZONE_MOVABLE pages. The
3929 * assumption is made that zones within a node are ordered in monotonic
3930 * increasing memory addresses so that the "highest" populated zone is used
3932 static void __init
find_usable_zone_for_movable(void)
3935 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
3936 if (zone_index
== ZONE_MOVABLE
)
3939 if (arch_zone_highest_possible_pfn
[zone_index
] >
3940 arch_zone_lowest_possible_pfn
[zone_index
])
3944 VM_BUG_ON(zone_index
== -1);
3945 movable_zone
= zone_index
;
3949 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
3950 * because it is sized independent of architecture. Unlike the other zones,
3951 * the starting point for ZONE_MOVABLE is not fixed. It may be different
3952 * in each node depending on the size of each node and how evenly kernelcore
3953 * is distributed. This helper function adjusts the zone ranges
3954 * provided by the architecture for a given node by using the end of the
3955 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
3956 * zones within a node are in order of monotonic increases memory addresses
3958 static void __meminit
adjust_zone_range_for_zone_movable(int nid
,
3959 unsigned long zone_type
,
3960 unsigned long node_start_pfn
,
3961 unsigned long node_end_pfn
,
3962 unsigned long *zone_start_pfn
,
3963 unsigned long *zone_end_pfn
)
3965 /* Only adjust if ZONE_MOVABLE is on this node */
3966 if (zone_movable_pfn
[nid
]) {
3967 /* Size ZONE_MOVABLE */
3968 if (zone_type
== ZONE_MOVABLE
) {
3969 *zone_start_pfn
= zone_movable_pfn
[nid
];
3970 *zone_end_pfn
= min(node_end_pfn
,
3971 arch_zone_highest_possible_pfn
[movable_zone
]);
3973 /* Adjust for ZONE_MOVABLE starting within this range */
3974 } else if (*zone_start_pfn
< zone_movable_pfn
[nid
] &&
3975 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
3976 *zone_end_pfn
= zone_movable_pfn
[nid
];
3978 /* Check if this whole range is within ZONE_MOVABLE */
3979 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
3980 *zone_start_pfn
= *zone_end_pfn
;
3985 * Return the number of pages a zone spans in a node, including holes
3986 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
3988 static unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
3989 unsigned long zone_type
,
3990 unsigned long *ignored
)
3992 unsigned long node_start_pfn
, node_end_pfn
;
3993 unsigned long zone_start_pfn
, zone_end_pfn
;
3995 /* Get the start and end of the node and zone */
3996 get_pfn_range_for_nid(nid
, &node_start_pfn
, &node_end_pfn
);
3997 zone_start_pfn
= arch_zone_lowest_possible_pfn
[zone_type
];
3998 zone_end_pfn
= arch_zone_highest_possible_pfn
[zone_type
];
3999 adjust_zone_range_for_zone_movable(nid
, zone_type
,
4000 node_start_pfn
, node_end_pfn
,
4001 &zone_start_pfn
, &zone_end_pfn
);
4003 /* Check that this node has pages within the zone's required range */
4004 if (zone_end_pfn
< node_start_pfn
|| zone_start_pfn
> node_end_pfn
)
4007 /* Move the zone boundaries inside the node if necessary */
4008 zone_end_pfn
= min(zone_end_pfn
, node_end_pfn
);
4009 zone_start_pfn
= max(zone_start_pfn
, node_start_pfn
);
4011 /* Return the spanned pages */
4012 return zone_end_pfn
- zone_start_pfn
;
4016 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4017 * then all holes in the requested range will be accounted for.
4019 unsigned long __meminit
__absent_pages_in_range(int nid
,
4020 unsigned long range_start_pfn
,
4021 unsigned long range_end_pfn
)
4024 unsigned long prev_end_pfn
= 0, hole_pages
= 0;
4025 unsigned long start_pfn
;
4027 /* Find the end_pfn of the first active range of pfns in the node */
4028 i
= first_active_region_index_in_nid(nid
);
4032 prev_end_pfn
= min(early_node_map
[i
].start_pfn
, range_end_pfn
);
4034 /* Account for ranges before physical memory on this node */
4035 if (early_node_map
[i
].start_pfn
> range_start_pfn
)
4036 hole_pages
= prev_end_pfn
- range_start_pfn
;
4038 /* Find all holes for the zone within the node */
4039 for (; i
!= -1; i
= next_active_region_index_in_nid(i
, nid
)) {
4041 /* No need to continue if prev_end_pfn is outside the zone */
4042 if (prev_end_pfn
>= range_end_pfn
)
4045 /* Make sure the end of the zone is not within the hole */
4046 start_pfn
= min(early_node_map
[i
].start_pfn
, range_end_pfn
);
4047 prev_end_pfn
= max(prev_end_pfn
, range_start_pfn
);
4049 /* Update the hole size cound and move on */
4050 if (start_pfn
> range_start_pfn
) {
4051 BUG_ON(prev_end_pfn
> start_pfn
);
4052 hole_pages
+= start_pfn
- prev_end_pfn
;
4054 prev_end_pfn
= early_node_map
[i
].end_pfn
;
4057 /* Account for ranges past physical memory on this node */
4058 if (range_end_pfn
> prev_end_pfn
)
4059 hole_pages
+= range_end_pfn
-
4060 max(range_start_pfn
, prev_end_pfn
);
4066 * absent_pages_in_range - Return number of page frames in holes within a range
4067 * @start_pfn: The start PFN to start searching for holes
4068 * @end_pfn: The end PFN to stop searching for holes
4070 * It returns the number of pages frames in memory holes within a range.
4072 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
4073 unsigned long end_pfn
)
4075 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
4078 /* Return the number of page frames in holes in a zone on a node */
4079 static unsigned long __meminit
zone_absent_pages_in_node(int nid
,
4080 unsigned long zone_type
,
4081 unsigned long *ignored
)
4083 unsigned long node_start_pfn
, node_end_pfn
;
4084 unsigned long zone_start_pfn
, zone_end_pfn
;
4086 get_pfn_range_for_nid(nid
, &node_start_pfn
, &node_end_pfn
);
4087 zone_start_pfn
= max(arch_zone_lowest_possible_pfn
[zone_type
],
4089 zone_end_pfn
= min(arch_zone_highest_possible_pfn
[zone_type
],
4092 adjust_zone_range_for_zone_movable(nid
, zone_type
,
4093 node_start_pfn
, node_end_pfn
,
4094 &zone_start_pfn
, &zone_end_pfn
);
4095 return __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
4099 static inline unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
4100 unsigned long zone_type
,
4101 unsigned long *zones_size
)
4103 return zones_size
[zone_type
];
4106 static inline unsigned long __meminit
zone_absent_pages_in_node(int nid
,
4107 unsigned long zone_type
,
4108 unsigned long *zholes_size
)
4113 return zholes_size
[zone_type
];
4118 static void __meminit
calculate_node_totalpages(struct pglist_data
*pgdat
,
4119 unsigned long *zones_size
, unsigned long *zholes_size
)
4121 unsigned long realtotalpages
, totalpages
= 0;
4124 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4125 totalpages
+= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
4127 pgdat
->node_spanned_pages
= totalpages
;
4129 realtotalpages
= totalpages
;
4130 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4132 zone_absent_pages_in_node(pgdat
->node_id
, i
,
4134 pgdat
->node_present_pages
= realtotalpages
;
4135 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
4139 #ifndef CONFIG_SPARSEMEM
4141 * Calculate the size of the zone->blockflags rounded to an unsigned long
4142 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4143 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4144 * round what is now in bits to nearest long in bits, then return it in
4147 static unsigned long __init
usemap_size(unsigned long zonesize
)
4149 unsigned long usemapsize
;
4151 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
4152 usemapsize
= usemapsize
>> pageblock_order
;
4153 usemapsize
*= NR_PAGEBLOCK_BITS
;
4154 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
4156 return usemapsize
/ 8;
4159 static void __init
setup_usemap(struct pglist_data
*pgdat
,
4160 struct zone
*zone
, unsigned long zonesize
)
4162 unsigned long usemapsize
= usemap_size(zonesize
);
4163 zone
->pageblock_flags
= NULL
;
4165 zone
->pageblock_flags
= alloc_bootmem_node_nopanic(pgdat
,
4169 static inline void setup_usemap(struct pglist_data
*pgdat
,
4170 struct zone
*zone
, unsigned long zonesize
) {}
4171 #endif /* CONFIG_SPARSEMEM */
4173 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4175 /* Return a sensible default order for the pageblock size. */
4176 static inline int pageblock_default_order(void)
4178 if (HPAGE_SHIFT
> PAGE_SHIFT
)
4179 return HUGETLB_PAGE_ORDER
;
4184 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4185 static inline void __init
set_pageblock_order(unsigned int order
)
4187 /* Check that pageblock_nr_pages has not already been setup */
4188 if (pageblock_order
)
4192 * Assume the largest contiguous order of interest is a huge page.
4193 * This value may be variable depending on boot parameters on IA64
4195 pageblock_order
= order
;
4197 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4200 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4201 * and pageblock_default_order() are unused as pageblock_order is set
4202 * at compile-time. See include/linux/pageblock-flags.h for the values of
4203 * pageblock_order based on the kernel config
4205 static inline int pageblock_default_order(unsigned int order
)
4209 #define set_pageblock_order(x) do {} while (0)
4211 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4214 * Set up the zone data structures:
4215 * - mark all pages reserved
4216 * - mark all memory queues empty
4217 * - clear the memory bitmaps
4219 static void __paginginit
free_area_init_core(struct pglist_data
*pgdat
,
4220 unsigned long *zones_size
, unsigned long *zholes_size
)
4223 int nid
= pgdat
->node_id
;
4224 unsigned long zone_start_pfn
= pgdat
->node_start_pfn
;
4227 pgdat_resize_init(pgdat
);
4228 pgdat
->nr_zones
= 0;
4229 init_waitqueue_head(&pgdat
->kswapd_wait
);
4230 pgdat
->kswapd_max_order
= 0;
4231 pgdat_page_cgroup_init(pgdat
);
4233 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
4234 struct zone
*zone
= pgdat
->node_zones
+ j
;
4235 unsigned long size
, realsize
, memmap_pages
;
4238 size
= zone_spanned_pages_in_node(nid
, j
, zones_size
);
4239 realsize
= size
- zone_absent_pages_in_node(nid
, j
,
4243 * Adjust realsize so that it accounts for how much memory
4244 * is used by this zone for memmap. This affects the watermark
4245 * and per-cpu initialisations
4248 PAGE_ALIGN(size
* sizeof(struct page
)) >> PAGE_SHIFT
;
4249 if (realsize
>= memmap_pages
) {
4250 realsize
-= memmap_pages
;
4253 " %s zone: %lu pages used for memmap\n",
4254 zone_names
[j
], memmap_pages
);
4257 " %s zone: %lu pages exceeds realsize %lu\n",
4258 zone_names
[j
], memmap_pages
, realsize
);
4260 /* Account for reserved pages */
4261 if (j
== 0 && realsize
> dma_reserve
) {
4262 realsize
-= dma_reserve
;
4263 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
4264 zone_names
[0], dma_reserve
);
4267 if (!is_highmem_idx(j
))
4268 nr_kernel_pages
+= realsize
;
4269 nr_all_pages
+= realsize
;
4271 zone
->spanned_pages
= size
;
4272 zone
->present_pages
= realsize
;
4275 zone
->min_unmapped_pages
= (realsize
*sysctl_min_unmapped_ratio
)
4277 zone
->min_slab_pages
= (realsize
* sysctl_min_slab_ratio
) / 100;
4279 zone
->name
= zone_names
[j
];
4280 spin_lock_init(&zone
->lock
);
4281 spin_lock_init(&zone
->lru_lock
);
4282 zone_seqlock_init(zone
);
4283 zone
->zone_pgdat
= pgdat
;
4285 zone_pcp_init(zone
);
4287 INIT_LIST_HEAD(&zone
->lru
[l
].list
);
4288 zone
->reclaim_stat
.nr_saved_scan
[l
] = 0;
4290 zone
->reclaim_stat
.recent_rotated
[0] = 0;
4291 zone
->reclaim_stat
.recent_rotated
[1] = 0;
4292 zone
->reclaim_stat
.recent_scanned
[0] = 0;
4293 zone
->reclaim_stat
.recent_scanned
[1] = 0;
4294 zap_zone_vm_stats(zone
);
4299 set_pageblock_order(pageblock_default_order());
4300 setup_usemap(pgdat
, zone
, size
);
4301 ret
= init_currently_empty_zone(zone
, zone_start_pfn
,
4302 size
, MEMMAP_EARLY
);
4304 memmap_init(size
, nid
, j
, zone_start_pfn
);
4305 zone_start_pfn
+= size
;
4309 static void __init_refok
alloc_node_mem_map(struct pglist_data
*pgdat
)
4311 /* Skip empty nodes */
4312 if (!pgdat
->node_spanned_pages
)
4315 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4316 /* ia64 gets its own node_mem_map, before this, without bootmem */
4317 if (!pgdat
->node_mem_map
) {
4318 unsigned long size
, start
, end
;
4322 * The zone's endpoints aren't required to be MAX_ORDER
4323 * aligned but the node_mem_map endpoints must be in order
4324 * for the buddy allocator to function correctly.
4326 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
4327 end
= pgdat
->node_start_pfn
+ pgdat
->node_spanned_pages
;
4328 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
4329 size
= (end
- start
) * sizeof(struct page
);
4330 map
= alloc_remap(pgdat
->node_id
, size
);
4332 map
= alloc_bootmem_node_nopanic(pgdat
, size
);
4333 pgdat
->node_mem_map
= map
+ (pgdat
->node_start_pfn
- start
);
4335 #ifndef CONFIG_NEED_MULTIPLE_NODES
4337 * With no DISCONTIG, the global mem_map is just set as node 0's
4339 if (pgdat
== NODE_DATA(0)) {
4340 mem_map
= NODE_DATA(0)->node_mem_map
;
4341 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
4342 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
4343 mem_map
-= (pgdat
->node_start_pfn
- ARCH_PFN_OFFSET
);
4344 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4347 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4350 void __paginginit
free_area_init_node(int nid
, unsigned long *zones_size
,
4351 unsigned long node_start_pfn
, unsigned long *zholes_size
)
4353 pg_data_t
*pgdat
= NODE_DATA(nid
);
4355 pgdat
->node_id
= nid
;
4356 pgdat
->node_start_pfn
= node_start_pfn
;
4357 calculate_node_totalpages(pgdat
, zones_size
, zholes_size
);
4359 alloc_node_mem_map(pgdat
);
4360 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4361 printk(KERN_DEBUG
"free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4362 nid
, (unsigned long)pgdat
,
4363 (unsigned long)pgdat
->node_mem_map
);
4366 free_area_init_core(pgdat
, zones_size
, zholes_size
);
4369 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
4371 #if MAX_NUMNODES > 1
4373 * Figure out the number of possible node ids.
4375 static void __init
setup_nr_node_ids(void)
4378 unsigned int highest
= 0;
4380 for_each_node_mask(node
, node_possible_map
)
4382 nr_node_ids
= highest
+ 1;
4385 static inline void setup_nr_node_ids(void)
4391 * add_active_range - Register a range of PFNs backed by physical memory
4392 * @nid: The node ID the range resides on
4393 * @start_pfn: The start PFN of the available physical memory
4394 * @end_pfn: The end PFN of the available physical memory
4396 * These ranges are stored in an early_node_map[] and later used by
4397 * free_area_init_nodes() to calculate zone sizes and holes. If the
4398 * range spans a memory hole, it is up to the architecture to ensure
4399 * the memory is not freed by the bootmem allocator. If possible
4400 * the range being registered will be merged with existing ranges.
4402 void __init
add_active_range(unsigned int nid
, unsigned long start_pfn
,
4403 unsigned long end_pfn
)
4407 mminit_dprintk(MMINIT_TRACE
, "memory_register",
4408 "Entering add_active_range(%d, %#lx, %#lx) "
4409 "%d entries of %d used\n",
4410 nid
, start_pfn
, end_pfn
,
4411 nr_nodemap_entries
, MAX_ACTIVE_REGIONS
);
4413 mminit_validate_memmodel_limits(&start_pfn
, &end_pfn
);
4415 /* Merge with existing active regions if possible */
4416 for (i
= 0; i
< nr_nodemap_entries
; i
++) {
4417 if (early_node_map
[i
].nid
!= nid
)
4420 /* Skip if an existing region covers this new one */
4421 if (start_pfn
>= early_node_map
[i
].start_pfn
&&
4422 end_pfn
<= early_node_map
[i
].end_pfn
)
4425 /* Merge forward if suitable */
4426 if (start_pfn
<= early_node_map
[i
].end_pfn
&&
4427 end_pfn
> early_node_map
[i
].end_pfn
) {
4428 early_node_map
[i
].end_pfn
= end_pfn
;
4432 /* Merge backward if suitable */
4433 if (start_pfn
< early_node_map
[i
].start_pfn
&&
4434 end_pfn
>= early_node_map
[i
].start_pfn
) {
4435 early_node_map
[i
].start_pfn
= start_pfn
;
4440 /* Check that early_node_map is large enough */
4441 if (i
>= MAX_ACTIVE_REGIONS
) {
4442 printk(KERN_CRIT
"More than %d memory regions, truncating\n",
4443 MAX_ACTIVE_REGIONS
);
4447 early_node_map
[i
].nid
= nid
;
4448 early_node_map
[i
].start_pfn
= start_pfn
;
4449 early_node_map
[i
].end_pfn
= end_pfn
;
4450 nr_nodemap_entries
= i
+ 1;
4454 * remove_active_range - Shrink an existing registered range of PFNs
4455 * @nid: The node id the range is on that should be shrunk
4456 * @start_pfn: The new PFN of the range
4457 * @end_pfn: The new PFN of the range
4459 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
4460 * The map is kept near the end physical page range that has already been
4461 * registered. This function allows an arch to shrink an existing registered
4464 void __init
remove_active_range(unsigned int nid
, unsigned long start_pfn
,
4465 unsigned long end_pfn
)
4470 printk(KERN_DEBUG
"remove_active_range (%d, %lu, %lu)\n",
4471 nid
, start_pfn
, end_pfn
);
4473 /* Find the old active region end and shrink */
4474 for_each_active_range_index_in_nid(i
, nid
) {
4475 if (early_node_map
[i
].start_pfn
>= start_pfn
&&
4476 early_node_map
[i
].end_pfn
<= end_pfn
) {
4478 early_node_map
[i
].start_pfn
= 0;
4479 early_node_map
[i
].end_pfn
= 0;
4483 if (early_node_map
[i
].start_pfn
< start_pfn
&&
4484 early_node_map
[i
].end_pfn
> start_pfn
) {
4485 unsigned long temp_end_pfn
= early_node_map
[i
].end_pfn
;
4486 early_node_map
[i
].end_pfn
= start_pfn
;
4487 if (temp_end_pfn
> end_pfn
)
4488 add_active_range(nid
, end_pfn
, temp_end_pfn
);
4491 if (early_node_map
[i
].start_pfn
>= start_pfn
&&
4492 early_node_map
[i
].end_pfn
> end_pfn
&&
4493 early_node_map
[i
].start_pfn
< end_pfn
) {
4494 early_node_map
[i
].start_pfn
= end_pfn
;
4502 /* remove the blank ones */
4503 for (i
= nr_nodemap_entries
- 1; i
> 0; i
--) {
4504 if (early_node_map
[i
].nid
!= nid
)
4506 if (early_node_map
[i
].end_pfn
)
4508 /* we found it, get rid of it */
4509 for (j
= i
; j
< nr_nodemap_entries
- 1; j
++)
4510 memcpy(&early_node_map
[j
], &early_node_map
[j
+1],
4511 sizeof(early_node_map
[j
]));
4512 j
= nr_nodemap_entries
- 1;
4513 memset(&early_node_map
[j
], 0, sizeof(early_node_map
[j
]));
4514 nr_nodemap_entries
--;
4519 * remove_all_active_ranges - Remove all currently registered regions
4521 * During discovery, it may be found that a table like SRAT is invalid
4522 * and an alternative discovery method must be used. This function removes
4523 * all currently registered regions.
4525 void __init
remove_all_active_ranges(void)
4527 memset(early_node_map
, 0, sizeof(early_node_map
));
4528 nr_nodemap_entries
= 0;
4531 /* Compare two active node_active_regions */
4532 static int __init
cmp_node_active_region(const void *a
, const void *b
)
4534 struct node_active_region
*arange
= (struct node_active_region
*)a
;
4535 struct node_active_region
*brange
= (struct node_active_region
*)b
;
4537 /* Done this way to avoid overflows */
4538 if (arange
->start_pfn
> brange
->start_pfn
)
4540 if (arange
->start_pfn
< brange
->start_pfn
)
4546 /* sort the node_map by start_pfn */
4547 void __init
sort_node_map(void)
4549 sort(early_node_map
, (size_t)nr_nodemap_entries
,
4550 sizeof(struct node_active_region
),
4551 cmp_node_active_region
, NULL
);
4554 /* Find the lowest pfn for a node */
4555 static unsigned long __init
find_min_pfn_for_node(int nid
)
4558 unsigned long min_pfn
= ULONG_MAX
;
4560 /* Assuming a sorted map, the first range found has the starting pfn */
4561 for_each_active_range_index_in_nid(i
, nid
)
4562 min_pfn
= min(min_pfn
, early_node_map
[i
].start_pfn
);
4564 if (min_pfn
== ULONG_MAX
) {
4566 "Could not find start_pfn for node %d\n", nid
);
4574 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4576 * It returns the minimum PFN based on information provided via
4577 * add_active_range().
4579 unsigned long __init
find_min_pfn_with_active_regions(void)
4581 return find_min_pfn_for_node(MAX_NUMNODES
);
4585 * early_calculate_totalpages()
4586 * Sum pages in active regions for movable zone.
4587 * Populate N_HIGH_MEMORY for calculating usable_nodes.
4589 static unsigned long __init
early_calculate_totalpages(void)
4592 unsigned long totalpages
= 0;
4594 for (i
= 0; i
< nr_nodemap_entries
; i
++) {
4595 unsigned long pages
= early_node_map
[i
].end_pfn
-
4596 early_node_map
[i
].start_pfn
;
4597 totalpages
+= pages
;
4599 node_set_state(early_node_map
[i
].nid
, N_HIGH_MEMORY
);
4605 * Find the PFN the Movable zone begins in each node. Kernel memory
4606 * is spread evenly between nodes as long as the nodes have enough
4607 * memory. When they don't, some nodes will have more kernelcore than
4610 static void __init
find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn
)
4613 unsigned long usable_startpfn
;
4614 unsigned long kernelcore_node
, kernelcore_remaining
;
4615 /* save the state before borrow the nodemask */
4616 nodemask_t saved_node_state
= node_states
[N_HIGH_MEMORY
];
4617 unsigned long totalpages
= early_calculate_totalpages();
4618 int usable_nodes
= nodes_weight(node_states
[N_HIGH_MEMORY
]);
4621 * If movablecore was specified, calculate what size of
4622 * kernelcore that corresponds so that memory usable for
4623 * any allocation type is evenly spread. If both kernelcore
4624 * and movablecore are specified, then the value of kernelcore
4625 * will be used for required_kernelcore if it's greater than
4626 * what movablecore would have allowed.
4628 if (required_movablecore
) {
4629 unsigned long corepages
;
4632 * Round-up so that ZONE_MOVABLE is at least as large as what
4633 * was requested by the user
4635 required_movablecore
=
4636 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
4637 corepages
= totalpages
- required_movablecore
;
4639 required_kernelcore
= max(required_kernelcore
, corepages
);
4642 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4643 if (!required_kernelcore
)
4646 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4647 find_usable_zone_for_movable();
4648 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
4651 /* Spread kernelcore memory as evenly as possible throughout nodes */
4652 kernelcore_node
= required_kernelcore
/ usable_nodes
;
4653 for_each_node_state(nid
, N_HIGH_MEMORY
) {
4655 * Recalculate kernelcore_node if the division per node
4656 * now exceeds what is necessary to satisfy the requested
4657 * amount of memory for the kernel
4659 if (required_kernelcore
< kernelcore_node
)
4660 kernelcore_node
= required_kernelcore
/ usable_nodes
;
4663 * As the map is walked, we track how much memory is usable
4664 * by the kernel using kernelcore_remaining. When it is
4665 * 0, the rest of the node is usable by ZONE_MOVABLE
4667 kernelcore_remaining
= kernelcore_node
;
4669 /* Go through each range of PFNs within this node */
4670 for_each_active_range_index_in_nid(i
, nid
) {
4671 unsigned long start_pfn
, end_pfn
;
4672 unsigned long size_pages
;
4674 start_pfn
= max(early_node_map
[i
].start_pfn
,
4675 zone_movable_pfn
[nid
]);
4676 end_pfn
= early_node_map
[i
].end_pfn
;
4677 if (start_pfn
>= end_pfn
)
4680 /* Account for what is only usable for kernelcore */
4681 if (start_pfn
< usable_startpfn
) {
4682 unsigned long kernel_pages
;
4683 kernel_pages
= min(end_pfn
, usable_startpfn
)
4686 kernelcore_remaining
-= min(kernel_pages
,
4687 kernelcore_remaining
);
4688 required_kernelcore
-= min(kernel_pages
,
4689 required_kernelcore
);
4691 /* Continue if range is now fully accounted */
4692 if (end_pfn
<= usable_startpfn
) {
4695 * Push zone_movable_pfn to the end so
4696 * that if we have to rebalance
4697 * kernelcore across nodes, we will
4698 * not double account here
4700 zone_movable_pfn
[nid
] = end_pfn
;
4703 start_pfn
= usable_startpfn
;
4707 * The usable PFN range for ZONE_MOVABLE is from
4708 * start_pfn->end_pfn. Calculate size_pages as the
4709 * number of pages used as kernelcore
4711 size_pages
= end_pfn
- start_pfn
;
4712 if (size_pages
> kernelcore_remaining
)
4713 size_pages
= kernelcore_remaining
;
4714 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
4717 * Some kernelcore has been met, update counts and
4718 * break if the kernelcore for this node has been
4721 required_kernelcore
-= min(required_kernelcore
,
4723 kernelcore_remaining
-= size_pages
;
4724 if (!kernelcore_remaining
)
4730 * If there is still required_kernelcore, we do another pass with one
4731 * less node in the count. This will push zone_movable_pfn[nid] further
4732 * along on the nodes that still have memory until kernelcore is
4736 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
4739 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4740 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
4741 zone_movable_pfn
[nid
] =
4742 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
4745 /* restore the node_state */
4746 node_states
[N_HIGH_MEMORY
] = saved_node_state
;
4749 /* Any regular memory on that node ? */
4750 static void check_for_regular_memory(pg_data_t
*pgdat
)
4752 #ifdef CONFIG_HIGHMEM
4753 enum zone_type zone_type
;
4755 for (zone_type
= 0; zone_type
<= ZONE_NORMAL
; zone_type
++) {
4756 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
4757 if (zone
->present_pages
)
4758 node_set_state(zone_to_nid(zone
), N_NORMAL_MEMORY
);
4764 * free_area_init_nodes - Initialise all pg_data_t and zone data
4765 * @max_zone_pfn: an array of max PFNs for each zone
4767 * This will call free_area_init_node() for each active node in the system.
4768 * Using the page ranges provided by add_active_range(), the size of each
4769 * zone in each node and their holes is calculated. If the maximum PFN
4770 * between two adjacent zones match, it is assumed that the zone is empty.
4771 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4772 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4773 * starts where the previous one ended. For example, ZONE_DMA32 starts
4774 * at arch_max_dma_pfn.
4776 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
4781 /* Sort early_node_map as initialisation assumes it is sorted */
4784 /* Record where the zone boundaries are */
4785 memset(arch_zone_lowest_possible_pfn
, 0,
4786 sizeof(arch_zone_lowest_possible_pfn
));
4787 memset(arch_zone_highest_possible_pfn
, 0,
4788 sizeof(arch_zone_highest_possible_pfn
));
4789 arch_zone_lowest_possible_pfn
[0] = find_min_pfn_with_active_regions();
4790 arch_zone_highest_possible_pfn
[0] = max_zone_pfn
[0];
4791 for (i
= 1; i
< MAX_NR_ZONES
; i
++) {
4792 if (i
== ZONE_MOVABLE
)
4794 arch_zone_lowest_possible_pfn
[i
] =
4795 arch_zone_highest_possible_pfn
[i
-1];
4796 arch_zone_highest_possible_pfn
[i
] =
4797 max(max_zone_pfn
[i
], arch_zone_lowest_possible_pfn
[i
]);
4799 arch_zone_lowest_possible_pfn
[ZONE_MOVABLE
] = 0;
4800 arch_zone_highest_possible_pfn
[ZONE_MOVABLE
] = 0;
4802 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4803 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
4804 find_zone_movable_pfns_for_nodes(zone_movable_pfn
);
4806 /* Print out the zone ranges */
4807 printk("Zone PFN ranges:\n");
4808 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
4809 if (i
== ZONE_MOVABLE
)
4811 printk(" %-8s ", zone_names
[i
]);
4812 if (arch_zone_lowest_possible_pfn
[i
] ==
4813 arch_zone_highest_possible_pfn
[i
])
4816 printk("%0#10lx -> %0#10lx\n",
4817 arch_zone_lowest_possible_pfn
[i
],
4818 arch_zone_highest_possible_pfn
[i
]);
4821 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4822 printk("Movable zone start PFN for each node\n");
4823 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
4824 if (zone_movable_pfn
[i
])
4825 printk(" Node %d: %lu\n", i
, zone_movable_pfn
[i
]);
4828 /* Print out the early_node_map[] */
4829 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries
);
4830 for (i
= 0; i
< nr_nodemap_entries
; i
++)
4831 printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map
[i
].nid
,
4832 early_node_map
[i
].start_pfn
,
4833 early_node_map
[i
].end_pfn
);
4835 /* Initialise every node */
4836 mminit_verify_pageflags_layout();
4837 setup_nr_node_ids();
4838 for_each_online_node(nid
) {
4839 pg_data_t
*pgdat
= NODE_DATA(nid
);
4840 free_area_init_node(nid
, NULL
,
4841 find_min_pfn_for_node(nid
), NULL
);
4843 /* Any memory on that node */
4844 if (pgdat
->node_present_pages
)
4845 node_set_state(nid
, N_HIGH_MEMORY
);
4846 check_for_regular_memory(pgdat
);
4850 static int __init
cmdline_parse_core(char *p
, unsigned long *core
)
4852 unsigned long long coremem
;
4856 coremem
= memparse(p
, &p
);
4857 *core
= coremem
>> PAGE_SHIFT
;
4859 /* Paranoid check that UL is enough for the coremem value */
4860 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
4866 * kernelcore=size sets the amount of memory for use for allocations that
4867 * cannot be reclaimed or migrated.
4869 static int __init
cmdline_parse_kernelcore(char *p
)
4871 return cmdline_parse_core(p
, &required_kernelcore
);
4875 * movablecore=size sets the amount of memory for use for allocations that
4876 * can be reclaimed or migrated.
4878 static int __init
cmdline_parse_movablecore(char *p
)
4880 return cmdline_parse_core(p
, &required_movablecore
);
4883 early_param("kernelcore", cmdline_parse_kernelcore
);
4884 early_param("movablecore", cmdline_parse_movablecore
);
4886 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4889 * set_dma_reserve - set the specified number of pages reserved in the first zone
4890 * @new_dma_reserve: The number of pages to mark reserved
4892 * The per-cpu batchsize and zone watermarks are determined by present_pages.
4893 * In the DMA zone, a significant percentage may be consumed by kernel image
4894 * and other unfreeable allocations which can skew the watermarks badly. This
4895 * function may optionally be used to account for unfreeable pages in the
4896 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
4897 * smaller per-cpu batchsize.
4899 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
4901 dma_reserve
= new_dma_reserve
;
4904 void __init
free_area_init(unsigned long *zones_size
)
4906 free_area_init_node(0, zones_size
,
4907 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
4910 static int page_alloc_cpu_notify(struct notifier_block
*self
,
4911 unsigned long action
, void *hcpu
)
4913 int cpu
= (unsigned long)hcpu
;
4915 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
) {
4919 * Spill the event counters of the dead processor
4920 * into the current processors event counters.
4921 * This artificially elevates the count of the current
4924 vm_events_fold_cpu(cpu
);
4927 * Zero the differential counters of the dead processor
4928 * so that the vm statistics are consistent.
4930 * This is only okay since the processor is dead and cannot
4931 * race with what we are doing.
4933 refresh_cpu_vm_stats(cpu
);
4938 void __init
page_alloc_init(void)
4940 hotcpu_notifier(page_alloc_cpu_notify
, 0);
4944 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
4945 * or min_free_kbytes changes.
4947 static void calculate_totalreserve_pages(void)
4949 struct pglist_data
*pgdat
;
4950 unsigned long reserve_pages
= 0;
4951 enum zone_type i
, j
;
4953 for_each_online_pgdat(pgdat
) {
4954 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
4955 struct zone
*zone
= pgdat
->node_zones
+ i
;
4956 unsigned long max
= 0;
4958 /* Find valid and maximum lowmem_reserve in the zone */
4959 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
4960 if (zone
->lowmem_reserve
[j
] > max
)
4961 max
= zone
->lowmem_reserve
[j
];
4964 /* we treat the high watermark as reserved pages. */
4965 max
+= high_wmark_pages(zone
);
4967 if (max
> zone
->present_pages
)
4968 max
= zone
->present_pages
;
4969 reserve_pages
+= max
;
4972 totalreserve_pages
= reserve_pages
;
4976 * setup_per_zone_lowmem_reserve - called whenever
4977 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
4978 * has a correct pages reserved value, so an adequate number of
4979 * pages are left in the zone after a successful __alloc_pages().
4981 static void setup_per_zone_lowmem_reserve(void)
4983 struct pglist_data
*pgdat
;
4984 enum zone_type j
, idx
;
4986 for_each_online_pgdat(pgdat
) {
4987 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
4988 struct zone
*zone
= pgdat
->node_zones
+ j
;
4989 unsigned long present_pages
= zone
->present_pages
;
4991 zone
->lowmem_reserve
[j
] = 0;
4995 struct zone
*lower_zone
;
4999 if (sysctl_lowmem_reserve_ratio
[idx
] < 1)
5000 sysctl_lowmem_reserve_ratio
[idx
] = 1;
5002 lower_zone
= pgdat
->node_zones
+ idx
;
5003 lower_zone
->lowmem_reserve
[j
] = present_pages
/
5004 sysctl_lowmem_reserve_ratio
[idx
];
5005 present_pages
+= lower_zone
->present_pages
;
5010 /* update totalreserve_pages */
5011 calculate_totalreserve_pages();
5015 * setup_per_zone_wmarks - called when min_free_kbytes changes
5016 * or when memory is hot-{added|removed}
5018 * Ensures that the watermark[min,low,high] values for each zone are set
5019 * correctly with respect to min_free_kbytes.
5021 void setup_per_zone_wmarks(void)
5023 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
5024 unsigned long lowmem_pages
= 0;
5026 unsigned long flags
;
5028 /* Calculate total number of !ZONE_HIGHMEM pages */
5029 for_each_zone(zone
) {
5030 if (!is_highmem(zone
))
5031 lowmem_pages
+= zone
->present_pages
;
5034 for_each_zone(zone
) {
5037 spin_lock_irqsave(&zone
->lock
, flags
);
5038 tmp
= (u64
)pages_min
* zone
->present_pages
;
5039 do_div(tmp
, lowmem_pages
);
5040 if (is_highmem(zone
)) {
5042 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5043 * need highmem pages, so cap pages_min to a small
5046 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5047 * deltas controls asynch page reclaim, and so should
5048 * not be capped for highmem.
5052 min_pages
= zone
->present_pages
/ 1024;
5053 if (min_pages
< SWAP_CLUSTER_MAX
)
5054 min_pages
= SWAP_CLUSTER_MAX
;
5055 if (min_pages
> 128)
5057 zone
->watermark
[WMARK_MIN
] = min_pages
;
5060 * If it's a lowmem zone, reserve a number of pages
5061 * proportionate to the zone's size.
5063 zone
->watermark
[WMARK_MIN
] = tmp
;
5066 zone
->watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + (tmp
>> 2);
5067 zone
->watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + (tmp
>> 1);
5068 setup_zone_migrate_reserve(zone
);
5069 spin_unlock_irqrestore(&zone
->lock
, flags
);
5072 /* update totalreserve_pages */
5073 calculate_totalreserve_pages();
5077 * The inactive anon list should be small enough that the VM never has to
5078 * do too much work, but large enough that each inactive page has a chance
5079 * to be referenced again before it is swapped out.
5081 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5082 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5083 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5084 * the anonymous pages are kept on the inactive list.
5087 * memory ratio inactive anon
5088 * -------------------------------------
5097 void __meminit
calculate_zone_inactive_ratio(struct zone
*zone
)
5099 unsigned int gb
, ratio
;
5101 /* Zone size in gigabytes */
5102 gb
= zone
->present_pages
>> (30 - PAGE_SHIFT
);
5104 ratio
= int_sqrt(10 * gb
);
5108 zone
->inactive_ratio
= ratio
;
5111 static void __meminit
setup_per_zone_inactive_ratio(void)
5116 calculate_zone_inactive_ratio(zone
);
5120 * Initialise min_free_kbytes.
5122 * For small machines we want it small (128k min). For large machines
5123 * we want it large (64MB max). But it is not linear, because network
5124 * bandwidth does not increase linearly with machine size. We use
5126 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5127 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5143 static int __init
init_per_zone_wmark_min(void)
5145 unsigned long lowmem_kbytes
;
5147 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
5149 min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
5150 if (min_free_kbytes
< 128)
5151 min_free_kbytes
= 128;
5152 if (min_free_kbytes
> 65536)
5153 min_free_kbytes
= 65536;
5154 setup_per_zone_wmarks();
5155 setup_per_zone_lowmem_reserve();
5156 setup_per_zone_inactive_ratio();
5159 module_init(init_per_zone_wmark_min
)
5162 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5163 * that we can call two helper functions whenever min_free_kbytes
5166 int min_free_kbytes_sysctl_handler(ctl_table
*table
, int write
,
5167 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5169 proc_dointvec(table
, write
, buffer
, length
, ppos
);
5171 setup_per_zone_wmarks();
5176 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table
*table
, int write
,
5177 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5182 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5187 zone
->min_unmapped_pages
= (zone
->present_pages
*
5188 sysctl_min_unmapped_ratio
) / 100;
5192 int sysctl_min_slab_ratio_sysctl_handler(ctl_table
*table
, int write
,
5193 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5198 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5203 zone
->min_slab_pages
= (zone
->present_pages
*
5204 sysctl_min_slab_ratio
) / 100;
5210 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5211 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5212 * whenever sysctl_lowmem_reserve_ratio changes.
5214 * The reserve ratio obviously has absolutely no relation with the
5215 * minimum watermarks. The lowmem reserve ratio can only make sense
5216 * if in function of the boot time zone sizes.
5218 int lowmem_reserve_ratio_sysctl_handler(ctl_table
*table
, int write
,
5219 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5221 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5222 setup_per_zone_lowmem_reserve();
5227 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5228 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
5229 * can have before it gets flushed back to buddy allocator.
5232 int percpu_pagelist_fraction_sysctl_handler(ctl_table
*table
, int write
,
5233 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5239 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5240 if (!write
|| (ret
== -EINVAL
))
5242 for_each_populated_zone(zone
) {
5243 for_each_possible_cpu(cpu
) {
5245 high
= zone
->present_pages
/ percpu_pagelist_fraction
;
5246 setup_pagelist_highmark(
5247 per_cpu_ptr(zone
->pageset
, cpu
), high
);
5253 int hashdist
= HASHDIST_DEFAULT
;
5256 static int __init
set_hashdist(char *str
)
5260 hashdist
= simple_strtoul(str
, &str
, 0);
5263 __setup("hashdist=", set_hashdist
);
5267 * allocate a large system hash table from bootmem
5268 * - it is assumed that the hash table must contain an exact power-of-2
5269 * quantity of entries
5270 * - limit is the number of hash buckets, not the total allocation size
5272 void *__init
alloc_large_system_hash(const char *tablename
,
5273 unsigned long bucketsize
,
5274 unsigned long numentries
,
5277 unsigned int *_hash_shift
,
5278 unsigned int *_hash_mask
,
5279 unsigned long limit
)
5281 unsigned long long max
= limit
;
5282 unsigned long log2qty
, size
;
5285 /* allow the kernel cmdline to have a say */
5287 /* round applicable memory size up to nearest megabyte */
5288 numentries
= nr_kernel_pages
;
5289 numentries
+= (1UL << (20 - PAGE_SHIFT
)) - 1;
5290 numentries
>>= 20 - PAGE_SHIFT
;
5291 numentries
<<= 20 - PAGE_SHIFT
;
5293 /* limit to 1 bucket per 2^scale bytes of low memory */
5294 if (scale
> PAGE_SHIFT
)
5295 numentries
>>= (scale
- PAGE_SHIFT
);
5297 numentries
<<= (PAGE_SHIFT
- scale
);
5299 /* Make sure we've got at least a 0-order allocation.. */
5300 if (unlikely(flags
& HASH_SMALL
)) {
5301 /* Makes no sense without HASH_EARLY */
5302 WARN_ON(!(flags
& HASH_EARLY
));
5303 if (!(numentries
>> *_hash_shift
)) {
5304 numentries
= 1UL << *_hash_shift
;
5305 BUG_ON(!numentries
);
5307 } else if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
5308 numentries
= PAGE_SIZE
/ bucketsize
;
5310 numentries
= roundup_pow_of_two(numentries
);
5312 /* limit allocation size to 1/16 total memory by default */
5314 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
5315 do_div(max
, bucketsize
);
5318 if (numentries
> max
)
5321 log2qty
= ilog2(numentries
);
5324 size
= bucketsize
<< log2qty
;
5325 if (flags
& HASH_EARLY
)
5326 table
= alloc_bootmem_nopanic(size
);
5328 table
= __vmalloc(size
, GFP_ATOMIC
, PAGE_KERNEL
);
5331 * If bucketsize is not a power-of-two, we may free
5332 * some pages at the end of hash table which
5333 * alloc_pages_exact() automatically does
5335 if (get_order(size
) < MAX_ORDER
) {
5336 table
= alloc_pages_exact(size
, GFP_ATOMIC
);
5337 kmemleak_alloc(table
, size
, 1, GFP_ATOMIC
);
5340 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
5343 panic("Failed to allocate %s hash table\n", tablename
);
5345 printk(KERN_INFO
"%s hash table entries: %ld (order: %d, %lu bytes)\n",
5348 ilog2(size
) - PAGE_SHIFT
,
5352 *_hash_shift
= log2qty
;
5354 *_hash_mask
= (1 << log2qty
) - 1;
5359 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5360 static inline unsigned long *get_pageblock_bitmap(struct zone
*zone
,
5363 #ifdef CONFIG_SPARSEMEM
5364 return __pfn_to_section(pfn
)->pageblock_flags
;
5366 return zone
->pageblock_flags
;
5367 #endif /* CONFIG_SPARSEMEM */
5370 static inline int pfn_to_bitidx(struct zone
*zone
, unsigned long pfn
)
5372 #ifdef CONFIG_SPARSEMEM
5373 pfn
&= (PAGES_PER_SECTION
-1);
5374 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
5376 pfn
= pfn
- zone
->zone_start_pfn
;
5377 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
5378 #endif /* CONFIG_SPARSEMEM */
5382 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5383 * @page: The page within the block of interest
5384 * @start_bitidx: The first bit of interest to retrieve
5385 * @end_bitidx: The last bit of interest
5386 * returns pageblock_bits flags
5388 unsigned long get_pageblock_flags_group(struct page
*page
,
5389 int start_bitidx
, int end_bitidx
)
5392 unsigned long *bitmap
;
5393 unsigned long pfn
, bitidx
;
5394 unsigned long flags
= 0;
5395 unsigned long value
= 1;
5397 zone
= page_zone(page
);
5398 pfn
= page_to_pfn(page
);
5399 bitmap
= get_pageblock_bitmap(zone
, pfn
);
5400 bitidx
= pfn_to_bitidx(zone
, pfn
);
5402 for (; start_bitidx
<= end_bitidx
; start_bitidx
++, value
<<= 1)
5403 if (test_bit(bitidx
+ start_bitidx
, bitmap
))
5410 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5411 * @page: The page within the block of interest
5412 * @start_bitidx: The first bit of interest
5413 * @end_bitidx: The last bit of interest
5414 * @flags: The flags to set
5416 void set_pageblock_flags_group(struct page
*page
, unsigned long flags
,
5417 int start_bitidx
, int end_bitidx
)
5420 unsigned long *bitmap
;
5421 unsigned long pfn
, bitidx
;
5422 unsigned long value
= 1;
5424 zone
= page_zone(page
);
5425 pfn
= page_to_pfn(page
);
5426 bitmap
= get_pageblock_bitmap(zone
, pfn
);
5427 bitidx
= pfn_to_bitidx(zone
, pfn
);
5428 VM_BUG_ON(pfn
< zone
->zone_start_pfn
);
5429 VM_BUG_ON(pfn
>= zone
->zone_start_pfn
+ zone
->spanned_pages
);
5431 for (; start_bitidx
<= end_bitidx
; start_bitidx
++, value
<<= 1)
5433 __set_bit(bitidx
+ start_bitidx
, bitmap
);
5435 __clear_bit(bitidx
+ start_bitidx
, bitmap
);
5439 * This is designed as sub function...plz see page_isolation.c also.
5440 * set/clear page block's type to be ISOLATE.
5441 * page allocater never alloc memory from ISOLATE block.
5445 __count_immobile_pages(struct zone
*zone
, struct page
*page
, int count
)
5447 unsigned long pfn
, iter
, found
;
5449 * For avoiding noise data, lru_add_drain_all() should be called
5450 * If ZONE_MOVABLE, the zone never contains immobile pages
5452 if (zone_idx(zone
) == ZONE_MOVABLE
)
5455 if (get_pageblock_migratetype(page
) == MIGRATE_MOVABLE
)
5458 pfn
= page_to_pfn(page
);
5459 for (found
= 0, iter
= 0; iter
< pageblock_nr_pages
; iter
++) {
5460 unsigned long check
= pfn
+ iter
;
5462 if (!pfn_valid_within(check
))
5465 page
= pfn_to_page(check
);
5466 if (!page_count(page
)) {
5467 if (PageBuddy(page
))
5468 iter
+= (1 << page_order(page
)) - 1;
5474 * If there are RECLAIMABLE pages, we need to check it.
5475 * But now, memory offline itself doesn't call shrink_slab()
5476 * and it still to be fixed.
5479 * If the page is not RAM, page_count()should be 0.
5480 * we don't need more check. This is an _used_ not-movable page.
5482 * The problematic thing here is PG_reserved pages. PG_reserved
5483 * is set to both of a memory hole page and a _used_ kernel
5492 bool is_pageblock_removable_nolock(struct page
*page
)
5494 struct zone
*zone
= page_zone(page
);
5495 return __count_immobile_pages(zone
, page
, 0);
5498 int set_migratetype_isolate(struct page
*page
)
5501 unsigned long flags
, pfn
;
5502 struct memory_isolate_notify arg
;
5506 zone
= page_zone(page
);
5508 spin_lock_irqsave(&zone
->lock
, flags
);
5510 pfn
= page_to_pfn(page
);
5511 arg
.start_pfn
= pfn
;
5512 arg
.nr_pages
= pageblock_nr_pages
;
5513 arg
.pages_found
= 0;
5516 * It may be possible to isolate a pageblock even if the
5517 * migratetype is not MIGRATE_MOVABLE. The memory isolation
5518 * notifier chain is used by balloon drivers to return the
5519 * number of pages in a range that are held by the balloon
5520 * driver to shrink memory. If all the pages are accounted for
5521 * by balloons, are free, or on the LRU, isolation can continue.
5522 * Later, for example, when memory hotplug notifier runs, these
5523 * pages reported as "can be isolated" should be isolated(freed)
5524 * by the balloon driver through the memory notifier chain.
5526 notifier_ret
= memory_isolate_notify(MEM_ISOLATE_COUNT
, &arg
);
5527 notifier_ret
= notifier_to_errno(notifier_ret
);
5531 * FIXME: Now, memory hotplug doesn't call shrink_slab() by itself.
5532 * We just check MOVABLE pages.
5534 if (__count_immobile_pages(zone
, page
, arg
.pages_found
))
5538 * immobile means "not-on-lru" paes. If immobile is larger than
5539 * removable-by-driver pages reported by notifier, we'll fail.
5544 set_pageblock_migratetype(page
, MIGRATE_ISOLATE
);
5545 move_freepages_block(zone
, page
, MIGRATE_ISOLATE
);
5548 spin_unlock_irqrestore(&zone
->lock
, flags
);
5554 void unset_migratetype_isolate(struct page
*page
)
5557 unsigned long flags
;
5558 zone
= page_zone(page
);
5559 spin_lock_irqsave(&zone
->lock
, flags
);
5560 if (get_pageblock_migratetype(page
) != MIGRATE_ISOLATE
)
5562 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
5563 move_freepages_block(zone
, page
, MIGRATE_MOVABLE
);
5565 spin_unlock_irqrestore(&zone
->lock
, flags
);
5568 #ifdef CONFIG_MEMORY_HOTREMOVE
5570 * All pages in the range must be isolated before calling this.
5573 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
5579 unsigned long flags
;
5580 /* find the first valid pfn */
5581 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
5586 zone
= page_zone(pfn_to_page(pfn
));
5587 spin_lock_irqsave(&zone
->lock
, flags
);
5589 while (pfn
< end_pfn
) {
5590 if (!pfn_valid(pfn
)) {
5594 page
= pfn_to_page(pfn
);
5595 BUG_ON(page_count(page
));
5596 BUG_ON(!PageBuddy(page
));
5597 order
= page_order(page
);
5598 #ifdef CONFIG_DEBUG_VM
5599 printk(KERN_INFO
"remove from free list %lx %d %lx\n",
5600 pfn
, 1 << order
, end_pfn
);
5602 list_del(&page
->lru
);
5603 rmv_page_order(page
);
5604 zone
->free_area
[order
].nr_free
--;
5605 __mod_zone_page_state(zone
, NR_FREE_PAGES
,
5607 for (i
= 0; i
< (1 << order
); i
++)
5608 SetPageReserved((page
+i
));
5609 pfn
+= (1 << order
);
5611 spin_unlock_irqrestore(&zone
->lock
, flags
);
5615 #ifdef CONFIG_MEMORY_FAILURE
5616 bool is_free_buddy_page(struct page
*page
)
5618 struct zone
*zone
= page_zone(page
);
5619 unsigned long pfn
= page_to_pfn(page
);
5620 unsigned long flags
;
5623 spin_lock_irqsave(&zone
->lock
, flags
);
5624 for (order
= 0; order
< MAX_ORDER
; order
++) {
5625 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
5627 if (PageBuddy(page_head
) && page_order(page_head
) >= order
)
5630 spin_unlock_irqrestore(&zone
->lock
, flags
);
5632 return order
< MAX_ORDER
;
5636 static struct trace_print_flags pageflag_names
[] = {
5637 {1UL << PG_locked
, "locked" },
5638 {1UL << PG_error
, "error" },
5639 {1UL << PG_referenced
, "referenced" },
5640 {1UL << PG_uptodate
, "uptodate" },
5641 {1UL << PG_dirty
, "dirty" },
5642 {1UL << PG_lru
, "lru" },
5643 {1UL << PG_active
, "active" },
5644 {1UL << PG_slab
, "slab" },
5645 {1UL << PG_owner_priv_1
, "owner_priv_1" },
5646 {1UL << PG_arch_1
, "arch_1" },
5647 {1UL << PG_reserved
, "reserved" },
5648 {1UL << PG_private
, "private" },
5649 {1UL << PG_private_2
, "private_2" },
5650 {1UL << PG_writeback
, "writeback" },
5651 #ifdef CONFIG_PAGEFLAGS_EXTENDED
5652 {1UL << PG_head
, "head" },
5653 {1UL << PG_tail
, "tail" },
5655 {1UL << PG_compound
, "compound" },
5657 {1UL << PG_swapcache
, "swapcache" },
5658 {1UL << PG_mappedtodisk
, "mappedtodisk" },
5659 {1UL << PG_reclaim
, "reclaim" },
5660 {1UL << PG_swapbacked
, "swapbacked" },
5661 {1UL << PG_unevictable
, "unevictable" },
5663 {1UL << PG_mlocked
, "mlocked" },
5665 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
5666 {1UL << PG_uncached
, "uncached" },
5668 #ifdef CONFIG_MEMORY_FAILURE
5669 {1UL << PG_hwpoison
, "hwpoison" },
5674 static void dump_page_flags(unsigned long flags
)
5676 const char *delim
= "";
5680 printk(KERN_ALERT
"page flags: %#lx(", flags
);
5682 /* remove zone id */
5683 flags
&= (1UL << NR_PAGEFLAGS
) - 1;
5685 for (i
= 0; pageflag_names
[i
].name
&& flags
; i
++) {
5687 mask
= pageflag_names
[i
].mask
;
5688 if ((flags
& mask
) != mask
)
5692 printk("%s%s", delim
, pageflag_names
[i
].name
);
5696 /* check for left over flags */
5698 printk("%s%#lx", delim
, flags
);
5703 void dump_page(struct page
*page
)
5706 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
5707 page
, atomic_read(&page
->_count
), page_mapcount(page
),
5708 page
->mapping
, page
->index
);
5709 dump_page_flags(page
->flags
);
5710 mem_cgroup_print_bad_page(page
);