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>
57 #include <asm/tlbflush.h>
58 #include <asm/div64.h>
61 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
62 DEFINE_PER_CPU(int, numa_node
);
63 EXPORT_PER_CPU_SYMBOL(numa_node
);
66 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
68 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
69 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
70 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
71 * defined in <linux/topology.h>.
73 DEFINE_PER_CPU(int, _numa_mem_
); /* Kernel "local memory" node */
74 EXPORT_PER_CPU_SYMBOL(_numa_mem_
);
78 * Array of node states.
80 nodemask_t node_states
[NR_NODE_STATES
] __read_mostly
= {
81 [N_POSSIBLE
] = NODE_MASK_ALL
,
82 [N_ONLINE
] = { { [0] = 1UL } },
84 [N_NORMAL_MEMORY
] = { { [0] = 1UL } },
86 [N_HIGH_MEMORY
] = { { [0] = 1UL } },
88 [N_CPU
] = { { [0] = 1UL } },
91 EXPORT_SYMBOL(node_states
);
93 unsigned long totalram_pages __read_mostly
;
94 unsigned long totalreserve_pages __read_mostly
;
95 int percpu_pagelist_fraction
;
96 gfp_t gfp_allowed_mask __read_mostly
= GFP_BOOT_MASK
;
98 #ifdef CONFIG_PM_SLEEP
100 * The following functions are used by the suspend/hibernate code to temporarily
101 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
102 * while devices are suspended. To avoid races with the suspend/hibernate code,
103 * they should always be called with pm_mutex held (gfp_allowed_mask also should
104 * only be modified with pm_mutex held, unless the suspend/hibernate code is
105 * guaranteed not to run in parallel with that modification).
108 static gfp_t saved_gfp_mask
;
110 void pm_restore_gfp_mask(void)
112 WARN_ON(!mutex_is_locked(&pm_mutex
));
113 if (saved_gfp_mask
) {
114 gfp_allowed_mask
= saved_gfp_mask
;
119 void pm_restrict_gfp_mask(void)
121 WARN_ON(!mutex_is_locked(&pm_mutex
));
122 WARN_ON(saved_gfp_mask
);
123 saved_gfp_mask
= gfp_allowed_mask
;
124 gfp_allowed_mask
&= ~GFP_IOFS
;
126 #endif /* CONFIG_PM_SLEEP */
128 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
129 int pageblock_order __read_mostly
;
132 static void __free_pages_ok(struct page
*page
, unsigned int order
);
135 * results with 256, 32 in the lowmem_reserve sysctl:
136 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
137 * 1G machine -> (16M dma, 784M normal, 224M high)
138 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
139 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
140 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
142 * TBD: should special case ZONE_DMA32 machines here - in those we normally
143 * don't need any ZONE_NORMAL reservation
145 int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
-1] = {
146 #ifdef CONFIG_ZONE_DMA
149 #ifdef CONFIG_ZONE_DMA32
152 #ifdef CONFIG_HIGHMEM
158 EXPORT_SYMBOL(totalram_pages
);
160 static char * const zone_names
[MAX_NR_ZONES
] = {
161 #ifdef CONFIG_ZONE_DMA
164 #ifdef CONFIG_ZONE_DMA32
168 #ifdef CONFIG_HIGHMEM
174 int min_free_kbytes
= 1024;
176 static unsigned long __meminitdata nr_kernel_pages
;
177 static unsigned long __meminitdata nr_all_pages
;
178 static unsigned long __meminitdata dma_reserve
;
180 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
182 * MAX_ACTIVE_REGIONS determines the maximum number of distinct
183 * ranges of memory (RAM) that may be registered with add_active_range().
184 * Ranges passed to add_active_range() will be merged if possible
185 * so the number of times add_active_range() can be called is
186 * related to the number of nodes and the number of holes
188 #ifdef CONFIG_MAX_ACTIVE_REGIONS
189 /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */
190 #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS
192 #if MAX_NUMNODES >= 32
193 /* If there can be many nodes, allow up to 50 holes per node */
194 #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50)
196 /* By default, allow up to 256 distinct regions */
197 #define MAX_ACTIVE_REGIONS 256
201 static struct node_active_region __meminitdata early_node_map
[MAX_ACTIVE_REGIONS
];
202 static int __meminitdata nr_nodemap_entries
;
203 static unsigned long __meminitdata arch_zone_lowest_possible_pfn
[MAX_NR_ZONES
];
204 static unsigned long __meminitdata arch_zone_highest_possible_pfn
[MAX_NR_ZONES
];
205 static unsigned long __initdata required_kernelcore
;
206 static unsigned long __initdata required_movablecore
;
207 static unsigned long __meminitdata zone_movable_pfn
[MAX_NUMNODES
];
209 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
211 EXPORT_SYMBOL(movable_zone
);
212 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
215 int nr_node_ids __read_mostly
= MAX_NUMNODES
;
216 int nr_online_nodes __read_mostly
= 1;
217 EXPORT_SYMBOL(nr_node_ids
);
218 EXPORT_SYMBOL(nr_online_nodes
);
221 int page_group_by_mobility_disabled __read_mostly
;
223 static void set_pageblock_migratetype(struct page
*page
, int migratetype
)
226 if (unlikely(page_group_by_mobility_disabled
))
227 migratetype
= MIGRATE_UNMOVABLE
;
229 set_pageblock_flags_group(page
, (unsigned long)migratetype
,
230 PB_migrate
, PB_migrate_end
);
233 bool oom_killer_disabled __read_mostly
;
235 #ifdef CONFIG_DEBUG_VM
236 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
240 unsigned long pfn
= page_to_pfn(page
);
243 seq
= zone_span_seqbegin(zone
);
244 if (pfn
>= zone
->zone_start_pfn
+ zone
->spanned_pages
)
246 else if (pfn
< zone
->zone_start_pfn
)
248 } while (zone_span_seqretry(zone
, seq
));
253 static int page_is_consistent(struct zone
*zone
, struct page
*page
)
255 if (!pfn_valid_within(page_to_pfn(page
)))
257 if (zone
!= page_zone(page
))
263 * Temporary debugging check for pages not lying within a given zone.
265 static int bad_range(struct zone
*zone
, struct page
*page
)
267 if (page_outside_zone_boundaries(zone
, page
))
269 if (!page_is_consistent(zone
, page
))
275 static inline int bad_range(struct zone
*zone
, struct page
*page
)
281 static void bad_page(struct page
*page
)
283 static unsigned long resume
;
284 static unsigned long nr_shown
;
285 static unsigned long nr_unshown
;
287 /* Don't complain about poisoned pages */
288 if (PageHWPoison(page
)) {
289 __ClearPageBuddy(page
);
294 * Allow a burst of 60 reports, then keep quiet for that minute;
295 * or allow a steady drip of one report per second.
297 if (nr_shown
== 60) {
298 if (time_before(jiffies
, resume
)) {
304 "BUG: Bad page state: %lu messages suppressed\n",
311 resume
= jiffies
+ 60 * HZ
;
313 printk(KERN_ALERT
"BUG: Bad page state in process %s pfn:%05lx\n",
314 current
->comm
, page_to_pfn(page
));
319 /* Leave bad fields for debug, except PageBuddy could make trouble */
320 __ClearPageBuddy(page
);
321 add_taint(TAINT_BAD_PAGE
);
325 * Higher-order pages are called "compound pages". They are structured thusly:
327 * The first PAGE_SIZE page is called the "head page".
329 * The remaining PAGE_SIZE pages are called "tail pages".
331 * All pages have PG_compound set. All pages have their ->private pointing at
332 * the head page (even the head page has this).
334 * The first tail page's ->lru.next holds the address of the compound page's
335 * put_page() function. Its ->lru.prev holds the order of allocation.
336 * This usage means that zero-order pages may not be compound.
339 static void free_compound_page(struct page
*page
)
341 __free_pages_ok(page
, compound_order(page
));
344 void prep_compound_page(struct page
*page
, unsigned long order
)
347 int nr_pages
= 1 << order
;
349 set_compound_page_dtor(page
, free_compound_page
);
350 set_compound_order(page
, order
);
352 for (i
= 1; i
< nr_pages
; i
++) {
353 struct page
*p
= page
+ i
;
356 p
->first_page
= page
;
360 /* update __split_huge_page_refcount if you change this function */
361 static int destroy_compound_page(struct page
*page
, unsigned long order
)
364 int nr_pages
= 1 << order
;
367 if (unlikely(compound_order(page
) != order
) ||
368 unlikely(!PageHead(page
))) {
373 __ClearPageHead(page
);
375 for (i
= 1; i
< nr_pages
; i
++) {
376 struct page
*p
= page
+ i
;
378 if (unlikely(!PageTail(p
) || (p
->first_page
!= page
))) {
388 static inline void prep_zero_page(struct page
*page
, int order
, gfp_t gfp_flags
)
393 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
394 * and __GFP_HIGHMEM from hard or soft interrupt context.
396 VM_BUG_ON((gfp_flags
& __GFP_HIGHMEM
) && in_interrupt());
397 for (i
= 0; i
< (1 << order
); i
++)
398 clear_highpage(page
+ i
);
401 static inline void set_page_order(struct page
*page
, int order
)
403 set_page_private(page
, order
);
404 __SetPageBuddy(page
);
407 static inline void rmv_page_order(struct page
*page
)
409 __ClearPageBuddy(page
);
410 set_page_private(page
, 0);
414 * Locate the struct page for both the matching buddy in our
415 * pair (buddy1) and the combined O(n+1) page they form (page).
417 * 1) Any buddy B1 will have an order O twin B2 which satisfies
418 * the following equation:
420 * For example, if the starting buddy (buddy2) is #8 its order
422 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
424 * 2) Any buddy B will have an order O+1 parent P which
425 * satisfies the following equation:
428 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
430 static inline unsigned long
431 __find_buddy_index(unsigned long page_idx
, unsigned int order
)
433 return page_idx
^ (1 << order
);
437 * This function checks whether a page is free && is the buddy
438 * we can do coalesce a page and its buddy if
439 * (a) the buddy is not in a hole &&
440 * (b) the buddy is in the buddy system &&
441 * (c) a page and its buddy have the same order &&
442 * (d) a page and its buddy are in the same zone.
444 * For recording whether a page is in the buddy system, we set ->_mapcount -2.
445 * Setting, clearing, and testing _mapcount -2 is serialized by zone->lock.
447 * For recording page's order, we use page_private(page).
449 static inline int page_is_buddy(struct page
*page
, struct page
*buddy
,
452 if (!pfn_valid_within(page_to_pfn(buddy
)))
455 if (page_zone_id(page
) != page_zone_id(buddy
))
458 if (PageBuddy(buddy
) && page_order(buddy
) == order
) {
459 VM_BUG_ON(page_count(buddy
) != 0);
466 * Freeing function for a buddy system allocator.
468 * The concept of a buddy system is to maintain direct-mapped table
469 * (containing bit values) for memory blocks of various "orders".
470 * The bottom level table contains the map for the smallest allocatable
471 * units of memory (here, pages), and each level above it describes
472 * pairs of units from the levels below, hence, "buddies".
473 * At a high level, all that happens here is marking the table entry
474 * at the bottom level available, and propagating the changes upward
475 * as necessary, plus some accounting needed to play nicely with other
476 * parts of the VM system.
477 * At each level, we keep a list of pages, which are heads of continuous
478 * free pages of length of (1 << order) and marked with _mapcount -2. Page's
479 * order is recorded in page_private(page) field.
480 * So when we are allocating or freeing one, we can derive the state of the
481 * other. That is, if we allocate a small block, and both were
482 * free, the remainder of the region must be split into blocks.
483 * If a block is freed, and its buddy is also free, then this
484 * triggers coalescing into a block of larger size.
489 static inline void __free_one_page(struct page
*page
,
490 struct zone
*zone
, unsigned int order
,
493 unsigned long page_idx
;
494 unsigned long combined_idx
;
495 unsigned long uninitialized_var(buddy_idx
);
498 if (unlikely(PageCompound(page
)))
499 if (unlikely(destroy_compound_page(page
, order
)))
502 VM_BUG_ON(migratetype
== -1);
504 page_idx
= page_to_pfn(page
) & ((1 << MAX_ORDER
) - 1);
506 VM_BUG_ON(page_idx
& ((1 << order
) - 1));
507 VM_BUG_ON(bad_range(zone
, page
));
509 while (order
< MAX_ORDER
-1) {
510 buddy_idx
= __find_buddy_index(page_idx
, order
);
511 buddy
= page
+ (buddy_idx
- page_idx
);
512 if (!page_is_buddy(page
, buddy
, order
))
515 /* Our buddy is free, merge with it and move up one order. */
516 list_del(&buddy
->lru
);
517 zone
->free_area
[order
].nr_free
--;
518 rmv_page_order(buddy
);
519 combined_idx
= buddy_idx
& page_idx
;
520 page
= page
+ (combined_idx
- page_idx
);
521 page_idx
= combined_idx
;
524 set_page_order(page
, order
);
527 * If this is not the largest possible page, check if the buddy
528 * of the next-highest order is free. If it is, it's possible
529 * that pages are being freed that will coalesce soon. In case,
530 * that is happening, add the free page to the tail of the list
531 * so it's less likely to be used soon and more likely to be merged
532 * as a higher order page
534 if ((order
< MAX_ORDER
-2) && pfn_valid_within(page_to_pfn(buddy
))) {
535 struct page
*higher_page
, *higher_buddy
;
536 combined_idx
= buddy_idx
& page_idx
;
537 higher_page
= page
+ (combined_idx
- page_idx
);
538 buddy_idx
= __find_buddy_index(combined_idx
, order
+ 1);
539 higher_buddy
= page
+ (buddy_idx
- combined_idx
);
540 if (page_is_buddy(higher_page
, higher_buddy
, order
+ 1)) {
541 list_add_tail(&page
->lru
,
542 &zone
->free_area
[order
].free_list
[migratetype
]);
547 list_add(&page
->lru
, &zone
->free_area
[order
].free_list
[migratetype
]);
549 zone
->free_area
[order
].nr_free
++;
553 * free_page_mlock() -- clean up attempts to free and mlocked() page.
554 * Page should not be on lru, so no need to fix that up.
555 * free_pages_check() will verify...
557 static inline void free_page_mlock(struct page
*page
)
559 __dec_zone_page_state(page
, NR_MLOCK
);
560 __count_vm_event(UNEVICTABLE_MLOCKFREED
);
563 static inline int free_pages_check(struct page
*page
)
565 if (unlikely(page_mapcount(page
) |
566 (page
->mapping
!= NULL
) |
567 (atomic_read(&page
->_count
) != 0) |
568 (page
->flags
& PAGE_FLAGS_CHECK_AT_FREE
))) {
572 if (page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)
573 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
578 * Frees a number of pages from the PCP lists
579 * Assumes all pages on list are in same zone, and of same order.
580 * count is the number of pages to free.
582 * If the zone was previously in an "all pages pinned" state then look to
583 * see if this freeing clears that state.
585 * And clear the zone's pages_scanned counter, to hold off the "all pages are
586 * pinned" detection logic.
588 static void free_pcppages_bulk(struct zone
*zone
, int count
,
589 struct per_cpu_pages
*pcp
)
595 spin_lock(&zone
->lock
);
596 zone
->all_unreclaimable
= 0;
597 zone
->pages_scanned
= 0;
601 struct list_head
*list
;
604 * Remove pages from lists in a round-robin fashion. A
605 * batch_free count is maintained that is incremented when an
606 * empty list is encountered. This is so more pages are freed
607 * off fuller lists instead of spinning excessively around empty
612 if (++migratetype
== MIGRATE_PCPTYPES
)
614 list
= &pcp
->lists
[migratetype
];
615 } while (list_empty(list
));
618 page
= list_entry(list
->prev
, struct page
, lru
);
619 /* must delete as __free_one_page list manipulates */
620 list_del(&page
->lru
);
621 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
622 __free_one_page(page
, zone
, 0, page_private(page
));
623 trace_mm_page_pcpu_drain(page
, 0, page_private(page
));
624 } while (--to_free
&& --batch_free
&& !list_empty(list
));
626 __mod_zone_page_state(zone
, NR_FREE_PAGES
, count
);
627 spin_unlock(&zone
->lock
);
630 static void free_one_page(struct zone
*zone
, struct page
*page
, int order
,
633 spin_lock(&zone
->lock
);
634 zone
->all_unreclaimable
= 0;
635 zone
->pages_scanned
= 0;
637 __free_one_page(page
, zone
, order
, migratetype
);
638 __mod_zone_page_state(zone
, NR_FREE_PAGES
, 1 << order
);
639 spin_unlock(&zone
->lock
);
642 static bool free_pages_prepare(struct page
*page
, unsigned int order
)
647 trace_mm_page_free_direct(page
, order
);
648 kmemcheck_free_shadow(page
, order
);
651 page
->mapping
= NULL
;
652 for (i
= 0; i
< (1 << order
); i
++)
653 bad
+= free_pages_check(page
+ i
);
657 if (!PageHighMem(page
)) {
658 debug_check_no_locks_freed(page_address(page
),PAGE_SIZE
<<order
);
659 debug_check_no_obj_freed(page_address(page
),
662 arch_free_page(page
, order
);
663 kernel_map_pages(page
, 1 << order
, 0);
668 static void __free_pages_ok(struct page
*page
, unsigned int order
)
671 int wasMlocked
= __TestClearPageMlocked(page
);
673 if (!free_pages_prepare(page
, order
))
676 local_irq_save(flags
);
677 if (unlikely(wasMlocked
))
678 free_page_mlock(page
);
679 __count_vm_events(PGFREE
, 1 << order
);
680 free_one_page(page_zone(page
), page
, order
,
681 get_pageblock_migratetype(page
));
682 local_irq_restore(flags
);
686 * permit the bootmem allocator to evade page validation on high-order frees
688 void __meminit
__free_pages_bootmem(struct page
*page
, unsigned int order
)
691 __ClearPageReserved(page
);
692 set_page_count(page
, 0);
693 set_page_refcounted(page
);
699 for (loop
= 0; loop
< BITS_PER_LONG
; loop
++) {
700 struct page
*p
= &page
[loop
];
702 if (loop
+ 1 < BITS_PER_LONG
)
704 __ClearPageReserved(p
);
705 set_page_count(p
, 0);
708 set_page_refcounted(page
);
709 __free_pages(page
, order
);
715 * The order of subdivision here is critical for the IO subsystem.
716 * Please do not alter this order without good reasons and regression
717 * testing. Specifically, as large blocks of memory are subdivided,
718 * the order in which smaller blocks are delivered depends on the order
719 * they're subdivided in this function. This is the primary factor
720 * influencing the order in which pages are delivered to the IO
721 * subsystem according to empirical testing, and this is also justified
722 * by considering the behavior of a buddy system containing a single
723 * large block of memory acted on by a series of small allocations.
724 * This behavior is a critical factor in sglist merging's success.
728 static inline void expand(struct zone
*zone
, struct page
*page
,
729 int low
, int high
, struct free_area
*area
,
732 unsigned long size
= 1 << high
;
738 VM_BUG_ON(bad_range(zone
, &page
[size
]));
739 list_add(&page
[size
].lru
, &area
->free_list
[migratetype
]);
741 set_page_order(&page
[size
], high
);
746 * This page is about to be returned from the page allocator
748 static inline int check_new_page(struct page
*page
)
750 if (unlikely(page_mapcount(page
) |
751 (page
->mapping
!= NULL
) |
752 (atomic_read(&page
->_count
) != 0) |
753 (page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
))) {
760 static int prep_new_page(struct page
*page
, int order
, gfp_t gfp_flags
)
764 for (i
= 0; i
< (1 << order
); i
++) {
765 struct page
*p
= page
+ i
;
766 if (unlikely(check_new_page(p
)))
770 set_page_private(page
, 0);
771 set_page_refcounted(page
);
773 arch_alloc_page(page
, order
);
774 kernel_map_pages(page
, 1 << order
, 1);
776 if (gfp_flags
& __GFP_ZERO
)
777 prep_zero_page(page
, order
, gfp_flags
);
779 if (order
&& (gfp_flags
& __GFP_COMP
))
780 prep_compound_page(page
, order
);
786 * Go through the free lists for the given migratetype and remove
787 * the smallest available page from the freelists
790 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
793 unsigned int current_order
;
794 struct free_area
* area
;
797 /* Find a page of the appropriate size in the preferred list */
798 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
799 area
= &(zone
->free_area
[current_order
]);
800 if (list_empty(&area
->free_list
[migratetype
]))
803 page
= list_entry(area
->free_list
[migratetype
].next
,
805 list_del(&page
->lru
);
806 rmv_page_order(page
);
808 expand(zone
, page
, order
, current_order
, area
, migratetype
);
817 * This array describes the order lists are fallen back to when
818 * the free lists for the desirable migrate type are depleted
820 static int fallbacks
[MIGRATE_TYPES
][MIGRATE_TYPES
-1] = {
821 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
822 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
823 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_RESERVE
},
824 [MIGRATE_RESERVE
] = { MIGRATE_RESERVE
, MIGRATE_RESERVE
, MIGRATE_RESERVE
}, /* Never used */
828 * Move the free pages in a range to the free lists of the requested type.
829 * Note that start_page and end_pages are not aligned on a pageblock
830 * boundary. If alignment is required, use move_freepages_block()
832 static int move_freepages(struct zone
*zone
,
833 struct page
*start_page
, struct page
*end_page
,
840 #ifndef CONFIG_HOLES_IN_ZONE
842 * page_zone is not safe to call in this context when
843 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
844 * anyway as we check zone boundaries in move_freepages_block().
845 * Remove at a later date when no bug reports exist related to
846 * grouping pages by mobility
848 BUG_ON(page_zone(start_page
) != page_zone(end_page
));
851 for (page
= start_page
; page
<= end_page
;) {
852 /* Make sure we are not inadvertently changing nodes */
853 VM_BUG_ON(page_to_nid(page
) != zone_to_nid(zone
));
855 if (!pfn_valid_within(page_to_pfn(page
))) {
860 if (!PageBuddy(page
)) {
865 order
= page_order(page
);
866 list_del(&page
->lru
);
868 &zone
->free_area
[order
].free_list
[migratetype
]);
870 pages_moved
+= 1 << order
;
876 static int move_freepages_block(struct zone
*zone
, struct page
*page
,
879 unsigned long start_pfn
, end_pfn
;
880 struct page
*start_page
, *end_page
;
882 start_pfn
= page_to_pfn(page
);
883 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
884 start_page
= pfn_to_page(start_pfn
);
885 end_page
= start_page
+ pageblock_nr_pages
- 1;
886 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
888 /* Do not cross zone boundaries */
889 if (start_pfn
< zone
->zone_start_pfn
)
891 if (end_pfn
>= zone
->zone_start_pfn
+ zone
->spanned_pages
)
894 return move_freepages(zone
, start_page
, end_page
, migratetype
);
897 static void change_pageblock_range(struct page
*pageblock_page
,
898 int start_order
, int migratetype
)
900 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
902 while (nr_pageblocks
--) {
903 set_pageblock_migratetype(pageblock_page
, migratetype
);
904 pageblock_page
+= pageblock_nr_pages
;
908 /* Remove an element from the buddy allocator from the fallback list */
909 static inline struct page
*
910 __rmqueue_fallback(struct zone
*zone
, int order
, int start_migratetype
)
912 struct free_area
* area
;
917 /* Find the largest possible block of pages in the other list */
918 for (current_order
= MAX_ORDER
-1; current_order
>= order
;
920 for (i
= 0; i
< MIGRATE_TYPES
- 1; i
++) {
921 migratetype
= fallbacks
[start_migratetype
][i
];
923 /* MIGRATE_RESERVE handled later if necessary */
924 if (migratetype
== MIGRATE_RESERVE
)
927 area
= &(zone
->free_area
[current_order
]);
928 if (list_empty(&area
->free_list
[migratetype
]))
931 page
= list_entry(area
->free_list
[migratetype
].next
,
936 * If breaking a large block of pages, move all free
937 * pages to the preferred allocation list. If falling
938 * back for a reclaimable kernel allocation, be more
939 * agressive about taking ownership of free pages
941 if (unlikely(current_order
>= (pageblock_order
>> 1)) ||
942 start_migratetype
== MIGRATE_RECLAIMABLE
||
943 page_group_by_mobility_disabled
) {
945 pages
= move_freepages_block(zone
, page
,
948 /* Claim the whole block if over half of it is free */
949 if (pages
>= (1 << (pageblock_order
-1)) ||
950 page_group_by_mobility_disabled
)
951 set_pageblock_migratetype(page
,
954 migratetype
= start_migratetype
;
957 /* Remove the page from the freelists */
958 list_del(&page
->lru
);
959 rmv_page_order(page
);
961 /* Take ownership for orders >= pageblock_order */
962 if (current_order
>= pageblock_order
)
963 change_pageblock_range(page
, current_order
,
966 expand(zone
, page
, order
, current_order
, area
, migratetype
);
968 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
969 start_migratetype
, migratetype
);
979 * Do the hard work of removing an element from the buddy allocator.
980 * Call me with the zone->lock already held.
982 static struct page
*__rmqueue(struct zone
*zone
, unsigned int order
,
988 page
= __rmqueue_smallest(zone
, order
, migratetype
);
990 if (unlikely(!page
) && migratetype
!= MIGRATE_RESERVE
) {
991 page
= __rmqueue_fallback(zone
, order
, migratetype
);
994 * Use MIGRATE_RESERVE rather than fail an allocation. goto
995 * is used because __rmqueue_smallest is an inline function
996 * and we want just one call site
999 migratetype
= MIGRATE_RESERVE
;
1004 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
1009 * Obtain a specified number of elements from the buddy allocator, all under
1010 * a single hold of the lock, for efficiency. Add them to the supplied list.
1011 * Returns the number of new pages which were placed at *list.
1013 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
1014 unsigned long count
, struct list_head
*list
,
1015 int migratetype
, int cold
)
1019 spin_lock(&zone
->lock
);
1020 for (i
= 0; i
< count
; ++i
) {
1021 struct page
*page
= __rmqueue(zone
, order
, migratetype
);
1022 if (unlikely(page
== NULL
))
1026 * Split buddy pages returned by expand() are received here
1027 * in physical page order. The page is added to the callers and
1028 * list and the list head then moves forward. From the callers
1029 * perspective, the linked list is ordered by page number in
1030 * some conditions. This is useful for IO devices that can
1031 * merge IO requests if the physical pages are ordered
1034 if (likely(cold
== 0))
1035 list_add(&page
->lru
, list
);
1037 list_add_tail(&page
->lru
, list
);
1038 set_page_private(page
, migratetype
);
1041 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
1042 spin_unlock(&zone
->lock
);
1048 * Called from the vmstat counter updater to drain pagesets of this
1049 * currently executing processor on remote nodes after they have
1052 * Note that this function must be called with the thread pinned to
1053 * a single processor.
1055 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
1057 unsigned long flags
;
1060 local_irq_save(flags
);
1061 if (pcp
->count
>= pcp
->batch
)
1062 to_drain
= pcp
->batch
;
1064 to_drain
= pcp
->count
;
1065 free_pcppages_bulk(zone
, to_drain
, pcp
);
1066 pcp
->count
-= to_drain
;
1067 local_irq_restore(flags
);
1072 * Drain pages of the indicated processor.
1074 * The processor must either be the current processor and the
1075 * thread pinned to the current processor or a processor that
1078 static void drain_pages(unsigned int cpu
)
1080 unsigned long flags
;
1083 for_each_populated_zone(zone
) {
1084 struct per_cpu_pageset
*pset
;
1085 struct per_cpu_pages
*pcp
;
1087 local_irq_save(flags
);
1088 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
1091 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
1093 local_irq_restore(flags
);
1098 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1100 void drain_local_pages(void *arg
)
1102 drain_pages(smp_processor_id());
1106 * Spill all the per-cpu pages from all CPUs back into the buddy allocator
1108 void drain_all_pages(void)
1110 on_each_cpu(drain_local_pages
, NULL
, 1);
1113 #ifdef CONFIG_HIBERNATION
1115 void mark_free_pages(struct zone
*zone
)
1117 unsigned long pfn
, max_zone_pfn
;
1118 unsigned long flags
;
1120 struct list_head
*curr
;
1122 if (!zone
->spanned_pages
)
1125 spin_lock_irqsave(&zone
->lock
, flags
);
1127 max_zone_pfn
= zone
->zone_start_pfn
+ zone
->spanned_pages
;
1128 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
1129 if (pfn_valid(pfn
)) {
1130 struct page
*page
= pfn_to_page(pfn
);
1132 if (!swsusp_page_is_forbidden(page
))
1133 swsusp_unset_page_free(page
);
1136 for_each_migratetype_order(order
, t
) {
1137 list_for_each(curr
, &zone
->free_area
[order
].free_list
[t
]) {
1140 pfn
= page_to_pfn(list_entry(curr
, struct page
, lru
));
1141 for (i
= 0; i
< (1UL << order
); i
++)
1142 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
1145 spin_unlock_irqrestore(&zone
->lock
, flags
);
1147 #endif /* CONFIG_PM */
1150 * Free a 0-order page
1151 * cold == 1 ? free a cold page : free a hot page
1153 void free_hot_cold_page(struct page
*page
, int cold
)
1155 struct zone
*zone
= page_zone(page
);
1156 struct per_cpu_pages
*pcp
;
1157 unsigned long flags
;
1159 int wasMlocked
= __TestClearPageMlocked(page
);
1161 if (!free_pages_prepare(page
, 0))
1164 migratetype
= get_pageblock_migratetype(page
);
1165 set_page_private(page
, migratetype
);
1166 local_irq_save(flags
);
1167 if (unlikely(wasMlocked
))
1168 free_page_mlock(page
);
1169 __count_vm_event(PGFREE
);
1172 * We only track unmovable, reclaimable and movable on pcp lists.
1173 * Free ISOLATE pages back to the allocator because they are being
1174 * offlined but treat RESERVE as movable pages so we can get those
1175 * areas back if necessary. Otherwise, we may have to free
1176 * excessively into the page allocator
1178 if (migratetype
>= MIGRATE_PCPTYPES
) {
1179 if (unlikely(migratetype
== MIGRATE_ISOLATE
)) {
1180 free_one_page(zone
, page
, 0, migratetype
);
1183 migratetype
= MIGRATE_MOVABLE
;
1186 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
1188 list_add_tail(&page
->lru
, &pcp
->lists
[migratetype
]);
1190 list_add(&page
->lru
, &pcp
->lists
[migratetype
]);
1192 if (pcp
->count
>= pcp
->high
) {
1193 free_pcppages_bulk(zone
, pcp
->batch
, pcp
);
1194 pcp
->count
-= pcp
->batch
;
1198 local_irq_restore(flags
);
1202 * split_page takes a non-compound higher-order page, and splits it into
1203 * n (1<<order) sub-pages: page[0..n]
1204 * Each sub-page must be freed individually.
1206 * Note: this is probably too low level an operation for use in drivers.
1207 * Please consult with lkml before using this in your driver.
1209 void split_page(struct page
*page
, unsigned int order
)
1213 VM_BUG_ON(PageCompound(page
));
1214 VM_BUG_ON(!page_count(page
));
1216 #ifdef CONFIG_KMEMCHECK
1218 * Split shadow pages too, because free(page[0]) would
1219 * otherwise free the whole shadow.
1221 if (kmemcheck_page_is_tracked(page
))
1222 split_page(virt_to_page(page
[0].shadow
), order
);
1225 for (i
= 1; i
< (1 << order
); i
++)
1226 set_page_refcounted(page
+ i
);
1230 * Similar to split_page except the page is already free. As this is only
1231 * being used for migration, the migratetype of the block also changes.
1232 * As this is called with interrupts disabled, the caller is responsible
1233 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1236 * Note: this is probably too low level an operation for use in drivers.
1237 * Please consult with lkml before using this in your driver.
1239 int split_free_page(struct page
*page
)
1242 unsigned long watermark
;
1245 BUG_ON(!PageBuddy(page
));
1247 zone
= page_zone(page
);
1248 order
= page_order(page
);
1250 /* Obey watermarks as if the page was being allocated */
1251 watermark
= low_wmark_pages(zone
) + (1 << order
);
1252 if (!zone_watermark_ok(zone
, 0, watermark
, 0, 0))
1255 /* Remove page from free list */
1256 list_del(&page
->lru
);
1257 zone
->free_area
[order
].nr_free
--;
1258 rmv_page_order(page
);
1259 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(1UL << order
));
1261 /* Split into individual pages */
1262 set_page_refcounted(page
);
1263 split_page(page
, order
);
1265 if (order
>= pageblock_order
- 1) {
1266 struct page
*endpage
= page
+ (1 << order
) - 1;
1267 for (; page
< endpage
; page
+= pageblock_nr_pages
)
1268 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
1275 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1276 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1280 struct page
*buffered_rmqueue(struct zone
*preferred_zone
,
1281 struct zone
*zone
, int order
, gfp_t gfp_flags
,
1284 unsigned long flags
;
1286 int cold
= !!(gfp_flags
& __GFP_COLD
);
1289 if (likely(order
== 0)) {
1290 struct per_cpu_pages
*pcp
;
1291 struct list_head
*list
;
1293 local_irq_save(flags
);
1294 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
1295 list
= &pcp
->lists
[migratetype
];
1296 if (list_empty(list
)) {
1297 pcp
->count
+= rmqueue_bulk(zone
, 0,
1300 if (unlikely(list_empty(list
)))
1305 page
= list_entry(list
->prev
, struct page
, lru
);
1307 page
= list_entry(list
->next
, struct page
, lru
);
1309 list_del(&page
->lru
);
1312 if (unlikely(gfp_flags
& __GFP_NOFAIL
)) {
1314 * __GFP_NOFAIL is not to be used in new code.
1316 * All __GFP_NOFAIL callers should be fixed so that they
1317 * properly detect and handle allocation failures.
1319 * We most definitely don't want callers attempting to
1320 * allocate greater than order-1 page units with
1323 WARN_ON_ONCE(order
> 1);
1325 spin_lock_irqsave(&zone
->lock
, flags
);
1326 page
= __rmqueue(zone
, order
, migratetype
);
1327 spin_unlock(&zone
->lock
);
1330 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(1 << order
));
1333 __count_zone_vm_events(PGALLOC
, zone
, 1 << order
);
1334 zone_statistics(preferred_zone
, zone
);
1335 local_irq_restore(flags
);
1337 VM_BUG_ON(bad_range(zone
, page
));
1338 if (prep_new_page(page
, order
, gfp_flags
))
1343 local_irq_restore(flags
);
1347 /* The ALLOC_WMARK bits are used as an index to zone->watermark */
1348 #define ALLOC_WMARK_MIN WMARK_MIN
1349 #define ALLOC_WMARK_LOW WMARK_LOW
1350 #define ALLOC_WMARK_HIGH WMARK_HIGH
1351 #define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */
1353 /* Mask to get the watermark bits */
1354 #define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1)
1356 #define ALLOC_HARDER 0x10 /* try to alloc harder */
1357 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
1358 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
1360 #ifdef CONFIG_FAIL_PAGE_ALLOC
1362 static struct fail_page_alloc_attr
{
1363 struct fault_attr attr
;
1365 u32 ignore_gfp_highmem
;
1366 u32 ignore_gfp_wait
;
1369 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1371 struct dentry
*ignore_gfp_highmem_file
;
1372 struct dentry
*ignore_gfp_wait_file
;
1373 struct dentry
*min_order_file
;
1375 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1377 } fail_page_alloc
= {
1378 .attr
= FAULT_ATTR_INITIALIZER
,
1379 .ignore_gfp_wait
= 1,
1380 .ignore_gfp_highmem
= 1,
1384 static int __init
setup_fail_page_alloc(char *str
)
1386 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
1388 __setup("fail_page_alloc=", setup_fail_page_alloc
);
1390 static int should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1392 if (order
< fail_page_alloc
.min_order
)
1394 if (gfp_mask
& __GFP_NOFAIL
)
1396 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
1398 if (fail_page_alloc
.ignore_gfp_wait
&& (gfp_mask
& __GFP_WAIT
))
1401 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
1404 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1406 static int __init
fail_page_alloc_debugfs(void)
1408 mode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
1412 err
= init_fault_attr_dentries(&fail_page_alloc
.attr
,
1416 dir
= fail_page_alloc
.attr
.dentries
.dir
;
1418 fail_page_alloc
.ignore_gfp_wait_file
=
1419 debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
1420 &fail_page_alloc
.ignore_gfp_wait
);
1422 fail_page_alloc
.ignore_gfp_highmem_file
=
1423 debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
1424 &fail_page_alloc
.ignore_gfp_highmem
);
1425 fail_page_alloc
.min_order_file
=
1426 debugfs_create_u32("min-order", mode
, dir
,
1427 &fail_page_alloc
.min_order
);
1429 if (!fail_page_alloc
.ignore_gfp_wait_file
||
1430 !fail_page_alloc
.ignore_gfp_highmem_file
||
1431 !fail_page_alloc
.min_order_file
) {
1433 debugfs_remove(fail_page_alloc
.ignore_gfp_wait_file
);
1434 debugfs_remove(fail_page_alloc
.ignore_gfp_highmem_file
);
1435 debugfs_remove(fail_page_alloc
.min_order_file
);
1436 cleanup_fault_attr_dentries(&fail_page_alloc
.attr
);
1442 late_initcall(fail_page_alloc_debugfs
);
1444 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1446 #else /* CONFIG_FAIL_PAGE_ALLOC */
1448 static inline int should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1453 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1456 * Return true if free pages are above 'mark'. This takes into account the order
1457 * of the allocation.
1459 static bool __zone_watermark_ok(struct zone
*z
, int order
, unsigned long mark
,
1460 int classzone_idx
, int alloc_flags
, long free_pages
)
1462 /* free_pages my go negative - that's OK */
1466 free_pages
-= (1 << order
) + 1;
1467 if (alloc_flags
& ALLOC_HIGH
)
1469 if (alloc_flags
& ALLOC_HARDER
)
1472 if (free_pages
<= min
+ z
->lowmem_reserve
[classzone_idx
])
1474 for (o
= 0; o
< order
; o
++) {
1475 /* At the next order, this order's pages become unavailable */
1476 free_pages
-= z
->free_area
[o
].nr_free
<< o
;
1478 /* Require fewer higher order pages to be free */
1481 if (free_pages
<= min
)
1487 bool zone_watermark_ok(struct zone
*z
, int order
, unsigned long mark
,
1488 int classzone_idx
, int alloc_flags
)
1490 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
1491 zone_page_state(z
, NR_FREE_PAGES
));
1494 bool zone_watermark_ok_safe(struct zone
*z
, int order
, unsigned long mark
,
1495 int classzone_idx
, int alloc_flags
)
1497 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
1499 if (z
->percpu_drift_mark
&& free_pages
< z
->percpu_drift_mark
)
1500 free_pages
= zone_page_state_snapshot(z
, NR_FREE_PAGES
);
1502 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
1508 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1509 * skip over zones that are not allowed by the cpuset, or that have
1510 * been recently (in last second) found to be nearly full. See further
1511 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1512 * that have to skip over a lot of full or unallowed zones.
1514 * If the zonelist cache is present in the passed in zonelist, then
1515 * returns a pointer to the allowed node mask (either the current
1516 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1518 * If the zonelist cache is not available for this zonelist, does
1519 * nothing and returns NULL.
1521 * If the fullzones BITMAP in the zonelist cache is stale (more than
1522 * a second since last zap'd) then we zap it out (clear its bits.)
1524 * We hold off even calling zlc_setup, until after we've checked the
1525 * first zone in the zonelist, on the theory that most allocations will
1526 * be satisfied from that first zone, so best to examine that zone as
1527 * quickly as we can.
1529 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1531 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1532 nodemask_t
*allowednodes
; /* zonelist_cache approximation */
1534 zlc
= zonelist
->zlcache_ptr
;
1538 if (time_after(jiffies
, zlc
->last_full_zap
+ HZ
)) {
1539 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
1540 zlc
->last_full_zap
= jiffies
;
1543 allowednodes
= !in_interrupt() && (alloc_flags
& ALLOC_CPUSET
) ?
1544 &cpuset_current_mems_allowed
:
1545 &node_states
[N_HIGH_MEMORY
];
1546 return allowednodes
;
1550 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1551 * if it is worth looking at further for free memory:
1552 * 1) Check that the zone isn't thought to be full (doesn't have its
1553 * bit set in the zonelist_cache fullzones BITMAP).
1554 * 2) Check that the zones node (obtained from the zonelist_cache
1555 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1556 * Return true (non-zero) if zone is worth looking at further, or
1557 * else return false (zero) if it is not.
1559 * This check -ignores- the distinction between various watermarks,
1560 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1561 * found to be full for any variation of these watermarks, it will
1562 * be considered full for up to one second by all requests, unless
1563 * we are so low on memory on all allowed nodes that we are forced
1564 * into the second scan of the zonelist.
1566 * In the second scan we ignore this zonelist cache and exactly
1567 * apply the watermarks to all zones, even it is slower to do so.
1568 * We are low on memory in the second scan, and should leave no stone
1569 * unturned looking for a free page.
1571 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1572 nodemask_t
*allowednodes
)
1574 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1575 int i
; /* index of *z in zonelist zones */
1576 int n
; /* node that zone *z is on */
1578 zlc
= zonelist
->zlcache_ptr
;
1582 i
= z
- zonelist
->_zonerefs
;
1585 /* This zone is worth trying if it is allowed but not full */
1586 return node_isset(n
, *allowednodes
) && !test_bit(i
, zlc
->fullzones
);
1590 * Given 'z' scanning a zonelist, set the corresponding bit in
1591 * zlc->fullzones, so that subsequent attempts to allocate a page
1592 * from that zone don't waste time re-examining it.
1594 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1596 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1597 int i
; /* index of *z in zonelist zones */
1599 zlc
= zonelist
->zlcache_ptr
;
1603 i
= z
- zonelist
->_zonerefs
;
1605 set_bit(i
, zlc
->fullzones
);
1608 #else /* CONFIG_NUMA */
1610 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1615 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1616 nodemask_t
*allowednodes
)
1621 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1624 #endif /* CONFIG_NUMA */
1627 * get_page_from_freelist goes through the zonelist trying to allocate
1630 static struct page
*
1631 get_page_from_freelist(gfp_t gfp_mask
, nodemask_t
*nodemask
, unsigned int order
,
1632 struct zonelist
*zonelist
, int high_zoneidx
, int alloc_flags
,
1633 struct zone
*preferred_zone
, int migratetype
)
1636 struct page
*page
= NULL
;
1639 nodemask_t
*allowednodes
= NULL
;/* zonelist_cache approximation */
1640 int zlc_active
= 0; /* set if using zonelist_cache */
1641 int did_zlc_setup
= 0; /* just call zlc_setup() one time */
1643 classzone_idx
= zone_idx(preferred_zone
);
1646 * Scan zonelist, looking for a zone with enough free.
1647 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1649 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
1650 high_zoneidx
, nodemask
) {
1651 if (NUMA_BUILD
&& zlc_active
&&
1652 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
1654 if ((alloc_flags
& ALLOC_CPUSET
) &&
1655 !cpuset_zone_allowed_softwall(zone
, gfp_mask
))
1658 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
1659 if (!(alloc_flags
& ALLOC_NO_WATERMARKS
)) {
1663 mark
= zone
->watermark
[alloc_flags
& ALLOC_WMARK_MASK
];
1664 if (zone_watermark_ok(zone
, order
, mark
,
1665 classzone_idx
, alloc_flags
))
1668 if (zone_reclaim_mode
== 0)
1669 goto this_zone_full
;
1671 ret
= zone_reclaim(zone
, gfp_mask
, order
);
1673 case ZONE_RECLAIM_NOSCAN
:
1676 case ZONE_RECLAIM_FULL
:
1677 /* scanned but unreclaimable */
1678 goto this_zone_full
;
1680 /* did we reclaim enough */
1681 if (!zone_watermark_ok(zone
, order
, mark
,
1682 classzone_idx
, alloc_flags
))
1683 goto this_zone_full
;
1688 page
= buffered_rmqueue(preferred_zone
, zone
, order
,
1689 gfp_mask
, migratetype
);
1694 zlc_mark_zone_full(zonelist
, z
);
1696 if (NUMA_BUILD
&& !did_zlc_setup
&& nr_online_nodes
> 1) {
1698 * we do zlc_setup after the first zone is tried but only
1699 * if there are multiple nodes make it worthwhile
1701 allowednodes
= zlc_setup(zonelist
, alloc_flags
);
1707 if (unlikely(NUMA_BUILD
&& page
== NULL
&& zlc_active
)) {
1708 /* Disable zlc cache for second zonelist scan */
1716 should_alloc_retry(gfp_t gfp_mask
, unsigned int order
,
1717 unsigned long pages_reclaimed
)
1719 /* Do not loop if specifically requested */
1720 if (gfp_mask
& __GFP_NORETRY
)
1724 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
1725 * means __GFP_NOFAIL, but that may not be true in other
1728 if (order
<= PAGE_ALLOC_COSTLY_ORDER
)
1732 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
1733 * specified, then we retry until we no longer reclaim any pages
1734 * (above), or we've reclaimed an order of pages at least as
1735 * large as the allocation's order. In both cases, if the
1736 * allocation still fails, we stop retrying.
1738 if (gfp_mask
& __GFP_REPEAT
&& pages_reclaimed
< (1 << order
))
1742 * Don't let big-order allocations loop unless the caller
1743 * explicitly requests that.
1745 if (gfp_mask
& __GFP_NOFAIL
)
1751 static inline struct page
*
1752 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
1753 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1754 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
1759 /* Acquire the OOM killer lock for the zones in zonelist */
1760 if (!try_set_zonelist_oom(zonelist
, gfp_mask
)) {
1761 schedule_timeout_uninterruptible(1);
1766 * Go through the zonelist yet one more time, keep very high watermark
1767 * here, this is only to catch a parallel oom killing, we must fail if
1768 * we're still under heavy pressure.
1770 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
,
1771 order
, zonelist
, high_zoneidx
,
1772 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
,
1773 preferred_zone
, migratetype
);
1777 if (!(gfp_mask
& __GFP_NOFAIL
)) {
1778 /* The OOM killer will not help higher order allocs */
1779 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
1781 /* The OOM killer does not needlessly kill tasks for lowmem */
1782 if (high_zoneidx
< ZONE_NORMAL
)
1785 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
1786 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
1787 * The caller should handle page allocation failure by itself if
1788 * it specifies __GFP_THISNODE.
1789 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
1791 if (gfp_mask
& __GFP_THISNODE
)
1794 /* Exhausted what can be done so it's blamo time */
1795 out_of_memory(zonelist
, gfp_mask
, order
, nodemask
);
1798 clear_zonelist_oom(zonelist
, gfp_mask
);
1802 #ifdef CONFIG_COMPACTION
1803 /* Try memory compaction for high-order allocations before reclaim */
1804 static struct page
*
1805 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
1806 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1807 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
1808 int migratetype
, unsigned long *did_some_progress
,
1809 bool sync_migration
)
1813 if (!order
|| compaction_deferred(preferred_zone
))
1816 current
->flags
|= PF_MEMALLOC
;
1817 *did_some_progress
= try_to_compact_pages(zonelist
, order
, gfp_mask
,
1818 nodemask
, sync_migration
);
1819 current
->flags
&= ~PF_MEMALLOC
;
1820 if (*did_some_progress
!= COMPACT_SKIPPED
) {
1822 /* Page migration frees to the PCP lists but we want merging */
1823 drain_pages(get_cpu());
1826 page
= get_page_from_freelist(gfp_mask
, nodemask
,
1827 order
, zonelist
, high_zoneidx
,
1828 alloc_flags
, preferred_zone
,
1831 preferred_zone
->compact_considered
= 0;
1832 preferred_zone
->compact_defer_shift
= 0;
1833 count_vm_event(COMPACTSUCCESS
);
1838 * It's bad if compaction run occurs and fails.
1839 * The most likely reason is that pages exist,
1840 * but not enough to satisfy watermarks.
1842 count_vm_event(COMPACTFAIL
);
1843 defer_compaction(preferred_zone
);
1851 static inline struct page
*
1852 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
1853 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1854 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
1855 int migratetype
, unsigned long *did_some_progress
,
1856 bool sync_migration
)
1860 #endif /* CONFIG_COMPACTION */
1862 /* The really slow allocator path where we enter direct reclaim */
1863 static inline struct page
*
1864 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
1865 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1866 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
1867 int migratetype
, unsigned long *did_some_progress
)
1869 struct page
*page
= NULL
;
1870 struct reclaim_state reclaim_state
;
1871 bool drained
= false;
1875 /* We now go into synchronous reclaim */
1876 cpuset_memory_pressure_bump();
1877 current
->flags
|= PF_MEMALLOC
;
1878 lockdep_set_current_reclaim_state(gfp_mask
);
1879 reclaim_state
.reclaimed_slab
= 0;
1880 current
->reclaim_state
= &reclaim_state
;
1882 *did_some_progress
= try_to_free_pages(zonelist
, order
, gfp_mask
, nodemask
);
1884 current
->reclaim_state
= NULL
;
1885 lockdep_clear_current_reclaim_state();
1886 current
->flags
&= ~PF_MEMALLOC
;
1890 if (unlikely(!(*did_some_progress
)))
1894 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
1895 zonelist
, high_zoneidx
,
1896 alloc_flags
, preferred_zone
,
1900 * If an allocation failed after direct reclaim, it could be because
1901 * pages are pinned on the per-cpu lists. Drain them and try again
1903 if (!page
&& !drained
) {
1913 * This is called in the allocator slow-path if the allocation request is of
1914 * sufficient urgency to ignore watermarks and take other desperate measures
1916 static inline struct page
*
1917 __alloc_pages_high_priority(gfp_t gfp_mask
, unsigned int order
,
1918 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1919 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
1925 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
1926 zonelist
, high_zoneidx
, ALLOC_NO_WATERMARKS
,
1927 preferred_zone
, migratetype
);
1929 if (!page
&& gfp_mask
& __GFP_NOFAIL
)
1930 wait_iff_congested(preferred_zone
, BLK_RW_ASYNC
, HZ
/50);
1931 } while (!page
&& (gfp_mask
& __GFP_NOFAIL
));
1937 void wake_all_kswapd(unsigned int order
, struct zonelist
*zonelist
,
1938 enum zone_type high_zoneidx
,
1939 enum zone_type classzone_idx
)
1944 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
)
1945 wakeup_kswapd(zone
, order
, classzone_idx
);
1949 gfp_to_alloc_flags(gfp_t gfp_mask
)
1951 int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
1952 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
1954 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
1955 BUILD_BUG_ON(__GFP_HIGH
!= (__force gfp_t
) ALLOC_HIGH
);
1958 * The caller may dip into page reserves a bit more if the caller
1959 * cannot run direct reclaim, or if the caller has realtime scheduling
1960 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
1961 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
1963 alloc_flags
|= (__force
int) (gfp_mask
& __GFP_HIGH
);
1967 * Not worth trying to allocate harder for
1968 * __GFP_NOMEMALLOC even if it can't schedule.
1970 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
1971 alloc_flags
|= ALLOC_HARDER
;
1973 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
1974 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1976 alloc_flags
&= ~ALLOC_CPUSET
;
1977 } else if (unlikely(rt_task(current
)) && !in_interrupt())
1978 alloc_flags
|= ALLOC_HARDER
;
1980 if (likely(!(gfp_mask
& __GFP_NOMEMALLOC
))) {
1981 if (!in_interrupt() &&
1982 ((current
->flags
& PF_MEMALLOC
) ||
1983 unlikely(test_thread_flag(TIF_MEMDIE
))))
1984 alloc_flags
|= ALLOC_NO_WATERMARKS
;
1990 static inline struct page
*
1991 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
1992 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1993 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
1996 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
1997 struct page
*page
= NULL
;
1999 unsigned long pages_reclaimed
= 0;
2000 unsigned long did_some_progress
;
2001 bool sync_migration
= false;
2004 * In the slowpath, we sanity check order to avoid ever trying to
2005 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2006 * be using allocators in order of preference for an area that is
2009 if (order
>= MAX_ORDER
) {
2010 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
2015 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2016 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2017 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2018 * using a larger set of nodes after it has established that the
2019 * allowed per node queues are empty and that nodes are
2022 if (NUMA_BUILD
&& (gfp_mask
& GFP_THISNODE
) == GFP_THISNODE
)
2026 if (!(gfp_mask
& __GFP_NO_KSWAPD
))
2027 wake_all_kswapd(order
, zonelist
, high_zoneidx
,
2028 zone_idx(preferred_zone
));
2031 * OK, we're below the kswapd watermark and have kicked background
2032 * reclaim. Now things get more complex, so set up alloc_flags according
2033 * to how we want to proceed.
2035 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
2038 * Find the true preferred zone if the allocation is unconstrained by
2041 if (!(alloc_flags
& ALLOC_CPUSET
) && !nodemask
)
2042 first_zones_zonelist(zonelist
, high_zoneidx
, NULL
,
2045 /* This is the last chance, in general, before the goto nopage. */
2046 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
, zonelist
,
2047 high_zoneidx
, alloc_flags
& ~ALLOC_NO_WATERMARKS
,
2048 preferred_zone
, migratetype
);
2053 /* Allocate without watermarks if the context allows */
2054 if (alloc_flags
& ALLOC_NO_WATERMARKS
) {
2055 page
= __alloc_pages_high_priority(gfp_mask
, order
,
2056 zonelist
, high_zoneidx
, nodemask
,
2057 preferred_zone
, migratetype
);
2062 /* Atomic allocations - we can't balance anything */
2066 /* Avoid recursion of direct reclaim */
2067 if (current
->flags
& PF_MEMALLOC
)
2070 /* Avoid allocations with no watermarks from looping endlessly */
2071 if (test_thread_flag(TIF_MEMDIE
) && !(gfp_mask
& __GFP_NOFAIL
))
2075 * Try direct compaction. The first pass is asynchronous. Subsequent
2076 * attempts after direct reclaim are synchronous
2078 page
= __alloc_pages_direct_compact(gfp_mask
, order
,
2079 zonelist
, high_zoneidx
,
2081 alloc_flags
, preferred_zone
,
2082 migratetype
, &did_some_progress
,
2086 sync_migration
= true;
2088 /* Try direct reclaim and then allocating */
2089 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
,
2090 zonelist
, high_zoneidx
,
2092 alloc_flags
, preferred_zone
,
2093 migratetype
, &did_some_progress
);
2098 * If we failed to make any progress reclaiming, then we are
2099 * running out of options and have to consider going OOM
2101 if (!did_some_progress
) {
2102 if ((gfp_mask
& __GFP_FS
) && !(gfp_mask
& __GFP_NORETRY
)) {
2103 if (oom_killer_disabled
)
2105 page
= __alloc_pages_may_oom(gfp_mask
, order
,
2106 zonelist
, high_zoneidx
,
2107 nodemask
, preferred_zone
,
2112 if (!(gfp_mask
& __GFP_NOFAIL
)) {
2114 * The oom killer is not called for high-order
2115 * allocations that may fail, so if no progress
2116 * is being made, there are no other options and
2117 * retrying is unlikely to help.
2119 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
2122 * The oom killer is not called for lowmem
2123 * allocations to prevent needlessly killing
2126 if (high_zoneidx
< ZONE_NORMAL
)
2134 /* Check if we should retry the allocation */
2135 pages_reclaimed
+= did_some_progress
;
2136 if (should_alloc_retry(gfp_mask
, order
, pages_reclaimed
)) {
2137 /* Wait for some write requests to complete then retry */
2138 wait_iff_congested(preferred_zone
, BLK_RW_ASYNC
, HZ
/50);
2142 * High-order allocations do not necessarily loop after
2143 * direct reclaim and reclaim/compaction depends on compaction
2144 * being called after reclaim so call directly if necessary
2146 page
= __alloc_pages_direct_compact(gfp_mask
, order
,
2147 zonelist
, high_zoneidx
,
2149 alloc_flags
, preferred_zone
,
2150 migratetype
, &did_some_progress
,
2157 if (!(gfp_mask
& __GFP_NOWARN
) && printk_ratelimit()) {
2158 printk(KERN_WARNING
"%s: page allocation failure."
2159 " order:%d, mode:0x%x\n",
2160 current
->comm
, order
, gfp_mask
);
2166 if (kmemcheck_enabled
)
2167 kmemcheck_pagealloc_alloc(page
, order
, gfp_mask
);
2173 * This is the 'heart' of the zoned buddy allocator.
2176 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
,
2177 struct zonelist
*zonelist
, nodemask_t
*nodemask
)
2179 enum zone_type high_zoneidx
= gfp_zone(gfp_mask
);
2180 struct zone
*preferred_zone
;
2182 int migratetype
= allocflags_to_migratetype(gfp_mask
);
2184 gfp_mask
&= gfp_allowed_mask
;
2186 lockdep_trace_alloc(gfp_mask
);
2188 might_sleep_if(gfp_mask
& __GFP_WAIT
);
2190 if (should_fail_alloc_page(gfp_mask
, order
))
2194 * Check the zones suitable for the gfp_mask contain at least one
2195 * valid zone. It's possible to have an empty zonelist as a result
2196 * of GFP_THISNODE and a memoryless node
2198 if (unlikely(!zonelist
->_zonerefs
->zone
))
2202 /* The preferred zone is used for statistics later */
2203 first_zones_zonelist(zonelist
, high_zoneidx
,
2204 nodemask
? : &cpuset_current_mems_allowed
,
2206 if (!preferred_zone
) {
2211 /* First allocation attempt */
2212 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
, order
,
2213 zonelist
, high_zoneidx
, ALLOC_WMARK_LOW
|ALLOC_CPUSET
,
2214 preferred_zone
, migratetype
);
2215 if (unlikely(!page
))
2216 page
= __alloc_pages_slowpath(gfp_mask
, order
,
2217 zonelist
, high_zoneidx
, nodemask
,
2218 preferred_zone
, migratetype
);
2221 trace_mm_page_alloc(page
, order
, gfp_mask
, migratetype
);
2224 EXPORT_SYMBOL(__alloc_pages_nodemask
);
2227 * Common helper functions.
2229 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
2234 * __get_free_pages() returns a 32-bit address, which cannot represent
2237 VM_BUG_ON((gfp_mask
& __GFP_HIGHMEM
) != 0);
2239 page
= alloc_pages(gfp_mask
, order
);
2242 return (unsigned long) page_address(page
);
2244 EXPORT_SYMBOL(__get_free_pages
);
2246 unsigned long get_zeroed_page(gfp_t gfp_mask
)
2248 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
2250 EXPORT_SYMBOL(get_zeroed_page
);
2252 void __pagevec_free(struct pagevec
*pvec
)
2254 int i
= pagevec_count(pvec
);
2257 trace_mm_pagevec_free(pvec
->pages
[i
], pvec
->cold
);
2258 free_hot_cold_page(pvec
->pages
[i
], pvec
->cold
);
2262 void __free_pages(struct page
*page
, unsigned int order
)
2264 if (put_page_testzero(page
)) {
2266 free_hot_cold_page(page
, 0);
2268 __free_pages_ok(page
, order
);
2272 EXPORT_SYMBOL(__free_pages
);
2274 void free_pages(unsigned long addr
, unsigned int order
)
2277 VM_BUG_ON(!virt_addr_valid((void *)addr
));
2278 __free_pages(virt_to_page((void *)addr
), order
);
2282 EXPORT_SYMBOL(free_pages
);
2285 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2286 * @size: the number of bytes to allocate
2287 * @gfp_mask: GFP flags for the allocation
2289 * This function is similar to alloc_pages(), except that it allocates the
2290 * minimum number of pages to satisfy the request. alloc_pages() can only
2291 * allocate memory in power-of-two pages.
2293 * This function is also limited by MAX_ORDER.
2295 * Memory allocated by this function must be released by free_pages_exact().
2297 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
2299 unsigned int order
= get_order(size
);
2302 addr
= __get_free_pages(gfp_mask
, order
);
2304 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
2305 unsigned long used
= addr
+ PAGE_ALIGN(size
);
2307 split_page(virt_to_page((void *)addr
), order
);
2308 while (used
< alloc_end
) {
2314 return (void *)addr
;
2316 EXPORT_SYMBOL(alloc_pages_exact
);
2319 * free_pages_exact - release memory allocated via alloc_pages_exact()
2320 * @virt: the value returned by alloc_pages_exact.
2321 * @size: size of allocation, same value as passed to alloc_pages_exact().
2323 * Release the memory allocated by a previous call to alloc_pages_exact.
2325 void free_pages_exact(void *virt
, size_t size
)
2327 unsigned long addr
= (unsigned long)virt
;
2328 unsigned long end
= addr
+ PAGE_ALIGN(size
);
2330 while (addr
< end
) {
2335 EXPORT_SYMBOL(free_pages_exact
);
2337 static unsigned int nr_free_zone_pages(int offset
)
2342 /* Just pick one node, since fallback list is circular */
2343 unsigned int sum
= 0;
2345 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
2347 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
2348 unsigned long size
= zone
->present_pages
;
2349 unsigned long high
= high_wmark_pages(zone
);
2358 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2360 unsigned int nr_free_buffer_pages(void)
2362 return nr_free_zone_pages(gfp_zone(GFP_USER
));
2364 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
2367 * Amount of free RAM allocatable within all zones
2369 unsigned int nr_free_pagecache_pages(void)
2371 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
2374 static inline void show_node(struct zone
*zone
)
2377 printk("Node %d ", zone_to_nid(zone
));
2380 void si_meminfo(struct sysinfo
*val
)
2382 val
->totalram
= totalram_pages
;
2384 val
->freeram
= global_page_state(NR_FREE_PAGES
);
2385 val
->bufferram
= nr_blockdev_pages();
2386 val
->totalhigh
= totalhigh_pages
;
2387 val
->freehigh
= nr_free_highpages();
2388 val
->mem_unit
= PAGE_SIZE
;
2391 EXPORT_SYMBOL(si_meminfo
);
2394 void si_meminfo_node(struct sysinfo
*val
, int nid
)
2396 pg_data_t
*pgdat
= NODE_DATA(nid
);
2398 val
->totalram
= pgdat
->node_present_pages
;
2399 val
->freeram
= node_page_state(nid
, NR_FREE_PAGES
);
2400 #ifdef CONFIG_HIGHMEM
2401 val
->totalhigh
= pgdat
->node_zones
[ZONE_HIGHMEM
].present_pages
;
2402 val
->freehigh
= zone_page_state(&pgdat
->node_zones
[ZONE_HIGHMEM
],
2408 val
->mem_unit
= PAGE_SIZE
;
2412 #define K(x) ((x) << (PAGE_SHIFT-10))
2415 * Show free area list (used inside shift_scroll-lock stuff)
2416 * We also calculate the percentage fragmentation. We do this by counting the
2417 * memory on each free list with the exception of the first item on the list.
2419 void show_free_areas(void)
2424 for_each_populated_zone(zone
) {
2426 printk("%s per-cpu:\n", zone
->name
);
2428 for_each_online_cpu(cpu
) {
2429 struct per_cpu_pageset
*pageset
;
2431 pageset
= per_cpu_ptr(zone
->pageset
, cpu
);
2433 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2434 cpu
, pageset
->pcp
.high
,
2435 pageset
->pcp
.batch
, pageset
->pcp
.count
);
2439 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2440 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2442 " dirty:%lu writeback:%lu unstable:%lu\n"
2443 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2444 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n",
2445 global_page_state(NR_ACTIVE_ANON
),
2446 global_page_state(NR_INACTIVE_ANON
),
2447 global_page_state(NR_ISOLATED_ANON
),
2448 global_page_state(NR_ACTIVE_FILE
),
2449 global_page_state(NR_INACTIVE_FILE
),
2450 global_page_state(NR_ISOLATED_FILE
),
2451 global_page_state(NR_UNEVICTABLE
),
2452 global_page_state(NR_FILE_DIRTY
),
2453 global_page_state(NR_WRITEBACK
),
2454 global_page_state(NR_UNSTABLE_NFS
),
2455 global_page_state(NR_FREE_PAGES
),
2456 global_page_state(NR_SLAB_RECLAIMABLE
),
2457 global_page_state(NR_SLAB_UNRECLAIMABLE
),
2458 global_page_state(NR_FILE_MAPPED
),
2459 global_page_state(NR_SHMEM
),
2460 global_page_state(NR_PAGETABLE
),
2461 global_page_state(NR_BOUNCE
));
2463 for_each_populated_zone(zone
) {
2472 " active_anon:%lukB"
2473 " inactive_anon:%lukB"
2474 " active_file:%lukB"
2475 " inactive_file:%lukB"
2476 " unevictable:%lukB"
2477 " isolated(anon):%lukB"
2478 " isolated(file):%lukB"
2485 " slab_reclaimable:%lukB"
2486 " slab_unreclaimable:%lukB"
2487 " kernel_stack:%lukB"
2491 " writeback_tmp:%lukB"
2492 " pages_scanned:%lu"
2493 " all_unreclaimable? %s"
2496 K(zone_page_state(zone
, NR_FREE_PAGES
)),
2497 K(min_wmark_pages(zone
)),
2498 K(low_wmark_pages(zone
)),
2499 K(high_wmark_pages(zone
)),
2500 K(zone_page_state(zone
, NR_ACTIVE_ANON
)),
2501 K(zone_page_state(zone
, NR_INACTIVE_ANON
)),
2502 K(zone_page_state(zone
, NR_ACTIVE_FILE
)),
2503 K(zone_page_state(zone
, NR_INACTIVE_FILE
)),
2504 K(zone_page_state(zone
, NR_UNEVICTABLE
)),
2505 K(zone_page_state(zone
, NR_ISOLATED_ANON
)),
2506 K(zone_page_state(zone
, NR_ISOLATED_FILE
)),
2507 K(zone
->present_pages
),
2508 K(zone_page_state(zone
, NR_MLOCK
)),
2509 K(zone_page_state(zone
, NR_FILE_DIRTY
)),
2510 K(zone_page_state(zone
, NR_WRITEBACK
)),
2511 K(zone_page_state(zone
, NR_FILE_MAPPED
)),
2512 K(zone_page_state(zone
, NR_SHMEM
)),
2513 K(zone_page_state(zone
, NR_SLAB_RECLAIMABLE
)),
2514 K(zone_page_state(zone
, NR_SLAB_UNRECLAIMABLE
)),
2515 zone_page_state(zone
, NR_KERNEL_STACK
) *
2517 K(zone_page_state(zone
, NR_PAGETABLE
)),
2518 K(zone_page_state(zone
, NR_UNSTABLE_NFS
)),
2519 K(zone_page_state(zone
, NR_BOUNCE
)),
2520 K(zone_page_state(zone
, NR_WRITEBACK_TEMP
)),
2521 zone
->pages_scanned
,
2522 (zone
->all_unreclaimable
? "yes" : "no")
2524 printk("lowmem_reserve[]:");
2525 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
2526 printk(" %lu", zone
->lowmem_reserve
[i
]);
2530 for_each_populated_zone(zone
) {
2531 unsigned long nr
[MAX_ORDER
], flags
, order
, total
= 0;
2534 printk("%s: ", zone
->name
);
2536 spin_lock_irqsave(&zone
->lock
, flags
);
2537 for (order
= 0; order
< MAX_ORDER
; order
++) {
2538 nr
[order
] = zone
->free_area
[order
].nr_free
;
2539 total
+= nr
[order
] << order
;
2541 spin_unlock_irqrestore(&zone
->lock
, flags
);
2542 for (order
= 0; order
< MAX_ORDER
; order
++)
2543 printk("%lu*%lukB ", nr
[order
], K(1UL) << order
);
2544 printk("= %lukB\n", K(total
));
2547 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES
));
2549 show_swap_cache_info();
2552 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
2554 zoneref
->zone
= zone
;
2555 zoneref
->zone_idx
= zone_idx(zone
);
2559 * Builds allocation fallback zone lists.
2561 * Add all populated zones of a node to the zonelist.
2563 static int build_zonelists_node(pg_data_t
*pgdat
, struct zonelist
*zonelist
,
2564 int nr_zones
, enum zone_type zone_type
)
2568 BUG_ON(zone_type
>= MAX_NR_ZONES
);
2573 zone
= pgdat
->node_zones
+ zone_type
;
2574 if (populated_zone(zone
)) {
2575 zoneref_set_zone(zone
,
2576 &zonelist
->_zonerefs
[nr_zones
++]);
2577 check_highest_zone(zone_type
);
2580 } while (zone_type
);
2587 * 0 = automatic detection of better ordering.
2588 * 1 = order by ([node] distance, -zonetype)
2589 * 2 = order by (-zonetype, [node] distance)
2591 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
2592 * the same zonelist. So only NUMA can configure this param.
2594 #define ZONELIST_ORDER_DEFAULT 0
2595 #define ZONELIST_ORDER_NODE 1
2596 #define ZONELIST_ORDER_ZONE 2
2598 /* zonelist order in the kernel.
2599 * set_zonelist_order() will set this to NODE or ZONE.
2601 static int current_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
2602 static char zonelist_order_name
[3][8] = {"Default", "Node", "Zone"};
2606 /* The value user specified ....changed by config */
2607 static int user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
2608 /* string for sysctl */
2609 #define NUMA_ZONELIST_ORDER_LEN 16
2610 char numa_zonelist_order
[16] = "default";
2613 * interface for configure zonelist ordering.
2614 * command line option "numa_zonelist_order"
2615 * = "[dD]efault - default, automatic configuration.
2616 * = "[nN]ode - order by node locality, then by zone within node
2617 * = "[zZ]one - order by zone, then by locality within zone
2620 static int __parse_numa_zonelist_order(char *s
)
2622 if (*s
== 'd' || *s
== 'D') {
2623 user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
2624 } else if (*s
== 'n' || *s
== 'N') {
2625 user_zonelist_order
= ZONELIST_ORDER_NODE
;
2626 } else if (*s
== 'z' || *s
== 'Z') {
2627 user_zonelist_order
= ZONELIST_ORDER_ZONE
;
2630 "Ignoring invalid numa_zonelist_order value: "
2637 static __init
int setup_numa_zonelist_order(char *s
)
2644 ret
= __parse_numa_zonelist_order(s
);
2646 strlcpy(numa_zonelist_order
, s
, NUMA_ZONELIST_ORDER_LEN
);
2650 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
2653 * sysctl handler for numa_zonelist_order
2655 int numa_zonelist_order_handler(ctl_table
*table
, int write
,
2656 void __user
*buffer
, size_t *length
,
2659 char saved_string
[NUMA_ZONELIST_ORDER_LEN
];
2661 static DEFINE_MUTEX(zl_order_mutex
);
2663 mutex_lock(&zl_order_mutex
);
2665 strcpy(saved_string
, (char*)table
->data
);
2666 ret
= proc_dostring(table
, write
, buffer
, length
, ppos
);
2670 int oldval
= user_zonelist_order
;
2671 if (__parse_numa_zonelist_order((char*)table
->data
)) {
2673 * bogus value. restore saved string
2675 strncpy((char*)table
->data
, saved_string
,
2676 NUMA_ZONELIST_ORDER_LEN
);
2677 user_zonelist_order
= oldval
;
2678 } else if (oldval
!= user_zonelist_order
) {
2679 mutex_lock(&zonelists_mutex
);
2680 build_all_zonelists(NULL
);
2681 mutex_unlock(&zonelists_mutex
);
2685 mutex_unlock(&zl_order_mutex
);
2690 #define MAX_NODE_LOAD (nr_online_nodes)
2691 static int node_load
[MAX_NUMNODES
];
2694 * find_next_best_node - find the next node that should appear in a given node's fallback list
2695 * @node: node whose fallback list we're appending
2696 * @used_node_mask: nodemask_t of already used nodes
2698 * We use a number of factors to determine which is the next node that should
2699 * appear on a given node's fallback list. The node should not have appeared
2700 * already in @node's fallback list, and it should be the next closest node
2701 * according to the distance array (which contains arbitrary distance values
2702 * from each node to each node in the system), and should also prefer nodes
2703 * with no CPUs, since presumably they'll have very little allocation pressure
2704 * on them otherwise.
2705 * It returns -1 if no node is found.
2707 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
2710 int min_val
= INT_MAX
;
2712 const struct cpumask
*tmp
= cpumask_of_node(0);
2714 /* Use the local node if we haven't already */
2715 if (!node_isset(node
, *used_node_mask
)) {
2716 node_set(node
, *used_node_mask
);
2720 for_each_node_state(n
, N_HIGH_MEMORY
) {
2722 /* Don't want a node to appear more than once */
2723 if (node_isset(n
, *used_node_mask
))
2726 /* Use the distance array to find the distance */
2727 val
= node_distance(node
, n
);
2729 /* Penalize nodes under us ("prefer the next node") */
2732 /* Give preference to headless and unused nodes */
2733 tmp
= cpumask_of_node(n
);
2734 if (!cpumask_empty(tmp
))
2735 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
2737 /* Slight preference for less loaded node */
2738 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
2739 val
+= node_load
[n
];
2741 if (val
< min_val
) {
2748 node_set(best_node
, *used_node_mask
);
2755 * Build zonelists ordered by node and zones within node.
2756 * This results in maximum locality--normal zone overflows into local
2757 * DMA zone, if any--but risks exhausting DMA zone.
2759 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int node
)
2762 struct zonelist
*zonelist
;
2764 zonelist
= &pgdat
->node_zonelists
[0];
2765 for (j
= 0; zonelist
->_zonerefs
[j
].zone
!= NULL
; j
++)
2767 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
2769 zonelist
->_zonerefs
[j
].zone
= NULL
;
2770 zonelist
->_zonerefs
[j
].zone_idx
= 0;
2774 * Build gfp_thisnode zonelists
2776 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
2779 struct zonelist
*zonelist
;
2781 zonelist
= &pgdat
->node_zonelists
[1];
2782 j
= build_zonelists_node(pgdat
, zonelist
, 0, MAX_NR_ZONES
- 1);
2783 zonelist
->_zonerefs
[j
].zone
= NULL
;
2784 zonelist
->_zonerefs
[j
].zone_idx
= 0;
2788 * Build zonelists ordered by zone and nodes within zones.
2789 * This results in conserving DMA zone[s] until all Normal memory is
2790 * exhausted, but results in overflowing to remote node while memory
2791 * may still exist in local DMA zone.
2793 static int node_order
[MAX_NUMNODES
];
2795 static void build_zonelists_in_zone_order(pg_data_t
*pgdat
, int nr_nodes
)
2798 int zone_type
; /* needs to be signed */
2800 struct zonelist
*zonelist
;
2802 zonelist
= &pgdat
->node_zonelists
[0];
2804 for (zone_type
= MAX_NR_ZONES
- 1; zone_type
>= 0; zone_type
--) {
2805 for (j
= 0; j
< nr_nodes
; j
++) {
2806 node
= node_order
[j
];
2807 z
= &NODE_DATA(node
)->node_zones
[zone_type
];
2808 if (populated_zone(z
)) {
2810 &zonelist
->_zonerefs
[pos
++]);
2811 check_highest_zone(zone_type
);
2815 zonelist
->_zonerefs
[pos
].zone
= NULL
;
2816 zonelist
->_zonerefs
[pos
].zone_idx
= 0;
2819 static int default_zonelist_order(void)
2822 unsigned long low_kmem_size
,total_size
;
2826 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
2827 * If they are really small and used heavily, the system can fall
2828 * into OOM very easily.
2829 * This function detect ZONE_DMA/DMA32 size and configures zone order.
2831 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
2834 for_each_online_node(nid
) {
2835 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
2836 z
= &NODE_DATA(nid
)->node_zones
[zone_type
];
2837 if (populated_zone(z
)) {
2838 if (zone_type
< ZONE_NORMAL
)
2839 low_kmem_size
+= z
->present_pages
;
2840 total_size
+= z
->present_pages
;
2841 } else if (zone_type
== ZONE_NORMAL
) {
2843 * If any node has only lowmem, then node order
2844 * is preferred to allow kernel allocations
2845 * locally; otherwise, they can easily infringe
2846 * on other nodes when there is an abundance of
2847 * lowmem available to allocate from.
2849 return ZONELIST_ORDER_NODE
;
2853 if (!low_kmem_size
|| /* there are no DMA area. */
2854 low_kmem_size
> total_size
/2) /* DMA/DMA32 is big. */
2855 return ZONELIST_ORDER_NODE
;
2857 * look into each node's config.
2858 * If there is a node whose DMA/DMA32 memory is very big area on
2859 * local memory, NODE_ORDER may be suitable.
2861 average_size
= total_size
/
2862 (nodes_weight(node_states
[N_HIGH_MEMORY
]) + 1);
2863 for_each_online_node(nid
) {
2866 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
2867 z
= &NODE_DATA(nid
)->node_zones
[zone_type
];
2868 if (populated_zone(z
)) {
2869 if (zone_type
< ZONE_NORMAL
)
2870 low_kmem_size
+= z
->present_pages
;
2871 total_size
+= z
->present_pages
;
2874 if (low_kmem_size
&&
2875 total_size
> average_size
&& /* ignore small node */
2876 low_kmem_size
> total_size
* 70/100)
2877 return ZONELIST_ORDER_NODE
;
2879 return ZONELIST_ORDER_ZONE
;
2882 static void set_zonelist_order(void)
2884 if (user_zonelist_order
== ZONELIST_ORDER_DEFAULT
)
2885 current_zonelist_order
= default_zonelist_order();
2887 current_zonelist_order
= user_zonelist_order
;
2890 static void build_zonelists(pg_data_t
*pgdat
)
2894 nodemask_t used_mask
;
2895 int local_node
, prev_node
;
2896 struct zonelist
*zonelist
;
2897 int order
= current_zonelist_order
;
2899 /* initialize zonelists */
2900 for (i
= 0; i
< MAX_ZONELISTS
; i
++) {
2901 zonelist
= pgdat
->node_zonelists
+ i
;
2902 zonelist
->_zonerefs
[0].zone
= NULL
;
2903 zonelist
->_zonerefs
[0].zone_idx
= 0;
2906 /* NUMA-aware ordering of nodes */
2907 local_node
= pgdat
->node_id
;
2908 load
= nr_online_nodes
;
2909 prev_node
= local_node
;
2910 nodes_clear(used_mask
);
2912 memset(node_order
, 0, sizeof(node_order
));
2915 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
2916 int distance
= node_distance(local_node
, node
);
2919 * If another node is sufficiently far away then it is better
2920 * to reclaim pages in a zone before going off node.
2922 if (distance
> RECLAIM_DISTANCE
)
2923 zone_reclaim_mode
= 1;
2926 * We don't want to pressure a particular node.
2927 * So adding penalty to the first node in same
2928 * distance group to make it round-robin.
2930 if (distance
!= node_distance(local_node
, prev_node
))
2931 node_load
[node
] = load
;
2935 if (order
== ZONELIST_ORDER_NODE
)
2936 build_zonelists_in_node_order(pgdat
, node
);
2938 node_order
[j
++] = node
; /* remember order */
2941 if (order
== ZONELIST_ORDER_ZONE
) {
2942 /* calculate node order -- i.e., DMA last! */
2943 build_zonelists_in_zone_order(pgdat
, j
);
2946 build_thisnode_zonelists(pgdat
);
2949 /* Construct the zonelist performance cache - see further mmzone.h */
2950 static void build_zonelist_cache(pg_data_t
*pgdat
)
2952 struct zonelist
*zonelist
;
2953 struct zonelist_cache
*zlc
;
2956 zonelist
= &pgdat
->node_zonelists
[0];
2957 zonelist
->zlcache_ptr
= zlc
= &zonelist
->zlcache
;
2958 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
2959 for (z
= zonelist
->_zonerefs
; z
->zone
; z
++)
2960 zlc
->z_to_n
[z
- zonelist
->_zonerefs
] = zonelist_node_idx(z
);
2963 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
2965 * Return node id of node used for "local" allocations.
2966 * I.e., first node id of first zone in arg node's generic zonelist.
2967 * Used for initializing percpu 'numa_mem', which is used primarily
2968 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
2970 int local_memory_node(int node
)
2974 (void)first_zones_zonelist(node_zonelist(node
, GFP_KERNEL
),
2975 gfp_zone(GFP_KERNEL
),
2982 #else /* CONFIG_NUMA */
2984 static void set_zonelist_order(void)
2986 current_zonelist_order
= ZONELIST_ORDER_ZONE
;
2989 static void build_zonelists(pg_data_t
*pgdat
)
2991 int node
, local_node
;
2993 struct zonelist
*zonelist
;
2995 local_node
= pgdat
->node_id
;
2997 zonelist
= &pgdat
->node_zonelists
[0];
2998 j
= build_zonelists_node(pgdat
, zonelist
, 0, MAX_NR_ZONES
- 1);
3001 * Now we build the zonelist so that it contains the zones
3002 * of all the other nodes.
3003 * We don't want to pressure a particular node, so when
3004 * building the zones for node N, we make sure that the
3005 * zones coming right after the local ones are those from
3006 * node N+1 (modulo N)
3008 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
3009 if (!node_online(node
))
3011 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
3014 for (node
= 0; node
< local_node
; node
++) {
3015 if (!node_online(node
))
3017 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
3021 zonelist
->_zonerefs
[j
].zone
= NULL
;
3022 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3025 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3026 static void build_zonelist_cache(pg_data_t
*pgdat
)
3028 pgdat
->node_zonelists
[0].zlcache_ptr
= NULL
;
3031 #endif /* CONFIG_NUMA */
3034 * Boot pageset table. One per cpu which is going to be used for all
3035 * zones and all nodes. The parameters will be set in such a way
3036 * that an item put on a list will immediately be handed over to
3037 * the buddy list. This is safe since pageset manipulation is done
3038 * with interrupts disabled.
3040 * The boot_pagesets must be kept even after bootup is complete for
3041 * unused processors and/or zones. They do play a role for bootstrapping
3042 * hotplugged processors.
3044 * zoneinfo_show() and maybe other functions do
3045 * not check if the processor is online before following the pageset pointer.
3046 * Other parts of the kernel may not check if the zone is available.
3048 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
);
3049 static DEFINE_PER_CPU(struct per_cpu_pageset
, boot_pageset
);
3050 static void setup_zone_pageset(struct zone
*zone
);
3053 * Global mutex to protect against size modification of zonelists
3054 * as well as to serialize pageset setup for the new populated zone.
3056 DEFINE_MUTEX(zonelists_mutex
);
3058 /* return values int ....just for stop_machine() */
3059 static __init_refok
int __build_all_zonelists(void *data
)
3065 memset(node_load
, 0, sizeof(node_load
));
3067 for_each_online_node(nid
) {
3068 pg_data_t
*pgdat
= NODE_DATA(nid
);
3070 build_zonelists(pgdat
);
3071 build_zonelist_cache(pgdat
);
3075 * Initialize the boot_pagesets that are going to be used
3076 * for bootstrapping processors. The real pagesets for
3077 * each zone will be allocated later when the per cpu
3078 * allocator is available.
3080 * boot_pagesets are used also for bootstrapping offline
3081 * cpus if the system is already booted because the pagesets
3082 * are needed to initialize allocators on a specific cpu too.
3083 * F.e. the percpu allocator needs the page allocator which
3084 * needs the percpu allocator in order to allocate its pagesets
3085 * (a chicken-egg dilemma).
3087 for_each_possible_cpu(cpu
) {
3088 setup_pageset(&per_cpu(boot_pageset
, cpu
), 0);
3090 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3092 * We now know the "local memory node" for each node--
3093 * i.e., the node of the first zone in the generic zonelist.
3094 * Set up numa_mem percpu variable for on-line cpus. During
3095 * boot, only the boot cpu should be on-line; we'll init the
3096 * secondary cpus' numa_mem as they come on-line. During
3097 * node/memory hotplug, we'll fixup all on-line cpus.
3099 if (cpu_online(cpu
))
3100 set_cpu_numa_mem(cpu
, local_memory_node(cpu_to_node(cpu
)));
3108 * Called with zonelists_mutex held always
3109 * unless system_state == SYSTEM_BOOTING.
3111 void build_all_zonelists(void *data
)
3113 set_zonelist_order();
3115 if (system_state
== SYSTEM_BOOTING
) {
3116 __build_all_zonelists(NULL
);
3117 mminit_verify_zonelist();
3118 cpuset_init_current_mems_allowed();
3120 /* we have to stop all cpus to guarantee there is no user
3122 #ifdef CONFIG_MEMORY_HOTPLUG
3124 setup_zone_pageset((struct zone
*)data
);
3126 stop_machine(__build_all_zonelists
, NULL
, NULL
);
3127 /* cpuset refresh routine should be here */
3129 vm_total_pages
= nr_free_pagecache_pages();
3131 * Disable grouping by mobility if the number of pages in the
3132 * system is too low to allow the mechanism to work. It would be
3133 * more accurate, but expensive to check per-zone. This check is
3134 * made on memory-hotadd so a system can start with mobility
3135 * disabled and enable it later
3137 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
3138 page_group_by_mobility_disabled
= 1;
3140 page_group_by_mobility_disabled
= 0;
3142 printk("Built %i zonelists in %s order, mobility grouping %s. "
3143 "Total pages: %ld\n",
3145 zonelist_order_name
[current_zonelist_order
],
3146 page_group_by_mobility_disabled
? "off" : "on",
3149 printk("Policy zone: %s\n", zone_names
[policy_zone
]);
3154 * Helper functions to size the waitqueue hash table.
3155 * Essentially these want to choose hash table sizes sufficiently
3156 * large so that collisions trying to wait on pages are rare.
3157 * But in fact, the number of active page waitqueues on typical
3158 * systems is ridiculously low, less than 200. So this is even
3159 * conservative, even though it seems large.
3161 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3162 * waitqueues, i.e. the size of the waitq table given the number of pages.
3164 #define PAGES_PER_WAITQUEUE 256
3166 #ifndef CONFIG_MEMORY_HOTPLUG
3167 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
3169 unsigned long size
= 1;
3171 pages
/= PAGES_PER_WAITQUEUE
;
3173 while (size
< pages
)
3177 * Once we have dozens or even hundreds of threads sleeping
3178 * on IO we've got bigger problems than wait queue collision.
3179 * Limit the size of the wait table to a reasonable size.
3181 size
= min(size
, 4096UL);
3183 return max(size
, 4UL);
3187 * A zone's size might be changed by hot-add, so it is not possible to determine
3188 * a suitable size for its wait_table. So we use the maximum size now.
3190 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3192 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3193 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3194 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3196 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3197 * or more by the traditional way. (See above). It equals:
3199 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3200 * ia64(16K page size) : = ( 8G + 4M)byte.
3201 * powerpc (64K page size) : = (32G +16M)byte.
3203 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
3210 * This is an integer logarithm so that shifts can be used later
3211 * to extract the more random high bits from the multiplicative
3212 * hash function before the remainder is taken.
3214 static inline unsigned long wait_table_bits(unsigned long size
)
3219 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3222 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3223 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3224 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3225 * higher will lead to a bigger reserve which will get freed as contiguous
3226 * blocks as reclaim kicks in
3228 static void setup_zone_migrate_reserve(struct zone
*zone
)
3230 unsigned long start_pfn
, pfn
, end_pfn
;
3232 unsigned long block_migratetype
;
3235 /* Get the start pfn, end pfn and the number of blocks to reserve */
3236 start_pfn
= zone
->zone_start_pfn
;
3237 end_pfn
= start_pfn
+ zone
->spanned_pages
;
3238 reserve
= roundup(min_wmark_pages(zone
), pageblock_nr_pages
) >>
3242 * Reserve blocks are generally in place to help high-order atomic
3243 * allocations that are short-lived. A min_free_kbytes value that
3244 * would result in more than 2 reserve blocks for atomic allocations
3245 * is assumed to be in place to help anti-fragmentation for the
3246 * future allocation of hugepages at runtime.
3248 reserve
= min(2, reserve
);
3250 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= pageblock_nr_pages
) {
3251 if (!pfn_valid(pfn
))
3253 page
= pfn_to_page(pfn
);
3255 /* Watch out for overlapping nodes */
3256 if (page_to_nid(page
) != zone_to_nid(zone
))
3259 /* Blocks with reserved pages will never free, skip them. */
3260 if (PageReserved(page
))
3263 block_migratetype
= get_pageblock_migratetype(page
);
3265 /* If this block is reserved, account for it */
3266 if (reserve
> 0 && block_migratetype
== MIGRATE_RESERVE
) {
3271 /* Suitable for reserving if this block is movable */
3272 if (reserve
> 0 && block_migratetype
== MIGRATE_MOVABLE
) {
3273 set_pageblock_migratetype(page
, MIGRATE_RESERVE
);
3274 move_freepages_block(zone
, page
, MIGRATE_RESERVE
);
3280 * If the reserve is met and this is a previous reserved block,
3283 if (block_migratetype
== MIGRATE_RESERVE
) {
3284 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
3285 move_freepages_block(zone
, page
, MIGRATE_MOVABLE
);
3291 * Initially all pages are reserved - free ones are freed
3292 * up by free_all_bootmem() once the early boot process is
3293 * done. Non-atomic initialization, single-pass.
3295 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
3296 unsigned long start_pfn
, enum memmap_context context
)
3299 unsigned long end_pfn
= start_pfn
+ size
;
3303 if (highest_memmap_pfn
< end_pfn
- 1)
3304 highest_memmap_pfn
= end_pfn
- 1;
3306 z
= &NODE_DATA(nid
)->node_zones
[zone
];
3307 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
3309 * There can be holes in boot-time mem_map[]s
3310 * handed to this function. They do not
3311 * exist on hotplugged memory.
3313 if (context
== MEMMAP_EARLY
) {
3314 if (!early_pfn_valid(pfn
))
3316 if (!early_pfn_in_nid(pfn
, nid
))
3319 page
= pfn_to_page(pfn
);
3320 set_page_links(page
, zone
, nid
, pfn
);
3321 mminit_verify_page_links(page
, zone
, nid
, pfn
);
3322 init_page_count(page
);
3323 reset_page_mapcount(page
);
3324 SetPageReserved(page
);
3326 * Mark the block movable so that blocks are reserved for
3327 * movable at startup. This will force kernel allocations
3328 * to reserve their blocks rather than leaking throughout
3329 * the address space during boot when many long-lived
3330 * kernel allocations are made. Later some blocks near
3331 * the start are marked MIGRATE_RESERVE by
3332 * setup_zone_migrate_reserve()
3334 * bitmap is created for zone's valid pfn range. but memmap
3335 * can be created for invalid pages (for alignment)
3336 * check here not to call set_pageblock_migratetype() against
3339 if ((z
->zone_start_pfn
<= pfn
)
3340 && (pfn
< z
->zone_start_pfn
+ z
->spanned_pages
)
3341 && !(pfn
& (pageblock_nr_pages
- 1)))
3342 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
3344 INIT_LIST_HEAD(&page
->lru
);
3345 #ifdef WANT_PAGE_VIRTUAL
3346 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3347 if (!is_highmem_idx(zone
))
3348 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
3353 static void __meminit
zone_init_free_lists(struct zone
*zone
)
3356 for_each_migratetype_order(order
, t
) {
3357 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
3358 zone
->free_area
[order
].nr_free
= 0;
3362 #ifndef __HAVE_ARCH_MEMMAP_INIT
3363 #define memmap_init(size, nid, zone, start_pfn) \
3364 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3367 static int zone_batchsize(struct zone
*zone
)
3373 * The per-cpu-pages pools are set to around 1000th of the
3374 * size of the zone. But no more than 1/2 of a meg.
3376 * OK, so we don't know how big the cache is. So guess.
3378 batch
= zone
->present_pages
/ 1024;
3379 if (batch
* PAGE_SIZE
> 512 * 1024)
3380 batch
= (512 * 1024) / PAGE_SIZE
;
3381 batch
/= 4; /* We effectively *= 4 below */
3386 * Clamp the batch to a 2^n - 1 value. Having a power
3387 * of 2 value was found to be more likely to have
3388 * suboptimal cache aliasing properties in some cases.
3390 * For example if 2 tasks are alternately allocating
3391 * batches of pages, one task can end up with a lot
3392 * of pages of one half of the possible page colors
3393 * and the other with pages of the other colors.
3395 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
3400 /* The deferral and batching of frees should be suppressed under NOMMU
3403 * The problem is that NOMMU needs to be able to allocate large chunks
3404 * of contiguous memory as there's no hardware page translation to
3405 * assemble apparent contiguous memory from discontiguous pages.
3407 * Queueing large contiguous runs of pages for batching, however,
3408 * causes the pages to actually be freed in smaller chunks. As there
3409 * can be a significant delay between the individual batches being
3410 * recycled, this leads to the once large chunks of space being
3411 * fragmented and becoming unavailable for high-order allocations.
3417 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
3419 struct per_cpu_pages
*pcp
;
3422 memset(p
, 0, sizeof(*p
));
3426 pcp
->high
= 6 * batch
;
3427 pcp
->batch
= max(1UL, 1 * batch
);
3428 for (migratetype
= 0; migratetype
< MIGRATE_PCPTYPES
; migratetype
++)
3429 INIT_LIST_HEAD(&pcp
->lists
[migratetype
]);
3433 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3434 * to the value high for the pageset p.
3437 static void setup_pagelist_highmark(struct per_cpu_pageset
*p
,
3440 struct per_cpu_pages
*pcp
;
3444 pcp
->batch
= max(1UL, high
/4);
3445 if ((high
/4) > (PAGE_SHIFT
* 8))
3446 pcp
->batch
= PAGE_SHIFT
* 8;
3449 static __meminit
void setup_zone_pageset(struct zone
*zone
)
3453 zone
->pageset
= alloc_percpu(struct per_cpu_pageset
);
3455 for_each_possible_cpu(cpu
) {
3456 struct per_cpu_pageset
*pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
3458 setup_pageset(pcp
, zone_batchsize(zone
));
3460 if (percpu_pagelist_fraction
)
3461 setup_pagelist_highmark(pcp
,
3462 (zone
->present_pages
/
3463 percpu_pagelist_fraction
));
3468 * Allocate per cpu pagesets and initialize them.
3469 * Before this call only boot pagesets were available.
3471 void __init
setup_per_cpu_pageset(void)
3475 for_each_populated_zone(zone
)
3476 setup_zone_pageset(zone
);
3479 static noinline __init_refok
3480 int zone_wait_table_init(struct zone
*zone
, unsigned long zone_size_pages
)
3483 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
3487 * The per-page waitqueue mechanism uses hashed waitqueues
3490 zone
->wait_table_hash_nr_entries
=
3491 wait_table_hash_nr_entries(zone_size_pages
);
3492 zone
->wait_table_bits
=
3493 wait_table_bits(zone
->wait_table_hash_nr_entries
);
3494 alloc_size
= zone
->wait_table_hash_nr_entries
3495 * sizeof(wait_queue_head_t
);
3497 if (!slab_is_available()) {
3498 zone
->wait_table
= (wait_queue_head_t
*)
3499 alloc_bootmem_node(pgdat
, alloc_size
);
3502 * This case means that a zone whose size was 0 gets new memory
3503 * via memory hot-add.
3504 * But it may be the case that a new node was hot-added. In
3505 * this case vmalloc() will not be able to use this new node's
3506 * memory - this wait_table must be initialized to use this new
3507 * node itself as well.
3508 * To use this new node's memory, further consideration will be
3511 zone
->wait_table
= vmalloc(alloc_size
);
3513 if (!zone
->wait_table
)
3516 for(i
= 0; i
< zone
->wait_table_hash_nr_entries
; ++i
)
3517 init_waitqueue_head(zone
->wait_table
+ i
);
3522 static int __zone_pcp_update(void *data
)
3524 struct zone
*zone
= data
;
3526 unsigned long batch
= zone_batchsize(zone
), flags
;
3528 for_each_possible_cpu(cpu
) {
3529 struct per_cpu_pageset
*pset
;
3530 struct per_cpu_pages
*pcp
;
3532 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
3535 local_irq_save(flags
);
3536 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
3537 setup_pageset(pset
, batch
);
3538 local_irq_restore(flags
);
3543 void zone_pcp_update(struct zone
*zone
)
3545 stop_machine(__zone_pcp_update
, zone
, NULL
);
3548 static __meminit
void zone_pcp_init(struct zone
*zone
)
3551 * per cpu subsystem is not up at this point. The following code
3552 * relies on the ability of the linker to provide the
3553 * offset of a (static) per cpu variable into the per cpu area.
3555 zone
->pageset
= &boot_pageset
;
3557 if (zone
->present_pages
)
3558 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%u\n",
3559 zone
->name
, zone
->present_pages
,
3560 zone_batchsize(zone
));
3563 __meminit
int init_currently_empty_zone(struct zone
*zone
,
3564 unsigned long zone_start_pfn
,
3566 enum memmap_context context
)
3568 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
3570 ret
= zone_wait_table_init(zone
, size
);
3573 pgdat
->nr_zones
= zone_idx(zone
) + 1;
3575 zone
->zone_start_pfn
= zone_start_pfn
;
3577 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
3578 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
3580 (unsigned long)zone_idx(zone
),
3581 zone_start_pfn
, (zone_start_pfn
+ size
));
3583 zone_init_free_lists(zone
);
3588 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3590 * Basic iterator support. Return the first range of PFNs for a node
3591 * Note: nid == MAX_NUMNODES returns first region regardless of node
3593 static int __meminit
first_active_region_index_in_nid(int nid
)
3597 for (i
= 0; i
< nr_nodemap_entries
; i
++)
3598 if (nid
== MAX_NUMNODES
|| early_node_map
[i
].nid
== nid
)
3605 * Basic iterator support. Return the next active range of PFNs for a node
3606 * Note: nid == MAX_NUMNODES returns next region regardless of node
3608 static int __meminit
next_active_region_index_in_nid(int index
, int nid
)
3610 for (index
= index
+ 1; index
< nr_nodemap_entries
; index
++)
3611 if (nid
== MAX_NUMNODES
|| early_node_map
[index
].nid
== nid
)
3617 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
3619 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
3620 * Architectures may implement their own version but if add_active_range()
3621 * was used and there are no special requirements, this is a convenient
3624 int __meminit
__early_pfn_to_nid(unsigned long pfn
)
3628 for (i
= 0; i
< nr_nodemap_entries
; i
++) {
3629 unsigned long start_pfn
= early_node_map
[i
].start_pfn
;
3630 unsigned long end_pfn
= early_node_map
[i
].end_pfn
;
3632 if (start_pfn
<= pfn
&& pfn
< end_pfn
)
3633 return early_node_map
[i
].nid
;
3635 /* This is a memory hole */
3638 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
3640 int __meminit
early_pfn_to_nid(unsigned long pfn
)
3644 nid
= __early_pfn_to_nid(pfn
);
3647 /* just returns 0 */
3651 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
3652 bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
3656 nid
= __early_pfn_to_nid(pfn
);
3657 if (nid
>= 0 && nid
!= node
)
3663 /* Basic iterator support to walk early_node_map[] */
3664 #define for_each_active_range_index_in_nid(i, nid) \
3665 for (i = first_active_region_index_in_nid(nid); i != -1; \
3666 i = next_active_region_index_in_nid(i, nid))
3669 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
3670 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
3671 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
3673 * If an architecture guarantees that all ranges registered with
3674 * add_active_ranges() contain no holes and may be freed, this
3675 * this function may be used instead of calling free_bootmem() manually.
3677 void __init
free_bootmem_with_active_regions(int nid
,
3678 unsigned long max_low_pfn
)
3682 for_each_active_range_index_in_nid(i
, nid
) {
3683 unsigned long size_pages
= 0;
3684 unsigned long end_pfn
= early_node_map
[i
].end_pfn
;
3686 if (early_node_map
[i
].start_pfn
>= max_low_pfn
)
3689 if (end_pfn
> max_low_pfn
)
3690 end_pfn
= max_low_pfn
;
3692 size_pages
= end_pfn
- early_node_map
[i
].start_pfn
;
3693 free_bootmem_node(NODE_DATA(early_node_map
[i
].nid
),
3694 PFN_PHYS(early_node_map
[i
].start_pfn
),
3695 size_pages
<< PAGE_SHIFT
);
3699 #ifdef CONFIG_HAVE_MEMBLOCK
3700 u64 __init
find_memory_core_early(int nid
, u64 size
, u64 align
,
3701 u64 goal
, u64 limit
)
3705 /* Need to go over early_node_map to find out good range for node */
3706 for_each_active_range_index_in_nid(i
, nid
) {
3708 u64 ei_start
, ei_last
;
3709 u64 final_start
, final_end
;
3711 ei_last
= early_node_map
[i
].end_pfn
;
3712 ei_last
<<= PAGE_SHIFT
;
3713 ei_start
= early_node_map
[i
].start_pfn
;
3714 ei_start
<<= PAGE_SHIFT
;
3716 final_start
= max(ei_start
, goal
);
3717 final_end
= min(ei_last
, limit
);
3719 if (final_start
>= final_end
)
3722 addr
= memblock_find_in_range(final_start
, final_end
, size
, align
);
3724 if (addr
== MEMBLOCK_ERROR
)
3730 return MEMBLOCK_ERROR
;
3734 int __init
add_from_early_node_map(struct range
*range
, int az
,
3735 int nr_range
, int nid
)
3740 /* need to go over early_node_map to find out good range for node */
3741 for_each_active_range_index_in_nid(i
, nid
) {
3742 start
= early_node_map
[i
].start_pfn
;
3743 end
= early_node_map
[i
].end_pfn
;
3744 nr_range
= add_range(range
, az
, nr_range
, start
, end
);
3749 #ifdef CONFIG_NO_BOOTMEM
3750 void * __init
__alloc_memory_core_early(int nid
, u64 size
, u64 align
,
3751 u64 goal
, u64 limit
)
3756 if (limit
> memblock
.current_limit
)
3757 limit
= memblock
.current_limit
;
3759 addr
= find_memory_core_early(nid
, size
, align
, goal
, limit
);
3761 if (addr
== MEMBLOCK_ERROR
)
3764 ptr
= phys_to_virt(addr
);
3765 memset(ptr
, 0, size
);
3766 memblock_x86_reserve_range(addr
, addr
+ size
, "BOOTMEM");
3768 * The min_count is set to 0 so that bootmem allocated blocks
3769 * are never reported as leaks.
3771 kmemleak_alloc(ptr
, size
, 0, 0);
3777 void __init
work_with_active_regions(int nid
, work_fn_t work_fn
, void *data
)
3782 for_each_active_range_index_in_nid(i
, nid
) {
3783 ret
= work_fn(early_node_map
[i
].start_pfn
,
3784 early_node_map
[i
].end_pfn
, data
);
3790 * sparse_memory_present_with_active_regions - Call memory_present for each active range
3791 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
3793 * If an architecture guarantees that all ranges registered with
3794 * add_active_ranges() contain no holes and may be freed, this
3795 * function may be used instead of calling memory_present() manually.
3797 void __init
sparse_memory_present_with_active_regions(int nid
)
3801 for_each_active_range_index_in_nid(i
, nid
)
3802 memory_present(early_node_map
[i
].nid
,
3803 early_node_map
[i
].start_pfn
,
3804 early_node_map
[i
].end_pfn
);
3808 * get_pfn_range_for_nid - Return the start and end page frames for a node
3809 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
3810 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
3811 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
3813 * It returns the start and end page frame of a node based on information
3814 * provided by an arch calling add_active_range(). If called for a node
3815 * with no available memory, a warning is printed and the start and end
3818 void __meminit
get_pfn_range_for_nid(unsigned int nid
,
3819 unsigned long *start_pfn
, unsigned long *end_pfn
)
3825 for_each_active_range_index_in_nid(i
, nid
) {
3826 *start_pfn
= min(*start_pfn
, early_node_map
[i
].start_pfn
);
3827 *end_pfn
= max(*end_pfn
, early_node_map
[i
].end_pfn
);
3830 if (*start_pfn
== -1UL)
3835 * This finds a zone that can be used for ZONE_MOVABLE pages. The
3836 * assumption is made that zones within a node are ordered in monotonic
3837 * increasing memory addresses so that the "highest" populated zone is used
3839 static void __init
find_usable_zone_for_movable(void)
3842 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
3843 if (zone_index
== ZONE_MOVABLE
)
3846 if (arch_zone_highest_possible_pfn
[zone_index
] >
3847 arch_zone_lowest_possible_pfn
[zone_index
])
3851 VM_BUG_ON(zone_index
== -1);
3852 movable_zone
= zone_index
;
3856 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
3857 * because it is sized independant of architecture. Unlike the other zones,
3858 * the starting point for ZONE_MOVABLE is not fixed. It may be different
3859 * in each node depending on the size of each node and how evenly kernelcore
3860 * is distributed. This helper function adjusts the zone ranges
3861 * provided by the architecture for a given node by using the end of the
3862 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
3863 * zones within a node are in order of monotonic increases memory addresses
3865 static void __meminit
adjust_zone_range_for_zone_movable(int nid
,
3866 unsigned long zone_type
,
3867 unsigned long node_start_pfn
,
3868 unsigned long node_end_pfn
,
3869 unsigned long *zone_start_pfn
,
3870 unsigned long *zone_end_pfn
)
3872 /* Only adjust if ZONE_MOVABLE is on this node */
3873 if (zone_movable_pfn
[nid
]) {
3874 /* Size ZONE_MOVABLE */
3875 if (zone_type
== ZONE_MOVABLE
) {
3876 *zone_start_pfn
= zone_movable_pfn
[nid
];
3877 *zone_end_pfn
= min(node_end_pfn
,
3878 arch_zone_highest_possible_pfn
[movable_zone
]);
3880 /* Adjust for ZONE_MOVABLE starting within this range */
3881 } else if (*zone_start_pfn
< zone_movable_pfn
[nid
] &&
3882 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
3883 *zone_end_pfn
= zone_movable_pfn
[nid
];
3885 /* Check if this whole range is within ZONE_MOVABLE */
3886 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
3887 *zone_start_pfn
= *zone_end_pfn
;
3892 * Return the number of pages a zone spans in a node, including holes
3893 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
3895 static unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
3896 unsigned long zone_type
,
3897 unsigned long *ignored
)
3899 unsigned long node_start_pfn
, node_end_pfn
;
3900 unsigned long zone_start_pfn
, zone_end_pfn
;
3902 /* Get the start and end of the node and zone */
3903 get_pfn_range_for_nid(nid
, &node_start_pfn
, &node_end_pfn
);
3904 zone_start_pfn
= arch_zone_lowest_possible_pfn
[zone_type
];
3905 zone_end_pfn
= arch_zone_highest_possible_pfn
[zone_type
];
3906 adjust_zone_range_for_zone_movable(nid
, zone_type
,
3907 node_start_pfn
, node_end_pfn
,
3908 &zone_start_pfn
, &zone_end_pfn
);
3910 /* Check that this node has pages within the zone's required range */
3911 if (zone_end_pfn
< node_start_pfn
|| zone_start_pfn
> node_end_pfn
)
3914 /* Move the zone boundaries inside the node if necessary */
3915 zone_end_pfn
= min(zone_end_pfn
, node_end_pfn
);
3916 zone_start_pfn
= max(zone_start_pfn
, node_start_pfn
);
3918 /* Return the spanned pages */
3919 return zone_end_pfn
- zone_start_pfn
;
3923 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
3924 * then all holes in the requested range will be accounted for.
3926 unsigned long __meminit
__absent_pages_in_range(int nid
,
3927 unsigned long range_start_pfn
,
3928 unsigned long range_end_pfn
)
3931 unsigned long prev_end_pfn
= 0, hole_pages
= 0;
3932 unsigned long start_pfn
;
3934 /* Find the end_pfn of the first active range of pfns in the node */
3935 i
= first_active_region_index_in_nid(nid
);
3939 prev_end_pfn
= min(early_node_map
[i
].start_pfn
, range_end_pfn
);
3941 /* Account for ranges before physical memory on this node */
3942 if (early_node_map
[i
].start_pfn
> range_start_pfn
)
3943 hole_pages
= prev_end_pfn
- range_start_pfn
;
3945 /* Find all holes for the zone within the node */
3946 for (; i
!= -1; i
= next_active_region_index_in_nid(i
, nid
)) {
3948 /* No need to continue if prev_end_pfn is outside the zone */
3949 if (prev_end_pfn
>= range_end_pfn
)
3952 /* Make sure the end of the zone is not within the hole */
3953 start_pfn
= min(early_node_map
[i
].start_pfn
, range_end_pfn
);
3954 prev_end_pfn
= max(prev_end_pfn
, range_start_pfn
);
3956 /* Update the hole size cound and move on */
3957 if (start_pfn
> range_start_pfn
) {
3958 BUG_ON(prev_end_pfn
> start_pfn
);
3959 hole_pages
+= start_pfn
- prev_end_pfn
;
3961 prev_end_pfn
= early_node_map
[i
].end_pfn
;
3964 /* Account for ranges past physical memory on this node */
3965 if (range_end_pfn
> prev_end_pfn
)
3966 hole_pages
+= range_end_pfn
-
3967 max(range_start_pfn
, prev_end_pfn
);
3973 * absent_pages_in_range - Return number of page frames in holes within a range
3974 * @start_pfn: The start PFN to start searching for holes
3975 * @end_pfn: The end PFN to stop searching for holes
3977 * It returns the number of pages frames in memory holes within a range.
3979 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
3980 unsigned long end_pfn
)
3982 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
3985 /* Return the number of page frames in holes in a zone on a node */
3986 static unsigned long __meminit
zone_absent_pages_in_node(int nid
,
3987 unsigned long zone_type
,
3988 unsigned long *ignored
)
3990 unsigned long node_start_pfn
, node_end_pfn
;
3991 unsigned long zone_start_pfn
, zone_end_pfn
;
3993 get_pfn_range_for_nid(nid
, &node_start_pfn
, &node_end_pfn
);
3994 zone_start_pfn
= max(arch_zone_lowest_possible_pfn
[zone_type
],
3996 zone_end_pfn
= min(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
);
4002 return __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
4006 static inline unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
4007 unsigned long zone_type
,
4008 unsigned long *zones_size
)
4010 return zones_size
[zone_type
];
4013 static inline unsigned long __meminit
zone_absent_pages_in_node(int nid
,
4014 unsigned long zone_type
,
4015 unsigned long *zholes_size
)
4020 return zholes_size
[zone_type
];
4025 static void __meminit
calculate_node_totalpages(struct pglist_data
*pgdat
,
4026 unsigned long *zones_size
, unsigned long *zholes_size
)
4028 unsigned long realtotalpages
, totalpages
= 0;
4031 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4032 totalpages
+= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
4034 pgdat
->node_spanned_pages
= totalpages
;
4036 realtotalpages
= totalpages
;
4037 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4039 zone_absent_pages_in_node(pgdat
->node_id
, i
,
4041 pgdat
->node_present_pages
= realtotalpages
;
4042 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
4046 #ifndef CONFIG_SPARSEMEM
4048 * Calculate the size of the zone->blockflags rounded to an unsigned long
4049 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4050 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4051 * round what is now in bits to nearest long in bits, then return it in
4054 static unsigned long __init
usemap_size(unsigned long zonesize
)
4056 unsigned long usemapsize
;
4058 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
4059 usemapsize
= usemapsize
>> pageblock_order
;
4060 usemapsize
*= NR_PAGEBLOCK_BITS
;
4061 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
4063 return usemapsize
/ 8;
4066 static void __init
setup_usemap(struct pglist_data
*pgdat
,
4067 struct zone
*zone
, unsigned long zonesize
)
4069 unsigned long usemapsize
= usemap_size(zonesize
);
4070 zone
->pageblock_flags
= NULL
;
4072 zone
->pageblock_flags
= alloc_bootmem_node(pgdat
, usemapsize
);
4075 static inline void setup_usemap(struct pglist_data
*pgdat
,
4076 struct zone
*zone
, unsigned long zonesize
) {}
4077 #endif /* CONFIG_SPARSEMEM */
4079 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4081 /* Return a sensible default order for the pageblock size. */
4082 static inline int pageblock_default_order(void)
4084 if (HPAGE_SHIFT
> PAGE_SHIFT
)
4085 return HUGETLB_PAGE_ORDER
;
4090 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4091 static inline void __init
set_pageblock_order(unsigned int order
)
4093 /* Check that pageblock_nr_pages has not already been setup */
4094 if (pageblock_order
)
4098 * Assume the largest contiguous order of interest is a huge page.
4099 * This value may be variable depending on boot parameters on IA64
4101 pageblock_order
= order
;
4103 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4106 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4107 * and pageblock_default_order() are unused as pageblock_order is set
4108 * at compile-time. See include/linux/pageblock-flags.h for the values of
4109 * pageblock_order based on the kernel config
4111 static inline int pageblock_default_order(unsigned int order
)
4115 #define set_pageblock_order(x) do {} while (0)
4117 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4120 * Set up the zone data structures:
4121 * - mark all pages reserved
4122 * - mark all memory queues empty
4123 * - clear the memory bitmaps
4125 static void __paginginit
free_area_init_core(struct pglist_data
*pgdat
,
4126 unsigned long *zones_size
, unsigned long *zholes_size
)
4129 int nid
= pgdat
->node_id
;
4130 unsigned long zone_start_pfn
= pgdat
->node_start_pfn
;
4133 pgdat_resize_init(pgdat
);
4134 pgdat
->nr_zones
= 0;
4135 init_waitqueue_head(&pgdat
->kswapd_wait
);
4136 pgdat
->kswapd_max_order
= 0;
4137 pgdat_page_cgroup_init(pgdat
);
4139 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
4140 struct zone
*zone
= pgdat
->node_zones
+ j
;
4141 unsigned long size
, realsize
, memmap_pages
;
4144 size
= zone_spanned_pages_in_node(nid
, j
, zones_size
);
4145 realsize
= size
- zone_absent_pages_in_node(nid
, j
,
4149 * Adjust realsize so that it accounts for how much memory
4150 * is used by this zone for memmap. This affects the watermark
4151 * and per-cpu initialisations
4154 PAGE_ALIGN(size
* sizeof(struct page
)) >> PAGE_SHIFT
;
4155 if (realsize
>= memmap_pages
) {
4156 realsize
-= memmap_pages
;
4159 " %s zone: %lu pages used for memmap\n",
4160 zone_names
[j
], memmap_pages
);
4163 " %s zone: %lu pages exceeds realsize %lu\n",
4164 zone_names
[j
], memmap_pages
, realsize
);
4166 /* Account for reserved pages */
4167 if (j
== 0 && realsize
> dma_reserve
) {
4168 realsize
-= dma_reserve
;
4169 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
4170 zone_names
[0], dma_reserve
);
4173 if (!is_highmem_idx(j
))
4174 nr_kernel_pages
+= realsize
;
4175 nr_all_pages
+= realsize
;
4177 zone
->spanned_pages
= size
;
4178 zone
->present_pages
= realsize
;
4181 zone
->min_unmapped_pages
= (realsize
*sysctl_min_unmapped_ratio
)
4183 zone
->min_slab_pages
= (realsize
* sysctl_min_slab_ratio
) / 100;
4185 zone
->name
= zone_names
[j
];
4186 spin_lock_init(&zone
->lock
);
4187 spin_lock_init(&zone
->lru_lock
);
4188 zone_seqlock_init(zone
);
4189 zone
->zone_pgdat
= pgdat
;
4191 zone_pcp_init(zone
);
4193 INIT_LIST_HEAD(&zone
->lru
[l
].list
);
4194 zone
->reclaim_stat
.nr_saved_scan
[l
] = 0;
4196 zone
->reclaim_stat
.recent_rotated
[0] = 0;
4197 zone
->reclaim_stat
.recent_rotated
[1] = 0;
4198 zone
->reclaim_stat
.recent_scanned
[0] = 0;
4199 zone
->reclaim_stat
.recent_scanned
[1] = 0;
4200 zap_zone_vm_stats(zone
);
4205 set_pageblock_order(pageblock_default_order());
4206 setup_usemap(pgdat
, zone
, size
);
4207 ret
= init_currently_empty_zone(zone
, zone_start_pfn
,
4208 size
, MEMMAP_EARLY
);
4210 memmap_init(size
, nid
, j
, zone_start_pfn
);
4211 zone_start_pfn
+= size
;
4215 static void __init_refok
alloc_node_mem_map(struct pglist_data
*pgdat
)
4217 /* Skip empty nodes */
4218 if (!pgdat
->node_spanned_pages
)
4221 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4222 /* ia64 gets its own node_mem_map, before this, without bootmem */
4223 if (!pgdat
->node_mem_map
) {
4224 unsigned long size
, start
, end
;
4228 * The zone's endpoints aren't required to be MAX_ORDER
4229 * aligned but the node_mem_map endpoints must be in order
4230 * for the buddy allocator to function correctly.
4232 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
4233 end
= pgdat
->node_start_pfn
+ pgdat
->node_spanned_pages
;
4234 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
4235 size
= (end
- start
) * sizeof(struct page
);
4236 map
= alloc_remap(pgdat
->node_id
, size
);
4238 map
= alloc_bootmem_node(pgdat
, size
);
4239 pgdat
->node_mem_map
= map
+ (pgdat
->node_start_pfn
- start
);
4241 #ifndef CONFIG_NEED_MULTIPLE_NODES
4243 * With no DISCONTIG, the global mem_map is just set as node 0's
4245 if (pgdat
== NODE_DATA(0)) {
4246 mem_map
= NODE_DATA(0)->node_mem_map
;
4247 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
4248 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
4249 mem_map
-= (pgdat
->node_start_pfn
- ARCH_PFN_OFFSET
);
4250 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4253 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4256 void __paginginit
free_area_init_node(int nid
, unsigned long *zones_size
,
4257 unsigned long node_start_pfn
, unsigned long *zholes_size
)
4259 pg_data_t
*pgdat
= NODE_DATA(nid
);
4261 pgdat
->node_id
= nid
;
4262 pgdat
->node_start_pfn
= node_start_pfn
;
4263 calculate_node_totalpages(pgdat
, zones_size
, zholes_size
);
4265 alloc_node_mem_map(pgdat
);
4266 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4267 printk(KERN_DEBUG
"free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4268 nid
, (unsigned long)pgdat
,
4269 (unsigned long)pgdat
->node_mem_map
);
4272 free_area_init_core(pgdat
, zones_size
, zholes_size
);
4275 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
4277 #if MAX_NUMNODES > 1
4279 * Figure out the number of possible node ids.
4281 static void __init
setup_nr_node_ids(void)
4284 unsigned int highest
= 0;
4286 for_each_node_mask(node
, node_possible_map
)
4288 nr_node_ids
= highest
+ 1;
4291 static inline void setup_nr_node_ids(void)
4297 * add_active_range - Register a range of PFNs backed by physical memory
4298 * @nid: The node ID the range resides on
4299 * @start_pfn: The start PFN of the available physical memory
4300 * @end_pfn: The end PFN of the available physical memory
4302 * These ranges are stored in an early_node_map[] and later used by
4303 * free_area_init_nodes() to calculate zone sizes and holes. If the
4304 * range spans a memory hole, it is up to the architecture to ensure
4305 * the memory is not freed by the bootmem allocator. If possible
4306 * the range being registered will be merged with existing ranges.
4308 void __init
add_active_range(unsigned int nid
, unsigned long start_pfn
,
4309 unsigned long end_pfn
)
4313 mminit_dprintk(MMINIT_TRACE
, "memory_register",
4314 "Entering add_active_range(%d, %#lx, %#lx) "
4315 "%d entries of %d used\n",
4316 nid
, start_pfn
, end_pfn
,
4317 nr_nodemap_entries
, MAX_ACTIVE_REGIONS
);
4319 mminit_validate_memmodel_limits(&start_pfn
, &end_pfn
);
4321 /* Merge with existing active regions if possible */
4322 for (i
= 0; i
< nr_nodemap_entries
; i
++) {
4323 if (early_node_map
[i
].nid
!= nid
)
4326 /* Skip if an existing region covers this new one */
4327 if (start_pfn
>= early_node_map
[i
].start_pfn
&&
4328 end_pfn
<= early_node_map
[i
].end_pfn
)
4331 /* Merge forward if suitable */
4332 if (start_pfn
<= early_node_map
[i
].end_pfn
&&
4333 end_pfn
> early_node_map
[i
].end_pfn
) {
4334 early_node_map
[i
].end_pfn
= end_pfn
;
4338 /* Merge backward if suitable */
4339 if (start_pfn
< early_node_map
[i
].start_pfn
&&
4340 end_pfn
>= early_node_map
[i
].start_pfn
) {
4341 early_node_map
[i
].start_pfn
= start_pfn
;
4346 /* Check that early_node_map is large enough */
4347 if (i
>= MAX_ACTIVE_REGIONS
) {
4348 printk(KERN_CRIT
"More than %d memory regions, truncating\n",
4349 MAX_ACTIVE_REGIONS
);
4353 early_node_map
[i
].nid
= nid
;
4354 early_node_map
[i
].start_pfn
= start_pfn
;
4355 early_node_map
[i
].end_pfn
= end_pfn
;
4356 nr_nodemap_entries
= i
+ 1;
4360 * remove_active_range - Shrink an existing registered range of PFNs
4361 * @nid: The node id the range is on that should be shrunk
4362 * @start_pfn: The new PFN of the range
4363 * @end_pfn: The new PFN of the range
4365 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
4366 * The map is kept near the end physical page range that has already been
4367 * registered. This function allows an arch to shrink an existing registered
4370 void __init
remove_active_range(unsigned int nid
, unsigned long start_pfn
,
4371 unsigned long end_pfn
)
4376 printk(KERN_DEBUG
"remove_active_range (%d, %lu, %lu)\n",
4377 nid
, start_pfn
, end_pfn
);
4379 /* Find the old active region end and shrink */
4380 for_each_active_range_index_in_nid(i
, nid
) {
4381 if (early_node_map
[i
].start_pfn
>= start_pfn
&&
4382 early_node_map
[i
].end_pfn
<= end_pfn
) {
4384 early_node_map
[i
].start_pfn
= 0;
4385 early_node_map
[i
].end_pfn
= 0;
4389 if (early_node_map
[i
].start_pfn
< start_pfn
&&
4390 early_node_map
[i
].end_pfn
> start_pfn
) {
4391 unsigned long temp_end_pfn
= early_node_map
[i
].end_pfn
;
4392 early_node_map
[i
].end_pfn
= start_pfn
;
4393 if (temp_end_pfn
> end_pfn
)
4394 add_active_range(nid
, end_pfn
, temp_end_pfn
);
4397 if (early_node_map
[i
].start_pfn
>= start_pfn
&&
4398 early_node_map
[i
].end_pfn
> end_pfn
&&
4399 early_node_map
[i
].start_pfn
< end_pfn
) {
4400 early_node_map
[i
].start_pfn
= end_pfn
;
4408 /* remove the blank ones */
4409 for (i
= nr_nodemap_entries
- 1; i
> 0; i
--) {
4410 if (early_node_map
[i
].nid
!= nid
)
4412 if (early_node_map
[i
].end_pfn
)
4414 /* we found it, get rid of it */
4415 for (j
= i
; j
< nr_nodemap_entries
- 1; j
++)
4416 memcpy(&early_node_map
[j
], &early_node_map
[j
+1],
4417 sizeof(early_node_map
[j
]));
4418 j
= nr_nodemap_entries
- 1;
4419 memset(&early_node_map
[j
], 0, sizeof(early_node_map
[j
]));
4420 nr_nodemap_entries
--;
4425 * remove_all_active_ranges - Remove all currently registered regions
4427 * During discovery, it may be found that a table like SRAT is invalid
4428 * and an alternative discovery method must be used. This function removes
4429 * all currently registered regions.
4431 void __init
remove_all_active_ranges(void)
4433 memset(early_node_map
, 0, sizeof(early_node_map
));
4434 nr_nodemap_entries
= 0;
4437 /* Compare two active node_active_regions */
4438 static int __init
cmp_node_active_region(const void *a
, const void *b
)
4440 struct node_active_region
*arange
= (struct node_active_region
*)a
;
4441 struct node_active_region
*brange
= (struct node_active_region
*)b
;
4443 /* Done this way to avoid overflows */
4444 if (arange
->start_pfn
> brange
->start_pfn
)
4446 if (arange
->start_pfn
< brange
->start_pfn
)
4452 /* sort the node_map by start_pfn */
4453 void __init
sort_node_map(void)
4455 sort(early_node_map
, (size_t)nr_nodemap_entries
,
4456 sizeof(struct node_active_region
),
4457 cmp_node_active_region
, NULL
);
4460 /* Find the lowest pfn for a node */
4461 static unsigned long __init
find_min_pfn_for_node(int nid
)
4464 unsigned long min_pfn
= ULONG_MAX
;
4466 /* Assuming a sorted map, the first range found has the starting pfn */
4467 for_each_active_range_index_in_nid(i
, nid
)
4468 min_pfn
= min(min_pfn
, early_node_map
[i
].start_pfn
);
4470 if (min_pfn
== ULONG_MAX
) {
4472 "Could not find start_pfn for node %d\n", nid
);
4480 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4482 * It returns the minimum PFN based on information provided via
4483 * add_active_range().
4485 unsigned long __init
find_min_pfn_with_active_regions(void)
4487 return find_min_pfn_for_node(MAX_NUMNODES
);
4491 * early_calculate_totalpages()
4492 * Sum pages in active regions for movable zone.
4493 * Populate N_HIGH_MEMORY for calculating usable_nodes.
4495 static unsigned long __init
early_calculate_totalpages(void)
4498 unsigned long totalpages
= 0;
4500 for (i
= 0; i
< nr_nodemap_entries
; i
++) {
4501 unsigned long pages
= early_node_map
[i
].end_pfn
-
4502 early_node_map
[i
].start_pfn
;
4503 totalpages
+= pages
;
4505 node_set_state(early_node_map
[i
].nid
, N_HIGH_MEMORY
);
4511 * Find the PFN the Movable zone begins in each node. Kernel memory
4512 * is spread evenly between nodes as long as the nodes have enough
4513 * memory. When they don't, some nodes will have more kernelcore than
4516 static void __init
find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn
)
4519 unsigned long usable_startpfn
;
4520 unsigned long kernelcore_node
, kernelcore_remaining
;
4521 /* save the state before borrow the nodemask */
4522 nodemask_t saved_node_state
= node_states
[N_HIGH_MEMORY
];
4523 unsigned long totalpages
= early_calculate_totalpages();
4524 int usable_nodes
= nodes_weight(node_states
[N_HIGH_MEMORY
]);
4527 * If movablecore was specified, calculate what size of
4528 * kernelcore that corresponds so that memory usable for
4529 * any allocation type is evenly spread. If both kernelcore
4530 * and movablecore are specified, then the value of kernelcore
4531 * will be used for required_kernelcore if it's greater than
4532 * what movablecore would have allowed.
4534 if (required_movablecore
) {
4535 unsigned long corepages
;
4538 * Round-up so that ZONE_MOVABLE is at least as large as what
4539 * was requested by the user
4541 required_movablecore
=
4542 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
4543 corepages
= totalpages
- required_movablecore
;
4545 required_kernelcore
= max(required_kernelcore
, corepages
);
4548 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4549 if (!required_kernelcore
)
4552 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4553 find_usable_zone_for_movable();
4554 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
4557 /* Spread kernelcore memory as evenly as possible throughout nodes */
4558 kernelcore_node
= required_kernelcore
/ usable_nodes
;
4559 for_each_node_state(nid
, N_HIGH_MEMORY
) {
4561 * Recalculate kernelcore_node if the division per node
4562 * now exceeds what is necessary to satisfy the requested
4563 * amount of memory for the kernel
4565 if (required_kernelcore
< kernelcore_node
)
4566 kernelcore_node
= required_kernelcore
/ usable_nodes
;
4569 * As the map is walked, we track how much memory is usable
4570 * by the kernel using kernelcore_remaining. When it is
4571 * 0, the rest of the node is usable by ZONE_MOVABLE
4573 kernelcore_remaining
= kernelcore_node
;
4575 /* Go through each range of PFNs within this node */
4576 for_each_active_range_index_in_nid(i
, nid
) {
4577 unsigned long start_pfn
, end_pfn
;
4578 unsigned long size_pages
;
4580 start_pfn
= max(early_node_map
[i
].start_pfn
,
4581 zone_movable_pfn
[nid
]);
4582 end_pfn
= early_node_map
[i
].end_pfn
;
4583 if (start_pfn
>= end_pfn
)
4586 /* Account for what is only usable for kernelcore */
4587 if (start_pfn
< usable_startpfn
) {
4588 unsigned long kernel_pages
;
4589 kernel_pages
= min(end_pfn
, usable_startpfn
)
4592 kernelcore_remaining
-= min(kernel_pages
,
4593 kernelcore_remaining
);
4594 required_kernelcore
-= min(kernel_pages
,
4595 required_kernelcore
);
4597 /* Continue if range is now fully accounted */
4598 if (end_pfn
<= usable_startpfn
) {
4601 * Push zone_movable_pfn to the end so
4602 * that if we have to rebalance
4603 * kernelcore across nodes, we will
4604 * not double account here
4606 zone_movable_pfn
[nid
] = end_pfn
;
4609 start_pfn
= usable_startpfn
;
4613 * The usable PFN range for ZONE_MOVABLE is from
4614 * start_pfn->end_pfn. Calculate size_pages as the
4615 * number of pages used as kernelcore
4617 size_pages
= end_pfn
- start_pfn
;
4618 if (size_pages
> kernelcore_remaining
)
4619 size_pages
= kernelcore_remaining
;
4620 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
4623 * Some kernelcore has been met, update counts and
4624 * break if the kernelcore for this node has been
4627 required_kernelcore
-= min(required_kernelcore
,
4629 kernelcore_remaining
-= size_pages
;
4630 if (!kernelcore_remaining
)
4636 * If there is still required_kernelcore, we do another pass with one
4637 * less node in the count. This will push zone_movable_pfn[nid] further
4638 * along on the nodes that still have memory until kernelcore is
4642 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
4645 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4646 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
4647 zone_movable_pfn
[nid
] =
4648 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
4651 /* restore the node_state */
4652 node_states
[N_HIGH_MEMORY
] = saved_node_state
;
4655 /* Any regular memory on that node ? */
4656 static void check_for_regular_memory(pg_data_t
*pgdat
)
4658 #ifdef CONFIG_HIGHMEM
4659 enum zone_type zone_type
;
4661 for (zone_type
= 0; zone_type
<= ZONE_NORMAL
; zone_type
++) {
4662 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
4663 if (zone
->present_pages
)
4664 node_set_state(zone_to_nid(zone
), N_NORMAL_MEMORY
);
4670 * free_area_init_nodes - Initialise all pg_data_t and zone data
4671 * @max_zone_pfn: an array of max PFNs for each zone
4673 * This will call free_area_init_node() for each active node in the system.
4674 * Using the page ranges provided by add_active_range(), the size of each
4675 * zone in each node and their holes is calculated. If the maximum PFN
4676 * between two adjacent zones match, it is assumed that the zone is empty.
4677 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4678 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4679 * starts where the previous one ended. For example, ZONE_DMA32 starts
4680 * at arch_max_dma_pfn.
4682 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
4687 /* Sort early_node_map as initialisation assumes it is sorted */
4690 /* Record where the zone boundaries are */
4691 memset(arch_zone_lowest_possible_pfn
, 0,
4692 sizeof(arch_zone_lowest_possible_pfn
));
4693 memset(arch_zone_highest_possible_pfn
, 0,
4694 sizeof(arch_zone_highest_possible_pfn
));
4695 arch_zone_lowest_possible_pfn
[0] = find_min_pfn_with_active_regions();
4696 arch_zone_highest_possible_pfn
[0] = max_zone_pfn
[0];
4697 for (i
= 1; i
< MAX_NR_ZONES
; i
++) {
4698 if (i
== ZONE_MOVABLE
)
4700 arch_zone_lowest_possible_pfn
[i
] =
4701 arch_zone_highest_possible_pfn
[i
-1];
4702 arch_zone_highest_possible_pfn
[i
] =
4703 max(max_zone_pfn
[i
], arch_zone_lowest_possible_pfn
[i
]);
4705 arch_zone_lowest_possible_pfn
[ZONE_MOVABLE
] = 0;
4706 arch_zone_highest_possible_pfn
[ZONE_MOVABLE
] = 0;
4708 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4709 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
4710 find_zone_movable_pfns_for_nodes(zone_movable_pfn
);
4712 /* Print out the zone ranges */
4713 printk("Zone PFN ranges:\n");
4714 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
4715 if (i
== ZONE_MOVABLE
)
4717 printk(" %-8s ", zone_names
[i
]);
4718 if (arch_zone_lowest_possible_pfn
[i
] ==
4719 arch_zone_highest_possible_pfn
[i
])
4722 printk("%0#10lx -> %0#10lx\n",
4723 arch_zone_lowest_possible_pfn
[i
],
4724 arch_zone_highest_possible_pfn
[i
]);
4727 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4728 printk("Movable zone start PFN for each node\n");
4729 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
4730 if (zone_movable_pfn
[i
])
4731 printk(" Node %d: %lu\n", i
, zone_movable_pfn
[i
]);
4734 /* Print out the early_node_map[] */
4735 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries
);
4736 for (i
= 0; i
< nr_nodemap_entries
; i
++)
4737 printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map
[i
].nid
,
4738 early_node_map
[i
].start_pfn
,
4739 early_node_map
[i
].end_pfn
);
4741 /* Initialise every node */
4742 mminit_verify_pageflags_layout();
4743 setup_nr_node_ids();
4744 for_each_online_node(nid
) {
4745 pg_data_t
*pgdat
= NODE_DATA(nid
);
4746 free_area_init_node(nid
, NULL
,
4747 find_min_pfn_for_node(nid
), NULL
);
4749 /* Any memory on that node */
4750 if (pgdat
->node_present_pages
)
4751 node_set_state(nid
, N_HIGH_MEMORY
);
4752 check_for_regular_memory(pgdat
);
4756 static int __init
cmdline_parse_core(char *p
, unsigned long *core
)
4758 unsigned long long coremem
;
4762 coremem
= memparse(p
, &p
);
4763 *core
= coremem
>> PAGE_SHIFT
;
4765 /* Paranoid check that UL is enough for the coremem value */
4766 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
4772 * kernelcore=size sets the amount of memory for use for allocations that
4773 * cannot be reclaimed or migrated.
4775 static int __init
cmdline_parse_kernelcore(char *p
)
4777 return cmdline_parse_core(p
, &required_kernelcore
);
4781 * movablecore=size sets the amount of memory for use for allocations that
4782 * can be reclaimed or migrated.
4784 static int __init
cmdline_parse_movablecore(char *p
)
4786 return cmdline_parse_core(p
, &required_movablecore
);
4789 early_param("kernelcore", cmdline_parse_kernelcore
);
4790 early_param("movablecore", cmdline_parse_movablecore
);
4792 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4795 * set_dma_reserve - set the specified number of pages reserved in the first zone
4796 * @new_dma_reserve: The number of pages to mark reserved
4798 * The per-cpu batchsize and zone watermarks are determined by present_pages.
4799 * In the DMA zone, a significant percentage may be consumed by kernel image
4800 * and other unfreeable allocations which can skew the watermarks badly. This
4801 * function may optionally be used to account for unfreeable pages in the
4802 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
4803 * smaller per-cpu batchsize.
4805 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
4807 dma_reserve
= new_dma_reserve
;
4810 #ifndef CONFIG_NEED_MULTIPLE_NODES
4811 struct pglist_data __refdata contig_page_data
= {
4812 #ifndef CONFIG_NO_BOOTMEM
4813 .bdata
= &bootmem_node_data
[0]
4816 EXPORT_SYMBOL(contig_page_data
);
4819 void __init
free_area_init(unsigned long *zones_size
)
4821 free_area_init_node(0, zones_size
,
4822 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
4825 static int page_alloc_cpu_notify(struct notifier_block
*self
,
4826 unsigned long action
, void *hcpu
)
4828 int cpu
= (unsigned long)hcpu
;
4830 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
) {
4834 * Spill the event counters of the dead processor
4835 * into the current processors event counters.
4836 * This artificially elevates the count of the current
4839 vm_events_fold_cpu(cpu
);
4842 * Zero the differential counters of the dead processor
4843 * so that the vm statistics are consistent.
4845 * This is only okay since the processor is dead and cannot
4846 * race with what we are doing.
4848 refresh_cpu_vm_stats(cpu
);
4853 void __init
page_alloc_init(void)
4855 hotcpu_notifier(page_alloc_cpu_notify
, 0);
4859 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
4860 * or min_free_kbytes changes.
4862 static void calculate_totalreserve_pages(void)
4864 struct pglist_data
*pgdat
;
4865 unsigned long reserve_pages
= 0;
4866 enum zone_type i
, j
;
4868 for_each_online_pgdat(pgdat
) {
4869 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
4870 struct zone
*zone
= pgdat
->node_zones
+ i
;
4871 unsigned long max
= 0;
4873 /* Find valid and maximum lowmem_reserve in the zone */
4874 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
4875 if (zone
->lowmem_reserve
[j
] > max
)
4876 max
= zone
->lowmem_reserve
[j
];
4879 /* we treat the high watermark as reserved pages. */
4880 max
+= high_wmark_pages(zone
);
4882 if (max
> zone
->present_pages
)
4883 max
= zone
->present_pages
;
4884 reserve_pages
+= max
;
4887 totalreserve_pages
= reserve_pages
;
4891 * setup_per_zone_lowmem_reserve - called whenever
4892 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
4893 * has a correct pages reserved value, so an adequate number of
4894 * pages are left in the zone after a successful __alloc_pages().
4896 static void setup_per_zone_lowmem_reserve(void)
4898 struct pglist_data
*pgdat
;
4899 enum zone_type j
, idx
;
4901 for_each_online_pgdat(pgdat
) {
4902 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
4903 struct zone
*zone
= pgdat
->node_zones
+ j
;
4904 unsigned long present_pages
= zone
->present_pages
;
4906 zone
->lowmem_reserve
[j
] = 0;
4910 struct zone
*lower_zone
;
4914 if (sysctl_lowmem_reserve_ratio
[idx
] < 1)
4915 sysctl_lowmem_reserve_ratio
[idx
] = 1;
4917 lower_zone
= pgdat
->node_zones
+ idx
;
4918 lower_zone
->lowmem_reserve
[j
] = present_pages
/
4919 sysctl_lowmem_reserve_ratio
[idx
];
4920 present_pages
+= lower_zone
->present_pages
;
4925 /* update totalreserve_pages */
4926 calculate_totalreserve_pages();
4930 * setup_per_zone_wmarks - called when min_free_kbytes changes
4931 * or when memory is hot-{added|removed}
4933 * Ensures that the watermark[min,low,high] values for each zone are set
4934 * correctly with respect to min_free_kbytes.
4936 void setup_per_zone_wmarks(void)
4938 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
4939 unsigned long lowmem_pages
= 0;
4941 unsigned long flags
;
4943 /* Calculate total number of !ZONE_HIGHMEM pages */
4944 for_each_zone(zone
) {
4945 if (!is_highmem(zone
))
4946 lowmem_pages
+= zone
->present_pages
;
4949 for_each_zone(zone
) {
4952 spin_lock_irqsave(&zone
->lock
, flags
);
4953 tmp
= (u64
)pages_min
* zone
->present_pages
;
4954 do_div(tmp
, lowmem_pages
);
4955 if (is_highmem(zone
)) {
4957 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
4958 * need highmem pages, so cap pages_min to a small
4961 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
4962 * deltas controls asynch page reclaim, and so should
4963 * not be capped for highmem.
4967 min_pages
= zone
->present_pages
/ 1024;
4968 if (min_pages
< SWAP_CLUSTER_MAX
)
4969 min_pages
= SWAP_CLUSTER_MAX
;
4970 if (min_pages
> 128)
4972 zone
->watermark
[WMARK_MIN
] = min_pages
;
4975 * If it's a lowmem zone, reserve a number of pages
4976 * proportionate to the zone's size.
4978 zone
->watermark
[WMARK_MIN
] = tmp
;
4981 zone
->watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + (tmp
>> 2);
4982 zone
->watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + (tmp
>> 1);
4983 setup_zone_migrate_reserve(zone
);
4984 spin_unlock_irqrestore(&zone
->lock
, flags
);
4987 /* update totalreserve_pages */
4988 calculate_totalreserve_pages();
4992 * The inactive anon list should be small enough that the VM never has to
4993 * do too much work, but large enough that each inactive page has a chance
4994 * to be referenced again before it is swapped out.
4996 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
4997 * INACTIVE_ANON pages on this zone's LRU, maintained by the
4998 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
4999 * the anonymous pages are kept on the inactive list.
5002 * memory ratio inactive anon
5003 * -------------------------------------
5012 void calculate_zone_inactive_ratio(struct zone
*zone
)
5014 unsigned int gb
, ratio
;
5016 /* Zone size in gigabytes */
5017 gb
= zone
->present_pages
>> (30 - PAGE_SHIFT
);
5019 ratio
= int_sqrt(10 * gb
);
5023 zone
->inactive_ratio
= ratio
;
5026 static void __init
setup_per_zone_inactive_ratio(void)
5031 calculate_zone_inactive_ratio(zone
);
5035 * Initialise min_free_kbytes.
5037 * For small machines we want it small (128k min). For large machines
5038 * we want it large (64MB max). But it is not linear, because network
5039 * bandwidth does not increase linearly with machine size. We use
5041 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5042 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5058 static int __init
init_per_zone_wmark_min(void)
5060 unsigned long lowmem_kbytes
;
5062 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
5064 min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
5065 if (min_free_kbytes
< 128)
5066 min_free_kbytes
= 128;
5067 if (min_free_kbytes
> 65536)
5068 min_free_kbytes
= 65536;
5069 setup_per_zone_wmarks();
5070 setup_per_zone_lowmem_reserve();
5071 setup_per_zone_inactive_ratio();
5074 module_init(init_per_zone_wmark_min
)
5077 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5078 * that we can call two helper functions whenever min_free_kbytes
5081 int min_free_kbytes_sysctl_handler(ctl_table
*table
, int write
,
5082 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5084 proc_dointvec(table
, write
, buffer
, length
, ppos
);
5086 setup_per_zone_wmarks();
5091 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table
*table
, int write
,
5092 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5097 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5102 zone
->min_unmapped_pages
= (zone
->present_pages
*
5103 sysctl_min_unmapped_ratio
) / 100;
5107 int sysctl_min_slab_ratio_sysctl_handler(ctl_table
*table
, int write
,
5108 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5113 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5118 zone
->min_slab_pages
= (zone
->present_pages
*
5119 sysctl_min_slab_ratio
) / 100;
5125 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5126 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5127 * whenever sysctl_lowmem_reserve_ratio changes.
5129 * The reserve ratio obviously has absolutely no relation with the
5130 * minimum watermarks. The lowmem reserve ratio can only make sense
5131 * if in function of the boot time zone sizes.
5133 int lowmem_reserve_ratio_sysctl_handler(ctl_table
*table
, int write
,
5134 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5136 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5137 setup_per_zone_lowmem_reserve();
5142 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5143 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
5144 * can have before it gets flushed back to buddy allocator.
5147 int percpu_pagelist_fraction_sysctl_handler(ctl_table
*table
, int write
,
5148 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5154 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5155 if (!write
|| (ret
== -EINVAL
))
5157 for_each_populated_zone(zone
) {
5158 for_each_possible_cpu(cpu
) {
5160 high
= zone
->present_pages
/ percpu_pagelist_fraction
;
5161 setup_pagelist_highmark(
5162 per_cpu_ptr(zone
->pageset
, cpu
), high
);
5168 int hashdist
= HASHDIST_DEFAULT
;
5171 static int __init
set_hashdist(char *str
)
5175 hashdist
= simple_strtoul(str
, &str
, 0);
5178 __setup("hashdist=", set_hashdist
);
5182 * allocate a large system hash table from bootmem
5183 * - it is assumed that the hash table must contain an exact power-of-2
5184 * quantity of entries
5185 * - limit is the number of hash buckets, not the total allocation size
5187 void *__init
alloc_large_system_hash(const char *tablename
,
5188 unsigned long bucketsize
,
5189 unsigned long numentries
,
5192 unsigned int *_hash_shift
,
5193 unsigned int *_hash_mask
,
5194 unsigned long limit
)
5196 unsigned long long max
= limit
;
5197 unsigned long log2qty
, size
;
5200 /* allow the kernel cmdline to have a say */
5202 /* round applicable memory size up to nearest megabyte */
5203 numentries
= nr_kernel_pages
;
5204 numentries
+= (1UL << (20 - PAGE_SHIFT
)) - 1;
5205 numentries
>>= 20 - PAGE_SHIFT
;
5206 numentries
<<= 20 - PAGE_SHIFT
;
5208 /* limit to 1 bucket per 2^scale bytes of low memory */
5209 if (scale
> PAGE_SHIFT
)
5210 numentries
>>= (scale
- PAGE_SHIFT
);
5212 numentries
<<= (PAGE_SHIFT
- scale
);
5214 /* Make sure we've got at least a 0-order allocation.. */
5215 if (unlikely(flags
& HASH_SMALL
)) {
5216 /* Makes no sense without HASH_EARLY */
5217 WARN_ON(!(flags
& HASH_EARLY
));
5218 if (!(numentries
>> *_hash_shift
)) {
5219 numentries
= 1UL << *_hash_shift
;
5220 BUG_ON(!numentries
);
5222 } else if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
5223 numentries
= PAGE_SIZE
/ bucketsize
;
5225 numentries
= roundup_pow_of_two(numentries
);
5227 /* limit allocation size to 1/16 total memory by default */
5229 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
5230 do_div(max
, bucketsize
);
5233 if (numentries
> max
)
5236 log2qty
= ilog2(numentries
);
5239 size
= bucketsize
<< log2qty
;
5240 if (flags
& HASH_EARLY
)
5241 table
= alloc_bootmem_nopanic(size
);
5243 table
= __vmalloc(size
, GFP_ATOMIC
, PAGE_KERNEL
);
5246 * If bucketsize is not a power-of-two, we may free
5247 * some pages at the end of hash table which
5248 * alloc_pages_exact() automatically does
5250 if (get_order(size
) < MAX_ORDER
) {
5251 table
= alloc_pages_exact(size
, GFP_ATOMIC
);
5252 kmemleak_alloc(table
, size
, 1, GFP_ATOMIC
);
5255 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
5258 panic("Failed to allocate %s hash table\n", tablename
);
5260 printk(KERN_INFO
"%s hash table entries: %ld (order: %d, %lu bytes)\n",
5263 ilog2(size
) - PAGE_SHIFT
,
5267 *_hash_shift
= log2qty
;
5269 *_hash_mask
= (1 << log2qty
) - 1;
5274 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5275 static inline unsigned long *get_pageblock_bitmap(struct zone
*zone
,
5278 #ifdef CONFIG_SPARSEMEM
5279 return __pfn_to_section(pfn
)->pageblock_flags
;
5281 return zone
->pageblock_flags
;
5282 #endif /* CONFIG_SPARSEMEM */
5285 static inline int pfn_to_bitidx(struct zone
*zone
, unsigned long pfn
)
5287 #ifdef CONFIG_SPARSEMEM
5288 pfn
&= (PAGES_PER_SECTION
-1);
5289 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
5291 pfn
= pfn
- zone
->zone_start_pfn
;
5292 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
5293 #endif /* CONFIG_SPARSEMEM */
5297 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5298 * @page: The page within the block of interest
5299 * @start_bitidx: The first bit of interest to retrieve
5300 * @end_bitidx: The last bit of interest
5301 * returns pageblock_bits flags
5303 unsigned long get_pageblock_flags_group(struct page
*page
,
5304 int start_bitidx
, int end_bitidx
)
5307 unsigned long *bitmap
;
5308 unsigned long pfn
, bitidx
;
5309 unsigned long flags
= 0;
5310 unsigned long value
= 1;
5312 zone
= page_zone(page
);
5313 pfn
= page_to_pfn(page
);
5314 bitmap
= get_pageblock_bitmap(zone
, pfn
);
5315 bitidx
= pfn_to_bitidx(zone
, pfn
);
5317 for (; start_bitidx
<= end_bitidx
; start_bitidx
++, value
<<= 1)
5318 if (test_bit(bitidx
+ start_bitidx
, bitmap
))
5325 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5326 * @page: The page within the block of interest
5327 * @start_bitidx: The first bit of interest
5328 * @end_bitidx: The last bit of interest
5329 * @flags: The flags to set
5331 void set_pageblock_flags_group(struct page
*page
, unsigned long flags
,
5332 int start_bitidx
, int end_bitidx
)
5335 unsigned long *bitmap
;
5336 unsigned long pfn
, bitidx
;
5337 unsigned long value
= 1;
5339 zone
= page_zone(page
);
5340 pfn
= page_to_pfn(page
);
5341 bitmap
= get_pageblock_bitmap(zone
, pfn
);
5342 bitidx
= pfn_to_bitidx(zone
, pfn
);
5343 VM_BUG_ON(pfn
< zone
->zone_start_pfn
);
5344 VM_BUG_ON(pfn
>= zone
->zone_start_pfn
+ zone
->spanned_pages
);
5346 for (; start_bitidx
<= end_bitidx
; start_bitidx
++, value
<<= 1)
5348 __set_bit(bitidx
+ start_bitidx
, bitmap
);
5350 __clear_bit(bitidx
+ start_bitidx
, bitmap
);
5354 * This is designed as sub function...plz see page_isolation.c also.
5355 * set/clear page block's type to be ISOLATE.
5356 * page allocater never alloc memory from ISOLATE block.
5360 __count_immobile_pages(struct zone
*zone
, struct page
*page
, int count
)
5362 unsigned long pfn
, iter
, found
;
5364 * For avoiding noise data, lru_add_drain_all() should be called
5365 * If ZONE_MOVABLE, the zone never contains immobile pages
5367 if (zone_idx(zone
) == ZONE_MOVABLE
)
5370 if (get_pageblock_migratetype(page
) == MIGRATE_MOVABLE
)
5373 pfn
= page_to_pfn(page
);
5374 for (found
= 0, iter
= 0; iter
< pageblock_nr_pages
; iter
++) {
5375 unsigned long check
= pfn
+ iter
;
5377 if (!pfn_valid_within(check
)) {
5381 page
= pfn_to_page(check
);
5382 if (!page_count(page
)) {
5383 if (PageBuddy(page
))
5384 iter
+= (1 << page_order(page
)) - 1;
5390 * If there are RECLAIMABLE pages, we need to check it.
5391 * But now, memory offline itself doesn't call shrink_slab()
5392 * and it still to be fixed.
5395 * If the page is not RAM, page_count()should be 0.
5396 * we don't need more check. This is an _used_ not-movable page.
5398 * The problematic thing here is PG_reserved pages. PG_reserved
5399 * is set to both of a memory hole page and a _used_ kernel
5408 bool is_pageblock_removable_nolock(struct page
*page
)
5410 struct zone
*zone
= page_zone(page
);
5411 return __count_immobile_pages(zone
, page
, 0);
5414 int set_migratetype_isolate(struct page
*page
)
5417 unsigned long flags
, pfn
;
5418 struct memory_isolate_notify arg
;
5423 zone
= page_zone(page
);
5424 zone_idx
= zone_idx(zone
);
5426 spin_lock_irqsave(&zone
->lock
, flags
);
5428 pfn
= page_to_pfn(page
);
5429 arg
.start_pfn
= pfn
;
5430 arg
.nr_pages
= pageblock_nr_pages
;
5431 arg
.pages_found
= 0;
5434 * It may be possible to isolate a pageblock even if the
5435 * migratetype is not MIGRATE_MOVABLE. The memory isolation
5436 * notifier chain is used by balloon drivers to return the
5437 * number of pages in a range that are held by the balloon
5438 * driver to shrink memory. If all the pages are accounted for
5439 * by balloons, are free, or on the LRU, isolation can continue.
5440 * Later, for example, when memory hotplug notifier runs, these
5441 * pages reported as "can be isolated" should be isolated(freed)
5442 * by the balloon driver through the memory notifier chain.
5444 notifier_ret
= memory_isolate_notify(MEM_ISOLATE_COUNT
, &arg
);
5445 notifier_ret
= notifier_to_errno(notifier_ret
);
5449 * FIXME: Now, memory hotplug doesn't call shrink_slab() by itself.
5450 * We just check MOVABLE pages.
5452 if (__count_immobile_pages(zone
, page
, arg
.pages_found
))
5456 * immobile means "not-on-lru" paes. If immobile is larger than
5457 * removable-by-driver pages reported by notifier, we'll fail.
5462 set_pageblock_migratetype(page
, MIGRATE_ISOLATE
);
5463 move_freepages_block(zone
, page
, MIGRATE_ISOLATE
);
5466 spin_unlock_irqrestore(&zone
->lock
, flags
);
5472 void unset_migratetype_isolate(struct page
*page
)
5475 unsigned long flags
;
5476 zone
= page_zone(page
);
5477 spin_lock_irqsave(&zone
->lock
, flags
);
5478 if (get_pageblock_migratetype(page
) != MIGRATE_ISOLATE
)
5480 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
5481 move_freepages_block(zone
, page
, MIGRATE_MOVABLE
);
5483 spin_unlock_irqrestore(&zone
->lock
, flags
);
5486 #ifdef CONFIG_MEMORY_HOTREMOVE
5488 * All pages in the range must be isolated before calling this.
5491 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
5497 unsigned long flags
;
5498 /* find the first valid pfn */
5499 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
5504 zone
= page_zone(pfn_to_page(pfn
));
5505 spin_lock_irqsave(&zone
->lock
, flags
);
5507 while (pfn
< end_pfn
) {
5508 if (!pfn_valid(pfn
)) {
5512 page
= pfn_to_page(pfn
);
5513 BUG_ON(page_count(page
));
5514 BUG_ON(!PageBuddy(page
));
5515 order
= page_order(page
);
5516 #ifdef CONFIG_DEBUG_VM
5517 printk(KERN_INFO
"remove from free list %lx %d %lx\n",
5518 pfn
, 1 << order
, end_pfn
);
5520 list_del(&page
->lru
);
5521 rmv_page_order(page
);
5522 zone
->free_area
[order
].nr_free
--;
5523 __mod_zone_page_state(zone
, NR_FREE_PAGES
,
5525 for (i
= 0; i
< (1 << order
); i
++)
5526 SetPageReserved((page
+i
));
5527 pfn
+= (1 << order
);
5529 spin_unlock_irqrestore(&zone
->lock
, flags
);
5533 #ifdef CONFIG_MEMORY_FAILURE
5534 bool is_free_buddy_page(struct page
*page
)
5536 struct zone
*zone
= page_zone(page
);
5537 unsigned long pfn
= page_to_pfn(page
);
5538 unsigned long flags
;
5541 spin_lock_irqsave(&zone
->lock
, flags
);
5542 for (order
= 0; order
< MAX_ORDER
; order
++) {
5543 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
5545 if (PageBuddy(page_head
) && page_order(page_head
) >= order
)
5548 spin_unlock_irqrestore(&zone
->lock
, flags
);
5550 return order
< MAX_ORDER
;
5554 static struct trace_print_flags pageflag_names
[] = {
5555 {1UL << PG_locked
, "locked" },
5556 {1UL << PG_error
, "error" },
5557 {1UL << PG_referenced
, "referenced" },
5558 {1UL << PG_uptodate
, "uptodate" },
5559 {1UL << PG_dirty
, "dirty" },
5560 {1UL << PG_lru
, "lru" },
5561 {1UL << PG_active
, "active" },
5562 {1UL << PG_slab
, "slab" },
5563 {1UL << PG_owner_priv_1
, "owner_priv_1" },
5564 {1UL << PG_arch_1
, "arch_1" },
5565 {1UL << PG_reserved
, "reserved" },
5566 {1UL << PG_private
, "private" },
5567 {1UL << PG_private_2
, "private_2" },
5568 {1UL << PG_writeback
, "writeback" },
5569 #ifdef CONFIG_PAGEFLAGS_EXTENDED
5570 {1UL << PG_head
, "head" },
5571 {1UL << PG_tail
, "tail" },
5573 {1UL << PG_compound
, "compound" },
5575 {1UL << PG_swapcache
, "swapcache" },
5576 {1UL << PG_mappedtodisk
, "mappedtodisk" },
5577 {1UL << PG_reclaim
, "reclaim" },
5578 {1UL << PG_swapbacked
, "swapbacked" },
5579 {1UL << PG_unevictable
, "unevictable" },
5581 {1UL << PG_mlocked
, "mlocked" },
5583 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
5584 {1UL << PG_uncached
, "uncached" },
5586 #ifdef CONFIG_MEMORY_FAILURE
5587 {1UL << PG_hwpoison
, "hwpoison" },
5592 static void dump_page_flags(unsigned long flags
)
5594 const char *delim
= "";
5598 printk(KERN_ALERT
"page flags: %#lx(", flags
);
5600 /* remove zone id */
5601 flags
&= (1UL << NR_PAGEFLAGS
) - 1;
5603 for (i
= 0; pageflag_names
[i
].name
&& flags
; i
++) {
5605 mask
= pageflag_names
[i
].mask
;
5606 if ((flags
& mask
) != mask
)
5610 printk("%s%s", delim
, pageflag_names
[i
].name
);
5614 /* check for left over flags */
5616 printk("%s%#lx", delim
, flags
);
5621 void dump_page(struct page
*page
)
5624 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
5625 page
, atomic_read(&page
->_count
), page_mapcount(page
),
5626 page
->mapping
, page
->index
);
5627 dump_page_flags(page
->flags
);