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 static int destroy_compound_page(struct page
*page
, unsigned long order
)
363 int nr_pages
= 1 << order
;
366 if (unlikely(compound_order(page
) != order
) ||
367 unlikely(!PageHead(page
))) {
372 __ClearPageHead(page
);
374 for (i
= 1; i
< nr_pages
; i
++) {
375 struct page
*p
= page
+ i
;
377 if (unlikely(!PageTail(p
) || (p
->first_page
!= page
))) {
387 static inline void prep_zero_page(struct page
*page
, int order
, gfp_t gfp_flags
)
392 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
393 * and __GFP_HIGHMEM from hard or soft interrupt context.
395 VM_BUG_ON((gfp_flags
& __GFP_HIGHMEM
) && in_interrupt());
396 for (i
= 0; i
< (1 << order
); i
++)
397 clear_highpage(page
+ i
);
400 static inline void set_page_order(struct page
*page
, int order
)
402 set_page_private(page
, order
);
403 __SetPageBuddy(page
);
406 static inline void rmv_page_order(struct page
*page
)
408 __ClearPageBuddy(page
);
409 set_page_private(page
, 0);
413 * Locate the struct page for both the matching buddy in our
414 * pair (buddy1) and the combined O(n+1) page they form (page).
416 * 1) Any buddy B1 will have an order O twin B2 which satisfies
417 * the following equation:
419 * For example, if the starting buddy (buddy2) is #8 its order
421 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
423 * 2) Any buddy B will have an order O+1 parent P which
424 * satisfies the following equation:
427 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
429 static inline struct page
*
430 __page_find_buddy(struct page
*page
, unsigned long page_idx
, unsigned int order
)
432 unsigned long buddy_idx
= page_idx
^ (1 << order
);
434 return page
+ (buddy_idx
- page_idx
);
437 static inline unsigned long
438 __find_combined_index(unsigned long page_idx
, unsigned int order
)
440 return (page_idx
& ~(1 << order
));
444 * This function checks whether a page is free && is the buddy
445 * we can do coalesce a page and its buddy if
446 * (a) the buddy is not in a hole &&
447 * (b) the buddy is in the buddy system &&
448 * (c) a page and its buddy have the same order &&
449 * (d) a page and its buddy are in the same zone.
451 * For recording whether a page is in the buddy system, we use PG_buddy.
452 * Setting, clearing, and testing PG_buddy is serialized by zone->lock.
454 * For recording page's order, we use page_private(page).
456 static inline int page_is_buddy(struct page
*page
, struct page
*buddy
,
459 if (!pfn_valid_within(page_to_pfn(buddy
)))
462 if (page_zone_id(page
) != page_zone_id(buddy
))
465 if (PageBuddy(buddy
) && page_order(buddy
) == order
) {
466 VM_BUG_ON(page_count(buddy
) != 0);
473 * Freeing function for a buddy system allocator.
475 * The concept of a buddy system is to maintain direct-mapped table
476 * (containing bit values) for memory blocks of various "orders".
477 * The bottom level table contains the map for the smallest allocatable
478 * units of memory (here, pages), and each level above it describes
479 * pairs of units from the levels below, hence, "buddies".
480 * At a high level, all that happens here is marking the table entry
481 * at the bottom level available, and propagating the changes upward
482 * as necessary, plus some accounting needed to play nicely with other
483 * parts of the VM system.
484 * At each level, we keep a list of pages, which are heads of continuous
485 * free pages of length of (1 << order) and marked with PG_buddy. Page's
486 * order is recorded in page_private(page) field.
487 * So when we are allocating or freeing one, we can derive the state of the
488 * other. That is, if we allocate a small block, and both were
489 * free, the remainder of the region must be split into blocks.
490 * If a block is freed, and its buddy is also free, then this
491 * triggers coalescing into a block of larger size.
496 static inline void __free_one_page(struct page
*page
,
497 struct zone
*zone
, unsigned int order
,
500 unsigned long page_idx
;
501 unsigned long combined_idx
;
504 if (unlikely(PageCompound(page
)))
505 if (unlikely(destroy_compound_page(page
, order
)))
508 VM_BUG_ON(migratetype
== -1);
510 page_idx
= page_to_pfn(page
) & ((1 << MAX_ORDER
) - 1);
512 VM_BUG_ON(page_idx
& ((1 << order
) - 1));
513 VM_BUG_ON(bad_range(zone
, page
));
515 while (order
< MAX_ORDER
-1) {
516 buddy
= __page_find_buddy(page
, page_idx
, order
);
517 if (!page_is_buddy(page
, buddy
, order
))
520 /* Our buddy is free, merge with it and move up one order. */
521 list_del(&buddy
->lru
);
522 zone
->free_area
[order
].nr_free
--;
523 rmv_page_order(buddy
);
524 combined_idx
= __find_combined_index(page_idx
, order
);
525 page
= page
+ (combined_idx
- page_idx
);
526 page_idx
= combined_idx
;
529 set_page_order(page
, order
);
532 * If this is not the largest possible page, check if the buddy
533 * of the next-highest order is free. If it is, it's possible
534 * that pages are being freed that will coalesce soon. In case,
535 * that is happening, add the free page to the tail of the list
536 * so it's less likely to be used soon and more likely to be merged
537 * as a higher order page
539 if ((order
< MAX_ORDER
-2) && pfn_valid_within(page_to_pfn(buddy
))) {
540 struct page
*higher_page
, *higher_buddy
;
541 combined_idx
= __find_combined_index(page_idx
, order
);
542 higher_page
= page
+ combined_idx
- page_idx
;
543 higher_buddy
= __page_find_buddy(higher_page
, combined_idx
, order
+ 1);
544 if (page_is_buddy(higher_page
, higher_buddy
, order
+ 1)) {
545 list_add_tail(&page
->lru
,
546 &zone
->free_area
[order
].free_list
[migratetype
]);
551 list_add(&page
->lru
, &zone
->free_area
[order
].free_list
[migratetype
]);
553 zone
->free_area
[order
].nr_free
++;
557 * free_page_mlock() -- clean up attempts to free and mlocked() page.
558 * Page should not be on lru, so no need to fix that up.
559 * free_pages_check() will verify...
561 static inline void free_page_mlock(struct page
*page
)
563 __dec_zone_page_state(page
, NR_MLOCK
);
564 __count_vm_event(UNEVICTABLE_MLOCKFREED
);
567 static inline int free_pages_check(struct page
*page
)
569 if (unlikely(page_mapcount(page
) |
570 (page
->mapping
!= NULL
) |
571 (atomic_read(&page
->_count
) != 0) |
572 (page
->flags
& PAGE_FLAGS_CHECK_AT_FREE
))) {
576 if (page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)
577 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
582 * Frees a number of pages from the PCP lists
583 * Assumes all pages on list are in same zone, and of same order.
584 * count is the number of pages to free.
586 * If the zone was previously in an "all pages pinned" state then look to
587 * see if this freeing clears that state.
589 * And clear the zone's pages_scanned counter, to hold off the "all pages are
590 * pinned" detection logic.
592 static void free_pcppages_bulk(struct zone
*zone
, int count
,
593 struct per_cpu_pages
*pcp
)
599 spin_lock(&zone
->lock
);
600 zone
->all_unreclaimable
= 0;
601 zone
->pages_scanned
= 0;
605 struct list_head
*list
;
608 * Remove pages from lists in a round-robin fashion. A
609 * batch_free count is maintained that is incremented when an
610 * empty list is encountered. This is so more pages are freed
611 * off fuller lists instead of spinning excessively around empty
616 if (++migratetype
== MIGRATE_PCPTYPES
)
618 list
= &pcp
->lists
[migratetype
];
619 } while (list_empty(list
));
622 page
= list_entry(list
->prev
, struct page
, lru
);
623 /* must delete as __free_one_page list manipulates */
624 list_del(&page
->lru
);
625 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
626 __free_one_page(page
, zone
, 0, page_private(page
));
627 trace_mm_page_pcpu_drain(page
, 0, page_private(page
));
628 } while (--to_free
&& --batch_free
&& !list_empty(list
));
630 __mod_zone_page_state(zone
, NR_FREE_PAGES
, count
);
631 spin_unlock(&zone
->lock
);
634 static void free_one_page(struct zone
*zone
, struct page
*page
, int order
,
637 spin_lock(&zone
->lock
);
638 zone
->all_unreclaimable
= 0;
639 zone
->pages_scanned
= 0;
641 __free_one_page(page
, zone
, order
, migratetype
);
642 __mod_zone_page_state(zone
, NR_FREE_PAGES
, 1 << order
);
643 spin_unlock(&zone
->lock
);
646 static bool free_pages_prepare(struct page
*page
, unsigned int order
)
651 trace_mm_page_free_direct(page
, order
);
652 kmemcheck_free_shadow(page
, order
);
654 for (i
= 0; i
< (1 << order
); i
++) {
655 struct page
*pg
= page
+ i
;
659 bad
+= free_pages_check(pg
);
664 if (!PageHighMem(page
)) {
665 debug_check_no_locks_freed(page_address(page
),PAGE_SIZE
<<order
);
666 debug_check_no_obj_freed(page_address(page
),
669 arch_free_page(page
, order
);
670 kernel_map_pages(page
, 1 << order
, 0);
675 static void __free_pages_ok(struct page
*page
, unsigned int order
)
678 int wasMlocked
= __TestClearPageMlocked(page
);
680 if (!free_pages_prepare(page
, order
))
683 local_irq_save(flags
);
684 if (unlikely(wasMlocked
))
685 free_page_mlock(page
);
686 __count_vm_events(PGFREE
, 1 << order
);
687 free_one_page(page_zone(page
), page
, order
,
688 get_pageblock_migratetype(page
));
689 local_irq_restore(flags
);
693 * permit the bootmem allocator to evade page validation on high-order frees
695 void __meminit
__free_pages_bootmem(struct page
*page
, unsigned int order
)
698 __ClearPageReserved(page
);
699 set_page_count(page
, 0);
700 set_page_refcounted(page
);
706 for (loop
= 0; loop
< BITS_PER_LONG
; loop
++) {
707 struct page
*p
= &page
[loop
];
709 if (loop
+ 1 < BITS_PER_LONG
)
711 __ClearPageReserved(p
);
712 set_page_count(p
, 0);
715 set_page_refcounted(page
);
716 __free_pages(page
, order
);
722 * The order of subdivision here is critical for the IO subsystem.
723 * Please do not alter this order without good reasons and regression
724 * testing. Specifically, as large blocks of memory are subdivided,
725 * the order in which smaller blocks are delivered depends on the order
726 * they're subdivided in this function. This is the primary factor
727 * influencing the order in which pages are delivered to the IO
728 * subsystem according to empirical testing, and this is also justified
729 * by considering the behavior of a buddy system containing a single
730 * large block of memory acted on by a series of small allocations.
731 * This behavior is a critical factor in sglist merging's success.
735 static inline void expand(struct zone
*zone
, struct page
*page
,
736 int low
, int high
, struct free_area
*area
,
739 unsigned long size
= 1 << high
;
745 VM_BUG_ON(bad_range(zone
, &page
[size
]));
746 list_add(&page
[size
].lru
, &area
->free_list
[migratetype
]);
748 set_page_order(&page
[size
], high
);
753 * This page is about to be returned from the page allocator
755 static inline int check_new_page(struct page
*page
)
757 if (unlikely(page_mapcount(page
) |
758 (page
->mapping
!= NULL
) |
759 (atomic_read(&page
->_count
) != 0) |
760 (page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
))) {
767 static int prep_new_page(struct page
*page
, int order
, gfp_t gfp_flags
)
771 for (i
= 0; i
< (1 << order
); i
++) {
772 struct page
*p
= page
+ i
;
773 if (unlikely(check_new_page(p
)))
777 set_page_private(page
, 0);
778 set_page_refcounted(page
);
780 arch_alloc_page(page
, order
);
781 kernel_map_pages(page
, 1 << order
, 1);
783 if (gfp_flags
& __GFP_ZERO
)
784 prep_zero_page(page
, order
, gfp_flags
);
786 if (order
&& (gfp_flags
& __GFP_COMP
))
787 prep_compound_page(page
, order
);
793 * Go through the free lists for the given migratetype and remove
794 * the smallest available page from the freelists
797 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
800 unsigned int current_order
;
801 struct free_area
* area
;
804 /* Find a page of the appropriate size in the preferred list */
805 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
806 area
= &(zone
->free_area
[current_order
]);
807 if (list_empty(&area
->free_list
[migratetype
]))
810 page
= list_entry(area
->free_list
[migratetype
].next
,
812 list_del(&page
->lru
);
813 rmv_page_order(page
);
815 expand(zone
, page
, order
, current_order
, area
, migratetype
);
824 * This array describes the order lists are fallen back to when
825 * the free lists for the desirable migrate type are depleted
827 static int fallbacks
[MIGRATE_TYPES
][MIGRATE_TYPES
-1] = {
828 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
829 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
830 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_RESERVE
},
831 [MIGRATE_RESERVE
] = { MIGRATE_RESERVE
, MIGRATE_RESERVE
, MIGRATE_RESERVE
}, /* Never used */
835 * Move the free pages in a range to the free lists of the requested type.
836 * Note that start_page and end_pages are not aligned on a pageblock
837 * boundary. If alignment is required, use move_freepages_block()
839 static int move_freepages(struct zone
*zone
,
840 struct page
*start_page
, struct page
*end_page
,
847 #ifndef CONFIG_HOLES_IN_ZONE
849 * page_zone is not safe to call in this context when
850 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
851 * anyway as we check zone boundaries in move_freepages_block().
852 * Remove at a later date when no bug reports exist related to
853 * grouping pages by mobility
855 BUG_ON(page_zone(start_page
) != page_zone(end_page
));
858 for (page
= start_page
; page
<= end_page
;) {
859 /* Make sure we are not inadvertently changing nodes */
860 VM_BUG_ON(page_to_nid(page
) != zone_to_nid(zone
));
862 if (!pfn_valid_within(page_to_pfn(page
))) {
867 if (!PageBuddy(page
)) {
872 order
= page_order(page
);
873 list_del(&page
->lru
);
875 &zone
->free_area
[order
].free_list
[migratetype
]);
877 pages_moved
+= 1 << order
;
883 static int move_freepages_block(struct zone
*zone
, struct page
*page
,
886 unsigned long start_pfn
, end_pfn
;
887 struct page
*start_page
, *end_page
;
889 start_pfn
= page_to_pfn(page
);
890 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
891 start_page
= pfn_to_page(start_pfn
);
892 end_page
= start_page
+ pageblock_nr_pages
- 1;
893 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
895 /* Do not cross zone boundaries */
896 if (start_pfn
< zone
->zone_start_pfn
)
898 if (end_pfn
>= zone
->zone_start_pfn
+ zone
->spanned_pages
)
901 return move_freepages(zone
, start_page
, end_page
, migratetype
);
904 static void change_pageblock_range(struct page
*pageblock_page
,
905 int start_order
, int migratetype
)
907 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
909 while (nr_pageblocks
--) {
910 set_pageblock_migratetype(pageblock_page
, migratetype
);
911 pageblock_page
+= pageblock_nr_pages
;
915 /* Remove an element from the buddy allocator from the fallback list */
916 static inline struct page
*
917 __rmqueue_fallback(struct zone
*zone
, int order
, int start_migratetype
)
919 struct free_area
* area
;
924 /* Find the largest possible block of pages in the other list */
925 for (current_order
= MAX_ORDER
-1; current_order
>= order
;
927 for (i
= 0; i
< MIGRATE_TYPES
- 1; i
++) {
928 migratetype
= fallbacks
[start_migratetype
][i
];
930 /* MIGRATE_RESERVE handled later if necessary */
931 if (migratetype
== MIGRATE_RESERVE
)
934 area
= &(zone
->free_area
[current_order
]);
935 if (list_empty(&area
->free_list
[migratetype
]))
938 page
= list_entry(area
->free_list
[migratetype
].next
,
943 * If breaking a large block of pages, move all free
944 * pages to the preferred allocation list. If falling
945 * back for a reclaimable kernel allocation, be more
946 * agressive about taking ownership of free pages
948 if (unlikely(current_order
>= (pageblock_order
>> 1)) ||
949 start_migratetype
== MIGRATE_RECLAIMABLE
||
950 page_group_by_mobility_disabled
) {
952 pages
= move_freepages_block(zone
, page
,
955 /* Claim the whole block if over half of it is free */
956 if (pages
>= (1 << (pageblock_order
-1)) ||
957 page_group_by_mobility_disabled
)
958 set_pageblock_migratetype(page
,
961 migratetype
= start_migratetype
;
964 /* Remove the page from the freelists */
965 list_del(&page
->lru
);
966 rmv_page_order(page
);
968 /* Take ownership for orders >= pageblock_order */
969 if (current_order
>= pageblock_order
)
970 change_pageblock_range(page
, current_order
,
973 expand(zone
, page
, order
, current_order
, area
, migratetype
);
975 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
976 start_migratetype
, migratetype
);
986 * Do the hard work of removing an element from the buddy allocator.
987 * Call me with the zone->lock already held.
989 static struct page
*__rmqueue(struct zone
*zone
, unsigned int order
,
995 page
= __rmqueue_smallest(zone
, order
, migratetype
);
997 if (unlikely(!page
) && migratetype
!= MIGRATE_RESERVE
) {
998 page
= __rmqueue_fallback(zone
, order
, migratetype
);
1001 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1002 * is used because __rmqueue_smallest is an inline function
1003 * and we want just one call site
1006 migratetype
= MIGRATE_RESERVE
;
1011 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
1016 * Obtain a specified number of elements from the buddy allocator, all under
1017 * a single hold of the lock, for efficiency. Add them to the supplied list.
1018 * Returns the number of new pages which were placed at *list.
1020 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
1021 unsigned long count
, struct list_head
*list
,
1022 int migratetype
, int cold
)
1026 spin_lock(&zone
->lock
);
1027 for (i
= 0; i
< count
; ++i
) {
1028 struct page
*page
= __rmqueue(zone
, order
, migratetype
);
1029 if (unlikely(page
== NULL
))
1033 * Split buddy pages returned by expand() are received here
1034 * in physical page order. The page is added to the callers and
1035 * list and the list head then moves forward. From the callers
1036 * perspective, the linked list is ordered by page number in
1037 * some conditions. This is useful for IO devices that can
1038 * merge IO requests if the physical pages are ordered
1041 if (likely(cold
== 0))
1042 list_add(&page
->lru
, list
);
1044 list_add_tail(&page
->lru
, list
);
1045 set_page_private(page
, migratetype
);
1048 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
1049 spin_unlock(&zone
->lock
);
1055 * Called from the vmstat counter updater to drain pagesets of this
1056 * currently executing processor on remote nodes after they have
1059 * Note that this function must be called with the thread pinned to
1060 * a single processor.
1062 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
1064 unsigned long flags
;
1067 local_irq_save(flags
);
1068 if (pcp
->count
>= pcp
->batch
)
1069 to_drain
= pcp
->batch
;
1071 to_drain
= pcp
->count
;
1072 free_pcppages_bulk(zone
, to_drain
, pcp
);
1073 pcp
->count
-= to_drain
;
1074 local_irq_restore(flags
);
1079 * Drain pages of the indicated processor.
1081 * The processor must either be the current processor and the
1082 * thread pinned to the current processor or a processor that
1085 static void drain_pages(unsigned int cpu
)
1087 unsigned long flags
;
1090 for_each_populated_zone(zone
) {
1091 struct per_cpu_pageset
*pset
;
1092 struct per_cpu_pages
*pcp
;
1094 local_irq_save(flags
);
1095 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
1098 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
1100 local_irq_restore(flags
);
1105 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1107 void drain_local_pages(void *arg
)
1109 drain_pages(smp_processor_id());
1113 * Spill all the per-cpu pages from all CPUs back into the buddy allocator
1115 void drain_all_pages(void)
1117 on_each_cpu(drain_local_pages
, NULL
, 1);
1120 #ifdef CONFIG_HIBERNATION
1122 void mark_free_pages(struct zone
*zone
)
1124 unsigned long pfn
, max_zone_pfn
;
1125 unsigned long flags
;
1127 struct list_head
*curr
;
1129 if (!zone
->spanned_pages
)
1132 spin_lock_irqsave(&zone
->lock
, flags
);
1134 max_zone_pfn
= zone
->zone_start_pfn
+ zone
->spanned_pages
;
1135 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
1136 if (pfn_valid(pfn
)) {
1137 struct page
*page
= pfn_to_page(pfn
);
1139 if (!swsusp_page_is_forbidden(page
))
1140 swsusp_unset_page_free(page
);
1143 for_each_migratetype_order(order
, t
) {
1144 list_for_each(curr
, &zone
->free_area
[order
].free_list
[t
]) {
1147 pfn
= page_to_pfn(list_entry(curr
, struct page
, lru
));
1148 for (i
= 0; i
< (1UL << order
); i
++)
1149 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
1152 spin_unlock_irqrestore(&zone
->lock
, flags
);
1154 #endif /* CONFIG_PM */
1157 * Free a 0-order page
1158 * cold == 1 ? free a cold page : free a hot page
1160 void free_hot_cold_page(struct page
*page
, int cold
)
1162 struct zone
*zone
= page_zone(page
);
1163 struct per_cpu_pages
*pcp
;
1164 unsigned long flags
;
1166 int wasMlocked
= __TestClearPageMlocked(page
);
1168 if (!free_pages_prepare(page
, 0))
1171 migratetype
= get_pageblock_migratetype(page
);
1172 set_page_private(page
, migratetype
);
1173 local_irq_save(flags
);
1174 if (unlikely(wasMlocked
))
1175 free_page_mlock(page
);
1176 __count_vm_event(PGFREE
);
1179 * We only track unmovable, reclaimable and movable on pcp lists.
1180 * Free ISOLATE pages back to the allocator because they are being
1181 * offlined but treat RESERVE as movable pages so we can get those
1182 * areas back if necessary. Otherwise, we may have to free
1183 * excessively into the page allocator
1185 if (migratetype
>= MIGRATE_PCPTYPES
) {
1186 if (unlikely(migratetype
== MIGRATE_ISOLATE
)) {
1187 free_one_page(zone
, page
, 0, migratetype
);
1190 migratetype
= MIGRATE_MOVABLE
;
1193 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
1195 list_add_tail(&page
->lru
, &pcp
->lists
[migratetype
]);
1197 list_add(&page
->lru
, &pcp
->lists
[migratetype
]);
1199 if (pcp
->count
>= pcp
->high
) {
1200 free_pcppages_bulk(zone
, pcp
->batch
, pcp
);
1201 pcp
->count
-= pcp
->batch
;
1205 local_irq_restore(flags
);
1209 * split_page takes a non-compound higher-order page, and splits it into
1210 * n (1<<order) sub-pages: page[0..n]
1211 * Each sub-page must be freed individually.
1213 * Note: this is probably too low level an operation for use in drivers.
1214 * Please consult with lkml before using this in your driver.
1216 void split_page(struct page
*page
, unsigned int order
)
1220 VM_BUG_ON(PageCompound(page
));
1221 VM_BUG_ON(!page_count(page
));
1223 #ifdef CONFIG_KMEMCHECK
1225 * Split shadow pages too, because free(page[0]) would
1226 * otherwise free the whole shadow.
1228 if (kmemcheck_page_is_tracked(page
))
1229 split_page(virt_to_page(page
[0].shadow
), order
);
1232 for (i
= 1; i
< (1 << order
); i
++)
1233 set_page_refcounted(page
+ i
);
1237 * Similar to split_page except the page is already free. As this is only
1238 * being used for migration, the migratetype of the block also changes.
1239 * As this is called with interrupts disabled, the caller is responsible
1240 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1243 * Note: this is probably too low level an operation for use in drivers.
1244 * Please consult with lkml before using this in your driver.
1246 int split_free_page(struct page
*page
)
1249 unsigned long watermark
;
1252 BUG_ON(!PageBuddy(page
));
1254 zone
= page_zone(page
);
1255 order
= page_order(page
);
1257 /* Obey watermarks as if the page was being allocated */
1258 watermark
= low_wmark_pages(zone
) + (1 << order
);
1259 if (!zone_watermark_ok(zone
, 0, watermark
, 0, 0))
1262 /* Remove page from free list */
1263 list_del(&page
->lru
);
1264 zone
->free_area
[order
].nr_free
--;
1265 rmv_page_order(page
);
1266 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(1UL << order
));
1268 /* Split into individual pages */
1269 set_page_refcounted(page
);
1270 split_page(page
, order
);
1272 if (order
>= pageblock_order
- 1) {
1273 struct page
*endpage
= page
+ (1 << order
) - 1;
1274 for (; page
< endpage
; page
+= pageblock_nr_pages
)
1275 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
1282 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1283 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1287 struct page
*buffered_rmqueue(struct zone
*preferred_zone
,
1288 struct zone
*zone
, int order
, gfp_t gfp_flags
,
1291 unsigned long flags
;
1293 int cold
= !!(gfp_flags
& __GFP_COLD
);
1296 if (likely(order
== 0)) {
1297 struct per_cpu_pages
*pcp
;
1298 struct list_head
*list
;
1300 local_irq_save(flags
);
1301 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
1302 list
= &pcp
->lists
[migratetype
];
1303 if (list_empty(list
)) {
1304 pcp
->count
+= rmqueue_bulk(zone
, 0,
1307 if (unlikely(list_empty(list
)))
1312 page
= list_entry(list
->prev
, struct page
, lru
);
1314 page
= list_entry(list
->next
, struct page
, lru
);
1316 list_del(&page
->lru
);
1319 if (unlikely(gfp_flags
& __GFP_NOFAIL
)) {
1321 * __GFP_NOFAIL is not to be used in new code.
1323 * All __GFP_NOFAIL callers should be fixed so that they
1324 * properly detect and handle allocation failures.
1326 * We most definitely don't want callers attempting to
1327 * allocate greater than order-1 page units with
1330 WARN_ON_ONCE(order
> 1);
1332 spin_lock_irqsave(&zone
->lock
, flags
);
1333 page
= __rmqueue(zone
, order
, migratetype
);
1334 spin_unlock(&zone
->lock
);
1337 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(1 << order
));
1340 __count_zone_vm_events(PGALLOC
, zone
, 1 << order
);
1341 zone_statistics(preferred_zone
, zone
);
1342 local_irq_restore(flags
);
1344 VM_BUG_ON(bad_range(zone
, page
));
1345 if (prep_new_page(page
, order
, gfp_flags
))
1350 local_irq_restore(flags
);
1354 /* The ALLOC_WMARK bits are used as an index to zone->watermark */
1355 #define ALLOC_WMARK_MIN WMARK_MIN
1356 #define ALLOC_WMARK_LOW WMARK_LOW
1357 #define ALLOC_WMARK_HIGH WMARK_HIGH
1358 #define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */
1360 /* Mask to get the watermark bits */
1361 #define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1)
1363 #define ALLOC_HARDER 0x10 /* try to alloc harder */
1364 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
1365 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
1367 #ifdef CONFIG_FAIL_PAGE_ALLOC
1369 static struct fail_page_alloc_attr
{
1370 struct fault_attr attr
;
1372 u32 ignore_gfp_highmem
;
1373 u32 ignore_gfp_wait
;
1376 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1378 struct dentry
*ignore_gfp_highmem_file
;
1379 struct dentry
*ignore_gfp_wait_file
;
1380 struct dentry
*min_order_file
;
1382 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1384 } fail_page_alloc
= {
1385 .attr
= FAULT_ATTR_INITIALIZER
,
1386 .ignore_gfp_wait
= 1,
1387 .ignore_gfp_highmem
= 1,
1391 static int __init
setup_fail_page_alloc(char *str
)
1393 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
1395 __setup("fail_page_alloc=", setup_fail_page_alloc
);
1397 static int should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1399 if (order
< fail_page_alloc
.min_order
)
1401 if (gfp_mask
& __GFP_NOFAIL
)
1403 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
1405 if (fail_page_alloc
.ignore_gfp_wait
&& (gfp_mask
& __GFP_WAIT
))
1408 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
1411 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1413 static int __init
fail_page_alloc_debugfs(void)
1415 mode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
1419 err
= init_fault_attr_dentries(&fail_page_alloc
.attr
,
1423 dir
= fail_page_alloc
.attr
.dentries
.dir
;
1425 fail_page_alloc
.ignore_gfp_wait_file
=
1426 debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
1427 &fail_page_alloc
.ignore_gfp_wait
);
1429 fail_page_alloc
.ignore_gfp_highmem_file
=
1430 debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
1431 &fail_page_alloc
.ignore_gfp_highmem
);
1432 fail_page_alloc
.min_order_file
=
1433 debugfs_create_u32("min-order", mode
, dir
,
1434 &fail_page_alloc
.min_order
);
1436 if (!fail_page_alloc
.ignore_gfp_wait_file
||
1437 !fail_page_alloc
.ignore_gfp_highmem_file
||
1438 !fail_page_alloc
.min_order_file
) {
1440 debugfs_remove(fail_page_alloc
.ignore_gfp_wait_file
);
1441 debugfs_remove(fail_page_alloc
.ignore_gfp_highmem_file
);
1442 debugfs_remove(fail_page_alloc
.min_order_file
);
1443 cleanup_fault_attr_dentries(&fail_page_alloc
.attr
);
1449 late_initcall(fail_page_alloc_debugfs
);
1451 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1453 #else /* CONFIG_FAIL_PAGE_ALLOC */
1455 static inline int should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1460 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1463 * Return true if free pages are above 'mark'. This takes into account the order
1464 * of the allocation.
1466 static bool __zone_watermark_ok(struct zone
*z
, int order
, unsigned long mark
,
1467 int classzone_idx
, int alloc_flags
, long free_pages
)
1469 /* free_pages my go negative - that's OK */
1473 free_pages
-= (1 << order
) + 1;
1474 if (alloc_flags
& ALLOC_HIGH
)
1476 if (alloc_flags
& ALLOC_HARDER
)
1479 if (free_pages
<= min
+ z
->lowmem_reserve
[classzone_idx
])
1481 for (o
= 0; o
< order
; o
++) {
1482 /* At the next order, this order's pages become unavailable */
1483 free_pages
-= z
->free_area
[o
].nr_free
<< o
;
1485 /* Require fewer higher order pages to be free */
1488 if (free_pages
<= min
)
1494 bool zone_watermark_ok(struct zone
*z
, int order
, unsigned long mark
,
1495 int classzone_idx
, int alloc_flags
)
1497 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
1498 zone_page_state(z
, NR_FREE_PAGES
));
1501 bool zone_watermark_ok_safe(struct zone
*z
, int order
, unsigned long mark
,
1502 int classzone_idx
, int alloc_flags
)
1504 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
1506 if (z
->percpu_drift_mark
&& free_pages
< z
->percpu_drift_mark
)
1507 free_pages
= zone_page_state_snapshot(z
, NR_FREE_PAGES
);
1509 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
1515 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1516 * skip over zones that are not allowed by the cpuset, or that have
1517 * been recently (in last second) found to be nearly full. See further
1518 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1519 * that have to skip over a lot of full or unallowed zones.
1521 * If the zonelist cache is present in the passed in zonelist, then
1522 * returns a pointer to the allowed node mask (either the current
1523 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1525 * If the zonelist cache is not available for this zonelist, does
1526 * nothing and returns NULL.
1528 * If the fullzones BITMAP in the zonelist cache is stale (more than
1529 * a second since last zap'd) then we zap it out (clear its bits.)
1531 * We hold off even calling zlc_setup, until after we've checked the
1532 * first zone in the zonelist, on the theory that most allocations will
1533 * be satisfied from that first zone, so best to examine that zone as
1534 * quickly as we can.
1536 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1538 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1539 nodemask_t
*allowednodes
; /* zonelist_cache approximation */
1541 zlc
= zonelist
->zlcache_ptr
;
1545 if (time_after(jiffies
, zlc
->last_full_zap
+ HZ
)) {
1546 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
1547 zlc
->last_full_zap
= jiffies
;
1550 allowednodes
= !in_interrupt() && (alloc_flags
& ALLOC_CPUSET
) ?
1551 &cpuset_current_mems_allowed
:
1552 &node_states
[N_HIGH_MEMORY
];
1553 return allowednodes
;
1557 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1558 * if it is worth looking at further for free memory:
1559 * 1) Check that the zone isn't thought to be full (doesn't have its
1560 * bit set in the zonelist_cache fullzones BITMAP).
1561 * 2) Check that the zones node (obtained from the zonelist_cache
1562 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1563 * Return true (non-zero) if zone is worth looking at further, or
1564 * else return false (zero) if it is not.
1566 * This check -ignores- the distinction between various watermarks,
1567 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1568 * found to be full for any variation of these watermarks, it will
1569 * be considered full for up to one second by all requests, unless
1570 * we are so low on memory on all allowed nodes that we are forced
1571 * into the second scan of the zonelist.
1573 * In the second scan we ignore this zonelist cache and exactly
1574 * apply the watermarks to all zones, even it is slower to do so.
1575 * We are low on memory in the second scan, and should leave no stone
1576 * unturned looking for a free page.
1578 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1579 nodemask_t
*allowednodes
)
1581 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1582 int i
; /* index of *z in zonelist zones */
1583 int n
; /* node that zone *z is on */
1585 zlc
= zonelist
->zlcache_ptr
;
1589 i
= z
- zonelist
->_zonerefs
;
1592 /* This zone is worth trying if it is allowed but not full */
1593 return node_isset(n
, *allowednodes
) && !test_bit(i
, zlc
->fullzones
);
1597 * Given 'z' scanning a zonelist, set the corresponding bit in
1598 * zlc->fullzones, so that subsequent attempts to allocate a page
1599 * from that zone don't waste time re-examining it.
1601 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1603 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1604 int i
; /* index of *z in zonelist zones */
1606 zlc
= zonelist
->zlcache_ptr
;
1610 i
= z
- zonelist
->_zonerefs
;
1612 set_bit(i
, zlc
->fullzones
);
1615 #else /* CONFIG_NUMA */
1617 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1622 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1623 nodemask_t
*allowednodes
)
1628 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1631 #endif /* CONFIG_NUMA */
1634 * get_page_from_freelist goes through the zonelist trying to allocate
1637 static struct page
*
1638 get_page_from_freelist(gfp_t gfp_mask
, nodemask_t
*nodemask
, unsigned int order
,
1639 struct zonelist
*zonelist
, int high_zoneidx
, int alloc_flags
,
1640 struct zone
*preferred_zone
, int migratetype
)
1643 struct page
*page
= NULL
;
1646 nodemask_t
*allowednodes
= NULL
;/* zonelist_cache approximation */
1647 int zlc_active
= 0; /* set if using zonelist_cache */
1648 int did_zlc_setup
= 0; /* just call zlc_setup() one time */
1650 classzone_idx
= zone_idx(preferred_zone
);
1653 * Scan zonelist, looking for a zone with enough free.
1654 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1656 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
1657 high_zoneidx
, nodemask
) {
1658 if (NUMA_BUILD
&& zlc_active
&&
1659 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
1661 if ((alloc_flags
& ALLOC_CPUSET
) &&
1662 !cpuset_zone_allowed_softwall(zone
, gfp_mask
))
1665 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
1666 if (!(alloc_flags
& ALLOC_NO_WATERMARKS
)) {
1670 mark
= zone
->watermark
[alloc_flags
& ALLOC_WMARK_MASK
];
1671 if (zone_watermark_ok(zone
, order
, mark
,
1672 classzone_idx
, alloc_flags
))
1675 if (zone_reclaim_mode
== 0)
1676 goto this_zone_full
;
1678 ret
= zone_reclaim(zone
, gfp_mask
, order
);
1680 case ZONE_RECLAIM_NOSCAN
:
1683 case ZONE_RECLAIM_FULL
:
1684 /* scanned but unreclaimable */
1685 goto this_zone_full
;
1687 /* did we reclaim enough */
1688 if (!zone_watermark_ok(zone
, order
, mark
,
1689 classzone_idx
, alloc_flags
))
1690 goto this_zone_full
;
1695 page
= buffered_rmqueue(preferred_zone
, zone
, order
,
1696 gfp_mask
, migratetype
);
1701 zlc_mark_zone_full(zonelist
, z
);
1703 if (NUMA_BUILD
&& !did_zlc_setup
&& nr_online_nodes
> 1) {
1705 * we do zlc_setup after the first zone is tried but only
1706 * if there are multiple nodes make it worthwhile
1708 allowednodes
= zlc_setup(zonelist
, alloc_flags
);
1714 if (unlikely(NUMA_BUILD
&& page
== NULL
&& zlc_active
)) {
1715 /* Disable zlc cache for second zonelist scan */
1723 should_alloc_retry(gfp_t gfp_mask
, unsigned int order
,
1724 unsigned long pages_reclaimed
)
1726 /* Do not loop if specifically requested */
1727 if (gfp_mask
& __GFP_NORETRY
)
1731 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
1732 * means __GFP_NOFAIL, but that may not be true in other
1735 if (order
<= PAGE_ALLOC_COSTLY_ORDER
)
1739 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
1740 * specified, then we retry until we no longer reclaim any pages
1741 * (above), or we've reclaimed an order of pages at least as
1742 * large as the allocation's order. In both cases, if the
1743 * allocation still fails, we stop retrying.
1745 if (gfp_mask
& __GFP_REPEAT
&& pages_reclaimed
< (1 << order
))
1749 * Don't let big-order allocations loop unless the caller
1750 * explicitly requests that.
1752 if (gfp_mask
& __GFP_NOFAIL
)
1758 static inline struct page
*
1759 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
1760 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1761 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
1766 /* Acquire the OOM killer lock for the zones in zonelist */
1767 if (!try_set_zonelist_oom(zonelist
, gfp_mask
)) {
1768 schedule_timeout_uninterruptible(1);
1773 * Go through the zonelist yet one more time, keep very high watermark
1774 * here, this is only to catch a parallel oom killing, we must fail if
1775 * we're still under heavy pressure.
1777 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
,
1778 order
, zonelist
, high_zoneidx
,
1779 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
,
1780 preferred_zone
, migratetype
);
1784 if (!(gfp_mask
& __GFP_NOFAIL
)) {
1785 /* The OOM killer will not help higher order allocs */
1786 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
1788 /* The OOM killer does not needlessly kill tasks for lowmem */
1789 if (high_zoneidx
< ZONE_NORMAL
)
1792 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
1793 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
1794 * The caller should handle page allocation failure by itself if
1795 * it specifies __GFP_THISNODE.
1796 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
1798 if (gfp_mask
& __GFP_THISNODE
)
1801 /* Exhausted what can be done so it's blamo time */
1802 out_of_memory(zonelist
, gfp_mask
, order
, nodemask
);
1805 clear_zonelist_oom(zonelist
, gfp_mask
);
1809 #ifdef CONFIG_COMPACTION
1810 /* Try memory compaction for high-order allocations before reclaim */
1811 static struct page
*
1812 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
1813 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1814 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
1815 int migratetype
, unsigned long *did_some_progress
,
1816 bool sync_migration
)
1819 struct task_struct
*tsk
= current
;
1821 if (!order
|| compaction_deferred(preferred_zone
))
1824 tsk
->flags
|= PF_MEMALLOC
;
1825 *did_some_progress
= try_to_compact_pages(zonelist
, order
, gfp_mask
,
1826 nodemask
, sync_migration
);
1827 tsk
->flags
&= ~PF_MEMALLOC
;
1828 if (*did_some_progress
!= COMPACT_SKIPPED
) {
1830 /* Page migration frees to the PCP lists but we want merging */
1831 drain_pages(get_cpu());
1834 page
= get_page_from_freelist(gfp_mask
, nodemask
,
1835 order
, zonelist
, high_zoneidx
,
1836 alloc_flags
, preferred_zone
,
1839 preferred_zone
->compact_considered
= 0;
1840 preferred_zone
->compact_defer_shift
= 0;
1841 count_vm_event(COMPACTSUCCESS
);
1846 * It's bad if compaction run occurs and fails.
1847 * The most likely reason is that pages exist,
1848 * but not enough to satisfy watermarks.
1850 count_vm_event(COMPACTFAIL
);
1851 defer_compaction(preferred_zone
);
1859 static inline struct page
*
1860 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
1861 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1862 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
1863 int migratetype
, unsigned long *did_some_progress
,
1864 bool sync_migration
)
1868 #endif /* CONFIG_COMPACTION */
1870 /* The really slow allocator path where we enter direct reclaim */
1871 static inline struct page
*
1872 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
1873 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1874 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
1875 int migratetype
, unsigned long *did_some_progress
)
1877 struct page
*page
= NULL
;
1878 struct reclaim_state reclaim_state
;
1879 struct task_struct
*p
= current
;
1880 bool drained
= false;
1884 /* We now go into synchronous reclaim */
1885 cpuset_memory_pressure_bump();
1886 p
->flags
|= PF_MEMALLOC
;
1887 lockdep_set_current_reclaim_state(gfp_mask
);
1888 reclaim_state
.reclaimed_slab
= 0;
1889 p
->reclaim_state
= &reclaim_state
;
1891 *did_some_progress
= try_to_free_pages(zonelist
, order
, gfp_mask
, nodemask
);
1893 p
->reclaim_state
= NULL
;
1894 lockdep_clear_current_reclaim_state();
1895 p
->flags
&= ~PF_MEMALLOC
;
1899 if (unlikely(!(*did_some_progress
)))
1903 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
1904 zonelist
, high_zoneidx
,
1905 alloc_flags
, preferred_zone
,
1909 * If an allocation failed after direct reclaim, it could be because
1910 * pages are pinned on the per-cpu lists. Drain them and try again
1912 if (!page
&& !drained
) {
1922 * This is called in the allocator slow-path if the allocation request is of
1923 * sufficient urgency to ignore watermarks and take other desperate measures
1925 static inline struct page
*
1926 __alloc_pages_high_priority(gfp_t gfp_mask
, unsigned int order
,
1927 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1928 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
1934 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
1935 zonelist
, high_zoneidx
, ALLOC_NO_WATERMARKS
,
1936 preferred_zone
, migratetype
);
1938 if (!page
&& gfp_mask
& __GFP_NOFAIL
)
1939 wait_iff_congested(preferred_zone
, BLK_RW_ASYNC
, HZ
/50);
1940 } while (!page
&& (gfp_mask
& __GFP_NOFAIL
));
1946 void wake_all_kswapd(unsigned int order
, struct zonelist
*zonelist
,
1947 enum zone_type high_zoneidx
,
1948 enum zone_type classzone_idx
)
1953 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
)
1954 wakeup_kswapd(zone
, order
, classzone_idx
);
1958 gfp_to_alloc_flags(gfp_t gfp_mask
)
1960 struct task_struct
*p
= current
;
1961 int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
1962 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
1964 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
1965 BUILD_BUG_ON(__GFP_HIGH
!= (__force gfp_t
) ALLOC_HIGH
);
1968 * The caller may dip into page reserves a bit more if the caller
1969 * cannot run direct reclaim, or if the caller has realtime scheduling
1970 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
1971 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
1973 alloc_flags
|= (__force
int) (gfp_mask
& __GFP_HIGH
);
1976 alloc_flags
|= ALLOC_HARDER
;
1978 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
1979 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1981 alloc_flags
&= ~ALLOC_CPUSET
;
1982 } else if (unlikely(rt_task(p
)) && !in_interrupt())
1983 alloc_flags
|= ALLOC_HARDER
;
1985 if (likely(!(gfp_mask
& __GFP_NOMEMALLOC
))) {
1986 if (!in_interrupt() &&
1987 ((p
->flags
& PF_MEMALLOC
) ||
1988 unlikely(test_thread_flag(TIF_MEMDIE
))))
1989 alloc_flags
|= ALLOC_NO_WATERMARKS
;
1995 static inline struct page
*
1996 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
1997 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1998 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
2001 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
2002 struct page
*page
= NULL
;
2004 unsigned long pages_reclaimed
= 0;
2005 unsigned long did_some_progress
;
2006 struct task_struct
*p
= current
;
2007 bool sync_migration
= false;
2010 * In the slowpath, we sanity check order to avoid ever trying to
2011 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2012 * be using allocators in order of preference for an area that is
2015 if (order
>= MAX_ORDER
) {
2016 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
2021 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2022 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2023 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2024 * using a larger set of nodes after it has established that the
2025 * allowed per node queues are empty and that nodes are
2028 if (NUMA_BUILD
&& (gfp_mask
& GFP_THISNODE
) == GFP_THISNODE
)
2032 wake_all_kswapd(order
, zonelist
, high_zoneidx
,
2033 zone_idx(preferred_zone
));
2036 * OK, we're below the kswapd watermark and have kicked background
2037 * reclaim. Now things get more complex, so set up alloc_flags according
2038 * to how we want to proceed.
2040 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
2042 /* This is the last chance, in general, before the goto nopage. */
2043 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
, zonelist
,
2044 high_zoneidx
, alloc_flags
& ~ALLOC_NO_WATERMARKS
,
2045 preferred_zone
, migratetype
);
2050 /* Allocate without watermarks if the context allows */
2051 if (alloc_flags
& ALLOC_NO_WATERMARKS
) {
2052 page
= __alloc_pages_high_priority(gfp_mask
, order
,
2053 zonelist
, high_zoneidx
, nodemask
,
2054 preferred_zone
, migratetype
);
2059 /* Atomic allocations - we can't balance anything */
2063 /* Avoid recursion of direct reclaim */
2064 if (p
->flags
& PF_MEMALLOC
)
2067 /* Avoid allocations with no watermarks from looping endlessly */
2068 if (test_thread_flag(TIF_MEMDIE
) && !(gfp_mask
& __GFP_NOFAIL
))
2072 * Try direct compaction. The first pass is asynchronous. Subsequent
2073 * attempts after direct reclaim are synchronous
2075 page
= __alloc_pages_direct_compact(gfp_mask
, order
,
2076 zonelist
, high_zoneidx
,
2078 alloc_flags
, preferred_zone
,
2079 migratetype
, &did_some_progress
,
2083 sync_migration
= true;
2085 /* Try direct reclaim and then allocating */
2086 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
,
2087 zonelist
, high_zoneidx
,
2089 alloc_flags
, preferred_zone
,
2090 migratetype
, &did_some_progress
);
2095 * If we failed to make any progress reclaiming, then we are
2096 * running out of options and have to consider going OOM
2098 if (!did_some_progress
) {
2099 if ((gfp_mask
& __GFP_FS
) && !(gfp_mask
& __GFP_NORETRY
)) {
2100 if (oom_killer_disabled
)
2102 page
= __alloc_pages_may_oom(gfp_mask
, order
,
2103 zonelist
, high_zoneidx
,
2104 nodemask
, preferred_zone
,
2109 if (!(gfp_mask
& __GFP_NOFAIL
)) {
2111 * The oom killer is not called for high-order
2112 * allocations that may fail, so if no progress
2113 * is being made, there are no other options and
2114 * retrying is unlikely to help.
2116 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
2119 * The oom killer is not called for lowmem
2120 * allocations to prevent needlessly killing
2123 if (high_zoneidx
< ZONE_NORMAL
)
2131 /* Check if we should retry the allocation */
2132 pages_reclaimed
+= did_some_progress
;
2133 if (should_alloc_retry(gfp_mask
, order
, pages_reclaimed
)) {
2134 /* Wait for some write requests to complete then retry */
2135 wait_iff_congested(preferred_zone
, BLK_RW_ASYNC
, HZ
/50);
2139 * High-order allocations do not necessarily loop after
2140 * direct reclaim and reclaim/compaction depends on compaction
2141 * being called after reclaim so call directly if necessary
2143 page
= __alloc_pages_direct_compact(gfp_mask
, order
,
2144 zonelist
, high_zoneidx
,
2146 alloc_flags
, preferred_zone
,
2147 migratetype
, &did_some_progress
,
2154 if (!(gfp_mask
& __GFP_NOWARN
) && printk_ratelimit()) {
2155 printk(KERN_WARNING
"%s: page allocation failure."
2156 " order:%d, mode:0x%x\n",
2157 p
->comm
, order
, gfp_mask
);
2163 if (kmemcheck_enabled
)
2164 kmemcheck_pagealloc_alloc(page
, order
, gfp_mask
);
2170 * This is the 'heart' of the zoned buddy allocator.
2173 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
,
2174 struct zonelist
*zonelist
, nodemask_t
*nodemask
)
2176 enum zone_type high_zoneidx
= gfp_zone(gfp_mask
);
2177 struct zone
*preferred_zone
;
2179 int migratetype
= allocflags_to_migratetype(gfp_mask
);
2181 gfp_mask
&= gfp_allowed_mask
;
2183 lockdep_trace_alloc(gfp_mask
);
2185 might_sleep_if(gfp_mask
& __GFP_WAIT
);
2187 if (should_fail_alloc_page(gfp_mask
, order
))
2191 * Check the zones suitable for the gfp_mask contain at least one
2192 * valid zone. It's possible to have an empty zonelist as a result
2193 * of GFP_THISNODE and a memoryless node
2195 if (unlikely(!zonelist
->_zonerefs
->zone
))
2199 /* The preferred zone is used for statistics later */
2200 first_zones_zonelist(zonelist
, high_zoneidx
, nodemask
, &preferred_zone
);
2201 if (!preferred_zone
) {
2206 /* First allocation attempt */
2207 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
, order
,
2208 zonelist
, high_zoneidx
, ALLOC_WMARK_LOW
|ALLOC_CPUSET
,
2209 preferred_zone
, migratetype
);
2210 if (unlikely(!page
))
2211 page
= __alloc_pages_slowpath(gfp_mask
, order
,
2212 zonelist
, high_zoneidx
, nodemask
,
2213 preferred_zone
, migratetype
);
2216 trace_mm_page_alloc(page
, order
, gfp_mask
, migratetype
);
2219 EXPORT_SYMBOL(__alloc_pages_nodemask
);
2222 * Common helper functions.
2224 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
2229 * __get_free_pages() returns a 32-bit address, which cannot represent
2232 VM_BUG_ON((gfp_mask
& __GFP_HIGHMEM
) != 0);
2234 page
= alloc_pages(gfp_mask
, order
);
2237 return (unsigned long) page_address(page
);
2239 EXPORT_SYMBOL(__get_free_pages
);
2241 unsigned long get_zeroed_page(gfp_t gfp_mask
)
2243 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
2245 EXPORT_SYMBOL(get_zeroed_page
);
2247 void __pagevec_free(struct pagevec
*pvec
)
2249 int i
= pagevec_count(pvec
);
2252 trace_mm_pagevec_free(pvec
->pages
[i
], pvec
->cold
);
2253 free_hot_cold_page(pvec
->pages
[i
], pvec
->cold
);
2257 void __free_pages(struct page
*page
, unsigned int order
)
2259 if (put_page_testzero(page
)) {
2261 free_hot_cold_page(page
, 0);
2263 __free_pages_ok(page
, order
);
2267 EXPORT_SYMBOL(__free_pages
);
2269 void free_pages(unsigned long addr
, unsigned int order
)
2272 VM_BUG_ON(!virt_addr_valid((void *)addr
));
2273 __free_pages(virt_to_page((void *)addr
), order
);
2277 EXPORT_SYMBOL(free_pages
);
2280 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2281 * @size: the number of bytes to allocate
2282 * @gfp_mask: GFP flags for the allocation
2284 * This function is similar to alloc_pages(), except that it allocates the
2285 * minimum number of pages to satisfy the request. alloc_pages() can only
2286 * allocate memory in power-of-two pages.
2288 * This function is also limited by MAX_ORDER.
2290 * Memory allocated by this function must be released by free_pages_exact().
2292 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
2294 unsigned int order
= get_order(size
);
2297 addr
= __get_free_pages(gfp_mask
, order
);
2299 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
2300 unsigned long used
= addr
+ PAGE_ALIGN(size
);
2302 split_page(virt_to_page((void *)addr
), order
);
2303 while (used
< alloc_end
) {
2309 return (void *)addr
;
2311 EXPORT_SYMBOL(alloc_pages_exact
);
2314 * free_pages_exact - release memory allocated via alloc_pages_exact()
2315 * @virt: the value returned by alloc_pages_exact.
2316 * @size: size of allocation, same value as passed to alloc_pages_exact().
2318 * Release the memory allocated by a previous call to alloc_pages_exact.
2320 void free_pages_exact(void *virt
, size_t size
)
2322 unsigned long addr
= (unsigned long)virt
;
2323 unsigned long end
= addr
+ PAGE_ALIGN(size
);
2325 while (addr
< end
) {
2330 EXPORT_SYMBOL(free_pages_exact
);
2332 static unsigned int nr_free_zone_pages(int offset
)
2337 /* Just pick one node, since fallback list is circular */
2338 unsigned int sum
= 0;
2340 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
2342 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
2343 unsigned long size
= zone
->present_pages
;
2344 unsigned long high
= high_wmark_pages(zone
);
2353 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2355 unsigned int nr_free_buffer_pages(void)
2357 return nr_free_zone_pages(gfp_zone(GFP_USER
));
2359 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
2362 * Amount of free RAM allocatable within all zones
2364 unsigned int nr_free_pagecache_pages(void)
2366 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
2369 static inline void show_node(struct zone
*zone
)
2372 printk("Node %d ", zone_to_nid(zone
));
2375 void si_meminfo(struct sysinfo
*val
)
2377 val
->totalram
= totalram_pages
;
2379 val
->freeram
= global_page_state(NR_FREE_PAGES
);
2380 val
->bufferram
= nr_blockdev_pages();
2381 val
->totalhigh
= totalhigh_pages
;
2382 val
->freehigh
= nr_free_highpages();
2383 val
->mem_unit
= PAGE_SIZE
;
2386 EXPORT_SYMBOL(si_meminfo
);
2389 void si_meminfo_node(struct sysinfo
*val
, int nid
)
2391 pg_data_t
*pgdat
= NODE_DATA(nid
);
2393 val
->totalram
= pgdat
->node_present_pages
;
2394 val
->freeram
= node_page_state(nid
, NR_FREE_PAGES
);
2395 #ifdef CONFIG_HIGHMEM
2396 val
->totalhigh
= pgdat
->node_zones
[ZONE_HIGHMEM
].present_pages
;
2397 val
->freehigh
= zone_page_state(&pgdat
->node_zones
[ZONE_HIGHMEM
],
2403 val
->mem_unit
= PAGE_SIZE
;
2407 #define K(x) ((x) << (PAGE_SHIFT-10))
2410 * Show free area list (used inside shift_scroll-lock stuff)
2411 * We also calculate the percentage fragmentation. We do this by counting the
2412 * memory on each free list with the exception of the first item on the list.
2414 void show_free_areas(void)
2419 for_each_populated_zone(zone
) {
2421 printk("%s per-cpu:\n", zone
->name
);
2423 for_each_online_cpu(cpu
) {
2424 struct per_cpu_pageset
*pageset
;
2426 pageset
= per_cpu_ptr(zone
->pageset
, cpu
);
2428 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2429 cpu
, pageset
->pcp
.high
,
2430 pageset
->pcp
.batch
, pageset
->pcp
.count
);
2434 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2435 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2437 " dirty:%lu writeback:%lu unstable:%lu\n"
2438 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2439 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n",
2440 global_page_state(NR_ACTIVE_ANON
),
2441 global_page_state(NR_INACTIVE_ANON
),
2442 global_page_state(NR_ISOLATED_ANON
),
2443 global_page_state(NR_ACTIVE_FILE
),
2444 global_page_state(NR_INACTIVE_FILE
),
2445 global_page_state(NR_ISOLATED_FILE
),
2446 global_page_state(NR_UNEVICTABLE
),
2447 global_page_state(NR_FILE_DIRTY
),
2448 global_page_state(NR_WRITEBACK
),
2449 global_page_state(NR_UNSTABLE_NFS
),
2450 global_page_state(NR_FREE_PAGES
),
2451 global_page_state(NR_SLAB_RECLAIMABLE
),
2452 global_page_state(NR_SLAB_UNRECLAIMABLE
),
2453 global_page_state(NR_FILE_MAPPED
),
2454 global_page_state(NR_SHMEM
),
2455 global_page_state(NR_PAGETABLE
),
2456 global_page_state(NR_BOUNCE
));
2458 for_each_populated_zone(zone
) {
2467 " active_anon:%lukB"
2468 " inactive_anon:%lukB"
2469 " active_file:%lukB"
2470 " inactive_file:%lukB"
2471 " unevictable:%lukB"
2472 " isolated(anon):%lukB"
2473 " isolated(file):%lukB"
2480 " slab_reclaimable:%lukB"
2481 " slab_unreclaimable:%lukB"
2482 " kernel_stack:%lukB"
2486 " writeback_tmp:%lukB"
2487 " pages_scanned:%lu"
2488 " all_unreclaimable? %s"
2491 K(zone_page_state(zone
, NR_FREE_PAGES
)),
2492 K(min_wmark_pages(zone
)),
2493 K(low_wmark_pages(zone
)),
2494 K(high_wmark_pages(zone
)),
2495 K(zone_page_state(zone
, NR_ACTIVE_ANON
)),
2496 K(zone_page_state(zone
, NR_INACTIVE_ANON
)),
2497 K(zone_page_state(zone
, NR_ACTIVE_FILE
)),
2498 K(zone_page_state(zone
, NR_INACTIVE_FILE
)),
2499 K(zone_page_state(zone
, NR_UNEVICTABLE
)),
2500 K(zone_page_state(zone
, NR_ISOLATED_ANON
)),
2501 K(zone_page_state(zone
, NR_ISOLATED_FILE
)),
2502 K(zone
->present_pages
),
2503 K(zone_page_state(zone
, NR_MLOCK
)),
2504 K(zone_page_state(zone
, NR_FILE_DIRTY
)),
2505 K(zone_page_state(zone
, NR_WRITEBACK
)),
2506 K(zone_page_state(zone
, NR_FILE_MAPPED
)),
2507 K(zone_page_state(zone
, NR_SHMEM
)),
2508 K(zone_page_state(zone
, NR_SLAB_RECLAIMABLE
)),
2509 K(zone_page_state(zone
, NR_SLAB_UNRECLAIMABLE
)),
2510 zone_page_state(zone
, NR_KERNEL_STACK
) *
2512 K(zone_page_state(zone
, NR_PAGETABLE
)),
2513 K(zone_page_state(zone
, NR_UNSTABLE_NFS
)),
2514 K(zone_page_state(zone
, NR_BOUNCE
)),
2515 K(zone_page_state(zone
, NR_WRITEBACK_TEMP
)),
2516 zone
->pages_scanned
,
2517 (zone
->all_unreclaimable
? "yes" : "no")
2519 printk("lowmem_reserve[]:");
2520 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
2521 printk(" %lu", zone
->lowmem_reserve
[i
]);
2525 for_each_populated_zone(zone
) {
2526 unsigned long nr
[MAX_ORDER
], flags
, order
, total
= 0;
2529 printk("%s: ", zone
->name
);
2531 spin_lock_irqsave(&zone
->lock
, flags
);
2532 for (order
= 0; order
< MAX_ORDER
; order
++) {
2533 nr
[order
] = zone
->free_area
[order
].nr_free
;
2534 total
+= nr
[order
] << order
;
2536 spin_unlock_irqrestore(&zone
->lock
, flags
);
2537 for (order
= 0; order
< MAX_ORDER
; order
++)
2538 printk("%lu*%lukB ", nr
[order
], K(1UL) << order
);
2539 printk("= %lukB\n", K(total
));
2542 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES
));
2544 show_swap_cache_info();
2547 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
2549 zoneref
->zone
= zone
;
2550 zoneref
->zone_idx
= zone_idx(zone
);
2554 * Builds allocation fallback zone lists.
2556 * Add all populated zones of a node to the zonelist.
2558 static int build_zonelists_node(pg_data_t
*pgdat
, struct zonelist
*zonelist
,
2559 int nr_zones
, enum zone_type zone_type
)
2563 BUG_ON(zone_type
>= MAX_NR_ZONES
);
2568 zone
= pgdat
->node_zones
+ zone_type
;
2569 if (populated_zone(zone
)) {
2570 zoneref_set_zone(zone
,
2571 &zonelist
->_zonerefs
[nr_zones
++]);
2572 check_highest_zone(zone_type
);
2575 } while (zone_type
);
2582 * 0 = automatic detection of better ordering.
2583 * 1 = order by ([node] distance, -zonetype)
2584 * 2 = order by (-zonetype, [node] distance)
2586 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
2587 * the same zonelist. So only NUMA can configure this param.
2589 #define ZONELIST_ORDER_DEFAULT 0
2590 #define ZONELIST_ORDER_NODE 1
2591 #define ZONELIST_ORDER_ZONE 2
2593 /* zonelist order in the kernel.
2594 * set_zonelist_order() will set this to NODE or ZONE.
2596 static int current_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
2597 static char zonelist_order_name
[3][8] = {"Default", "Node", "Zone"};
2601 /* The value user specified ....changed by config */
2602 static int user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
2603 /* string for sysctl */
2604 #define NUMA_ZONELIST_ORDER_LEN 16
2605 char numa_zonelist_order
[16] = "default";
2608 * interface for configure zonelist ordering.
2609 * command line option "numa_zonelist_order"
2610 * = "[dD]efault - default, automatic configuration.
2611 * = "[nN]ode - order by node locality, then by zone within node
2612 * = "[zZ]one - order by zone, then by locality within zone
2615 static int __parse_numa_zonelist_order(char *s
)
2617 if (*s
== 'd' || *s
== 'D') {
2618 user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
2619 } else if (*s
== 'n' || *s
== 'N') {
2620 user_zonelist_order
= ZONELIST_ORDER_NODE
;
2621 } else if (*s
== 'z' || *s
== 'Z') {
2622 user_zonelist_order
= ZONELIST_ORDER_ZONE
;
2625 "Ignoring invalid numa_zonelist_order value: "
2632 static __init
int setup_numa_zonelist_order(char *s
)
2635 return __parse_numa_zonelist_order(s
);
2638 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
2641 * sysctl handler for numa_zonelist_order
2643 int numa_zonelist_order_handler(ctl_table
*table
, int write
,
2644 void __user
*buffer
, size_t *length
,
2647 char saved_string
[NUMA_ZONELIST_ORDER_LEN
];
2649 static DEFINE_MUTEX(zl_order_mutex
);
2651 mutex_lock(&zl_order_mutex
);
2653 strcpy(saved_string
, (char*)table
->data
);
2654 ret
= proc_dostring(table
, write
, buffer
, length
, ppos
);
2658 int oldval
= user_zonelist_order
;
2659 if (__parse_numa_zonelist_order((char*)table
->data
)) {
2661 * bogus value. restore saved string
2663 strncpy((char*)table
->data
, saved_string
,
2664 NUMA_ZONELIST_ORDER_LEN
);
2665 user_zonelist_order
= oldval
;
2666 } else if (oldval
!= user_zonelist_order
) {
2667 mutex_lock(&zonelists_mutex
);
2668 build_all_zonelists(NULL
);
2669 mutex_unlock(&zonelists_mutex
);
2673 mutex_unlock(&zl_order_mutex
);
2678 #define MAX_NODE_LOAD (nr_online_nodes)
2679 static int node_load
[MAX_NUMNODES
];
2682 * find_next_best_node - find the next node that should appear in a given node's fallback list
2683 * @node: node whose fallback list we're appending
2684 * @used_node_mask: nodemask_t of already used nodes
2686 * We use a number of factors to determine which is the next node that should
2687 * appear on a given node's fallback list. The node should not have appeared
2688 * already in @node's fallback list, and it should be the next closest node
2689 * according to the distance array (which contains arbitrary distance values
2690 * from each node to each node in the system), and should also prefer nodes
2691 * with no CPUs, since presumably they'll have very little allocation pressure
2692 * on them otherwise.
2693 * It returns -1 if no node is found.
2695 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
2698 int min_val
= INT_MAX
;
2700 const struct cpumask
*tmp
= cpumask_of_node(0);
2702 /* Use the local node if we haven't already */
2703 if (!node_isset(node
, *used_node_mask
)) {
2704 node_set(node
, *used_node_mask
);
2708 for_each_node_state(n
, N_HIGH_MEMORY
) {
2710 /* Don't want a node to appear more than once */
2711 if (node_isset(n
, *used_node_mask
))
2714 /* Use the distance array to find the distance */
2715 val
= node_distance(node
, n
);
2717 /* Penalize nodes under us ("prefer the next node") */
2720 /* Give preference to headless and unused nodes */
2721 tmp
= cpumask_of_node(n
);
2722 if (!cpumask_empty(tmp
))
2723 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
2725 /* Slight preference for less loaded node */
2726 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
2727 val
+= node_load
[n
];
2729 if (val
< min_val
) {
2736 node_set(best_node
, *used_node_mask
);
2743 * Build zonelists ordered by node and zones within node.
2744 * This results in maximum locality--normal zone overflows into local
2745 * DMA zone, if any--but risks exhausting DMA zone.
2747 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int node
)
2750 struct zonelist
*zonelist
;
2752 zonelist
= &pgdat
->node_zonelists
[0];
2753 for (j
= 0; zonelist
->_zonerefs
[j
].zone
!= NULL
; j
++)
2755 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
2757 zonelist
->_zonerefs
[j
].zone
= NULL
;
2758 zonelist
->_zonerefs
[j
].zone_idx
= 0;
2762 * Build gfp_thisnode zonelists
2764 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
2767 struct zonelist
*zonelist
;
2769 zonelist
= &pgdat
->node_zonelists
[1];
2770 j
= build_zonelists_node(pgdat
, zonelist
, 0, MAX_NR_ZONES
- 1);
2771 zonelist
->_zonerefs
[j
].zone
= NULL
;
2772 zonelist
->_zonerefs
[j
].zone_idx
= 0;
2776 * Build zonelists ordered by zone and nodes within zones.
2777 * This results in conserving DMA zone[s] until all Normal memory is
2778 * exhausted, but results in overflowing to remote node while memory
2779 * may still exist in local DMA zone.
2781 static int node_order
[MAX_NUMNODES
];
2783 static void build_zonelists_in_zone_order(pg_data_t
*pgdat
, int nr_nodes
)
2786 int zone_type
; /* needs to be signed */
2788 struct zonelist
*zonelist
;
2790 zonelist
= &pgdat
->node_zonelists
[0];
2792 for (zone_type
= MAX_NR_ZONES
- 1; zone_type
>= 0; zone_type
--) {
2793 for (j
= 0; j
< nr_nodes
; j
++) {
2794 node
= node_order
[j
];
2795 z
= &NODE_DATA(node
)->node_zones
[zone_type
];
2796 if (populated_zone(z
)) {
2798 &zonelist
->_zonerefs
[pos
++]);
2799 check_highest_zone(zone_type
);
2803 zonelist
->_zonerefs
[pos
].zone
= NULL
;
2804 zonelist
->_zonerefs
[pos
].zone_idx
= 0;
2807 static int default_zonelist_order(void)
2810 unsigned long low_kmem_size
,total_size
;
2814 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
2815 * If they are really small and used heavily, the system can fall
2816 * into OOM very easily.
2817 * This function detect ZONE_DMA/DMA32 size and configures zone order.
2819 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
2822 for_each_online_node(nid
) {
2823 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
2824 z
= &NODE_DATA(nid
)->node_zones
[zone_type
];
2825 if (populated_zone(z
)) {
2826 if (zone_type
< ZONE_NORMAL
)
2827 low_kmem_size
+= z
->present_pages
;
2828 total_size
+= z
->present_pages
;
2829 } else if (zone_type
== ZONE_NORMAL
) {
2831 * If any node has only lowmem, then node order
2832 * is preferred to allow kernel allocations
2833 * locally; otherwise, they can easily infringe
2834 * on other nodes when there is an abundance of
2835 * lowmem available to allocate from.
2837 return ZONELIST_ORDER_NODE
;
2841 if (!low_kmem_size
|| /* there are no DMA area. */
2842 low_kmem_size
> total_size
/2) /* DMA/DMA32 is big. */
2843 return ZONELIST_ORDER_NODE
;
2845 * look into each node's config.
2846 * If there is a node whose DMA/DMA32 memory is very big area on
2847 * local memory, NODE_ORDER may be suitable.
2849 average_size
= total_size
/
2850 (nodes_weight(node_states
[N_HIGH_MEMORY
]) + 1);
2851 for_each_online_node(nid
) {
2854 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
2855 z
= &NODE_DATA(nid
)->node_zones
[zone_type
];
2856 if (populated_zone(z
)) {
2857 if (zone_type
< ZONE_NORMAL
)
2858 low_kmem_size
+= z
->present_pages
;
2859 total_size
+= z
->present_pages
;
2862 if (low_kmem_size
&&
2863 total_size
> average_size
&& /* ignore small node */
2864 low_kmem_size
> total_size
* 70/100)
2865 return ZONELIST_ORDER_NODE
;
2867 return ZONELIST_ORDER_ZONE
;
2870 static void set_zonelist_order(void)
2872 if (user_zonelist_order
== ZONELIST_ORDER_DEFAULT
)
2873 current_zonelist_order
= default_zonelist_order();
2875 current_zonelist_order
= user_zonelist_order
;
2878 static void build_zonelists(pg_data_t
*pgdat
)
2882 nodemask_t used_mask
;
2883 int local_node
, prev_node
;
2884 struct zonelist
*zonelist
;
2885 int order
= current_zonelist_order
;
2887 /* initialize zonelists */
2888 for (i
= 0; i
< MAX_ZONELISTS
; i
++) {
2889 zonelist
= pgdat
->node_zonelists
+ i
;
2890 zonelist
->_zonerefs
[0].zone
= NULL
;
2891 zonelist
->_zonerefs
[0].zone_idx
= 0;
2894 /* NUMA-aware ordering of nodes */
2895 local_node
= pgdat
->node_id
;
2896 load
= nr_online_nodes
;
2897 prev_node
= local_node
;
2898 nodes_clear(used_mask
);
2900 memset(node_order
, 0, sizeof(node_order
));
2903 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
2904 int distance
= node_distance(local_node
, node
);
2907 * If another node is sufficiently far away then it is better
2908 * to reclaim pages in a zone before going off node.
2910 if (distance
> RECLAIM_DISTANCE
)
2911 zone_reclaim_mode
= 1;
2914 * We don't want to pressure a particular node.
2915 * So adding penalty to the first node in same
2916 * distance group to make it round-robin.
2918 if (distance
!= node_distance(local_node
, prev_node
))
2919 node_load
[node
] = load
;
2923 if (order
== ZONELIST_ORDER_NODE
)
2924 build_zonelists_in_node_order(pgdat
, node
);
2926 node_order
[j
++] = node
; /* remember order */
2929 if (order
== ZONELIST_ORDER_ZONE
) {
2930 /* calculate node order -- i.e., DMA last! */
2931 build_zonelists_in_zone_order(pgdat
, j
);
2934 build_thisnode_zonelists(pgdat
);
2937 /* Construct the zonelist performance cache - see further mmzone.h */
2938 static void build_zonelist_cache(pg_data_t
*pgdat
)
2940 struct zonelist
*zonelist
;
2941 struct zonelist_cache
*zlc
;
2944 zonelist
= &pgdat
->node_zonelists
[0];
2945 zonelist
->zlcache_ptr
= zlc
= &zonelist
->zlcache
;
2946 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
2947 for (z
= zonelist
->_zonerefs
; z
->zone
; z
++)
2948 zlc
->z_to_n
[z
- zonelist
->_zonerefs
] = zonelist_node_idx(z
);
2951 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
2953 * Return node id of node used for "local" allocations.
2954 * I.e., first node id of first zone in arg node's generic zonelist.
2955 * Used for initializing percpu 'numa_mem', which is used primarily
2956 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
2958 int local_memory_node(int node
)
2962 (void)first_zones_zonelist(node_zonelist(node
, GFP_KERNEL
),
2963 gfp_zone(GFP_KERNEL
),
2970 #else /* CONFIG_NUMA */
2972 static void set_zonelist_order(void)
2974 current_zonelist_order
= ZONELIST_ORDER_ZONE
;
2977 static void build_zonelists(pg_data_t
*pgdat
)
2979 int node
, local_node
;
2981 struct zonelist
*zonelist
;
2983 local_node
= pgdat
->node_id
;
2985 zonelist
= &pgdat
->node_zonelists
[0];
2986 j
= build_zonelists_node(pgdat
, zonelist
, 0, MAX_NR_ZONES
- 1);
2989 * Now we build the zonelist so that it contains the zones
2990 * of all the other nodes.
2991 * We don't want to pressure a particular node, so when
2992 * building the zones for node N, we make sure that the
2993 * zones coming right after the local ones are those from
2994 * node N+1 (modulo N)
2996 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
2997 if (!node_online(node
))
2999 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
3002 for (node
= 0; node
< local_node
; node
++) {
3003 if (!node_online(node
))
3005 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
3009 zonelist
->_zonerefs
[j
].zone
= NULL
;
3010 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3013 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3014 static void build_zonelist_cache(pg_data_t
*pgdat
)
3016 pgdat
->node_zonelists
[0].zlcache_ptr
= NULL
;
3019 #endif /* CONFIG_NUMA */
3022 * Boot pageset table. One per cpu which is going to be used for all
3023 * zones and all nodes. The parameters will be set in such a way
3024 * that an item put on a list will immediately be handed over to
3025 * the buddy list. This is safe since pageset manipulation is done
3026 * with interrupts disabled.
3028 * The boot_pagesets must be kept even after bootup is complete for
3029 * unused processors and/or zones. They do play a role for bootstrapping
3030 * hotplugged processors.
3032 * zoneinfo_show() and maybe other functions do
3033 * not check if the processor is online before following the pageset pointer.
3034 * Other parts of the kernel may not check if the zone is available.
3036 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
);
3037 static DEFINE_PER_CPU(struct per_cpu_pageset
, boot_pageset
);
3038 static void setup_zone_pageset(struct zone
*zone
);
3041 * Global mutex to protect against size modification of zonelists
3042 * as well as to serialize pageset setup for the new populated zone.
3044 DEFINE_MUTEX(zonelists_mutex
);
3046 /* return values int ....just for stop_machine() */
3047 static __init_refok
int __build_all_zonelists(void *data
)
3053 memset(node_load
, 0, sizeof(node_load
));
3055 for_each_online_node(nid
) {
3056 pg_data_t
*pgdat
= NODE_DATA(nid
);
3058 build_zonelists(pgdat
);
3059 build_zonelist_cache(pgdat
);
3063 * Initialize the boot_pagesets that are going to be used
3064 * for bootstrapping processors. The real pagesets for
3065 * each zone will be allocated later when the per cpu
3066 * allocator is available.
3068 * boot_pagesets are used also for bootstrapping offline
3069 * cpus if the system is already booted because the pagesets
3070 * are needed to initialize allocators on a specific cpu too.
3071 * F.e. the percpu allocator needs the page allocator which
3072 * needs the percpu allocator in order to allocate its pagesets
3073 * (a chicken-egg dilemma).
3075 for_each_possible_cpu(cpu
) {
3076 setup_pageset(&per_cpu(boot_pageset
, cpu
), 0);
3078 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3080 * We now know the "local memory node" for each node--
3081 * i.e., the node of the first zone in the generic zonelist.
3082 * Set up numa_mem percpu variable for on-line cpus. During
3083 * boot, only the boot cpu should be on-line; we'll init the
3084 * secondary cpus' numa_mem as they come on-line. During
3085 * node/memory hotplug, we'll fixup all on-line cpus.
3087 if (cpu_online(cpu
))
3088 set_cpu_numa_mem(cpu
, local_memory_node(cpu_to_node(cpu
)));
3096 * Called with zonelists_mutex held always
3097 * unless system_state == SYSTEM_BOOTING.
3099 void build_all_zonelists(void *data
)
3101 set_zonelist_order();
3103 if (system_state
== SYSTEM_BOOTING
) {
3104 __build_all_zonelists(NULL
);
3105 mminit_verify_zonelist();
3106 cpuset_init_current_mems_allowed();
3108 /* we have to stop all cpus to guarantee there is no user
3110 #ifdef CONFIG_MEMORY_HOTPLUG
3112 setup_zone_pageset((struct zone
*)data
);
3114 stop_machine(__build_all_zonelists
, NULL
, NULL
);
3115 /* cpuset refresh routine should be here */
3117 vm_total_pages
= nr_free_pagecache_pages();
3119 * Disable grouping by mobility if the number of pages in the
3120 * system is too low to allow the mechanism to work. It would be
3121 * more accurate, but expensive to check per-zone. This check is
3122 * made on memory-hotadd so a system can start with mobility
3123 * disabled and enable it later
3125 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
3126 page_group_by_mobility_disabled
= 1;
3128 page_group_by_mobility_disabled
= 0;
3130 printk("Built %i zonelists in %s order, mobility grouping %s. "
3131 "Total pages: %ld\n",
3133 zonelist_order_name
[current_zonelist_order
],
3134 page_group_by_mobility_disabled
? "off" : "on",
3137 printk("Policy zone: %s\n", zone_names
[policy_zone
]);
3142 * Helper functions to size the waitqueue hash table.
3143 * Essentially these want to choose hash table sizes sufficiently
3144 * large so that collisions trying to wait on pages are rare.
3145 * But in fact, the number of active page waitqueues on typical
3146 * systems is ridiculously low, less than 200. So this is even
3147 * conservative, even though it seems large.
3149 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3150 * waitqueues, i.e. the size of the waitq table given the number of pages.
3152 #define PAGES_PER_WAITQUEUE 256
3154 #ifndef CONFIG_MEMORY_HOTPLUG
3155 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
3157 unsigned long size
= 1;
3159 pages
/= PAGES_PER_WAITQUEUE
;
3161 while (size
< pages
)
3165 * Once we have dozens or even hundreds of threads sleeping
3166 * on IO we've got bigger problems than wait queue collision.
3167 * Limit the size of the wait table to a reasonable size.
3169 size
= min(size
, 4096UL);
3171 return max(size
, 4UL);
3175 * A zone's size might be changed by hot-add, so it is not possible to determine
3176 * a suitable size for its wait_table. So we use the maximum size now.
3178 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3180 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3181 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3182 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3184 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3185 * or more by the traditional way. (See above). It equals:
3187 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3188 * ia64(16K page size) : = ( 8G + 4M)byte.
3189 * powerpc (64K page size) : = (32G +16M)byte.
3191 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
3198 * This is an integer logarithm so that shifts can be used later
3199 * to extract the more random high bits from the multiplicative
3200 * hash function before the remainder is taken.
3202 static inline unsigned long wait_table_bits(unsigned long size
)
3207 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3210 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3211 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3212 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3213 * higher will lead to a bigger reserve which will get freed as contiguous
3214 * blocks as reclaim kicks in
3216 static void setup_zone_migrate_reserve(struct zone
*zone
)
3218 unsigned long start_pfn
, pfn
, end_pfn
;
3220 unsigned long block_migratetype
;
3223 /* Get the start pfn, end pfn and the number of blocks to reserve */
3224 start_pfn
= zone
->zone_start_pfn
;
3225 end_pfn
= start_pfn
+ zone
->spanned_pages
;
3226 reserve
= roundup(min_wmark_pages(zone
), pageblock_nr_pages
) >>
3230 * Reserve blocks are generally in place to help high-order atomic
3231 * allocations that are short-lived. A min_free_kbytes value that
3232 * would result in more than 2 reserve blocks for atomic allocations
3233 * is assumed to be in place to help anti-fragmentation for the
3234 * future allocation of hugepages at runtime.
3236 reserve
= min(2, reserve
);
3238 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= pageblock_nr_pages
) {
3239 if (!pfn_valid(pfn
))
3241 page
= pfn_to_page(pfn
);
3243 /* Watch out for overlapping nodes */
3244 if (page_to_nid(page
) != zone_to_nid(zone
))
3247 /* Blocks with reserved pages will never free, skip them. */
3248 if (PageReserved(page
))
3251 block_migratetype
= get_pageblock_migratetype(page
);
3253 /* If this block is reserved, account for it */
3254 if (reserve
> 0 && block_migratetype
== MIGRATE_RESERVE
) {
3259 /* Suitable for reserving if this block is movable */
3260 if (reserve
> 0 && block_migratetype
== MIGRATE_MOVABLE
) {
3261 set_pageblock_migratetype(page
, MIGRATE_RESERVE
);
3262 move_freepages_block(zone
, page
, MIGRATE_RESERVE
);
3268 * If the reserve is met and this is a previous reserved block,
3271 if (block_migratetype
== MIGRATE_RESERVE
) {
3272 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
3273 move_freepages_block(zone
, page
, MIGRATE_MOVABLE
);
3279 * Initially all pages are reserved - free ones are freed
3280 * up by free_all_bootmem() once the early boot process is
3281 * done. Non-atomic initialization, single-pass.
3283 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
3284 unsigned long start_pfn
, enum memmap_context context
)
3287 unsigned long end_pfn
= start_pfn
+ size
;
3291 if (highest_memmap_pfn
< end_pfn
- 1)
3292 highest_memmap_pfn
= end_pfn
- 1;
3294 z
= &NODE_DATA(nid
)->node_zones
[zone
];
3295 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
3297 * There can be holes in boot-time mem_map[]s
3298 * handed to this function. They do not
3299 * exist on hotplugged memory.
3301 if (context
== MEMMAP_EARLY
) {
3302 if (!early_pfn_valid(pfn
))
3304 if (!early_pfn_in_nid(pfn
, nid
))
3307 page
= pfn_to_page(pfn
);
3308 set_page_links(page
, zone
, nid
, pfn
);
3309 mminit_verify_page_links(page
, zone
, nid
, pfn
);
3310 init_page_count(page
);
3311 reset_page_mapcount(page
);
3312 SetPageReserved(page
);
3314 * Mark the block movable so that blocks are reserved for
3315 * movable at startup. This will force kernel allocations
3316 * to reserve their blocks rather than leaking throughout
3317 * the address space during boot when many long-lived
3318 * kernel allocations are made. Later some blocks near
3319 * the start are marked MIGRATE_RESERVE by
3320 * setup_zone_migrate_reserve()
3322 * bitmap is created for zone's valid pfn range. but memmap
3323 * can be created for invalid pages (for alignment)
3324 * check here not to call set_pageblock_migratetype() against
3327 if ((z
->zone_start_pfn
<= pfn
)
3328 && (pfn
< z
->zone_start_pfn
+ z
->spanned_pages
)
3329 && !(pfn
& (pageblock_nr_pages
- 1)))
3330 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
3332 INIT_LIST_HEAD(&page
->lru
);
3333 #ifdef WANT_PAGE_VIRTUAL
3334 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3335 if (!is_highmem_idx(zone
))
3336 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
3341 static void __meminit
zone_init_free_lists(struct zone
*zone
)
3344 for_each_migratetype_order(order
, t
) {
3345 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
3346 zone
->free_area
[order
].nr_free
= 0;
3350 #ifndef __HAVE_ARCH_MEMMAP_INIT
3351 #define memmap_init(size, nid, zone, start_pfn) \
3352 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3355 static int zone_batchsize(struct zone
*zone
)
3361 * The per-cpu-pages pools are set to around 1000th of the
3362 * size of the zone. But no more than 1/2 of a meg.
3364 * OK, so we don't know how big the cache is. So guess.
3366 batch
= zone
->present_pages
/ 1024;
3367 if (batch
* PAGE_SIZE
> 512 * 1024)
3368 batch
= (512 * 1024) / PAGE_SIZE
;
3369 batch
/= 4; /* We effectively *= 4 below */
3374 * Clamp the batch to a 2^n - 1 value. Having a power
3375 * of 2 value was found to be more likely to have
3376 * suboptimal cache aliasing properties in some cases.
3378 * For example if 2 tasks are alternately allocating
3379 * batches of pages, one task can end up with a lot
3380 * of pages of one half of the possible page colors
3381 * and the other with pages of the other colors.
3383 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
3388 /* The deferral and batching of frees should be suppressed under NOMMU
3391 * The problem is that NOMMU needs to be able to allocate large chunks
3392 * of contiguous memory as there's no hardware page translation to
3393 * assemble apparent contiguous memory from discontiguous pages.
3395 * Queueing large contiguous runs of pages for batching, however,
3396 * causes the pages to actually be freed in smaller chunks. As there
3397 * can be a significant delay between the individual batches being
3398 * recycled, this leads to the once large chunks of space being
3399 * fragmented and becoming unavailable for high-order allocations.
3405 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
3407 struct per_cpu_pages
*pcp
;
3410 memset(p
, 0, sizeof(*p
));
3414 pcp
->high
= 6 * batch
;
3415 pcp
->batch
= max(1UL, 1 * batch
);
3416 for (migratetype
= 0; migratetype
< MIGRATE_PCPTYPES
; migratetype
++)
3417 INIT_LIST_HEAD(&pcp
->lists
[migratetype
]);
3421 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3422 * to the value high for the pageset p.
3425 static void setup_pagelist_highmark(struct per_cpu_pageset
*p
,
3428 struct per_cpu_pages
*pcp
;
3432 pcp
->batch
= max(1UL, high
/4);
3433 if ((high
/4) > (PAGE_SHIFT
* 8))
3434 pcp
->batch
= PAGE_SHIFT
* 8;
3437 static __meminit
void setup_zone_pageset(struct zone
*zone
)
3441 zone
->pageset
= alloc_percpu(struct per_cpu_pageset
);
3443 for_each_possible_cpu(cpu
) {
3444 struct per_cpu_pageset
*pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
3446 setup_pageset(pcp
, zone_batchsize(zone
));
3448 if (percpu_pagelist_fraction
)
3449 setup_pagelist_highmark(pcp
,
3450 (zone
->present_pages
/
3451 percpu_pagelist_fraction
));
3456 * Allocate per cpu pagesets and initialize them.
3457 * Before this call only boot pagesets were available.
3459 void __init
setup_per_cpu_pageset(void)
3463 for_each_populated_zone(zone
)
3464 setup_zone_pageset(zone
);
3467 static noinline __init_refok
3468 int zone_wait_table_init(struct zone
*zone
, unsigned long zone_size_pages
)
3471 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
3475 * The per-page waitqueue mechanism uses hashed waitqueues
3478 zone
->wait_table_hash_nr_entries
=
3479 wait_table_hash_nr_entries(zone_size_pages
);
3480 zone
->wait_table_bits
=
3481 wait_table_bits(zone
->wait_table_hash_nr_entries
);
3482 alloc_size
= zone
->wait_table_hash_nr_entries
3483 * sizeof(wait_queue_head_t
);
3485 if (!slab_is_available()) {
3486 zone
->wait_table
= (wait_queue_head_t
*)
3487 alloc_bootmem_node(pgdat
, alloc_size
);
3490 * This case means that a zone whose size was 0 gets new memory
3491 * via memory hot-add.
3492 * But it may be the case that a new node was hot-added. In
3493 * this case vmalloc() will not be able to use this new node's
3494 * memory - this wait_table must be initialized to use this new
3495 * node itself as well.
3496 * To use this new node's memory, further consideration will be
3499 zone
->wait_table
= vmalloc(alloc_size
);
3501 if (!zone
->wait_table
)
3504 for(i
= 0; i
< zone
->wait_table_hash_nr_entries
; ++i
)
3505 init_waitqueue_head(zone
->wait_table
+ i
);
3510 static int __zone_pcp_update(void *data
)
3512 struct zone
*zone
= data
;
3514 unsigned long batch
= zone_batchsize(zone
), flags
;
3516 for_each_possible_cpu(cpu
) {
3517 struct per_cpu_pageset
*pset
;
3518 struct per_cpu_pages
*pcp
;
3520 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
3523 local_irq_save(flags
);
3524 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
3525 setup_pageset(pset
, batch
);
3526 local_irq_restore(flags
);
3531 void zone_pcp_update(struct zone
*zone
)
3533 stop_machine(__zone_pcp_update
, zone
, NULL
);
3536 static __meminit
void zone_pcp_init(struct zone
*zone
)
3539 * per cpu subsystem is not up at this point. The following code
3540 * relies on the ability of the linker to provide the
3541 * offset of a (static) per cpu variable into the per cpu area.
3543 zone
->pageset
= &boot_pageset
;
3545 if (zone
->present_pages
)
3546 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%u\n",
3547 zone
->name
, zone
->present_pages
,
3548 zone_batchsize(zone
));
3551 __meminit
int init_currently_empty_zone(struct zone
*zone
,
3552 unsigned long zone_start_pfn
,
3554 enum memmap_context context
)
3556 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
3558 ret
= zone_wait_table_init(zone
, size
);
3561 pgdat
->nr_zones
= zone_idx(zone
) + 1;
3563 zone
->zone_start_pfn
= zone_start_pfn
;
3565 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
3566 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
3568 (unsigned long)zone_idx(zone
),
3569 zone_start_pfn
, (zone_start_pfn
+ size
));
3571 zone_init_free_lists(zone
);
3576 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3578 * Basic iterator support. Return the first range of PFNs for a node
3579 * Note: nid == MAX_NUMNODES returns first region regardless of node
3581 static int __meminit
first_active_region_index_in_nid(int nid
)
3585 for (i
= 0; i
< nr_nodemap_entries
; i
++)
3586 if (nid
== MAX_NUMNODES
|| early_node_map
[i
].nid
== nid
)
3593 * Basic iterator support. Return the next active range of PFNs for a node
3594 * Note: nid == MAX_NUMNODES returns next region regardless of node
3596 static int __meminit
next_active_region_index_in_nid(int index
, int nid
)
3598 for (index
= index
+ 1; index
< nr_nodemap_entries
; index
++)
3599 if (nid
== MAX_NUMNODES
|| early_node_map
[index
].nid
== nid
)
3605 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
3607 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
3608 * Architectures may implement their own version but if add_active_range()
3609 * was used and there are no special requirements, this is a convenient
3612 int __meminit
__early_pfn_to_nid(unsigned long pfn
)
3616 for (i
= 0; i
< nr_nodemap_entries
; i
++) {
3617 unsigned long start_pfn
= early_node_map
[i
].start_pfn
;
3618 unsigned long end_pfn
= early_node_map
[i
].end_pfn
;
3620 if (start_pfn
<= pfn
&& pfn
< end_pfn
)
3621 return early_node_map
[i
].nid
;
3623 /* This is a memory hole */
3626 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
3628 int __meminit
early_pfn_to_nid(unsigned long pfn
)
3632 nid
= __early_pfn_to_nid(pfn
);
3635 /* just returns 0 */
3639 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
3640 bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
3644 nid
= __early_pfn_to_nid(pfn
);
3645 if (nid
>= 0 && nid
!= node
)
3651 /* Basic iterator support to walk early_node_map[] */
3652 #define for_each_active_range_index_in_nid(i, nid) \
3653 for (i = first_active_region_index_in_nid(nid); i != -1; \
3654 i = next_active_region_index_in_nid(i, nid))
3657 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
3658 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
3659 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
3661 * If an architecture guarantees that all ranges registered with
3662 * add_active_ranges() contain no holes and may be freed, this
3663 * this function may be used instead of calling free_bootmem() manually.
3665 void __init
free_bootmem_with_active_regions(int nid
,
3666 unsigned long max_low_pfn
)
3670 for_each_active_range_index_in_nid(i
, nid
) {
3671 unsigned long size_pages
= 0;
3672 unsigned long end_pfn
= early_node_map
[i
].end_pfn
;
3674 if (early_node_map
[i
].start_pfn
>= max_low_pfn
)
3677 if (end_pfn
> max_low_pfn
)
3678 end_pfn
= max_low_pfn
;
3680 size_pages
= end_pfn
- early_node_map
[i
].start_pfn
;
3681 free_bootmem_node(NODE_DATA(early_node_map
[i
].nid
),
3682 PFN_PHYS(early_node_map
[i
].start_pfn
),
3683 size_pages
<< PAGE_SHIFT
);
3687 #ifdef CONFIG_HAVE_MEMBLOCK
3688 u64 __init
find_memory_core_early(int nid
, u64 size
, u64 align
,
3689 u64 goal
, u64 limit
)
3693 /* Need to go over early_node_map to find out good range for node */
3694 for_each_active_range_index_in_nid(i
, nid
) {
3696 u64 ei_start
, ei_last
;
3697 u64 final_start
, final_end
;
3699 ei_last
= early_node_map
[i
].end_pfn
;
3700 ei_last
<<= PAGE_SHIFT
;
3701 ei_start
= early_node_map
[i
].start_pfn
;
3702 ei_start
<<= PAGE_SHIFT
;
3704 final_start
= max(ei_start
, goal
);
3705 final_end
= min(ei_last
, limit
);
3707 if (final_start
>= final_end
)
3710 addr
= memblock_find_in_range(final_start
, final_end
, size
, align
);
3712 if (addr
== MEMBLOCK_ERROR
)
3718 return MEMBLOCK_ERROR
;
3722 int __init
add_from_early_node_map(struct range
*range
, int az
,
3723 int nr_range
, int nid
)
3728 /* need to go over early_node_map to find out good range for node */
3729 for_each_active_range_index_in_nid(i
, nid
) {
3730 start
= early_node_map
[i
].start_pfn
;
3731 end
= early_node_map
[i
].end_pfn
;
3732 nr_range
= add_range(range
, az
, nr_range
, start
, end
);
3737 #ifdef CONFIG_NO_BOOTMEM
3738 void * __init
__alloc_memory_core_early(int nid
, u64 size
, u64 align
,
3739 u64 goal
, u64 limit
)
3744 if (limit
> memblock
.current_limit
)
3745 limit
= memblock
.current_limit
;
3747 addr
= find_memory_core_early(nid
, size
, align
, goal
, limit
);
3749 if (addr
== MEMBLOCK_ERROR
)
3752 ptr
= phys_to_virt(addr
);
3753 memset(ptr
, 0, size
);
3754 memblock_x86_reserve_range(addr
, addr
+ size
, "BOOTMEM");
3756 * The min_count is set to 0 so that bootmem allocated blocks
3757 * are never reported as leaks.
3759 kmemleak_alloc(ptr
, size
, 0, 0);
3765 void __init
work_with_active_regions(int nid
, work_fn_t work_fn
, void *data
)
3770 for_each_active_range_index_in_nid(i
, nid
) {
3771 ret
= work_fn(early_node_map
[i
].start_pfn
,
3772 early_node_map
[i
].end_pfn
, data
);
3778 * sparse_memory_present_with_active_regions - Call memory_present for each active range
3779 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
3781 * If an architecture guarantees that all ranges registered with
3782 * add_active_ranges() contain no holes and may be freed, this
3783 * function may be used instead of calling memory_present() manually.
3785 void __init
sparse_memory_present_with_active_regions(int nid
)
3789 for_each_active_range_index_in_nid(i
, nid
)
3790 memory_present(early_node_map
[i
].nid
,
3791 early_node_map
[i
].start_pfn
,
3792 early_node_map
[i
].end_pfn
);
3796 * get_pfn_range_for_nid - Return the start and end page frames for a node
3797 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
3798 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
3799 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
3801 * It returns the start and end page frame of a node based on information
3802 * provided by an arch calling add_active_range(). If called for a node
3803 * with no available memory, a warning is printed and the start and end
3806 void __meminit
get_pfn_range_for_nid(unsigned int nid
,
3807 unsigned long *start_pfn
, unsigned long *end_pfn
)
3813 for_each_active_range_index_in_nid(i
, nid
) {
3814 *start_pfn
= min(*start_pfn
, early_node_map
[i
].start_pfn
);
3815 *end_pfn
= max(*end_pfn
, early_node_map
[i
].end_pfn
);
3818 if (*start_pfn
== -1UL)
3823 * This finds a zone that can be used for ZONE_MOVABLE pages. The
3824 * assumption is made that zones within a node are ordered in monotonic
3825 * increasing memory addresses so that the "highest" populated zone is used
3827 static void __init
find_usable_zone_for_movable(void)
3830 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
3831 if (zone_index
== ZONE_MOVABLE
)
3834 if (arch_zone_highest_possible_pfn
[zone_index
] >
3835 arch_zone_lowest_possible_pfn
[zone_index
])
3839 VM_BUG_ON(zone_index
== -1);
3840 movable_zone
= zone_index
;
3844 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
3845 * because it is sized independant of architecture. Unlike the other zones,
3846 * the starting point for ZONE_MOVABLE is not fixed. It may be different
3847 * in each node depending on the size of each node and how evenly kernelcore
3848 * is distributed. This helper function adjusts the zone ranges
3849 * provided by the architecture for a given node by using the end of the
3850 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
3851 * zones within a node are in order of monotonic increases memory addresses
3853 static void __meminit
adjust_zone_range_for_zone_movable(int nid
,
3854 unsigned long zone_type
,
3855 unsigned long node_start_pfn
,
3856 unsigned long node_end_pfn
,
3857 unsigned long *zone_start_pfn
,
3858 unsigned long *zone_end_pfn
)
3860 /* Only adjust if ZONE_MOVABLE is on this node */
3861 if (zone_movable_pfn
[nid
]) {
3862 /* Size ZONE_MOVABLE */
3863 if (zone_type
== ZONE_MOVABLE
) {
3864 *zone_start_pfn
= zone_movable_pfn
[nid
];
3865 *zone_end_pfn
= min(node_end_pfn
,
3866 arch_zone_highest_possible_pfn
[movable_zone
]);
3868 /* Adjust for ZONE_MOVABLE starting within this range */
3869 } else if (*zone_start_pfn
< zone_movable_pfn
[nid
] &&
3870 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
3871 *zone_end_pfn
= zone_movable_pfn
[nid
];
3873 /* Check if this whole range is within ZONE_MOVABLE */
3874 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
3875 *zone_start_pfn
= *zone_end_pfn
;
3880 * Return the number of pages a zone spans in a node, including holes
3881 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
3883 static unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
3884 unsigned long zone_type
,
3885 unsigned long *ignored
)
3887 unsigned long node_start_pfn
, node_end_pfn
;
3888 unsigned long zone_start_pfn
, zone_end_pfn
;
3890 /* Get the start and end of the node and zone */
3891 get_pfn_range_for_nid(nid
, &node_start_pfn
, &node_end_pfn
);
3892 zone_start_pfn
= arch_zone_lowest_possible_pfn
[zone_type
];
3893 zone_end_pfn
= arch_zone_highest_possible_pfn
[zone_type
];
3894 adjust_zone_range_for_zone_movable(nid
, zone_type
,
3895 node_start_pfn
, node_end_pfn
,
3896 &zone_start_pfn
, &zone_end_pfn
);
3898 /* Check that this node has pages within the zone's required range */
3899 if (zone_end_pfn
< node_start_pfn
|| zone_start_pfn
> node_end_pfn
)
3902 /* Move the zone boundaries inside the node if necessary */
3903 zone_end_pfn
= min(zone_end_pfn
, node_end_pfn
);
3904 zone_start_pfn
= max(zone_start_pfn
, node_start_pfn
);
3906 /* Return the spanned pages */
3907 return zone_end_pfn
- zone_start_pfn
;
3911 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
3912 * then all holes in the requested range will be accounted for.
3914 unsigned long __meminit
__absent_pages_in_range(int nid
,
3915 unsigned long range_start_pfn
,
3916 unsigned long range_end_pfn
)
3919 unsigned long prev_end_pfn
= 0, hole_pages
= 0;
3920 unsigned long start_pfn
;
3922 /* Find the end_pfn of the first active range of pfns in the node */
3923 i
= first_active_region_index_in_nid(nid
);
3927 prev_end_pfn
= min(early_node_map
[i
].start_pfn
, range_end_pfn
);
3929 /* Account for ranges before physical memory on this node */
3930 if (early_node_map
[i
].start_pfn
> range_start_pfn
)
3931 hole_pages
= prev_end_pfn
- range_start_pfn
;
3933 /* Find all holes for the zone within the node */
3934 for (; i
!= -1; i
= next_active_region_index_in_nid(i
, nid
)) {
3936 /* No need to continue if prev_end_pfn is outside the zone */
3937 if (prev_end_pfn
>= range_end_pfn
)
3940 /* Make sure the end of the zone is not within the hole */
3941 start_pfn
= min(early_node_map
[i
].start_pfn
, range_end_pfn
);
3942 prev_end_pfn
= max(prev_end_pfn
, range_start_pfn
);
3944 /* Update the hole size cound and move on */
3945 if (start_pfn
> range_start_pfn
) {
3946 BUG_ON(prev_end_pfn
> start_pfn
);
3947 hole_pages
+= start_pfn
- prev_end_pfn
;
3949 prev_end_pfn
= early_node_map
[i
].end_pfn
;
3952 /* Account for ranges past physical memory on this node */
3953 if (range_end_pfn
> prev_end_pfn
)
3954 hole_pages
+= range_end_pfn
-
3955 max(range_start_pfn
, prev_end_pfn
);
3961 * absent_pages_in_range - Return number of page frames in holes within a range
3962 * @start_pfn: The start PFN to start searching for holes
3963 * @end_pfn: The end PFN to stop searching for holes
3965 * It returns the number of pages frames in memory holes within a range.
3967 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
3968 unsigned long end_pfn
)
3970 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
3973 /* Return the number of page frames in holes in a zone on a node */
3974 static unsigned long __meminit
zone_absent_pages_in_node(int nid
,
3975 unsigned long zone_type
,
3976 unsigned long *ignored
)
3978 unsigned long node_start_pfn
, node_end_pfn
;
3979 unsigned long zone_start_pfn
, zone_end_pfn
;
3981 get_pfn_range_for_nid(nid
, &node_start_pfn
, &node_end_pfn
);
3982 zone_start_pfn
= max(arch_zone_lowest_possible_pfn
[zone_type
],
3984 zone_end_pfn
= min(arch_zone_highest_possible_pfn
[zone_type
],
3987 adjust_zone_range_for_zone_movable(nid
, zone_type
,
3988 node_start_pfn
, node_end_pfn
,
3989 &zone_start_pfn
, &zone_end_pfn
);
3990 return __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
3994 static inline unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
3995 unsigned long zone_type
,
3996 unsigned long *zones_size
)
3998 return zones_size
[zone_type
];
4001 static inline unsigned long __meminit
zone_absent_pages_in_node(int nid
,
4002 unsigned long zone_type
,
4003 unsigned long *zholes_size
)
4008 return zholes_size
[zone_type
];
4013 static void __meminit
calculate_node_totalpages(struct pglist_data
*pgdat
,
4014 unsigned long *zones_size
, unsigned long *zholes_size
)
4016 unsigned long realtotalpages
, totalpages
= 0;
4019 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4020 totalpages
+= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
4022 pgdat
->node_spanned_pages
= totalpages
;
4024 realtotalpages
= totalpages
;
4025 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4027 zone_absent_pages_in_node(pgdat
->node_id
, i
,
4029 pgdat
->node_present_pages
= realtotalpages
;
4030 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
4034 #ifndef CONFIG_SPARSEMEM
4036 * Calculate the size of the zone->blockflags rounded to an unsigned long
4037 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4038 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4039 * round what is now in bits to nearest long in bits, then return it in
4042 static unsigned long __init
usemap_size(unsigned long zonesize
)
4044 unsigned long usemapsize
;
4046 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
4047 usemapsize
= usemapsize
>> pageblock_order
;
4048 usemapsize
*= NR_PAGEBLOCK_BITS
;
4049 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
4051 return usemapsize
/ 8;
4054 static void __init
setup_usemap(struct pglist_data
*pgdat
,
4055 struct zone
*zone
, unsigned long zonesize
)
4057 unsigned long usemapsize
= usemap_size(zonesize
);
4058 zone
->pageblock_flags
= NULL
;
4060 zone
->pageblock_flags
= alloc_bootmem_node(pgdat
, usemapsize
);
4063 static inline void setup_usemap(struct pglist_data
*pgdat
,
4064 struct zone
*zone
, unsigned long zonesize
) {}
4065 #endif /* CONFIG_SPARSEMEM */
4067 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4069 /* Return a sensible default order for the pageblock size. */
4070 static inline int pageblock_default_order(void)
4072 if (HPAGE_SHIFT
> PAGE_SHIFT
)
4073 return HUGETLB_PAGE_ORDER
;
4078 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4079 static inline void __init
set_pageblock_order(unsigned int order
)
4081 /* Check that pageblock_nr_pages has not already been setup */
4082 if (pageblock_order
)
4086 * Assume the largest contiguous order of interest is a huge page.
4087 * This value may be variable depending on boot parameters on IA64
4089 pageblock_order
= order
;
4091 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4094 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4095 * and pageblock_default_order() are unused as pageblock_order is set
4096 * at compile-time. See include/linux/pageblock-flags.h for the values of
4097 * pageblock_order based on the kernel config
4099 static inline int pageblock_default_order(unsigned int order
)
4103 #define set_pageblock_order(x) do {} while (0)
4105 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4108 * Set up the zone data structures:
4109 * - mark all pages reserved
4110 * - mark all memory queues empty
4111 * - clear the memory bitmaps
4113 static void __paginginit
free_area_init_core(struct pglist_data
*pgdat
,
4114 unsigned long *zones_size
, unsigned long *zholes_size
)
4117 int nid
= pgdat
->node_id
;
4118 unsigned long zone_start_pfn
= pgdat
->node_start_pfn
;
4121 pgdat_resize_init(pgdat
);
4122 pgdat
->nr_zones
= 0;
4123 init_waitqueue_head(&pgdat
->kswapd_wait
);
4124 pgdat
->kswapd_max_order
= 0;
4125 pgdat_page_cgroup_init(pgdat
);
4127 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
4128 struct zone
*zone
= pgdat
->node_zones
+ j
;
4129 unsigned long size
, realsize
, memmap_pages
;
4132 size
= zone_spanned_pages_in_node(nid
, j
, zones_size
);
4133 realsize
= size
- zone_absent_pages_in_node(nid
, j
,
4137 * Adjust realsize so that it accounts for how much memory
4138 * is used by this zone for memmap. This affects the watermark
4139 * and per-cpu initialisations
4142 PAGE_ALIGN(size
* sizeof(struct page
)) >> PAGE_SHIFT
;
4143 if (realsize
>= memmap_pages
) {
4144 realsize
-= memmap_pages
;
4147 " %s zone: %lu pages used for memmap\n",
4148 zone_names
[j
], memmap_pages
);
4151 " %s zone: %lu pages exceeds realsize %lu\n",
4152 zone_names
[j
], memmap_pages
, realsize
);
4154 /* Account for reserved pages */
4155 if (j
== 0 && realsize
> dma_reserve
) {
4156 realsize
-= dma_reserve
;
4157 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
4158 zone_names
[0], dma_reserve
);
4161 if (!is_highmem_idx(j
))
4162 nr_kernel_pages
+= realsize
;
4163 nr_all_pages
+= realsize
;
4165 zone
->spanned_pages
= size
;
4166 zone
->present_pages
= realsize
;
4169 zone
->min_unmapped_pages
= (realsize
*sysctl_min_unmapped_ratio
)
4171 zone
->min_slab_pages
= (realsize
* sysctl_min_slab_ratio
) / 100;
4173 zone
->name
= zone_names
[j
];
4174 spin_lock_init(&zone
->lock
);
4175 spin_lock_init(&zone
->lru_lock
);
4176 zone_seqlock_init(zone
);
4177 zone
->zone_pgdat
= pgdat
;
4179 zone_pcp_init(zone
);
4181 INIT_LIST_HEAD(&zone
->lru
[l
].list
);
4182 zone
->reclaim_stat
.nr_saved_scan
[l
] = 0;
4184 zone
->reclaim_stat
.recent_rotated
[0] = 0;
4185 zone
->reclaim_stat
.recent_rotated
[1] = 0;
4186 zone
->reclaim_stat
.recent_scanned
[0] = 0;
4187 zone
->reclaim_stat
.recent_scanned
[1] = 0;
4188 zap_zone_vm_stats(zone
);
4193 set_pageblock_order(pageblock_default_order());
4194 setup_usemap(pgdat
, zone
, size
);
4195 ret
= init_currently_empty_zone(zone
, zone_start_pfn
,
4196 size
, MEMMAP_EARLY
);
4198 memmap_init(size
, nid
, j
, zone_start_pfn
);
4199 zone_start_pfn
+= size
;
4203 static void __init_refok
alloc_node_mem_map(struct pglist_data
*pgdat
)
4205 /* Skip empty nodes */
4206 if (!pgdat
->node_spanned_pages
)
4209 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4210 /* ia64 gets its own node_mem_map, before this, without bootmem */
4211 if (!pgdat
->node_mem_map
) {
4212 unsigned long size
, start
, end
;
4216 * The zone's endpoints aren't required to be MAX_ORDER
4217 * aligned but the node_mem_map endpoints must be in order
4218 * for the buddy allocator to function correctly.
4220 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
4221 end
= pgdat
->node_start_pfn
+ pgdat
->node_spanned_pages
;
4222 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
4223 size
= (end
- start
) * sizeof(struct page
);
4224 map
= alloc_remap(pgdat
->node_id
, size
);
4226 map
= alloc_bootmem_node(pgdat
, size
);
4227 pgdat
->node_mem_map
= map
+ (pgdat
->node_start_pfn
- start
);
4229 #ifndef CONFIG_NEED_MULTIPLE_NODES
4231 * With no DISCONTIG, the global mem_map is just set as node 0's
4233 if (pgdat
== NODE_DATA(0)) {
4234 mem_map
= NODE_DATA(0)->node_mem_map
;
4235 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
4236 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
4237 mem_map
-= (pgdat
->node_start_pfn
- ARCH_PFN_OFFSET
);
4238 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4241 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4244 void __paginginit
free_area_init_node(int nid
, unsigned long *zones_size
,
4245 unsigned long node_start_pfn
, unsigned long *zholes_size
)
4247 pg_data_t
*pgdat
= NODE_DATA(nid
);
4249 pgdat
->node_id
= nid
;
4250 pgdat
->node_start_pfn
= node_start_pfn
;
4251 calculate_node_totalpages(pgdat
, zones_size
, zholes_size
);
4253 alloc_node_mem_map(pgdat
);
4254 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4255 printk(KERN_DEBUG
"free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4256 nid
, (unsigned long)pgdat
,
4257 (unsigned long)pgdat
->node_mem_map
);
4260 free_area_init_core(pgdat
, zones_size
, zholes_size
);
4263 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
4265 #if MAX_NUMNODES > 1
4267 * Figure out the number of possible node ids.
4269 static void __init
setup_nr_node_ids(void)
4272 unsigned int highest
= 0;
4274 for_each_node_mask(node
, node_possible_map
)
4276 nr_node_ids
= highest
+ 1;
4279 static inline void setup_nr_node_ids(void)
4285 * add_active_range - Register a range of PFNs backed by physical memory
4286 * @nid: The node ID the range resides on
4287 * @start_pfn: The start PFN of the available physical memory
4288 * @end_pfn: The end PFN of the available physical memory
4290 * These ranges are stored in an early_node_map[] and later used by
4291 * free_area_init_nodes() to calculate zone sizes and holes. If the
4292 * range spans a memory hole, it is up to the architecture to ensure
4293 * the memory is not freed by the bootmem allocator. If possible
4294 * the range being registered will be merged with existing ranges.
4296 void __init
add_active_range(unsigned int nid
, unsigned long start_pfn
,
4297 unsigned long end_pfn
)
4301 mminit_dprintk(MMINIT_TRACE
, "memory_register",
4302 "Entering add_active_range(%d, %#lx, %#lx) "
4303 "%d entries of %d used\n",
4304 nid
, start_pfn
, end_pfn
,
4305 nr_nodemap_entries
, MAX_ACTIVE_REGIONS
);
4307 mminit_validate_memmodel_limits(&start_pfn
, &end_pfn
);
4309 /* Merge with existing active regions if possible */
4310 for (i
= 0; i
< nr_nodemap_entries
; i
++) {
4311 if (early_node_map
[i
].nid
!= nid
)
4314 /* Skip if an existing region covers this new one */
4315 if (start_pfn
>= early_node_map
[i
].start_pfn
&&
4316 end_pfn
<= early_node_map
[i
].end_pfn
)
4319 /* Merge forward if suitable */
4320 if (start_pfn
<= early_node_map
[i
].end_pfn
&&
4321 end_pfn
> early_node_map
[i
].end_pfn
) {
4322 early_node_map
[i
].end_pfn
= end_pfn
;
4326 /* Merge backward if suitable */
4327 if (start_pfn
< early_node_map
[i
].start_pfn
&&
4328 end_pfn
>= early_node_map
[i
].start_pfn
) {
4329 early_node_map
[i
].start_pfn
= start_pfn
;
4334 /* Check that early_node_map is large enough */
4335 if (i
>= MAX_ACTIVE_REGIONS
) {
4336 printk(KERN_CRIT
"More than %d memory regions, truncating\n",
4337 MAX_ACTIVE_REGIONS
);
4341 early_node_map
[i
].nid
= nid
;
4342 early_node_map
[i
].start_pfn
= start_pfn
;
4343 early_node_map
[i
].end_pfn
= end_pfn
;
4344 nr_nodemap_entries
= i
+ 1;
4348 * remove_active_range - Shrink an existing registered range of PFNs
4349 * @nid: The node id the range is on that should be shrunk
4350 * @start_pfn: The new PFN of the range
4351 * @end_pfn: The new PFN of the range
4353 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
4354 * The map is kept near the end physical page range that has already been
4355 * registered. This function allows an arch to shrink an existing registered
4358 void __init
remove_active_range(unsigned int nid
, unsigned long start_pfn
,
4359 unsigned long end_pfn
)
4364 printk(KERN_DEBUG
"remove_active_range (%d, %lu, %lu)\n",
4365 nid
, start_pfn
, end_pfn
);
4367 /* Find the old active region end and shrink */
4368 for_each_active_range_index_in_nid(i
, nid
) {
4369 if (early_node_map
[i
].start_pfn
>= start_pfn
&&
4370 early_node_map
[i
].end_pfn
<= end_pfn
) {
4372 early_node_map
[i
].start_pfn
= 0;
4373 early_node_map
[i
].end_pfn
= 0;
4377 if (early_node_map
[i
].start_pfn
< start_pfn
&&
4378 early_node_map
[i
].end_pfn
> start_pfn
) {
4379 unsigned long temp_end_pfn
= early_node_map
[i
].end_pfn
;
4380 early_node_map
[i
].end_pfn
= start_pfn
;
4381 if (temp_end_pfn
> end_pfn
)
4382 add_active_range(nid
, end_pfn
, temp_end_pfn
);
4385 if (early_node_map
[i
].start_pfn
>= start_pfn
&&
4386 early_node_map
[i
].end_pfn
> end_pfn
&&
4387 early_node_map
[i
].start_pfn
< end_pfn
) {
4388 early_node_map
[i
].start_pfn
= end_pfn
;
4396 /* remove the blank ones */
4397 for (i
= nr_nodemap_entries
- 1; i
> 0; i
--) {
4398 if (early_node_map
[i
].nid
!= nid
)
4400 if (early_node_map
[i
].end_pfn
)
4402 /* we found it, get rid of it */
4403 for (j
= i
; j
< nr_nodemap_entries
- 1; j
++)
4404 memcpy(&early_node_map
[j
], &early_node_map
[j
+1],
4405 sizeof(early_node_map
[j
]));
4406 j
= nr_nodemap_entries
- 1;
4407 memset(&early_node_map
[j
], 0, sizeof(early_node_map
[j
]));
4408 nr_nodemap_entries
--;
4413 * remove_all_active_ranges - Remove all currently registered regions
4415 * During discovery, it may be found that a table like SRAT is invalid
4416 * and an alternative discovery method must be used. This function removes
4417 * all currently registered regions.
4419 void __init
remove_all_active_ranges(void)
4421 memset(early_node_map
, 0, sizeof(early_node_map
));
4422 nr_nodemap_entries
= 0;
4425 /* Compare two active node_active_regions */
4426 static int __init
cmp_node_active_region(const void *a
, const void *b
)
4428 struct node_active_region
*arange
= (struct node_active_region
*)a
;
4429 struct node_active_region
*brange
= (struct node_active_region
*)b
;
4431 /* Done this way to avoid overflows */
4432 if (arange
->start_pfn
> brange
->start_pfn
)
4434 if (arange
->start_pfn
< brange
->start_pfn
)
4440 /* sort the node_map by start_pfn */
4441 void __init
sort_node_map(void)
4443 sort(early_node_map
, (size_t)nr_nodemap_entries
,
4444 sizeof(struct node_active_region
),
4445 cmp_node_active_region
, NULL
);
4448 /* Find the lowest pfn for a node */
4449 static unsigned long __init
find_min_pfn_for_node(int nid
)
4452 unsigned long min_pfn
= ULONG_MAX
;
4454 /* Assuming a sorted map, the first range found has the starting pfn */
4455 for_each_active_range_index_in_nid(i
, nid
)
4456 min_pfn
= min(min_pfn
, early_node_map
[i
].start_pfn
);
4458 if (min_pfn
== ULONG_MAX
) {
4460 "Could not find start_pfn for node %d\n", nid
);
4468 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4470 * It returns the minimum PFN based on information provided via
4471 * add_active_range().
4473 unsigned long __init
find_min_pfn_with_active_regions(void)
4475 return find_min_pfn_for_node(MAX_NUMNODES
);
4479 * early_calculate_totalpages()
4480 * Sum pages in active regions for movable zone.
4481 * Populate N_HIGH_MEMORY for calculating usable_nodes.
4483 static unsigned long __init
early_calculate_totalpages(void)
4486 unsigned long totalpages
= 0;
4488 for (i
= 0; i
< nr_nodemap_entries
; i
++) {
4489 unsigned long pages
= early_node_map
[i
].end_pfn
-
4490 early_node_map
[i
].start_pfn
;
4491 totalpages
+= pages
;
4493 node_set_state(early_node_map
[i
].nid
, N_HIGH_MEMORY
);
4499 * Find the PFN the Movable zone begins in each node. Kernel memory
4500 * is spread evenly between nodes as long as the nodes have enough
4501 * memory. When they don't, some nodes will have more kernelcore than
4504 static void __init
find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn
)
4507 unsigned long usable_startpfn
;
4508 unsigned long kernelcore_node
, kernelcore_remaining
;
4509 /* save the state before borrow the nodemask */
4510 nodemask_t saved_node_state
= node_states
[N_HIGH_MEMORY
];
4511 unsigned long totalpages
= early_calculate_totalpages();
4512 int usable_nodes
= nodes_weight(node_states
[N_HIGH_MEMORY
]);
4515 * If movablecore was specified, calculate what size of
4516 * kernelcore that corresponds so that memory usable for
4517 * any allocation type is evenly spread. If both kernelcore
4518 * and movablecore are specified, then the value of kernelcore
4519 * will be used for required_kernelcore if it's greater than
4520 * what movablecore would have allowed.
4522 if (required_movablecore
) {
4523 unsigned long corepages
;
4526 * Round-up so that ZONE_MOVABLE is at least as large as what
4527 * was requested by the user
4529 required_movablecore
=
4530 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
4531 corepages
= totalpages
- required_movablecore
;
4533 required_kernelcore
= max(required_kernelcore
, corepages
);
4536 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4537 if (!required_kernelcore
)
4540 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4541 find_usable_zone_for_movable();
4542 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
4545 /* Spread kernelcore memory as evenly as possible throughout nodes */
4546 kernelcore_node
= required_kernelcore
/ usable_nodes
;
4547 for_each_node_state(nid
, N_HIGH_MEMORY
) {
4549 * Recalculate kernelcore_node if the division per node
4550 * now exceeds what is necessary to satisfy the requested
4551 * amount of memory for the kernel
4553 if (required_kernelcore
< kernelcore_node
)
4554 kernelcore_node
= required_kernelcore
/ usable_nodes
;
4557 * As the map is walked, we track how much memory is usable
4558 * by the kernel using kernelcore_remaining. When it is
4559 * 0, the rest of the node is usable by ZONE_MOVABLE
4561 kernelcore_remaining
= kernelcore_node
;
4563 /* Go through each range of PFNs within this node */
4564 for_each_active_range_index_in_nid(i
, nid
) {
4565 unsigned long start_pfn
, end_pfn
;
4566 unsigned long size_pages
;
4568 start_pfn
= max(early_node_map
[i
].start_pfn
,
4569 zone_movable_pfn
[nid
]);
4570 end_pfn
= early_node_map
[i
].end_pfn
;
4571 if (start_pfn
>= end_pfn
)
4574 /* Account for what is only usable for kernelcore */
4575 if (start_pfn
< usable_startpfn
) {
4576 unsigned long kernel_pages
;
4577 kernel_pages
= min(end_pfn
, usable_startpfn
)
4580 kernelcore_remaining
-= min(kernel_pages
,
4581 kernelcore_remaining
);
4582 required_kernelcore
-= min(kernel_pages
,
4583 required_kernelcore
);
4585 /* Continue if range is now fully accounted */
4586 if (end_pfn
<= usable_startpfn
) {
4589 * Push zone_movable_pfn to the end so
4590 * that if we have to rebalance
4591 * kernelcore across nodes, we will
4592 * not double account here
4594 zone_movable_pfn
[nid
] = end_pfn
;
4597 start_pfn
= usable_startpfn
;
4601 * The usable PFN range for ZONE_MOVABLE is from
4602 * start_pfn->end_pfn. Calculate size_pages as the
4603 * number of pages used as kernelcore
4605 size_pages
= end_pfn
- start_pfn
;
4606 if (size_pages
> kernelcore_remaining
)
4607 size_pages
= kernelcore_remaining
;
4608 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
4611 * Some kernelcore has been met, update counts and
4612 * break if the kernelcore for this node has been
4615 required_kernelcore
-= min(required_kernelcore
,
4617 kernelcore_remaining
-= size_pages
;
4618 if (!kernelcore_remaining
)
4624 * If there is still required_kernelcore, we do another pass with one
4625 * less node in the count. This will push zone_movable_pfn[nid] further
4626 * along on the nodes that still have memory until kernelcore is
4630 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
4633 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4634 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
4635 zone_movable_pfn
[nid
] =
4636 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
4639 /* restore the node_state */
4640 node_states
[N_HIGH_MEMORY
] = saved_node_state
;
4643 /* Any regular memory on that node ? */
4644 static void check_for_regular_memory(pg_data_t
*pgdat
)
4646 #ifdef CONFIG_HIGHMEM
4647 enum zone_type zone_type
;
4649 for (zone_type
= 0; zone_type
<= ZONE_NORMAL
; zone_type
++) {
4650 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
4651 if (zone
->present_pages
)
4652 node_set_state(zone_to_nid(zone
), N_NORMAL_MEMORY
);
4658 * free_area_init_nodes - Initialise all pg_data_t and zone data
4659 * @max_zone_pfn: an array of max PFNs for each zone
4661 * This will call free_area_init_node() for each active node in the system.
4662 * Using the page ranges provided by add_active_range(), the size of each
4663 * zone in each node and their holes is calculated. If the maximum PFN
4664 * between two adjacent zones match, it is assumed that the zone is empty.
4665 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4666 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4667 * starts where the previous one ended. For example, ZONE_DMA32 starts
4668 * at arch_max_dma_pfn.
4670 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
4675 /* Sort early_node_map as initialisation assumes it is sorted */
4678 /* Record where the zone boundaries are */
4679 memset(arch_zone_lowest_possible_pfn
, 0,
4680 sizeof(arch_zone_lowest_possible_pfn
));
4681 memset(arch_zone_highest_possible_pfn
, 0,
4682 sizeof(arch_zone_highest_possible_pfn
));
4683 arch_zone_lowest_possible_pfn
[0] = find_min_pfn_with_active_regions();
4684 arch_zone_highest_possible_pfn
[0] = max_zone_pfn
[0];
4685 for (i
= 1; i
< MAX_NR_ZONES
; i
++) {
4686 if (i
== ZONE_MOVABLE
)
4688 arch_zone_lowest_possible_pfn
[i
] =
4689 arch_zone_highest_possible_pfn
[i
-1];
4690 arch_zone_highest_possible_pfn
[i
] =
4691 max(max_zone_pfn
[i
], arch_zone_lowest_possible_pfn
[i
]);
4693 arch_zone_lowest_possible_pfn
[ZONE_MOVABLE
] = 0;
4694 arch_zone_highest_possible_pfn
[ZONE_MOVABLE
] = 0;
4696 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4697 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
4698 find_zone_movable_pfns_for_nodes(zone_movable_pfn
);
4700 /* Print out the zone ranges */
4701 printk("Zone PFN ranges:\n");
4702 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
4703 if (i
== ZONE_MOVABLE
)
4705 printk(" %-8s ", zone_names
[i
]);
4706 if (arch_zone_lowest_possible_pfn
[i
] ==
4707 arch_zone_highest_possible_pfn
[i
])
4710 printk("%0#10lx -> %0#10lx\n",
4711 arch_zone_lowest_possible_pfn
[i
],
4712 arch_zone_highest_possible_pfn
[i
]);
4715 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4716 printk("Movable zone start PFN for each node\n");
4717 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
4718 if (zone_movable_pfn
[i
])
4719 printk(" Node %d: %lu\n", i
, zone_movable_pfn
[i
]);
4722 /* Print out the early_node_map[] */
4723 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries
);
4724 for (i
= 0; i
< nr_nodemap_entries
; i
++)
4725 printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map
[i
].nid
,
4726 early_node_map
[i
].start_pfn
,
4727 early_node_map
[i
].end_pfn
);
4729 /* Initialise every node */
4730 mminit_verify_pageflags_layout();
4731 setup_nr_node_ids();
4732 for_each_online_node(nid
) {
4733 pg_data_t
*pgdat
= NODE_DATA(nid
);
4734 free_area_init_node(nid
, NULL
,
4735 find_min_pfn_for_node(nid
), NULL
);
4737 /* Any memory on that node */
4738 if (pgdat
->node_present_pages
)
4739 node_set_state(nid
, N_HIGH_MEMORY
);
4740 check_for_regular_memory(pgdat
);
4744 static int __init
cmdline_parse_core(char *p
, unsigned long *core
)
4746 unsigned long long coremem
;
4750 coremem
= memparse(p
, &p
);
4751 *core
= coremem
>> PAGE_SHIFT
;
4753 /* Paranoid check that UL is enough for the coremem value */
4754 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
4760 * kernelcore=size sets the amount of memory for use for allocations that
4761 * cannot be reclaimed or migrated.
4763 static int __init
cmdline_parse_kernelcore(char *p
)
4765 return cmdline_parse_core(p
, &required_kernelcore
);
4769 * movablecore=size sets the amount of memory for use for allocations that
4770 * can be reclaimed or migrated.
4772 static int __init
cmdline_parse_movablecore(char *p
)
4774 return cmdline_parse_core(p
, &required_movablecore
);
4777 early_param("kernelcore", cmdline_parse_kernelcore
);
4778 early_param("movablecore", cmdline_parse_movablecore
);
4780 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4783 * set_dma_reserve - set the specified number of pages reserved in the first zone
4784 * @new_dma_reserve: The number of pages to mark reserved
4786 * The per-cpu batchsize and zone watermarks are determined by present_pages.
4787 * In the DMA zone, a significant percentage may be consumed by kernel image
4788 * and other unfreeable allocations which can skew the watermarks badly. This
4789 * function may optionally be used to account for unfreeable pages in the
4790 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
4791 * smaller per-cpu batchsize.
4793 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
4795 dma_reserve
= new_dma_reserve
;
4798 #ifndef CONFIG_NEED_MULTIPLE_NODES
4799 struct pglist_data __refdata contig_page_data
= {
4800 #ifndef CONFIG_NO_BOOTMEM
4801 .bdata
= &bootmem_node_data
[0]
4804 EXPORT_SYMBOL(contig_page_data
);
4807 void __init
free_area_init(unsigned long *zones_size
)
4809 free_area_init_node(0, zones_size
,
4810 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
4813 static int page_alloc_cpu_notify(struct notifier_block
*self
,
4814 unsigned long action
, void *hcpu
)
4816 int cpu
= (unsigned long)hcpu
;
4818 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
) {
4822 * Spill the event counters of the dead processor
4823 * into the current processors event counters.
4824 * This artificially elevates the count of the current
4827 vm_events_fold_cpu(cpu
);
4830 * Zero the differential counters of the dead processor
4831 * so that the vm statistics are consistent.
4833 * This is only okay since the processor is dead and cannot
4834 * race with what we are doing.
4836 refresh_cpu_vm_stats(cpu
);
4841 void __init
page_alloc_init(void)
4843 hotcpu_notifier(page_alloc_cpu_notify
, 0);
4847 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
4848 * or min_free_kbytes changes.
4850 static void calculate_totalreserve_pages(void)
4852 struct pglist_data
*pgdat
;
4853 unsigned long reserve_pages
= 0;
4854 enum zone_type i
, j
;
4856 for_each_online_pgdat(pgdat
) {
4857 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
4858 struct zone
*zone
= pgdat
->node_zones
+ i
;
4859 unsigned long max
= 0;
4861 /* Find valid and maximum lowmem_reserve in the zone */
4862 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
4863 if (zone
->lowmem_reserve
[j
] > max
)
4864 max
= zone
->lowmem_reserve
[j
];
4867 /* we treat the high watermark as reserved pages. */
4868 max
+= high_wmark_pages(zone
);
4870 if (max
> zone
->present_pages
)
4871 max
= zone
->present_pages
;
4872 reserve_pages
+= max
;
4875 totalreserve_pages
= reserve_pages
;
4879 * setup_per_zone_lowmem_reserve - called whenever
4880 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
4881 * has a correct pages reserved value, so an adequate number of
4882 * pages are left in the zone after a successful __alloc_pages().
4884 static void setup_per_zone_lowmem_reserve(void)
4886 struct pglist_data
*pgdat
;
4887 enum zone_type j
, idx
;
4889 for_each_online_pgdat(pgdat
) {
4890 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
4891 struct zone
*zone
= pgdat
->node_zones
+ j
;
4892 unsigned long present_pages
= zone
->present_pages
;
4894 zone
->lowmem_reserve
[j
] = 0;
4898 struct zone
*lower_zone
;
4902 if (sysctl_lowmem_reserve_ratio
[idx
] < 1)
4903 sysctl_lowmem_reserve_ratio
[idx
] = 1;
4905 lower_zone
= pgdat
->node_zones
+ idx
;
4906 lower_zone
->lowmem_reserve
[j
] = present_pages
/
4907 sysctl_lowmem_reserve_ratio
[idx
];
4908 present_pages
+= lower_zone
->present_pages
;
4913 /* update totalreserve_pages */
4914 calculate_totalreserve_pages();
4918 * setup_per_zone_wmarks - called when min_free_kbytes changes
4919 * or when memory is hot-{added|removed}
4921 * Ensures that the watermark[min,low,high] values for each zone are set
4922 * correctly with respect to min_free_kbytes.
4924 void setup_per_zone_wmarks(void)
4926 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
4927 unsigned long lowmem_pages
= 0;
4929 unsigned long flags
;
4931 /* Calculate total number of !ZONE_HIGHMEM pages */
4932 for_each_zone(zone
) {
4933 if (!is_highmem(zone
))
4934 lowmem_pages
+= zone
->present_pages
;
4937 for_each_zone(zone
) {
4940 spin_lock_irqsave(&zone
->lock
, flags
);
4941 tmp
= (u64
)pages_min
* zone
->present_pages
;
4942 do_div(tmp
, lowmem_pages
);
4943 if (is_highmem(zone
)) {
4945 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
4946 * need highmem pages, so cap pages_min to a small
4949 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
4950 * deltas controls asynch page reclaim, and so should
4951 * not be capped for highmem.
4955 min_pages
= zone
->present_pages
/ 1024;
4956 if (min_pages
< SWAP_CLUSTER_MAX
)
4957 min_pages
= SWAP_CLUSTER_MAX
;
4958 if (min_pages
> 128)
4960 zone
->watermark
[WMARK_MIN
] = min_pages
;
4963 * If it's a lowmem zone, reserve a number of pages
4964 * proportionate to the zone's size.
4966 zone
->watermark
[WMARK_MIN
] = tmp
;
4969 zone
->watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + (tmp
>> 2);
4970 zone
->watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + (tmp
>> 1);
4971 setup_zone_migrate_reserve(zone
);
4972 spin_unlock_irqrestore(&zone
->lock
, flags
);
4975 /* update totalreserve_pages */
4976 calculate_totalreserve_pages();
4980 * The inactive anon list should be small enough that the VM never has to
4981 * do too much work, but large enough that each inactive page has a chance
4982 * to be referenced again before it is swapped out.
4984 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
4985 * INACTIVE_ANON pages on this zone's LRU, maintained by the
4986 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
4987 * the anonymous pages are kept on the inactive list.
4990 * memory ratio inactive anon
4991 * -------------------------------------
5000 void calculate_zone_inactive_ratio(struct zone
*zone
)
5002 unsigned int gb
, ratio
;
5004 /* Zone size in gigabytes */
5005 gb
= zone
->present_pages
>> (30 - PAGE_SHIFT
);
5007 ratio
= int_sqrt(10 * gb
);
5011 zone
->inactive_ratio
= ratio
;
5014 static void __init
setup_per_zone_inactive_ratio(void)
5019 calculate_zone_inactive_ratio(zone
);
5023 * Initialise min_free_kbytes.
5025 * For small machines we want it small (128k min). For large machines
5026 * we want it large (64MB max). But it is not linear, because network
5027 * bandwidth does not increase linearly with machine size. We use
5029 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5030 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5046 static int __init
init_per_zone_wmark_min(void)
5048 unsigned long lowmem_kbytes
;
5050 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
5052 min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
5053 if (min_free_kbytes
< 128)
5054 min_free_kbytes
= 128;
5055 if (min_free_kbytes
> 65536)
5056 min_free_kbytes
= 65536;
5057 setup_per_zone_wmarks();
5058 setup_per_zone_lowmem_reserve();
5059 setup_per_zone_inactive_ratio();
5062 module_init(init_per_zone_wmark_min
)
5065 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5066 * that we can call two helper functions whenever min_free_kbytes
5069 int min_free_kbytes_sysctl_handler(ctl_table
*table
, int write
,
5070 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5072 proc_dointvec(table
, write
, buffer
, length
, ppos
);
5074 setup_per_zone_wmarks();
5079 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table
*table
, int write
,
5080 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5085 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5090 zone
->min_unmapped_pages
= (zone
->present_pages
*
5091 sysctl_min_unmapped_ratio
) / 100;
5095 int sysctl_min_slab_ratio_sysctl_handler(ctl_table
*table
, int write
,
5096 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5101 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5106 zone
->min_slab_pages
= (zone
->present_pages
*
5107 sysctl_min_slab_ratio
) / 100;
5113 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5114 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5115 * whenever sysctl_lowmem_reserve_ratio changes.
5117 * The reserve ratio obviously has absolutely no relation with the
5118 * minimum watermarks. The lowmem reserve ratio can only make sense
5119 * if in function of the boot time zone sizes.
5121 int lowmem_reserve_ratio_sysctl_handler(ctl_table
*table
, int write
,
5122 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5124 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5125 setup_per_zone_lowmem_reserve();
5130 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5131 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
5132 * can have before it gets flushed back to buddy allocator.
5135 int percpu_pagelist_fraction_sysctl_handler(ctl_table
*table
, int write
,
5136 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5142 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5143 if (!write
|| (ret
== -EINVAL
))
5145 for_each_populated_zone(zone
) {
5146 for_each_possible_cpu(cpu
) {
5148 high
= zone
->present_pages
/ percpu_pagelist_fraction
;
5149 setup_pagelist_highmark(
5150 per_cpu_ptr(zone
->pageset
, cpu
), high
);
5156 int hashdist
= HASHDIST_DEFAULT
;
5159 static int __init
set_hashdist(char *str
)
5163 hashdist
= simple_strtoul(str
, &str
, 0);
5166 __setup("hashdist=", set_hashdist
);
5170 * allocate a large system hash table from bootmem
5171 * - it is assumed that the hash table must contain an exact power-of-2
5172 * quantity of entries
5173 * - limit is the number of hash buckets, not the total allocation size
5175 void *__init
alloc_large_system_hash(const char *tablename
,
5176 unsigned long bucketsize
,
5177 unsigned long numentries
,
5180 unsigned int *_hash_shift
,
5181 unsigned int *_hash_mask
,
5182 unsigned long limit
)
5184 unsigned long long max
= limit
;
5185 unsigned long log2qty
, size
;
5188 /* allow the kernel cmdline to have a say */
5190 /* round applicable memory size up to nearest megabyte */
5191 numentries
= nr_kernel_pages
;
5192 numentries
+= (1UL << (20 - PAGE_SHIFT
)) - 1;
5193 numentries
>>= 20 - PAGE_SHIFT
;
5194 numentries
<<= 20 - PAGE_SHIFT
;
5196 /* limit to 1 bucket per 2^scale bytes of low memory */
5197 if (scale
> PAGE_SHIFT
)
5198 numentries
>>= (scale
- PAGE_SHIFT
);
5200 numentries
<<= (PAGE_SHIFT
- scale
);
5202 /* Make sure we've got at least a 0-order allocation.. */
5203 if (unlikely(flags
& HASH_SMALL
)) {
5204 /* Makes no sense without HASH_EARLY */
5205 WARN_ON(!(flags
& HASH_EARLY
));
5206 if (!(numentries
>> *_hash_shift
)) {
5207 numentries
= 1UL << *_hash_shift
;
5208 BUG_ON(!numentries
);
5210 } else if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
5211 numentries
= PAGE_SIZE
/ bucketsize
;
5213 numentries
= roundup_pow_of_two(numentries
);
5215 /* limit allocation size to 1/16 total memory by default */
5217 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
5218 do_div(max
, bucketsize
);
5221 if (numentries
> max
)
5224 log2qty
= ilog2(numentries
);
5227 size
= bucketsize
<< log2qty
;
5228 if (flags
& HASH_EARLY
)
5229 table
= alloc_bootmem_nopanic(size
);
5231 table
= __vmalloc(size
, GFP_ATOMIC
, PAGE_KERNEL
);
5234 * If bucketsize is not a power-of-two, we may free
5235 * some pages at the end of hash table which
5236 * alloc_pages_exact() automatically does
5238 if (get_order(size
) < MAX_ORDER
) {
5239 table
= alloc_pages_exact(size
, GFP_ATOMIC
);
5240 kmemleak_alloc(table
, size
, 1, GFP_ATOMIC
);
5243 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
5246 panic("Failed to allocate %s hash table\n", tablename
);
5248 printk(KERN_INFO
"%s hash table entries: %ld (order: %d, %lu bytes)\n",
5251 ilog2(size
) - PAGE_SHIFT
,
5255 *_hash_shift
= log2qty
;
5257 *_hash_mask
= (1 << log2qty
) - 1;
5262 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5263 static inline unsigned long *get_pageblock_bitmap(struct zone
*zone
,
5266 #ifdef CONFIG_SPARSEMEM
5267 return __pfn_to_section(pfn
)->pageblock_flags
;
5269 return zone
->pageblock_flags
;
5270 #endif /* CONFIG_SPARSEMEM */
5273 static inline int pfn_to_bitidx(struct zone
*zone
, unsigned long pfn
)
5275 #ifdef CONFIG_SPARSEMEM
5276 pfn
&= (PAGES_PER_SECTION
-1);
5277 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
5279 pfn
= pfn
- zone
->zone_start_pfn
;
5280 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
5281 #endif /* CONFIG_SPARSEMEM */
5285 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5286 * @page: The page within the block of interest
5287 * @start_bitidx: The first bit of interest to retrieve
5288 * @end_bitidx: The last bit of interest
5289 * returns pageblock_bits flags
5291 unsigned long get_pageblock_flags_group(struct page
*page
,
5292 int start_bitidx
, int end_bitidx
)
5295 unsigned long *bitmap
;
5296 unsigned long pfn
, bitidx
;
5297 unsigned long flags
= 0;
5298 unsigned long value
= 1;
5300 zone
= page_zone(page
);
5301 pfn
= page_to_pfn(page
);
5302 bitmap
= get_pageblock_bitmap(zone
, pfn
);
5303 bitidx
= pfn_to_bitidx(zone
, pfn
);
5305 for (; start_bitidx
<= end_bitidx
; start_bitidx
++, value
<<= 1)
5306 if (test_bit(bitidx
+ start_bitidx
, bitmap
))
5313 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5314 * @page: The page within the block of interest
5315 * @start_bitidx: The first bit of interest
5316 * @end_bitidx: The last bit of interest
5317 * @flags: The flags to set
5319 void set_pageblock_flags_group(struct page
*page
, unsigned long flags
,
5320 int start_bitidx
, int end_bitidx
)
5323 unsigned long *bitmap
;
5324 unsigned long pfn
, bitidx
;
5325 unsigned long value
= 1;
5327 zone
= page_zone(page
);
5328 pfn
= page_to_pfn(page
);
5329 bitmap
= get_pageblock_bitmap(zone
, pfn
);
5330 bitidx
= pfn_to_bitidx(zone
, pfn
);
5331 VM_BUG_ON(pfn
< zone
->zone_start_pfn
);
5332 VM_BUG_ON(pfn
>= zone
->zone_start_pfn
+ zone
->spanned_pages
);
5334 for (; start_bitidx
<= end_bitidx
; start_bitidx
++, value
<<= 1)
5336 __set_bit(bitidx
+ start_bitidx
, bitmap
);
5338 __clear_bit(bitidx
+ start_bitidx
, bitmap
);
5342 * This is designed as sub function...plz see page_isolation.c also.
5343 * set/clear page block's type to be ISOLATE.
5344 * page allocater never alloc memory from ISOLATE block.
5348 __count_immobile_pages(struct zone
*zone
, struct page
*page
, int count
)
5350 unsigned long pfn
, iter
, found
;
5352 * For avoiding noise data, lru_add_drain_all() should be called
5353 * If ZONE_MOVABLE, the zone never contains immobile pages
5355 if (zone_idx(zone
) == ZONE_MOVABLE
)
5358 if (get_pageblock_migratetype(page
) == MIGRATE_MOVABLE
)
5361 pfn
= page_to_pfn(page
);
5362 for (found
= 0, iter
= 0; iter
< pageblock_nr_pages
; iter
++) {
5363 unsigned long check
= pfn
+ iter
;
5365 if (!pfn_valid_within(check
)) {
5369 page
= pfn_to_page(check
);
5370 if (!page_count(page
)) {
5371 if (PageBuddy(page
))
5372 iter
+= (1 << page_order(page
)) - 1;
5378 * If there are RECLAIMABLE pages, we need to check it.
5379 * But now, memory offline itself doesn't call shrink_slab()
5380 * and it still to be fixed.
5383 * If the page is not RAM, page_count()should be 0.
5384 * we don't need more check. This is an _used_ not-movable page.
5386 * The problematic thing here is PG_reserved pages. PG_reserved
5387 * is set to both of a memory hole page and a _used_ kernel
5396 bool is_pageblock_removable_nolock(struct page
*page
)
5398 struct zone
*zone
= page_zone(page
);
5399 return __count_immobile_pages(zone
, page
, 0);
5402 int set_migratetype_isolate(struct page
*page
)
5405 unsigned long flags
, pfn
;
5406 struct memory_isolate_notify arg
;
5411 zone
= page_zone(page
);
5412 zone_idx
= zone_idx(zone
);
5414 spin_lock_irqsave(&zone
->lock
, flags
);
5416 pfn
= page_to_pfn(page
);
5417 arg
.start_pfn
= pfn
;
5418 arg
.nr_pages
= pageblock_nr_pages
;
5419 arg
.pages_found
= 0;
5422 * It may be possible to isolate a pageblock even if the
5423 * migratetype is not MIGRATE_MOVABLE. The memory isolation
5424 * notifier chain is used by balloon drivers to return the
5425 * number of pages in a range that are held by the balloon
5426 * driver to shrink memory. If all the pages are accounted for
5427 * by balloons, are free, or on the LRU, isolation can continue.
5428 * Later, for example, when memory hotplug notifier runs, these
5429 * pages reported as "can be isolated" should be isolated(freed)
5430 * by the balloon driver through the memory notifier chain.
5432 notifier_ret
= memory_isolate_notify(MEM_ISOLATE_COUNT
, &arg
);
5433 notifier_ret
= notifier_to_errno(notifier_ret
);
5437 * FIXME: Now, memory hotplug doesn't call shrink_slab() by itself.
5438 * We just check MOVABLE pages.
5440 if (__count_immobile_pages(zone
, page
, arg
.pages_found
))
5444 * immobile means "not-on-lru" paes. If immobile is larger than
5445 * removable-by-driver pages reported by notifier, we'll fail.
5450 set_pageblock_migratetype(page
, MIGRATE_ISOLATE
);
5451 move_freepages_block(zone
, page
, MIGRATE_ISOLATE
);
5454 spin_unlock_irqrestore(&zone
->lock
, flags
);
5460 void unset_migratetype_isolate(struct page
*page
)
5463 unsigned long flags
;
5464 zone
= page_zone(page
);
5465 spin_lock_irqsave(&zone
->lock
, flags
);
5466 if (get_pageblock_migratetype(page
) != MIGRATE_ISOLATE
)
5468 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
5469 move_freepages_block(zone
, page
, MIGRATE_MOVABLE
);
5471 spin_unlock_irqrestore(&zone
->lock
, flags
);
5474 #ifdef CONFIG_MEMORY_HOTREMOVE
5476 * All pages in the range must be isolated before calling this.
5479 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
5485 unsigned long flags
;
5486 /* find the first valid pfn */
5487 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
5492 zone
= page_zone(pfn_to_page(pfn
));
5493 spin_lock_irqsave(&zone
->lock
, flags
);
5495 while (pfn
< end_pfn
) {
5496 if (!pfn_valid(pfn
)) {
5500 page
= pfn_to_page(pfn
);
5501 BUG_ON(page_count(page
));
5502 BUG_ON(!PageBuddy(page
));
5503 order
= page_order(page
);
5504 #ifdef CONFIG_DEBUG_VM
5505 printk(KERN_INFO
"remove from free list %lx %d %lx\n",
5506 pfn
, 1 << order
, end_pfn
);
5508 list_del(&page
->lru
);
5509 rmv_page_order(page
);
5510 zone
->free_area
[order
].nr_free
--;
5511 __mod_zone_page_state(zone
, NR_FREE_PAGES
,
5513 for (i
= 0; i
< (1 << order
); i
++)
5514 SetPageReserved((page
+i
));
5515 pfn
+= (1 << order
);
5517 spin_unlock_irqrestore(&zone
->lock
, flags
);
5521 #ifdef CONFIG_MEMORY_FAILURE
5522 bool is_free_buddy_page(struct page
*page
)
5524 struct zone
*zone
= page_zone(page
);
5525 unsigned long pfn
= page_to_pfn(page
);
5526 unsigned long flags
;
5529 spin_lock_irqsave(&zone
->lock
, flags
);
5530 for (order
= 0; order
< MAX_ORDER
; order
++) {
5531 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
5533 if (PageBuddy(page_head
) && page_order(page_head
) >= order
)
5536 spin_unlock_irqrestore(&zone
->lock
, flags
);
5538 return order
< MAX_ORDER
;
5542 static struct trace_print_flags pageflag_names
[] = {
5543 {1UL << PG_locked
, "locked" },
5544 {1UL << PG_error
, "error" },
5545 {1UL << PG_referenced
, "referenced" },
5546 {1UL << PG_uptodate
, "uptodate" },
5547 {1UL << PG_dirty
, "dirty" },
5548 {1UL << PG_lru
, "lru" },
5549 {1UL << PG_active
, "active" },
5550 {1UL << PG_slab
, "slab" },
5551 {1UL << PG_owner_priv_1
, "owner_priv_1" },
5552 {1UL << PG_arch_1
, "arch_1" },
5553 {1UL << PG_reserved
, "reserved" },
5554 {1UL << PG_private
, "private" },
5555 {1UL << PG_private_2
, "private_2" },
5556 {1UL << PG_writeback
, "writeback" },
5557 #ifdef CONFIG_PAGEFLAGS_EXTENDED
5558 {1UL << PG_head
, "head" },
5559 {1UL << PG_tail
, "tail" },
5561 {1UL << PG_compound
, "compound" },
5563 {1UL << PG_swapcache
, "swapcache" },
5564 {1UL << PG_mappedtodisk
, "mappedtodisk" },
5565 {1UL << PG_reclaim
, "reclaim" },
5566 {1UL << PG_buddy
, "buddy" },
5567 {1UL << PG_swapbacked
, "swapbacked" },
5568 {1UL << PG_unevictable
, "unevictable" },
5570 {1UL << PG_mlocked
, "mlocked" },
5572 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
5573 {1UL << PG_uncached
, "uncached" },
5575 #ifdef CONFIG_MEMORY_FAILURE
5576 {1UL << PG_hwpoison
, "hwpoison" },
5581 static void dump_page_flags(unsigned long flags
)
5583 const char *delim
= "";
5587 printk(KERN_ALERT
"page flags: %#lx(", flags
);
5589 /* remove zone id */
5590 flags
&= (1UL << NR_PAGEFLAGS
) - 1;
5592 for (i
= 0; pageflag_names
[i
].name
&& flags
; i
++) {
5594 mask
= pageflag_names
[i
].mask
;
5595 if ((flags
& mask
) != mask
)
5599 printk("%s%s", delim
, pageflag_names
[i
].name
);
5603 /* check for left over flags */
5605 printk("%s%#lx", delim
, flags
);
5610 void dump_page(struct page
*page
)
5613 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
5614 page
, page_count(page
), page_mapcount(page
),
5615 page
->mapping
, page
->index
);
5616 dump_page_flags(page
->flags
);