2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kmemcheck.h>
28 #include <linux/module.h>
29 #include <linux/suspend.h>
30 #include <linux/pagevec.h>
31 #include <linux/blkdev.h>
32 #include <linux/slab.h>
33 #include <linux/oom.h>
34 #include <linux/notifier.h>
35 #include <linux/topology.h>
36 #include <linux/sysctl.h>
37 #include <linux/cpu.h>
38 #include <linux/cpuset.h>
39 #include <linux/memory_hotplug.h>
40 #include <linux/nodemask.h>
41 #include <linux/vmalloc.h>
42 #include <linux/mempolicy.h>
43 #include <linux/stop_machine.h>
44 #include <linux/sort.h>
45 #include <linux/pfn.h>
46 #include <linux/backing-dev.h>
47 #include <linux/fault-inject.h>
48 #include <linux/page-isolation.h>
49 #include <linux/page_cgroup.h>
50 #include <linux/debugobjects.h>
51 #include <linux/kmemleak.h>
52 #include <linux/memory.h>
53 #include <linux/compaction.h>
54 #include <trace/events/kmem.h>
55 #include <linux/ftrace_event.h>
56 #include <linux/memcontrol.h>
58 #include <asm/tlbflush.h>
59 #include <asm/div64.h>
62 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
63 DEFINE_PER_CPU(int, numa_node
);
64 EXPORT_PER_CPU_SYMBOL(numa_node
);
67 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
69 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
70 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
71 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
72 * defined in <linux/topology.h>.
74 DEFINE_PER_CPU(int, _numa_mem_
); /* Kernel "local memory" node */
75 EXPORT_PER_CPU_SYMBOL(_numa_mem_
);
79 * Array of node states.
81 nodemask_t node_states
[NR_NODE_STATES
] __read_mostly
= {
82 [N_POSSIBLE
] = NODE_MASK_ALL
,
83 [N_ONLINE
] = { { [0] = 1UL } },
85 [N_NORMAL_MEMORY
] = { { [0] = 1UL } },
87 [N_HIGH_MEMORY
] = { { [0] = 1UL } },
89 [N_CPU
] = { { [0] = 1UL } },
92 EXPORT_SYMBOL(node_states
);
94 unsigned long totalram_pages __read_mostly
;
95 unsigned long totalreserve_pages __read_mostly
;
96 int percpu_pagelist_fraction
;
97 gfp_t gfp_allowed_mask __read_mostly
= GFP_BOOT_MASK
;
99 #ifdef CONFIG_PM_SLEEP
101 * The following functions are used by the suspend/hibernate code to temporarily
102 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
103 * while devices are suspended. To avoid races with the suspend/hibernate code,
104 * they should always be called with pm_mutex held (gfp_allowed_mask also should
105 * only be modified with pm_mutex held, unless the suspend/hibernate code is
106 * guaranteed not to run in parallel with that modification).
109 static gfp_t saved_gfp_mask
;
111 void pm_restore_gfp_mask(void)
113 WARN_ON(!mutex_is_locked(&pm_mutex
));
114 if (saved_gfp_mask
) {
115 gfp_allowed_mask
= saved_gfp_mask
;
120 void pm_restrict_gfp_mask(void)
122 WARN_ON(!mutex_is_locked(&pm_mutex
));
123 WARN_ON(saved_gfp_mask
);
124 saved_gfp_mask
= gfp_allowed_mask
;
125 gfp_allowed_mask
&= ~GFP_IOFS
;
127 #endif /* CONFIG_PM_SLEEP */
129 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
130 int pageblock_order __read_mostly
;
133 static void __free_pages_ok(struct page
*page
, unsigned int order
);
136 * results with 256, 32 in the lowmem_reserve sysctl:
137 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
138 * 1G machine -> (16M dma, 784M normal, 224M high)
139 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
140 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
141 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
143 * TBD: should special case ZONE_DMA32 machines here - in those we normally
144 * don't need any ZONE_NORMAL reservation
146 int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
-1] = {
147 #ifdef CONFIG_ZONE_DMA
150 #ifdef CONFIG_ZONE_DMA32
153 #ifdef CONFIG_HIGHMEM
159 EXPORT_SYMBOL(totalram_pages
);
161 static char * const zone_names
[MAX_NR_ZONES
] = {
162 #ifdef CONFIG_ZONE_DMA
165 #ifdef CONFIG_ZONE_DMA32
169 #ifdef CONFIG_HIGHMEM
175 int min_free_kbytes
= 1024;
177 static unsigned long __meminitdata nr_kernel_pages
;
178 static unsigned long __meminitdata nr_all_pages
;
179 static unsigned long __meminitdata dma_reserve
;
181 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
183 * MAX_ACTIVE_REGIONS determines the maximum number of distinct
184 * ranges of memory (RAM) that may be registered with add_active_range().
185 * Ranges passed to add_active_range() will be merged if possible
186 * so the number of times add_active_range() can be called is
187 * related to the number of nodes and the number of holes
189 #ifdef CONFIG_MAX_ACTIVE_REGIONS
190 /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */
191 #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS
193 #if MAX_NUMNODES >= 32
194 /* If there can be many nodes, allow up to 50 holes per node */
195 #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50)
197 /* By default, allow up to 256 distinct regions */
198 #define MAX_ACTIVE_REGIONS 256
202 static struct node_active_region __meminitdata early_node_map
[MAX_ACTIVE_REGIONS
];
203 static int __meminitdata nr_nodemap_entries
;
204 static unsigned long __meminitdata arch_zone_lowest_possible_pfn
[MAX_NR_ZONES
];
205 static unsigned long __meminitdata arch_zone_highest_possible_pfn
[MAX_NR_ZONES
];
206 static unsigned long __initdata required_kernelcore
;
207 static unsigned long __initdata required_movablecore
;
208 static unsigned long __meminitdata zone_movable_pfn
[MAX_NUMNODES
];
210 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
212 EXPORT_SYMBOL(movable_zone
);
213 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
216 int nr_node_ids __read_mostly
= MAX_NUMNODES
;
217 int nr_online_nodes __read_mostly
= 1;
218 EXPORT_SYMBOL(nr_node_ids
);
219 EXPORT_SYMBOL(nr_online_nodes
);
222 int page_group_by_mobility_disabled __read_mostly
;
224 static void set_pageblock_migratetype(struct page
*page
, int migratetype
)
227 if (unlikely(page_group_by_mobility_disabled
))
228 migratetype
= MIGRATE_UNMOVABLE
;
230 set_pageblock_flags_group(page
, (unsigned long)migratetype
,
231 PB_migrate
, PB_migrate_end
);
234 bool oom_killer_disabled __read_mostly
;
236 #ifdef CONFIG_DEBUG_VM
237 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
241 unsigned long pfn
= page_to_pfn(page
);
244 seq
= zone_span_seqbegin(zone
);
245 if (pfn
>= zone
->zone_start_pfn
+ zone
->spanned_pages
)
247 else if (pfn
< zone
->zone_start_pfn
)
249 } while (zone_span_seqretry(zone
, seq
));
254 static int page_is_consistent(struct zone
*zone
, struct page
*page
)
256 if (!pfn_valid_within(page_to_pfn(page
)))
258 if (zone
!= page_zone(page
))
264 * Temporary debugging check for pages not lying within a given zone.
266 static int bad_range(struct zone
*zone
, struct page
*page
)
268 if (page_outside_zone_boundaries(zone
, page
))
270 if (!page_is_consistent(zone
, page
))
276 static inline int bad_range(struct zone
*zone
, struct page
*page
)
282 static void bad_page(struct page
*page
)
284 static unsigned long resume
;
285 static unsigned long nr_shown
;
286 static unsigned long nr_unshown
;
288 /* Don't complain about poisoned pages */
289 if (PageHWPoison(page
)) {
290 reset_page_mapcount(page
); /* remove PageBuddy */
295 * Allow a burst of 60 reports, then keep quiet for that minute;
296 * or allow a steady drip of one report per second.
298 if (nr_shown
== 60) {
299 if (time_before(jiffies
, resume
)) {
305 "BUG: Bad page state: %lu messages suppressed\n",
312 resume
= jiffies
+ 60 * HZ
;
314 printk(KERN_ALERT
"BUG: Bad page state in process %s pfn:%05lx\n",
315 current
->comm
, page_to_pfn(page
));
320 /* Leave bad fields for debug, except PageBuddy could make trouble */
321 reset_page_mapcount(page
); /* remove PageBuddy */
322 add_taint(TAINT_BAD_PAGE
);
326 * Higher-order pages are called "compound pages". They are structured thusly:
328 * The first PAGE_SIZE page is called the "head page".
330 * The remaining PAGE_SIZE pages are called "tail pages".
332 * All pages have PG_compound set. All pages have their ->private pointing at
333 * the head page (even the head page has this).
335 * The first tail page's ->lru.next holds the address of the compound page's
336 * put_page() function. Its ->lru.prev holds the order of allocation.
337 * This usage means that zero-order pages may not be compound.
340 static void free_compound_page(struct page
*page
)
342 __free_pages_ok(page
, compound_order(page
));
345 void prep_compound_page(struct page
*page
, unsigned long order
)
348 int nr_pages
= 1 << order
;
350 set_compound_page_dtor(page
, free_compound_page
);
351 set_compound_order(page
, order
);
353 for (i
= 1; i
< nr_pages
; i
++) {
354 struct page
*p
= page
+ i
;
357 p
->first_page
= page
;
361 /* update __split_huge_page_refcount if you change this function */
362 static int destroy_compound_page(struct page
*page
, unsigned long order
)
365 int nr_pages
= 1 << order
;
368 if (unlikely(compound_order(page
) != order
) ||
369 unlikely(!PageHead(page
))) {
374 __ClearPageHead(page
);
376 for (i
= 1; i
< nr_pages
; i
++) {
377 struct page
*p
= page
+ i
;
379 if (unlikely(!PageTail(p
) || (p
->first_page
!= page
))) {
389 static inline void prep_zero_page(struct page
*page
, int order
, gfp_t gfp_flags
)
394 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
395 * and __GFP_HIGHMEM from hard or soft interrupt context.
397 VM_BUG_ON((gfp_flags
& __GFP_HIGHMEM
) && in_interrupt());
398 for (i
= 0; i
< (1 << order
); i
++)
399 clear_highpage(page
+ i
);
402 static inline void set_page_order(struct page
*page
, int order
)
404 set_page_private(page
, order
);
405 __SetPageBuddy(page
);
408 static inline void rmv_page_order(struct page
*page
)
410 __ClearPageBuddy(page
);
411 set_page_private(page
, 0);
415 * Locate the struct page for both the matching buddy in our
416 * pair (buddy1) and the combined O(n+1) page they form (page).
418 * 1) Any buddy B1 will have an order O twin B2 which satisfies
419 * the following equation:
421 * For example, if the starting buddy (buddy2) is #8 its order
423 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
425 * 2) Any buddy B will have an order O+1 parent P which
426 * satisfies the following equation:
429 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
431 static inline unsigned long
432 __find_buddy_index(unsigned long page_idx
, unsigned int order
)
434 return page_idx
^ (1 << order
);
438 * This function checks whether a page is free && is the buddy
439 * we can do coalesce a page and its buddy if
440 * (a) the buddy is not in a hole &&
441 * (b) the buddy is in the buddy system &&
442 * (c) a page and its buddy have the same order &&
443 * (d) a page and its buddy are in the same zone.
445 * For recording whether a page is in the buddy system, we set ->_mapcount -2.
446 * Setting, clearing, and testing _mapcount -2 is serialized by zone->lock.
448 * For recording page's order, we use page_private(page).
450 static inline int page_is_buddy(struct page
*page
, struct page
*buddy
,
453 if (!pfn_valid_within(page_to_pfn(buddy
)))
456 if (page_zone_id(page
) != page_zone_id(buddy
))
459 if (PageBuddy(buddy
) && page_order(buddy
) == order
) {
460 VM_BUG_ON(page_count(buddy
) != 0);
467 * Freeing function for a buddy system allocator.
469 * The concept of a buddy system is to maintain direct-mapped table
470 * (containing bit values) for memory blocks of various "orders".
471 * The bottom level table contains the map for the smallest allocatable
472 * units of memory (here, pages), and each level above it describes
473 * pairs of units from the levels below, hence, "buddies".
474 * At a high level, all that happens here is marking the table entry
475 * at the bottom level available, and propagating the changes upward
476 * as necessary, plus some accounting needed to play nicely with other
477 * parts of the VM system.
478 * At each level, we keep a list of pages, which are heads of continuous
479 * free pages of length of (1 << order) and marked with _mapcount -2. Page's
480 * order is recorded in page_private(page) field.
481 * So when we are allocating or freeing one, we can derive the state of the
482 * other. That is, if we allocate a small block, and both were
483 * free, the remainder of the region must be split into blocks.
484 * If a block is freed, and its buddy is also free, then this
485 * triggers coalescing into a block of larger size.
490 static inline void __free_one_page(struct page
*page
,
491 struct zone
*zone
, unsigned int order
,
494 unsigned long page_idx
;
495 unsigned long combined_idx
;
496 unsigned long uninitialized_var(buddy_idx
);
499 if (unlikely(PageCompound(page
)))
500 if (unlikely(destroy_compound_page(page
, order
)))
503 VM_BUG_ON(migratetype
== -1);
505 page_idx
= page_to_pfn(page
) & ((1 << MAX_ORDER
) - 1);
507 VM_BUG_ON(page_idx
& ((1 << order
) - 1));
508 VM_BUG_ON(bad_range(zone
, page
));
510 while (order
< MAX_ORDER
-1) {
511 buddy_idx
= __find_buddy_index(page_idx
, order
);
512 buddy
= page
+ (buddy_idx
- page_idx
);
513 if (!page_is_buddy(page
, buddy
, order
))
516 /* Our buddy is free, merge with it and move up one order. */
517 list_del(&buddy
->lru
);
518 zone
->free_area
[order
].nr_free
--;
519 rmv_page_order(buddy
);
520 combined_idx
= buddy_idx
& page_idx
;
521 page
= page
+ (combined_idx
- page_idx
);
522 page_idx
= combined_idx
;
525 set_page_order(page
, order
);
528 * If this is not the largest possible page, check if the buddy
529 * of the next-highest order is free. If it is, it's possible
530 * that pages are being freed that will coalesce soon. In case,
531 * that is happening, add the free page to the tail of the list
532 * so it's less likely to be used soon and more likely to be merged
533 * as a higher order page
535 if ((order
< MAX_ORDER
-2) && pfn_valid_within(page_to_pfn(buddy
))) {
536 struct page
*higher_page
, *higher_buddy
;
537 combined_idx
= buddy_idx
& page_idx
;
538 higher_page
= page
+ (combined_idx
- page_idx
);
539 buddy_idx
= __find_buddy_index(combined_idx
, order
+ 1);
540 higher_buddy
= page
+ (buddy_idx
- combined_idx
);
541 if (page_is_buddy(higher_page
, higher_buddy
, order
+ 1)) {
542 list_add_tail(&page
->lru
,
543 &zone
->free_area
[order
].free_list
[migratetype
]);
548 list_add(&page
->lru
, &zone
->free_area
[order
].free_list
[migratetype
]);
550 zone
->free_area
[order
].nr_free
++;
554 * free_page_mlock() -- clean up attempts to free and mlocked() page.
555 * Page should not be on lru, so no need to fix that up.
556 * free_pages_check() will verify...
558 static inline void free_page_mlock(struct page
*page
)
560 __dec_zone_page_state(page
, NR_MLOCK
);
561 __count_vm_event(UNEVICTABLE_MLOCKFREED
);
564 static inline int free_pages_check(struct page
*page
)
566 if (unlikely(page_mapcount(page
) |
567 (page
->mapping
!= NULL
) |
568 (atomic_read(&page
->_count
) != 0) |
569 (page
->flags
& PAGE_FLAGS_CHECK_AT_FREE
) |
570 (mem_cgroup_bad_page_check(page
)))) {
574 if (page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)
575 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
580 * Frees a number of pages from the PCP lists
581 * Assumes all pages on list are in same zone, and of same order.
582 * count is the number of pages to free.
584 * If the zone was previously in an "all pages pinned" state then look to
585 * see if this freeing clears that state.
587 * And clear the zone's pages_scanned counter, to hold off the "all pages are
588 * pinned" detection logic.
590 static void free_pcppages_bulk(struct zone
*zone
, int count
,
591 struct per_cpu_pages
*pcp
)
597 spin_lock(&zone
->lock
);
598 zone
->all_unreclaimable
= 0;
599 zone
->pages_scanned
= 0;
603 struct list_head
*list
;
606 * Remove pages from lists in a round-robin fashion. A
607 * batch_free count is maintained that is incremented when an
608 * empty list is encountered. This is so more pages are freed
609 * off fuller lists instead of spinning excessively around empty
614 if (++migratetype
== MIGRATE_PCPTYPES
)
616 list
= &pcp
->lists
[migratetype
];
617 } while (list_empty(list
));
619 /* This is the only non-empty list. Free them all. */
620 if (batch_free
== MIGRATE_PCPTYPES
)
621 batch_free
= to_free
;
624 page
= list_entry(list
->prev
, struct page
, lru
);
625 /* must delete as __free_one_page list manipulates */
626 list_del(&page
->lru
);
627 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
628 __free_one_page(page
, zone
, 0, page_private(page
));
629 trace_mm_page_pcpu_drain(page
, 0, page_private(page
));
630 } while (--to_free
&& --batch_free
&& !list_empty(list
));
632 __mod_zone_page_state(zone
, NR_FREE_PAGES
, count
);
633 spin_unlock(&zone
->lock
);
636 static void free_one_page(struct zone
*zone
, struct page
*page
, int order
,
639 spin_lock(&zone
->lock
);
640 zone
->all_unreclaimable
= 0;
641 zone
->pages_scanned
= 0;
643 __free_one_page(page
, zone
, order
, migratetype
);
644 __mod_zone_page_state(zone
, NR_FREE_PAGES
, 1 << order
);
645 spin_unlock(&zone
->lock
);
648 static bool free_pages_prepare(struct page
*page
, unsigned int order
)
653 trace_mm_page_free_direct(page
, order
);
654 kmemcheck_free_shadow(page
, order
);
657 page
->mapping
= NULL
;
658 for (i
= 0; i
< (1 << order
); i
++)
659 bad
+= free_pages_check(page
+ i
);
663 if (!PageHighMem(page
)) {
664 debug_check_no_locks_freed(page_address(page
),PAGE_SIZE
<<order
);
665 debug_check_no_obj_freed(page_address(page
),
668 arch_free_page(page
, order
);
669 kernel_map_pages(page
, 1 << order
, 0);
674 static void __free_pages_ok(struct page
*page
, unsigned int order
)
677 int wasMlocked
= __TestClearPageMlocked(page
);
679 if (!free_pages_prepare(page
, order
))
682 local_irq_save(flags
);
683 if (unlikely(wasMlocked
))
684 free_page_mlock(page
);
685 __count_vm_events(PGFREE
, 1 << order
);
686 free_one_page(page_zone(page
), page
, order
,
687 get_pageblock_migratetype(page
));
688 local_irq_restore(flags
);
692 * permit the bootmem allocator to evade page validation on high-order frees
694 void __meminit
__free_pages_bootmem(struct page
*page
, unsigned int order
)
697 __ClearPageReserved(page
);
698 set_page_count(page
, 0);
699 set_page_refcounted(page
);
705 for (loop
= 0; loop
< BITS_PER_LONG
; loop
++) {
706 struct page
*p
= &page
[loop
];
708 if (loop
+ 1 < BITS_PER_LONG
)
710 __ClearPageReserved(p
);
711 set_page_count(p
, 0);
714 set_page_refcounted(page
);
715 __free_pages(page
, order
);
721 * The order of subdivision here is critical for the IO subsystem.
722 * Please do not alter this order without good reasons and regression
723 * testing. Specifically, as large blocks of memory are subdivided,
724 * the order in which smaller blocks are delivered depends on the order
725 * they're subdivided in this function. This is the primary factor
726 * influencing the order in which pages are delivered to the IO
727 * subsystem according to empirical testing, and this is also justified
728 * by considering the behavior of a buddy system containing a single
729 * large block of memory acted on by a series of small allocations.
730 * This behavior is a critical factor in sglist merging's success.
734 static inline void expand(struct zone
*zone
, struct page
*page
,
735 int low
, int high
, struct free_area
*area
,
738 unsigned long size
= 1 << high
;
744 VM_BUG_ON(bad_range(zone
, &page
[size
]));
745 list_add(&page
[size
].lru
, &area
->free_list
[migratetype
]);
747 set_page_order(&page
[size
], high
);
752 * This page is about to be returned from the page allocator
754 static inline int check_new_page(struct page
*page
)
756 if (unlikely(page_mapcount(page
) |
757 (page
->mapping
!= NULL
) |
758 (atomic_read(&page
->_count
) != 0) |
759 (page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
) |
760 (mem_cgroup_bad_page_check(page
)))) {
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_move(&page
->lru
,
874 &zone
->free_area
[order
].free_list
[migratetype
]);
876 pages_moved
+= 1 << order
;
882 static int move_freepages_block(struct zone
*zone
, struct page
*page
,
885 unsigned long start_pfn
, end_pfn
;
886 struct page
*start_page
, *end_page
;
888 start_pfn
= page_to_pfn(page
);
889 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
890 start_page
= pfn_to_page(start_pfn
);
891 end_page
= start_page
+ pageblock_nr_pages
- 1;
892 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
894 /* Do not cross zone boundaries */
895 if (start_pfn
< zone
->zone_start_pfn
)
897 if (end_pfn
>= zone
->zone_start_pfn
+ zone
->spanned_pages
)
900 return move_freepages(zone
, start_page
, end_page
, migratetype
);
903 static void change_pageblock_range(struct page
*pageblock_page
,
904 int start_order
, int migratetype
)
906 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
908 while (nr_pageblocks
--) {
909 set_pageblock_migratetype(pageblock_page
, migratetype
);
910 pageblock_page
+= pageblock_nr_pages
;
914 /* Remove an element from the buddy allocator from the fallback list */
915 static inline struct page
*
916 __rmqueue_fallback(struct zone
*zone
, int order
, int start_migratetype
)
918 struct free_area
* area
;
923 /* Find the largest possible block of pages in the other list */
924 for (current_order
= MAX_ORDER
-1; current_order
>= order
;
926 for (i
= 0; i
< MIGRATE_TYPES
- 1; i
++) {
927 migratetype
= fallbacks
[start_migratetype
][i
];
929 /* MIGRATE_RESERVE handled later if necessary */
930 if (migratetype
== MIGRATE_RESERVE
)
933 area
= &(zone
->free_area
[current_order
]);
934 if (list_empty(&area
->free_list
[migratetype
]))
937 page
= list_entry(area
->free_list
[migratetype
].next
,
942 * If breaking a large block of pages, move all free
943 * pages to the preferred allocation list. If falling
944 * back for a reclaimable kernel allocation, be more
945 * aggressive about taking ownership of free pages
947 if (unlikely(current_order
>= (pageblock_order
>> 1)) ||
948 start_migratetype
== MIGRATE_RECLAIMABLE
||
949 page_group_by_mobility_disabled
) {
951 pages
= move_freepages_block(zone
, page
,
954 /* Claim the whole block if over half of it is free */
955 if (pages
>= (1 << (pageblock_order
-1)) ||
956 page_group_by_mobility_disabled
)
957 set_pageblock_migratetype(page
,
960 migratetype
= start_migratetype
;
963 /* Remove the page from the freelists */
964 list_del(&page
->lru
);
965 rmv_page_order(page
);
967 /* Take ownership for orders >= pageblock_order */
968 if (current_order
>= pageblock_order
)
969 change_pageblock_range(page
, current_order
,
972 expand(zone
, page
, order
, current_order
, area
, migratetype
);
974 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
975 start_migratetype
, migratetype
);
985 * Do the hard work of removing an element from the buddy allocator.
986 * Call me with the zone->lock already held.
988 static struct page
*__rmqueue(struct zone
*zone
, unsigned int order
,
994 page
= __rmqueue_smallest(zone
, order
, migratetype
);
996 if (unlikely(!page
) && migratetype
!= MIGRATE_RESERVE
) {
997 page
= __rmqueue_fallback(zone
, order
, migratetype
);
1000 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1001 * is used because __rmqueue_smallest is an inline function
1002 * and we want just one call site
1005 migratetype
= MIGRATE_RESERVE
;
1010 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
1015 * Obtain a specified number of elements from the buddy allocator, all under
1016 * a single hold of the lock, for efficiency. Add them to the supplied list.
1017 * Returns the number of new pages which were placed at *list.
1019 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
1020 unsigned long count
, struct list_head
*list
,
1021 int migratetype
, int cold
)
1025 spin_lock(&zone
->lock
);
1026 for (i
= 0; i
< count
; ++i
) {
1027 struct page
*page
= __rmqueue(zone
, order
, migratetype
);
1028 if (unlikely(page
== NULL
))
1032 * Split buddy pages returned by expand() are received here
1033 * in physical page order. The page is added to the callers and
1034 * list and the list head then moves forward. From the callers
1035 * perspective, the linked list is ordered by page number in
1036 * some conditions. This is useful for IO devices that can
1037 * merge IO requests if the physical pages are ordered
1040 if (likely(cold
== 0))
1041 list_add(&page
->lru
, list
);
1043 list_add_tail(&page
->lru
, list
);
1044 set_page_private(page
, migratetype
);
1047 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
1048 spin_unlock(&zone
->lock
);
1054 * Called from the vmstat counter updater to drain pagesets of this
1055 * currently executing processor on remote nodes after they have
1058 * Note that this function must be called with the thread pinned to
1059 * a single processor.
1061 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
1063 unsigned long flags
;
1066 local_irq_save(flags
);
1067 if (pcp
->count
>= pcp
->batch
)
1068 to_drain
= pcp
->batch
;
1070 to_drain
= pcp
->count
;
1071 free_pcppages_bulk(zone
, to_drain
, pcp
);
1072 pcp
->count
-= to_drain
;
1073 local_irq_restore(flags
);
1078 * Drain pages of the indicated processor.
1080 * The processor must either be the current processor and the
1081 * thread pinned to the current processor or a processor that
1084 static void drain_pages(unsigned int cpu
)
1086 unsigned long flags
;
1089 for_each_populated_zone(zone
) {
1090 struct per_cpu_pageset
*pset
;
1091 struct per_cpu_pages
*pcp
;
1093 local_irq_save(flags
);
1094 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
1098 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
1101 local_irq_restore(flags
);
1106 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1108 void drain_local_pages(void *arg
)
1110 drain_pages(smp_processor_id());
1114 * Spill all the per-cpu pages from all CPUs back into the buddy allocator
1116 void drain_all_pages(void)
1118 on_each_cpu(drain_local_pages
, NULL
, 1);
1121 #ifdef CONFIG_HIBERNATION
1123 void mark_free_pages(struct zone
*zone
)
1125 unsigned long pfn
, max_zone_pfn
;
1126 unsigned long flags
;
1128 struct list_head
*curr
;
1130 if (!zone
->spanned_pages
)
1133 spin_lock_irqsave(&zone
->lock
, flags
);
1135 max_zone_pfn
= zone
->zone_start_pfn
+ zone
->spanned_pages
;
1136 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
1137 if (pfn_valid(pfn
)) {
1138 struct page
*page
= pfn_to_page(pfn
);
1140 if (!swsusp_page_is_forbidden(page
))
1141 swsusp_unset_page_free(page
);
1144 for_each_migratetype_order(order
, t
) {
1145 list_for_each(curr
, &zone
->free_area
[order
].free_list
[t
]) {
1148 pfn
= page_to_pfn(list_entry(curr
, struct page
, lru
));
1149 for (i
= 0; i
< (1UL << order
); i
++)
1150 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
1153 spin_unlock_irqrestore(&zone
->lock
, flags
);
1155 #endif /* CONFIG_PM */
1158 * Free a 0-order page
1159 * cold == 1 ? free a cold page : free a hot page
1161 void free_hot_cold_page(struct page
*page
, int cold
)
1163 struct zone
*zone
= page_zone(page
);
1164 struct per_cpu_pages
*pcp
;
1165 unsigned long flags
;
1167 int wasMlocked
= __TestClearPageMlocked(page
);
1169 if (!free_pages_prepare(page
, 0))
1172 migratetype
= get_pageblock_migratetype(page
);
1173 set_page_private(page
, migratetype
);
1174 local_irq_save(flags
);
1175 if (unlikely(wasMlocked
))
1176 free_page_mlock(page
);
1177 __count_vm_event(PGFREE
);
1180 * We only track unmovable, reclaimable and movable on pcp lists.
1181 * Free ISOLATE pages back to the allocator because they are being
1182 * offlined but treat RESERVE as movable pages so we can get those
1183 * areas back if necessary. Otherwise, we may have to free
1184 * excessively into the page allocator
1186 if (migratetype
>= MIGRATE_PCPTYPES
) {
1187 if (unlikely(migratetype
== MIGRATE_ISOLATE
)) {
1188 free_one_page(zone
, page
, 0, migratetype
);
1191 migratetype
= MIGRATE_MOVABLE
;
1194 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
1196 list_add_tail(&page
->lru
, &pcp
->lists
[migratetype
]);
1198 list_add(&page
->lru
, &pcp
->lists
[migratetype
]);
1200 if (pcp
->count
>= pcp
->high
) {
1201 free_pcppages_bulk(zone
, pcp
->batch
, pcp
);
1202 pcp
->count
-= pcp
->batch
;
1206 local_irq_restore(flags
);
1210 * split_page takes a non-compound higher-order page, and splits it into
1211 * n (1<<order) sub-pages: page[0..n]
1212 * Each sub-page must be freed individually.
1214 * Note: this is probably too low level an operation for use in drivers.
1215 * Please consult with lkml before using this in your driver.
1217 void split_page(struct page
*page
, unsigned int order
)
1221 VM_BUG_ON(PageCompound(page
));
1222 VM_BUG_ON(!page_count(page
));
1224 #ifdef CONFIG_KMEMCHECK
1226 * Split shadow pages too, because free(page[0]) would
1227 * otherwise free the whole shadow.
1229 if (kmemcheck_page_is_tracked(page
))
1230 split_page(virt_to_page(page
[0].shadow
), order
);
1233 for (i
= 1; i
< (1 << order
); i
++)
1234 set_page_refcounted(page
+ i
);
1238 * Similar to split_page except the page is already free. As this is only
1239 * being used for migration, the migratetype of the block also changes.
1240 * As this is called with interrupts disabled, the caller is responsible
1241 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1244 * Note: this is probably too low level an operation for use in drivers.
1245 * Please consult with lkml before using this in your driver.
1247 int split_free_page(struct page
*page
)
1250 unsigned long watermark
;
1253 BUG_ON(!PageBuddy(page
));
1255 zone
= page_zone(page
);
1256 order
= page_order(page
);
1258 /* Obey watermarks as if the page was being allocated */
1259 watermark
= low_wmark_pages(zone
) + (1 << order
);
1260 if (!zone_watermark_ok(zone
, 0, watermark
, 0, 0))
1263 /* Remove page from free list */
1264 list_del(&page
->lru
);
1265 zone
->free_area
[order
].nr_free
--;
1266 rmv_page_order(page
);
1267 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(1UL << order
));
1269 /* Split into individual pages */
1270 set_page_refcounted(page
);
1271 split_page(page
, order
);
1273 if (order
>= pageblock_order
- 1) {
1274 struct page
*endpage
= page
+ (1 << order
) - 1;
1275 for (; page
< endpage
; page
+= pageblock_nr_pages
)
1276 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
1283 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1284 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1288 struct page
*buffered_rmqueue(struct zone
*preferred_zone
,
1289 struct zone
*zone
, int order
, gfp_t gfp_flags
,
1292 unsigned long flags
;
1294 int cold
= !!(gfp_flags
& __GFP_COLD
);
1297 if (likely(order
== 0)) {
1298 struct per_cpu_pages
*pcp
;
1299 struct list_head
*list
;
1301 local_irq_save(flags
);
1302 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
1303 list
= &pcp
->lists
[migratetype
];
1304 if (list_empty(list
)) {
1305 pcp
->count
+= rmqueue_bulk(zone
, 0,
1308 if (unlikely(list_empty(list
)))
1313 page
= list_entry(list
->prev
, struct page
, lru
);
1315 page
= list_entry(list
->next
, struct page
, lru
);
1317 list_del(&page
->lru
);
1320 if (unlikely(gfp_flags
& __GFP_NOFAIL
)) {
1322 * __GFP_NOFAIL is not to be used in new code.
1324 * All __GFP_NOFAIL callers should be fixed so that they
1325 * properly detect and handle allocation failures.
1327 * We most definitely don't want callers attempting to
1328 * allocate greater than order-1 page units with
1331 WARN_ON_ONCE(order
> 1);
1333 spin_lock_irqsave(&zone
->lock
, flags
);
1334 page
= __rmqueue(zone
, order
, migratetype
);
1335 spin_unlock(&zone
->lock
);
1338 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(1 << order
));
1341 __count_zone_vm_events(PGALLOC
, zone
, 1 << order
);
1342 zone_statistics(preferred_zone
, zone
, gfp_flags
);
1343 local_irq_restore(flags
);
1345 VM_BUG_ON(bad_range(zone
, page
));
1346 if (prep_new_page(page
, order
, gfp_flags
))
1351 local_irq_restore(flags
);
1355 /* The ALLOC_WMARK bits are used as an index to zone->watermark */
1356 #define ALLOC_WMARK_MIN WMARK_MIN
1357 #define ALLOC_WMARK_LOW WMARK_LOW
1358 #define ALLOC_WMARK_HIGH WMARK_HIGH
1359 #define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */
1361 /* Mask to get the watermark bits */
1362 #define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1)
1364 #define ALLOC_HARDER 0x10 /* try to alloc harder */
1365 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
1366 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
1368 #ifdef CONFIG_FAIL_PAGE_ALLOC
1370 static struct fail_page_alloc_attr
{
1371 struct fault_attr attr
;
1373 u32 ignore_gfp_highmem
;
1374 u32 ignore_gfp_wait
;
1377 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1379 struct dentry
*ignore_gfp_highmem_file
;
1380 struct dentry
*ignore_gfp_wait_file
;
1381 struct dentry
*min_order_file
;
1383 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1385 } fail_page_alloc
= {
1386 .attr
= FAULT_ATTR_INITIALIZER
,
1387 .ignore_gfp_wait
= 1,
1388 .ignore_gfp_highmem
= 1,
1392 static int __init
setup_fail_page_alloc(char *str
)
1394 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
1396 __setup("fail_page_alloc=", setup_fail_page_alloc
);
1398 static int should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1400 if (order
< fail_page_alloc
.min_order
)
1402 if (gfp_mask
& __GFP_NOFAIL
)
1404 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
1406 if (fail_page_alloc
.ignore_gfp_wait
&& (gfp_mask
& __GFP_WAIT
))
1409 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
1412 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1414 static int __init
fail_page_alloc_debugfs(void)
1416 mode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
1420 err
= init_fault_attr_dentries(&fail_page_alloc
.attr
,
1424 dir
= fail_page_alloc
.attr
.dentries
.dir
;
1426 fail_page_alloc
.ignore_gfp_wait_file
=
1427 debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
1428 &fail_page_alloc
.ignore_gfp_wait
);
1430 fail_page_alloc
.ignore_gfp_highmem_file
=
1431 debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
1432 &fail_page_alloc
.ignore_gfp_highmem
);
1433 fail_page_alloc
.min_order_file
=
1434 debugfs_create_u32("min-order", mode
, dir
,
1435 &fail_page_alloc
.min_order
);
1437 if (!fail_page_alloc
.ignore_gfp_wait_file
||
1438 !fail_page_alloc
.ignore_gfp_highmem_file
||
1439 !fail_page_alloc
.min_order_file
) {
1441 debugfs_remove(fail_page_alloc
.ignore_gfp_wait_file
);
1442 debugfs_remove(fail_page_alloc
.ignore_gfp_highmem_file
);
1443 debugfs_remove(fail_page_alloc
.min_order_file
);
1444 cleanup_fault_attr_dentries(&fail_page_alloc
.attr
);
1450 late_initcall(fail_page_alloc_debugfs
);
1452 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1454 #else /* CONFIG_FAIL_PAGE_ALLOC */
1456 static inline int should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1461 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1464 * Return true if free pages are above 'mark'. This takes into account the order
1465 * of the allocation.
1467 static bool __zone_watermark_ok(struct zone
*z
, int order
, unsigned long mark
,
1468 int classzone_idx
, int alloc_flags
, long free_pages
)
1470 /* free_pages my go negative - that's OK */
1474 free_pages
-= (1 << order
) + 1;
1475 if (alloc_flags
& ALLOC_HIGH
)
1477 if (alloc_flags
& ALLOC_HARDER
)
1480 if (free_pages
<= min
+ z
->lowmem_reserve
[classzone_idx
])
1482 for (o
= 0; o
< order
; o
++) {
1483 /* At the next order, this order's pages become unavailable */
1484 free_pages
-= z
->free_area
[o
].nr_free
<< o
;
1486 /* Require fewer higher order pages to be free */
1489 if (free_pages
<= min
)
1495 bool zone_watermark_ok(struct zone
*z
, int order
, unsigned long mark
,
1496 int classzone_idx
, int alloc_flags
)
1498 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
1499 zone_page_state(z
, NR_FREE_PAGES
));
1502 bool zone_watermark_ok_safe(struct zone
*z
, int order
, unsigned long mark
,
1503 int classzone_idx
, int alloc_flags
)
1505 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
1507 if (z
->percpu_drift_mark
&& free_pages
< z
->percpu_drift_mark
)
1508 free_pages
= zone_page_state_snapshot(z
, NR_FREE_PAGES
);
1510 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
1516 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1517 * skip over zones that are not allowed by the cpuset, or that have
1518 * been recently (in last second) found to be nearly full. See further
1519 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1520 * that have to skip over a lot of full or unallowed zones.
1522 * If the zonelist cache is present in the passed in zonelist, then
1523 * returns a pointer to the allowed node mask (either the current
1524 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1526 * If the zonelist cache is not available for this zonelist, does
1527 * nothing and returns NULL.
1529 * If the fullzones BITMAP in the zonelist cache is stale (more than
1530 * a second since last zap'd) then we zap it out (clear its bits.)
1532 * We hold off even calling zlc_setup, until after we've checked the
1533 * first zone in the zonelist, on the theory that most allocations will
1534 * be satisfied from that first zone, so best to examine that zone as
1535 * quickly as we can.
1537 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1539 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1540 nodemask_t
*allowednodes
; /* zonelist_cache approximation */
1542 zlc
= zonelist
->zlcache_ptr
;
1546 if (time_after(jiffies
, zlc
->last_full_zap
+ HZ
)) {
1547 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
1548 zlc
->last_full_zap
= jiffies
;
1551 allowednodes
= !in_interrupt() && (alloc_flags
& ALLOC_CPUSET
) ?
1552 &cpuset_current_mems_allowed
:
1553 &node_states
[N_HIGH_MEMORY
];
1554 return allowednodes
;
1558 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1559 * if it is worth looking at further for free memory:
1560 * 1) Check that the zone isn't thought to be full (doesn't have its
1561 * bit set in the zonelist_cache fullzones BITMAP).
1562 * 2) Check that the zones node (obtained from the zonelist_cache
1563 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1564 * Return true (non-zero) if zone is worth looking at further, or
1565 * else return false (zero) if it is not.
1567 * This check -ignores- the distinction between various watermarks,
1568 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1569 * found to be full for any variation of these watermarks, it will
1570 * be considered full for up to one second by all requests, unless
1571 * we are so low on memory on all allowed nodes that we are forced
1572 * into the second scan of the zonelist.
1574 * In the second scan we ignore this zonelist cache and exactly
1575 * apply the watermarks to all zones, even it is slower to do so.
1576 * We are low on memory in the second scan, and should leave no stone
1577 * unturned looking for a free page.
1579 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1580 nodemask_t
*allowednodes
)
1582 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1583 int i
; /* index of *z in zonelist zones */
1584 int n
; /* node that zone *z is on */
1586 zlc
= zonelist
->zlcache_ptr
;
1590 i
= z
- zonelist
->_zonerefs
;
1593 /* This zone is worth trying if it is allowed but not full */
1594 return node_isset(n
, *allowednodes
) && !test_bit(i
, zlc
->fullzones
);
1598 * Given 'z' scanning a zonelist, set the corresponding bit in
1599 * zlc->fullzones, so that subsequent attempts to allocate a page
1600 * from that zone don't waste time re-examining it.
1602 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1604 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1605 int i
; /* index of *z in zonelist zones */
1607 zlc
= zonelist
->zlcache_ptr
;
1611 i
= z
- zonelist
->_zonerefs
;
1613 set_bit(i
, zlc
->fullzones
);
1616 #else /* CONFIG_NUMA */
1618 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1623 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1624 nodemask_t
*allowednodes
)
1629 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1632 #endif /* CONFIG_NUMA */
1635 * get_page_from_freelist goes through the zonelist trying to allocate
1638 static struct page
*
1639 get_page_from_freelist(gfp_t gfp_mask
, nodemask_t
*nodemask
, unsigned int order
,
1640 struct zonelist
*zonelist
, int high_zoneidx
, int alloc_flags
,
1641 struct zone
*preferred_zone
, int migratetype
)
1644 struct page
*page
= NULL
;
1647 nodemask_t
*allowednodes
= NULL
;/* zonelist_cache approximation */
1648 int zlc_active
= 0; /* set if using zonelist_cache */
1649 int did_zlc_setup
= 0; /* just call zlc_setup() one time */
1651 classzone_idx
= zone_idx(preferred_zone
);
1654 * Scan zonelist, looking for a zone with enough free.
1655 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1657 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
1658 high_zoneidx
, nodemask
) {
1659 if (NUMA_BUILD
&& zlc_active
&&
1660 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
1662 if ((alloc_flags
& ALLOC_CPUSET
) &&
1663 !cpuset_zone_allowed_softwall(zone
, gfp_mask
))
1666 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
1667 if (!(alloc_flags
& ALLOC_NO_WATERMARKS
)) {
1671 mark
= zone
->watermark
[alloc_flags
& ALLOC_WMARK_MASK
];
1672 if (zone_watermark_ok(zone
, order
, mark
,
1673 classzone_idx
, alloc_flags
))
1676 if (zone_reclaim_mode
== 0)
1677 goto this_zone_full
;
1679 ret
= zone_reclaim(zone
, gfp_mask
, order
);
1681 case ZONE_RECLAIM_NOSCAN
:
1684 case ZONE_RECLAIM_FULL
:
1685 /* scanned but unreclaimable */
1686 goto this_zone_full
;
1688 /* did we reclaim enough */
1689 if (!zone_watermark_ok(zone
, order
, mark
,
1690 classzone_idx
, alloc_flags
))
1691 goto this_zone_full
;
1696 page
= buffered_rmqueue(preferred_zone
, zone
, order
,
1697 gfp_mask
, migratetype
);
1702 zlc_mark_zone_full(zonelist
, z
);
1704 if (NUMA_BUILD
&& !did_zlc_setup
&& nr_online_nodes
> 1) {
1706 * we do zlc_setup after the first zone is tried but only
1707 * if there are multiple nodes make it worthwhile
1709 allowednodes
= zlc_setup(zonelist
, alloc_flags
);
1715 if (unlikely(NUMA_BUILD
&& page
== NULL
&& zlc_active
)) {
1716 /* Disable zlc cache for second zonelist scan */
1724 * Large machines with many possible nodes should not always dump per-node
1725 * meminfo in irq context.
1727 static inline bool should_suppress_show_mem(void)
1732 ret
= in_interrupt();
1738 should_alloc_retry(gfp_t gfp_mask
, unsigned int order
,
1739 unsigned long pages_reclaimed
)
1741 /* Do not loop if specifically requested */
1742 if (gfp_mask
& __GFP_NORETRY
)
1746 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
1747 * means __GFP_NOFAIL, but that may not be true in other
1750 if (order
<= PAGE_ALLOC_COSTLY_ORDER
)
1754 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
1755 * specified, then we retry until we no longer reclaim any pages
1756 * (above), or we've reclaimed an order of pages at least as
1757 * large as the allocation's order. In both cases, if the
1758 * allocation still fails, we stop retrying.
1760 if (gfp_mask
& __GFP_REPEAT
&& pages_reclaimed
< (1 << order
))
1764 * Don't let big-order allocations loop unless the caller
1765 * explicitly requests that.
1767 if (gfp_mask
& __GFP_NOFAIL
)
1773 static inline struct page
*
1774 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
1775 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1776 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
1781 /* Acquire the OOM killer lock for the zones in zonelist */
1782 if (!try_set_zonelist_oom(zonelist
, gfp_mask
)) {
1783 schedule_timeout_uninterruptible(1);
1788 * Go through the zonelist yet one more time, keep very high watermark
1789 * here, this is only to catch a parallel oom killing, we must fail if
1790 * we're still under heavy pressure.
1792 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
,
1793 order
, zonelist
, high_zoneidx
,
1794 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
,
1795 preferred_zone
, migratetype
);
1799 if (!(gfp_mask
& __GFP_NOFAIL
)) {
1800 /* The OOM killer will not help higher order allocs */
1801 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
1803 /* The OOM killer does not needlessly kill tasks for lowmem */
1804 if (high_zoneidx
< ZONE_NORMAL
)
1807 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
1808 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
1809 * The caller should handle page allocation failure by itself if
1810 * it specifies __GFP_THISNODE.
1811 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
1813 if (gfp_mask
& __GFP_THISNODE
)
1816 /* Exhausted what can be done so it's blamo time */
1817 out_of_memory(zonelist
, gfp_mask
, order
, nodemask
);
1820 clear_zonelist_oom(zonelist
, gfp_mask
);
1824 #ifdef CONFIG_COMPACTION
1825 /* Try memory compaction for high-order allocations before reclaim */
1826 static struct page
*
1827 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
1828 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1829 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
1830 int migratetype
, unsigned long *did_some_progress
,
1831 bool sync_migration
)
1835 if (!order
|| compaction_deferred(preferred_zone
))
1838 current
->flags
|= PF_MEMALLOC
;
1839 *did_some_progress
= try_to_compact_pages(zonelist
, order
, gfp_mask
,
1840 nodemask
, sync_migration
);
1841 current
->flags
&= ~PF_MEMALLOC
;
1842 if (*did_some_progress
!= COMPACT_SKIPPED
) {
1844 /* Page migration frees to the PCP lists but we want merging */
1845 drain_pages(get_cpu());
1848 page
= get_page_from_freelist(gfp_mask
, nodemask
,
1849 order
, zonelist
, high_zoneidx
,
1850 alloc_flags
, preferred_zone
,
1853 preferred_zone
->compact_considered
= 0;
1854 preferred_zone
->compact_defer_shift
= 0;
1855 count_vm_event(COMPACTSUCCESS
);
1860 * It's bad if compaction run occurs and fails.
1861 * The most likely reason is that pages exist,
1862 * but not enough to satisfy watermarks.
1864 count_vm_event(COMPACTFAIL
);
1865 defer_compaction(preferred_zone
);
1873 static inline struct page
*
1874 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
1875 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1876 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
1877 int migratetype
, unsigned long *did_some_progress
,
1878 bool sync_migration
)
1882 #endif /* CONFIG_COMPACTION */
1884 /* The really slow allocator path where we enter direct reclaim */
1885 static inline struct page
*
1886 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
1887 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1888 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
1889 int migratetype
, unsigned long *did_some_progress
)
1891 struct page
*page
= NULL
;
1892 struct reclaim_state reclaim_state
;
1893 bool drained
= false;
1897 /* We now go into synchronous reclaim */
1898 cpuset_memory_pressure_bump();
1899 current
->flags
|= PF_MEMALLOC
;
1900 lockdep_set_current_reclaim_state(gfp_mask
);
1901 reclaim_state
.reclaimed_slab
= 0;
1902 current
->reclaim_state
= &reclaim_state
;
1904 *did_some_progress
= try_to_free_pages(zonelist
, order
, gfp_mask
, nodemask
);
1906 current
->reclaim_state
= NULL
;
1907 lockdep_clear_current_reclaim_state();
1908 current
->flags
&= ~PF_MEMALLOC
;
1912 if (unlikely(!(*did_some_progress
)))
1916 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
1917 zonelist
, high_zoneidx
,
1918 alloc_flags
, preferred_zone
,
1922 * If an allocation failed after direct reclaim, it could be because
1923 * pages are pinned on the per-cpu lists. Drain them and try again
1925 if (!page
&& !drained
) {
1935 * This is called in the allocator slow-path if the allocation request is of
1936 * sufficient urgency to ignore watermarks and take other desperate measures
1938 static inline struct page
*
1939 __alloc_pages_high_priority(gfp_t gfp_mask
, unsigned int order
,
1940 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1941 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
1947 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
1948 zonelist
, high_zoneidx
, ALLOC_NO_WATERMARKS
,
1949 preferred_zone
, migratetype
);
1951 if (!page
&& gfp_mask
& __GFP_NOFAIL
)
1952 wait_iff_congested(preferred_zone
, BLK_RW_ASYNC
, HZ
/50);
1953 } while (!page
&& (gfp_mask
& __GFP_NOFAIL
));
1959 void wake_all_kswapd(unsigned int order
, struct zonelist
*zonelist
,
1960 enum zone_type high_zoneidx
,
1961 enum zone_type classzone_idx
)
1966 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
)
1967 wakeup_kswapd(zone
, order
, classzone_idx
);
1971 gfp_to_alloc_flags(gfp_t gfp_mask
)
1973 int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
1974 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
1976 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
1977 BUILD_BUG_ON(__GFP_HIGH
!= (__force gfp_t
) ALLOC_HIGH
);
1980 * The caller may dip into page reserves a bit more if the caller
1981 * cannot run direct reclaim, or if the caller has realtime scheduling
1982 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
1983 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
1985 alloc_flags
|= (__force
int) (gfp_mask
& __GFP_HIGH
);
1989 * Not worth trying to allocate harder for
1990 * __GFP_NOMEMALLOC even if it can't schedule.
1992 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
1993 alloc_flags
|= ALLOC_HARDER
;
1995 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
1996 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1998 alloc_flags
&= ~ALLOC_CPUSET
;
1999 } else if (unlikely(rt_task(current
)) && !in_interrupt())
2000 alloc_flags
|= ALLOC_HARDER
;
2002 if (likely(!(gfp_mask
& __GFP_NOMEMALLOC
))) {
2003 if (!in_interrupt() &&
2004 ((current
->flags
& PF_MEMALLOC
) ||
2005 unlikely(test_thread_flag(TIF_MEMDIE
))))
2006 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2012 static inline struct page
*
2013 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
2014 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2015 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
2018 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
2019 struct page
*page
= NULL
;
2021 unsigned long pages_reclaimed
= 0;
2022 unsigned long did_some_progress
;
2023 bool sync_migration
= false;
2026 * In the slowpath, we sanity check order to avoid ever trying to
2027 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2028 * be using allocators in order of preference for an area that is
2031 if (order
>= MAX_ORDER
) {
2032 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
2037 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2038 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2039 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2040 * using a larger set of nodes after it has established that the
2041 * allowed per node queues are empty and that nodes are
2044 if (NUMA_BUILD
&& (gfp_mask
& GFP_THISNODE
) == GFP_THISNODE
)
2048 if (!(gfp_mask
& __GFP_NO_KSWAPD
))
2049 wake_all_kswapd(order
, zonelist
, high_zoneidx
,
2050 zone_idx(preferred_zone
));
2053 * OK, we're below the kswapd watermark and have kicked background
2054 * reclaim. Now things get more complex, so set up alloc_flags according
2055 * to how we want to proceed.
2057 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
2060 * Find the true preferred zone if the allocation is unconstrained by
2063 if (!(alloc_flags
& ALLOC_CPUSET
) && !nodemask
)
2064 first_zones_zonelist(zonelist
, high_zoneidx
, NULL
,
2067 /* This is the last chance, in general, before the goto nopage. */
2068 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
, zonelist
,
2069 high_zoneidx
, alloc_flags
& ~ALLOC_NO_WATERMARKS
,
2070 preferred_zone
, migratetype
);
2075 /* Allocate without watermarks if the context allows */
2076 if (alloc_flags
& ALLOC_NO_WATERMARKS
) {
2077 page
= __alloc_pages_high_priority(gfp_mask
, order
,
2078 zonelist
, high_zoneidx
, nodemask
,
2079 preferred_zone
, migratetype
);
2084 /* Atomic allocations - we can't balance anything */
2088 /* Avoid recursion of direct reclaim */
2089 if (current
->flags
& PF_MEMALLOC
)
2092 /* Avoid allocations with no watermarks from looping endlessly */
2093 if (test_thread_flag(TIF_MEMDIE
) && !(gfp_mask
& __GFP_NOFAIL
))
2097 * Try direct compaction. The first pass is asynchronous. Subsequent
2098 * attempts after direct reclaim are synchronous
2100 page
= __alloc_pages_direct_compact(gfp_mask
, order
,
2101 zonelist
, high_zoneidx
,
2103 alloc_flags
, preferred_zone
,
2104 migratetype
, &did_some_progress
,
2108 sync_migration
= !(gfp_mask
& __GFP_NO_KSWAPD
);
2110 /* Try direct reclaim and then allocating */
2111 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
,
2112 zonelist
, high_zoneidx
,
2114 alloc_flags
, preferred_zone
,
2115 migratetype
, &did_some_progress
);
2120 * If we failed to make any progress reclaiming, then we are
2121 * running out of options and have to consider going OOM
2123 if (!did_some_progress
) {
2124 if ((gfp_mask
& __GFP_FS
) && !(gfp_mask
& __GFP_NORETRY
)) {
2125 if (oom_killer_disabled
)
2127 page
= __alloc_pages_may_oom(gfp_mask
, order
,
2128 zonelist
, high_zoneidx
,
2129 nodemask
, preferred_zone
,
2134 if (!(gfp_mask
& __GFP_NOFAIL
)) {
2136 * The oom killer is not called for high-order
2137 * allocations that may fail, so if no progress
2138 * is being made, there are no other options and
2139 * retrying is unlikely to help.
2141 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
2144 * The oom killer is not called for lowmem
2145 * allocations to prevent needlessly killing
2148 if (high_zoneidx
< ZONE_NORMAL
)
2156 /* Check if we should retry the allocation */
2157 pages_reclaimed
+= did_some_progress
;
2158 if (should_alloc_retry(gfp_mask
, order
, pages_reclaimed
)) {
2159 /* Wait for some write requests to complete then retry */
2160 wait_iff_congested(preferred_zone
, BLK_RW_ASYNC
, HZ
/50);
2164 * High-order allocations do not necessarily loop after
2165 * direct reclaim and reclaim/compaction depends on compaction
2166 * being called after reclaim so call directly if necessary
2168 page
= __alloc_pages_direct_compact(gfp_mask
, order
,
2169 zonelist
, high_zoneidx
,
2171 alloc_flags
, preferred_zone
,
2172 migratetype
, &did_some_progress
,
2179 if (!(gfp_mask
& __GFP_NOWARN
) && printk_ratelimit()) {
2180 unsigned int filter
= SHOW_MEM_FILTER_NODES
;
2183 * This documents exceptions given to allocations in certain
2184 * contexts that are allowed to allocate outside current's set
2187 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
2188 if (test_thread_flag(TIF_MEMDIE
) ||
2189 (current
->flags
& (PF_MEMALLOC
| PF_EXITING
)))
2190 filter
&= ~SHOW_MEM_FILTER_NODES
;
2191 if (in_interrupt() || !wait
)
2192 filter
&= ~SHOW_MEM_FILTER_NODES
;
2194 pr_warning("%s: page allocation failure. order:%d, mode:0x%x\n",
2195 current
->comm
, order
, gfp_mask
);
2197 if (!should_suppress_show_mem())
2202 if (kmemcheck_enabled
)
2203 kmemcheck_pagealloc_alloc(page
, order
, gfp_mask
);
2209 * This is the 'heart' of the zoned buddy allocator.
2212 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
,
2213 struct zonelist
*zonelist
, nodemask_t
*nodemask
)
2215 enum zone_type high_zoneidx
= gfp_zone(gfp_mask
);
2216 struct zone
*preferred_zone
;
2218 int migratetype
= allocflags_to_migratetype(gfp_mask
);
2220 gfp_mask
&= gfp_allowed_mask
;
2222 lockdep_trace_alloc(gfp_mask
);
2224 might_sleep_if(gfp_mask
& __GFP_WAIT
);
2226 if (should_fail_alloc_page(gfp_mask
, order
))
2230 * Check the zones suitable for the gfp_mask contain at least one
2231 * valid zone. It's possible to have an empty zonelist as a result
2232 * of GFP_THISNODE and a memoryless node
2234 if (unlikely(!zonelist
->_zonerefs
->zone
))
2238 /* The preferred zone is used for statistics later */
2239 first_zones_zonelist(zonelist
, high_zoneidx
,
2240 nodemask
? : &cpuset_current_mems_allowed
,
2242 if (!preferred_zone
) {
2247 /* First allocation attempt */
2248 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
, order
,
2249 zonelist
, high_zoneidx
, ALLOC_WMARK_LOW
|ALLOC_CPUSET
,
2250 preferred_zone
, migratetype
);
2251 if (unlikely(!page
))
2252 page
= __alloc_pages_slowpath(gfp_mask
, order
,
2253 zonelist
, high_zoneidx
, nodemask
,
2254 preferred_zone
, migratetype
);
2257 trace_mm_page_alloc(page
, order
, gfp_mask
, migratetype
);
2260 EXPORT_SYMBOL(__alloc_pages_nodemask
);
2263 * Common helper functions.
2265 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
2270 * __get_free_pages() returns a 32-bit address, which cannot represent
2273 VM_BUG_ON((gfp_mask
& __GFP_HIGHMEM
) != 0);
2275 page
= alloc_pages(gfp_mask
, order
);
2278 return (unsigned long) page_address(page
);
2280 EXPORT_SYMBOL(__get_free_pages
);
2282 unsigned long get_zeroed_page(gfp_t gfp_mask
)
2284 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
2286 EXPORT_SYMBOL(get_zeroed_page
);
2288 void __pagevec_free(struct pagevec
*pvec
)
2290 int i
= pagevec_count(pvec
);
2293 trace_mm_pagevec_free(pvec
->pages
[i
], pvec
->cold
);
2294 free_hot_cold_page(pvec
->pages
[i
], pvec
->cold
);
2298 void __free_pages(struct page
*page
, unsigned int order
)
2300 if (put_page_testzero(page
)) {
2302 free_hot_cold_page(page
, 0);
2304 __free_pages_ok(page
, order
);
2308 EXPORT_SYMBOL(__free_pages
);
2310 void free_pages(unsigned long addr
, unsigned int order
)
2313 VM_BUG_ON(!virt_addr_valid((void *)addr
));
2314 __free_pages(virt_to_page((void *)addr
), order
);
2318 EXPORT_SYMBOL(free_pages
);
2320 static void *make_alloc_exact(unsigned long addr
, unsigned order
, size_t size
)
2323 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
2324 unsigned long used
= addr
+ PAGE_ALIGN(size
);
2326 split_page(virt_to_page((void *)addr
), order
);
2327 while (used
< alloc_end
) {
2332 return (void *)addr
;
2336 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2337 * @size: the number of bytes to allocate
2338 * @gfp_mask: GFP flags for the allocation
2340 * This function is similar to alloc_pages(), except that it allocates the
2341 * minimum number of pages to satisfy the request. alloc_pages() can only
2342 * allocate memory in power-of-two pages.
2344 * This function is also limited by MAX_ORDER.
2346 * Memory allocated by this function must be released by free_pages_exact().
2348 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
2350 unsigned int order
= get_order(size
);
2353 addr
= __get_free_pages(gfp_mask
, order
);
2354 return make_alloc_exact(addr
, order
, size
);
2356 EXPORT_SYMBOL(alloc_pages_exact
);
2359 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2361 * @size: the number of bytes to allocate
2362 * @gfp_mask: GFP flags for the allocation
2364 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2366 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2369 void *alloc_pages_exact_nid(int nid
, size_t size
, gfp_t gfp_mask
)
2371 unsigned order
= get_order(size
);
2372 struct page
*p
= alloc_pages_node(nid
, gfp_mask
, order
);
2375 return make_alloc_exact((unsigned long)page_address(p
), order
, size
);
2377 EXPORT_SYMBOL(alloc_pages_exact_nid
);
2380 * free_pages_exact - release memory allocated via alloc_pages_exact()
2381 * @virt: the value returned by alloc_pages_exact.
2382 * @size: size of allocation, same value as passed to alloc_pages_exact().
2384 * Release the memory allocated by a previous call to alloc_pages_exact.
2386 void free_pages_exact(void *virt
, size_t size
)
2388 unsigned long addr
= (unsigned long)virt
;
2389 unsigned long end
= addr
+ PAGE_ALIGN(size
);
2391 while (addr
< end
) {
2396 EXPORT_SYMBOL(free_pages_exact
);
2398 static unsigned int nr_free_zone_pages(int offset
)
2403 /* Just pick one node, since fallback list is circular */
2404 unsigned int sum
= 0;
2406 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
2408 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
2409 unsigned long size
= zone
->present_pages
;
2410 unsigned long high
= high_wmark_pages(zone
);
2419 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2421 unsigned int nr_free_buffer_pages(void)
2423 return nr_free_zone_pages(gfp_zone(GFP_USER
));
2425 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
2428 * Amount of free RAM allocatable within all zones
2430 unsigned int nr_free_pagecache_pages(void)
2432 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
2435 static inline void show_node(struct zone
*zone
)
2438 printk("Node %d ", zone_to_nid(zone
));
2441 void si_meminfo(struct sysinfo
*val
)
2443 val
->totalram
= totalram_pages
;
2445 val
->freeram
= global_page_state(NR_FREE_PAGES
);
2446 val
->bufferram
= nr_blockdev_pages();
2447 val
->totalhigh
= totalhigh_pages
;
2448 val
->freehigh
= nr_free_highpages();
2449 val
->mem_unit
= PAGE_SIZE
;
2452 EXPORT_SYMBOL(si_meminfo
);
2455 void si_meminfo_node(struct sysinfo
*val
, int nid
)
2457 pg_data_t
*pgdat
= NODE_DATA(nid
);
2459 val
->totalram
= pgdat
->node_present_pages
;
2460 val
->freeram
= node_page_state(nid
, NR_FREE_PAGES
);
2461 #ifdef CONFIG_HIGHMEM
2462 val
->totalhigh
= pgdat
->node_zones
[ZONE_HIGHMEM
].present_pages
;
2463 val
->freehigh
= zone_page_state(&pgdat
->node_zones
[ZONE_HIGHMEM
],
2469 val
->mem_unit
= PAGE_SIZE
;
2474 * Determine whether the zone's node should be displayed or not, depending on
2475 * whether SHOW_MEM_FILTER_NODES was passed to __show_free_areas().
2477 static bool skip_free_areas_zone(unsigned int flags
, const struct zone
*zone
)
2481 if (!(flags
& SHOW_MEM_FILTER_NODES
))
2485 ret
= !node_isset(zone
->zone_pgdat
->node_id
,
2486 cpuset_current_mems_allowed
);
2492 #define K(x) ((x) << (PAGE_SHIFT-10))
2495 * Show free area list (used inside shift_scroll-lock stuff)
2496 * We also calculate the percentage fragmentation. We do this by counting the
2497 * memory on each free list with the exception of the first item on the list.
2498 * Suppresses nodes that are not allowed by current's cpuset if
2499 * SHOW_MEM_FILTER_NODES is passed.
2501 void __show_free_areas(unsigned int filter
)
2506 for_each_populated_zone(zone
) {
2507 if (skip_free_areas_zone(filter
, zone
))
2510 printk("%s per-cpu:\n", zone
->name
);
2512 for_each_online_cpu(cpu
) {
2513 struct per_cpu_pageset
*pageset
;
2515 pageset
= per_cpu_ptr(zone
->pageset
, cpu
);
2517 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2518 cpu
, pageset
->pcp
.high
,
2519 pageset
->pcp
.batch
, pageset
->pcp
.count
);
2523 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2524 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2526 " dirty:%lu writeback:%lu unstable:%lu\n"
2527 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2528 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n",
2529 global_page_state(NR_ACTIVE_ANON
),
2530 global_page_state(NR_INACTIVE_ANON
),
2531 global_page_state(NR_ISOLATED_ANON
),
2532 global_page_state(NR_ACTIVE_FILE
),
2533 global_page_state(NR_INACTIVE_FILE
),
2534 global_page_state(NR_ISOLATED_FILE
),
2535 global_page_state(NR_UNEVICTABLE
),
2536 global_page_state(NR_FILE_DIRTY
),
2537 global_page_state(NR_WRITEBACK
),
2538 global_page_state(NR_UNSTABLE_NFS
),
2539 global_page_state(NR_FREE_PAGES
),
2540 global_page_state(NR_SLAB_RECLAIMABLE
),
2541 global_page_state(NR_SLAB_UNRECLAIMABLE
),
2542 global_page_state(NR_FILE_MAPPED
),
2543 global_page_state(NR_SHMEM
),
2544 global_page_state(NR_PAGETABLE
),
2545 global_page_state(NR_BOUNCE
));
2547 for_each_populated_zone(zone
) {
2550 if (skip_free_areas_zone(filter
, zone
))
2558 " active_anon:%lukB"
2559 " inactive_anon:%lukB"
2560 " active_file:%lukB"
2561 " inactive_file:%lukB"
2562 " unevictable:%lukB"
2563 " isolated(anon):%lukB"
2564 " isolated(file):%lukB"
2571 " slab_reclaimable:%lukB"
2572 " slab_unreclaimable:%lukB"
2573 " kernel_stack:%lukB"
2577 " writeback_tmp:%lukB"
2578 " pages_scanned:%lu"
2579 " all_unreclaimable? %s"
2582 K(zone_page_state(zone
, NR_FREE_PAGES
)),
2583 K(min_wmark_pages(zone
)),
2584 K(low_wmark_pages(zone
)),
2585 K(high_wmark_pages(zone
)),
2586 K(zone_page_state(zone
, NR_ACTIVE_ANON
)),
2587 K(zone_page_state(zone
, NR_INACTIVE_ANON
)),
2588 K(zone_page_state(zone
, NR_ACTIVE_FILE
)),
2589 K(zone_page_state(zone
, NR_INACTIVE_FILE
)),
2590 K(zone_page_state(zone
, NR_UNEVICTABLE
)),
2591 K(zone_page_state(zone
, NR_ISOLATED_ANON
)),
2592 K(zone_page_state(zone
, NR_ISOLATED_FILE
)),
2593 K(zone
->present_pages
),
2594 K(zone_page_state(zone
, NR_MLOCK
)),
2595 K(zone_page_state(zone
, NR_FILE_DIRTY
)),
2596 K(zone_page_state(zone
, NR_WRITEBACK
)),
2597 K(zone_page_state(zone
, NR_FILE_MAPPED
)),
2598 K(zone_page_state(zone
, NR_SHMEM
)),
2599 K(zone_page_state(zone
, NR_SLAB_RECLAIMABLE
)),
2600 K(zone_page_state(zone
, NR_SLAB_UNRECLAIMABLE
)),
2601 zone_page_state(zone
, NR_KERNEL_STACK
) *
2603 K(zone_page_state(zone
, NR_PAGETABLE
)),
2604 K(zone_page_state(zone
, NR_UNSTABLE_NFS
)),
2605 K(zone_page_state(zone
, NR_BOUNCE
)),
2606 K(zone_page_state(zone
, NR_WRITEBACK_TEMP
)),
2607 zone
->pages_scanned
,
2608 (zone
->all_unreclaimable
? "yes" : "no")
2610 printk("lowmem_reserve[]:");
2611 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
2612 printk(" %lu", zone
->lowmem_reserve
[i
]);
2616 for_each_populated_zone(zone
) {
2617 unsigned long nr
[MAX_ORDER
], flags
, order
, total
= 0;
2619 if (skip_free_areas_zone(filter
, zone
))
2622 printk("%s: ", zone
->name
);
2624 spin_lock_irqsave(&zone
->lock
, flags
);
2625 for (order
= 0; order
< MAX_ORDER
; order
++) {
2626 nr
[order
] = zone
->free_area
[order
].nr_free
;
2627 total
+= nr
[order
] << order
;
2629 spin_unlock_irqrestore(&zone
->lock
, flags
);
2630 for (order
= 0; order
< MAX_ORDER
; order
++)
2631 printk("%lu*%lukB ", nr
[order
], K(1UL) << order
);
2632 printk("= %lukB\n", K(total
));
2635 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES
));
2637 show_swap_cache_info();
2640 void show_free_areas(void)
2642 __show_free_areas(0);
2645 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
2647 zoneref
->zone
= zone
;
2648 zoneref
->zone_idx
= zone_idx(zone
);
2652 * Builds allocation fallback zone lists.
2654 * Add all populated zones of a node to the zonelist.
2656 static int build_zonelists_node(pg_data_t
*pgdat
, struct zonelist
*zonelist
,
2657 int nr_zones
, enum zone_type zone_type
)
2661 BUG_ON(zone_type
>= MAX_NR_ZONES
);
2666 zone
= pgdat
->node_zones
+ zone_type
;
2667 if (populated_zone(zone
)) {
2668 zoneref_set_zone(zone
,
2669 &zonelist
->_zonerefs
[nr_zones
++]);
2670 check_highest_zone(zone_type
);
2673 } while (zone_type
);
2680 * 0 = automatic detection of better ordering.
2681 * 1 = order by ([node] distance, -zonetype)
2682 * 2 = order by (-zonetype, [node] distance)
2684 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
2685 * the same zonelist. So only NUMA can configure this param.
2687 #define ZONELIST_ORDER_DEFAULT 0
2688 #define ZONELIST_ORDER_NODE 1
2689 #define ZONELIST_ORDER_ZONE 2
2691 /* zonelist order in the kernel.
2692 * set_zonelist_order() will set this to NODE or ZONE.
2694 static int current_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
2695 static char zonelist_order_name
[3][8] = {"Default", "Node", "Zone"};
2699 /* The value user specified ....changed by config */
2700 static int user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
2701 /* string for sysctl */
2702 #define NUMA_ZONELIST_ORDER_LEN 16
2703 char numa_zonelist_order
[16] = "default";
2706 * interface for configure zonelist ordering.
2707 * command line option "numa_zonelist_order"
2708 * = "[dD]efault - default, automatic configuration.
2709 * = "[nN]ode - order by node locality, then by zone within node
2710 * = "[zZ]one - order by zone, then by locality within zone
2713 static int __parse_numa_zonelist_order(char *s
)
2715 if (*s
== 'd' || *s
== 'D') {
2716 user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
2717 } else if (*s
== 'n' || *s
== 'N') {
2718 user_zonelist_order
= ZONELIST_ORDER_NODE
;
2719 } else if (*s
== 'z' || *s
== 'Z') {
2720 user_zonelist_order
= ZONELIST_ORDER_ZONE
;
2723 "Ignoring invalid numa_zonelist_order value: "
2730 static __init
int setup_numa_zonelist_order(char *s
)
2737 ret
= __parse_numa_zonelist_order(s
);
2739 strlcpy(numa_zonelist_order
, s
, NUMA_ZONELIST_ORDER_LEN
);
2743 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
2746 * sysctl handler for numa_zonelist_order
2748 int numa_zonelist_order_handler(ctl_table
*table
, int write
,
2749 void __user
*buffer
, size_t *length
,
2752 char saved_string
[NUMA_ZONELIST_ORDER_LEN
];
2754 static DEFINE_MUTEX(zl_order_mutex
);
2756 mutex_lock(&zl_order_mutex
);
2758 strcpy(saved_string
, (char*)table
->data
);
2759 ret
= proc_dostring(table
, write
, buffer
, length
, ppos
);
2763 int oldval
= user_zonelist_order
;
2764 if (__parse_numa_zonelist_order((char*)table
->data
)) {
2766 * bogus value. restore saved string
2768 strncpy((char*)table
->data
, saved_string
,
2769 NUMA_ZONELIST_ORDER_LEN
);
2770 user_zonelist_order
= oldval
;
2771 } else if (oldval
!= user_zonelist_order
) {
2772 mutex_lock(&zonelists_mutex
);
2773 build_all_zonelists(NULL
);
2774 mutex_unlock(&zonelists_mutex
);
2778 mutex_unlock(&zl_order_mutex
);
2783 #define MAX_NODE_LOAD (nr_online_nodes)
2784 static int node_load
[MAX_NUMNODES
];
2787 * find_next_best_node - find the next node that should appear in a given node's fallback list
2788 * @node: node whose fallback list we're appending
2789 * @used_node_mask: nodemask_t of already used nodes
2791 * We use a number of factors to determine which is the next node that should
2792 * appear on a given node's fallback list. The node should not have appeared
2793 * already in @node's fallback list, and it should be the next closest node
2794 * according to the distance array (which contains arbitrary distance values
2795 * from each node to each node in the system), and should also prefer nodes
2796 * with no CPUs, since presumably they'll have very little allocation pressure
2797 * on them otherwise.
2798 * It returns -1 if no node is found.
2800 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
2803 int min_val
= INT_MAX
;
2805 const struct cpumask
*tmp
= cpumask_of_node(0);
2807 /* Use the local node if we haven't already */
2808 if (!node_isset(node
, *used_node_mask
)) {
2809 node_set(node
, *used_node_mask
);
2813 for_each_node_state(n
, N_HIGH_MEMORY
) {
2815 /* Don't want a node to appear more than once */
2816 if (node_isset(n
, *used_node_mask
))
2819 /* Use the distance array to find the distance */
2820 val
= node_distance(node
, n
);
2822 /* Penalize nodes under us ("prefer the next node") */
2825 /* Give preference to headless and unused nodes */
2826 tmp
= cpumask_of_node(n
);
2827 if (!cpumask_empty(tmp
))
2828 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
2830 /* Slight preference for less loaded node */
2831 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
2832 val
+= node_load
[n
];
2834 if (val
< min_val
) {
2841 node_set(best_node
, *used_node_mask
);
2848 * Build zonelists ordered by node and zones within node.
2849 * This results in maximum locality--normal zone overflows into local
2850 * DMA zone, if any--but risks exhausting DMA zone.
2852 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int node
)
2855 struct zonelist
*zonelist
;
2857 zonelist
= &pgdat
->node_zonelists
[0];
2858 for (j
= 0; zonelist
->_zonerefs
[j
].zone
!= NULL
; j
++)
2860 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
2862 zonelist
->_zonerefs
[j
].zone
= NULL
;
2863 zonelist
->_zonerefs
[j
].zone_idx
= 0;
2867 * Build gfp_thisnode zonelists
2869 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
2872 struct zonelist
*zonelist
;
2874 zonelist
= &pgdat
->node_zonelists
[1];
2875 j
= build_zonelists_node(pgdat
, zonelist
, 0, MAX_NR_ZONES
- 1);
2876 zonelist
->_zonerefs
[j
].zone
= NULL
;
2877 zonelist
->_zonerefs
[j
].zone_idx
= 0;
2881 * Build zonelists ordered by zone and nodes within zones.
2882 * This results in conserving DMA zone[s] until all Normal memory is
2883 * exhausted, but results in overflowing to remote node while memory
2884 * may still exist in local DMA zone.
2886 static int node_order
[MAX_NUMNODES
];
2888 static void build_zonelists_in_zone_order(pg_data_t
*pgdat
, int nr_nodes
)
2891 int zone_type
; /* needs to be signed */
2893 struct zonelist
*zonelist
;
2895 zonelist
= &pgdat
->node_zonelists
[0];
2897 for (zone_type
= MAX_NR_ZONES
- 1; zone_type
>= 0; zone_type
--) {
2898 for (j
= 0; j
< nr_nodes
; j
++) {
2899 node
= node_order
[j
];
2900 z
= &NODE_DATA(node
)->node_zones
[zone_type
];
2901 if (populated_zone(z
)) {
2903 &zonelist
->_zonerefs
[pos
++]);
2904 check_highest_zone(zone_type
);
2908 zonelist
->_zonerefs
[pos
].zone
= NULL
;
2909 zonelist
->_zonerefs
[pos
].zone_idx
= 0;
2912 static int default_zonelist_order(void)
2915 unsigned long low_kmem_size
,total_size
;
2919 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
2920 * If they are really small and used heavily, the system can fall
2921 * into OOM very easily.
2922 * This function detect ZONE_DMA/DMA32 size and configures zone order.
2924 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
2927 for_each_online_node(nid
) {
2928 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
2929 z
= &NODE_DATA(nid
)->node_zones
[zone_type
];
2930 if (populated_zone(z
)) {
2931 if (zone_type
< ZONE_NORMAL
)
2932 low_kmem_size
+= z
->present_pages
;
2933 total_size
+= z
->present_pages
;
2934 } else if (zone_type
== ZONE_NORMAL
) {
2936 * If any node has only lowmem, then node order
2937 * is preferred to allow kernel allocations
2938 * locally; otherwise, they can easily infringe
2939 * on other nodes when there is an abundance of
2940 * lowmem available to allocate from.
2942 return ZONELIST_ORDER_NODE
;
2946 if (!low_kmem_size
|| /* there are no DMA area. */
2947 low_kmem_size
> total_size
/2) /* DMA/DMA32 is big. */
2948 return ZONELIST_ORDER_NODE
;
2950 * look into each node's config.
2951 * If there is a node whose DMA/DMA32 memory is very big area on
2952 * local memory, NODE_ORDER may be suitable.
2954 average_size
= total_size
/
2955 (nodes_weight(node_states
[N_HIGH_MEMORY
]) + 1);
2956 for_each_online_node(nid
) {
2959 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
2960 z
= &NODE_DATA(nid
)->node_zones
[zone_type
];
2961 if (populated_zone(z
)) {
2962 if (zone_type
< ZONE_NORMAL
)
2963 low_kmem_size
+= z
->present_pages
;
2964 total_size
+= z
->present_pages
;
2967 if (low_kmem_size
&&
2968 total_size
> average_size
&& /* ignore small node */
2969 low_kmem_size
> total_size
* 70/100)
2970 return ZONELIST_ORDER_NODE
;
2972 return ZONELIST_ORDER_ZONE
;
2975 static void set_zonelist_order(void)
2977 if (user_zonelist_order
== ZONELIST_ORDER_DEFAULT
)
2978 current_zonelist_order
= default_zonelist_order();
2980 current_zonelist_order
= user_zonelist_order
;
2983 static void build_zonelists(pg_data_t
*pgdat
)
2987 nodemask_t used_mask
;
2988 int local_node
, prev_node
;
2989 struct zonelist
*zonelist
;
2990 int order
= current_zonelist_order
;
2992 /* initialize zonelists */
2993 for (i
= 0; i
< MAX_ZONELISTS
; i
++) {
2994 zonelist
= pgdat
->node_zonelists
+ i
;
2995 zonelist
->_zonerefs
[0].zone
= NULL
;
2996 zonelist
->_zonerefs
[0].zone_idx
= 0;
2999 /* NUMA-aware ordering of nodes */
3000 local_node
= pgdat
->node_id
;
3001 load
= nr_online_nodes
;
3002 prev_node
= local_node
;
3003 nodes_clear(used_mask
);
3005 memset(node_order
, 0, sizeof(node_order
));
3008 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
3009 int distance
= node_distance(local_node
, node
);
3012 * If another node is sufficiently far away then it is better
3013 * to reclaim pages in a zone before going off node.
3015 if (distance
> RECLAIM_DISTANCE
)
3016 zone_reclaim_mode
= 1;
3019 * We don't want to pressure a particular node.
3020 * So adding penalty to the first node in same
3021 * distance group to make it round-robin.
3023 if (distance
!= node_distance(local_node
, prev_node
))
3024 node_load
[node
] = load
;
3028 if (order
== ZONELIST_ORDER_NODE
)
3029 build_zonelists_in_node_order(pgdat
, node
);
3031 node_order
[j
++] = node
; /* remember order */
3034 if (order
== ZONELIST_ORDER_ZONE
) {
3035 /* calculate node order -- i.e., DMA last! */
3036 build_zonelists_in_zone_order(pgdat
, j
);
3039 build_thisnode_zonelists(pgdat
);
3042 /* Construct the zonelist performance cache - see further mmzone.h */
3043 static void build_zonelist_cache(pg_data_t
*pgdat
)
3045 struct zonelist
*zonelist
;
3046 struct zonelist_cache
*zlc
;
3049 zonelist
= &pgdat
->node_zonelists
[0];
3050 zonelist
->zlcache_ptr
= zlc
= &zonelist
->zlcache
;
3051 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
3052 for (z
= zonelist
->_zonerefs
; z
->zone
; z
++)
3053 zlc
->z_to_n
[z
- zonelist
->_zonerefs
] = zonelist_node_idx(z
);
3056 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3058 * Return node id of node used for "local" allocations.
3059 * I.e., first node id of first zone in arg node's generic zonelist.
3060 * Used for initializing percpu 'numa_mem', which is used primarily
3061 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3063 int local_memory_node(int node
)
3067 (void)first_zones_zonelist(node_zonelist(node
, GFP_KERNEL
),
3068 gfp_zone(GFP_KERNEL
),
3075 #else /* CONFIG_NUMA */
3077 static void set_zonelist_order(void)
3079 current_zonelist_order
= ZONELIST_ORDER_ZONE
;
3082 static void build_zonelists(pg_data_t
*pgdat
)
3084 int node
, local_node
;
3086 struct zonelist
*zonelist
;
3088 local_node
= pgdat
->node_id
;
3090 zonelist
= &pgdat
->node_zonelists
[0];
3091 j
= build_zonelists_node(pgdat
, zonelist
, 0, MAX_NR_ZONES
- 1);
3094 * Now we build the zonelist so that it contains the zones
3095 * of all the other nodes.
3096 * We don't want to pressure a particular node, so when
3097 * building the zones for node N, we make sure that the
3098 * zones coming right after the local ones are those from
3099 * node N+1 (modulo N)
3101 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
3102 if (!node_online(node
))
3104 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
3107 for (node
= 0; node
< local_node
; node
++) {
3108 if (!node_online(node
))
3110 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
3114 zonelist
->_zonerefs
[j
].zone
= NULL
;
3115 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3118 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3119 static void build_zonelist_cache(pg_data_t
*pgdat
)
3121 pgdat
->node_zonelists
[0].zlcache_ptr
= NULL
;
3124 #endif /* CONFIG_NUMA */
3127 * Boot pageset table. One per cpu which is going to be used for all
3128 * zones and all nodes. The parameters will be set in such a way
3129 * that an item put on a list will immediately be handed over to
3130 * the buddy list. This is safe since pageset manipulation is done
3131 * with interrupts disabled.
3133 * The boot_pagesets must be kept even after bootup is complete for
3134 * unused processors and/or zones. They do play a role for bootstrapping
3135 * hotplugged processors.
3137 * zoneinfo_show() and maybe other functions do
3138 * not check if the processor is online before following the pageset pointer.
3139 * Other parts of the kernel may not check if the zone is available.
3141 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
);
3142 static DEFINE_PER_CPU(struct per_cpu_pageset
, boot_pageset
);
3143 static void setup_zone_pageset(struct zone
*zone
);
3146 * Global mutex to protect against size modification of zonelists
3147 * as well as to serialize pageset setup for the new populated zone.
3149 DEFINE_MUTEX(zonelists_mutex
);
3151 /* return values int ....just for stop_machine() */
3152 static __init_refok
int __build_all_zonelists(void *data
)
3158 memset(node_load
, 0, sizeof(node_load
));
3160 for_each_online_node(nid
) {
3161 pg_data_t
*pgdat
= NODE_DATA(nid
);
3163 build_zonelists(pgdat
);
3164 build_zonelist_cache(pgdat
);
3168 * Initialize the boot_pagesets that are going to be used
3169 * for bootstrapping processors. The real pagesets for
3170 * each zone will be allocated later when the per cpu
3171 * allocator is available.
3173 * boot_pagesets are used also for bootstrapping offline
3174 * cpus if the system is already booted because the pagesets
3175 * are needed to initialize allocators on a specific cpu too.
3176 * F.e. the percpu allocator needs the page allocator which
3177 * needs the percpu allocator in order to allocate its pagesets
3178 * (a chicken-egg dilemma).
3180 for_each_possible_cpu(cpu
) {
3181 setup_pageset(&per_cpu(boot_pageset
, cpu
), 0);
3183 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3185 * We now know the "local memory node" for each node--
3186 * i.e., the node of the first zone in the generic zonelist.
3187 * Set up numa_mem percpu variable for on-line cpus. During
3188 * boot, only the boot cpu should be on-line; we'll init the
3189 * secondary cpus' numa_mem as they come on-line. During
3190 * node/memory hotplug, we'll fixup all on-line cpus.
3192 if (cpu_online(cpu
))
3193 set_cpu_numa_mem(cpu
, local_memory_node(cpu_to_node(cpu
)));
3201 * Called with zonelists_mutex held always
3202 * unless system_state == SYSTEM_BOOTING.
3204 void __ref
build_all_zonelists(void *data
)
3206 set_zonelist_order();
3208 if (system_state
== SYSTEM_BOOTING
) {
3209 __build_all_zonelists(NULL
);
3210 mminit_verify_zonelist();
3211 cpuset_init_current_mems_allowed();
3213 /* we have to stop all cpus to guarantee there is no user
3215 #ifdef CONFIG_MEMORY_HOTPLUG
3217 setup_zone_pageset((struct zone
*)data
);
3219 stop_machine(__build_all_zonelists
, NULL
, NULL
);
3220 /* cpuset refresh routine should be here */
3222 vm_total_pages
= nr_free_pagecache_pages();
3224 * Disable grouping by mobility if the number of pages in the
3225 * system is too low to allow the mechanism to work. It would be
3226 * more accurate, but expensive to check per-zone. This check is
3227 * made on memory-hotadd so a system can start with mobility
3228 * disabled and enable it later
3230 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
3231 page_group_by_mobility_disabled
= 1;
3233 page_group_by_mobility_disabled
= 0;
3235 printk("Built %i zonelists in %s order, mobility grouping %s. "
3236 "Total pages: %ld\n",
3238 zonelist_order_name
[current_zonelist_order
],
3239 page_group_by_mobility_disabled
? "off" : "on",
3242 printk("Policy zone: %s\n", zone_names
[policy_zone
]);
3247 * Helper functions to size the waitqueue hash table.
3248 * Essentially these want to choose hash table sizes sufficiently
3249 * large so that collisions trying to wait on pages are rare.
3250 * But in fact, the number of active page waitqueues on typical
3251 * systems is ridiculously low, less than 200. So this is even
3252 * conservative, even though it seems large.
3254 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3255 * waitqueues, i.e. the size of the waitq table given the number of pages.
3257 #define PAGES_PER_WAITQUEUE 256
3259 #ifndef CONFIG_MEMORY_HOTPLUG
3260 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
3262 unsigned long size
= 1;
3264 pages
/= PAGES_PER_WAITQUEUE
;
3266 while (size
< pages
)
3270 * Once we have dozens or even hundreds of threads sleeping
3271 * on IO we've got bigger problems than wait queue collision.
3272 * Limit the size of the wait table to a reasonable size.
3274 size
= min(size
, 4096UL);
3276 return max(size
, 4UL);
3280 * A zone's size might be changed by hot-add, so it is not possible to determine
3281 * a suitable size for its wait_table. So we use the maximum size now.
3283 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3285 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3286 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3287 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3289 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3290 * or more by the traditional way. (See above). It equals:
3292 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3293 * ia64(16K page size) : = ( 8G + 4M)byte.
3294 * powerpc (64K page size) : = (32G +16M)byte.
3296 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
3303 * This is an integer logarithm so that shifts can be used later
3304 * to extract the more random high bits from the multiplicative
3305 * hash function before the remainder is taken.
3307 static inline unsigned long wait_table_bits(unsigned long size
)
3312 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3315 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3316 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3317 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3318 * higher will lead to a bigger reserve which will get freed as contiguous
3319 * blocks as reclaim kicks in
3321 static void setup_zone_migrate_reserve(struct zone
*zone
)
3323 unsigned long start_pfn
, pfn
, end_pfn
;
3325 unsigned long block_migratetype
;
3328 /* Get the start pfn, end pfn and the number of blocks to reserve */
3329 start_pfn
= zone
->zone_start_pfn
;
3330 end_pfn
= start_pfn
+ zone
->spanned_pages
;
3331 reserve
= roundup(min_wmark_pages(zone
), pageblock_nr_pages
) >>
3335 * Reserve blocks are generally in place to help high-order atomic
3336 * allocations that are short-lived. A min_free_kbytes value that
3337 * would result in more than 2 reserve blocks for atomic allocations
3338 * is assumed to be in place to help anti-fragmentation for the
3339 * future allocation of hugepages at runtime.
3341 reserve
= min(2, reserve
);
3343 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= pageblock_nr_pages
) {
3344 if (!pfn_valid(pfn
))
3346 page
= pfn_to_page(pfn
);
3348 /* Watch out for overlapping nodes */
3349 if (page_to_nid(page
) != zone_to_nid(zone
))
3352 /* Blocks with reserved pages will never free, skip them. */
3353 if (PageReserved(page
))
3356 block_migratetype
= get_pageblock_migratetype(page
);
3358 /* If this block is reserved, account for it */
3359 if (reserve
> 0 && block_migratetype
== MIGRATE_RESERVE
) {
3364 /* Suitable for reserving if this block is movable */
3365 if (reserve
> 0 && block_migratetype
== MIGRATE_MOVABLE
) {
3366 set_pageblock_migratetype(page
, MIGRATE_RESERVE
);
3367 move_freepages_block(zone
, page
, MIGRATE_RESERVE
);
3373 * If the reserve is met and this is a previous reserved block,
3376 if (block_migratetype
== MIGRATE_RESERVE
) {
3377 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
3378 move_freepages_block(zone
, page
, MIGRATE_MOVABLE
);
3384 * Initially all pages are reserved - free ones are freed
3385 * up by free_all_bootmem() once the early boot process is
3386 * done. Non-atomic initialization, single-pass.
3388 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
3389 unsigned long start_pfn
, enum memmap_context context
)
3392 unsigned long end_pfn
= start_pfn
+ size
;
3396 if (highest_memmap_pfn
< end_pfn
- 1)
3397 highest_memmap_pfn
= end_pfn
- 1;
3399 z
= &NODE_DATA(nid
)->node_zones
[zone
];
3400 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
3402 * There can be holes in boot-time mem_map[]s
3403 * handed to this function. They do not
3404 * exist on hotplugged memory.
3406 if (context
== MEMMAP_EARLY
) {
3407 if (!early_pfn_valid(pfn
))
3409 if (!early_pfn_in_nid(pfn
, nid
))
3412 page
= pfn_to_page(pfn
);
3413 set_page_links(page
, zone
, nid
, pfn
);
3414 mminit_verify_page_links(page
, zone
, nid
, pfn
);
3415 init_page_count(page
);
3416 reset_page_mapcount(page
);
3417 SetPageReserved(page
);
3419 * Mark the block movable so that blocks are reserved for
3420 * movable at startup. This will force kernel allocations
3421 * to reserve their blocks rather than leaking throughout
3422 * the address space during boot when many long-lived
3423 * kernel allocations are made. Later some blocks near
3424 * the start are marked MIGRATE_RESERVE by
3425 * setup_zone_migrate_reserve()
3427 * bitmap is created for zone's valid pfn range. but memmap
3428 * can be created for invalid pages (for alignment)
3429 * check here not to call set_pageblock_migratetype() against
3432 if ((z
->zone_start_pfn
<= pfn
)
3433 && (pfn
< z
->zone_start_pfn
+ z
->spanned_pages
)
3434 && !(pfn
& (pageblock_nr_pages
- 1)))
3435 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
3437 INIT_LIST_HEAD(&page
->lru
);
3438 #ifdef WANT_PAGE_VIRTUAL
3439 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3440 if (!is_highmem_idx(zone
))
3441 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
3446 static void __meminit
zone_init_free_lists(struct zone
*zone
)
3449 for_each_migratetype_order(order
, t
) {
3450 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
3451 zone
->free_area
[order
].nr_free
= 0;
3455 #ifndef __HAVE_ARCH_MEMMAP_INIT
3456 #define memmap_init(size, nid, zone, start_pfn) \
3457 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3460 static int zone_batchsize(struct zone
*zone
)
3466 * The per-cpu-pages pools are set to around 1000th of the
3467 * size of the zone. But no more than 1/2 of a meg.
3469 * OK, so we don't know how big the cache is. So guess.
3471 batch
= zone
->present_pages
/ 1024;
3472 if (batch
* PAGE_SIZE
> 512 * 1024)
3473 batch
= (512 * 1024) / PAGE_SIZE
;
3474 batch
/= 4; /* We effectively *= 4 below */
3479 * Clamp the batch to a 2^n - 1 value. Having a power
3480 * of 2 value was found to be more likely to have
3481 * suboptimal cache aliasing properties in some cases.
3483 * For example if 2 tasks are alternately allocating
3484 * batches of pages, one task can end up with a lot
3485 * of pages of one half of the possible page colors
3486 * and the other with pages of the other colors.
3488 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
3493 /* The deferral and batching of frees should be suppressed under NOMMU
3496 * The problem is that NOMMU needs to be able to allocate large chunks
3497 * of contiguous memory as there's no hardware page translation to
3498 * assemble apparent contiguous memory from discontiguous pages.
3500 * Queueing large contiguous runs of pages for batching, however,
3501 * causes the pages to actually be freed in smaller chunks. As there
3502 * can be a significant delay between the individual batches being
3503 * recycled, this leads to the once large chunks of space being
3504 * fragmented and becoming unavailable for high-order allocations.
3510 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
3512 struct per_cpu_pages
*pcp
;
3515 memset(p
, 0, sizeof(*p
));
3519 pcp
->high
= 6 * batch
;
3520 pcp
->batch
= max(1UL, 1 * batch
);
3521 for (migratetype
= 0; migratetype
< MIGRATE_PCPTYPES
; migratetype
++)
3522 INIT_LIST_HEAD(&pcp
->lists
[migratetype
]);
3526 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3527 * to the value high for the pageset p.
3530 static void setup_pagelist_highmark(struct per_cpu_pageset
*p
,
3533 struct per_cpu_pages
*pcp
;
3537 pcp
->batch
= max(1UL, high
/4);
3538 if ((high
/4) > (PAGE_SHIFT
* 8))
3539 pcp
->batch
= PAGE_SHIFT
* 8;
3542 static __meminit
void setup_zone_pageset(struct zone
*zone
)
3546 zone
->pageset
= alloc_percpu(struct per_cpu_pageset
);
3548 for_each_possible_cpu(cpu
) {
3549 struct per_cpu_pageset
*pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
3551 setup_pageset(pcp
, zone_batchsize(zone
));
3553 if (percpu_pagelist_fraction
)
3554 setup_pagelist_highmark(pcp
,
3555 (zone
->present_pages
/
3556 percpu_pagelist_fraction
));
3561 * Allocate per cpu pagesets and initialize them.
3562 * Before this call only boot pagesets were available.
3564 void __init
setup_per_cpu_pageset(void)
3568 for_each_populated_zone(zone
)
3569 setup_zone_pageset(zone
);
3572 static noinline __init_refok
3573 int zone_wait_table_init(struct zone
*zone
, unsigned long zone_size_pages
)
3576 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
3580 * The per-page waitqueue mechanism uses hashed waitqueues
3583 zone
->wait_table_hash_nr_entries
=
3584 wait_table_hash_nr_entries(zone_size_pages
);
3585 zone
->wait_table_bits
=
3586 wait_table_bits(zone
->wait_table_hash_nr_entries
);
3587 alloc_size
= zone
->wait_table_hash_nr_entries
3588 * sizeof(wait_queue_head_t
);
3590 if (!slab_is_available()) {
3591 zone
->wait_table
= (wait_queue_head_t
*)
3592 alloc_bootmem_node_nopanic(pgdat
, alloc_size
);
3595 * This case means that a zone whose size was 0 gets new memory
3596 * via memory hot-add.
3597 * But it may be the case that a new node was hot-added. In
3598 * this case vmalloc() will not be able to use this new node's
3599 * memory - this wait_table must be initialized to use this new
3600 * node itself as well.
3601 * To use this new node's memory, further consideration will be
3604 zone
->wait_table
= vmalloc(alloc_size
);
3606 if (!zone
->wait_table
)
3609 for(i
= 0; i
< zone
->wait_table_hash_nr_entries
; ++i
)
3610 init_waitqueue_head(zone
->wait_table
+ i
);
3615 static int __zone_pcp_update(void *data
)
3617 struct zone
*zone
= data
;
3619 unsigned long batch
= zone_batchsize(zone
), flags
;
3621 for_each_possible_cpu(cpu
) {
3622 struct per_cpu_pageset
*pset
;
3623 struct per_cpu_pages
*pcp
;
3625 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
3628 local_irq_save(flags
);
3629 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
3630 setup_pageset(pset
, batch
);
3631 local_irq_restore(flags
);
3636 void zone_pcp_update(struct zone
*zone
)
3638 stop_machine(__zone_pcp_update
, zone
, NULL
);
3641 static __meminit
void zone_pcp_init(struct zone
*zone
)
3644 * per cpu subsystem is not up at this point. The following code
3645 * relies on the ability of the linker to provide the
3646 * offset of a (static) per cpu variable into the per cpu area.
3648 zone
->pageset
= &boot_pageset
;
3650 if (zone
->present_pages
)
3651 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%u\n",
3652 zone
->name
, zone
->present_pages
,
3653 zone_batchsize(zone
));
3656 __meminit
int init_currently_empty_zone(struct zone
*zone
,
3657 unsigned long zone_start_pfn
,
3659 enum memmap_context context
)
3661 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
3663 ret
= zone_wait_table_init(zone
, size
);
3666 pgdat
->nr_zones
= zone_idx(zone
) + 1;
3668 zone
->zone_start_pfn
= zone_start_pfn
;
3670 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
3671 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
3673 (unsigned long)zone_idx(zone
),
3674 zone_start_pfn
, (zone_start_pfn
+ size
));
3676 zone_init_free_lists(zone
);
3681 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3683 * Basic iterator support. Return the first range of PFNs for a node
3684 * Note: nid == MAX_NUMNODES returns first region regardless of node
3686 static int __meminit
first_active_region_index_in_nid(int nid
)
3690 for (i
= 0; i
< nr_nodemap_entries
; i
++)
3691 if (nid
== MAX_NUMNODES
|| early_node_map
[i
].nid
== nid
)
3698 * Basic iterator support. Return the next active range of PFNs for a node
3699 * Note: nid == MAX_NUMNODES returns next region regardless of node
3701 static int __meminit
next_active_region_index_in_nid(int index
, int nid
)
3703 for (index
= index
+ 1; index
< nr_nodemap_entries
; index
++)
3704 if (nid
== MAX_NUMNODES
|| early_node_map
[index
].nid
== nid
)
3710 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
3712 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
3713 * Architectures may implement their own version but if add_active_range()
3714 * was used and there are no special requirements, this is a convenient
3717 int __meminit
__early_pfn_to_nid(unsigned long pfn
)
3721 for (i
= 0; i
< nr_nodemap_entries
; i
++) {
3722 unsigned long start_pfn
= early_node_map
[i
].start_pfn
;
3723 unsigned long end_pfn
= early_node_map
[i
].end_pfn
;
3725 if (start_pfn
<= pfn
&& pfn
< end_pfn
)
3726 return early_node_map
[i
].nid
;
3728 /* This is a memory hole */
3731 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
3733 int __meminit
early_pfn_to_nid(unsigned long pfn
)
3737 nid
= __early_pfn_to_nid(pfn
);
3740 /* just returns 0 */
3744 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
3745 bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
3749 nid
= __early_pfn_to_nid(pfn
);
3750 if (nid
>= 0 && nid
!= node
)
3756 /* Basic iterator support to walk early_node_map[] */
3757 #define for_each_active_range_index_in_nid(i, nid) \
3758 for (i = first_active_region_index_in_nid(nid); i != -1; \
3759 i = next_active_region_index_in_nid(i, nid))
3762 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
3763 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
3764 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
3766 * If an architecture guarantees that all ranges registered with
3767 * add_active_ranges() contain no holes and may be freed, this
3768 * this function may be used instead of calling free_bootmem() manually.
3770 void __init
free_bootmem_with_active_regions(int nid
,
3771 unsigned long max_low_pfn
)
3775 for_each_active_range_index_in_nid(i
, nid
) {
3776 unsigned long size_pages
= 0;
3777 unsigned long end_pfn
= early_node_map
[i
].end_pfn
;
3779 if (early_node_map
[i
].start_pfn
>= max_low_pfn
)
3782 if (end_pfn
> max_low_pfn
)
3783 end_pfn
= max_low_pfn
;
3785 size_pages
= end_pfn
- early_node_map
[i
].start_pfn
;
3786 free_bootmem_node(NODE_DATA(early_node_map
[i
].nid
),
3787 PFN_PHYS(early_node_map
[i
].start_pfn
),
3788 size_pages
<< PAGE_SHIFT
);
3792 #ifdef CONFIG_HAVE_MEMBLOCK
3794 * Basic iterator support. Return the last range of PFNs for a node
3795 * Note: nid == MAX_NUMNODES returns last region regardless of node
3797 static int __meminit
last_active_region_index_in_nid(int nid
)
3801 for (i
= nr_nodemap_entries
- 1; i
>= 0; i
--)
3802 if (nid
== MAX_NUMNODES
|| early_node_map
[i
].nid
== nid
)
3809 * Basic iterator support. Return the previous active range of PFNs for a node
3810 * Note: nid == MAX_NUMNODES returns next region regardless of node
3812 static int __meminit
previous_active_region_index_in_nid(int index
, int nid
)
3814 for (index
= index
- 1; index
>= 0; index
--)
3815 if (nid
== MAX_NUMNODES
|| early_node_map
[index
].nid
== nid
)
3821 #define for_each_active_range_index_in_nid_reverse(i, nid) \
3822 for (i = last_active_region_index_in_nid(nid); i != -1; \
3823 i = previous_active_region_index_in_nid(i, nid))
3825 u64 __init
find_memory_core_early(int nid
, u64 size
, u64 align
,
3826 u64 goal
, u64 limit
)
3830 /* Need to go over early_node_map to find out good range for node */
3831 for_each_active_range_index_in_nid_reverse(i
, nid
) {
3833 u64 ei_start
, ei_last
;
3834 u64 final_start
, final_end
;
3836 ei_last
= early_node_map
[i
].end_pfn
;
3837 ei_last
<<= PAGE_SHIFT
;
3838 ei_start
= early_node_map
[i
].start_pfn
;
3839 ei_start
<<= PAGE_SHIFT
;
3841 final_start
= max(ei_start
, goal
);
3842 final_end
= min(ei_last
, limit
);
3844 if (final_start
>= final_end
)
3847 addr
= memblock_find_in_range(final_start
, final_end
, size
, align
);
3849 if (addr
== MEMBLOCK_ERROR
)
3855 return MEMBLOCK_ERROR
;
3859 int __init
add_from_early_node_map(struct range
*range
, int az
,
3860 int nr_range
, int nid
)
3865 /* need to go over early_node_map to find out good range for node */
3866 for_each_active_range_index_in_nid(i
, nid
) {
3867 start
= early_node_map
[i
].start_pfn
;
3868 end
= early_node_map
[i
].end_pfn
;
3869 nr_range
= add_range(range
, az
, nr_range
, start
, end
);
3874 void __init
work_with_active_regions(int nid
, work_fn_t work_fn
, void *data
)
3879 for_each_active_range_index_in_nid(i
, nid
) {
3880 ret
= work_fn(early_node_map
[i
].start_pfn
,
3881 early_node_map
[i
].end_pfn
, data
);
3887 * sparse_memory_present_with_active_regions - Call memory_present for each active range
3888 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
3890 * If an architecture guarantees that all ranges registered with
3891 * add_active_ranges() contain no holes and may be freed, this
3892 * function may be used instead of calling memory_present() manually.
3894 void __init
sparse_memory_present_with_active_regions(int nid
)
3898 for_each_active_range_index_in_nid(i
, nid
)
3899 memory_present(early_node_map
[i
].nid
,
3900 early_node_map
[i
].start_pfn
,
3901 early_node_map
[i
].end_pfn
);
3905 * get_pfn_range_for_nid - Return the start and end page frames for a node
3906 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
3907 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
3908 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
3910 * It returns the start and end page frame of a node based on information
3911 * provided by an arch calling add_active_range(). If called for a node
3912 * with no available memory, a warning is printed and the start and end
3915 void __meminit
get_pfn_range_for_nid(unsigned int nid
,
3916 unsigned long *start_pfn
, unsigned long *end_pfn
)
3922 for_each_active_range_index_in_nid(i
, nid
) {
3923 *start_pfn
= min(*start_pfn
, early_node_map
[i
].start_pfn
);
3924 *end_pfn
= max(*end_pfn
, early_node_map
[i
].end_pfn
);
3927 if (*start_pfn
== -1UL)
3932 * This finds a zone that can be used for ZONE_MOVABLE pages. The
3933 * assumption is made that zones within a node are ordered in monotonic
3934 * increasing memory addresses so that the "highest" populated zone is used
3936 static void __init
find_usable_zone_for_movable(void)
3939 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
3940 if (zone_index
== ZONE_MOVABLE
)
3943 if (arch_zone_highest_possible_pfn
[zone_index
] >
3944 arch_zone_lowest_possible_pfn
[zone_index
])
3948 VM_BUG_ON(zone_index
== -1);
3949 movable_zone
= zone_index
;
3953 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
3954 * because it is sized independent of architecture. Unlike the other zones,
3955 * the starting point for ZONE_MOVABLE is not fixed. It may be different
3956 * in each node depending on the size of each node and how evenly kernelcore
3957 * is distributed. This helper function adjusts the zone ranges
3958 * provided by the architecture for a given node by using the end of the
3959 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
3960 * zones within a node are in order of monotonic increases memory addresses
3962 static void __meminit
adjust_zone_range_for_zone_movable(int nid
,
3963 unsigned long zone_type
,
3964 unsigned long node_start_pfn
,
3965 unsigned long node_end_pfn
,
3966 unsigned long *zone_start_pfn
,
3967 unsigned long *zone_end_pfn
)
3969 /* Only adjust if ZONE_MOVABLE is on this node */
3970 if (zone_movable_pfn
[nid
]) {
3971 /* Size ZONE_MOVABLE */
3972 if (zone_type
== ZONE_MOVABLE
) {
3973 *zone_start_pfn
= zone_movable_pfn
[nid
];
3974 *zone_end_pfn
= min(node_end_pfn
,
3975 arch_zone_highest_possible_pfn
[movable_zone
]);
3977 /* Adjust for ZONE_MOVABLE starting within this range */
3978 } else if (*zone_start_pfn
< zone_movable_pfn
[nid
] &&
3979 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
3980 *zone_end_pfn
= zone_movable_pfn
[nid
];
3982 /* Check if this whole range is within ZONE_MOVABLE */
3983 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
3984 *zone_start_pfn
= *zone_end_pfn
;
3989 * Return the number of pages a zone spans in a node, including holes
3990 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
3992 static unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
3993 unsigned long zone_type
,
3994 unsigned long *ignored
)
3996 unsigned long node_start_pfn
, node_end_pfn
;
3997 unsigned long zone_start_pfn
, zone_end_pfn
;
3999 /* Get the start and end of the node and zone */
4000 get_pfn_range_for_nid(nid
, &node_start_pfn
, &node_end_pfn
);
4001 zone_start_pfn
= arch_zone_lowest_possible_pfn
[zone_type
];
4002 zone_end_pfn
= arch_zone_highest_possible_pfn
[zone_type
];
4003 adjust_zone_range_for_zone_movable(nid
, zone_type
,
4004 node_start_pfn
, node_end_pfn
,
4005 &zone_start_pfn
, &zone_end_pfn
);
4007 /* Check that this node has pages within the zone's required range */
4008 if (zone_end_pfn
< node_start_pfn
|| zone_start_pfn
> node_end_pfn
)
4011 /* Move the zone boundaries inside the node if necessary */
4012 zone_end_pfn
= min(zone_end_pfn
, node_end_pfn
);
4013 zone_start_pfn
= max(zone_start_pfn
, node_start_pfn
);
4015 /* Return the spanned pages */
4016 return zone_end_pfn
- zone_start_pfn
;
4020 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4021 * then all holes in the requested range will be accounted for.
4023 unsigned long __meminit
__absent_pages_in_range(int nid
,
4024 unsigned long range_start_pfn
,
4025 unsigned long range_end_pfn
)
4028 unsigned long prev_end_pfn
= 0, hole_pages
= 0;
4029 unsigned long start_pfn
;
4031 /* Find the end_pfn of the first active range of pfns in the node */
4032 i
= first_active_region_index_in_nid(nid
);
4036 prev_end_pfn
= min(early_node_map
[i
].start_pfn
, range_end_pfn
);
4038 /* Account for ranges before physical memory on this node */
4039 if (early_node_map
[i
].start_pfn
> range_start_pfn
)
4040 hole_pages
= prev_end_pfn
- range_start_pfn
;
4042 /* Find all holes for the zone within the node */
4043 for (; i
!= -1; i
= next_active_region_index_in_nid(i
, nid
)) {
4045 /* No need to continue if prev_end_pfn is outside the zone */
4046 if (prev_end_pfn
>= range_end_pfn
)
4049 /* Make sure the end of the zone is not within the hole */
4050 start_pfn
= min(early_node_map
[i
].start_pfn
, range_end_pfn
);
4051 prev_end_pfn
= max(prev_end_pfn
, range_start_pfn
);
4053 /* Update the hole size cound and move on */
4054 if (start_pfn
> range_start_pfn
) {
4055 BUG_ON(prev_end_pfn
> start_pfn
);
4056 hole_pages
+= start_pfn
- prev_end_pfn
;
4058 prev_end_pfn
= early_node_map
[i
].end_pfn
;
4061 /* Account for ranges past physical memory on this node */
4062 if (range_end_pfn
> prev_end_pfn
)
4063 hole_pages
+= range_end_pfn
-
4064 max(range_start_pfn
, prev_end_pfn
);
4070 * absent_pages_in_range - Return number of page frames in holes within a range
4071 * @start_pfn: The start PFN to start searching for holes
4072 * @end_pfn: The end PFN to stop searching for holes
4074 * It returns the number of pages frames in memory holes within a range.
4076 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
4077 unsigned long end_pfn
)
4079 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
4082 /* Return the number of page frames in holes in a zone on a node */
4083 static unsigned long __meminit
zone_absent_pages_in_node(int nid
,
4084 unsigned long zone_type
,
4085 unsigned long *ignored
)
4087 unsigned long node_start_pfn
, node_end_pfn
;
4088 unsigned long zone_start_pfn
, zone_end_pfn
;
4090 get_pfn_range_for_nid(nid
, &node_start_pfn
, &node_end_pfn
);
4091 zone_start_pfn
= max(arch_zone_lowest_possible_pfn
[zone_type
],
4093 zone_end_pfn
= min(arch_zone_highest_possible_pfn
[zone_type
],
4096 adjust_zone_range_for_zone_movable(nid
, zone_type
,
4097 node_start_pfn
, node_end_pfn
,
4098 &zone_start_pfn
, &zone_end_pfn
);
4099 return __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
4103 static inline unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
4104 unsigned long zone_type
,
4105 unsigned long *zones_size
)
4107 return zones_size
[zone_type
];
4110 static inline unsigned long __meminit
zone_absent_pages_in_node(int nid
,
4111 unsigned long zone_type
,
4112 unsigned long *zholes_size
)
4117 return zholes_size
[zone_type
];
4122 static void __meminit
calculate_node_totalpages(struct pglist_data
*pgdat
,
4123 unsigned long *zones_size
, unsigned long *zholes_size
)
4125 unsigned long realtotalpages
, totalpages
= 0;
4128 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4129 totalpages
+= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
4131 pgdat
->node_spanned_pages
= totalpages
;
4133 realtotalpages
= totalpages
;
4134 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4136 zone_absent_pages_in_node(pgdat
->node_id
, i
,
4138 pgdat
->node_present_pages
= realtotalpages
;
4139 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
4143 #ifndef CONFIG_SPARSEMEM
4145 * Calculate the size of the zone->blockflags rounded to an unsigned long
4146 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4147 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4148 * round what is now in bits to nearest long in bits, then return it in
4151 static unsigned long __init
usemap_size(unsigned long zonesize
)
4153 unsigned long usemapsize
;
4155 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
4156 usemapsize
= usemapsize
>> pageblock_order
;
4157 usemapsize
*= NR_PAGEBLOCK_BITS
;
4158 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
4160 return usemapsize
/ 8;
4163 static void __init
setup_usemap(struct pglist_data
*pgdat
,
4164 struct zone
*zone
, unsigned long zonesize
)
4166 unsigned long usemapsize
= usemap_size(zonesize
);
4167 zone
->pageblock_flags
= NULL
;
4169 zone
->pageblock_flags
= alloc_bootmem_node_nopanic(pgdat
,
4173 static inline void setup_usemap(struct pglist_data
*pgdat
,
4174 struct zone
*zone
, unsigned long zonesize
) {}
4175 #endif /* CONFIG_SPARSEMEM */
4177 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4179 /* Return a sensible default order for the pageblock size. */
4180 static inline int pageblock_default_order(void)
4182 if (HPAGE_SHIFT
> PAGE_SHIFT
)
4183 return HUGETLB_PAGE_ORDER
;
4188 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4189 static inline void __init
set_pageblock_order(unsigned int order
)
4191 /* Check that pageblock_nr_pages has not already been setup */
4192 if (pageblock_order
)
4196 * Assume the largest contiguous order of interest is a huge page.
4197 * This value may be variable depending on boot parameters on IA64
4199 pageblock_order
= order
;
4201 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4204 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4205 * and pageblock_default_order() are unused as pageblock_order is set
4206 * at compile-time. See include/linux/pageblock-flags.h for the values of
4207 * pageblock_order based on the kernel config
4209 static inline int pageblock_default_order(unsigned int order
)
4213 #define set_pageblock_order(x) do {} while (0)
4215 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4218 * Set up the zone data structures:
4219 * - mark all pages reserved
4220 * - mark all memory queues empty
4221 * - clear the memory bitmaps
4223 static void __paginginit
free_area_init_core(struct pglist_data
*pgdat
,
4224 unsigned long *zones_size
, unsigned long *zholes_size
)
4227 int nid
= pgdat
->node_id
;
4228 unsigned long zone_start_pfn
= pgdat
->node_start_pfn
;
4231 pgdat_resize_init(pgdat
);
4232 pgdat
->nr_zones
= 0;
4233 init_waitqueue_head(&pgdat
->kswapd_wait
);
4234 pgdat
->kswapd_max_order
= 0;
4235 pgdat_page_cgroup_init(pgdat
);
4237 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
4238 struct zone
*zone
= pgdat
->node_zones
+ j
;
4239 unsigned long size
, realsize
, memmap_pages
;
4242 size
= zone_spanned_pages_in_node(nid
, j
, zones_size
);
4243 realsize
= size
- zone_absent_pages_in_node(nid
, j
,
4247 * Adjust realsize so that it accounts for how much memory
4248 * is used by this zone for memmap. This affects the watermark
4249 * and per-cpu initialisations
4252 PAGE_ALIGN(size
* sizeof(struct page
)) >> PAGE_SHIFT
;
4253 if (realsize
>= memmap_pages
) {
4254 realsize
-= memmap_pages
;
4257 " %s zone: %lu pages used for memmap\n",
4258 zone_names
[j
], memmap_pages
);
4261 " %s zone: %lu pages exceeds realsize %lu\n",
4262 zone_names
[j
], memmap_pages
, realsize
);
4264 /* Account for reserved pages */
4265 if (j
== 0 && realsize
> dma_reserve
) {
4266 realsize
-= dma_reserve
;
4267 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
4268 zone_names
[0], dma_reserve
);
4271 if (!is_highmem_idx(j
))
4272 nr_kernel_pages
+= realsize
;
4273 nr_all_pages
+= realsize
;
4275 zone
->spanned_pages
= size
;
4276 zone
->present_pages
= realsize
;
4279 zone
->min_unmapped_pages
= (realsize
*sysctl_min_unmapped_ratio
)
4281 zone
->min_slab_pages
= (realsize
* sysctl_min_slab_ratio
) / 100;
4283 zone
->name
= zone_names
[j
];
4284 spin_lock_init(&zone
->lock
);
4285 spin_lock_init(&zone
->lru_lock
);
4286 zone_seqlock_init(zone
);
4287 zone
->zone_pgdat
= pgdat
;
4289 zone_pcp_init(zone
);
4291 INIT_LIST_HEAD(&zone
->lru
[l
].list
);
4292 zone
->reclaim_stat
.nr_saved_scan
[l
] = 0;
4294 zone
->reclaim_stat
.recent_rotated
[0] = 0;
4295 zone
->reclaim_stat
.recent_rotated
[1] = 0;
4296 zone
->reclaim_stat
.recent_scanned
[0] = 0;
4297 zone
->reclaim_stat
.recent_scanned
[1] = 0;
4298 zap_zone_vm_stats(zone
);
4303 set_pageblock_order(pageblock_default_order());
4304 setup_usemap(pgdat
, zone
, size
);
4305 ret
= init_currently_empty_zone(zone
, zone_start_pfn
,
4306 size
, MEMMAP_EARLY
);
4308 memmap_init(size
, nid
, j
, zone_start_pfn
);
4309 zone_start_pfn
+= size
;
4313 static void __init_refok
alloc_node_mem_map(struct pglist_data
*pgdat
)
4315 /* Skip empty nodes */
4316 if (!pgdat
->node_spanned_pages
)
4319 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4320 /* ia64 gets its own node_mem_map, before this, without bootmem */
4321 if (!pgdat
->node_mem_map
) {
4322 unsigned long size
, start
, end
;
4326 * The zone's endpoints aren't required to be MAX_ORDER
4327 * aligned but the node_mem_map endpoints must be in order
4328 * for the buddy allocator to function correctly.
4330 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
4331 end
= pgdat
->node_start_pfn
+ pgdat
->node_spanned_pages
;
4332 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
4333 size
= (end
- start
) * sizeof(struct page
);
4334 map
= alloc_remap(pgdat
->node_id
, size
);
4336 map
= alloc_bootmem_node_nopanic(pgdat
, size
);
4337 pgdat
->node_mem_map
= map
+ (pgdat
->node_start_pfn
- start
);
4339 #ifndef CONFIG_NEED_MULTIPLE_NODES
4341 * With no DISCONTIG, the global mem_map is just set as node 0's
4343 if (pgdat
== NODE_DATA(0)) {
4344 mem_map
= NODE_DATA(0)->node_mem_map
;
4345 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
4346 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
4347 mem_map
-= (pgdat
->node_start_pfn
- ARCH_PFN_OFFSET
);
4348 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4351 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4354 void __paginginit
free_area_init_node(int nid
, unsigned long *zones_size
,
4355 unsigned long node_start_pfn
, unsigned long *zholes_size
)
4357 pg_data_t
*pgdat
= NODE_DATA(nid
);
4359 pgdat
->node_id
= nid
;
4360 pgdat
->node_start_pfn
= node_start_pfn
;
4361 calculate_node_totalpages(pgdat
, zones_size
, zholes_size
);
4363 alloc_node_mem_map(pgdat
);
4364 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4365 printk(KERN_DEBUG
"free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4366 nid
, (unsigned long)pgdat
,
4367 (unsigned long)pgdat
->node_mem_map
);
4370 free_area_init_core(pgdat
, zones_size
, zholes_size
);
4373 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
4375 #if MAX_NUMNODES > 1
4377 * Figure out the number of possible node ids.
4379 static void __init
setup_nr_node_ids(void)
4382 unsigned int highest
= 0;
4384 for_each_node_mask(node
, node_possible_map
)
4386 nr_node_ids
= highest
+ 1;
4389 static inline void setup_nr_node_ids(void)
4395 * add_active_range - Register a range of PFNs backed by physical memory
4396 * @nid: The node ID the range resides on
4397 * @start_pfn: The start PFN of the available physical memory
4398 * @end_pfn: The end PFN of the available physical memory
4400 * These ranges are stored in an early_node_map[] and later used by
4401 * free_area_init_nodes() to calculate zone sizes and holes. If the
4402 * range spans a memory hole, it is up to the architecture to ensure
4403 * the memory is not freed by the bootmem allocator. If possible
4404 * the range being registered will be merged with existing ranges.
4406 void __init
add_active_range(unsigned int nid
, unsigned long start_pfn
,
4407 unsigned long end_pfn
)
4411 mminit_dprintk(MMINIT_TRACE
, "memory_register",
4412 "Entering add_active_range(%d, %#lx, %#lx) "
4413 "%d entries of %d used\n",
4414 nid
, start_pfn
, end_pfn
,
4415 nr_nodemap_entries
, MAX_ACTIVE_REGIONS
);
4417 mminit_validate_memmodel_limits(&start_pfn
, &end_pfn
);
4419 /* Merge with existing active regions if possible */
4420 for (i
= 0; i
< nr_nodemap_entries
; i
++) {
4421 if (early_node_map
[i
].nid
!= nid
)
4424 /* Skip if an existing region covers this new one */
4425 if (start_pfn
>= early_node_map
[i
].start_pfn
&&
4426 end_pfn
<= early_node_map
[i
].end_pfn
)
4429 /* Merge forward if suitable */
4430 if (start_pfn
<= early_node_map
[i
].end_pfn
&&
4431 end_pfn
> early_node_map
[i
].end_pfn
) {
4432 early_node_map
[i
].end_pfn
= end_pfn
;
4436 /* Merge backward if suitable */
4437 if (start_pfn
< early_node_map
[i
].start_pfn
&&
4438 end_pfn
>= early_node_map
[i
].start_pfn
) {
4439 early_node_map
[i
].start_pfn
= start_pfn
;
4444 /* Check that early_node_map is large enough */
4445 if (i
>= MAX_ACTIVE_REGIONS
) {
4446 printk(KERN_CRIT
"More than %d memory regions, truncating\n",
4447 MAX_ACTIVE_REGIONS
);
4451 early_node_map
[i
].nid
= nid
;
4452 early_node_map
[i
].start_pfn
= start_pfn
;
4453 early_node_map
[i
].end_pfn
= end_pfn
;
4454 nr_nodemap_entries
= i
+ 1;
4458 * remove_active_range - Shrink an existing registered range of PFNs
4459 * @nid: The node id the range is on that should be shrunk
4460 * @start_pfn: The new PFN of the range
4461 * @end_pfn: The new PFN of the range
4463 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
4464 * The map is kept near the end physical page range that has already been
4465 * registered. This function allows an arch to shrink an existing registered
4468 void __init
remove_active_range(unsigned int nid
, unsigned long start_pfn
,
4469 unsigned long end_pfn
)
4474 printk(KERN_DEBUG
"remove_active_range (%d, %lu, %lu)\n",
4475 nid
, start_pfn
, end_pfn
);
4477 /* Find the old active region end and shrink */
4478 for_each_active_range_index_in_nid(i
, nid
) {
4479 if (early_node_map
[i
].start_pfn
>= start_pfn
&&
4480 early_node_map
[i
].end_pfn
<= end_pfn
) {
4482 early_node_map
[i
].start_pfn
= 0;
4483 early_node_map
[i
].end_pfn
= 0;
4487 if (early_node_map
[i
].start_pfn
< start_pfn
&&
4488 early_node_map
[i
].end_pfn
> start_pfn
) {
4489 unsigned long temp_end_pfn
= early_node_map
[i
].end_pfn
;
4490 early_node_map
[i
].end_pfn
= start_pfn
;
4491 if (temp_end_pfn
> end_pfn
)
4492 add_active_range(nid
, end_pfn
, temp_end_pfn
);
4495 if (early_node_map
[i
].start_pfn
>= start_pfn
&&
4496 early_node_map
[i
].end_pfn
> end_pfn
&&
4497 early_node_map
[i
].start_pfn
< end_pfn
) {
4498 early_node_map
[i
].start_pfn
= end_pfn
;
4506 /* remove the blank ones */
4507 for (i
= nr_nodemap_entries
- 1; i
> 0; i
--) {
4508 if (early_node_map
[i
].nid
!= nid
)
4510 if (early_node_map
[i
].end_pfn
)
4512 /* we found it, get rid of it */
4513 for (j
= i
; j
< nr_nodemap_entries
- 1; j
++)
4514 memcpy(&early_node_map
[j
], &early_node_map
[j
+1],
4515 sizeof(early_node_map
[j
]));
4516 j
= nr_nodemap_entries
- 1;
4517 memset(&early_node_map
[j
], 0, sizeof(early_node_map
[j
]));
4518 nr_nodemap_entries
--;
4523 * remove_all_active_ranges - Remove all currently registered regions
4525 * During discovery, it may be found that a table like SRAT is invalid
4526 * and an alternative discovery method must be used. This function removes
4527 * all currently registered regions.
4529 void __init
remove_all_active_ranges(void)
4531 memset(early_node_map
, 0, sizeof(early_node_map
));
4532 nr_nodemap_entries
= 0;
4535 /* Compare two active node_active_regions */
4536 static int __init
cmp_node_active_region(const void *a
, const void *b
)
4538 struct node_active_region
*arange
= (struct node_active_region
*)a
;
4539 struct node_active_region
*brange
= (struct node_active_region
*)b
;
4541 /* Done this way to avoid overflows */
4542 if (arange
->start_pfn
> brange
->start_pfn
)
4544 if (arange
->start_pfn
< brange
->start_pfn
)
4550 /* sort the node_map by start_pfn */
4551 void __init
sort_node_map(void)
4553 sort(early_node_map
, (size_t)nr_nodemap_entries
,
4554 sizeof(struct node_active_region
),
4555 cmp_node_active_region
, NULL
);
4558 /* Find the lowest pfn for a node */
4559 static unsigned long __init
find_min_pfn_for_node(int nid
)
4562 unsigned long min_pfn
= ULONG_MAX
;
4564 /* Assuming a sorted map, the first range found has the starting pfn */
4565 for_each_active_range_index_in_nid(i
, nid
)
4566 min_pfn
= min(min_pfn
, early_node_map
[i
].start_pfn
);
4568 if (min_pfn
== ULONG_MAX
) {
4570 "Could not find start_pfn for node %d\n", nid
);
4578 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4580 * It returns the minimum PFN based on information provided via
4581 * add_active_range().
4583 unsigned long __init
find_min_pfn_with_active_regions(void)
4585 return find_min_pfn_for_node(MAX_NUMNODES
);
4589 * early_calculate_totalpages()
4590 * Sum pages in active regions for movable zone.
4591 * Populate N_HIGH_MEMORY for calculating usable_nodes.
4593 static unsigned long __init
early_calculate_totalpages(void)
4596 unsigned long totalpages
= 0;
4598 for (i
= 0; i
< nr_nodemap_entries
; i
++) {
4599 unsigned long pages
= early_node_map
[i
].end_pfn
-
4600 early_node_map
[i
].start_pfn
;
4601 totalpages
+= pages
;
4603 node_set_state(early_node_map
[i
].nid
, N_HIGH_MEMORY
);
4609 * Find the PFN the Movable zone begins in each node. Kernel memory
4610 * is spread evenly between nodes as long as the nodes have enough
4611 * memory. When they don't, some nodes will have more kernelcore than
4614 static void __init
find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn
)
4617 unsigned long usable_startpfn
;
4618 unsigned long kernelcore_node
, kernelcore_remaining
;
4619 /* save the state before borrow the nodemask */
4620 nodemask_t saved_node_state
= node_states
[N_HIGH_MEMORY
];
4621 unsigned long totalpages
= early_calculate_totalpages();
4622 int usable_nodes
= nodes_weight(node_states
[N_HIGH_MEMORY
]);
4625 * If movablecore was specified, calculate what size of
4626 * kernelcore that corresponds so that memory usable for
4627 * any allocation type is evenly spread. If both kernelcore
4628 * and movablecore are specified, then the value of kernelcore
4629 * will be used for required_kernelcore if it's greater than
4630 * what movablecore would have allowed.
4632 if (required_movablecore
) {
4633 unsigned long corepages
;
4636 * Round-up so that ZONE_MOVABLE is at least as large as what
4637 * was requested by the user
4639 required_movablecore
=
4640 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
4641 corepages
= totalpages
- required_movablecore
;
4643 required_kernelcore
= max(required_kernelcore
, corepages
);
4646 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4647 if (!required_kernelcore
)
4650 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4651 find_usable_zone_for_movable();
4652 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
4655 /* Spread kernelcore memory as evenly as possible throughout nodes */
4656 kernelcore_node
= required_kernelcore
/ usable_nodes
;
4657 for_each_node_state(nid
, N_HIGH_MEMORY
) {
4659 * Recalculate kernelcore_node if the division per node
4660 * now exceeds what is necessary to satisfy the requested
4661 * amount of memory for the kernel
4663 if (required_kernelcore
< kernelcore_node
)
4664 kernelcore_node
= required_kernelcore
/ usable_nodes
;
4667 * As the map is walked, we track how much memory is usable
4668 * by the kernel using kernelcore_remaining. When it is
4669 * 0, the rest of the node is usable by ZONE_MOVABLE
4671 kernelcore_remaining
= kernelcore_node
;
4673 /* Go through each range of PFNs within this node */
4674 for_each_active_range_index_in_nid(i
, nid
) {
4675 unsigned long start_pfn
, end_pfn
;
4676 unsigned long size_pages
;
4678 start_pfn
= max(early_node_map
[i
].start_pfn
,
4679 zone_movable_pfn
[nid
]);
4680 end_pfn
= early_node_map
[i
].end_pfn
;
4681 if (start_pfn
>= end_pfn
)
4684 /* Account for what is only usable for kernelcore */
4685 if (start_pfn
< usable_startpfn
) {
4686 unsigned long kernel_pages
;
4687 kernel_pages
= min(end_pfn
, usable_startpfn
)
4690 kernelcore_remaining
-= min(kernel_pages
,
4691 kernelcore_remaining
);
4692 required_kernelcore
-= min(kernel_pages
,
4693 required_kernelcore
);
4695 /* Continue if range is now fully accounted */
4696 if (end_pfn
<= usable_startpfn
) {
4699 * Push zone_movable_pfn to the end so
4700 * that if we have to rebalance
4701 * kernelcore across nodes, we will
4702 * not double account here
4704 zone_movable_pfn
[nid
] = end_pfn
;
4707 start_pfn
= usable_startpfn
;
4711 * The usable PFN range for ZONE_MOVABLE is from
4712 * start_pfn->end_pfn. Calculate size_pages as the
4713 * number of pages used as kernelcore
4715 size_pages
= end_pfn
- start_pfn
;
4716 if (size_pages
> kernelcore_remaining
)
4717 size_pages
= kernelcore_remaining
;
4718 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
4721 * Some kernelcore has been met, update counts and
4722 * break if the kernelcore for this node has been
4725 required_kernelcore
-= min(required_kernelcore
,
4727 kernelcore_remaining
-= size_pages
;
4728 if (!kernelcore_remaining
)
4734 * If there is still required_kernelcore, we do another pass with one
4735 * less node in the count. This will push zone_movable_pfn[nid] further
4736 * along on the nodes that still have memory until kernelcore is
4740 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
4743 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4744 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
4745 zone_movable_pfn
[nid
] =
4746 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
4749 /* restore the node_state */
4750 node_states
[N_HIGH_MEMORY
] = saved_node_state
;
4753 /* Any regular memory on that node ? */
4754 static void check_for_regular_memory(pg_data_t
*pgdat
)
4756 #ifdef CONFIG_HIGHMEM
4757 enum zone_type zone_type
;
4759 for (zone_type
= 0; zone_type
<= ZONE_NORMAL
; zone_type
++) {
4760 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
4761 if (zone
->present_pages
)
4762 node_set_state(zone_to_nid(zone
), N_NORMAL_MEMORY
);
4768 * free_area_init_nodes - Initialise all pg_data_t and zone data
4769 * @max_zone_pfn: an array of max PFNs for each zone
4771 * This will call free_area_init_node() for each active node in the system.
4772 * Using the page ranges provided by add_active_range(), the size of each
4773 * zone in each node and their holes is calculated. If the maximum PFN
4774 * between two adjacent zones match, it is assumed that the zone is empty.
4775 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4776 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4777 * starts where the previous one ended. For example, ZONE_DMA32 starts
4778 * at arch_max_dma_pfn.
4780 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
4785 /* Sort early_node_map as initialisation assumes it is sorted */
4788 /* Record where the zone boundaries are */
4789 memset(arch_zone_lowest_possible_pfn
, 0,
4790 sizeof(arch_zone_lowest_possible_pfn
));
4791 memset(arch_zone_highest_possible_pfn
, 0,
4792 sizeof(arch_zone_highest_possible_pfn
));
4793 arch_zone_lowest_possible_pfn
[0] = find_min_pfn_with_active_regions();
4794 arch_zone_highest_possible_pfn
[0] = max_zone_pfn
[0];
4795 for (i
= 1; i
< MAX_NR_ZONES
; i
++) {
4796 if (i
== ZONE_MOVABLE
)
4798 arch_zone_lowest_possible_pfn
[i
] =
4799 arch_zone_highest_possible_pfn
[i
-1];
4800 arch_zone_highest_possible_pfn
[i
] =
4801 max(max_zone_pfn
[i
], arch_zone_lowest_possible_pfn
[i
]);
4803 arch_zone_lowest_possible_pfn
[ZONE_MOVABLE
] = 0;
4804 arch_zone_highest_possible_pfn
[ZONE_MOVABLE
] = 0;
4806 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4807 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
4808 find_zone_movable_pfns_for_nodes(zone_movable_pfn
);
4810 /* Print out the zone ranges */
4811 printk("Zone PFN ranges:\n");
4812 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
4813 if (i
== ZONE_MOVABLE
)
4815 printk(" %-8s ", zone_names
[i
]);
4816 if (arch_zone_lowest_possible_pfn
[i
] ==
4817 arch_zone_highest_possible_pfn
[i
])
4820 printk("%0#10lx -> %0#10lx\n",
4821 arch_zone_lowest_possible_pfn
[i
],
4822 arch_zone_highest_possible_pfn
[i
]);
4825 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4826 printk("Movable zone start PFN for each node\n");
4827 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
4828 if (zone_movable_pfn
[i
])
4829 printk(" Node %d: %lu\n", i
, zone_movable_pfn
[i
]);
4832 /* Print out the early_node_map[] */
4833 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries
);
4834 for (i
= 0; i
< nr_nodemap_entries
; i
++)
4835 printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map
[i
].nid
,
4836 early_node_map
[i
].start_pfn
,
4837 early_node_map
[i
].end_pfn
);
4839 /* Initialise every node */
4840 mminit_verify_pageflags_layout();
4841 setup_nr_node_ids();
4842 for_each_online_node(nid
) {
4843 pg_data_t
*pgdat
= NODE_DATA(nid
);
4844 free_area_init_node(nid
, NULL
,
4845 find_min_pfn_for_node(nid
), NULL
);
4847 /* Any memory on that node */
4848 if (pgdat
->node_present_pages
)
4849 node_set_state(nid
, N_HIGH_MEMORY
);
4850 check_for_regular_memory(pgdat
);
4854 static int __init
cmdline_parse_core(char *p
, unsigned long *core
)
4856 unsigned long long coremem
;
4860 coremem
= memparse(p
, &p
);
4861 *core
= coremem
>> PAGE_SHIFT
;
4863 /* Paranoid check that UL is enough for the coremem value */
4864 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
4870 * kernelcore=size sets the amount of memory for use for allocations that
4871 * cannot be reclaimed or migrated.
4873 static int __init
cmdline_parse_kernelcore(char *p
)
4875 return cmdline_parse_core(p
, &required_kernelcore
);
4879 * movablecore=size sets the amount of memory for use for allocations that
4880 * can be reclaimed or migrated.
4882 static int __init
cmdline_parse_movablecore(char *p
)
4884 return cmdline_parse_core(p
, &required_movablecore
);
4887 early_param("kernelcore", cmdline_parse_kernelcore
);
4888 early_param("movablecore", cmdline_parse_movablecore
);
4890 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4893 * set_dma_reserve - set the specified number of pages reserved in the first zone
4894 * @new_dma_reserve: The number of pages to mark reserved
4896 * The per-cpu batchsize and zone watermarks are determined by present_pages.
4897 * In the DMA zone, a significant percentage may be consumed by kernel image
4898 * and other unfreeable allocations which can skew the watermarks badly. This
4899 * function may optionally be used to account for unfreeable pages in the
4900 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
4901 * smaller per-cpu batchsize.
4903 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
4905 dma_reserve
= new_dma_reserve
;
4908 void __init
free_area_init(unsigned long *zones_size
)
4910 free_area_init_node(0, zones_size
,
4911 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
4914 static int page_alloc_cpu_notify(struct notifier_block
*self
,
4915 unsigned long action
, void *hcpu
)
4917 int cpu
= (unsigned long)hcpu
;
4919 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
) {
4923 * Spill the event counters of the dead processor
4924 * into the current processors event counters.
4925 * This artificially elevates the count of the current
4928 vm_events_fold_cpu(cpu
);
4931 * Zero the differential counters of the dead processor
4932 * so that the vm statistics are consistent.
4934 * This is only okay since the processor is dead and cannot
4935 * race with what we are doing.
4937 refresh_cpu_vm_stats(cpu
);
4942 void __init
page_alloc_init(void)
4944 hotcpu_notifier(page_alloc_cpu_notify
, 0);
4948 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
4949 * or min_free_kbytes changes.
4951 static void calculate_totalreserve_pages(void)
4953 struct pglist_data
*pgdat
;
4954 unsigned long reserve_pages
= 0;
4955 enum zone_type i
, j
;
4957 for_each_online_pgdat(pgdat
) {
4958 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
4959 struct zone
*zone
= pgdat
->node_zones
+ i
;
4960 unsigned long max
= 0;
4962 /* Find valid and maximum lowmem_reserve in the zone */
4963 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
4964 if (zone
->lowmem_reserve
[j
] > max
)
4965 max
= zone
->lowmem_reserve
[j
];
4968 /* we treat the high watermark as reserved pages. */
4969 max
+= high_wmark_pages(zone
);
4971 if (max
> zone
->present_pages
)
4972 max
= zone
->present_pages
;
4973 reserve_pages
+= max
;
4976 totalreserve_pages
= reserve_pages
;
4980 * setup_per_zone_lowmem_reserve - called whenever
4981 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
4982 * has a correct pages reserved value, so an adequate number of
4983 * pages are left in the zone after a successful __alloc_pages().
4985 static void setup_per_zone_lowmem_reserve(void)
4987 struct pglist_data
*pgdat
;
4988 enum zone_type j
, idx
;
4990 for_each_online_pgdat(pgdat
) {
4991 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
4992 struct zone
*zone
= pgdat
->node_zones
+ j
;
4993 unsigned long present_pages
= zone
->present_pages
;
4995 zone
->lowmem_reserve
[j
] = 0;
4999 struct zone
*lower_zone
;
5003 if (sysctl_lowmem_reserve_ratio
[idx
] < 1)
5004 sysctl_lowmem_reserve_ratio
[idx
] = 1;
5006 lower_zone
= pgdat
->node_zones
+ idx
;
5007 lower_zone
->lowmem_reserve
[j
] = present_pages
/
5008 sysctl_lowmem_reserve_ratio
[idx
];
5009 present_pages
+= lower_zone
->present_pages
;
5014 /* update totalreserve_pages */
5015 calculate_totalreserve_pages();
5019 * setup_per_zone_wmarks - called when min_free_kbytes changes
5020 * or when memory is hot-{added|removed}
5022 * Ensures that the watermark[min,low,high] values for each zone are set
5023 * correctly with respect to min_free_kbytes.
5025 void setup_per_zone_wmarks(void)
5027 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
5028 unsigned long lowmem_pages
= 0;
5030 unsigned long flags
;
5032 /* Calculate total number of !ZONE_HIGHMEM pages */
5033 for_each_zone(zone
) {
5034 if (!is_highmem(zone
))
5035 lowmem_pages
+= zone
->present_pages
;
5038 for_each_zone(zone
) {
5041 spin_lock_irqsave(&zone
->lock
, flags
);
5042 tmp
= (u64
)pages_min
* zone
->present_pages
;
5043 do_div(tmp
, lowmem_pages
);
5044 if (is_highmem(zone
)) {
5046 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5047 * need highmem pages, so cap pages_min to a small
5050 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5051 * deltas controls asynch page reclaim, and so should
5052 * not be capped for highmem.
5056 min_pages
= zone
->present_pages
/ 1024;
5057 if (min_pages
< SWAP_CLUSTER_MAX
)
5058 min_pages
= SWAP_CLUSTER_MAX
;
5059 if (min_pages
> 128)
5061 zone
->watermark
[WMARK_MIN
] = min_pages
;
5064 * If it's a lowmem zone, reserve a number of pages
5065 * proportionate to the zone's size.
5067 zone
->watermark
[WMARK_MIN
] = tmp
;
5070 zone
->watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + (tmp
>> 2);
5071 zone
->watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + (tmp
>> 1);
5072 setup_zone_migrate_reserve(zone
);
5073 spin_unlock_irqrestore(&zone
->lock
, flags
);
5076 /* update totalreserve_pages */
5077 calculate_totalreserve_pages();
5081 * The inactive anon list should be small enough that the VM never has to
5082 * do too much work, but large enough that each inactive page has a chance
5083 * to be referenced again before it is swapped out.
5085 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5086 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5087 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5088 * the anonymous pages are kept on the inactive list.
5091 * memory ratio inactive anon
5092 * -------------------------------------
5101 void calculate_zone_inactive_ratio(struct zone
*zone
)
5103 unsigned int gb
, ratio
;
5105 /* Zone size in gigabytes */
5106 gb
= zone
->present_pages
>> (30 - PAGE_SHIFT
);
5108 ratio
= int_sqrt(10 * gb
);
5112 zone
->inactive_ratio
= ratio
;
5115 static void __init
setup_per_zone_inactive_ratio(void)
5120 calculate_zone_inactive_ratio(zone
);
5124 * Initialise min_free_kbytes.
5126 * For small machines we want it small (128k min). For large machines
5127 * we want it large (64MB max). But it is not linear, because network
5128 * bandwidth does not increase linearly with machine size. We use
5130 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5131 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5147 static int __init
init_per_zone_wmark_min(void)
5149 unsigned long lowmem_kbytes
;
5151 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
5153 min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
5154 if (min_free_kbytes
< 128)
5155 min_free_kbytes
= 128;
5156 if (min_free_kbytes
> 65536)
5157 min_free_kbytes
= 65536;
5158 setup_per_zone_wmarks();
5159 setup_per_zone_lowmem_reserve();
5160 setup_per_zone_inactive_ratio();
5163 module_init(init_per_zone_wmark_min
)
5166 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5167 * that we can call two helper functions whenever min_free_kbytes
5170 int min_free_kbytes_sysctl_handler(ctl_table
*table
, int write
,
5171 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5173 proc_dointvec(table
, write
, buffer
, length
, ppos
);
5175 setup_per_zone_wmarks();
5180 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table
*table
, int write
,
5181 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5186 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5191 zone
->min_unmapped_pages
= (zone
->present_pages
*
5192 sysctl_min_unmapped_ratio
) / 100;
5196 int sysctl_min_slab_ratio_sysctl_handler(ctl_table
*table
, int write
,
5197 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5202 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5207 zone
->min_slab_pages
= (zone
->present_pages
*
5208 sysctl_min_slab_ratio
) / 100;
5214 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5215 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5216 * whenever sysctl_lowmem_reserve_ratio changes.
5218 * The reserve ratio obviously has absolutely no relation with the
5219 * minimum watermarks. The lowmem reserve ratio can only make sense
5220 * if in function of the boot time zone sizes.
5222 int lowmem_reserve_ratio_sysctl_handler(ctl_table
*table
, int write
,
5223 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5225 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5226 setup_per_zone_lowmem_reserve();
5231 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5232 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
5233 * can have before it gets flushed back to buddy allocator.
5236 int percpu_pagelist_fraction_sysctl_handler(ctl_table
*table
, int write
,
5237 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5243 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5244 if (!write
|| (ret
== -EINVAL
))
5246 for_each_populated_zone(zone
) {
5247 for_each_possible_cpu(cpu
) {
5249 high
= zone
->present_pages
/ percpu_pagelist_fraction
;
5250 setup_pagelist_highmark(
5251 per_cpu_ptr(zone
->pageset
, cpu
), high
);
5257 int hashdist
= HASHDIST_DEFAULT
;
5260 static int __init
set_hashdist(char *str
)
5264 hashdist
= simple_strtoul(str
, &str
, 0);
5267 __setup("hashdist=", set_hashdist
);
5271 * allocate a large system hash table from bootmem
5272 * - it is assumed that the hash table must contain an exact power-of-2
5273 * quantity of entries
5274 * - limit is the number of hash buckets, not the total allocation size
5276 void *__init
alloc_large_system_hash(const char *tablename
,
5277 unsigned long bucketsize
,
5278 unsigned long numentries
,
5281 unsigned int *_hash_shift
,
5282 unsigned int *_hash_mask
,
5283 unsigned long limit
)
5285 unsigned long long max
= limit
;
5286 unsigned long log2qty
, size
;
5289 /* allow the kernel cmdline to have a say */
5291 /* round applicable memory size up to nearest megabyte */
5292 numentries
= nr_kernel_pages
;
5293 numentries
+= (1UL << (20 - PAGE_SHIFT
)) - 1;
5294 numentries
>>= 20 - PAGE_SHIFT
;
5295 numentries
<<= 20 - PAGE_SHIFT
;
5297 /* limit to 1 bucket per 2^scale bytes of low memory */
5298 if (scale
> PAGE_SHIFT
)
5299 numentries
>>= (scale
- PAGE_SHIFT
);
5301 numentries
<<= (PAGE_SHIFT
- scale
);
5303 /* Make sure we've got at least a 0-order allocation.. */
5304 if (unlikely(flags
& HASH_SMALL
)) {
5305 /* Makes no sense without HASH_EARLY */
5306 WARN_ON(!(flags
& HASH_EARLY
));
5307 if (!(numentries
>> *_hash_shift
)) {
5308 numentries
= 1UL << *_hash_shift
;
5309 BUG_ON(!numentries
);
5311 } else if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
5312 numentries
= PAGE_SIZE
/ bucketsize
;
5314 numentries
= roundup_pow_of_two(numentries
);
5316 /* limit allocation size to 1/16 total memory by default */
5318 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
5319 do_div(max
, bucketsize
);
5322 if (numentries
> max
)
5325 log2qty
= ilog2(numentries
);
5328 size
= bucketsize
<< log2qty
;
5329 if (flags
& HASH_EARLY
)
5330 table
= alloc_bootmem_nopanic(size
);
5332 table
= __vmalloc(size
, GFP_ATOMIC
, PAGE_KERNEL
);
5335 * If bucketsize is not a power-of-two, we may free
5336 * some pages at the end of hash table which
5337 * alloc_pages_exact() automatically does
5339 if (get_order(size
) < MAX_ORDER
) {
5340 table
= alloc_pages_exact(size
, GFP_ATOMIC
);
5341 kmemleak_alloc(table
, size
, 1, GFP_ATOMIC
);
5344 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
5347 panic("Failed to allocate %s hash table\n", tablename
);
5349 printk(KERN_INFO
"%s hash table entries: %ld (order: %d, %lu bytes)\n",
5352 ilog2(size
) - PAGE_SHIFT
,
5356 *_hash_shift
= log2qty
;
5358 *_hash_mask
= (1 << log2qty
) - 1;
5363 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5364 static inline unsigned long *get_pageblock_bitmap(struct zone
*zone
,
5367 #ifdef CONFIG_SPARSEMEM
5368 return __pfn_to_section(pfn
)->pageblock_flags
;
5370 return zone
->pageblock_flags
;
5371 #endif /* CONFIG_SPARSEMEM */
5374 static inline int pfn_to_bitidx(struct zone
*zone
, unsigned long pfn
)
5376 #ifdef CONFIG_SPARSEMEM
5377 pfn
&= (PAGES_PER_SECTION
-1);
5378 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
5380 pfn
= pfn
- zone
->zone_start_pfn
;
5381 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
5382 #endif /* CONFIG_SPARSEMEM */
5386 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5387 * @page: The page within the block of interest
5388 * @start_bitidx: The first bit of interest to retrieve
5389 * @end_bitidx: The last bit of interest
5390 * returns pageblock_bits flags
5392 unsigned long get_pageblock_flags_group(struct page
*page
,
5393 int start_bitidx
, int end_bitidx
)
5396 unsigned long *bitmap
;
5397 unsigned long pfn
, bitidx
;
5398 unsigned long flags
= 0;
5399 unsigned long value
= 1;
5401 zone
= page_zone(page
);
5402 pfn
= page_to_pfn(page
);
5403 bitmap
= get_pageblock_bitmap(zone
, pfn
);
5404 bitidx
= pfn_to_bitidx(zone
, pfn
);
5406 for (; start_bitidx
<= end_bitidx
; start_bitidx
++, value
<<= 1)
5407 if (test_bit(bitidx
+ start_bitidx
, bitmap
))
5414 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5415 * @page: The page within the block of interest
5416 * @start_bitidx: The first bit of interest
5417 * @end_bitidx: The last bit of interest
5418 * @flags: The flags to set
5420 void set_pageblock_flags_group(struct page
*page
, unsigned long flags
,
5421 int start_bitidx
, int end_bitidx
)
5424 unsigned long *bitmap
;
5425 unsigned long pfn
, bitidx
;
5426 unsigned long value
= 1;
5428 zone
= page_zone(page
);
5429 pfn
= page_to_pfn(page
);
5430 bitmap
= get_pageblock_bitmap(zone
, pfn
);
5431 bitidx
= pfn_to_bitidx(zone
, pfn
);
5432 VM_BUG_ON(pfn
< zone
->zone_start_pfn
);
5433 VM_BUG_ON(pfn
>= zone
->zone_start_pfn
+ zone
->spanned_pages
);
5435 for (; start_bitidx
<= end_bitidx
; start_bitidx
++, value
<<= 1)
5437 __set_bit(bitidx
+ start_bitidx
, bitmap
);
5439 __clear_bit(bitidx
+ start_bitidx
, bitmap
);
5443 * This is designed as sub function...plz see page_isolation.c also.
5444 * set/clear page block's type to be ISOLATE.
5445 * page allocater never alloc memory from ISOLATE block.
5449 __count_immobile_pages(struct zone
*zone
, struct page
*page
, int count
)
5451 unsigned long pfn
, iter
, found
;
5453 * For avoiding noise data, lru_add_drain_all() should be called
5454 * If ZONE_MOVABLE, the zone never contains immobile pages
5456 if (zone_idx(zone
) == ZONE_MOVABLE
)
5459 if (get_pageblock_migratetype(page
) == MIGRATE_MOVABLE
)
5462 pfn
= page_to_pfn(page
);
5463 for (found
= 0, iter
= 0; iter
< pageblock_nr_pages
; iter
++) {
5464 unsigned long check
= pfn
+ iter
;
5466 if (!pfn_valid_within(check
))
5469 page
= pfn_to_page(check
);
5470 if (!page_count(page
)) {
5471 if (PageBuddy(page
))
5472 iter
+= (1 << page_order(page
)) - 1;
5478 * If there are RECLAIMABLE pages, we need to check it.
5479 * But now, memory offline itself doesn't call shrink_slab()
5480 * and it still to be fixed.
5483 * If the page is not RAM, page_count()should be 0.
5484 * we don't need more check. This is an _used_ not-movable page.
5486 * The problematic thing here is PG_reserved pages. PG_reserved
5487 * is set to both of a memory hole page and a _used_ kernel
5496 bool is_pageblock_removable_nolock(struct page
*page
)
5498 struct zone
*zone
= page_zone(page
);
5499 return __count_immobile_pages(zone
, page
, 0);
5502 int set_migratetype_isolate(struct page
*page
)
5505 unsigned long flags
, pfn
;
5506 struct memory_isolate_notify arg
;
5511 zone
= page_zone(page
);
5512 zone_idx
= zone_idx(zone
);
5514 spin_lock_irqsave(&zone
->lock
, flags
);
5516 pfn
= page_to_pfn(page
);
5517 arg
.start_pfn
= pfn
;
5518 arg
.nr_pages
= pageblock_nr_pages
;
5519 arg
.pages_found
= 0;
5522 * It may be possible to isolate a pageblock even if the
5523 * migratetype is not MIGRATE_MOVABLE. The memory isolation
5524 * notifier chain is used by balloon drivers to return the
5525 * number of pages in a range that are held by the balloon
5526 * driver to shrink memory. If all the pages are accounted for
5527 * by balloons, are free, or on the LRU, isolation can continue.
5528 * Later, for example, when memory hotplug notifier runs, these
5529 * pages reported as "can be isolated" should be isolated(freed)
5530 * by the balloon driver through the memory notifier chain.
5532 notifier_ret
= memory_isolate_notify(MEM_ISOLATE_COUNT
, &arg
);
5533 notifier_ret
= notifier_to_errno(notifier_ret
);
5537 * FIXME: Now, memory hotplug doesn't call shrink_slab() by itself.
5538 * We just check MOVABLE pages.
5540 if (__count_immobile_pages(zone
, page
, arg
.pages_found
))
5544 * immobile means "not-on-lru" paes. If immobile is larger than
5545 * removable-by-driver pages reported by notifier, we'll fail.
5550 set_pageblock_migratetype(page
, MIGRATE_ISOLATE
);
5551 move_freepages_block(zone
, page
, MIGRATE_ISOLATE
);
5554 spin_unlock_irqrestore(&zone
->lock
, flags
);
5560 void unset_migratetype_isolate(struct page
*page
)
5563 unsigned long flags
;
5564 zone
= page_zone(page
);
5565 spin_lock_irqsave(&zone
->lock
, flags
);
5566 if (get_pageblock_migratetype(page
) != MIGRATE_ISOLATE
)
5568 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
5569 move_freepages_block(zone
, page
, MIGRATE_MOVABLE
);
5571 spin_unlock_irqrestore(&zone
->lock
, flags
);
5574 #ifdef CONFIG_MEMORY_HOTREMOVE
5576 * All pages in the range must be isolated before calling this.
5579 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
5585 unsigned long flags
;
5586 /* find the first valid pfn */
5587 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
5592 zone
= page_zone(pfn_to_page(pfn
));
5593 spin_lock_irqsave(&zone
->lock
, flags
);
5595 while (pfn
< end_pfn
) {
5596 if (!pfn_valid(pfn
)) {
5600 page
= pfn_to_page(pfn
);
5601 BUG_ON(page_count(page
));
5602 BUG_ON(!PageBuddy(page
));
5603 order
= page_order(page
);
5604 #ifdef CONFIG_DEBUG_VM
5605 printk(KERN_INFO
"remove from free list %lx %d %lx\n",
5606 pfn
, 1 << order
, end_pfn
);
5608 list_del(&page
->lru
);
5609 rmv_page_order(page
);
5610 zone
->free_area
[order
].nr_free
--;
5611 __mod_zone_page_state(zone
, NR_FREE_PAGES
,
5613 for (i
= 0; i
< (1 << order
); i
++)
5614 SetPageReserved((page
+i
));
5615 pfn
+= (1 << order
);
5617 spin_unlock_irqrestore(&zone
->lock
, flags
);
5621 #ifdef CONFIG_MEMORY_FAILURE
5622 bool is_free_buddy_page(struct page
*page
)
5624 struct zone
*zone
= page_zone(page
);
5625 unsigned long pfn
= page_to_pfn(page
);
5626 unsigned long flags
;
5629 spin_lock_irqsave(&zone
->lock
, flags
);
5630 for (order
= 0; order
< MAX_ORDER
; order
++) {
5631 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
5633 if (PageBuddy(page_head
) && page_order(page_head
) >= order
)
5636 spin_unlock_irqrestore(&zone
->lock
, flags
);
5638 return order
< MAX_ORDER
;
5642 static struct trace_print_flags pageflag_names
[] = {
5643 {1UL << PG_locked
, "locked" },
5644 {1UL << PG_error
, "error" },
5645 {1UL << PG_referenced
, "referenced" },
5646 {1UL << PG_uptodate
, "uptodate" },
5647 {1UL << PG_dirty
, "dirty" },
5648 {1UL << PG_lru
, "lru" },
5649 {1UL << PG_active
, "active" },
5650 {1UL << PG_slab
, "slab" },
5651 {1UL << PG_owner_priv_1
, "owner_priv_1" },
5652 {1UL << PG_arch_1
, "arch_1" },
5653 {1UL << PG_reserved
, "reserved" },
5654 {1UL << PG_private
, "private" },
5655 {1UL << PG_private_2
, "private_2" },
5656 {1UL << PG_writeback
, "writeback" },
5657 #ifdef CONFIG_PAGEFLAGS_EXTENDED
5658 {1UL << PG_head
, "head" },
5659 {1UL << PG_tail
, "tail" },
5661 {1UL << PG_compound
, "compound" },
5663 {1UL << PG_swapcache
, "swapcache" },
5664 {1UL << PG_mappedtodisk
, "mappedtodisk" },
5665 {1UL << PG_reclaim
, "reclaim" },
5666 {1UL << PG_swapbacked
, "swapbacked" },
5667 {1UL << PG_unevictable
, "unevictable" },
5669 {1UL << PG_mlocked
, "mlocked" },
5671 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
5672 {1UL << PG_uncached
, "uncached" },
5674 #ifdef CONFIG_MEMORY_FAILURE
5675 {1UL << PG_hwpoison
, "hwpoison" },
5680 static void dump_page_flags(unsigned long flags
)
5682 const char *delim
= "";
5686 printk(KERN_ALERT
"page flags: %#lx(", flags
);
5688 /* remove zone id */
5689 flags
&= (1UL << NR_PAGEFLAGS
) - 1;
5691 for (i
= 0; pageflag_names
[i
].name
&& flags
; i
++) {
5693 mask
= pageflag_names
[i
].mask
;
5694 if ((flags
& mask
) != mask
)
5698 printk("%s%s", delim
, pageflag_names
[i
].name
);
5702 /* check for left over flags */
5704 printk("%s%#lx", delim
, flags
);
5709 void dump_page(struct page
*page
)
5712 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
5713 page
, atomic_read(&page
->_count
), page_mapcount(page
),
5714 page
->mapping
, page
->index
);
5715 dump_page_flags(page
->flags
);
5716 mem_cgroup_print_bad_page(page
);