2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kmemcheck.h>
28 #include <linux/module.h>
29 #include <linux/suspend.h>
30 #include <linux/pagevec.h>
31 #include <linux/blkdev.h>
32 #include <linux/slab.h>
33 #include <linux/oom.h>
34 #include <linux/notifier.h>
35 #include <linux/topology.h>
36 #include <linux/sysctl.h>
37 #include <linux/cpu.h>
38 #include <linux/cpuset.h>
39 #include <linux/memory_hotplug.h>
40 #include <linux/nodemask.h>
41 #include <linux/vmalloc.h>
42 #include <linux/mempolicy.h>
43 #include <linux/stop_machine.h>
44 #include <linux/sort.h>
45 #include <linux/pfn.h>
46 #include <linux/backing-dev.h>
47 #include <linux/fault-inject.h>
48 #include <linux/page-isolation.h>
49 #include <linux/page_cgroup.h>
50 #include <linux/debugobjects.h>
51 #include <linux/kmemleak.h>
52 #include <linux/memory.h>
53 #include <linux/compaction.h>
54 #include <trace/events/kmem.h>
55 #include <linux/ftrace_event.h>
57 #include <asm/tlbflush.h>
58 #include <asm/div64.h>
61 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
62 DEFINE_PER_CPU(int, numa_node
);
63 EXPORT_PER_CPU_SYMBOL(numa_node
);
66 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
68 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
69 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
70 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
71 * defined in <linux/topology.h>.
73 DEFINE_PER_CPU(int, _numa_mem_
); /* Kernel "local memory" node */
74 EXPORT_PER_CPU_SYMBOL(_numa_mem_
);
78 * Array of node states.
80 nodemask_t node_states
[NR_NODE_STATES
] __read_mostly
= {
81 [N_POSSIBLE
] = NODE_MASK_ALL
,
82 [N_ONLINE
] = { { [0] = 1UL } },
84 [N_NORMAL_MEMORY
] = { { [0] = 1UL } },
86 [N_HIGH_MEMORY
] = { { [0] = 1UL } },
88 [N_CPU
] = { { [0] = 1UL } },
91 EXPORT_SYMBOL(node_states
);
93 unsigned long totalram_pages __read_mostly
;
94 unsigned long totalreserve_pages __read_mostly
;
95 int percpu_pagelist_fraction
;
96 gfp_t gfp_allowed_mask __read_mostly
= GFP_BOOT_MASK
;
98 #ifdef CONFIG_PM_SLEEP
100 * The following functions are used by the suspend/hibernate code to temporarily
101 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
102 * while devices are suspended. To avoid races with the suspend/hibernate code,
103 * they should always be called with pm_mutex held (gfp_allowed_mask also should
104 * only be modified with pm_mutex held, unless the suspend/hibernate code is
105 * guaranteed not to run in parallel with that modification).
108 static gfp_t saved_gfp_mask
;
110 void pm_restore_gfp_mask(void)
112 WARN_ON(!mutex_is_locked(&pm_mutex
));
113 if (saved_gfp_mask
) {
114 gfp_allowed_mask
= saved_gfp_mask
;
119 void pm_restrict_gfp_mask(void)
121 WARN_ON(!mutex_is_locked(&pm_mutex
));
122 WARN_ON(saved_gfp_mask
);
123 saved_gfp_mask
= gfp_allowed_mask
;
124 gfp_allowed_mask
&= ~GFP_IOFS
;
126 #endif /* CONFIG_PM_SLEEP */
128 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
129 int pageblock_order __read_mostly
;
132 static void __free_pages_ok(struct page
*page
, unsigned int order
);
135 * results with 256, 32 in the lowmem_reserve sysctl:
136 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
137 * 1G machine -> (16M dma, 784M normal, 224M high)
138 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
139 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
140 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
142 * TBD: should special case ZONE_DMA32 machines here - in those we normally
143 * don't need any ZONE_NORMAL reservation
145 int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
-1] = {
146 #ifdef CONFIG_ZONE_DMA
149 #ifdef CONFIG_ZONE_DMA32
152 #ifdef CONFIG_HIGHMEM
158 EXPORT_SYMBOL(totalram_pages
);
160 static char * const zone_names
[MAX_NR_ZONES
] = {
161 #ifdef CONFIG_ZONE_DMA
164 #ifdef CONFIG_ZONE_DMA32
168 #ifdef CONFIG_HIGHMEM
174 int min_free_kbytes
= 1024;
176 static unsigned long __meminitdata nr_kernel_pages
;
177 static unsigned long __meminitdata nr_all_pages
;
178 static unsigned long __meminitdata dma_reserve
;
180 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
182 * MAX_ACTIVE_REGIONS determines the maximum number of distinct
183 * ranges of memory (RAM) that may be registered with add_active_range().
184 * Ranges passed to add_active_range() will be merged if possible
185 * so the number of times add_active_range() can be called is
186 * related to the number of nodes and the number of holes
188 #ifdef CONFIG_MAX_ACTIVE_REGIONS
189 /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */
190 #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS
192 #if MAX_NUMNODES >= 32
193 /* If there can be many nodes, allow up to 50 holes per node */
194 #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50)
196 /* By default, allow up to 256 distinct regions */
197 #define MAX_ACTIVE_REGIONS 256
201 static struct node_active_region __meminitdata early_node_map
[MAX_ACTIVE_REGIONS
];
202 static int __meminitdata nr_nodemap_entries
;
203 static unsigned long __meminitdata arch_zone_lowest_possible_pfn
[MAX_NR_ZONES
];
204 static unsigned long __meminitdata arch_zone_highest_possible_pfn
[MAX_NR_ZONES
];
205 static unsigned long __initdata required_kernelcore
;
206 static unsigned long __initdata required_movablecore
;
207 static unsigned long __meminitdata zone_movable_pfn
[MAX_NUMNODES
];
209 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
211 EXPORT_SYMBOL(movable_zone
);
212 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
215 int nr_node_ids __read_mostly
= MAX_NUMNODES
;
216 int nr_online_nodes __read_mostly
= 1;
217 EXPORT_SYMBOL(nr_node_ids
);
218 EXPORT_SYMBOL(nr_online_nodes
);
221 int page_group_by_mobility_disabled __read_mostly
;
223 static void set_pageblock_migratetype(struct page
*page
, int migratetype
)
226 if (unlikely(page_group_by_mobility_disabled
))
227 migratetype
= MIGRATE_UNMOVABLE
;
229 set_pageblock_flags_group(page
, (unsigned long)migratetype
,
230 PB_migrate
, PB_migrate_end
);
233 bool oom_killer_disabled __read_mostly
;
235 #ifdef CONFIG_DEBUG_VM
236 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
240 unsigned long pfn
= page_to_pfn(page
);
243 seq
= zone_span_seqbegin(zone
);
244 if (pfn
>= zone
->zone_start_pfn
+ zone
->spanned_pages
)
246 else if (pfn
< zone
->zone_start_pfn
)
248 } while (zone_span_seqretry(zone
, seq
));
253 static int page_is_consistent(struct zone
*zone
, struct page
*page
)
255 if (!pfn_valid_within(page_to_pfn(page
)))
257 if (zone
!= page_zone(page
))
263 * Temporary debugging check for pages not lying within a given zone.
265 static int bad_range(struct zone
*zone
, struct page
*page
)
267 if (page_outside_zone_boundaries(zone
, page
))
269 if (!page_is_consistent(zone
, page
))
275 static inline int bad_range(struct zone
*zone
, struct page
*page
)
281 static void bad_page(struct page
*page
)
283 static unsigned long resume
;
284 static unsigned long nr_shown
;
285 static unsigned long nr_unshown
;
287 /* Don't complain about poisoned pages */
288 if (PageHWPoison(page
)) {
289 reset_page_mapcount(page
); /* remove PageBuddy */
294 * Allow a burst of 60 reports, then keep quiet for that minute;
295 * or allow a steady drip of one report per second.
297 if (nr_shown
== 60) {
298 if (time_before(jiffies
, resume
)) {
304 "BUG: Bad page state: %lu messages suppressed\n",
311 resume
= jiffies
+ 60 * HZ
;
313 printk(KERN_ALERT
"BUG: Bad page state in process %s pfn:%05lx\n",
314 current
->comm
, page_to_pfn(page
));
319 /* Leave bad fields for debug, except PageBuddy could make trouble */
320 reset_page_mapcount(page
); /* remove PageBuddy */
321 add_taint(TAINT_BAD_PAGE
);
325 * Higher-order pages are called "compound pages". They are structured thusly:
327 * The first PAGE_SIZE page is called the "head page".
329 * The remaining PAGE_SIZE pages are called "tail pages".
331 * All pages have PG_compound set. All pages have their ->private pointing at
332 * the head page (even the head page has this).
334 * The first tail page's ->lru.next holds the address of the compound page's
335 * put_page() function. Its ->lru.prev holds the order of allocation.
336 * This usage means that zero-order pages may not be compound.
339 static void free_compound_page(struct page
*page
)
341 __free_pages_ok(page
, compound_order(page
));
344 void prep_compound_page(struct page
*page
, unsigned long order
)
347 int nr_pages
= 1 << order
;
349 set_compound_page_dtor(page
, free_compound_page
);
350 set_compound_order(page
, order
);
352 for (i
= 1; i
< nr_pages
; i
++) {
353 struct page
*p
= page
+ i
;
356 p
->first_page
= page
;
360 /* update __split_huge_page_refcount if you change this function */
361 static int destroy_compound_page(struct page
*page
, unsigned long order
)
364 int nr_pages
= 1 << order
;
367 if (unlikely(compound_order(page
) != order
) ||
368 unlikely(!PageHead(page
))) {
373 __ClearPageHead(page
);
375 for (i
= 1; i
< nr_pages
; i
++) {
376 struct page
*p
= page
+ i
;
378 if (unlikely(!PageTail(p
) || (p
->first_page
!= page
))) {
388 static inline void prep_zero_page(struct page
*page
, int order
, gfp_t gfp_flags
)
393 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
394 * and __GFP_HIGHMEM from hard or soft interrupt context.
396 VM_BUG_ON((gfp_flags
& __GFP_HIGHMEM
) && in_interrupt());
397 for (i
= 0; i
< (1 << order
); i
++)
398 clear_highpage(page
+ i
);
401 static inline void set_page_order(struct page
*page
, int order
)
403 set_page_private(page
, order
);
404 __SetPageBuddy(page
);
407 static inline void rmv_page_order(struct page
*page
)
409 __ClearPageBuddy(page
);
410 set_page_private(page
, 0);
414 * Locate the struct page for both the matching buddy in our
415 * pair (buddy1) and the combined O(n+1) page they form (page).
417 * 1) Any buddy B1 will have an order O twin B2 which satisfies
418 * the following equation:
420 * For example, if the starting buddy (buddy2) is #8 its order
422 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
424 * 2) Any buddy B will have an order O+1 parent P which
425 * satisfies the following equation:
428 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
430 static inline unsigned long
431 __find_buddy_index(unsigned long page_idx
, unsigned int order
)
433 return page_idx
^ (1 << order
);
437 * This function checks whether a page is free && is the buddy
438 * we can do coalesce a page and its buddy if
439 * (a) the buddy is not in a hole &&
440 * (b) the buddy is in the buddy system &&
441 * (c) a page and its buddy have the same order &&
442 * (d) a page and its buddy are in the same zone.
444 * For recording whether a page is in the buddy system, we set ->_mapcount -2.
445 * Setting, clearing, and testing _mapcount -2 is serialized by zone->lock.
447 * For recording page's order, we use page_private(page).
449 static inline int page_is_buddy(struct page
*page
, struct page
*buddy
,
452 if (!pfn_valid_within(page_to_pfn(buddy
)))
455 if (page_zone_id(page
) != page_zone_id(buddy
))
458 if (PageBuddy(buddy
) && page_order(buddy
) == order
) {
459 VM_BUG_ON(page_count(buddy
) != 0);
466 * Freeing function for a buddy system allocator.
468 * The concept of a buddy system is to maintain direct-mapped table
469 * (containing bit values) for memory blocks of various "orders".
470 * The bottom level table contains the map for the smallest allocatable
471 * units of memory (here, pages), and each level above it describes
472 * pairs of units from the levels below, hence, "buddies".
473 * At a high level, all that happens here is marking the table entry
474 * at the bottom level available, and propagating the changes upward
475 * as necessary, plus some accounting needed to play nicely with other
476 * parts of the VM system.
477 * At each level, we keep a list of pages, which are heads of continuous
478 * free pages of length of (1 << order) and marked with _mapcount -2. Page's
479 * order is recorded in page_private(page) field.
480 * So when we are allocating or freeing one, we can derive the state of the
481 * other. That is, if we allocate a small block, and both were
482 * free, the remainder of the region must be split into blocks.
483 * If a block is freed, and its buddy is also free, then this
484 * triggers coalescing into a block of larger size.
489 static inline void __free_one_page(struct page
*page
,
490 struct zone
*zone
, unsigned int order
,
493 unsigned long page_idx
;
494 unsigned long combined_idx
;
495 unsigned long uninitialized_var(buddy_idx
);
498 if (unlikely(PageCompound(page
)))
499 if (unlikely(destroy_compound_page(page
, order
)))
502 VM_BUG_ON(migratetype
== -1);
504 page_idx
= page_to_pfn(page
) & ((1 << MAX_ORDER
) - 1);
506 VM_BUG_ON(page_idx
& ((1 << order
) - 1));
507 VM_BUG_ON(bad_range(zone
, page
));
509 while (order
< MAX_ORDER
-1) {
510 buddy_idx
= __find_buddy_index(page_idx
, order
);
511 buddy
= page
+ (buddy_idx
- page_idx
);
512 if (!page_is_buddy(page
, buddy
, order
))
515 /* Our buddy is free, merge with it and move up one order. */
516 list_del(&buddy
->lru
);
517 zone
->free_area
[order
].nr_free
--;
518 rmv_page_order(buddy
);
519 combined_idx
= buddy_idx
& page_idx
;
520 page
= page
+ (combined_idx
- page_idx
);
521 page_idx
= combined_idx
;
524 set_page_order(page
, order
);
527 * If this is not the largest possible page, check if the buddy
528 * of the next-highest order is free. If it is, it's possible
529 * that pages are being freed that will coalesce soon. In case,
530 * that is happening, add the free page to the tail of the list
531 * so it's less likely to be used soon and more likely to be merged
532 * as a higher order page
534 if ((order
< MAX_ORDER
-2) && pfn_valid_within(page_to_pfn(buddy
))) {
535 struct page
*higher_page
, *higher_buddy
;
536 combined_idx
= buddy_idx
& page_idx
;
537 higher_page
= page
+ (combined_idx
- page_idx
);
538 buddy_idx
= __find_buddy_index(combined_idx
, order
+ 1);
539 higher_buddy
= page
+ (buddy_idx
- combined_idx
);
540 if (page_is_buddy(higher_page
, higher_buddy
, order
+ 1)) {
541 list_add_tail(&page
->lru
,
542 &zone
->free_area
[order
].free_list
[migratetype
]);
547 list_add(&page
->lru
, &zone
->free_area
[order
].free_list
[migratetype
]);
549 zone
->free_area
[order
].nr_free
++;
553 * free_page_mlock() -- clean up attempts to free and mlocked() page.
554 * Page should not be on lru, so no need to fix that up.
555 * free_pages_check() will verify...
557 static inline void free_page_mlock(struct page
*page
)
559 __dec_zone_page_state(page
, NR_MLOCK
);
560 __count_vm_event(UNEVICTABLE_MLOCKFREED
);
563 static inline int free_pages_check(struct page
*page
)
565 if (unlikely(page_mapcount(page
) |
566 (page
->mapping
!= NULL
) |
567 (atomic_read(&page
->_count
) != 0) |
568 (page
->flags
& PAGE_FLAGS_CHECK_AT_FREE
))) {
572 if (page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)
573 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
578 * Frees a number of pages from the PCP lists
579 * Assumes all pages on list are in same zone, and of same order.
580 * count is the number of pages to free.
582 * If the zone was previously in an "all pages pinned" state then look to
583 * see if this freeing clears that state.
585 * And clear the zone's pages_scanned counter, to hold off the "all pages are
586 * pinned" detection logic.
588 static void free_pcppages_bulk(struct zone
*zone
, int count
,
589 struct per_cpu_pages
*pcp
)
595 spin_lock(&zone
->lock
);
596 zone
->all_unreclaimable
= 0;
597 zone
->pages_scanned
= 0;
601 struct list_head
*list
;
604 * Remove pages from lists in a round-robin fashion. A
605 * batch_free count is maintained that is incremented when an
606 * empty list is encountered. This is so more pages are freed
607 * off fuller lists instead of spinning excessively around empty
612 if (++migratetype
== MIGRATE_PCPTYPES
)
614 list
= &pcp
->lists
[migratetype
];
615 } while (list_empty(list
));
617 /* This is the only non-empty list. Free them all. */
618 if (batch_free
== MIGRATE_PCPTYPES
)
619 batch_free
= to_free
;
622 page
= list_entry(list
->prev
, struct page
, lru
);
623 /* must delete as __free_one_page list manipulates */
624 list_del(&page
->lru
);
625 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
626 __free_one_page(page
, zone
, 0, page_private(page
));
627 trace_mm_page_pcpu_drain(page
, 0, page_private(page
));
628 } while (--to_free
&& --batch_free
&& !list_empty(list
));
630 __mod_zone_page_state(zone
, NR_FREE_PAGES
, count
);
631 spin_unlock(&zone
->lock
);
634 static void free_one_page(struct zone
*zone
, struct page
*page
, int order
,
637 spin_lock(&zone
->lock
);
638 zone
->all_unreclaimable
= 0;
639 zone
->pages_scanned
= 0;
641 __free_one_page(page
, zone
, order
, migratetype
);
642 __mod_zone_page_state(zone
, NR_FREE_PAGES
, 1 << order
);
643 spin_unlock(&zone
->lock
);
646 static bool free_pages_prepare(struct page
*page
, unsigned int order
)
651 trace_mm_page_free_direct(page
, order
);
652 kmemcheck_free_shadow(page
, order
);
655 page
->mapping
= NULL
;
656 for (i
= 0; i
< (1 << order
); i
++)
657 bad
+= free_pages_check(page
+ i
);
661 if (!PageHighMem(page
)) {
662 debug_check_no_locks_freed(page_address(page
),PAGE_SIZE
<<order
);
663 debug_check_no_obj_freed(page_address(page
),
666 arch_free_page(page
, order
);
667 kernel_map_pages(page
, 1 << order
, 0);
672 static void __free_pages_ok(struct page
*page
, unsigned int order
)
675 int wasMlocked
= __TestClearPageMlocked(page
);
677 if (!free_pages_prepare(page
, order
))
680 local_irq_save(flags
);
681 if (unlikely(wasMlocked
))
682 free_page_mlock(page
);
683 __count_vm_events(PGFREE
, 1 << order
);
684 free_one_page(page_zone(page
), page
, order
,
685 get_pageblock_migratetype(page
));
686 local_irq_restore(flags
);
690 * permit the bootmem allocator to evade page validation on high-order frees
692 void __meminit
__free_pages_bootmem(struct page
*page
, unsigned int order
)
695 __ClearPageReserved(page
);
696 set_page_count(page
, 0);
697 set_page_refcounted(page
);
703 for (loop
= 0; loop
< BITS_PER_LONG
; loop
++) {
704 struct page
*p
= &page
[loop
];
706 if (loop
+ 1 < BITS_PER_LONG
)
708 __ClearPageReserved(p
);
709 set_page_count(p
, 0);
712 set_page_refcounted(page
);
713 __free_pages(page
, order
);
719 * The order of subdivision here is critical for the IO subsystem.
720 * Please do not alter this order without good reasons and regression
721 * testing. Specifically, as large blocks of memory are subdivided,
722 * the order in which smaller blocks are delivered depends on the order
723 * they're subdivided in this function. This is the primary factor
724 * influencing the order in which pages are delivered to the IO
725 * subsystem according to empirical testing, and this is also justified
726 * by considering the behavior of a buddy system containing a single
727 * large block of memory acted on by a series of small allocations.
728 * This behavior is a critical factor in sglist merging's success.
732 static inline void expand(struct zone
*zone
, struct page
*page
,
733 int low
, int high
, struct free_area
*area
,
736 unsigned long size
= 1 << high
;
742 VM_BUG_ON(bad_range(zone
, &page
[size
]));
743 list_add(&page
[size
].lru
, &area
->free_list
[migratetype
]);
745 set_page_order(&page
[size
], high
);
750 * This page is about to be returned from the page allocator
752 static inline int check_new_page(struct page
*page
)
754 if (unlikely(page_mapcount(page
) |
755 (page
->mapping
!= NULL
) |
756 (atomic_read(&page
->_count
) != 0) |
757 (page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
))) {
764 static int prep_new_page(struct page
*page
, int order
, gfp_t gfp_flags
)
768 for (i
= 0; i
< (1 << order
); i
++) {
769 struct page
*p
= page
+ i
;
770 if (unlikely(check_new_page(p
)))
774 set_page_private(page
, 0);
775 set_page_refcounted(page
);
777 arch_alloc_page(page
, order
);
778 kernel_map_pages(page
, 1 << order
, 1);
780 if (gfp_flags
& __GFP_ZERO
)
781 prep_zero_page(page
, order
, gfp_flags
);
783 if (order
&& (gfp_flags
& __GFP_COMP
))
784 prep_compound_page(page
, order
);
790 * Go through the free lists for the given migratetype and remove
791 * the smallest available page from the freelists
794 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
797 unsigned int current_order
;
798 struct free_area
* area
;
801 /* Find a page of the appropriate size in the preferred list */
802 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
803 area
= &(zone
->free_area
[current_order
]);
804 if (list_empty(&area
->free_list
[migratetype
]))
807 page
= list_entry(area
->free_list
[migratetype
].next
,
809 list_del(&page
->lru
);
810 rmv_page_order(page
);
812 expand(zone
, page
, order
, current_order
, area
, migratetype
);
821 * This array describes the order lists are fallen back to when
822 * the free lists for the desirable migrate type are depleted
824 static int fallbacks
[MIGRATE_TYPES
][MIGRATE_TYPES
-1] = {
825 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
826 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
827 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_RESERVE
},
828 [MIGRATE_RESERVE
] = { MIGRATE_RESERVE
, MIGRATE_RESERVE
, MIGRATE_RESERVE
}, /* Never used */
832 * Move the free pages in a range to the free lists of the requested type.
833 * Note that start_page and end_pages are not aligned on a pageblock
834 * boundary. If alignment is required, use move_freepages_block()
836 static int move_freepages(struct zone
*zone
,
837 struct page
*start_page
, struct page
*end_page
,
844 #ifndef CONFIG_HOLES_IN_ZONE
846 * page_zone is not safe to call in this context when
847 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
848 * anyway as we check zone boundaries in move_freepages_block().
849 * Remove at a later date when no bug reports exist related to
850 * grouping pages by mobility
852 BUG_ON(page_zone(start_page
) != page_zone(end_page
));
855 for (page
= start_page
; page
<= end_page
;) {
856 /* Make sure we are not inadvertently changing nodes */
857 VM_BUG_ON(page_to_nid(page
) != zone_to_nid(zone
));
859 if (!pfn_valid_within(page_to_pfn(page
))) {
864 if (!PageBuddy(page
)) {
869 order
= page_order(page
);
870 list_del(&page
->lru
);
872 &zone
->free_area
[order
].free_list
[migratetype
]);
874 pages_moved
+= 1 << order
;
880 static int move_freepages_block(struct zone
*zone
, struct page
*page
,
883 unsigned long start_pfn
, end_pfn
;
884 struct page
*start_page
, *end_page
;
886 start_pfn
= page_to_pfn(page
);
887 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
888 start_page
= pfn_to_page(start_pfn
);
889 end_page
= start_page
+ pageblock_nr_pages
- 1;
890 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
892 /* Do not cross zone boundaries */
893 if (start_pfn
< zone
->zone_start_pfn
)
895 if (end_pfn
>= zone
->zone_start_pfn
+ zone
->spanned_pages
)
898 return move_freepages(zone
, start_page
, end_page
, migratetype
);
901 static void change_pageblock_range(struct page
*pageblock_page
,
902 int start_order
, int migratetype
)
904 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
906 while (nr_pageblocks
--) {
907 set_pageblock_migratetype(pageblock_page
, migratetype
);
908 pageblock_page
+= pageblock_nr_pages
;
912 /* Remove an element from the buddy allocator from the fallback list */
913 static inline struct page
*
914 __rmqueue_fallback(struct zone
*zone
, int order
, int start_migratetype
)
916 struct free_area
* area
;
921 /* Find the largest possible block of pages in the other list */
922 for (current_order
= MAX_ORDER
-1; current_order
>= order
;
924 for (i
= 0; i
< MIGRATE_TYPES
- 1; i
++) {
925 migratetype
= fallbacks
[start_migratetype
][i
];
927 /* MIGRATE_RESERVE handled later if necessary */
928 if (migratetype
== MIGRATE_RESERVE
)
931 area
= &(zone
->free_area
[current_order
]);
932 if (list_empty(&area
->free_list
[migratetype
]))
935 page
= list_entry(area
->free_list
[migratetype
].next
,
940 * If breaking a large block of pages, move all free
941 * pages to the preferred allocation list. If falling
942 * back for a reclaimable kernel allocation, be more
943 * agressive about taking ownership of free pages
945 if (unlikely(current_order
>= (pageblock_order
>> 1)) ||
946 start_migratetype
== MIGRATE_RECLAIMABLE
||
947 page_group_by_mobility_disabled
) {
949 pages
= move_freepages_block(zone
, page
,
952 /* Claim the whole block if over half of it is free */
953 if (pages
>= (1 << (pageblock_order
-1)) ||
954 page_group_by_mobility_disabled
)
955 set_pageblock_migratetype(page
,
958 migratetype
= start_migratetype
;
961 /* Remove the page from the freelists */
962 list_del(&page
->lru
);
963 rmv_page_order(page
);
965 /* Take ownership for orders >= pageblock_order */
966 if (current_order
>= pageblock_order
)
967 change_pageblock_range(page
, current_order
,
970 expand(zone
, page
, order
, current_order
, area
, migratetype
);
972 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
973 start_migratetype
, migratetype
);
983 * Do the hard work of removing an element from the buddy allocator.
984 * Call me with the zone->lock already held.
986 static struct page
*__rmqueue(struct zone
*zone
, unsigned int order
,
992 page
= __rmqueue_smallest(zone
, order
, migratetype
);
994 if (unlikely(!page
) && migratetype
!= MIGRATE_RESERVE
) {
995 page
= __rmqueue_fallback(zone
, order
, migratetype
);
998 * Use MIGRATE_RESERVE rather than fail an allocation. goto
999 * is used because __rmqueue_smallest is an inline function
1000 * and we want just one call site
1003 migratetype
= MIGRATE_RESERVE
;
1008 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
1013 * Obtain a specified number of elements from the buddy allocator, all under
1014 * a single hold of the lock, for efficiency. Add them to the supplied list.
1015 * Returns the number of new pages which were placed at *list.
1017 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
1018 unsigned long count
, struct list_head
*list
,
1019 int migratetype
, int cold
)
1023 spin_lock(&zone
->lock
);
1024 for (i
= 0; i
< count
; ++i
) {
1025 struct page
*page
= __rmqueue(zone
, order
, migratetype
);
1026 if (unlikely(page
== NULL
))
1030 * Split buddy pages returned by expand() are received here
1031 * in physical page order. The page is added to the callers and
1032 * list and the list head then moves forward. From the callers
1033 * perspective, the linked list is ordered by page number in
1034 * some conditions. This is useful for IO devices that can
1035 * merge IO requests if the physical pages are ordered
1038 if (likely(cold
== 0))
1039 list_add(&page
->lru
, list
);
1041 list_add_tail(&page
->lru
, list
);
1042 set_page_private(page
, migratetype
);
1045 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
1046 spin_unlock(&zone
->lock
);
1052 * Called from the vmstat counter updater to drain pagesets of this
1053 * currently executing processor on remote nodes after they have
1056 * Note that this function must be called with the thread pinned to
1057 * a single processor.
1059 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
1061 unsigned long flags
;
1064 local_irq_save(flags
);
1065 if (pcp
->count
>= pcp
->batch
)
1066 to_drain
= pcp
->batch
;
1068 to_drain
= pcp
->count
;
1069 free_pcppages_bulk(zone
, to_drain
, pcp
);
1070 pcp
->count
-= to_drain
;
1071 local_irq_restore(flags
);
1076 * Drain pages of the indicated processor.
1078 * The processor must either be the current processor and the
1079 * thread pinned to the current processor or a processor that
1082 static void drain_pages(unsigned int cpu
)
1084 unsigned long flags
;
1087 for_each_populated_zone(zone
) {
1088 struct per_cpu_pageset
*pset
;
1089 struct per_cpu_pages
*pcp
;
1091 local_irq_save(flags
);
1092 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
1096 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
1099 local_irq_restore(flags
);
1104 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1106 void drain_local_pages(void *arg
)
1108 drain_pages(smp_processor_id());
1112 * Spill all the per-cpu pages from all CPUs back into the buddy allocator
1114 void drain_all_pages(void)
1116 on_each_cpu(drain_local_pages
, NULL
, 1);
1119 #ifdef CONFIG_HIBERNATION
1121 void mark_free_pages(struct zone
*zone
)
1123 unsigned long pfn
, max_zone_pfn
;
1124 unsigned long flags
;
1126 struct list_head
*curr
;
1128 if (!zone
->spanned_pages
)
1131 spin_lock_irqsave(&zone
->lock
, flags
);
1133 max_zone_pfn
= zone
->zone_start_pfn
+ zone
->spanned_pages
;
1134 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
1135 if (pfn_valid(pfn
)) {
1136 struct page
*page
= pfn_to_page(pfn
);
1138 if (!swsusp_page_is_forbidden(page
))
1139 swsusp_unset_page_free(page
);
1142 for_each_migratetype_order(order
, t
) {
1143 list_for_each(curr
, &zone
->free_area
[order
].free_list
[t
]) {
1146 pfn
= page_to_pfn(list_entry(curr
, struct page
, lru
));
1147 for (i
= 0; i
< (1UL << order
); i
++)
1148 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
1151 spin_unlock_irqrestore(&zone
->lock
, flags
);
1153 #endif /* CONFIG_PM */
1156 * Free a 0-order page
1157 * cold == 1 ? free a cold page : free a hot page
1159 void free_hot_cold_page(struct page
*page
, int cold
)
1161 struct zone
*zone
= page_zone(page
);
1162 struct per_cpu_pages
*pcp
;
1163 unsigned long flags
;
1165 int wasMlocked
= __TestClearPageMlocked(page
);
1167 if (!free_pages_prepare(page
, 0))
1170 migratetype
= get_pageblock_migratetype(page
);
1171 set_page_private(page
, migratetype
);
1172 local_irq_save(flags
);
1173 if (unlikely(wasMlocked
))
1174 free_page_mlock(page
);
1175 __count_vm_event(PGFREE
);
1178 * We only track unmovable, reclaimable and movable on pcp lists.
1179 * Free ISOLATE pages back to the allocator because they are being
1180 * offlined but treat RESERVE as movable pages so we can get those
1181 * areas back if necessary. Otherwise, we may have to free
1182 * excessively into the page allocator
1184 if (migratetype
>= MIGRATE_PCPTYPES
) {
1185 if (unlikely(migratetype
== MIGRATE_ISOLATE
)) {
1186 free_one_page(zone
, page
, 0, migratetype
);
1189 migratetype
= MIGRATE_MOVABLE
;
1192 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
1194 list_add_tail(&page
->lru
, &pcp
->lists
[migratetype
]);
1196 list_add(&page
->lru
, &pcp
->lists
[migratetype
]);
1198 if (pcp
->count
>= pcp
->high
) {
1199 free_pcppages_bulk(zone
, pcp
->batch
, pcp
);
1200 pcp
->count
-= pcp
->batch
;
1204 local_irq_restore(flags
);
1208 * split_page takes a non-compound higher-order page, and splits it into
1209 * n (1<<order) sub-pages: page[0..n]
1210 * Each sub-page must be freed individually.
1212 * Note: this is probably too low level an operation for use in drivers.
1213 * Please consult with lkml before using this in your driver.
1215 void split_page(struct page
*page
, unsigned int order
)
1219 VM_BUG_ON(PageCompound(page
));
1220 VM_BUG_ON(!page_count(page
));
1222 #ifdef CONFIG_KMEMCHECK
1224 * Split shadow pages too, because free(page[0]) would
1225 * otherwise free the whole shadow.
1227 if (kmemcheck_page_is_tracked(page
))
1228 split_page(virt_to_page(page
[0].shadow
), order
);
1231 for (i
= 1; i
< (1 << order
); i
++)
1232 set_page_refcounted(page
+ i
);
1236 * Similar to split_page except the page is already free. As this is only
1237 * being used for migration, the migratetype of the block also changes.
1238 * As this is called with interrupts disabled, the caller is responsible
1239 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1242 * Note: this is probably too low level an operation for use in drivers.
1243 * Please consult with lkml before using this in your driver.
1245 int split_free_page(struct page
*page
)
1248 unsigned long watermark
;
1251 BUG_ON(!PageBuddy(page
));
1253 zone
= page_zone(page
);
1254 order
= page_order(page
);
1256 /* Obey watermarks as if the page was being allocated */
1257 watermark
= low_wmark_pages(zone
) + (1 << order
);
1258 if (!zone_watermark_ok(zone
, 0, watermark
, 0, 0))
1261 /* Remove page from free list */
1262 list_del(&page
->lru
);
1263 zone
->free_area
[order
].nr_free
--;
1264 rmv_page_order(page
);
1265 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(1UL << order
));
1267 /* Split into individual pages */
1268 set_page_refcounted(page
);
1269 split_page(page
, order
);
1271 if (order
>= pageblock_order
- 1) {
1272 struct page
*endpage
= page
+ (1 << order
) - 1;
1273 for (; page
< endpage
; page
+= pageblock_nr_pages
)
1274 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
1281 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1282 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1286 struct page
*buffered_rmqueue(struct zone
*preferred_zone
,
1287 struct zone
*zone
, int order
, gfp_t gfp_flags
,
1290 unsigned long flags
;
1292 int cold
= !!(gfp_flags
& __GFP_COLD
);
1295 if (likely(order
== 0)) {
1296 struct per_cpu_pages
*pcp
;
1297 struct list_head
*list
;
1299 local_irq_save(flags
);
1300 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
1301 list
= &pcp
->lists
[migratetype
];
1302 if (list_empty(list
)) {
1303 pcp
->count
+= rmqueue_bulk(zone
, 0,
1306 if (unlikely(list_empty(list
)))
1311 page
= list_entry(list
->prev
, struct page
, lru
);
1313 page
= list_entry(list
->next
, struct page
, lru
);
1315 list_del(&page
->lru
);
1318 if (unlikely(gfp_flags
& __GFP_NOFAIL
)) {
1320 * __GFP_NOFAIL is not to be used in new code.
1322 * All __GFP_NOFAIL callers should be fixed so that they
1323 * properly detect and handle allocation failures.
1325 * We most definitely don't want callers attempting to
1326 * allocate greater than order-1 page units with
1329 WARN_ON_ONCE(order
> 1);
1331 spin_lock_irqsave(&zone
->lock
, flags
);
1332 page
= __rmqueue(zone
, order
, migratetype
);
1333 spin_unlock(&zone
->lock
);
1336 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(1 << order
));
1339 __count_zone_vm_events(PGALLOC
, zone
, 1 << order
);
1340 zone_statistics(preferred_zone
, zone
, gfp_flags
);
1341 local_irq_restore(flags
);
1343 VM_BUG_ON(bad_range(zone
, page
));
1344 if (prep_new_page(page
, order
, gfp_flags
))
1349 local_irq_restore(flags
);
1353 /* The ALLOC_WMARK bits are used as an index to zone->watermark */
1354 #define ALLOC_WMARK_MIN WMARK_MIN
1355 #define ALLOC_WMARK_LOW WMARK_LOW
1356 #define ALLOC_WMARK_HIGH WMARK_HIGH
1357 #define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */
1359 /* Mask to get the watermark bits */
1360 #define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1)
1362 #define ALLOC_HARDER 0x10 /* try to alloc harder */
1363 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
1364 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
1366 #ifdef CONFIG_FAIL_PAGE_ALLOC
1368 static struct fail_page_alloc_attr
{
1369 struct fault_attr attr
;
1371 u32 ignore_gfp_highmem
;
1372 u32 ignore_gfp_wait
;
1375 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1377 struct dentry
*ignore_gfp_highmem_file
;
1378 struct dentry
*ignore_gfp_wait_file
;
1379 struct dentry
*min_order_file
;
1381 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1383 } fail_page_alloc
= {
1384 .attr
= FAULT_ATTR_INITIALIZER
,
1385 .ignore_gfp_wait
= 1,
1386 .ignore_gfp_highmem
= 1,
1390 static int __init
setup_fail_page_alloc(char *str
)
1392 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
1394 __setup("fail_page_alloc=", setup_fail_page_alloc
);
1396 static int should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1398 if (order
< fail_page_alloc
.min_order
)
1400 if (gfp_mask
& __GFP_NOFAIL
)
1402 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
1404 if (fail_page_alloc
.ignore_gfp_wait
&& (gfp_mask
& __GFP_WAIT
))
1407 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
1410 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1412 static int __init
fail_page_alloc_debugfs(void)
1414 mode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
1418 err
= init_fault_attr_dentries(&fail_page_alloc
.attr
,
1422 dir
= fail_page_alloc
.attr
.dentries
.dir
;
1424 fail_page_alloc
.ignore_gfp_wait_file
=
1425 debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
1426 &fail_page_alloc
.ignore_gfp_wait
);
1428 fail_page_alloc
.ignore_gfp_highmem_file
=
1429 debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
1430 &fail_page_alloc
.ignore_gfp_highmem
);
1431 fail_page_alloc
.min_order_file
=
1432 debugfs_create_u32("min-order", mode
, dir
,
1433 &fail_page_alloc
.min_order
);
1435 if (!fail_page_alloc
.ignore_gfp_wait_file
||
1436 !fail_page_alloc
.ignore_gfp_highmem_file
||
1437 !fail_page_alloc
.min_order_file
) {
1439 debugfs_remove(fail_page_alloc
.ignore_gfp_wait_file
);
1440 debugfs_remove(fail_page_alloc
.ignore_gfp_highmem_file
);
1441 debugfs_remove(fail_page_alloc
.min_order_file
);
1442 cleanup_fault_attr_dentries(&fail_page_alloc
.attr
);
1448 late_initcall(fail_page_alloc_debugfs
);
1450 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1452 #else /* CONFIG_FAIL_PAGE_ALLOC */
1454 static inline int should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1459 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1462 * Return true if free pages are above 'mark'. This takes into account the order
1463 * of the allocation.
1465 static bool __zone_watermark_ok(struct zone
*z
, int order
, unsigned long mark
,
1466 int classzone_idx
, int alloc_flags
, long free_pages
)
1468 /* free_pages my go negative - that's OK */
1472 free_pages
-= (1 << order
) + 1;
1473 if (alloc_flags
& ALLOC_HIGH
)
1475 if (alloc_flags
& ALLOC_HARDER
)
1478 if (free_pages
<= min
+ z
->lowmem_reserve
[classzone_idx
])
1480 for (o
= 0; o
< order
; o
++) {
1481 /* At the next order, this order's pages become unavailable */
1482 free_pages
-= z
->free_area
[o
].nr_free
<< o
;
1484 /* Require fewer higher order pages to be free */
1487 if (free_pages
<= min
)
1493 bool zone_watermark_ok(struct zone
*z
, int order
, unsigned long mark
,
1494 int classzone_idx
, int alloc_flags
)
1496 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
1497 zone_page_state(z
, NR_FREE_PAGES
));
1500 bool zone_watermark_ok_safe(struct zone
*z
, int order
, unsigned long mark
,
1501 int classzone_idx
, int alloc_flags
)
1503 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
1505 if (z
->percpu_drift_mark
&& free_pages
< z
->percpu_drift_mark
)
1506 free_pages
= zone_page_state_snapshot(z
, NR_FREE_PAGES
);
1508 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
1514 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1515 * skip over zones that are not allowed by the cpuset, or that have
1516 * been recently (in last second) found to be nearly full. See further
1517 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1518 * that have to skip over a lot of full or unallowed zones.
1520 * If the zonelist cache is present in the passed in zonelist, then
1521 * returns a pointer to the allowed node mask (either the current
1522 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1524 * If the zonelist cache is not available for this zonelist, does
1525 * nothing and returns NULL.
1527 * If the fullzones BITMAP in the zonelist cache is stale (more than
1528 * a second since last zap'd) then we zap it out (clear its bits.)
1530 * We hold off even calling zlc_setup, until after we've checked the
1531 * first zone in the zonelist, on the theory that most allocations will
1532 * be satisfied from that first zone, so best to examine that zone as
1533 * quickly as we can.
1535 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1537 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1538 nodemask_t
*allowednodes
; /* zonelist_cache approximation */
1540 zlc
= zonelist
->zlcache_ptr
;
1544 if (time_after(jiffies
, zlc
->last_full_zap
+ HZ
)) {
1545 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
1546 zlc
->last_full_zap
= jiffies
;
1549 allowednodes
= !in_interrupt() && (alloc_flags
& ALLOC_CPUSET
) ?
1550 &cpuset_current_mems_allowed
:
1551 &node_states
[N_HIGH_MEMORY
];
1552 return allowednodes
;
1556 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1557 * if it is worth looking at further for free memory:
1558 * 1) Check that the zone isn't thought to be full (doesn't have its
1559 * bit set in the zonelist_cache fullzones BITMAP).
1560 * 2) Check that the zones node (obtained from the zonelist_cache
1561 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1562 * Return true (non-zero) if zone is worth looking at further, or
1563 * else return false (zero) if it is not.
1565 * This check -ignores- the distinction between various watermarks,
1566 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1567 * found to be full for any variation of these watermarks, it will
1568 * be considered full for up to one second by all requests, unless
1569 * we are so low on memory on all allowed nodes that we are forced
1570 * into the second scan of the zonelist.
1572 * In the second scan we ignore this zonelist cache and exactly
1573 * apply the watermarks to all zones, even it is slower to do so.
1574 * We are low on memory in the second scan, and should leave no stone
1575 * unturned looking for a free page.
1577 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1578 nodemask_t
*allowednodes
)
1580 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1581 int i
; /* index of *z in zonelist zones */
1582 int n
; /* node that zone *z is on */
1584 zlc
= zonelist
->zlcache_ptr
;
1588 i
= z
- zonelist
->_zonerefs
;
1591 /* This zone is worth trying if it is allowed but not full */
1592 return node_isset(n
, *allowednodes
) && !test_bit(i
, zlc
->fullzones
);
1596 * Given 'z' scanning a zonelist, set the corresponding bit in
1597 * zlc->fullzones, so that subsequent attempts to allocate a page
1598 * from that zone don't waste time re-examining it.
1600 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1602 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1603 int i
; /* index of *z in zonelist zones */
1605 zlc
= zonelist
->zlcache_ptr
;
1609 i
= z
- zonelist
->_zonerefs
;
1611 set_bit(i
, zlc
->fullzones
);
1614 #else /* CONFIG_NUMA */
1616 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1621 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1622 nodemask_t
*allowednodes
)
1627 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1630 #endif /* CONFIG_NUMA */
1633 * get_page_from_freelist goes through the zonelist trying to allocate
1636 static struct page
*
1637 get_page_from_freelist(gfp_t gfp_mask
, nodemask_t
*nodemask
, unsigned int order
,
1638 struct zonelist
*zonelist
, int high_zoneidx
, int alloc_flags
,
1639 struct zone
*preferred_zone
, int migratetype
)
1642 struct page
*page
= NULL
;
1645 nodemask_t
*allowednodes
= NULL
;/* zonelist_cache approximation */
1646 int zlc_active
= 0; /* set if using zonelist_cache */
1647 int did_zlc_setup
= 0; /* just call zlc_setup() one time */
1649 classzone_idx
= zone_idx(preferred_zone
);
1652 * Scan zonelist, looking for a zone with enough free.
1653 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1655 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
1656 high_zoneidx
, nodemask
) {
1657 if (NUMA_BUILD
&& zlc_active
&&
1658 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
1660 if ((alloc_flags
& ALLOC_CPUSET
) &&
1661 !cpuset_zone_allowed_softwall(zone
, gfp_mask
))
1664 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
1665 if (!(alloc_flags
& ALLOC_NO_WATERMARKS
)) {
1669 mark
= zone
->watermark
[alloc_flags
& ALLOC_WMARK_MASK
];
1670 if (zone_watermark_ok(zone
, order
, mark
,
1671 classzone_idx
, alloc_flags
))
1674 if (zone_reclaim_mode
== 0)
1675 goto this_zone_full
;
1677 ret
= zone_reclaim(zone
, gfp_mask
, order
);
1679 case ZONE_RECLAIM_NOSCAN
:
1682 case ZONE_RECLAIM_FULL
:
1683 /* scanned but unreclaimable */
1684 goto this_zone_full
;
1686 /* did we reclaim enough */
1687 if (!zone_watermark_ok(zone
, order
, mark
,
1688 classzone_idx
, alloc_flags
))
1689 goto this_zone_full
;
1694 page
= buffered_rmqueue(preferred_zone
, zone
, order
,
1695 gfp_mask
, migratetype
);
1700 zlc_mark_zone_full(zonelist
, z
);
1702 if (NUMA_BUILD
&& !did_zlc_setup
&& nr_online_nodes
> 1) {
1704 * we do zlc_setup after the first zone is tried but only
1705 * if there are multiple nodes make it worthwhile
1707 allowednodes
= zlc_setup(zonelist
, alloc_flags
);
1713 if (unlikely(NUMA_BUILD
&& page
== NULL
&& zlc_active
)) {
1714 /* Disable zlc cache for second zonelist scan */
1722 * Large machines with many possible nodes should not always dump per-node
1723 * meminfo in irq context.
1725 static inline bool should_suppress_show_mem(void)
1730 ret
= in_interrupt();
1736 should_alloc_retry(gfp_t gfp_mask
, unsigned int order
,
1737 unsigned long pages_reclaimed
)
1739 /* Do not loop if specifically requested */
1740 if (gfp_mask
& __GFP_NORETRY
)
1744 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
1745 * means __GFP_NOFAIL, but that may not be true in other
1748 if (order
<= PAGE_ALLOC_COSTLY_ORDER
)
1752 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
1753 * specified, then we retry until we no longer reclaim any pages
1754 * (above), or we've reclaimed an order of pages at least as
1755 * large as the allocation's order. In both cases, if the
1756 * allocation still fails, we stop retrying.
1758 if (gfp_mask
& __GFP_REPEAT
&& pages_reclaimed
< (1 << order
))
1762 * Don't let big-order allocations loop unless the caller
1763 * explicitly requests that.
1765 if (gfp_mask
& __GFP_NOFAIL
)
1771 static inline struct page
*
1772 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
1773 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1774 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
1779 /* Acquire the OOM killer lock for the zones in zonelist */
1780 if (!try_set_zonelist_oom(zonelist
, gfp_mask
)) {
1781 schedule_timeout_uninterruptible(1);
1786 * Go through the zonelist yet one more time, keep very high watermark
1787 * here, this is only to catch a parallel oom killing, we must fail if
1788 * we're still under heavy pressure.
1790 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
,
1791 order
, zonelist
, high_zoneidx
,
1792 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
,
1793 preferred_zone
, migratetype
);
1797 if (!(gfp_mask
& __GFP_NOFAIL
)) {
1798 /* The OOM killer will not help higher order allocs */
1799 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
1801 /* The OOM killer does not needlessly kill tasks for lowmem */
1802 if (high_zoneidx
< ZONE_NORMAL
)
1805 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
1806 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
1807 * The caller should handle page allocation failure by itself if
1808 * it specifies __GFP_THISNODE.
1809 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
1811 if (gfp_mask
& __GFP_THISNODE
)
1814 /* Exhausted what can be done so it's blamo time */
1815 out_of_memory(zonelist
, gfp_mask
, order
, nodemask
);
1818 clear_zonelist_oom(zonelist
, gfp_mask
);
1822 #ifdef CONFIG_COMPACTION
1823 /* Try memory compaction for high-order allocations before reclaim */
1824 static struct page
*
1825 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
1826 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1827 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
1828 int migratetype
, unsigned long *did_some_progress
,
1829 bool sync_migration
)
1833 if (!order
|| compaction_deferred(preferred_zone
))
1836 current
->flags
|= PF_MEMALLOC
;
1837 *did_some_progress
= try_to_compact_pages(zonelist
, order
, gfp_mask
,
1838 nodemask
, sync_migration
);
1839 current
->flags
&= ~PF_MEMALLOC
;
1840 if (*did_some_progress
!= COMPACT_SKIPPED
) {
1842 /* Page migration frees to the PCP lists but we want merging */
1843 drain_pages(get_cpu());
1846 page
= get_page_from_freelist(gfp_mask
, nodemask
,
1847 order
, zonelist
, high_zoneidx
,
1848 alloc_flags
, preferred_zone
,
1851 preferred_zone
->compact_considered
= 0;
1852 preferred_zone
->compact_defer_shift
= 0;
1853 count_vm_event(COMPACTSUCCESS
);
1858 * It's bad if compaction run occurs and fails.
1859 * The most likely reason is that pages exist,
1860 * but not enough to satisfy watermarks.
1862 count_vm_event(COMPACTFAIL
);
1863 defer_compaction(preferred_zone
);
1871 static inline struct page
*
1872 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
1873 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1874 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
1875 int migratetype
, unsigned long *did_some_progress
,
1876 bool sync_migration
)
1880 #endif /* CONFIG_COMPACTION */
1882 /* The really slow allocator path where we enter direct reclaim */
1883 static inline struct page
*
1884 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
1885 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1886 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
1887 int migratetype
, unsigned long *did_some_progress
)
1889 struct page
*page
= NULL
;
1890 struct reclaim_state reclaim_state
;
1891 bool drained
= false;
1895 /* We now go into synchronous reclaim */
1896 cpuset_memory_pressure_bump();
1897 current
->flags
|= PF_MEMALLOC
;
1898 lockdep_set_current_reclaim_state(gfp_mask
);
1899 reclaim_state
.reclaimed_slab
= 0;
1900 current
->reclaim_state
= &reclaim_state
;
1902 *did_some_progress
= try_to_free_pages(zonelist
, order
, gfp_mask
, nodemask
);
1904 current
->reclaim_state
= NULL
;
1905 lockdep_clear_current_reclaim_state();
1906 current
->flags
&= ~PF_MEMALLOC
;
1910 if (unlikely(!(*did_some_progress
)))
1914 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
1915 zonelist
, high_zoneidx
,
1916 alloc_flags
, preferred_zone
,
1920 * If an allocation failed after direct reclaim, it could be because
1921 * pages are pinned on the per-cpu lists. Drain them and try again
1923 if (!page
&& !drained
) {
1933 * This is called in the allocator slow-path if the allocation request is of
1934 * sufficient urgency to ignore watermarks and take other desperate measures
1936 static inline struct page
*
1937 __alloc_pages_high_priority(gfp_t gfp_mask
, unsigned int order
,
1938 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1939 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
1945 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
1946 zonelist
, high_zoneidx
, ALLOC_NO_WATERMARKS
,
1947 preferred_zone
, migratetype
);
1949 if (!page
&& gfp_mask
& __GFP_NOFAIL
)
1950 wait_iff_congested(preferred_zone
, BLK_RW_ASYNC
, HZ
/50);
1951 } while (!page
&& (gfp_mask
& __GFP_NOFAIL
));
1957 void wake_all_kswapd(unsigned int order
, struct zonelist
*zonelist
,
1958 enum zone_type high_zoneidx
,
1959 enum zone_type classzone_idx
)
1964 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
)
1965 wakeup_kswapd(zone
, order
, classzone_idx
);
1969 gfp_to_alloc_flags(gfp_t gfp_mask
)
1971 int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
1972 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
1974 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
1975 BUILD_BUG_ON(__GFP_HIGH
!= (__force gfp_t
) ALLOC_HIGH
);
1978 * The caller may dip into page reserves a bit more if the caller
1979 * cannot run direct reclaim, or if the caller has realtime scheduling
1980 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
1981 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
1983 alloc_flags
|= (__force
int) (gfp_mask
& __GFP_HIGH
);
1987 * Not worth trying to allocate harder for
1988 * __GFP_NOMEMALLOC even if it can't schedule.
1990 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
1991 alloc_flags
|= ALLOC_HARDER
;
1993 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
1994 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1996 alloc_flags
&= ~ALLOC_CPUSET
;
1997 } else if (unlikely(rt_task(current
)) && !in_interrupt())
1998 alloc_flags
|= ALLOC_HARDER
;
2000 if (likely(!(gfp_mask
& __GFP_NOMEMALLOC
))) {
2001 if (!in_interrupt() &&
2002 ((current
->flags
& PF_MEMALLOC
) ||
2003 unlikely(test_thread_flag(TIF_MEMDIE
))))
2004 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2010 static inline struct page
*
2011 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
2012 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2013 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
2016 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
2017 struct page
*page
= NULL
;
2019 unsigned long pages_reclaimed
= 0;
2020 unsigned long did_some_progress
;
2021 bool sync_migration
= false;
2024 * In the slowpath, we sanity check order to avoid ever trying to
2025 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2026 * be using allocators in order of preference for an area that is
2029 if (order
>= MAX_ORDER
) {
2030 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
2035 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2036 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2037 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2038 * using a larger set of nodes after it has established that the
2039 * allowed per node queues are empty and that nodes are
2042 if (NUMA_BUILD
&& (gfp_mask
& GFP_THISNODE
) == GFP_THISNODE
)
2046 if (!(gfp_mask
& __GFP_NO_KSWAPD
))
2047 wake_all_kswapd(order
, zonelist
, high_zoneidx
,
2048 zone_idx(preferred_zone
));
2051 * OK, we're below the kswapd watermark and have kicked background
2052 * reclaim. Now things get more complex, so set up alloc_flags according
2053 * to how we want to proceed.
2055 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
2058 * Find the true preferred zone if the allocation is unconstrained by
2061 if (!(alloc_flags
& ALLOC_CPUSET
) && !nodemask
)
2062 first_zones_zonelist(zonelist
, high_zoneidx
, NULL
,
2065 /* This is the last chance, in general, before the goto nopage. */
2066 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
, zonelist
,
2067 high_zoneidx
, alloc_flags
& ~ALLOC_NO_WATERMARKS
,
2068 preferred_zone
, migratetype
);
2073 /* Allocate without watermarks if the context allows */
2074 if (alloc_flags
& ALLOC_NO_WATERMARKS
) {
2075 page
= __alloc_pages_high_priority(gfp_mask
, order
,
2076 zonelist
, high_zoneidx
, nodemask
,
2077 preferred_zone
, migratetype
);
2082 /* Atomic allocations - we can't balance anything */
2086 /* Avoid recursion of direct reclaim */
2087 if (current
->flags
& PF_MEMALLOC
)
2090 /* Avoid allocations with no watermarks from looping endlessly */
2091 if (test_thread_flag(TIF_MEMDIE
) && !(gfp_mask
& __GFP_NOFAIL
))
2095 * Try direct compaction. The first pass is asynchronous. Subsequent
2096 * attempts after direct reclaim are synchronous
2098 page
= __alloc_pages_direct_compact(gfp_mask
, order
,
2099 zonelist
, high_zoneidx
,
2101 alloc_flags
, preferred_zone
,
2102 migratetype
, &did_some_progress
,
2106 sync_migration
= !(gfp_mask
& __GFP_NO_KSWAPD
);
2108 /* Try direct reclaim and then allocating */
2109 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
,
2110 zonelist
, high_zoneidx
,
2112 alloc_flags
, preferred_zone
,
2113 migratetype
, &did_some_progress
);
2118 * If we failed to make any progress reclaiming, then we are
2119 * running out of options and have to consider going OOM
2121 if (!did_some_progress
) {
2122 if ((gfp_mask
& __GFP_FS
) && !(gfp_mask
& __GFP_NORETRY
)) {
2123 if (oom_killer_disabled
)
2125 page
= __alloc_pages_may_oom(gfp_mask
, order
,
2126 zonelist
, high_zoneidx
,
2127 nodemask
, preferred_zone
,
2132 if (!(gfp_mask
& __GFP_NOFAIL
)) {
2134 * The oom killer is not called for high-order
2135 * allocations that may fail, so if no progress
2136 * is being made, there are no other options and
2137 * retrying is unlikely to help.
2139 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
2142 * The oom killer is not called for lowmem
2143 * allocations to prevent needlessly killing
2146 if (high_zoneidx
< ZONE_NORMAL
)
2154 /* Check if we should retry the allocation */
2155 pages_reclaimed
+= did_some_progress
;
2156 if (should_alloc_retry(gfp_mask
, order
, pages_reclaimed
)) {
2157 /* Wait for some write requests to complete then retry */
2158 wait_iff_congested(preferred_zone
, BLK_RW_ASYNC
, HZ
/50);
2162 * High-order allocations do not necessarily loop after
2163 * direct reclaim and reclaim/compaction depends on compaction
2164 * being called after reclaim so call directly if necessary
2166 page
= __alloc_pages_direct_compact(gfp_mask
, order
,
2167 zonelist
, high_zoneidx
,
2169 alloc_flags
, preferred_zone
,
2170 migratetype
, &did_some_progress
,
2177 if (!(gfp_mask
& __GFP_NOWARN
) && printk_ratelimit()) {
2178 unsigned int filter
= SHOW_MEM_FILTER_NODES
;
2181 * This documents exceptions given to allocations in certain
2182 * contexts that are allowed to allocate outside current's set
2185 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
2186 if (test_thread_flag(TIF_MEMDIE
) ||
2187 (current
->flags
& (PF_MEMALLOC
| PF_EXITING
)))
2188 filter
&= ~SHOW_MEM_FILTER_NODES
;
2189 if (in_interrupt() || !wait
)
2190 filter
&= ~SHOW_MEM_FILTER_NODES
;
2192 pr_warning("%s: page allocation failure. order:%d, mode:0x%x\n",
2193 current
->comm
, order
, gfp_mask
);
2195 if (!should_suppress_show_mem())
2200 if (kmemcheck_enabled
)
2201 kmemcheck_pagealloc_alloc(page
, order
, gfp_mask
);
2207 * This is the 'heart' of the zoned buddy allocator.
2210 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
,
2211 struct zonelist
*zonelist
, nodemask_t
*nodemask
)
2213 enum zone_type high_zoneidx
= gfp_zone(gfp_mask
);
2214 struct zone
*preferred_zone
;
2216 int migratetype
= allocflags_to_migratetype(gfp_mask
);
2218 gfp_mask
&= gfp_allowed_mask
;
2220 lockdep_trace_alloc(gfp_mask
);
2222 might_sleep_if(gfp_mask
& __GFP_WAIT
);
2224 if (should_fail_alloc_page(gfp_mask
, order
))
2228 * Check the zones suitable for the gfp_mask contain at least one
2229 * valid zone. It's possible to have an empty zonelist as a result
2230 * of GFP_THISNODE and a memoryless node
2232 if (unlikely(!zonelist
->_zonerefs
->zone
))
2236 /* The preferred zone is used for statistics later */
2237 first_zones_zonelist(zonelist
, high_zoneidx
,
2238 nodemask
? : &cpuset_current_mems_allowed
,
2240 if (!preferred_zone
) {
2245 /* First allocation attempt */
2246 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
, order
,
2247 zonelist
, high_zoneidx
, ALLOC_WMARK_LOW
|ALLOC_CPUSET
,
2248 preferred_zone
, migratetype
);
2249 if (unlikely(!page
))
2250 page
= __alloc_pages_slowpath(gfp_mask
, order
,
2251 zonelist
, high_zoneidx
, nodemask
,
2252 preferred_zone
, migratetype
);
2255 trace_mm_page_alloc(page
, order
, gfp_mask
, migratetype
);
2258 EXPORT_SYMBOL(__alloc_pages_nodemask
);
2261 * Common helper functions.
2263 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
2268 * __get_free_pages() returns a 32-bit address, which cannot represent
2271 VM_BUG_ON((gfp_mask
& __GFP_HIGHMEM
) != 0);
2273 page
= alloc_pages(gfp_mask
, order
);
2276 return (unsigned long) page_address(page
);
2278 EXPORT_SYMBOL(__get_free_pages
);
2280 unsigned long get_zeroed_page(gfp_t gfp_mask
)
2282 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
2284 EXPORT_SYMBOL(get_zeroed_page
);
2286 void __pagevec_free(struct pagevec
*pvec
)
2288 int i
= pagevec_count(pvec
);
2291 trace_mm_pagevec_free(pvec
->pages
[i
], pvec
->cold
);
2292 free_hot_cold_page(pvec
->pages
[i
], pvec
->cold
);
2296 void __free_pages(struct page
*page
, unsigned int order
)
2298 if (put_page_testzero(page
)) {
2300 free_hot_cold_page(page
, 0);
2302 __free_pages_ok(page
, order
);
2306 EXPORT_SYMBOL(__free_pages
);
2308 void free_pages(unsigned long addr
, unsigned int order
)
2311 VM_BUG_ON(!virt_addr_valid((void *)addr
));
2312 __free_pages(virt_to_page((void *)addr
), order
);
2316 EXPORT_SYMBOL(free_pages
);
2319 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2320 * @size: the number of bytes to allocate
2321 * @gfp_mask: GFP flags for the allocation
2323 * This function is similar to alloc_pages(), except that it allocates the
2324 * minimum number of pages to satisfy the request. alloc_pages() can only
2325 * allocate memory in power-of-two pages.
2327 * This function is also limited by MAX_ORDER.
2329 * Memory allocated by this function must be released by free_pages_exact().
2331 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
2333 unsigned int order
= get_order(size
);
2336 addr
= __get_free_pages(gfp_mask
, order
);
2338 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
2339 unsigned long used
= addr
+ PAGE_ALIGN(size
);
2341 split_page(virt_to_page((void *)addr
), order
);
2342 while (used
< alloc_end
) {
2348 return (void *)addr
;
2350 EXPORT_SYMBOL(alloc_pages_exact
);
2353 * free_pages_exact - release memory allocated via alloc_pages_exact()
2354 * @virt: the value returned by alloc_pages_exact.
2355 * @size: size of allocation, same value as passed to alloc_pages_exact().
2357 * Release the memory allocated by a previous call to alloc_pages_exact.
2359 void free_pages_exact(void *virt
, size_t size
)
2361 unsigned long addr
= (unsigned long)virt
;
2362 unsigned long end
= addr
+ PAGE_ALIGN(size
);
2364 while (addr
< end
) {
2369 EXPORT_SYMBOL(free_pages_exact
);
2371 static unsigned int nr_free_zone_pages(int offset
)
2376 /* Just pick one node, since fallback list is circular */
2377 unsigned int sum
= 0;
2379 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
2381 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
2382 unsigned long size
= zone
->present_pages
;
2383 unsigned long high
= high_wmark_pages(zone
);
2392 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2394 unsigned int nr_free_buffer_pages(void)
2396 return nr_free_zone_pages(gfp_zone(GFP_USER
));
2398 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
2401 * Amount of free RAM allocatable within all zones
2403 unsigned int nr_free_pagecache_pages(void)
2405 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
2408 static inline void show_node(struct zone
*zone
)
2411 printk("Node %d ", zone_to_nid(zone
));
2414 void si_meminfo(struct sysinfo
*val
)
2416 val
->totalram
= totalram_pages
;
2418 val
->freeram
= global_page_state(NR_FREE_PAGES
);
2419 val
->bufferram
= nr_blockdev_pages();
2420 val
->totalhigh
= totalhigh_pages
;
2421 val
->freehigh
= nr_free_highpages();
2422 val
->mem_unit
= PAGE_SIZE
;
2425 EXPORT_SYMBOL(si_meminfo
);
2428 void si_meminfo_node(struct sysinfo
*val
, int nid
)
2430 pg_data_t
*pgdat
= NODE_DATA(nid
);
2432 val
->totalram
= pgdat
->node_present_pages
;
2433 val
->freeram
= node_page_state(nid
, NR_FREE_PAGES
);
2434 #ifdef CONFIG_HIGHMEM
2435 val
->totalhigh
= pgdat
->node_zones
[ZONE_HIGHMEM
].present_pages
;
2436 val
->freehigh
= zone_page_state(&pgdat
->node_zones
[ZONE_HIGHMEM
],
2442 val
->mem_unit
= PAGE_SIZE
;
2447 * Determine whether the zone's node should be displayed or not, depending on
2448 * whether SHOW_MEM_FILTER_NODES was passed to __show_free_areas().
2450 static bool skip_free_areas_zone(unsigned int flags
, const struct zone
*zone
)
2454 if (!(flags
& SHOW_MEM_FILTER_NODES
))
2458 ret
= !node_isset(zone
->zone_pgdat
->node_id
,
2459 cpuset_current_mems_allowed
);
2465 #define K(x) ((x) << (PAGE_SHIFT-10))
2468 * Show free area list (used inside shift_scroll-lock stuff)
2469 * We also calculate the percentage fragmentation. We do this by counting the
2470 * memory on each free list with the exception of the first item on the list.
2471 * Suppresses nodes that are not allowed by current's cpuset if
2472 * SHOW_MEM_FILTER_NODES is passed.
2474 void __show_free_areas(unsigned int filter
)
2479 for_each_populated_zone(zone
) {
2480 if (skip_free_areas_zone(filter
, zone
))
2483 printk("%s per-cpu:\n", zone
->name
);
2485 for_each_online_cpu(cpu
) {
2486 struct per_cpu_pageset
*pageset
;
2488 pageset
= per_cpu_ptr(zone
->pageset
, cpu
);
2490 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2491 cpu
, pageset
->pcp
.high
,
2492 pageset
->pcp
.batch
, pageset
->pcp
.count
);
2496 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2497 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2499 " dirty:%lu writeback:%lu unstable:%lu\n"
2500 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2501 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n",
2502 global_page_state(NR_ACTIVE_ANON
),
2503 global_page_state(NR_INACTIVE_ANON
),
2504 global_page_state(NR_ISOLATED_ANON
),
2505 global_page_state(NR_ACTIVE_FILE
),
2506 global_page_state(NR_INACTIVE_FILE
),
2507 global_page_state(NR_ISOLATED_FILE
),
2508 global_page_state(NR_UNEVICTABLE
),
2509 global_page_state(NR_FILE_DIRTY
),
2510 global_page_state(NR_WRITEBACK
),
2511 global_page_state(NR_UNSTABLE_NFS
),
2512 global_page_state(NR_FREE_PAGES
),
2513 global_page_state(NR_SLAB_RECLAIMABLE
),
2514 global_page_state(NR_SLAB_UNRECLAIMABLE
),
2515 global_page_state(NR_FILE_MAPPED
),
2516 global_page_state(NR_SHMEM
),
2517 global_page_state(NR_PAGETABLE
),
2518 global_page_state(NR_BOUNCE
));
2520 for_each_populated_zone(zone
) {
2523 if (skip_free_areas_zone(filter
, zone
))
2531 " active_anon:%lukB"
2532 " inactive_anon:%lukB"
2533 " active_file:%lukB"
2534 " inactive_file:%lukB"
2535 " unevictable:%lukB"
2536 " isolated(anon):%lukB"
2537 " isolated(file):%lukB"
2544 " slab_reclaimable:%lukB"
2545 " slab_unreclaimable:%lukB"
2546 " kernel_stack:%lukB"
2550 " writeback_tmp:%lukB"
2551 " pages_scanned:%lu"
2552 " all_unreclaimable? %s"
2555 K(zone_page_state(zone
, NR_FREE_PAGES
)),
2556 K(min_wmark_pages(zone
)),
2557 K(low_wmark_pages(zone
)),
2558 K(high_wmark_pages(zone
)),
2559 K(zone_page_state(zone
, NR_ACTIVE_ANON
)),
2560 K(zone_page_state(zone
, NR_INACTIVE_ANON
)),
2561 K(zone_page_state(zone
, NR_ACTIVE_FILE
)),
2562 K(zone_page_state(zone
, NR_INACTIVE_FILE
)),
2563 K(zone_page_state(zone
, NR_UNEVICTABLE
)),
2564 K(zone_page_state(zone
, NR_ISOLATED_ANON
)),
2565 K(zone_page_state(zone
, NR_ISOLATED_FILE
)),
2566 K(zone
->present_pages
),
2567 K(zone_page_state(zone
, NR_MLOCK
)),
2568 K(zone_page_state(zone
, NR_FILE_DIRTY
)),
2569 K(zone_page_state(zone
, NR_WRITEBACK
)),
2570 K(zone_page_state(zone
, NR_FILE_MAPPED
)),
2571 K(zone_page_state(zone
, NR_SHMEM
)),
2572 K(zone_page_state(zone
, NR_SLAB_RECLAIMABLE
)),
2573 K(zone_page_state(zone
, NR_SLAB_UNRECLAIMABLE
)),
2574 zone_page_state(zone
, NR_KERNEL_STACK
) *
2576 K(zone_page_state(zone
, NR_PAGETABLE
)),
2577 K(zone_page_state(zone
, NR_UNSTABLE_NFS
)),
2578 K(zone_page_state(zone
, NR_BOUNCE
)),
2579 K(zone_page_state(zone
, NR_WRITEBACK_TEMP
)),
2580 zone
->pages_scanned
,
2581 (zone
->all_unreclaimable
? "yes" : "no")
2583 printk("lowmem_reserve[]:");
2584 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
2585 printk(" %lu", zone
->lowmem_reserve
[i
]);
2589 for_each_populated_zone(zone
) {
2590 unsigned long nr
[MAX_ORDER
], flags
, order
, total
= 0;
2592 if (skip_free_areas_zone(filter
, zone
))
2595 printk("%s: ", zone
->name
);
2597 spin_lock_irqsave(&zone
->lock
, flags
);
2598 for (order
= 0; order
< MAX_ORDER
; order
++) {
2599 nr
[order
] = zone
->free_area
[order
].nr_free
;
2600 total
+= nr
[order
] << order
;
2602 spin_unlock_irqrestore(&zone
->lock
, flags
);
2603 for (order
= 0; order
< MAX_ORDER
; order
++)
2604 printk("%lu*%lukB ", nr
[order
], K(1UL) << order
);
2605 printk("= %lukB\n", K(total
));
2608 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES
));
2610 show_swap_cache_info();
2613 void show_free_areas(void)
2615 __show_free_areas(0);
2618 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
2620 zoneref
->zone
= zone
;
2621 zoneref
->zone_idx
= zone_idx(zone
);
2625 * Builds allocation fallback zone lists.
2627 * Add all populated zones of a node to the zonelist.
2629 static int build_zonelists_node(pg_data_t
*pgdat
, struct zonelist
*zonelist
,
2630 int nr_zones
, enum zone_type zone_type
)
2634 BUG_ON(zone_type
>= MAX_NR_ZONES
);
2639 zone
= pgdat
->node_zones
+ zone_type
;
2640 if (populated_zone(zone
)) {
2641 zoneref_set_zone(zone
,
2642 &zonelist
->_zonerefs
[nr_zones
++]);
2643 check_highest_zone(zone_type
);
2646 } while (zone_type
);
2653 * 0 = automatic detection of better ordering.
2654 * 1 = order by ([node] distance, -zonetype)
2655 * 2 = order by (-zonetype, [node] distance)
2657 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
2658 * the same zonelist. So only NUMA can configure this param.
2660 #define ZONELIST_ORDER_DEFAULT 0
2661 #define ZONELIST_ORDER_NODE 1
2662 #define ZONELIST_ORDER_ZONE 2
2664 /* zonelist order in the kernel.
2665 * set_zonelist_order() will set this to NODE or ZONE.
2667 static int current_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
2668 static char zonelist_order_name
[3][8] = {"Default", "Node", "Zone"};
2672 /* The value user specified ....changed by config */
2673 static int user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
2674 /* string for sysctl */
2675 #define NUMA_ZONELIST_ORDER_LEN 16
2676 char numa_zonelist_order
[16] = "default";
2679 * interface for configure zonelist ordering.
2680 * command line option "numa_zonelist_order"
2681 * = "[dD]efault - default, automatic configuration.
2682 * = "[nN]ode - order by node locality, then by zone within node
2683 * = "[zZ]one - order by zone, then by locality within zone
2686 static int __parse_numa_zonelist_order(char *s
)
2688 if (*s
== 'd' || *s
== 'D') {
2689 user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
2690 } else if (*s
== 'n' || *s
== 'N') {
2691 user_zonelist_order
= ZONELIST_ORDER_NODE
;
2692 } else if (*s
== 'z' || *s
== 'Z') {
2693 user_zonelist_order
= ZONELIST_ORDER_ZONE
;
2696 "Ignoring invalid numa_zonelist_order value: "
2703 static __init
int setup_numa_zonelist_order(char *s
)
2710 ret
= __parse_numa_zonelist_order(s
);
2712 strlcpy(numa_zonelist_order
, s
, NUMA_ZONELIST_ORDER_LEN
);
2716 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
2719 * sysctl handler for numa_zonelist_order
2721 int numa_zonelist_order_handler(ctl_table
*table
, int write
,
2722 void __user
*buffer
, size_t *length
,
2725 char saved_string
[NUMA_ZONELIST_ORDER_LEN
];
2727 static DEFINE_MUTEX(zl_order_mutex
);
2729 mutex_lock(&zl_order_mutex
);
2731 strcpy(saved_string
, (char*)table
->data
);
2732 ret
= proc_dostring(table
, write
, buffer
, length
, ppos
);
2736 int oldval
= user_zonelist_order
;
2737 if (__parse_numa_zonelist_order((char*)table
->data
)) {
2739 * bogus value. restore saved string
2741 strncpy((char*)table
->data
, saved_string
,
2742 NUMA_ZONELIST_ORDER_LEN
);
2743 user_zonelist_order
= oldval
;
2744 } else if (oldval
!= user_zonelist_order
) {
2745 mutex_lock(&zonelists_mutex
);
2746 build_all_zonelists(NULL
);
2747 mutex_unlock(&zonelists_mutex
);
2751 mutex_unlock(&zl_order_mutex
);
2756 #define MAX_NODE_LOAD (nr_online_nodes)
2757 static int node_load
[MAX_NUMNODES
];
2760 * find_next_best_node - find the next node that should appear in a given node's fallback list
2761 * @node: node whose fallback list we're appending
2762 * @used_node_mask: nodemask_t of already used nodes
2764 * We use a number of factors to determine which is the next node that should
2765 * appear on a given node's fallback list. The node should not have appeared
2766 * already in @node's fallback list, and it should be the next closest node
2767 * according to the distance array (which contains arbitrary distance values
2768 * from each node to each node in the system), and should also prefer nodes
2769 * with no CPUs, since presumably they'll have very little allocation pressure
2770 * on them otherwise.
2771 * It returns -1 if no node is found.
2773 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
2776 int min_val
= INT_MAX
;
2778 const struct cpumask
*tmp
= cpumask_of_node(0);
2780 /* Use the local node if we haven't already */
2781 if (!node_isset(node
, *used_node_mask
)) {
2782 node_set(node
, *used_node_mask
);
2786 for_each_node_state(n
, N_HIGH_MEMORY
) {
2788 /* Don't want a node to appear more than once */
2789 if (node_isset(n
, *used_node_mask
))
2792 /* Use the distance array to find the distance */
2793 val
= node_distance(node
, n
);
2795 /* Penalize nodes under us ("prefer the next node") */
2798 /* Give preference to headless and unused nodes */
2799 tmp
= cpumask_of_node(n
);
2800 if (!cpumask_empty(tmp
))
2801 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
2803 /* Slight preference for less loaded node */
2804 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
2805 val
+= node_load
[n
];
2807 if (val
< min_val
) {
2814 node_set(best_node
, *used_node_mask
);
2821 * Build zonelists ordered by node and zones within node.
2822 * This results in maximum locality--normal zone overflows into local
2823 * DMA zone, if any--but risks exhausting DMA zone.
2825 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int node
)
2828 struct zonelist
*zonelist
;
2830 zonelist
= &pgdat
->node_zonelists
[0];
2831 for (j
= 0; zonelist
->_zonerefs
[j
].zone
!= NULL
; j
++)
2833 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
2835 zonelist
->_zonerefs
[j
].zone
= NULL
;
2836 zonelist
->_zonerefs
[j
].zone_idx
= 0;
2840 * Build gfp_thisnode zonelists
2842 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
2845 struct zonelist
*zonelist
;
2847 zonelist
= &pgdat
->node_zonelists
[1];
2848 j
= build_zonelists_node(pgdat
, zonelist
, 0, MAX_NR_ZONES
- 1);
2849 zonelist
->_zonerefs
[j
].zone
= NULL
;
2850 zonelist
->_zonerefs
[j
].zone_idx
= 0;
2854 * Build zonelists ordered by zone and nodes within zones.
2855 * This results in conserving DMA zone[s] until all Normal memory is
2856 * exhausted, but results in overflowing to remote node while memory
2857 * may still exist in local DMA zone.
2859 static int node_order
[MAX_NUMNODES
];
2861 static void build_zonelists_in_zone_order(pg_data_t
*pgdat
, int nr_nodes
)
2864 int zone_type
; /* needs to be signed */
2866 struct zonelist
*zonelist
;
2868 zonelist
= &pgdat
->node_zonelists
[0];
2870 for (zone_type
= MAX_NR_ZONES
- 1; zone_type
>= 0; zone_type
--) {
2871 for (j
= 0; j
< nr_nodes
; j
++) {
2872 node
= node_order
[j
];
2873 z
= &NODE_DATA(node
)->node_zones
[zone_type
];
2874 if (populated_zone(z
)) {
2876 &zonelist
->_zonerefs
[pos
++]);
2877 check_highest_zone(zone_type
);
2881 zonelist
->_zonerefs
[pos
].zone
= NULL
;
2882 zonelist
->_zonerefs
[pos
].zone_idx
= 0;
2885 static int default_zonelist_order(void)
2888 unsigned long low_kmem_size
,total_size
;
2892 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
2893 * If they are really small and used heavily, the system can fall
2894 * into OOM very easily.
2895 * This function detect ZONE_DMA/DMA32 size and configures zone order.
2897 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
2900 for_each_online_node(nid
) {
2901 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
2902 z
= &NODE_DATA(nid
)->node_zones
[zone_type
];
2903 if (populated_zone(z
)) {
2904 if (zone_type
< ZONE_NORMAL
)
2905 low_kmem_size
+= z
->present_pages
;
2906 total_size
+= z
->present_pages
;
2907 } else if (zone_type
== ZONE_NORMAL
) {
2909 * If any node has only lowmem, then node order
2910 * is preferred to allow kernel allocations
2911 * locally; otherwise, they can easily infringe
2912 * on other nodes when there is an abundance of
2913 * lowmem available to allocate from.
2915 return ZONELIST_ORDER_NODE
;
2919 if (!low_kmem_size
|| /* there are no DMA area. */
2920 low_kmem_size
> total_size
/2) /* DMA/DMA32 is big. */
2921 return ZONELIST_ORDER_NODE
;
2923 * look into each node's config.
2924 * If there is a node whose DMA/DMA32 memory is very big area on
2925 * local memory, NODE_ORDER may be suitable.
2927 average_size
= total_size
/
2928 (nodes_weight(node_states
[N_HIGH_MEMORY
]) + 1);
2929 for_each_online_node(nid
) {
2932 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
2933 z
= &NODE_DATA(nid
)->node_zones
[zone_type
];
2934 if (populated_zone(z
)) {
2935 if (zone_type
< ZONE_NORMAL
)
2936 low_kmem_size
+= z
->present_pages
;
2937 total_size
+= z
->present_pages
;
2940 if (low_kmem_size
&&
2941 total_size
> average_size
&& /* ignore small node */
2942 low_kmem_size
> total_size
* 70/100)
2943 return ZONELIST_ORDER_NODE
;
2945 return ZONELIST_ORDER_ZONE
;
2948 static void set_zonelist_order(void)
2950 if (user_zonelist_order
== ZONELIST_ORDER_DEFAULT
)
2951 current_zonelist_order
= default_zonelist_order();
2953 current_zonelist_order
= user_zonelist_order
;
2956 static void build_zonelists(pg_data_t
*pgdat
)
2960 nodemask_t used_mask
;
2961 int local_node
, prev_node
;
2962 struct zonelist
*zonelist
;
2963 int order
= current_zonelist_order
;
2965 /* initialize zonelists */
2966 for (i
= 0; i
< MAX_ZONELISTS
; i
++) {
2967 zonelist
= pgdat
->node_zonelists
+ i
;
2968 zonelist
->_zonerefs
[0].zone
= NULL
;
2969 zonelist
->_zonerefs
[0].zone_idx
= 0;
2972 /* NUMA-aware ordering of nodes */
2973 local_node
= pgdat
->node_id
;
2974 load
= nr_online_nodes
;
2975 prev_node
= local_node
;
2976 nodes_clear(used_mask
);
2978 memset(node_order
, 0, sizeof(node_order
));
2981 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
2982 int distance
= node_distance(local_node
, node
);
2985 * If another node is sufficiently far away then it is better
2986 * to reclaim pages in a zone before going off node.
2988 if (distance
> RECLAIM_DISTANCE
)
2989 zone_reclaim_mode
= 1;
2992 * We don't want to pressure a particular node.
2993 * So adding penalty to the first node in same
2994 * distance group to make it round-robin.
2996 if (distance
!= node_distance(local_node
, prev_node
))
2997 node_load
[node
] = load
;
3001 if (order
== ZONELIST_ORDER_NODE
)
3002 build_zonelists_in_node_order(pgdat
, node
);
3004 node_order
[j
++] = node
; /* remember order */
3007 if (order
== ZONELIST_ORDER_ZONE
) {
3008 /* calculate node order -- i.e., DMA last! */
3009 build_zonelists_in_zone_order(pgdat
, j
);
3012 build_thisnode_zonelists(pgdat
);
3015 /* Construct the zonelist performance cache - see further mmzone.h */
3016 static void build_zonelist_cache(pg_data_t
*pgdat
)
3018 struct zonelist
*zonelist
;
3019 struct zonelist_cache
*zlc
;
3022 zonelist
= &pgdat
->node_zonelists
[0];
3023 zonelist
->zlcache_ptr
= zlc
= &zonelist
->zlcache
;
3024 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
3025 for (z
= zonelist
->_zonerefs
; z
->zone
; z
++)
3026 zlc
->z_to_n
[z
- zonelist
->_zonerefs
] = zonelist_node_idx(z
);
3029 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3031 * Return node id of node used for "local" allocations.
3032 * I.e., first node id of first zone in arg node's generic zonelist.
3033 * Used for initializing percpu 'numa_mem', which is used primarily
3034 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3036 int local_memory_node(int node
)
3040 (void)first_zones_zonelist(node_zonelist(node
, GFP_KERNEL
),
3041 gfp_zone(GFP_KERNEL
),
3048 #else /* CONFIG_NUMA */
3050 static void set_zonelist_order(void)
3052 current_zonelist_order
= ZONELIST_ORDER_ZONE
;
3055 static void build_zonelists(pg_data_t
*pgdat
)
3057 int node
, local_node
;
3059 struct zonelist
*zonelist
;
3061 local_node
= pgdat
->node_id
;
3063 zonelist
= &pgdat
->node_zonelists
[0];
3064 j
= build_zonelists_node(pgdat
, zonelist
, 0, MAX_NR_ZONES
- 1);
3067 * Now we build the zonelist so that it contains the zones
3068 * of all the other nodes.
3069 * We don't want to pressure a particular node, so when
3070 * building the zones for node N, we make sure that the
3071 * zones coming right after the local ones are those from
3072 * node N+1 (modulo N)
3074 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
3075 if (!node_online(node
))
3077 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
3080 for (node
= 0; node
< local_node
; node
++) {
3081 if (!node_online(node
))
3083 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
3087 zonelist
->_zonerefs
[j
].zone
= NULL
;
3088 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3091 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3092 static void build_zonelist_cache(pg_data_t
*pgdat
)
3094 pgdat
->node_zonelists
[0].zlcache_ptr
= NULL
;
3097 #endif /* CONFIG_NUMA */
3100 * Boot pageset table. One per cpu which is going to be used for all
3101 * zones and all nodes. The parameters will be set in such a way
3102 * that an item put on a list will immediately be handed over to
3103 * the buddy list. This is safe since pageset manipulation is done
3104 * with interrupts disabled.
3106 * The boot_pagesets must be kept even after bootup is complete for
3107 * unused processors and/or zones. They do play a role for bootstrapping
3108 * hotplugged processors.
3110 * zoneinfo_show() and maybe other functions do
3111 * not check if the processor is online before following the pageset pointer.
3112 * Other parts of the kernel may not check if the zone is available.
3114 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
);
3115 static DEFINE_PER_CPU(struct per_cpu_pageset
, boot_pageset
);
3116 static void setup_zone_pageset(struct zone
*zone
);
3119 * Global mutex to protect against size modification of zonelists
3120 * as well as to serialize pageset setup for the new populated zone.
3122 DEFINE_MUTEX(zonelists_mutex
);
3124 /* return values int ....just for stop_machine() */
3125 static __init_refok
int __build_all_zonelists(void *data
)
3131 memset(node_load
, 0, sizeof(node_load
));
3133 for_each_online_node(nid
) {
3134 pg_data_t
*pgdat
= NODE_DATA(nid
);
3136 build_zonelists(pgdat
);
3137 build_zonelist_cache(pgdat
);
3141 * Initialize the boot_pagesets that are going to be used
3142 * for bootstrapping processors. The real pagesets for
3143 * each zone will be allocated later when the per cpu
3144 * allocator is available.
3146 * boot_pagesets are used also for bootstrapping offline
3147 * cpus if the system is already booted because the pagesets
3148 * are needed to initialize allocators on a specific cpu too.
3149 * F.e. the percpu allocator needs the page allocator which
3150 * needs the percpu allocator in order to allocate its pagesets
3151 * (a chicken-egg dilemma).
3153 for_each_possible_cpu(cpu
) {
3154 setup_pageset(&per_cpu(boot_pageset
, cpu
), 0);
3156 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3158 * We now know the "local memory node" for each node--
3159 * i.e., the node of the first zone in the generic zonelist.
3160 * Set up numa_mem percpu variable for on-line cpus. During
3161 * boot, only the boot cpu should be on-line; we'll init the
3162 * secondary cpus' numa_mem as they come on-line. During
3163 * node/memory hotplug, we'll fixup all on-line cpus.
3165 if (cpu_online(cpu
))
3166 set_cpu_numa_mem(cpu
, local_memory_node(cpu_to_node(cpu
)));
3174 * Called with zonelists_mutex held always
3175 * unless system_state == SYSTEM_BOOTING.
3177 void build_all_zonelists(void *data
)
3179 set_zonelist_order();
3181 if (system_state
== SYSTEM_BOOTING
) {
3182 __build_all_zonelists(NULL
);
3183 mminit_verify_zonelist();
3184 cpuset_init_current_mems_allowed();
3186 /* we have to stop all cpus to guarantee there is no user
3188 #ifdef CONFIG_MEMORY_HOTPLUG
3190 setup_zone_pageset((struct zone
*)data
);
3192 stop_machine(__build_all_zonelists
, NULL
, NULL
);
3193 /* cpuset refresh routine should be here */
3195 vm_total_pages
= nr_free_pagecache_pages();
3197 * Disable grouping by mobility if the number of pages in the
3198 * system is too low to allow the mechanism to work. It would be
3199 * more accurate, but expensive to check per-zone. This check is
3200 * made on memory-hotadd so a system can start with mobility
3201 * disabled and enable it later
3203 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
3204 page_group_by_mobility_disabled
= 1;
3206 page_group_by_mobility_disabled
= 0;
3208 printk("Built %i zonelists in %s order, mobility grouping %s. "
3209 "Total pages: %ld\n",
3211 zonelist_order_name
[current_zonelist_order
],
3212 page_group_by_mobility_disabled
? "off" : "on",
3215 printk("Policy zone: %s\n", zone_names
[policy_zone
]);
3220 * Helper functions to size the waitqueue hash table.
3221 * Essentially these want to choose hash table sizes sufficiently
3222 * large so that collisions trying to wait on pages are rare.
3223 * But in fact, the number of active page waitqueues on typical
3224 * systems is ridiculously low, less than 200. So this is even
3225 * conservative, even though it seems large.
3227 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3228 * waitqueues, i.e. the size of the waitq table given the number of pages.
3230 #define PAGES_PER_WAITQUEUE 256
3232 #ifndef CONFIG_MEMORY_HOTPLUG
3233 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
3235 unsigned long size
= 1;
3237 pages
/= PAGES_PER_WAITQUEUE
;
3239 while (size
< pages
)
3243 * Once we have dozens or even hundreds of threads sleeping
3244 * on IO we've got bigger problems than wait queue collision.
3245 * Limit the size of the wait table to a reasonable size.
3247 size
= min(size
, 4096UL);
3249 return max(size
, 4UL);
3253 * A zone's size might be changed by hot-add, so it is not possible to determine
3254 * a suitable size for its wait_table. So we use the maximum size now.
3256 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3258 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3259 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3260 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3262 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3263 * or more by the traditional way. (See above). It equals:
3265 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3266 * ia64(16K page size) : = ( 8G + 4M)byte.
3267 * powerpc (64K page size) : = (32G +16M)byte.
3269 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
3276 * This is an integer logarithm so that shifts can be used later
3277 * to extract the more random high bits from the multiplicative
3278 * hash function before the remainder is taken.
3280 static inline unsigned long wait_table_bits(unsigned long size
)
3285 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3288 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3289 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3290 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3291 * higher will lead to a bigger reserve which will get freed as contiguous
3292 * blocks as reclaim kicks in
3294 static void setup_zone_migrate_reserve(struct zone
*zone
)
3296 unsigned long start_pfn
, pfn
, end_pfn
;
3298 unsigned long block_migratetype
;
3301 /* Get the start pfn, end pfn and the number of blocks to reserve */
3302 start_pfn
= zone
->zone_start_pfn
;
3303 end_pfn
= start_pfn
+ zone
->spanned_pages
;
3304 reserve
= roundup(min_wmark_pages(zone
), pageblock_nr_pages
) >>
3308 * Reserve blocks are generally in place to help high-order atomic
3309 * allocations that are short-lived. A min_free_kbytes value that
3310 * would result in more than 2 reserve blocks for atomic allocations
3311 * is assumed to be in place to help anti-fragmentation for the
3312 * future allocation of hugepages at runtime.
3314 reserve
= min(2, reserve
);
3316 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= pageblock_nr_pages
) {
3317 if (!pfn_valid(pfn
))
3319 page
= pfn_to_page(pfn
);
3321 /* Watch out for overlapping nodes */
3322 if (page_to_nid(page
) != zone_to_nid(zone
))
3325 /* Blocks with reserved pages will never free, skip them. */
3326 if (PageReserved(page
))
3329 block_migratetype
= get_pageblock_migratetype(page
);
3331 /* If this block is reserved, account for it */
3332 if (reserve
> 0 && block_migratetype
== MIGRATE_RESERVE
) {
3337 /* Suitable for reserving if this block is movable */
3338 if (reserve
> 0 && block_migratetype
== MIGRATE_MOVABLE
) {
3339 set_pageblock_migratetype(page
, MIGRATE_RESERVE
);
3340 move_freepages_block(zone
, page
, MIGRATE_RESERVE
);
3346 * If the reserve is met and this is a previous reserved block,
3349 if (block_migratetype
== MIGRATE_RESERVE
) {
3350 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
3351 move_freepages_block(zone
, page
, MIGRATE_MOVABLE
);
3357 * Initially all pages are reserved - free ones are freed
3358 * up by free_all_bootmem() once the early boot process is
3359 * done. Non-atomic initialization, single-pass.
3361 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
3362 unsigned long start_pfn
, enum memmap_context context
)
3365 unsigned long end_pfn
= start_pfn
+ size
;
3369 if (highest_memmap_pfn
< end_pfn
- 1)
3370 highest_memmap_pfn
= end_pfn
- 1;
3372 z
= &NODE_DATA(nid
)->node_zones
[zone
];
3373 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
3375 * There can be holes in boot-time mem_map[]s
3376 * handed to this function. They do not
3377 * exist on hotplugged memory.
3379 if (context
== MEMMAP_EARLY
) {
3380 if (!early_pfn_valid(pfn
))
3382 if (!early_pfn_in_nid(pfn
, nid
))
3385 page
= pfn_to_page(pfn
);
3386 set_page_links(page
, zone
, nid
, pfn
);
3387 mminit_verify_page_links(page
, zone
, nid
, pfn
);
3388 init_page_count(page
);
3389 reset_page_mapcount(page
);
3390 SetPageReserved(page
);
3392 * Mark the block movable so that blocks are reserved for
3393 * movable at startup. This will force kernel allocations
3394 * to reserve their blocks rather than leaking throughout
3395 * the address space during boot when many long-lived
3396 * kernel allocations are made. Later some blocks near
3397 * the start are marked MIGRATE_RESERVE by
3398 * setup_zone_migrate_reserve()
3400 * bitmap is created for zone's valid pfn range. but memmap
3401 * can be created for invalid pages (for alignment)
3402 * check here not to call set_pageblock_migratetype() against
3405 if ((z
->zone_start_pfn
<= pfn
)
3406 && (pfn
< z
->zone_start_pfn
+ z
->spanned_pages
)
3407 && !(pfn
& (pageblock_nr_pages
- 1)))
3408 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
3410 INIT_LIST_HEAD(&page
->lru
);
3411 #ifdef WANT_PAGE_VIRTUAL
3412 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3413 if (!is_highmem_idx(zone
))
3414 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
3419 static void __meminit
zone_init_free_lists(struct zone
*zone
)
3422 for_each_migratetype_order(order
, t
) {
3423 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
3424 zone
->free_area
[order
].nr_free
= 0;
3428 #ifndef __HAVE_ARCH_MEMMAP_INIT
3429 #define memmap_init(size, nid, zone, start_pfn) \
3430 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3433 static int zone_batchsize(struct zone
*zone
)
3439 * The per-cpu-pages pools are set to around 1000th of the
3440 * size of the zone. But no more than 1/2 of a meg.
3442 * OK, so we don't know how big the cache is. So guess.
3444 batch
= zone
->present_pages
/ 1024;
3445 if (batch
* PAGE_SIZE
> 512 * 1024)
3446 batch
= (512 * 1024) / PAGE_SIZE
;
3447 batch
/= 4; /* We effectively *= 4 below */
3452 * Clamp the batch to a 2^n - 1 value. Having a power
3453 * of 2 value was found to be more likely to have
3454 * suboptimal cache aliasing properties in some cases.
3456 * For example if 2 tasks are alternately allocating
3457 * batches of pages, one task can end up with a lot
3458 * of pages of one half of the possible page colors
3459 * and the other with pages of the other colors.
3461 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
3466 /* The deferral and batching of frees should be suppressed under NOMMU
3469 * The problem is that NOMMU needs to be able to allocate large chunks
3470 * of contiguous memory as there's no hardware page translation to
3471 * assemble apparent contiguous memory from discontiguous pages.
3473 * Queueing large contiguous runs of pages for batching, however,
3474 * causes the pages to actually be freed in smaller chunks. As there
3475 * can be a significant delay between the individual batches being
3476 * recycled, this leads to the once large chunks of space being
3477 * fragmented and becoming unavailable for high-order allocations.
3483 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
3485 struct per_cpu_pages
*pcp
;
3488 memset(p
, 0, sizeof(*p
));
3492 pcp
->high
= 6 * batch
;
3493 pcp
->batch
= max(1UL, 1 * batch
);
3494 for (migratetype
= 0; migratetype
< MIGRATE_PCPTYPES
; migratetype
++)
3495 INIT_LIST_HEAD(&pcp
->lists
[migratetype
]);
3499 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3500 * to the value high for the pageset p.
3503 static void setup_pagelist_highmark(struct per_cpu_pageset
*p
,
3506 struct per_cpu_pages
*pcp
;
3510 pcp
->batch
= max(1UL, high
/4);
3511 if ((high
/4) > (PAGE_SHIFT
* 8))
3512 pcp
->batch
= PAGE_SHIFT
* 8;
3515 static __meminit
void setup_zone_pageset(struct zone
*zone
)
3519 zone
->pageset
= alloc_percpu(struct per_cpu_pageset
);
3521 for_each_possible_cpu(cpu
) {
3522 struct per_cpu_pageset
*pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
3524 setup_pageset(pcp
, zone_batchsize(zone
));
3526 if (percpu_pagelist_fraction
)
3527 setup_pagelist_highmark(pcp
,
3528 (zone
->present_pages
/
3529 percpu_pagelist_fraction
));
3534 * Allocate per cpu pagesets and initialize them.
3535 * Before this call only boot pagesets were available.
3537 void __init
setup_per_cpu_pageset(void)
3541 for_each_populated_zone(zone
)
3542 setup_zone_pageset(zone
);
3545 static noinline __init_refok
3546 int zone_wait_table_init(struct zone
*zone
, unsigned long zone_size_pages
)
3549 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
3553 * The per-page waitqueue mechanism uses hashed waitqueues
3556 zone
->wait_table_hash_nr_entries
=
3557 wait_table_hash_nr_entries(zone_size_pages
);
3558 zone
->wait_table_bits
=
3559 wait_table_bits(zone
->wait_table_hash_nr_entries
);
3560 alloc_size
= zone
->wait_table_hash_nr_entries
3561 * sizeof(wait_queue_head_t
);
3563 if (!slab_is_available()) {
3564 zone
->wait_table
= (wait_queue_head_t
*)
3565 alloc_bootmem_node(pgdat
, alloc_size
);
3568 * This case means that a zone whose size was 0 gets new memory
3569 * via memory hot-add.
3570 * But it may be the case that a new node was hot-added. In
3571 * this case vmalloc() will not be able to use this new node's
3572 * memory - this wait_table must be initialized to use this new
3573 * node itself as well.
3574 * To use this new node's memory, further consideration will be
3577 zone
->wait_table
= vmalloc(alloc_size
);
3579 if (!zone
->wait_table
)
3582 for(i
= 0; i
< zone
->wait_table_hash_nr_entries
; ++i
)
3583 init_waitqueue_head(zone
->wait_table
+ i
);
3588 static int __zone_pcp_update(void *data
)
3590 struct zone
*zone
= data
;
3592 unsigned long batch
= zone_batchsize(zone
), flags
;
3594 for_each_possible_cpu(cpu
) {
3595 struct per_cpu_pageset
*pset
;
3596 struct per_cpu_pages
*pcp
;
3598 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
3601 local_irq_save(flags
);
3602 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
3603 setup_pageset(pset
, batch
);
3604 local_irq_restore(flags
);
3609 void zone_pcp_update(struct zone
*zone
)
3611 stop_machine(__zone_pcp_update
, zone
, NULL
);
3614 static __meminit
void zone_pcp_init(struct zone
*zone
)
3617 * per cpu subsystem is not up at this point. The following code
3618 * relies on the ability of the linker to provide the
3619 * offset of a (static) per cpu variable into the per cpu area.
3621 zone
->pageset
= &boot_pageset
;
3623 if (zone
->present_pages
)
3624 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%u\n",
3625 zone
->name
, zone
->present_pages
,
3626 zone_batchsize(zone
));
3629 __meminit
int init_currently_empty_zone(struct zone
*zone
,
3630 unsigned long zone_start_pfn
,
3632 enum memmap_context context
)
3634 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
3636 ret
= zone_wait_table_init(zone
, size
);
3639 pgdat
->nr_zones
= zone_idx(zone
) + 1;
3641 zone
->zone_start_pfn
= zone_start_pfn
;
3643 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
3644 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
3646 (unsigned long)zone_idx(zone
),
3647 zone_start_pfn
, (zone_start_pfn
+ size
));
3649 zone_init_free_lists(zone
);
3654 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3656 * Basic iterator support. Return the first range of PFNs for a node
3657 * Note: nid == MAX_NUMNODES returns first region regardless of node
3659 static int __meminit
first_active_region_index_in_nid(int nid
)
3663 for (i
= 0; i
< nr_nodemap_entries
; i
++)
3664 if (nid
== MAX_NUMNODES
|| early_node_map
[i
].nid
== nid
)
3671 * Basic iterator support. Return the next active range of PFNs for a node
3672 * Note: nid == MAX_NUMNODES returns next region regardless of node
3674 static int __meminit
next_active_region_index_in_nid(int index
, int nid
)
3676 for (index
= index
+ 1; index
< nr_nodemap_entries
; index
++)
3677 if (nid
== MAX_NUMNODES
|| early_node_map
[index
].nid
== nid
)
3683 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
3685 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
3686 * Architectures may implement their own version but if add_active_range()
3687 * was used and there are no special requirements, this is a convenient
3690 int __meminit
__early_pfn_to_nid(unsigned long pfn
)
3694 for (i
= 0; i
< nr_nodemap_entries
; i
++) {
3695 unsigned long start_pfn
= early_node_map
[i
].start_pfn
;
3696 unsigned long end_pfn
= early_node_map
[i
].end_pfn
;
3698 if (start_pfn
<= pfn
&& pfn
< end_pfn
)
3699 return early_node_map
[i
].nid
;
3701 /* This is a memory hole */
3704 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
3706 int __meminit
early_pfn_to_nid(unsigned long pfn
)
3710 nid
= __early_pfn_to_nid(pfn
);
3713 /* just returns 0 */
3717 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
3718 bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
3722 nid
= __early_pfn_to_nid(pfn
);
3723 if (nid
>= 0 && nid
!= node
)
3729 /* Basic iterator support to walk early_node_map[] */
3730 #define for_each_active_range_index_in_nid(i, nid) \
3731 for (i = first_active_region_index_in_nid(nid); i != -1; \
3732 i = next_active_region_index_in_nid(i, nid))
3735 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
3736 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
3737 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
3739 * If an architecture guarantees that all ranges registered with
3740 * add_active_ranges() contain no holes and may be freed, this
3741 * this function may be used instead of calling free_bootmem() manually.
3743 void __init
free_bootmem_with_active_regions(int nid
,
3744 unsigned long max_low_pfn
)
3748 for_each_active_range_index_in_nid(i
, nid
) {
3749 unsigned long size_pages
= 0;
3750 unsigned long end_pfn
= early_node_map
[i
].end_pfn
;
3752 if (early_node_map
[i
].start_pfn
>= max_low_pfn
)
3755 if (end_pfn
> max_low_pfn
)
3756 end_pfn
= max_low_pfn
;
3758 size_pages
= end_pfn
- early_node_map
[i
].start_pfn
;
3759 free_bootmem_node(NODE_DATA(early_node_map
[i
].nid
),
3760 PFN_PHYS(early_node_map
[i
].start_pfn
),
3761 size_pages
<< PAGE_SHIFT
);
3765 #ifdef CONFIG_HAVE_MEMBLOCK
3767 * Basic iterator support. Return the last range of PFNs for a node
3768 * Note: nid == MAX_NUMNODES returns last region regardless of node
3770 static int __meminit
last_active_region_index_in_nid(int nid
)
3774 for (i
= nr_nodemap_entries
- 1; i
>= 0; i
--)
3775 if (nid
== MAX_NUMNODES
|| early_node_map
[i
].nid
== nid
)
3782 * Basic iterator support. Return the previous active range of PFNs for a node
3783 * Note: nid == MAX_NUMNODES returns next region regardless of node
3785 static int __meminit
previous_active_region_index_in_nid(int index
, int nid
)
3787 for (index
= index
- 1; index
>= 0; index
--)
3788 if (nid
== MAX_NUMNODES
|| early_node_map
[index
].nid
== nid
)
3794 #define for_each_active_range_index_in_nid_reverse(i, nid) \
3795 for (i = last_active_region_index_in_nid(nid); i != -1; \
3796 i = previous_active_region_index_in_nid(i, nid))
3798 u64 __init
find_memory_core_early(int nid
, u64 size
, u64 align
,
3799 u64 goal
, u64 limit
)
3803 /* Need to go over early_node_map to find out good range for node */
3804 for_each_active_range_index_in_nid_reverse(i
, nid
) {
3806 u64 ei_start
, ei_last
;
3807 u64 final_start
, final_end
;
3809 ei_last
= early_node_map
[i
].end_pfn
;
3810 ei_last
<<= PAGE_SHIFT
;
3811 ei_start
= early_node_map
[i
].start_pfn
;
3812 ei_start
<<= PAGE_SHIFT
;
3814 final_start
= max(ei_start
, goal
);
3815 final_end
= min(ei_last
, limit
);
3817 if (final_start
>= final_end
)
3820 addr
= memblock_find_in_range(final_start
, final_end
, size
, align
);
3822 if (addr
== MEMBLOCK_ERROR
)
3828 return MEMBLOCK_ERROR
;
3832 int __init
add_from_early_node_map(struct range
*range
, int az
,
3833 int nr_range
, int nid
)
3838 /* need to go over early_node_map to find out good range for node */
3839 for_each_active_range_index_in_nid(i
, nid
) {
3840 start
= early_node_map
[i
].start_pfn
;
3841 end
= early_node_map
[i
].end_pfn
;
3842 nr_range
= add_range(range
, az
, nr_range
, start
, end
);
3847 void __init
work_with_active_regions(int nid
, work_fn_t work_fn
, void *data
)
3852 for_each_active_range_index_in_nid(i
, nid
) {
3853 ret
= work_fn(early_node_map
[i
].start_pfn
,
3854 early_node_map
[i
].end_pfn
, data
);
3860 * sparse_memory_present_with_active_regions - Call memory_present for each active range
3861 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
3863 * If an architecture guarantees that all ranges registered with
3864 * add_active_ranges() contain no holes and may be freed, this
3865 * function may be used instead of calling memory_present() manually.
3867 void __init
sparse_memory_present_with_active_regions(int nid
)
3871 for_each_active_range_index_in_nid(i
, nid
)
3872 memory_present(early_node_map
[i
].nid
,
3873 early_node_map
[i
].start_pfn
,
3874 early_node_map
[i
].end_pfn
);
3878 * get_pfn_range_for_nid - Return the start and end page frames for a node
3879 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
3880 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
3881 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
3883 * It returns the start and end page frame of a node based on information
3884 * provided by an arch calling add_active_range(). If called for a node
3885 * with no available memory, a warning is printed and the start and end
3888 void __meminit
get_pfn_range_for_nid(unsigned int nid
,
3889 unsigned long *start_pfn
, unsigned long *end_pfn
)
3895 for_each_active_range_index_in_nid(i
, nid
) {
3896 *start_pfn
= min(*start_pfn
, early_node_map
[i
].start_pfn
);
3897 *end_pfn
= max(*end_pfn
, early_node_map
[i
].end_pfn
);
3900 if (*start_pfn
== -1UL)
3905 * This finds a zone that can be used for ZONE_MOVABLE pages. The
3906 * assumption is made that zones within a node are ordered in monotonic
3907 * increasing memory addresses so that the "highest" populated zone is used
3909 static void __init
find_usable_zone_for_movable(void)
3912 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
3913 if (zone_index
== ZONE_MOVABLE
)
3916 if (arch_zone_highest_possible_pfn
[zone_index
] >
3917 arch_zone_lowest_possible_pfn
[zone_index
])
3921 VM_BUG_ON(zone_index
== -1);
3922 movable_zone
= zone_index
;
3926 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
3927 * because it is sized independant of architecture. Unlike the other zones,
3928 * the starting point for ZONE_MOVABLE is not fixed. It may be different
3929 * in each node depending on the size of each node and how evenly kernelcore
3930 * is distributed. This helper function adjusts the zone ranges
3931 * provided by the architecture for a given node by using the end of the
3932 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
3933 * zones within a node are in order of monotonic increases memory addresses
3935 static void __meminit
adjust_zone_range_for_zone_movable(int nid
,
3936 unsigned long zone_type
,
3937 unsigned long node_start_pfn
,
3938 unsigned long node_end_pfn
,
3939 unsigned long *zone_start_pfn
,
3940 unsigned long *zone_end_pfn
)
3942 /* Only adjust if ZONE_MOVABLE is on this node */
3943 if (zone_movable_pfn
[nid
]) {
3944 /* Size ZONE_MOVABLE */
3945 if (zone_type
== ZONE_MOVABLE
) {
3946 *zone_start_pfn
= zone_movable_pfn
[nid
];
3947 *zone_end_pfn
= min(node_end_pfn
,
3948 arch_zone_highest_possible_pfn
[movable_zone
]);
3950 /* Adjust for ZONE_MOVABLE starting within this range */
3951 } else if (*zone_start_pfn
< zone_movable_pfn
[nid
] &&
3952 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
3953 *zone_end_pfn
= zone_movable_pfn
[nid
];
3955 /* Check if this whole range is within ZONE_MOVABLE */
3956 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
3957 *zone_start_pfn
= *zone_end_pfn
;
3962 * Return the number of pages a zone spans in a node, including holes
3963 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
3965 static unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
3966 unsigned long zone_type
,
3967 unsigned long *ignored
)
3969 unsigned long node_start_pfn
, node_end_pfn
;
3970 unsigned long zone_start_pfn
, zone_end_pfn
;
3972 /* Get the start and end of the node and zone */
3973 get_pfn_range_for_nid(nid
, &node_start_pfn
, &node_end_pfn
);
3974 zone_start_pfn
= arch_zone_lowest_possible_pfn
[zone_type
];
3975 zone_end_pfn
= arch_zone_highest_possible_pfn
[zone_type
];
3976 adjust_zone_range_for_zone_movable(nid
, zone_type
,
3977 node_start_pfn
, node_end_pfn
,
3978 &zone_start_pfn
, &zone_end_pfn
);
3980 /* Check that this node has pages within the zone's required range */
3981 if (zone_end_pfn
< node_start_pfn
|| zone_start_pfn
> node_end_pfn
)
3984 /* Move the zone boundaries inside the node if necessary */
3985 zone_end_pfn
= min(zone_end_pfn
, node_end_pfn
);
3986 zone_start_pfn
= max(zone_start_pfn
, node_start_pfn
);
3988 /* Return the spanned pages */
3989 return zone_end_pfn
- zone_start_pfn
;
3993 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
3994 * then all holes in the requested range will be accounted for.
3996 unsigned long __meminit
__absent_pages_in_range(int nid
,
3997 unsigned long range_start_pfn
,
3998 unsigned long range_end_pfn
)
4001 unsigned long prev_end_pfn
= 0, hole_pages
= 0;
4002 unsigned long start_pfn
;
4004 /* Find the end_pfn of the first active range of pfns in the node */
4005 i
= first_active_region_index_in_nid(nid
);
4009 prev_end_pfn
= min(early_node_map
[i
].start_pfn
, range_end_pfn
);
4011 /* Account for ranges before physical memory on this node */
4012 if (early_node_map
[i
].start_pfn
> range_start_pfn
)
4013 hole_pages
= prev_end_pfn
- range_start_pfn
;
4015 /* Find all holes for the zone within the node */
4016 for (; i
!= -1; i
= next_active_region_index_in_nid(i
, nid
)) {
4018 /* No need to continue if prev_end_pfn is outside the zone */
4019 if (prev_end_pfn
>= range_end_pfn
)
4022 /* Make sure the end of the zone is not within the hole */
4023 start_pfn
= min(early_node_map
[i
].start_pfn
, range_end_pfn
);
4024 prev_end_pfn
= max(prev_end_pfn
, range_start_pfn
);
4026 /* Update the hole size cound and move on */
4027 if (start_pfn
> range_start_pfn
) {
4028 BUG_ON(prev_end_pfn
> start_pfn
);
4029 hole_pages
+= start_pfn
- prev_end_pfn
;
4031 prev_end_pfn
= early_node_map
[i
].end_pfn
;
4034 /* Account for ranges past physical memory on this node */
4035 if (range_end_pfn
> prev_end_pfn
)
4036 hole_pages
+= range_end_pfn
-
4037 max(range_start_pfn
, prev_end_pfn
);
4043 * absent_pages_in_range - Return number of page frames in holes within a range
4044 * @start_pfn: The start PFN to start searching for holes
4045 * @end_pfn: The end PFN to stop searching for holes
4047 * It returns the number of pages frames in memory holes within a range.
4049 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
4050 unsigned long end_pfn
)
4052 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
4055 /* Return the number of page frames in holes in a zone on a node */
4056 static unsigned long __meminit
zone_absent_pages_in_node(int nid
,
4057 unsigned long zone_type
,
4058 unsigned long *ignored
)
4060 unsigned long node_start_pfn
, node_end_pfn
;
4061 unsigned long zone_start_pfn
, zone_end_pfn
;
4063 get_pfn_range_for_nid(nid
, &node_start_pfn
, &node_end_pfn
);
4064 zone_start_pfn
= max(arch_zone_lowest_possible_pfn
[zone_type
],
4066 zone_end_pfn
= min(arch_zone_highest_possible_pfn
[zone_type
],
4069 adjust_zone_range_for_zone_movable(nid
, zone_type
,
4070 node_start_pfn
, node_end_pfn
,
4071 &zone_start_pfn
, &zone_end_pfn
);
4072 return __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
4076 static inline unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
4077 unsigned long zone_type
,
4078 unsigned long *zones_size
)
4080 return zones_size
[zone_type
];
4083 static inline unsigned long __meminit
zone_absent_pages_in_node(int nid
,
4084 unsigned long zone_type
,
4085 unsigned long *zholes_size
)
4090 return zholes_size
[zone_type
];
4095 static void __meminit
calculate_node_totalpages(struct pglist_data
*pgdat
,
4096 unsigned long *zones_size
, unsigned long *zholes_size
)
4098 unsigned long realtotalpages
, totalpages
= 0;
4101 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4102 totalpages
+= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
4104 pgdat
->node_spanned_pages
= totalpages
;
4106 realtotalpages
= totalpages
;
4107 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4109 zone_absent_pages_in_node(pgdat
->node_id
, i
,
4111 pgdat
->node_present_pages
= realtotalpages
;
4112 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
4116 #ifndef CONFIG_SPARSEMEM
4118 * Calculate the size of the zone->blockflags rounded to an unsigned long
4119 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4120 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4121 * round what is now in bits to nearest long in bits, then return it in
4124 static unsigned long __init
usemap_size(unsigned long zonesize
)
4126 unsigned long usemapsize
;
4128 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
4129 usemapsize
= usemapsize
>> pageblock_order
;
4130 usemapsize
*= NR_PAGEBLOCK_BITS
;
4131 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
4133 return usemapsize
/ 8;
4136 static void __init
setup_usemap(struct pglist_data
*pgdat
,
4137 struct zone
*zone
, unsigned long zonesize
)
4139 unsigned long usemapsize
= usemap_size(zonesize
);
4140 zone
->pageblock_flags
= NULL
;
4142 zone
->pageblock_flags
= alloc_bootmem_node(pgdat
, usemapsize
);
4145 static inline void setup_usemap(struct pglist_data
*pgdat
,
4146 struct zone
*zone
, unsigned long zonesize
) {}
4147 #endif /* CONFIG_SPARSEMEM */
4149 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4151 /* Return a sensible default order for the pageblock size. */
4152 static inline int pageblock_default_order(void)
4154 if (HPAGE_SHIFT
> PAGE_SHIFT
)
4155 return HUGETLB_PAGE_ORDER
;
4160 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4161 static inline void __init
set_pageblock_order(unsigned int order
)
4163 /* Check that pageblock_nr_pages has not already been setup */
4164 if (pageblock_order
)
4168 * Assume the largest contiguous order of interest is a huge page.
4169 * This value may be variable depending on boot parameters on IA64
4171 pageblock_order
= order
;
4173 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4176 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4177 * and pageblock_default_order() are unused as pageblock_order is set
4178 * at compile-time. See include/linux/pageblock-flags.h for the values of
4179 * pageblock_order based on the kernel config
4181 static inline int pageblock_default_order(unsigned int order
)
4185 #define set_pageblock_order(x) do {} while (0)
4187 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4190 * Set up the zone data structures:
4191 * - mark all pages reserved
4192 * - mark all memory queues empty
4193 * - clear the memory bitmaps
4195 static void __paginginit
free_area_init_core(struct pglist_data
*pgdat
,
4196 unsigned long *zones_size
, unsigned long *zholes_size
)
4199 int nid
= pgdat
->node_id
;
4200 unsigned long zone_start_pfn
= pgdat
->node_start_pfn
;
4203 pgdat_resize_init(pgdat
);
4204 pgdat
->nr_zones
= 0;
4205 init_waitqueue_head(&pgdat
->kswapd_wait
);
4206 pgdat
->kswapd_max_order
= 0;
4207 pgdat_page_cgroup_init(pgdat
);
4209 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
4210 struct zone
*zone
= pgdat
->node_zones
+ j
;
4211 unsigned long size
, realsize
, memmap_pages
;
4214 size
= zone_spanned_pages_in_node(nid
, j
, zones_size
);
4215 realsize
= size
- zone_absent_pages_in_node(nid
, j
,
4219 * Adjust realsize so that it accounts for how much memory
4220 * is used by this zone for memmap. This affects the watermark
4221 * and per-cpu initialisations
4224 PAGE_ALIGN(size
* sizeof(struct page
)) >> PAGE_SHIFT
;
4225 if (realsize
>= memmap_pages
) {
4226 realsize
-= memmap_pages
;
4229 " %s zone: %lu pages used for memmap\n",
4230 zone_names
[j
], memmap_pages
);
4233 " %s zone: %lu pages exceeds realsize %lu\n",
4234 zone_names
[j
], memmap_pages
, realsize
);
4236 /* Account for reserved pages */
4237 if (j
== 0 && realsize
> dma_reserve
) {
4238 realsize
-= dma_reserve
;
4239 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
4240 zone_names
[0], dma_reserve
);
4243 if (!is_highmem_idx(j
))
4244 nr_kernel_pages
+= realsize
;
4245 nr_all_pages
+= realsize
;
4247 zone
->spanned_pages
= size
;
4248 zone
->present_pages
= realsize
;
4251 zone
->min_unmapped_pages
= (realsize
*sysctl_min_unmapped_ratio
)
4253 zone
->min_slab_pages
= (realsize
* sysctl_min_slab_ratio
) / 100;
4255 zone
->name
= zone_names
[j
];
4256 spin_lock_init(&zone
->lock
);
4257 spin_lock_init(&zone
->lru_lock
);
4258 zone_seqlock_init(zone
);
4259 zone
->zone_pgdat
= pgdat
;
4261 zone_pcp_init(zone
);
4263 INIT_LIST_HEAD(&zone
->lru
[l
].list
);
4264 zone
->reclaim_stat
.nr_saved_scan
[l
] = 0;
4266 zone
->reclaim_stat
.recent_rotated
[0] = 0;
4267 zone
->reclaim_stat
.recent_rotated
[1] = 0;
4268 zone
->reclaim_stat
.recent_scanned
[0] = 0;
4269 zone
->reclaim_stat
.recent_scanned
[1] = 0;
4270 zap_zone_vm_stats(zone
);
4275 set_pageblock_order(pageblock_default_order());
4276 setup_usemap(pgdat
, zone
, size
);
4277 ret
= init_currently_empty_zone(zone
, zone_start_pfn
,
4278 size
, MEMMAP_EARLY
);
4280 memmap_init(size
, nid
, j
, zone_start_pfn
);
4281 zone_start_pfn
+= size
;
4285 static void __init_refok
alloc_node_mem_map(struct pglist_data
*pgdat
)
4287 /* Skip empty nodes */
4288 if (!pgdat
->node_spanned_pages
)
4291 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4292 /* ia64 gets its own node_mem_map, before this, without bootmem */
4293 if (!pgdat
->node_mem_map
) {
4294 unsigned long size
, start
, end
;
4298 * The zone's endpoints aren't required to be MAX_ORDER
4299 * aligned but the node_mem_map endpoints must be in order
4300 * for the buddy allocator to function correctly.
4302 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
4303 end
= pgdat
->node_start_pfn
+ pgdat
->node_spanned_pages
;
4304 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
4305 size
= (end
- start
) * sizeof(struct page
);
4306 map
= alloc_remap(pgdat
->node_id
, size
);
4308 map
= alloc_bootmem_node(pgdat
, size
);
4309 pgdat
->node_mem_map
= map
+ (pgdat
->node_start_pfn
- start
);
4311 #ifndef CONFIG_NEED_MULTIPLE_NODES
4313 * With no DISCONTIG, the global mem_map is just set as node 0's
4315 if (pgdat
== NODE_DATA(0)) {
4316 mem_map
= NODE_DATA(0)->node_mem_map
;
4317 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
4318 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
4319 mem_map
-= (pgdat
->node_start_pfn
- ARCH_PFN_OFFSET
);
4320 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4323 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4326 void __paginginit
free_area_init_node(int nid
, unsigned long *zones_size
,
4327 unsigned long node_start_pfn
, unsigned long *zholes_size
)
4329 pg_data_t
*pgdat
= NODE_DATA(nid
);
4331 pgdat
->node_id
= nid
;
4332 pgdat
->node_start_pfn
= node_start_pfn
;
4333 calculate_node_totalpages(pgdat
, zones_size
, zholes_size
);
4335 alloc_node_mem_map(pgdat
);
4336 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4337 printk(KERN_DEBUG
"free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4338 nid
, (unsigned long)pgdat
,
4339 (unsigned long)pgdat
->node_mem_map
);
4342 free_area_init_core(pgdat
, zones_size
, zholes_size
);
4345 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
4347 #if MAX_NUMNODES > 1
4349 * Figure out the number of possible node ids.
4351 static void __init
setup_nr_node_ids(void)
4354 unsigned int highest
= 0;
4356 for_each_node_mask(node
, node_possible_map
)
4358 nr_node_ids
= highest
+ 1;
4361 static inline void setup_nr_node_ids(void)
4367 * add_active_range - Register a range of PFNs backed by physical memory
4368 * @nid: The node ID the range resides on
4369 * @start_pfn: The start PFN of the available physical memory
4370 * @end_pfn: The end PFN of the available physical memory
4372 * These ranges are stored in an early_node_map[] and later used by
4373 * free_area_init_nodes() to calculate zone sizes and holes. If the
4374 * range spans a memory hole, it is up to the architecture to ensure
4375 * the memory is not freed by the bootmem allocator. If possible
4376 * the range being registered will be merged with existing ranges.
4378 void __init
add_active_range(unsigned int nid
, unsigned long start_pfn
,
4379 unsigned long end_pfn
)
4383 mminit_dprintk(MMINIT_TRACE
, "memory_register",
4384 "Entering add_active_range(%d, %#lx, %#lx) "
4385 "%d entries of %d used\n",
4386 nid
, start_pfn
, end_pfn
,
4387 nr_nodemap_entries
, MAX_ACTIVE_REGIONS
);
4389 mminit_validate_memmodel_limits(&start_pfn
, &end_pfn
);
4391 /* Merge with existing active regions if possible */
4392 for (i
= 0; i
< nr_nodemap_entries
; i
++) {
4393 if (early_node_map
[i
].nid
!= nid
)
4396 /* Skip if an existing region covers this new one */
4397 if (start_pfn
>= early_node_map
[i
].start_pfn
&&
4398 end_pfn
<= early_node_map
[i
].end_pfn
)
4401 /* Merge forward if suitable */
4402 if (start_pfn
<= early_node_map
[i
].end_pfn
&&
4403 end_pfn
> early_node_map
[i
].end_pfn
) {
4404 early_node_map
[i
].end_pfn
= end_pfn
;
4408 /* Merge backward if suitable */
4409 if (start_pfn
< early_node_map
[i
].start_pfn
&&
4410 end_pfn
>= early_node_map
[i
].start_pfn
) {
4411 early_node_map
[i
].start_pfn
= start_pfn
;
4416 /* Check that early_node_map is large enough */
4417 if (i
>= MAX_ACTIVE_REGIONS
) {
4418 printk(KERN_CRIT
"More than %d memory regions, truncating\n",
4419 MAX_ACTIVE_REGIONS
);
4423 early_node_map
[i
].nid
= nid
;
4424 early_node_map
[i
].start_pfn
= start_pfn
;
4425 early_node_map
[i
].end_pfn
= end_pfn
;
4426 nr_nodemap_entries
= i
+ 1;
4430 * remove_active_range - Shrink an existing registered range of PFNs
4431 * @nid: The node id the range is on that should be shrunk
4432 * @start_pfn: The new PFN of the range
4433 * @end_pfn: The new PFN of the range
4435 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
4436 * The map is kept near the end physical page range that has already been
4437 * registered. This function allows an arch to shrink an existing registered
4440 void __init
remove_active_range(unsigned int nid
, unsigned long start_pfn
,
4441 unsigned long end_pfn
)
4446 printk(KERN_DEBUG
"remove_active_range (%d, %lu, %lu)\n",
4447 nid
, start_pfn
, end_pfn
);
4449 /* Find the old active region end and shrink */
4450 for_each_active_range_index_in_nid(i
, nid
) {
4451 if (early_node_map
[i
].start_pfn
>= start_pfn
&&
4452 early_node_map
[i
].end_pfn
<= end_pfn
) {
4454 early_node_map
[i
].start_pfn
= 0;
4455 early_node_map
[i
].end_pfn
= 0;
4459 if (early_node_map
[i
].start_pfn
< start_pfn
&&
4460 early_node_map
[i
].end_pfn
> start_pfn
) {
4461 unsigned long temp_end_pfn
= early_node_map
[i
].end_pfn
;
4462 early_node_map
[i
].end_pfn
= start_pfn
;
4463 if (temp_end_pfn
> end_pfn
)
4464 add_active_range(nid
, end_pfn
, temp_end_pfn
);
4467 if (early_node_map
[i
].start_pfn
>= start_pfn
&&
4468 early_node_map
[i
].end_pfn
> end_pfn
&&
4469 early_node_map
[i
].start_pfn
< end_pfn
) {
4470 early_node_map
[i
].start_pfn
= end_pfn
;
4478 /* remove the blank ones */
4479 for (i
= nr_nodemap_entries
- 1; i
> 0; i
--) {
4480 if (early_node_map
[i
].nid
!= nid
)
4482 if (early_node_map
[i
].end_pfn
)
4484 /* we found it, get rid of it */
4485 for (j
= i
; j
< nr_nodemap_entries
- 1; j
++)
4486 memcpy(&early_node_map
[j
], &early_node_map
[j
+1],
4487 sizeof(early_node_map
[j
]));
4488 j
= nr_nodemap_entries
- 1;
4489 memset(&early_node_map
[j
], 0, sizeof(early_node_map
[j
]));
4490 nr_nodemap_entries
--;
4495 * remove_all_active_ranges - Remove all currently registered regions
4497 * During discovery, it may be found that a table like SRAT is invalid
4498 * and an alternative discovery method must be used. This function removes
4499 * all currently registered regions.
4501 void __init
remove_all_active_ranges(void)
4503 memset(early_node_map
, 0, sizeof(early_node_map
));
4504 nr_nodemap_entries
= 0;
4507 /* Compare two active node_active_regions */
4508 static int __init
cmp_node_active_region(const void *a
, const void *b
)
4510 struct node_active_region
*arange
= (struct node_active_region
*)a
;
4511 struct node_active_region
*brange
= (struct node_active_region
*)b
;
4513 /* Done this way to avoid overflows */
4514 if (arange
->start_pfn
> brange
->start_pfn
)
4516 if (arange
->start_pfn
< brange
->start_pfn
)
4522 /* sort the node_map by start_pfn */
4523 void __init
sort_node_map(void)
4525 sort(early_node_map
, (size_t)nr_nodemap_entries
,
4526 sizeof(struct node_active_region
),
4527 cmp_node_active_region
, NULL
);
4530 /* Find the lowest pfn for a node */
4531 static unsigned long __init
find_min_pfn_for_node(int nid
)
4534 unsigned long min_pfn
= ULONG_MAX
;
4536 /* Assuming a sorted map, the first range found has the starting pfn */
4537 for_each_active_range_index_in_nid(i
, nid
)
4538 min_pfn
= min(min_pfn
, early_node_map
[i
].start_pfn
);
4540 if (min_pfn
== ULONG_MAX
) {
4542 "Could not find start_pfn for node %d\n", nid
);
4550 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4552 * It returns the minimum PFN based on information provided via
4553 * add_active_range().
4555 unsigned long __init
find_min_pfn_with_active_regions(void)
4557 return find_min_pfn_for_node(MAX_NUMNODES
);
4561 * early_calculate_totalpages()
4562 * Sum pages in active regions for movable zone.
4563 * Populate N_HIGH_MEMORY for calculating usable_nodes.
4565 static unsigned long __init
early_calculate_totalpages(void)
4568 unsigned long totalpages
= 0;
4570 for (i
= 0; i
< nr_nodemap_entries
; i
++) {
4571 unsigned long pages
= early_node_map
[i
].end_pfn
-
4572 early_node_map
[i
].start_pfn
;
4573 totalpages
+= pages
;
4575 node_set_state(early_node_map
[i
].nid
, N_HIGH_MEMORY
);
4581 * Find the PFN the Movable zone begins in each node. Kernel memory
4582 * is spread evenly between nodes as long as the nodes have enough
4583 * memory. When they don't, some nodes will have more kernelcore than
4586 static void __init
find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn
)
4589 unsigned long usable_startpfn
;
4590 unsigned long kernelcore_node
, kernelcore_remaining
;
4591 /* save the state before borrow the nodemask */
4592 nodemask_t saved_node_state
= node_states
[N_HIGH_MEMORY
];
4593 unsigned long totalpages
= early_calculate_totalpages();
4594 int usable_nodes
= nodes_weight(node_states
[N_HIGH_MEMORY
]);
4597 * If movablecore was specified, calculate what size of
4598 * kernelcore that corresponds so that memory usable for
4599 * any allocation type is evenly spread. If both kernelcore
4600 * and movablecore are specified, then the value of kernelcore
4601 * will be used for required_kernelcore if it's greater than
4602 * what movablecore would have allowed.
4604 if (required_movablecore
) {
4605 unsigned long corepages
;
4608 * Round-up so that ZONE_MOVABLE is at least as large as what
4609 * was requested by the user
4611 required_movablecore
=
4612 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
4613 corepages
= totalpages
- required_movablecore
;
4615 required_kernelcore
= max(required_kernelcore
, corepages
);
4618 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4619 if (!required_kernelcore
)
4622 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4623 find_usable_zone_for_movable();
4624 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
4627 /* Spread kernelcore memory as evenly as possible throughout nodes */
4628 kernelcore_node
= required_kernelcore
/ usable_nodes
;
4629 for_each_node_state(nid
, N_HIGH_MEMORY
) {
4631 * Recalculate kernelcore_node if the division per node
4632 * now exceeds what is necessary to satisfy the requested
4633 * amount of memory for the kernel
4635 if (required_kernelcore
< kernelcore_node
)
4636 kernelcore_node
= required_kernelcore
/ usable_nodes
;
4639 * As the map is walked, we track how much memory is usable
4640 * by the kernel using kernelcore_remaining. When it is
4641 * 0, the rest of the node is usable by ZONE_MOVABLE
4643 kernelcore_remaining
= kernelcore_node
;
4645 /* Go through each range of PFNs within this node */
4646 for_each_active_range_index_in_nid(i
, nid
) {
4647 unsigned long start_pfn
, end_pfn
;
4648 unsigned long size_pages
;
4650 start_pfn
= max(early_node_map
[i
].start_pfn
,
4651 zone_movable_pfn
[nid
]);
4652 end_pfn
= early_node_map
[i
].end_pfn
;
4653 if (start_pfn
>= end_pfn
)
4656 /* Account for what is only usable for kernelcore */
4657 if (start_pfn
< usable_startpfn
) {
4658 unsigned long kernel_pages
;
4659 kernel_pages
= min(end_pfn
, usable_startpfn
)
4662 kernelcore_remaining
-= min(kernel_pages
,
4663 kernelcore_remaining
);
4664 required_kernelcore
-= min(kernel_pages
,
4665 required_kernelcore
);
4667 /* Continue if range is now fully accounted */
4668 if (end_pfn
<= usable_startpfn
) {
4671 * Push zone_movable_pfn to the end so
4672 * that if we have to rebalance
4673 * kernelcore across nodes, we will
4674 * not double account here
4676 zone_movable_pfn
[nid
] = end_pfn
;
4679 start_pfn
= usable_startpfn
;
4683 * The usable PFN range for ZONE_MOVABLE is from
4684 * start_pfn->end_pfn. Calculate size_pages as the
4685 * number of pages used as kernelcore
4687 size_pages
= end_pfn
- start_pfn
;
4688 if (size_pages
> kernelcore_remaining
)
4689 size_pages
= kernelcore_remaining
;
4690 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
4693 * Some kernelcore has been met, update counts and
4694 * break if the kernelcore for this node has been
4697 required_kernelcore
-= min(required_kernelcore
,
4699 kernelcore_remaining
-= size_pages
;
4700 if (!kernelcore_remaining
)
4706 * If there is still required_kernelcore, we do another pass with one
4707 * less node in the count. This will push zone_movable_pfn[nid] further
4708 * along on the nodes that still have memory until kernelcore is
4712 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
4715 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4716 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
4717 zone_movable_pfn
[nid
] =
4718 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
4721 /* restore the node_state */
4722 node_states
[N_HIGH_MEMORY
] = saved_node_state
;
4725 /* Any regular memory on that node ? */
4726 static void check_for_regular_memory(pg_data_t
*pgdat
)
4728 #ifdef CONFIG_HIGHMEM
4729 enum zone_type zone_type
;
4731 for (zone_type
= 0; zone_type
<= ZONE_NORMAL
; zone_type
++) {
4732 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
4733 if (zone
->present_pages
)
4734 node_set_state(zone_to_nid(zone
), N_NORMAL_MEMORY
);
4740 * free_area_init_nodes - Initialise all pg_data_t and zone data
4741 * @max_zone_pfn: an array of max PFNs for each zone
4743 * This will call free_area_init_node() for each active node in the system.
4744 * Using the page ranges provided by add_active_range(), the size of each
4745 * zone in each node and their holes is calculated. If the maximum PFN
4746 * between two adjacent zones match, it is assumed that the zone is empty.
4747 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4748 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4749 * starts where the previous one ended. For example, ZONE_DMA32 starts
4750 * at arch_max_dma_pfn.
4752 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
4757 /* Sort early_node_map as initialisation assumes it is sorted */
4760 /* Record where the zone boundaries are */
4761 memset(arch_zone_lowest_possible_pfn
, 0,
4762 sizeof(arch_zone_lowest_possible_pfn
));
4763 memset(arch_zone_highest_possible_pfn
, 0,
4764 sizeof(arch_zone_highest_possible_pfn
));
4765 arch_zone_lowest_possible_pfn
[0] = find_min_pfn_with_active_regions();
4766 arch_zone_highest_possible_pfn
[0] = max_zone_pfn
[0];
4767 for (i
= 1; i
< MAX_NR_ZONES
; i
++) {
4768 if (i
== ZONE_MOVABLE
)
4770 arch_zone_lowest_possible_pfn
[i
] =
4771 arch_zone_highest_possible_pfn
[i
-1];
4772 arch_zone_highest_possible_pfn
[i
] =
4773 max(max_zone_pfn
[i
], arch_zone_lowest_possible_pfn
[i
]);
4775 arch_zone_lowest_possible_pfn
[ZONE_MOVABLE
] = 0;
4776 arch_zone_highest_possible_pfn
[ZONE_MOVABLE
] = 0;
4778 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4779 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
4780 find_zone_movable_pfns_for_nodes(zone_movable_pfn
);
4782 /* Print out the zone ranges */
4783 printk("Zone PFN ranges:\n");
4784 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
4785 if (i
== ZONE_MOVABLE
)
4787 printk(" %-8s ", zone_names
[i
]);
4788 if (arch_zone_lowest_possible_pfn
[i
] ==
4789 arch_zone_highest_possible_pfn
[i
])
4792 printk("%0#10lx -> %0#10lx\n",
4793 arch_zone_lowest_possible_pfn
[i
],
4794 arch_zone_highest_possible_pfn
[i
]);
4797 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4798 printk("Movable zone start PFN for each node\n");
4799 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
4800 if (zone_movable_pfn
[i
])
4801 printk(" Node %d: %lu\n", i
, zone_movable_pfn
[i
]);
4804 /* Print out the early_node_map[] */
4805 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries
);
4806 for (i
= 0; i
< nr_nodemap_entries
; i
++)
4807 printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map
[i
].nid
,
4808 early_node_map
[i
].start_pfn
,
4809 early_node_map
[i
].end_pfn
);
4811 /* Initialise every node */
4812 mminit_verify_pageflags_layout();
4813 setup_nr_node_ids();
4814 for_each_online_node(nid
) {
4815 pg_data_t
*pgdat
= NODE_DATA(nid
);
4816 free_area_init_node(nid
, NULL
,
4817 find_min_pfn_for_node(nid
), NULL
);
4819 /* Any memory on that node */
4820 if (pgdat
->node_present_pages
)
4821 node_set_state(nid
, N_HIGH_MEMORY
);
4822 check_for_regular_memory(pgdat
);
4826 static int __init
cmdline_parse_core(char *p
, unsigned long *core
)
4828 unsigned long long coremem
;
4832 coremem
= memparse(p
, &p
);
4833 *core
= coremem
>> PAGE_SHIFT
;
4835 /* Paranoid check that UL is enough for the coremem value */
4836 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
4842 * kernelcore=size sets the amount of memory for use for allocations that
4843 * cannot be reclaimed or migrated.
4845 static int __init
cmdline_parse_kernelcore(char *p
)
4847 return cmdline_parse_core(p
, &required_kernelcore
);
4851 * movablecore=size sets the amount of memory for use for allocations that
4852 * can be reclaimed or migrated.
4854 static int __init
cmdline_parse_movablecore(char *p
)
4856 return cmdline_parse_core(p
, &required_movablecore
);
4859 early_param("kernelcore", cmdline_parse_kernelcore
);
4860 early_param("movablecore", cmdline_parse_movablecore
);
4862 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4865 * set_dma_reserve - set the specified number of pages reserved in the first zone
4866 * @new_dma_reserve: The number of pages to mark reserved
4868 * The per-cpu batchsize and zone watermarks are determined by present_pages.
4869 * In the DMA zone, a significant percentage may be consumed by kernel image
4870 * and other unfreeable allocations which can skew the watermarks badly. This
4871 * function may optionally be used to account for unfreeable pages in the
4872 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
4873 * smaller per-cpu batchsize.
4875 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
4877 dma_reserve
= new_dma_reserve
;
4880 void __init
free_area_init(unsigned long *zones_size
)
4882 free_area_init_node(0, zones_size
,
4883 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
4886 static int page_alloc_cpu_notify(struct notifier_block
*self
,
4887 unsigned long action
, void *hcpu
)
4889 int cpu
= (unsigned long)hcpu
;
4891 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
) {
4895 * Spill the event counters of the dead processor
4896 * into the current processors event counters.
4897 * This artificially elevates the count of the current
4900 vm_events_fold_cpu(cpu
);
4903 * Zero the differential counters of the dead processor
4904 * so that the vm statistics are consistent.
4906 * This is only okay since the processor is dead and cannot
4907 * race with what we are doing.
4909 refresh_cpu_vm_stats(cpu
);
4914 void __init
page_alloc_init(void)
4916 hotcpu_notifier(page_alloc_cpu_notify
, 0);
4920 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
4921 * or min_free_kbytes changes.
4923 static void calculate_totalreserve_pages(void)
4925 struct pglist_data
*pgdat
;
4926 unsigned long reserve_pages
= 0;
4927 enum zone_type i
, j
;
4929 for_each_online_pgdat(pgdat
) {
4930 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
4931 struct zone
*zone
= pgdat
->node_zones
+ i
;
4932 unsigned long max
= 0;
4934 /* Find valid and maximum lowmem_reserve in the zone */
4935 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
4936 if (zone
->lowmem_reserve
[j
] > max
)
4937 max
= zone
->lowmem_reserve
[j
];
4940 /* we treat the high watermark as reserved pages. */
4941 max
+= high_wmark_pages(zone
);
4943 if (max
> zone
->present_pages
)
4944 max
= zone
->present_pages
;
4945 reserve_pages
+= max
;
4948 totalreserve_pages
= reserve_pages
;
4952 * setup_per_zone_lowmem_reserve - called whenever
4953 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
4954 * has a correct pages reserved value, so an adequate number of
4955 * pages are left in the zone after a successful __alloc_pages().
4957 static void setup_per_zone_lowmem_reserve(void)
4959 struct pglist_data
*pgdat
;
4960 enum zone_type j
, idx
;
4962 for_each_online_pgdat(pgdat
) {
4963 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
4964 struct zone
*zone
= pgdat
->node_zones
+ j
;
4965 unsigned long present_pages
= zone
->present_pages
;
4967 zone
->lowmem_reserve
[j
] = 0;
4971 struct zone
*lower_zone
;
4975 if (sysctl_lowmem_reserve_ratio
[idx
] < 1)
4976 sysctl_lowmem_reserve_ratio
[idx
] = 1;
4978 lower_zone
= pgdat
->node_zones
+ idx
;
4979 lower_zone
->lowmem_reserve
[j
] = present_pages
/
4980 sysctl_lowmem_reserve_ratio
[idx
];
4981 present_pages
+= lower_zone
->present_pages
;
4986 /* update totalreserve_pages */
4987 calculate_totalreserve_pages();
4991 * setup_per_zone_wmarks - called when min_free_kbytes changes
4992 * or when memory is hot-{added|removed}
4994 * Ensures that the watermark[min,low,high] values for each zone are set
4995 * correctly with respect to min_free_kbytes.
4997 void setup_per_zone_wmarks(void)
4999 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
5000 unsigned long lowmem_pages
= 0;
5002 unsigned long flags
;
5004 /* Calculate total number of !ZONE_HIGHMEM pages */
5005 for_each_zone(zone
) {
5006 if (!is_highmem(zone
))
5007 lowmem_pages
+= zone
->present_pages
;
5010 for_each_zone(zone
) {
5013 spin_lock_irqsave(&zone
->lock
, flags
);
5014 tmp
= (u64
)pages_min
* zone
->present_pages
;
5015 do_div(tmp
, lowmem_pages
);
5016 if (is_highmem(zone
)) {
5018 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5019 * need highmem pages, so cap pages_min to a small
5022 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5023 * deltas controls asynch page reclaim, and so should
5024 * not be capped for highmem.
5028 min_pages
= zone
->present_pages
/ 1024;
5029 if (min_pages
< SWAP_CLUSTER_MAX
)
5030 min_pages
= SWAP_CLUSTER_MAX
;
5031 if (min_pages
> 128)
5033 zone
->watermark
[WMARK_MIN
] = min_pages
;
5036 * If it's a lowmem zone, reserve a number of pages
5037 * proportionate to the zone's size.
5039 zone
->watermark
[WMARK_MIN
] = tmp
;
5042 zone
->watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + (tmp
>> 2);
5043 zone
->watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + (tmp
>> 1);
5044 setup_zone_migrate_reserve(zone
);
5045 spin_unlock_irqrestore(&zone
->lock
, flags
);
5048 /* update totalreserve_pages */
5049 calculate_totalreserve_pages();
5053 * The inactive anon list should be small enough that the VM never has to
5054 * do too much work, but large enough that each inactive page has a chance
5055 * to be referenced again before it is swapped out.
5057 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5058 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5059 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5060 * the anonymous pages are kept on the inactive list.
5063 * memory ratio inactive anon
5064 * -------------------------------------
5073 void calculate_zone_inactive_ratio(struct zone
*zone
)
5075 unsigned int gb
, ratio
;
5077 /* Zone size in gigabytes */
5078 gb
= zone
->present_pages
>> (30 - PAGE_SHIFT
);
5080 ratio
= int_sqrt(10 * gb
);
5084 zone
->inactive_ratio
= ratio
;
5087 static void __init
setup_per_zone_inactive_ratio(void)
5092 calculate_zone_inactive_ratio(zone
);
5096 * Initialise min_free_kbytes.
5098 * For small machines we want it small (128k min). For large machines
5099 * we want it large (64MB max). But it is not linear, because network
5100 * bandwidth does not increase linearly with machine size. We use
5102 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5103 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5119 static int __init
init_per_zone_wmark_min(void)
5121 unsigned long lowmem_kbytes
;
5123 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
5125 min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
5126 if (min_free_kbytes
< 128)
5127 min_free_kbytes
= 128;
5128 if (min_free_kbytes
> 65536)
5129 min_free_kbytes
= 65536;
5130 setup_per_zone_wmarks();
5131 setup_per_zone_lowmem_reserve();
5132 setup_per_zone_inactive_ratio();
5135 module_init(init_per_zone_wmark_min
)
5138 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5139 * that we can call two helper functions whenever min_free_kbytes
5142 int min_free_kbytes_sysctl_handler(ctl_table
*table
, int write
,
5143 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5145 proc_dointvec(table
, write
, buffer
, length
, ppos
);
5147 setup_per_zone_wmarks();
5152 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table
*table
, int write
,
5153 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5158 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5163 zone
->min_unmapped_pages
= (zone
->present_pages
*
5164 sysctl_min_unmapped_ratio
) / 100;
5168 int sysctl_min_slab_ratio_sysctl_handler(ctl_table
*table
, int write
,
5169 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5174 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5179 zone
->min_slab_pages
= (zone
->present_pages
*
5180 sysctl_min_slab_ratio
) / 100;
5186 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5187 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5188 * whenever sysctl_lowmem_reserve_ratio changes.
5190 * The reserve ratio obviously has absolutely no relation with the
5191 * minimum watermarks. The lowmem reserve ratio can only make sense
5192 * if in function of the boot time zone sizes.
5194 int lowmem_reserve_ratio_sysctl_handler(ctl_table
*table
, int write
,
5195 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5197 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5198 setup_per_zone_lowmem_reserve();
5203 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5204 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
5205 * can have before it gets flushed back to buddy allocator.
5208 int percpu_pagelist_fraction_sysctl_handler(ctl_table
*table
, int write
,
5209 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5215 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5216 if (!write
|| (ret
== -EINVAL
))
5218 for_each_populated_zone(zone
) {
5219 for_each_possible_cpu(cpu
) {
5221 high
= zone
->present_pages
/ percpu_pagelist_fraction
;
5222 setup_pagelist_highmark(
5223 per_cpu_ptr(zone
->pageset
, cpu
), high
);
5229 int hashdist
= HASHDIST_DEFAULT
;
5232 static int __init
set_hashdist(char *str
)
5236 hashdist
= simple_strtoul(str
, &str
, 0);
5239 __setup("hashdist=", set_hashdist
);
5243 * allocate a large system hash table from bootmem
5244 * - it is assumed that the hash table must contain an exact power-of-2
5245 * quantity of entries
5246 * - limit is the number of hash buckets, not the total allocation size
5248 void *__init
alloc_large_system_hash(const char *tablename
,
5249 unsigned long bucketsize
,
5250 unsigned long numentries
,
5253 unsigned int *_hash_shift
,
5254 unsigned int *_hash_mask
,
5255 unsigned long limit
)
5257 unsigned long long max
= limit
;
5258 unsigned long log2qty
, size
;
5261 /* allow the kernel cmdline to have a say */
5263 /* round applicable memory size up to nearest megabyte */
5264 numentries
= nr_kernel_pages
;
5265 numentries
+= (1UL << (20 - PAGE_SHIFT
)) - 1;
5266 numentries
>>= 20 - PAGE_SHIFT
;
5267 numentries
<<= 20 - PAGE_SHIFT
;
5269 /* limit to 1 bucket per 2^scale bytes of low memory */
5270 if (scale
> PAGE_SHIFT
)
5271 numentries
>>= (scale
- PAGE_SHIFT
);
5273 numentries
<<= (PAGE_SHIFT
- scale
);
5275 /* Make sure we've got at least a 0-order allocation.. */
5276 if (unlikely(flags
& HASH_SMALL
)) {
5277 /* Makes no sense without HASH_EARLY */
5278 WARN_ON(!(flags
& HASH_EARLY
));
5279 if (!(numentries
>> *_hash_shift
)) {
5280 numentries
= 1UL << *_hash_shift
;
5281 BUG_ON(!numentries
);
5283 } else if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
5284 numentries
= PAGE_SIZE
/ bucketsize
;
5286 numentries
= roundup_pow_of_two(numentries
);
5288 /* limit allocation size to 1/16 total memory by default */
5290 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
5291 do_div(max
, bucketsize
);
5294 if (numentries
> max
)
5297 log2qty
= ilog2(numentries
);
5300 size
= bucketsize
<< log2qty
;
5301 if (flags
& HASH_EARLY
)
5302 table
= alloc_bootmem_nopanic(size
);
5304 table
= __vmalloc(size
, GFP_ATOMIC
, PAGE_KERNEL
);
5307 * If bucketsize is not a power-of-two, we may free
5308 * some pages at the end of hash table which
5309 * alloc_pages_exact() automatically does
5311 if (get_order(size
) < MAX_ORDER
) {
5312 table
= alloc_pages_exact(size
, GFP_ATOMIC
);
5313 kmemleak_alloc(table
, size
, 1, GFP_ATOMIC
);
5316 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
5319 panic("Failed to allocate %s hash table\n", tablename
);
5321 printk(KERN_INFO
"%s hash table entries: %ld (order: %d, %lu bytes)\n",
5324 ilog2(size
) - PAGE_SHIFT
,
5328 *_hash_shift
= log2qty
;
5330 *_hash_mask
= (1 << log2qty
) - 1;
5335 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5336 static inline unsigned long *get_pageblock_bitmap(struct zone
*zone
,
5339 #ifdef CONFIG_SPARSEMEM
5340 return __pfn_to_section(pfn
)->pageblock_flags
;
5342 return zone
->pageblock_flags
;
5343 #endif /* CONFIG_SPARSEMEM */
5346 static inline int pfn_to_bitidx(struct zone
*zone
, unsigned long pfn
)
5348 #ifdef CONFIG_SPARSEMEM
5349 pfn
&= (PAGES_PER_SECTION
-1);
5350 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
5352 pfn
= pfn
- zone
->zone_start_pfn
;
5353 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
5354 #endif /* CONFIG_SPARSEMEM */
5358 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5359 * @page: The page within the block of interest
5360 * @start_bitidx: The first bit of interest to retrieve
5361 * @end_bitidx: The last bit of interest
5362 * returns pageblock_bits flags
5364 unsigned long get_pageblock_flags_group(struct page
*page
,
5365 int start_bitidx
, int end_bitidx
)
5368 unsigned long *bitmap
;
5369 unsigned long pfn
, bitidx
;
5370 unsigned long flags
= 0;
5371 unsigned long value
= 1;
5373 zone
= page_zone(page
);
5374 pfn
= page_to_pfn(page
);
5375 bitmap
= get_pageblock_bitmap(zone
, pfn
);
5376 bitidx
= pfn_to_bitidx(zone
, pfn
);
5378 for (; start_bitidx
<= end_bitidx
; start_bitidx
++, value
<<= 1)
5379 if (test_bit(bitidx
+ start_bitidx
, bitmap
))
5386 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5387 * @page: The page within the block of interest
5388 * @start_bitidx: The first bit of interest
5389 * @end_bitidx: The last bit of interest
5390 * @flags: The flags to set
5392 void set_pageblock_flags_group(struct page
*page
, unsigned long flags
,
5393 int start_bitidx
, int end_bitidx
)
5396 unsigned long *bitmap
;
5397 unsigned long pfn
, bitidx
;
5398 unsigned long value
= 1;
5400 zone
= page_zone(page
);
5401 pfn
= page_to_pfn(page
);
5402 bitmap
= get_pageblock_bitmap(zone
, pfn
);
5403 bitidx
= pfn_to_bitidx(zone
, pfn
);
5404 VM_BUG_ON(pfn
< zone
->zone_start_pfn
);
5405 VM_BUG_ON(pfn
>= zone
->zone_start_pfn
+ zone
->spanned_pages
);
5407 for (; start_bitidx
<= end_bitidx
; start_bitidx
++, value
<<= 1)
5409 __set_bit(bitidx
+ start_bitidx
, bitmap
);
5411 __clear_bit(bitidx
+ start_bitidx
, bitmap
);
5415 * This is designed as sub function...plz see page_isolation.c also.
5416 * set/clear page block's type to be ISOLATE.
5417 * page allocater never alloc memory from ISOLATE block.
5421 __count_immobile_pages(struct zone
*zone
, struct page
*page
, int count
)
5423 unsigned long pfn
, iter
, found
;
5425 * For avoiding noise data, lru_add_drain_all() should be called
5426 * If ZONE_MOVABLE, the zone never contains immobile pages
5428 if (zone_idx(zone
) == ZONE_MOVABLE
)
5431 if (get_pageblock_migratetype(page
) == MIGRATE_MOVABLE
)
5434 pfn
= page_to_pfn(page
);
5435 for (found
= 0, iter
= 0; iter
< pageblock_nr_pages
; iter
++) {
5436 unsigned long check
= pfn
+ iter
;
5438 if (!pfn_valid_within(check
))
5441 page
= pfn_to_page(check
);
5442 if (!page_count(page
)) {
5443 if (PageBuddy(page
))
5444 iter
+= (1 << page_order(page
)) - 1;
5450 * If there are RECLAIMABLE pages, we need to check it.
5451 * But now, memory offline itself doesn't call shrink_slab()
5452 * and it still to be fixed.
5455 * If the page is not RAM, page_count()should be 0.
5456 * we don't need more check. This is an _used_ not-movable page.
5458 * The problematic thing here is PG_reserved pages. PG_reserved
5459 * is set to both of a memory hole page and a _used_ kernel
5468 bool is_pageblock_removable_nolock(struct page
*page
)
5470 struct zone
*zone
= page_zone(page
);
5471 return __count_immobile_pages(zone
, page
, 0);
5474 int set_migratetype_isolate(struct page
*page
)
5477 unsigned long flags
, pfn
;
5478 struct memory_isolate_notify arg
;
5483 zone
= page_zone(page
);
5484 zone_idx
= zone_idx(zone
);
5486 spin_lock_irqsave(&zone
->lock
, flags
);
5488 pfn
= page_to_pfn(page
);
5489 arg
.start_pfn
= pfn
;
5490 arg
.nr_pages
= pageblock_nr_pages
;
5491 arg
.pages_found
= 0;
5494 * It may be possible to isolate a pageblock even if the
5495 * migratetype is not MIGRATE_MOVABLE. The memory isolation
5496 * notifier chain is used by balloon drivers to return the
5497 * number of pages in a range that are held by the balloon
5498 * driver to shrink memory. If all the pages are accounted for
5499 * by balloons, are free, or on the LRU, isolation can continue.
5500 * Later, for example, when memory hotplug notifier runs, these
5501 * pages reported as "can be isolated" should be isolated(freed)
5502 * by the balloon driver through the memory notifier chain.
5504 notifier_ret
= memory_isolate_notify(MEM_ISOLATE_COUNT
, &arg
);
5505 notifier_ret
= notifier_to_errno(notifier_ret
);
5509 * FIXME: Now, memory hotplug doesn't call shrink_slab() by itself.
5510 * We just check MOVABLE pages.
5512 if (__count_immobile_pages(zone
, page
, arg
.pages_found
))
5516 * immobile means "not-on-lru" paes. If immobile is larger than
5517 * removable-by-driver pages reported by notifier, we'll fail.
5522 set_pageblock_migratetype(page
, MIGRATE_ISOLATE
);
5523 move_freepages_block(zone
, page
, MIGRATE_ISOLATE
);
5526 spin_unlock_irqrestore(&zone
->lock
, flags
);
5532 void unset_migratetype_isolate(struct page
*page
)
5535 unsigned long flags
;
5536 zone
= page_zone(page
);
5537 spin_lock_irqsave(&zone
->lock
, flags
);
5538 if (get_pageblock_migratetype(page
) != MIGRATE_ISOLATE
)
5540 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
5541 move_freepages_block(zone
, page
, MIGRATE_MOVABLE
);
5543 spin_unlock_irqrestore(&zone
->lock
, flags
);
5546 #ifdef CONFIG_MEMORY_HOTREMOVE
5548 * All pages in the range must be isolated before calling this.
5551 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
5557 unsigned long flags
;
5558 /* find the first valid pfn */
5559 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
5564 zone
= page_zone(pfn_to_page(pfn
));
5565 spin_lock_irqsave(&zone
->lock
, flags
);
5567 while (pfn
< end_pfn
) {
5568 if (!pfn_valid(pfn
)) {
5572 page
= pfn_to_page(pfn
);
5573 BUG_ON(page_count(page
));
5574 BUG_ON(!PageBuddy(page
));
5575 order
= page_order(page
);
5576 #ifdef CONFIG_DEBUG_VM
5577 printk(KERN_INFO
"remove from free list %lx %d %lx\n",
5578 pfn
, 1 << order
, end_pfn
);
5580 list_del(&page
->lru
);
5581 rmv_page_order(page
);
5582 zone
->free_area
[order
].nr_free
--;
5583 __mod_zone_page_state(zone
, NR_FREE_PAGES
,
5585 for (i
= 0; i
< (1 << order
); i
++)
5586 SetPageReserved((page
+i
));
5587 pfn
+= (1 << order
);
5589 spin_unlock_irqrestore(&zone
->lock
, flags
);
5593 #ifdef CONFIG_MEMORY_FAILURE
5594 bool is_free_buddy_page(struct page
*page
)
5596 struct zone
*zone
= page_zone(page
);
5597 unsigned long pfn
= page_to_pfn(page
);
5598 unsigned long flags
;
5601 spin_lock_irqsave(&zone
->lock
, flags
);
5602 for (order
= 0; order
< MAX_ORDER
; order
++) {
5603 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
5605 if (PageBuddy(page_head
) && page_order(page_head
) >= order
)
5608 spin_unlock_irqrestore(&zone
->lock
, flags
);
5610 return order
< MAX_ORDER
;
5614 static struct trace_print_flags pageflag_names
[] = {
5615 {1UL << PG_locked
, "locked" },
5616 {1UL << PG_error
, "error" },
5617 {1UL << PG_referenced
, "referenced" },
5618 {1UL << PG_uptodate
, "uptodate" },
5619 {1UL << PG_dirty
, "dirty" },
5620 {1UL << PG_lru
, "lru" },
5621 {1UL << PG_active
, "active" },
5622 {1UL << PG_slab
, "slab" },
5623 {1UL << PG_owner_priv_1
, "owner_priv_1" },
5624 {1UL << PG_arch_1
, "arch_1" },
5625 {1UL << PG_reserved
, "reserved" },
5626 {1UL << PG_private
, "private" },
5627 {1UL << PG_private_2
, "private_2" },
5628 {1UL << PG_writeback
, "writeback" },
5629 #ifdef CONFIG_PAGEFLAGS_EXTENDED
5630 {1UL << PG_head
, "head" },
5631 {1UL << PG_tail
, "tail" },
5633 {1UL << PG_compound
, "compound" },
5635 {1UL << PG_swapcache
, "swapcache" },
5636 {1UL << PG_mappedtodisk
, "mappedtodisk" },
5637 {1UL << PG_reclaim
, "reclaim" },
5638 {1UL << PG_swapbacked
, "swapbacked" },
5639 {1UL << PG_unevictable
, "unevictable" },
5641 {1UL << PG_mlocked
, "mlocked" },
5643 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
5644 {1UL << PG_uncached
, "uncached" },
5646 #ifdef CONFIG_MEMORY_FAILURE
5647 {1UL << PG_hwpoison
, "hwpoison" },
5652 static void dump_page_flags(unsigned long flags
)
5654 const char *delim
= "";
5658 printk(KERN_ALERT
"page flags: %#lx(", flags
);
5660 /* remove zone id */
5661 flags
&= (1UL << NR_PAGEFLAGS
) - 1;
5663 for (i
= 0; pageflag_names
[i
].name
&& flags
; i
++) {
5665 mask
= pageflag_names
[i
].mask
;
5666 if ((flags
& mask
) != mask
)
5670 printk("%s%s", delim
, pageflag_names
[i
].name
);
5674 /* check for left over flags */
5676 printk("%s%#lx", delim
, flags
);
5681 void dump_page(struct page
*page
)
5684 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
5685 page
, atomic_read(&page
->_count
), page_mapcount(page
),
5686 page
->mapping
, page
->index
);
5687 dump_page_flags(page
->flags
);