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/bootmem.h>
23 #include <linux/compiler.h>
24 #include <linux/kernel.h>
25 #include <linux/module.h>
26 #include <linux/suspend.h>
27 #include <linux/pagevec.h>
28 #include <linux/blkdev.h>
29 #include <linux/slab.h>
30 #include <linux/notifier.h>
31 #include <linux/topology.h>
32 #include <linux/sysctl.h>
33 #include <linux/cpu.h>
34 #include <linux/cpuset.h>
35 #include <linux/memory_hotplug.h>
36 #include <linux/nodemask.h>
37 #include <linux/vmalloc.h>
38 #include <linux/mempolicy.h>
39 #include <linux/stop_machine.h>
40 #include <linux/sort.h>
41 #include <linux/pfn.h>
42 #include <linux/backing-dev.h>
43 #include <linux/fault-inject.h>
45 #include <asm/tlbflush.h>
46 #include <asm/div64.h>
50 * Array of node states.
52 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
53 [N_POSSIBLE] = NODE_MASK_ALL,
54 [N_ONLINE] = { { [0] = 1UL } },
56 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
58 [N_HIGH_MEMORY] = { { [0] = 1UL } },
60 [N_CPU] = { { [0] = 1UL } },
63 EXPORT_SYMBOL(node_states);
65 unsigned long totalram_pages __read_mostly;
66 unsigned long totalreserve_pages __read_mostly;
68 int percpu_pagelist_fraction;
70 static void __free_pages_ok(struct page *page, unsigned int order);
73 * results with 256, 32 in the lowmem_reserve sysctl:
74 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
75 * 1G machine -> (16M dma, 784M normal, 224M high)
76 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
77 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
78 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
80 * TBD: should special case ZONE_DMA32 machines here - in those we normally
81 * don't need any ZONE_NORMAL reservation
83 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
84 #ifdef CONFIG_ZONE_DMA
87 #ifdef CONFIG_ZONE_DMA32
96 EXPORT_SYMBOL(totalram_pages);
98 static char * const zone_names[MAX_NR_ZONES] = {
99 #ifdef CONFIG_ZONE_DMA
102 #ifdef CONFIG_ZONE_DMA32
106 #ifdef CONFIG_HIGHMEM
112 int min_free_kbytes = 1024;
114 unsigned long __meminitdata nr_kernel_pages;
115 unsigned long __meminitdata nr_all_pages;
116 static unsigned long __meminitdata dma_reserve;
118 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
120 * MAX_ACTIVE_REGIONS determines the maxmimum number of distinct
121 * ranges of memory (RAM) that may be registered with add_active_range().
122 * Ranges passed to add_active_range() will be merged if possible
123 * so the number of times add_active_range() can be called is
124 * related to the number of nodes and the number of holes
126 #ifdef CONFIG_MAX_ACTIVE_REGIONS
127 /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */
128 #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS
130 #if MAX_NUMNODES >= 32
131 /* If there can be many nodes, allow up to 50 holes per node */
132 #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50)
134 /* By default, allow up to 256 distinct regions */
135 #define MAX_ACTIVE_REGIONS 256
139 static struct node_active_region __meminitdata early_node_map[MAX_ACTIVE_REGIONS];
140 static int __meminitdata nr_nodemap_entries;
141 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
142 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
143 #ifdef CONFIG_MEMORY_HOTPLUG_RESERVE
144 static unsigned long __meminitdata node_boundary_start_pfn[MAX_NUMNODES];
145 static unsigned long __meminitdata node_boundary_end_pfn[MAX_NUMNODES];
146 #endif /* CONFIG_MEMORY_HOTPLUG_RESERVE */
147 unsigned long __initdata required_kernelcore;
148 unsigned long __initdata required_movablecore;
149 unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
151 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
153 EXPORT_SYMBOL(movable_zone);
154 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
157 int nr_node_ids __read_mostly = MAX_NUMNODES;
158 EXPORT_SYMBOL(nr_node_ids);
161 #ifdef CONFIG_DEBUG_VM
162 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
166 unsigned long pfn = page_to_pfn(page);
169 seq = zone_span_seqbegin(zone);
170 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
172 else if (pfn < zone->zone_start_pfn)
174 } while (zone_span_seqretry(zone, seq));
179 static int page_is_consistent(struct zone *zone, struct page *page)
181 if (!pfn_valid_within(page_to_pfn(page)))
183 if (zone != page_zone(page))
189 * Temporary debugging check for pages not lying within a given zone.
191 static int bad_range(struct zone *zone, struct page *page)
193 if (page_outside_zone_boundaries(zone, page))
195 if (!page_is_consistent(zone, page))
201 static inline int bad_range(struct zone *zone, struct page *page)
207 static void bad_page(struct page *page)
209 printk(KERN_EMERG "Bad page state in process '%s'\n"
210 KERN_EMERG "page:%p flags:0x%0*lx mapping:%p mapcount:%d count:%d\n"
211 KERN_EMERG "Trying to fix it up, but a reboot is needed\n"
212 KERN_EMERG "Backtrace:\n",
213 current->comm, page, (int)(2*sizeof(unsigned long)),
214 (unsigned long)page->flags, page->mapping,
215 page_mapcount(page), page_count(page));
217 page->flags &= ~(1 << PG_lru |
227 set_page_count(page, 0);
228 reset_page_mapcount(page);
229 page->mapping = NULL;
230 add_taint(TAINT_BAD_PAGE);
234 * Higher-order pages are called "compound pages". They are structured thusly:
236 * The first PAGE_SIZE page is called the "head page".
238 * The remaining PAGE_SIZE pages are called "tail pages".
240 * All pages have PG_compound set. All pages have their ->private pointing at
241 * the head page (even the head page has this).
243 * The first tail page's ->lru.next holds the address of the compound page's
244 * put_page() function. Its ->lru.prev holds the order of allocation.
245 * This usage means that zero-order pages may not be compound.
248 static void free_compound_page(struct page *page)
250 __free_pages_ok(page, compound_order(page));
253 static void prep_compound_page(struct page *page, unsigned long order)
256 int nr_pages = 1 << order;
258 set_compound_page_dtor(page, free_compound_page);
259 set_compound_order(page, order);
261 for (i = 1; i < nr_pages; i++) {
262 struct page *p = page + i;
265 p->first_page = page;
269 static void destroy_compound_page(struct page *page, unsigned long order)
272 int nr_pages = 1 << order;
274 if (unlikely(compound_order(page) != order))
277 if (unlikely(!PageHead(page)))
279 __ClearPageHead(page);
280 for (i = 1; i < nr_pages; i++) {
281 struct page *p = page + i;
283 if (unlikely(!PageTail(p) |
284 (p->first_page != page)))
290 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
294 VM_BUG_ON((gfp_flags & (__GFP_WAIT | __GFP_HIGHMEM)) == __GFP_HIGHMEM);
296 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
297 * and __GFP_HIGHMEM from hard or soft interrupt context.
299 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
300 for (i = 0; i < (1 << order); i++)
301 clear_highpage(page + i);
305 * function for dealing with page's order in buddy system.
306 * zone->lock is already acquired when we use these.
307 * So, we don't need atomic page->flags operations here.
309 static inline unsigned long page_order(struct page *page)
311 return page_private(page);
314 static inline void set_page_order(struct page *page, int order)
316 set_page_private(page, order);
317 __SetPageBuddy(page);
320 static inline void rmv_page_order(struct page *page)
322 __ClearPageBuddy(page);
323 set_page_private(page, 0);
327 * Locate the struct page for both the matching buddy in our
328 * pair (buddy1) and the combined O(n+1) page they form (page).
330 * 1) Any buddy B1 will have an order O twin B2 which satisfies
331 * the following equation:
333 * For example, if the starting buddy (buddy2) is #8 its order
335 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
337 * 2) Any buddy B will have an order O+1 parent P which
338 * satisfies the following equation:
341 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
343 static inline struct page *
344 __page_find_buddy(struct page *page, unsigned long page_idx, unsigned int order)
346 unsigned long buddy_idx = page_idx ^ (1 << order);
348 return page + (buddy_idx - page_idx);
351 static inline unsigned long
352 __find_combined_index(unsigned long page_idx, unsigned int order)
354 return (page_idx & ~(1 << order));
358 * This function checks whether a page is free && is the buddy
359 * we can do coalesce a page and its buddy if
360 * (a) the buddy is not in a hole &&
361 * (b) the buddy is in the buddy system &&
362 * (c) a page and its buddy have the same order &&
363 * (d) a page and its buddy are in the same zone.
365 * For recording whether a page is in the buddy system, we use PG_buddy.
366 * Setting, clearing, and testing PG_buddy is serialized by zone->lock.
368 * For recording page's order, we use page_private(page).
370 static inline int page_is_buddy(struct page *page, struct page *buddy,
373 if (!pfn_valid_within(page_to_pfn(buddy)))
376 if (page_zone_id(page) != page_zone_id(buddy))
379 if (PageBuddy(buddy) && page_order(buddy) == order) {
380 BUG_ON(page_count(buddy) != 0);
387 * Freeing function for a buddy system allocator.
389 * The concept of a buddy system is to maintain direct-mapped table
390 * (containing bit values) for memory blocks of various "orders".
391 * The bottom level table contains the map for the smallest allocatable
392 * units of memory (here, pages), and each level above it describes
393 * pairs of units from the levels below, hence, "buddies".
394 * At a high level, all that happens here is marking the table entry
395 * at the bottom level available, and propagating the changes upward
396 * as necessary, plus some accounting needed to play nicely with other
397 * parts of the VM system.
398 * At each level, we keep a list of pages, which are heads of continuous
399 * free pages of length of (1 << order) and marked with PG_buddy. Page's
400 * order is recorded in page_private(page) field.
401 * So when we are allocating or freeing one, we can derive the state of the
402 * other. That is, if we allocate a small block, and both were
403 * free, the remainder of the region must be split into blocks.
404 * If a block is freed, and its buddy is also free, then this
405 * triggers coalescing into a block of larger size.
410 static inline void __free_one_page(struct page *page,
411 struct zone *zone, unsigned int order)
413 unsigned long page_idx;
414 int order_size = 1 << order;
416 if (unlikely(PageCompound(page)))
417 destroy_compound_page(page, order);
419 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
421 VM_BUG_ON(page_idx & (order_size - 1));
422 VM_BUG_ON(bad_range(zone, page));
424 __mod_zone_page_state(zone, NR_FREE_PAGES, order_size);
425 while (order < MAX_ORDER-1) {
426 unsigned long combined_idx;
427 struct free_area *area;
430 buddy = __page_find_buddy(page, page_idx, order);
431 if (!page_is_buddy(page, buddy, order))
432 break; /* Move the buddy up one level. */
434 list_del(&buddy->lru);
435 area = zone->free_area + order;
437 rmv_page_order(buddy);
438 combined_idx = __find_combined_index(page_idx, order);
439 page = page + (combined_idx - page_idx);
440 page_idx = combined_idx;
443 set_page_order(page, order);
444 list_add(&page->lru, &zone->free_area[order].free_list);
445 zone->free_area[order].nr_free++;
448 static inline int free_pages_check(struct page *page)
450 if (unlikely(page_mapcount(page) |
451 (page->mapping != NULL) |
452 (page_count(page) != 0) |
465 __ClearPageDirty(page);
467 * For now, we report if PG_reserved was found set, but do not
468 * clear it, and do not free the page. But we shall soon need
469 * to do more, for when the ZERO_PAGE count wraps negative.
471 return PageReserved(page);
475 * Frees a list of pages.
476 * Assumes all pages on list are in same zone, and of same order.
477 * count is the number of pages to free.
479 * If the zone was previously in an "all pages pinned" state then look to
480 * see if this freeing clears that state.
482 * And clear the zone's pages_scanned counter, to hold off the "all pages are
483 * pinned" detection logic.
485 static void free_pages_bulk(struct zone *zone, int count,
486 struct list_head *list, int order)
488 spin_lock(&zone->lock);
489 zone->all_unreclaimable = 0;
490 zone->pages_scanned = 0;
494 VM_BUG_ON(list_empty(list));
495 page = list_entry(list->prev, struct page, lru);
496 /* have to delete it as __free_one_page list manipulates */
497 list_del(&page->lru);
498 __free_one_page(page, zone, order);
500 spin_unlock(&zone->lock);
503 static void free_one_page(struct zone *zone, struct page *page, int order)
505 spin_lock(&zone->lock);
506 zone->all_unreclaimable = 0;
507 zone->pages_scanned = 0;
508 __free_one_page(page, zone, order);
509 spin_unlock(&zone->lock);
512 static void __free_pages_ok(struct page *page, unsigned int order)
518 for (i = 0 ; i < (1 << order) ; ++i)
519 reserved += free_pages_check(page + i);
523 if (!PageHighMem(page))
524 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
525 arch_free_page(page, order);
526 kernel_map_pages(page, 1 << order, 0);
528 local_irq_save(flags);
529 __count_vm_events(PGFREE, 1 << order);
530 free_one_page(page_zone(page), page, order);
531 local_irq_restore(flags);
535 * permit the bootmem allocator to evade page validation on high-order frees
537 void fastcall __init __free_pages_bootmem(struct page *page, unsigned int order)
540 __ClearPageReserved(page);
541 set_page_count(page, 0);
542 set_page_refcounted(page);
548 for (loop = 0; loop < BITS_PER_LONG; loop++) {
549 struct page *p = &page[loop];
551 if (loop + 1 < BITS_PER_LONG)
553 __ClearPageReserved(p);
554 set_page_count(p, 0);
557 set_page_refcounted(page);
558 __free_pages(page, order);
564 * The order of subdivision here is critical for the IO subsystem.
565 * Please do not alter this order without good reasons and regression
566 * testing. Specifically, as large blocks of memory are subdivided,
567 * the order in which smaller blocks are delivered depends on the order
568 * they're subdivided in this function. This is the primary factor
569 * influencing the order in which pages are delivered to the IO
570 * subsystem according to empirical testing, and this is also justified
571 * by considering the behavior of a buddy system containing a single
572 * large block of memory acted on by a series of small allocations.
573 * This behavior is a critical factor in sglist merging's success.
577 static inline void expand(struct zone *zone, struct page *page,
578 int low, int high, struct free_area *area)
580 unsigned long size = 1 << high;
586 VM_BUG_ON(bad_range(zone, &page[size]));
587 list_add(&page[size].lru, &area->free_list);
589 set_page_order(&page[size], high);
594 * This page is about to be returned from the page allocator
596 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
598 if (unlikely(page_mapcount(page) |
599 (page->mapping != NULL) |
600 (page_count(page) != 0) |
615 * For now, we report if PG_reserved was found set, but do not
616 * clear it, and do not allocate the page: as a safety net.
618 if (PageReserved(page))
621 page->flags &= ~(1 << PG_uptodate | 1 << PG_error | 1 << PG_readahead |
622 1 << PG_referenced | 1 << PG_arch_1 |
623 1 << PG_owner_priv_1 | 1 << PG_mappedtodisk);
624 set_page_private(page, 0);
625 set_page_refcounted(page);
627 arch_alloc_page(page, order);
628 kernel_map_pages(page, 1 << order, 1);
630 if (gfp_flags & __GFP_ZERO)
631 prep_zero_page(page, order, gfp_flags);
633 if (order && (gfp_flags & __GFP_COMP))
634 prep_compound_page(page, order);
640 * Do the hard work of removing an element from the buddy allocator.
641 * Call me with the zone->lock already held.
643 static struct page *__rmqueue(struct zone *zone, unsigned int order)
645 struct free_area * area;
646 unsigned int current_order;
649 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
650 area = zone->free_area + current_order;
651 if (list_empty(&area->free_list))
654 page = list_entry(area->free_list.next, struct page, lru);
655 list_del(&page->lru);
656 rmv_page_order(page);
658 __mod_zone_page_state(zone, NR_FREE_PAGES, - (1UL << order));
659 expand(zone, page, order, current_order, area);
667 * Obtain a specified number of elements from the buddy allocator, all under
668 * a single hold of the lock, for efficiency. Add them to the supplied list.
669 * Returns the number of new pages which were placed at *list.
671 static int rmqueue_bulk(struct zone *zone, unsigned int order,
672 unsigned long count, struct list_head *list)
676 spin_lock(&zone->lock);
677 for (i = 0; i < count; ++i) {
678 struct page *page = __rmqueue(zone, order);
679 if (unlikely(page == NULL))
681 list_add_tail(&page->lru, list);
683 spin_unlock(&zone->lock);
689 * Called from the vmstat counter updater to drain pagesets of this
690 * currently executing processor on remote nodes after they have
693 * Note that this function must be called with the thread pinned to
694 * a single processor.
696 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
701 local_irq_save(flags);
702 if (pcp->count >= pcp->batch)
703 to_drain = pcp->batch;
705 to_drain = pcp->count;
706 free_pages_bulk(zone, to_drain, &pcp->list, 0);
707 pcp->count -= to_drain;
708 local_irq_restore(flags);
712 static void __drain_pages(unsigned int cpu)
718 for_each_zone(zone) {
719 struct per_cpu_pageset *pset;
721 if (!populated_zone(zone))
724 pset = zone_pcp(zone, cpu);
725 for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
726 struct per_cpu_pages *pcp;
729 local_irq_save(flags);
730 free_pages_bulk(zone, pcp->count, &pcp->list, 0);
732 local_irq_restore(flags);
737 #ifdef CONFIG_HIBERNATION
739 void mark_free_pages(struct zone *zone)
741 unsigned long pfn, max_zone_pfn;
744 struct list_head *curr;
746 if (!zone->spanned_pages)
749 spin_lock_irqsave(&zone->lock, flags);
751 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
752 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
753 if (pfn_valid(pfn)) {
754 struct page *page = pfn_to_page(pfn);
756 if (!swsusp_page_is_forbidden(page))
757 swsusp_unset_page_free(page);
760 for (order = MAX_ORDER - 1; order >= 0; --order)
761 list_for_each(curr, &zone->free_area[order].free_list) {
764 pfn = page_to_pfn(list_entry(curr, struct page, lru));
765 for (i = 0; i < (1UL << order); i++)
766 swsusp_set_page_free(pfn_to_page(pfn + i));
769 spin_unlock_irqrestore(&zone->lock, flags);
773 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
775 void drain_local_pages(void)
779 local_irq_save(flags);
780 __drain_pages(smp_processor_id());
781 local_irq_restore(flags);
783 #endif /* CONFIG_HIBERNATION */
786 * Free a 0-order page
788 static void fastcall free_hot_cold_page(struct page *page, int cold)
790 struct zone *zone = page_zone(page);
791 struct per_cpu_pages *pcp;
795 page->mapping = NULL;
796 if (free_pages_check(page))
799 if (!PageHighMem(page))
800 debug_check_no_locks_freed(page_address(page), PAGE_SIZE);
801 arch_free_page(page, 0);
802 kernel_map_pages(page, 1, 0);
804 pcp = &zone_pcp(zone, get_cpu())->pcp[cold];
805 local_irq_save(flags);
806 __count_vm_event(PGFREE);
807 list_add(&page->lru, &pcp->list);
809 if (pcp->count >= pcp->high) {
810 free_pages_bulk(zone, pcp->batch, &pcp->list, 0);
811 pcp->count -= pcp->batch;
813 local_irq_restore(flags);
817 void fastcall free_hot_page(struct page *page)
819 free_hot_cold_page(page, 0);
822 void fastcall free_cold_page(struct page *page)
824 free_hot_cold_page(page, 1);
828 * split_page takes a non-compound higher-order page, and splits it into
829 * n (1<<order) sub-pages: page[0..n]
830 * Each sub-page must be freed individually.
832 * Note: this is probably too low level an operation for use in drivers.
833 * Please consult with lkml before using this in your driver.
835 void split_page(struct page *page, unsigned int order)
839 VM_BUG_ON(PageCompound(page));
840 VM_BUG_ON(!page_count(page));
841 for (i = 1; i < (1 << order); i++)
842 set_page_refcounted(page + i);
846 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
847 * we cheat by calling it from here, in the order > 0 path. Saves a branch
850 static struct page *buffered_rmqueue(struct zonelist *zonelist,
851 struct zone *zone, int order, gfp_t gfp_flags)
855 int cold = !!(gfp_flags & __GFP_COLD);
860 if (likely(order == 0)) {
861 struct per_cpu_pages *pcp;
863 pcp = &zone_pcp(zone, cpu)->pcp[cold];
864 local_irq_save(flags);
866 pcp->count = rmqueue_bulk(zone, 0,
867 pcp->batch, &pcp->list);
868 if (unlikely(!pcp->count))
871 page = list_entry(pcp->list.next, struct page, lru);
872 list_del(&page->lru);
875 spin_lock_irqsave(&zone->lock, flags);
876 page = __rmqueue(zone, order);
877 spin_unlock(&zone->lock);
882 __count_zone_vm_events(PGALLOC, zone, 1 << order);
883 zone_statistics(zonelist, zone);
884 local_irq_restore(flags);
887 VM_BUG_ON(bad_range(zone, page));
888 if (prep_new_page(page, order, gfp_flags))
893 local_irq_restore(flags);
898 #define ALLOC_NO_WATERMARKS 0x01 /* don't check watermarks at all */
899 #define ALLOC_WMARK_MIN 0x02 /* use pages_min watermark */
900 #define ALLOC_WMARK_LOW 0x04 /* use pages_low watermark */
901 #define ALLOC_WMARK_HIGH 0x08 /* use pages_high watermark */
902 #define ALLOC_HARDER 0x10 /* try to alloc harder */
903 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
904 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
906 #ifdef CONFIG_FAIL_PAGE_ALLOC
908 static struct fail_page_alloc_attr {
909 struct fault_attr attr;
911 u32 ignore_gfp_highmem;
915 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
917 struct dentry *ignore_gfp_highmem_file;
918 struct dentry *ignore_gfp_wait_file;
919 struct dentry *min_order_file;
921 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
923 } fail_page_alloc = {
924 .attr = FAULT_ATTR_INITIALIZER,
925 .ignore_gfp_wait = 1,
926 .ignore_gfp_highmem = 1,
930 static int __init setup_fail_page_alloc(char *str)
932 return setup_fault_attr(&fail_page_alloc.attr, str);
934 __setup("fail_page_alloc=", setup_fail_page_alloc);
936 static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
938 if (order < fail_page_alloc.min_order)
940 if (gfp_mask & __GFP_NOFAIL)
942 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
944 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
947 return should_fail(&fail_page_alloc.attr, 1 << order);
950 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
952 static int __init fail_page_alloc_debugfs(void)
954 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
958 err = init_fault_attr_dentries(&fail_page_alloc.attr,
962 dir = fail_page_alloc.attr.dentries.dir;
964 fail_page_alloc.ignore_gfp_wait_file =
965 debugfs_create_bool("ignore-gfp-wait", mode, dir,
966 &fail_page_alloc.ignore_gfp_wait);
968 fail_page_alloc.ignore_gfp_highmem_file =
969 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
970 &fail_page_alloc.ignore_gfp_highmem);
971 fail_page_alloc.min_order_file =
972 debugfs_create_u32("min-order", mode, dir,
973 &fail_page_alloc.min_order);
975 if (!fail_page_alloc.ignore_gfp_wait_file ||
976 !fail_page_alloc.ignore_gfp_highmem_file ||
977 !fail_page_alloc.min_order_file) {
979 debugfs_remove(fail_page_alloc.ignore_gfp_wait_file);
980 debugfs_remove(fail_page_alloc.ignore_gfp_highmem_file);
981 debugfs_remove(fail_page_alloc.min_order_file);
982 cleanup_fault_attr_dentries(&fail_page_alloc.attr);
988 late_initcall(fail_page_alloc_debugfs);
990 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
992 #else /* CONFIG_FAIL_PAGE_ALLOC */
994 static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
999 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1002 * Return 1 if free pages are above 'mark'. This takes into account the order
1003 * of the allocation.
1005 int zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1006 int classzone_idx, int alloc_flags)
1008 /* free_pages my go negative - that's OK */
1010 long free_pages = zone_page_state(z, NR_FREE_PAGES) - (1 << order) + 1;
1013 if (alloc_flags & ALLOC_HIGH)
1015 if (alloc_flags & ALLOC_HARDER)
1018 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
1020 for (o = 0; o < order; o++) {
1021 /* At the next order, this order's pages become unavailable */
1022 free_pages -= z->free_area[o].nr_free << o;
1024 /* Require fewer higher order pages to be free */
1027 if (free_pages <= min)
1035 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1036 * skip over zones that are not allowed by the cpuset, or that have
1037 * been recently (in last second) found to be nearly full. See further
1038 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1039 * that have to skip over alot of full or unallowed zones.
1041 * If the zonelist cache is present in the passed in zonelist, then
1042 * returns a pointer to the allowed node mask (either the current
1043 * tasks mems_allowed, or node_online_map.)
1045 * If the zonelist cache is not available for this zonelist, does
1046 * nothing and returns NULL.
1048 * If the fullzones BITMAP in the zonelist cache is stale (more than
1049 * a second since last zap'd) then we zap it out (clear its bits.)
1051 * We hold off even calling zlc_setup, until after we've checked the
1052 * first zone in the zonelist, on the theory that most allocations will
1053 * be satisfied from that first zone, so best to examine that zone as
1054 * quickly as we can.
1056 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1058 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1059 nodemask_t *allowednodes; /* zonelist_cache approximation */
1061 zlc = zonelist->zlcache_ptr;
1065 if (jiffies - zlc->last_full_zap > 1 * HZ) {
1066 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1067 zlc->last_full_zap = jiffies;
1070 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1071 &cpuset_current_mems_allowed :
1073 return allowednodes;
1077 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1078 * if it is worth looking at further for free memory:
1079 * 1) Check that the zone isn't thought to be full (doesn't have its
1080 * bit set in the zonelist_cache fullzones BITMAP).
1081 * 2) Check that the zones node (obtained from the zonelist_cache
1082 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1083 * Return true (non-zero) if zone is worth looking at further, or
1084 * else return false (zero) if it is not.
1086 * This check -ignores- the distinction between various watermarks,
1087 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1088 * found to be full for any variation of these watermarks, it will
1089 * be considered full for up to one second by all requests, unless
1090 * we are so low on memory on all allowed nodes that we are forced
1091 * into the second scan of the zonelist.
1093 * In the second scan we ignore this zonelist cache and exactly
1094 * apply the watermarks to all zones, even it is slower to do so.
1095 * We are low on memory in the second scan, and should leave no stone
1096 * unturned looking for a free page.
1098 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zone **z,
1099 nodemask_t *allowednodes)
1101 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1102 int i; /* index of *z in zonelist zones */
1103 int n; /* node that zone *z is on */
1105 zlc = zonelist->zlcache_ptr;
1109 i = z - zonelist->zones;
1112 /* This zone is worth trying if it is allowed but not full */
1113 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1117 * Given 'z' scanning a zonelist, set the corresponding bit in
1118 * zlc->fullzones, so that subsequent attempts to allocate a page
1119 * from that zone don't waste time re-examining it.
1121 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zone **z)
1123 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1124 int i; /* index of *z in zonelist zones */
1126 zlc = zonelist->zlcache_ptr;
1130 i = z - zonelist->zones;
1132 set_bit(i, zlc->fullzones);
1135 #else /* CONFIG_NUMA */
1137 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1142 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zone **z,
1143 nodemask_t *allowednodes)
1148 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zone **z)
1151 #endif /* CONFIG_NUMA */
1154 * get_page_from_freelist goes through the zonelist trying to allocate
1157 static struct page *
1158 get_page_from_freelist(gfp_t gfp_mask, unsigned int order,
1159 struct zonelist *zonelist, int alloc_flags)
1162 struct page *page = NULL;
1163 int classzone_idx = zone_idx(zonelist->zones[0]);
1165 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1166 int zlc_active = 0; /* set if using zonelist_cache */
1167 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1168 enum zone_type highest_zoneidx = -1; /* Gets set for policy zonelists */
1172 * Scan zonelist, looking for a zone with enough free.
1173 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1175 z = zonelist->zones;
1179 * In NUMA, this could be a policy zonelist which contains
1180 * zones that may not be allowed by the current gfp_mask.
1181 * Check the zone is allowed by the current flags
1183 if (unlikely(alloc_should_filter_zonelist(zonelist))) {
1184 if (highest_zoneidx == -1)
1185 highest_zoneidx = gfp_zone(gfp_mask);
1186 if (zone_idx(*z) > highest_zoneidx)
1190 if (NUMA_BUILD && zlc_active &&
1191 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1194 if ((alloc_flags & ALLOC_CPUSET) &&
1195 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1198 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1200 if (alloc_flags & ALLOC_WMARK_MIN)
1201 mark = zone->pages_min;
1202 else if (alloc_flags & ALLOC_WMARK_LOW)
1203 mark = zone->pages_low;
1205 mark = zone->pages_high;
1206 if (!zone_watermark_ok(zone, order, mark,
1207 classzone_idx, alloc_flags)) {
1208 if (!zone_reclaim_mode ||
1209 !zone_reclaim(zone, gfp_mask, order))
1210 goto this_zone_full;
1214 page = buffered_rmqueue(zonelist, zone, order, gfp_mask);
1219 zlc_mark_zone_full(zonelist, z);
1221 if (NUMA_BUILD && !did_zlc_setup) {
1222 /* we do zlc_setup after the first zone is tried */
1223 allowednodes = zlc_setup(zonelist, alloc_flags);
1227 } while (*(++z) != NULL);
1229 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) {
1230 /* Disable zlc cache for second zonelist scan */
1238 * This is the 'heart' of the zoned buddy allocator.
1240 struct page * fastcall
1241 __alloc_pages(gfp_t gfp_mask, unsigned int order,
1242 struct zonelist *zonelist)
1244 const gfp_t wait = gfp_mask & __GFP_WAIT;
1247 struct reclaim_state reclaim_state;
1248 struct task_struct *p = current;
1251 int did_some_progress;
1253 might_sleep_if(wait);
1255 if (should_fail_alloc_page(gfp_mask, order))
1259 z = zonelist->zones; /* the list of zones suitable for gfp_mask */
1261 if (unlikely(*z == NULL)) {
1263 * Happens if we have an empty zonelist as a result of
1264 * GFP_THISNODE being used on a memoryless node
1269 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, order,
1270 zonelist, ALLOC_WMARK_LOW|ALLOC_CPUSET);
1275 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
1276 * __GFP_NOWARN set) should not cause reclaim since the subsystem
1277 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
1278 * using a larger set of nodes after it has established that the
1279 * allowed per node queues are empty and that nodes are
1282 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
1285 for (z = zonelist->zones; *z; z++)
1286 wakeup_kswapd(*z, order);
1289 * OK, we're below the kswapd watermark and have kicked background
1290 * reclaim. Now things get more complex, so set up alloc_flags according
1291 * to how we want to proceed.
1293 * The caller may dip into page reserves a bit more if the caller
1294 * cannot run direct reclaim, or if the caller has realtime scheduling
1295 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
1296 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
1298 alloc_flags = ALLOC_WMARK_MIN;
1299 if ((unlikely(rt_task(p)) && !in_interrupt()) || !wait)
1300 alloc_flags |= ALLOC_HARDER;
1301 if (gfp_mask & __GFP_HIGH)
1302 alloc_flags |= ALLOC_HIGH;
1304 alloc_flags |= ALLOC_CPUSET;
1307 * Go through the zonelist again. Let __GFP_HIGH and allocations
1308 * coming from realtime tasks go deeper into reserves.
1310 * This is the last chance, in general, before the goto nopage.
1311 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
1312 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1314 page = get_page_from_freelist(gfp_mask, order, zonelist, alloc_flags);
1318 /* This allocation should allow future memory freeing. */
1321 if (((p->flags & PF_MEMALLOC) || unlikely(test_thread_flag(TIF_MEMDIE)))
1322 && !in_interrupt()) {
1323 if (!(gfp_mask & __GFP_NOMEMALLOC)) {
1325 /* go through the zonelist yet again, ignoring mins */
1326 page = get_page_from_freelist(gfp_mask, order,
1327 zonelist, ALLOC_NO_WATERMARKS);
1330 if (gfp_mask & __GFP_NOFAIL) {
1331 congestion_wait(WRITE, HZ/50);
1338 /* Atomic allocations - we can't balance anything */
1344 /* We now go into synchronous reclaim */
1345 cpuset_memory_pressure_bump();
1346 p->flags |= PF_MEMALLOC;
1347 reclaim_state.reclaimed_slab = 0;
1348 p->reclaim_state = &reclaim_state;
1350 did_some_progress = try_to_free_pages(zonelist->zones, order, gfp_mask);
1352 p->reclaim_state = NULL;
1353 p->flags &= ~PF_MEMALLOC;
1357 if (likely(did_some_progress)) {
1358 page = get_page_from_freelist(gfp_mask, order,
1359 zonelist, alloc_flags);
1362 } else if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
1364 * Go through the zonelist yet one more time, keep
1365 * very high watermark here, this is only to catch
1366 * a parallel oom killing, we must fail if we're still
1367 * under heavy pressure.
1369 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, order,
1370 zonelist, ALLOC_WMARK_HIGH|ALLOC_CPUSET);
1374 /* The OOM killer will not help higher order allocs so fail */
1375 if (order > PAGE_ALLOC_COSTLY_ORDER)
1378 out_of_memory(zonelist, gfp_mask, order);
1383 * Don't let big-order allocations loop unless the caller explicitly
1384 * requests that. Wait for some write requests to complete then retry.
1386 * In this implementation, __GFP_REPEAT means __GFP_NOFAIL for order
1387 * <= 3, but that may not be true in other implementations.
1390 if (!(gfp_mask & __GFP_NORETRY)) {
1391 if ((order <= PAGE_ALLOC_COSTLY_ORDER) ||
1392 (gfp_mask & __GFP_REPEAT))
1394 if (gfp_mask & __GFP_NOFAIL)
1398 congestion_wait(WRITE, HZ/50);
1403 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
1404 printk(KERN_WARNING "%s: page allocation failure."
1405 " order:%d, mode:0x%x\n",
1406 p->comm, order, gfp_mask);
1414 EXPORT_SYMBOL(__alloc_pages);
1417 * Common helper functions.
1419 fastcall unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
1422 page = alloc_pages(gfp_mask, order);
1425 return (unsigned long) page_address(page);
1428 EXPORT_SYMBOL(__get_free_pages);
1430 fastcall unsigned long get_zeroed_page(gfp_t gfp_mask)
1435 * get_zeroed_page() returns a 32-bit address, which cannot represent
1438 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
1440 page = alloc_pages(gfp_mask | __GFP_ZERO, 0);
1442 return (unsigned long) page_address(page);
1446 EXPORT_SYMBOL(get_zeroed_page);
1448 void __pagevec_free(struct pagevec *pvec)
1450 int i = pagevec_count(pvec);
1453 free_hot_cold_page(pvec->pages[i], pvec->cold);
1456 fastcall void __free_pages(struct page *page, unsigned int order)
1458 if (put_page_testzero(page)) {
1460 free_hot_page(page);
1462 __free_pages_ok(page, order);
1466 EXPORT_SYMBOL(__free_pages);
1468 fastcall void free_pages(unsigned long addr, unsigned int order)
1471 VM_BUG_ON(!virt_addr_valid((void *)addr));
1472 __free_pages(virt_to_page((void *)addr), order);
1476 EXPORT_SYMBOL(free_pages);
1478 static unsigned int nr_free_zone_pages(int offset)
1480 /* Just pick one node, since fallback list is circular */
1481 pg_data_t *pgdat = NODE_DATA(numa_node_id());
1482 unsigned int sum = 0;
1484 struct zonelist *zonelist = pgdat->node_zonelists + offset;
1485 struct zone **zonep = zonelist->zones;
1488 for (zone = *zonep++; zone; zone = *zonep++) {
1489 unsigned long size = zone->present_pages;
1490 unsigned long high = zone->pages_high;
1499 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
1501 unsigned int nr_free_buffer_pages(void)
1503 return nr_free_zone_pages(gfp_zone(GFP_USER));
1505 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
1508 * Amount of free RAM allocatable within all zones
1510 unsigned int nr_free_pagecache_pages(void)
1512 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
1515 static inline void show_node(struct zone *zone)
1518 printk("Node %d ", zone_to_nid(zone));
1521 void si_meminfo(struct sysinfo *val)
1523 val->totalram = totalram_pages;
1525 val->freeram = global_page_state(NR_FREE_PAGES);
1526 val->bufferram = nr_blockdev_pages();
1527 val->totalhigh = totalhigh_pages;
1528 val->freehigh = nr_free_highpages();
1529 val->mem_unit = PAGE_SIZE;
1532 EXPORT_SYMBOL(si_meminfo);
1535 void si_meminfo_node(struct sysinfo *val, int nid)
1537 pg_data_t *pgdat = NODE_DATA(nid);
1539 val->totalram = pgdat->node_present_pages;
1540 val->freeram = node_page_state(nid, NR_FREE_PAGES);
1541 #ifdef CONFIG_HIGHMEM
1542 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
1543 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
1549 val->mem_unit = PAGE_SIZE;
1553 #define K(x) ((x) << (PAGE_SHIFT-10))
1556 * Show free area list (used inside shift_scroll-lock stuff)
1557 * We also calculate the percentage fragmentation. We do this by counting the
1558 * memory on each free list with the exception of the first item on the list.
1560 void show_free_areas(void)
1565 for_each_zone(zone) {
1566 if (!populated_zone(zone))
1570 printk("%s per-cpu:\n", zone->name);
1572 for_each_online_cpu(cpu) {
1573 struct per_cpu_pageset *pageset;
1575 pageset = zone_pcp(zone, cpu);
1577 printk("CPU %4d: Hot: hi:%5d, btch:%4d usd:%4d "
1578 "Cold: hi:%5d, btch:%4d usd:%4d\n",
1579 cpu, pageset->pcp[0].high,
1580 pageset->pcp[0].batch, pageset->pcp[0].count,
1581 pageset->pcp[1].high, pageset->pcp[1].batch,
1582 pageset->pcp[1].count);
1586 printk("Active:%lu inactive:%lu dirty:%lu writeback:%lu unstable:%lu\n"
1587 " free:%lu slab:%lu mapped:%lu pagetables:%lu bounce:%lu\n",
1588 global_page_state(NR_ACTIVE),
1589 global_page_state(NR_INACTIVE),
1590 global_page_state(NR_FILE_DIRTY),
1591 global_page_state(NR_WRITEBACK),
1592 global_page_state(NR_UNSTABLE_NFS),
1593 global_page_state(NR_FREE_PAGES),
1594 global_page_state(NR_SLAB_RECLAIMABLE) +
1595 global_page_state(NR_SLAB_UNRECLAIMABLE),
1596 global_page_state(NR_FILE_MAPPED),
1597 global_page_state(NR_PAGETABLE),
1598 global_page_state(NR_BOUNCE));
1600 for_each_zone(zone) {
1603 if (!populated_zone(zone))
1615 " pages_scanned:%lu"
1616 " all_unreclaimable? %s"
1619 K(zone_page_state(zone, NR_FREE_PAGES)),
1622 K(zone->pages_high),
1623 K(zone_page_state(zone, NR_ACTIVE)),
1624 K(zone_page_state(zone, NR_INACTIVE)),
1625 K(zone->present_pages),
1626 zone->pages_scanned,
1627 (zone->all_unreclaimable ? "yes" : "no")
1629 printk("lowmem_reserve[]:");
1630 for (i = 0; i < MAX_NR_ZONES; i++)
1631 printk(" %lu", zone->lowmem_reserve[i]);
1635 for_each_zone(zone) {
1636 unsigned long nr[MAX_ORDER], flags, order, total = 0;
1638 if (!populated_zone(zone))
1642 printk("%s: ", zone->name);
1644 spin_lock_irqsave(&zone->lock, flags);
1645 for (order = 0; order < MAX_ORDER; order++) {
1646 nr[order] = zone->free_area[order].nr_free;
1647 total += nr[order] << order;
1649 spin_unlock_irqrestore(&zone->lock, flags);
1650 for (order = 0; order < MAX_ORDER; order++)
1651 printk("%lu*%lukB ", nr[order], K(1UL) << order);
1652 printk("= %lukB\n", K(total));
1655 show_swap_cache_info();
1659 * Builds allocation fallback zone lists.
1661 * Add all populated zones of a node to the zonelist.
1663 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
1664 int nr_zones, enum zone_type zone_type)
1668 BUG_ON(zone_type >= MAX_NR_ZONES);
1673 zone = pgdat->node_zones + zone_type;
1674 if (populated_zone(zone)) {
1675 zonelist->zones[nr_zones++] = zone;
1676 check_highest_zone(zone_type);
1679 } while (zone_type);
1686 * 0 = automatic detection of better ordering.
1687 * 1 = order by ([node] distance, -zonetype)
1688 * 2 = order by (-zonetype, [node] distance)
1690 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
1691 * the same zonelist. So only NUMA can configure this param.
1693 #define ZONELIST_ORDER_DEFAULT 0
1694 #define ZONELIST_ORDER_NODE 1
1695 #define ZONELIST_ORDER_ZONE 2
1697 /* zonelist order in the kernel.
1698 * set_zonelist_order() will set this to NODE or ZONE.
1700 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
1701 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
1705 /* The value user specified ....changed by config */
1706 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
1707 /* string for sysctl */
1708 #define NUMA_ZONELIST_ORDER_LEN 16
1709 char numa_zonelist_order[16] = "default";
1712 * interface for configure zonelist ordering.
1713 * command line option "numa_zonelist_order"
1714 * = "[dD]efault - default, automatic configuration.
1715 * = "[nN]ode - order by node locality, then by zone within node
1716 * = "[zZ]one - order by zone, then by locality within zone
1719 static int __parse_numa_zonelist_order(char *s)
1721 if (*s == 'd' || *s == 'D') {
1722 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
1723 } else if (*s == 'n' || *s == 'N') {
1724 user_zonelist_order = ZONELIST_ORDER_NODE;
1725 } else if (*s == 'z' || *s == 'Z') {
1726 user_zonelist_order = ZONELIST_ORDER_ZONE;
1729 "Ignoring invalid numa_zonelist_order value: "
1736 static __init int setup_numa_zonelist_order(char *s)
1739 return __parse_numa_zonelist_order(s);
1742 early_param("numa_zonelist_order", setup_numa_zonelist_order);
1745 * sysctl handler for numa_zonelist_order
1747 int numa_zonelist_order_handler(ctl_table *table, int write,
1748 struct file *file, void __user *buffer, size_t *length,
1751 char saved_string[NUMA_ZONELIST_ORDER_LEN];
1755 strncpy(saved_string, (char*)table->data,
1756 NUMA_ZONELIST_ORDER_LEN);
1757 ret = proc_dostring(table, write, file, buffer, length, ppos);
1761 int oldval = user_zonelist_order;
1762 if (__parse_numa_zonelist_order((char*)table->data)) {
1764 * bogus value. restore saved string
1766 strncpy((char*)table->data, saved_string,
1767 NUMA_ZONELIST_ORDER_LEN);
1768 user_zonelist_order = oldval;
1769 } else if (oldval != user_zonelist_order)
1770 build_all_zonelists();
1776 #define MAX_NODE_LOAD (num_online_nodes())
1777 static int node_load[MAX_NUMNODES];
1780 * find_next_best_node - find the next node that should appear in a given node's fallback list
1781 * @node: node whose fallback list we're appending
1782 * @used_node_mask: nodemask_t of already used nodes
1784 * We use a number of factors to determine which is the next node that should
1785 * appear on a given node's fallback list. The node should not have appeared
1786 * already in @node's fallback list, and it should be the next closest node
1787 * according to the distance array (which contains arbitrary distance values
1788 * from each node to each node in the system), and should also prefer nodes
1789 * with no CPUs, since presumably they'll have very little allocation pressure
1790 * on them otherwise.
1791 * It returns -1 if no node is found.
1793 static int find_next_best_node(int node, nodemask_t *used_node_mask)
1796 int min_val = INT_MAX;
1799 /* Use the local node if we haven't already */
1800 if (!node_isset(node, *used_node_mask)) {
1801 node_set(node, *used_node_mask);
1805 for_each_online_node(n) {
1808 /* Don't want a node to appear more than once */
1809 if (node_isset(n, *used_node_mask))
1812 /* Use the distance array to find the distance */
1813 val = node_distance(node, n);
1815 /* Penalize nodes under us ("prefer the next node") */
1818 /* Give preference to headless and unused nodes */
1819 tmp = node_to_cpumask(n);
1820 if (!cpus_empty(tmp))
1821 val += PENALTY_FOR_NODE_WITH_CPUS;
1823 /* Slight preference for less loaded node */
1824 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
1825 val += node_load[n];
1827 if (val < min_val) {
1834 node_set(best_node, *used_node_mask);
1841 * Build zonelists ordered by node and zones within node.
1842 * This results in maximum locality--normal zone overflows into local
1843 * DMA zone, if any--but risks exhausting DMA zone.
1845 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
1849 struct zonelist *zonelist;
1851 for (i = 0; i < MAX_NR_ZONES; i++) {
1852 zonelist = pgdat->node_zonelists + i;
1853 for (j = 0; zonelist->zones[j] != NULL; j++)
1855 j = build_zonelists_node(NODE_DATA(node), zonelist, j, i);
1856 zonelist->zones[j] = NULL;
1861 * Build gfp_thisnode zonelists
1863 static void build_thisnode_zonelists(pg_data_t *pgdat)
1867 struct zonelist *zonelist;
1869 for (i = 0; i < MAX_NR_ZONES; i++) {
1870 zonelist = pgdat->node_zonelists + MAX_NR_ZONES + i;
1871 j = build_zonelists_node(pgdat, zonelist, 0, i);
1872 zonelist->zones[j] = NULL;
1877 * Build zonelists ordered by zone and nodes within zones.
1878 * This results in conserving DMA zone[s] until all Normal memory is
1879 * exhausted, but results in overflowing to remote node while memory
1880 * may still exist in local DMA zone.
1882 static int node_order[MAX_NUMNODES];
1884 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
1888 int zone_type; /* needs to be signed */
1890 struct zonelist *zonelist;
1892 for (i = 0; i < MAX_NR_ZONES; i++) {
1893 zonelist = pgdat->node_zonelists + i;
1895 for (zone_type = i; zone_type >= 0; zone_type--) {
1896 for (j = 0; j < nr_nodes; j++) {
1897 node = node_order[j];
1898 z = &NODE_DATA(node)->node_zones[zone_type];
1899 if (populated_zone(z)) {
1900 zonelist->zones[pos++] = z;
1901 check_highest_zone(zone_type);
1905 zonelist->zones[pos] = NULL;
1909 static int default_zonelist_order(void)
1912 unsigned long low_kmem_size,total_size;
1916 * ZONE_DMA and ZONE_DMA32 can be very small area in the sytem.
1917 * If they are really small and used heavily, the system can fall
1918 * into OOM very easily.
1919 * This function detect ZONE_DMA/DMA32 size and confgigures zone order.
1921 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
1924 for_each_online_node(nid) {
1925 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
1926 z = &NODE_DATA(nid)->node_zones[zone_type];
1927 if (populated_zone(z)) {
1928 if (zone_type < ZONE_NORMAL)
1929 low_kmem_size += z->present_pages;
1930 total_size += z->present_pages;
1934 if (!low_kmem_size || /* there are no DMA area. */
1935 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
1936 return ZONELIST_ORDER_NODE;
1938 * look into each node's config.
1939 * If there is a node whose DMA/DMA32 memory is very big area on
1940 * local memory, NODE_ORDER may be suitable.
1942 average_size = total_size / (num_online_nodes() + 1);
1943 for_each_online_node(nid) {
1946 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
1947 z = &NODE_DATA(nid)->node_zones[zone_type];
1948 if (populated_zone(z)) {
1949 if (zone_type < ZONE_NORMAL)
1950 low_kmem_size += z->present_pages;
1951 total_size += z->present_pages;
1954 if (low_kmem_size &&
1955 total_size > average_size && /* ignore small node */
1956 low_kmem_size > total_size * 70/100)
1957 return ZONELIST_ORDER_NODE;
1959 return ZONELIST_ORDER_ZONE;
1962 static void set_zonelist_order(void)
1964 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
1965 current_zonelist_order = default_zonelist_order();
1967 current_zonelist_order = user_zonelist_order;
1970 static void build_zonelists(pg_data_t *pgdat)
1974 nodemask_t used_mask;
1975 int local_node, prev_node;
1976 struct zonelist *zonelist;
1977 int order = current_zonelist_order;
1979 /* initialize zonelists */
1980 for (i = 0; i < MAX_ZONELISTS; i++) {
1981 zonelist = pgdat->node_zonelists + i;
1982 zonelist->zones[0] = NULL;
1985 /* NUMA-aware ordering of nodes */
1986 local_node = pgdat->node_id;
1987 load = num_online_nodes();
1988 prev_node = local_node;
1989 nodes_clear(used_mask);
1991 memset(node_load, 0, sizeof(node_load));
1992 memset(node_order, 0, sizeof(node_order));
1995 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
1996 int distance = node_distance(local_node, node);
1999 * If another node is sufficiently far away then it is better
2000 * to reclaim pages in a zone before going off node.
2002 if (distance > RECLAIM_DISTANCE)
2003 zone_reclaim_mode = 1;
2006 * We don't want to pressure a particular node.
2007 * So adding penalty to the first node in same
2008 * distance group to make it round-robin.
2010 if (distance != node_distance(local_node, prev_node))
2011 node_load[node] = load;
2015 if (order == ZONELIST_ORDER_NODE)
2016 build_zonelists_in_node_order(pgdat, node);
2018 node_order[j++] = node; /* remember order */
2021 if (order == ZONELIST_ORDER_ZONE) {
2022 /* calculate node order -- i.e., DMA last! */
2023 build_zonelists_in_zone_order(pgdat, j);
2026 build_thisnode_zonelists(pgdat);
2029 /* Construct the zonelist performance cache - see further mmzone.h */
2030 static void build_zonelist_cache(pg_data_t *pgdat)
2034 for (i = 0; i < MAX_NR_ZONES; i++) {
2035 struct zonelist *zonelist;
2036 struct zonelist_cache *zlc;
2039 zonelist = pgdat->node_zonelists + i;
2040 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
2041 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
2042 for (z = zonelist->zones; *z; z++)
2043 zlc->z_to_n[z - zonelist->zones] = zone_to_nid(*z);
2048 #else /* CONFIG_NUMA */
2050 static void set_zonelist_order(void)
2052 current_zonelist_order = ZONELIST_ORDER_ZONE;
2055 static void build_zonelists(pg_data_t *pgdat)
2057 int node, local_node;
2060 local_node = pgdat->node_id;
2061 for (i = 0; i < MAX_NR_ZONES; i++) {
2062 struct zonelist *zonelist;
2064 zonelist = pgdat->node_zonelists + i;
2066 j = build_zonelists_node(pgdat, zonelist, 0, i);
2068 * Now we build the zonelist so that it contains the zones
2069 * of all the other nodes.
2070 * We don't want to pressure a particular node, so when
2071 * building the zones for node N, we make sure that the
2072 * zones coming right after the local ones are those from
2073 * node N+1 (modulo N)
2075 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
2076 if (!node_online(node))
2078 j = build_zonelists_node(NODE_DATA(node), zonelist, j, i);
2080 for (node = 0; node < local_node; node++) {
2081 if (!node_online(node))
2083 j = build_zonelists_node(NODE_DATA(node), zonelist, j, i);
2086 zonelist->zones[j] = NULL;
2090 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
2091 static void build_zonelist_cache(pg_data_t *pgdat)
2095 for (i = 0; i < MAX_NR_ZONES; i++)
2096 pgdat->node_zonelists[i].zlcache_ptr = NULL;
2099 #endif /* CONFIG_NUMA */
2101 /* Any regular memory on that node ? */
2102 static void check_for_regular_memory(pg_data_t *pgdat)
2104 #ifdef CONFIG_HIGHMEM
2105 enum zone_type zone_type;
2107 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) {
2108 struct zone *zone = &pgdat->node_zones[zone_type];
2109 if (zone->present_pages)
2110 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY);
2115 /* return values int ....just for stop_machine_run() */
2116 static int __build_all_zonelists(void *dummy)
2120 for_each_online_node(nid) {
2121 pg_data_t *pgdat = NODE_DATA(nid);
2123 build_zonelists(pgdat);
2124 build_zonelist_cache(pgdat);
2126 /* Any memory on that node */
2127 if (pgdat->node_present_pages)
2128 node_set_state(nid, N_HIGH_MEMORY);
2129 check_for_regular_memory(pgdat);
2134 void build_all_zonelists(void)
2136 set_zonelist_order();
2138 if (system_state == SYSTEM_BOOTING) {
2139 __build_all_zonelists(NULL);
2140 cpuset_init_current_mems_allowed();
2142 /* we have to stop all cpus to guaranntee there is no user
2144 stop_machine_run(__build_all_zonelists, NULL, NR_CPUS);
2145 /* cpuset refresh routine should be here */
2147 vm_total_pages = nr_free_pagecache_pages();
2148 printk("Built %i zonelists in %s order. Total pages: %ld\n",
2150 zonelist_order_name[current_zonelist_order],
2153 printk("Policy zone: %s\n", zone_names[policy_zone]);
2158 * Helper functions to size the waitqueue hash table.
2159 * Essentially these want to choose hash table sizes sufficiently
2160 * large so that collisions trying to wait on pages are rare.
2161 * But in fact, the number of active page waitqueues on typical
2162 * systems is ridiculously low, less than 200. So this is even
2163 * conservative, even though it seems large.
2165 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
2166 * waitqueues, i.e. the size of the waitq table given the number of pages.
2168 #define PAGES_PER_WAITQUEUE 256
2170 #ifndef CONFIG_MEMORY_HOTPLUG
2171 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
2173 unsigned long size = 1;
2175 pages /= PAGES_PER_WAITQUEUE;
2177 while (size < pages)
2181 * Once we have dozens or even hundreds of threads sleeping
2182 * on IO we've got bigger problems than wait queue collision.
2183 * Limit the size of the wait table to a reasonable size.
2185 size = min(size, 4096UL);
2187 return max(size, 4UL);
2191 * A zone's size might be changed by hot-add, so it is not possible to determine
2192 * a suitable size for its wait_table. So we use the maximum size now.
2194 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
2196 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
2197 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
2198 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
2200 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
2201 * or more by the traditional way. (See above). It equals:
2203 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
2204 * ia64(16K page size) : = ( 8G + 4M)byte.
2205 * powerpc (64K page size) : = (32G +16M)byte.
2207 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
2214 * This is an integer logarithm so that shifts can be used later
2215 * to extract the more random high bits from the multiplicative
2216 * hash function before the remainder is taken.
2218 static inline unsigned long wait_table_bits(unsigned long size)
2223 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
2226 * Initially all pages are reserved - free ones are freed
2227 * up by free_all_bootmem() once the early boot process is
2228 * done. Non-atomic initialization, single-pass.
2230 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
2231 unsigned long start_pfn, enum memmap_context context)
2234 unsigned long end_pfn = start_pfn + size;
2237 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
2239 * There can be holes in boot-time mem_map[]s
2240 * handed to this function. They do not
2241 * exist on hotplugged memory.
2243 if (context == MEMMAP_EARLY) {
2244 if (!early_pfn_valid(pfn))
2246 if (!early_pfn_in_nid(pfn, nid))
2249 page = pfn_to_page(pfn);
2250 set_page_links(page, zone, nid, pfn);
2251 init_page_count(page);
2252 reset_page_mapcount(page);
2253 SetPageReserved(page);
2254 INIT_LIST_HEAD(&page->lru);
2255 #ifdef WANT_PAGE_VIRTUAL
2256 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
2257 if (!is_highmem_idx(zone))
2258 set_page_address(page, __va(pfn << PAGE_SHIFT));
2263 static void __meminit zone_init_free_lists(struct pglist_data *pgdat,
2264 struct zone *zone, unsigned long size)
2267 for (order = 0; order < MAX_ORDER ; order++) {
2268 INIT_LIST_HEAD(&zone->free_area[order].free_list);
2269 zone->free_area[order].nr_free = 0;
2273 #ifndef __HAVE_ARCH_MEMMAP_INIT
2274 #define memmap_init(size, nid, zone, start_pfn) \
2275 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
2278 static int __devinit zone_batchsize(struct zone *zone)
2283 * The per-cpu-pages pools are set to around 1000th of the
2284 * size of the zone. But no more than 1/2 of a meg.
2286 * OK, so we don't know how big the cache is. So guess.
2288 batch = zone->present_pages / 1024;
2289 if (batch * PAGE_SIZE > 512 * 1024)
2290 batch = (512 * 1024) / PAGE_SIZE;
2291 batch /= 4; /* We effectively *= 4 below */
2296 * Clamp the batch to a 2^n - 1 value. Having a power
2297 * of 2 value was found to be more likely to have
2298 * suboptimal cache aliasing properties in some cases.
2300 * For example if 2 tasks are alternately allocating
2301 * batches of pages, one task can end up with a lot
2302 * of pages of one half of the possible page colors
2303 * and the other with pages of the other colors.
2305 batch = (1 << (fls(batch + batch/2)-1)) - 1;
2310 inline void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
2312 struct per_cpu_pages *pcp;
2314 memset(p, 0, sizeof(*p));
2316 pcp = &p->pcp[0]; /* hot */
2318 pcp->high = 6 * batch;
2319 pcp->batch = max(1UL, 1 * batch);
2320 INIT_LIST_HEAD(&pcp->list);
2322 pcp = &p->pcp[1]; /* cold*/
2324 pcp->high = 2 * batch;
2325 pcp->batch = max(1UL, batch/2);
2326 INIT_LIST_HEAD(&pcp->list);
2330 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
2331 * to the value high for the pageset p.
2334 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
2337 struct per_cpu_pages *pcp;
2339 pcp = &p->pcp[0]; /* hot list */
2341 pcp->batch = max(1UL, high/4);
2342 if ((high/4) > (PAGE_SHIFT * 8))
2343 pcp->batch = PAGE_SHIFT * 8;
2349 * Boot pageset table. One per cpu which is going to be used for all
2350 * zones and all nodes. The parameters will be set in such a way
2351 * that an item put on a list will immediately be handed over to
2352 * the buddy list. This is safe since pageset manipulation is done
2353 * with interrupts disabled.
2355 * Some NUMA counter updates may also be caught by the boot pagesets.
2357 * The boot_pagesets must be kept even after bootup is complete for
2358 * unused processors and/or zones. They do play a role for bootstrapping
2359 * hotplugged processors.
2361 * zoneinfo_show() and maybe other functions do
2362 * not check if the processor is online before following the pageset pointer.
2363 * Other parts of the kernel may not check if the zone is available.
2365 static struct per_cpu_pageset boot_pageset[NR_CPUS];
2368 * Dynamically allocate memory for the
2369 * per cpu pageset array in struct zone.
2371 static int __cpuinit process_zones(int cpu)
2373 struct zone *zone, *dzone;
2374 int node = cpu_to_node(cpu);
2376 node_set_state(node, N_CPU); /* this node has a cpu */
2378 for_each_zone(zone) {
2380 if (!populated_zone(zone))
2383 zone_pcp(zone, cpu) = kmalloc_node(sizeof(struct per_cpu_pageset),
2385 if (!zone_pcp(zone, cpu))
2388 setup_pageset(zone_pcp(zone, cpu), zone_batchsize(zone));
2390 if (percpu_pagelist_fraction)
2391 setup_pagelist_highmark(zone_pcp(zone, cpu),
2392 (zone->present_pages / percpu_pagelist_fraction));
2397 for_each_zone(dzone) {
2398 if (!populated_zone(dzone))
2402 kfree(zone_pcp(dzone, cpu));
2403 zone_pcp(dzone, cpu) = NULL;
2408 static inline void free_zone_pagesets(int cpu)
2412 for_each_zone(zone) {
2413 struct per_cpu_pageset *pset = zone_pcp(zone, cpu);
2415 /* Free per_cpu_pageset if it is slab allocated */
2416 if (pset != &boot_pageset[cpu])
2418 zone_pcp(zone, cpu) = NULL;
2422 static int __cpuinit pageset_cpuup_callback(struct notifier_block *nfb,
2423 unsigned long action,
2426 int cpu = (long)hcpu;
2427 int ret = NOTIFY_OK;
2430 case CPU_UP_PREPARE:
2431 case CPU_UP_PREPARE_FROZEN:
2432 if (process_zones(cpu))
2435 case CPU_UP_CANCELED:
2436 case CPU_UP_CANCELED_FROZEN:
2438 case CPU_DEAD_FROZEN:
2439 free_zone_pagesets(cpu);
2447 static struct notifier_block __cpuinitdata pageset_notifier =
2448 { &pageset_cpuup_callback, NULL, 0 };
2450 void __init setup_per_cpu_pageset(void)
2454 /* Initialize per_cpu_pageset for cpu 0.
2455 * A cpuup callback will do this for every cpu
2456 * as it comes online
2458 err = process_zones(smp_processor_id());
2460 register_cpu_notifier(&pageset_notifier);
2465 static noinline __init_refok
2466 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
2469 struct pglist_data *pgdat = zone->zone_pgdat;
2473 * The per-page waitqueue mechanism uses hashed waitqueues
2476 zone->wait_table_hash_nr_entries =
2477 wait_table_hash_nr_entries(zone_size_pages);
2478 zone->wait_table_bits =
2479 wait_table_bits(zone->wait_table_hash_nr_entries);
2480 alloc_size = zone->wait_table_hash_nr_entries
2481 * sizeof(wait_queue_head_t);
2483 if (system_state == SYSTEM_BOOTING) {
2484 zone->wait_table = (wait_queue_head_t *)
2485 alloc_bootmem_node(pgdat, alloc_size);
2488 * This case means that a zone whose size was 0 gets new memory
2489 * via memory hot-add.
2490 * But it may be the case that a new node was hot-added. In
2491 * this case vmalloc() will not be able to use this new node's
2492 * memory - this wait_table must be initialized to use this new
2493 * node itself as well.
2494 * To use this new node's memory, further consideration will be
2497 zone->wait_table = vmalloc(alloc_size);
2499 if (!zone->wait_table)
2502 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
2503 init_waitqueue_head(zone->wait_table + i);
2508 static __meminit void zone_pcp_init(struct zone *zone)
2511 unsigned long batch = zone_batchsize(zone);
2513 for (cpu = 0; cpu < NR_CPUS; cpu++) {
2515 /* Early boot. Slab allocator not functional yet */
2516 zone_pcp(zone, cpu) = &boot_pageset[cpu];
2517 setup_pageset(&boot_pageset[cpu],0);
2519 setup_pageset(zone_pcp(zone,cpu), batch);
2522 if (zone->present_pages)
2523 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%lu\n",
2524 zone->name, zone->present_pages, batch);
2527 __meminit int init_currently_empty_zone(struct zone *zone,
2528 unsigned long zone_start_pfn,
2530 enum memmap_context context)
2532 struct pglist_data *pgdat = zone->zone_pgdat;
2534 ret = zone_wait_table_init(zone, size);
2537 pgdat->nr_zones = zone_idx(zone) + 1;
2539 zone->zone_start_pfn = zone_start_pfn;
2541 memmap_init(size, pgdat->node_id, zone_idx(zone), zone_start_pfn);
2543 zone_init_free_lists(pgdat, zone, zone->spanned_pages);
2548 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
2550 * Basic iterator support. Return the first range of PFNs for a node
2551 * Note: nid == MAX_NUMNODES returns first region regardless of node
2553 static int __meminit first_active_region_index_in_nid(int nid)
2557 for (i = 0; i < nr_nodemap_entries; i++)
2558 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
2565 * Basic iterator support. Return the next active range of PFNs for a node
2566 * Note: nid == MAX_NUMNODES returns next region regardles of node
2568 static int __meminit next_active_region_index_in_nid(int index, int nid)
2570 for (index = index + 1; index < nr_nodemap_entries; index++)
2571 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
2577 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
2579 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
2580 * Architectures may implement their own version but if add_active_range()
2581 * was used and there are no special requirements, this is a convenient
2584 int __meminit early_pfn_to_nid(unsigned long pfn)
2588 for (i = 0; i < nr_nodemap_entries; i++) {
2589 unsigned long start_pfn = early_node_map[i].start_pfn;
2590 unsigned long end_pfn = early_node_map[i].end_pfn;
2592 if (start_pfn <= pfn && pfn < end_pfn)
2593 return early_node_map[i].nid;
2598 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
2600 /* Basic iterator support to walk early_node_map[] */
2601 #define for_each_active_range_index_in_nid(i, nid) \
2602 for (i = first_active_region_index_in_nid(nid); i != -1; \
2603 i = next_active_region_index_in_nid(i, nid))
2606 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
2607 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
2608 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
2610 * If an architecture guarantees that all ranges registered with
2611 * add_active_ranges() contain no holes and may be freed, this
2612 * this function may be used instead of calling free_bootmem() manually.
2614 void __init free_bootmem_with_active_regions(int nid,
2615 unsigned long max_low_pfn)
2619 for_each_active_range_index_in_nid(i, nid) {
2620 unsigned long size_pages = 0;
2621 unsigned long end_pfn = early_node_map[i].end_pfn;
2623 if (early_node_map[i].start_pfn >= max_low_pfn)
2626 if (end_pfn > max_low_pfn)
2627 end_pfn = max_low_pfn;
2629 size_pages = end_pfn - early_node_map[i].start_pfn;
2630 free_bootmem_node(NODE_DATA(early_node_map[i].nid),
2631 PFN_PHYS(early_node_map[i].start_pfn),
2632 size_pages << PAGE_SHIFT);
2637 * sparse_memory_present_with_active_regions - Call memory_present for each active range
2638 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
2640 * If an architecture guarantees that all ranges registered with
2641 * add_active_ranges() contain no holes and may be freed, this
2642 * function may be used instead of calling memory_present() manually.
2644 void __init sparse_memory_present_with_active_regions(int nid)
2648 for_each_active_range_index_in_nid(i, nid)
2649 memory_present(early_node_map[i].nid,
2650 early_node_map[i].start_pfn,
2651 early_node_map[i].end_pfn);
2655 * push_node_boundaries - Push node boundaries to at least the requested boundary
2656 * @nid: The nid of the node to push the boundary for
2657 * @start_pfn: The start pfn of the node
2658 * @end_pfn: The end pfn of the node
2660 * In reserve-based hot-add, mem_map is allocated that is unused until hotadd
2661 * time. Specifically, on x86_64, SRAT will report ranges that can potentially
2662 * be hotplugged even though no physical memory exists. This function allows
2663 * an arch to push out the node boundaries so mem_map is allocated that can
2666 #ifdef CONFIG_MEMORY_HOTPLUG_RESERVE
2667 void __init push_node_boundaries(unsigned int nid,
2668 unsigned long start_pfn, unsigned long end_pfn)
2670 printk(KERN_DEBUG "Entering push_node_boundaries(%u, %lu, %lu)\n",
2671 nid, start_pfn, end_pfn);
2673 /* Initialise the boundary for this node if necessary */
2674 if (node_boundary_end_pfn[nid] == 0)
2675 node_boundary_start_pfn[nid] = -1UL;
2677 /* Update the boundaries */
2678 if (node_boundary_start_pfn[nid] > start_pfn)
2679 node_boundary_start_pfn[nid] = start_pfn;
2680 if (node_boundary_end_pfn[nid] < end_pfn)
2681 node_boundary_end_pfn[nid] = end_pfn;
2684 /* If necessary, push the node boundary out for reserve hotadd */
2685 static void __meminit account_node_boundary(unsigned int nid,
2686 unsigned long *start_pfn, unsigned long *end_pfn)
2688 printk(KERN_DEBUG "Entering account_node_boundary(%u, %lu, %lu)\n",
2689 nid, *start_pfn, *end_pfn);
2691 /* Return if boundary information has not been provided */
2692 if (node_boundary_end_pfn[nid] == 0)
2695 /* Check the boundaries and update if necessary */
2696 if (node_boundary_start_pfn[nid] < *start_pfn)
2697 *start_pfn = node_boundary_start_pfn[nid];
2698 if (node_boundary_end_pfn[nid] > *end_pfn)
2699 *end_pfn = node_boundary_end_pfn[nid];
2702 void __init push_node_boundaries(unsigned int nid,
2703 unsigned long start_pfn, unsigned long end_pfn) {}
2705 static void __meminit account_node_boundary(unsigned int nid,
2706 unsigned long *start_pfn, unsigned long *end_pfn) {}
2711 * get_pfn_range_for_nid - Return the start and end page frames for a node
2712 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
2713 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
2714 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
2716 * It returns the start and end page frame of a node based on information
2717 * provided by an arch calling add_active_range(). If called for a node
2718 * with no available memory, a warning is printed and the start and end
2721 void __meminit get_pfn_range_for_nid(unsigned int nid,
2722 unsigned long *start_pfn, unsigned long *end_pfn)
2728 for_each_active_range_index_in_nid(i, nid) {
2729 *start_pfn = min(*start_pfn, early_node_map[i].start_pfn);
2730 *end_pfn = max(*end_pfn, early_node_map[i].end_pfn);
2733 if (*start_pfn == -1UL)
2736 /* Push the node boundaries out if requested */
2737 account_node_boundary(nid, start_pfn, end_pfn);
2741 * This finds a zone that can be used for ZONE_MOVABLE pages. The
2742 * assumption is made that zones within a node are ordered in monotonic
2743 * increasing memory addresses so that the "highest" populated zone is used
2745 void __init find_usable_zone_for_movable(void)
2748 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
2749 if (zone_index == ZONE_MOVABLE)
2752 if (arch_zone_highest_possible_pfn[zone_index] >
2753 arch_zone_lowest_possible_pfn[zone_index])
2757 VM_BUG_ON(zone_index == -1);
2758 movable_zone = zone_index;
2762 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
2763 * because it is sized independant of architecture. Unlike the other zones,
2764 * the starting point for ZONE_MOVABLE is not fixed. It may be different
2765 * in each node depending on the size of each node and how evenly kernelcore
2766 * is distributed. This helper function adjusts the zone ranges
2767 * provided by the architecture for a given node by using the end of the
2768 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
2769 * zones within a node are in order of monotonic increases memory addresses
2771 void __meminit adjust_zone_range_for_zone_movable(int nid,
2772 unsigned long zone_type,
2773 unsigned long node_start_pfn,
2774 unsigned long node_end_pfn,
2775 unsigned long *zone_start_pfn,
2776 unsigned long *zone_end_pfn)
2778 /* Only adjust if ZONE_MOVABLE is on this node */
2779 if (zone_movable_pfn[nid]) {
2780 /* Size ZONE_MOVABLE */
2781 if (zone_type == ZONE_MOVABLE) {
2782 *zone_start_pfn = zone_movable_pfn[nid];
2783 *zone_end_pfn = min(node_end_pfn,
2784 arch_zone_highest_possible_pfn[movable_zone]);
2786 /* Adjust for ZONE_MOVABLE starting within this range */
2787 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
2788 *zone_end_pfn > zone_movable_pfn[nid]) {
2789 *zone_end_pfn = zone_movable_pfn[nid];
2791 /* Check if this whole range is within ZONE_MOVABLE */
2792 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
2793 *zone_start_pfn = *zone_end_pfn;
2798 * Return the number of pages a zone spans in a node, including holes
2799 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
2801 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
2802 unsigned long zone_type,
2803 unsigned long *ignored)
2805 unsigned long node_start_pfn, node_end_pfn;
2806 unsigned long zone_start_pfn, zone_end_pfn;
2808 /* Get the start and end of the node and zone */
2809 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
2810 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
2811 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
2812 adjust_zone_range_for_zone_movable(nid, zone_type,
2813 node_start_pfn, node_end_pfn,
2814 &zone_start_pfn, &zone_end_pfn);
2816 /* Check that this node has pages within the zone's required range */
2817 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
2820 /* Move the zone boundaries inside the node if necessary */
2821 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
2822 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
2824 /* Return the spanned pages */
2825 return zone_end_pfn - zone_start_pfn;
2829 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
2830 * then all holes in the requested range will be accounted for.
2832 unsigned long __meminit __absent_pages_in_range(int nid,
2833 unsigned long range_start_pfn,
2834 unsigned long range_end_pfn)
2837 unsigned long prev_end_pfn = 0, hole_pages = 0;
2838 unsigned long start_pfn;
2840 /* Find the end_pfn of the first active range of pfns in the node */
2841 i = first_active_region_index_in_nid(nid);
2845 prev_end_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
2847 /* Account for ranges before physical memory on this node */
2848 if (early_node_map[i].start_pfn > range_start_pfn)
2849 hole_pages = prev_end_pfn - range_start_pfn;
2851 /* Find all holes for the zone within the node */
2852 for (; i != -1; i = next_active_region_index_in_nid(i, nid)) {
2854 /* No need to continue if prev_end_pfn is outside the zone */
2855 if (prev_end_pfn >= range_end_pfn)
2858 /* Make sure the end of the zone is not within the hole */
2859 start_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
2860 prev_end_pfn = max(prev_end_pfn, range_start_pfn);
2862 /* Update the hole size cound and move on */
2863 if (start_pfn > range_start_pfn) {
2864 BUG_ON(prev_end_pfn > start_pfn);
2865 hole_pages += start_pfn - prev_end_pfn;
2867 prev_end_pfn = early_node_map[i].end_pfn;
2870 /* Account for ranges past physical memory on this node */
2871 if (range_end_pfn > prev_end_pfn)
2872 hole_pages += range_end_pfn -
2873 max(range_start_pfn, prev_end_pfn);
2879 * absent_pages_in_range - Return number of page frames in holes within a range
2880 * @start_pfn: The start PFN to start searching for holes
2881 * @end_pfn: The end PFN to stop searching for holes
2883 * It returns the number of pages frames in memory holes within a range.
2885 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
2886 unsigned long end_pfn)
2888 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
2891 /* Return the number of page frames in holes in a zone on a node */
2892 static unsigned long __meminit zone_absent_pages_in_node(int nid,
2893 unsigned long zone_type,
2894 unsigned long *ignored)
2896 unsigned long node_start_pfn, node_end_pfn;
2897 unsigned long zone_start_pfn, zone_end_pfn;
2899 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
2900 zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type],
2902 zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type],
2905 adjust_zone_range_for_zone_movable(nid, zone_type,
2906 node_start_pfn, node_end_pfn,
2907 &zone_start_pfn, &zone_end_pfn);
2908 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
2912 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
2913 unsigned long zone_type,
2914 unsigned long *zones_size)
2916 return zones_size[zone_type];
2919 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
2920 unsigned long zone_type,
2921 unsigned long *zholes_size)
2926 return zholes_size[zone_type];
2931 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
2932 unsigned long *zones_size, unsigned long *zholes_size)
2934 unsigned long realtotalpages, totalpages = 0;
2937 for (i = 0; i < MAX_NR_ZONES; i++)
2938 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
2940 pgdat->node_spanned_pages = totalpages;
2942 realtotalpages = totalpages;
2943 for (i = 0; i < MAX_NR_ZONES; i++)
2945 zone_absent_pages_in_node(pgdat->node_id, i,
2947 pgdat->node_present_pages = realtotalpages;
2948 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
2953 * Set up the zone data structures:
2954 * - mark all pages reserved
2955 * - mark all memory queues empty
2956 * - clear the memory bitmaps
2958 static void __meminit free_area_init_core(struct pglist_data *pgdat,
2959 unsigned long *zones_size, unsigned long *zholes_size)
2962 int nid = pgdat->node_id;
2963 unsigned long zone_start_pfn = pgdat->node_start_pfn;
2966 pgdat_resize_init(pgdat);
2967 pgdat->nr_zones = 0;
2968 init_waitqueue_head(&pgdat->kswapd_wait);
2969 pgdat->kswapd_max_order = 0;
2971 for (j = 0; j < MAX_NR_ZONES; j++) {
2972 struct zone *zone = pgdat->node_zones + j;
2973 unsigned long size, realsize, memmap_pages;
2975 size = zone_spanned_pages_in_node(nid, j, zones_size);
2976 realsize = size - zone_absent_pages_in_node(nid, j,
2980 * Adjust realsize so that it accounts for how much memory
2981 * is used by this zone for memmap. This affects the watermark
2982 * and per-cpu initialisations
2984 memmap_pages = (size * sizeof(struct page)) >> PAGE_SHIFT;
2985 if (realsize >= memmap_pages) {
2986 realsize -= memmap_pages;
2988 " %s zone: %lu pages used for memmap\n",
2989 zone_names[j], memmap_pages);
2992 " %s zone: %lu pages exceeds realsize %lu\n",
2993 zone_names[j], memmap_pages, realsize);
2995 /* Account for reserved pages */
2996 if (j == 0 && realsize > dma_reserve) {
2997 realsize -= dma_reserve;
2998 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
2999 zone_names[0], dma_reserve);
3002 if (!is_highmem_idx(j))
3003 nr_kernel_pages += realsize;
3004 nr_all_pages += realsize;
3006 zone->spanned_pages = size;
3007 zone->present_pages = realsize;
3010 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
3012 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
3014 zone->name = zone_names[j];
3015 spin_lock_init(&zone->lock);
3016 spin_lock_init(&zone->lru_lock);
3017 zone_seqlock_init(zone);
3018 zone->zone_pgdat = pgdat;
3020 zone->prev_priority = DEF_PRIORITY;
3022 zone_pcp_init(zone);
3023 INIT_LIST_HEAD(&zone->active_list);
3024 INIT_LIST_HEAD(&zone->inactive_list);
3025 zone->nr_scan_active = 0;
3026 zone->nr_scan_inactive = 0;
3027 zap_zone_vm_stats(zone);
3028 atomic_set(&zone->reclaim_in_progress, 0);
3032 ret = init_currently_empty_zone(zone, zone_start_pfn,
3033 size, MEMMAP_EARLY);
3035 zone_start_pfn += size;
3039 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
3041 /* Skip empty nodes */
3042 if (!pgdat->node_spanned_pages)
3045 #ifdef CONFIG_FLAT_NODE_MEM_MAP
3046 /* ia64 gets its own node_mem_map, before this, without bootmem */
3047 if (!pgdat->node_mem_map) {
3048 unsigned long size, start, end;
3052 * The zone's endpoints aren't required to be MAX_ORDER
3053 * aligned but the node_mem_map endpoints must be in order
3054 * for the buddy allocator to function correctly.
3056 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
3057 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
3058 end = ALIGN(end, MAX_ORDER_NR_PAGES);
3059 size = (end - start) * sizeof(struct page);
3060 map = alloc_remap(pgdat->node_id, size);
3062 map = alloc_bootmem_node(pgdat, size);
3063 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
3065 #ifndef CONFIG_NEED_MULTIPLE_NODES
3067 * With no DISCONTIG, the global mem_map is just set as node 0's
3069 if (pgdat == NODE_DATA(0)) {
3070 mem_map = NODE_DATA(0)->node_mem_map;
3071 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3072 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
3073 mem_map -= pgdat->node_start_pfn;
3074 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
3077 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
3080 void __meminit free_area_init_node(int nid, struct pglist_data *pgdat,
3081 unsigned long *zones_size, unsigned long node_start_pfn,
3082 unsigned long *zholes_size)
3084 pgdat->node_id = nid;
3085 pgdat->node_start_pfn = node_start_pfn;
3086 calculate_node_totalpages(pgdat, zones_size, zholes_size);
3088 alloc_node_mem_map(pgdat);
3090 free_area_init_core(pgdat, zones_size, zholes_size);
3093 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3095 #if MAX_NUMNODES > 1
3097 * Figure out the number of possible node ids.
3099 static void __init setup_nr_node_ids(void)
3102 unsigned int highest = 0;
3104 for_each_node_mask(node, node_possible_map)
3106 nr_node_ids = highest + 1;
3109 static inline void setup_nr_node_ids(void)
3115 * add_active_range - Register a range of PFNs backed by physical memory
3116 * @nid: The node ID the range resides on
3117 * @start_pfn: The start PFN of the available physical memory
3118 * @end_pfn: The end PFN of the available physical memory
3120 * These ranges are stored in an early_node_map[] and later used by
3121 * free_area_init_nodes() to calculate zone sizes and holes. If the
3122 * range spans a memory hole, it is up to the architecture to ensure
3123 * the memory is not freed by the bootmem allocator. If possible
3124 * the range being registered will be merged with existing ranges.
3126 void __init add_active_range(unsigned int nid, unsigned long start_pfn,
3127 unsigned long end_pfn)
3131 printk(KERN_DEBUG "Entering add_active_range(%d, %lu, %lu) "
3132 "%d entries of %d used\n",
3133 nid, start_pfn, end_pfn,
3134 nr_nodemap_entries, MAX_ACTIVE_REGIONS);
3136 /* Merge with existing active regions if possible */
3137 for (i = 0; i < nr_nodemap_entries; i++) {
3138 if (early_node_map[i].nid != nid)
3141 /* Skip if an existing region covers this new one */
3142 if (start_pfn >= early_node_map[i].start_pfn &&
3143 end_pfn <= early_node_map[i].end_pfn)
3146 /* Merge forward if suitable */
3147 if (start_pfn <= early_node_map[i].end_pfn &&
3148 end_pfn > early_node_map[i].end_pfn) {
3149 early_node_map[i].end_pfn = end_pfn;
3153 /* Merge backward if suitable */
3154 if (start_pfn < early_node_map[i].end_pfn &&
3155 end_pfn >= early_node_map[i].start_pfn) {
3156 early_node_map[i].start_pfn = start_pfn;
3161 /* Check that early_node_map is large enough */
3162 if (i >= MAX_ACTIVE_REGIONS) {
3163 printk(KERN_CRIT "More than %d memory regions, truncating\n",
3164 MAX_ACTIVE_REGIONS);
3168 early_node_map[i].nid = nid;
3169 early_node_map[i].start_pfn = start_pfn;
3170 early_node_map[i].end_pfn = end_pfn;
3171 nr_nodemap_entries = i + 1;
3175 * shrink_active_range - Shrink an existing registered range of PFNs
3176 * @nid: The node id the range is on that should be shrunk
3177 * @old_end_pfn: The old end PFN of the range
3178 * @new_end_pfn: The new PFN of the range
3180 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
3181 * The map is kept at the end physical page range that has already been
3182 * registered with add_active_range(). This function allows an arch to shrink
3183 * an existing registered range.
3185 void __init shrink_active_range(unsigned int nid, unsigned long old_end_pfn,
3186 unsigned long new_end_pfn)
3190 /* Find the old active region end and shrink */
3191 for_each_active_range_index_in_nid(i, nid)
3192 if (early_node_map[i].end_pfn == old_end_pfn) {
3193 early_node_map[i].end_pfn = new_end_pfn;
3199 * remove_all_active_ranges - Remove all currently registered regions
3201 * During discovery, it may be found that a table like SRAT is invalid
3202 * and an alternative discovery method must be used. This function removes
3203 * all currently registered regions.
3205 void __init remove_all_active_ranges(void)
3207 memset(early_node_map, 0, sizeof(early_node_map));
3208 nr_nodemap_entries = 0;
3209 #ifdef CONFIG_MEMORY_HOTPLUG_RESERVE
3210 memset(node_boundary_start_pfn, 0, sizeof(node_boundary_start_pfn));
3211 memset(node_boundary_end_pfn, 0, sizeof(node_boundary_end_pfn));
3212 #endif /* CONFIG_MEMORY_HOTPLUG_RESERVE */
3215 /* Compare two active node_active_regions */
3216 static int __init cmp_node_active_region(const void *a, const void *b)
3218 struct node_active_region *arange = (struct node_active_region *)a;
3219 struct node_active_region *brange = (struct node_active_region *)b;
3221 /* Done this way to avoid overflows */
3222 if (arange->start_pfn > brange->start_pfn)
3224 if (arange->start_pfn < brange->start_pfn)
3230 /* sort the node_map by start_pfn */
3231 static void __init sort_node_map(void)
3233 sort(early_node_map, (size_t)nr_nodemap_entries,
3234 sizeof(struct node_active_region),
3235 cmp_node_active_region, NULL);
3238 /* Find the lowest pfn for a node */
3239 unsigned long __init find_min_pfn_for_node(unsigned long nid)
3242 unsigned long min_pfn = ULONG_MAX;
3244 /* Assuming a sorted map, the first range found has the starting pfn */
3245 for_each_active_range_index_in_nid(i, nid)
3246 min_pfn = min(min_pfn, early_node_map[i].start_pfn);
3248 if (min_pfn == ULONG_MAX) {
3250 "Could not find start_pfn for node %lu\n", nid);
3258 * find_min_pfn_with_active_regions - Find the minimum PFN registered
3260 * It returns the minimum PFN based on information provided via
3261 * add_active_range().
3263 unsigned long __init find_min_pfn_with_active_regions(void)
3265 return find_min_pfn_for_node(MAX_NUMNODES);
3269 * find_max_pfn_with_active_regions - Find the maximum PFN registered
3271 * It returns the maximum PFN based on information provided via
3272 * add_active_range().
3274 unsigned long __init find_max_pfn_with_active_regions(void)
3277 unsigned long max_pfn = 0;
3279 for (i = 0; i < nr_nodemap_entries; i++)
3280 max_pfn = max(max_pfn, early_node_map[i].end_pfn);
3285 unsigned long __init early_calculate_totalpages(void)
3288 unsigned long totalpages = 0;
3290 for (i = 0; i < nr_nodemap_entries; i++)
3291 totalpages += early_node_map[i].end_pfn -
3292 early_node_map[i].start_pfn;
3298 * Find the PFN the Movable zone begins in each node. Kernel memory
3299 * is spread evenly between nodes as long as the nodes have enough
3300 * memory. When they don't, some nodes will have more kernelcore than
3303 void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn)
3306 unsigned long usable_startpfn;
3307 unsigned long kernelcore_node, kernelcore_remaining;
3308 int usable_nodes = num_online_nodes();
3311 * If movablecore was specified, calculate what size of
3312 * kernelcore that corresponds so that memory usable for
3313 * any allocation type is evenly spread. If both kernelcore
3314 * and movablecore are specified, then the value of kernelcore
3315 * will be used for required_kernelcore if it's greater than
3316 * what movablecore would have allowed.
3318 if (required_movablecore) {
3319 unsigned long totalpages = early_calculate_totalpages();
3320 unsigned long corepages;
3323 * Round-up so that ZONE_MOVABLE is at least as large as what
3324 * was requested by the user
3326 required_movablecore =
3327 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
3328 corepages = totalpages - required_movablecore;
3330 required_kernelcore = max(required_kernelcore, corepages);
3333 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
3334 if (!required_kernelcore)
3337 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
3338 find_usable_zone_for_movable();
3339 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
3342 /* Spread kernelcore memory as evenly as possible throughout nodes */
3343 kernelcore_node = required_kernelcore / usable_nodes;
3344 for_each_online_node(nid) {
3346 * Recalculate kernelcore_node if the division per node
3347 * now exceeds what is necessary to satisfy the requested
3348 * amount of memory for the kernel
3350 if (required_kernelcore < kernelcore_node)
3351 kernelcore_node = required_kernelcore / usable_nodes;
3354 * As the map is walked, we track how much memory is usable
3355 * by the kernel using kernelcore_remaining. When it is
3356 * 0, the rest of the node is usable by ZONE_MOVABLE
3358 kernelcore_remaining = kernelcore_node;
3360 /* Go through each range of PFNs within this node */
3361 for_each_active_range_index_in_nid(i, nid) {
3362 unsigned long start_pfn, end_pfn;
3363 unsigned long size_pages;
3365 start_pfn = max(early_node_map[i].start_pfn,
3366 zone_movable_pfn[nid]);
3367 end_pfn = early_node_map[i].end_pfn;
3368 if (start_pfn >= end_pfn)
3371 /* Account for what is only usable for kernelcore */
3372 if (start_pfn < usable_startpfn) {
3373 unsigned long kernel_pages;
3374 kernel_pages = min(end_pfn, usable_startpfn)
3377 kernelcore_remaining -= min(kernel_pages,
3378 kernelcore_remaining);
3379 required_kernelcore -= min(kernel_pages,
3380 required_kernelcore);
3382 /* Continue if range is now fully accounted */
3383 if (end_pfn <= usable_startpfn) {
3386 * Push zone_movable_pfn to the end so
3387 * that if we have to rebalance
3388 * kernelcore across nodes, we will
3389 * not double account here
3391 zone_movable_pfn[nid] = end_pfn;
3394 start_pfn = usable_startpfn;
3398 * The usable PFN range for ZONE_MOVABLE is from
3399 * start_pfn->end_pfn. Calculate size_pages as the
3400 * number of pages used as kernelcore
3402 size_pages = end_pfn - start_pfn;
3403 if (size_pages > kernelcore_remaining)
3404 size_pages = kernelcore_remaining;
3405 zone_movable_pfn[nid] = start_pfn + size_pages;
3408 * Some kernelcore has been met, update counts and
3409 * break if the kernelcore for this node has been
3412 required_kernelcore -= min(required_kernelcore,
3414 kernelcore_remaining -= size_pages;
3415 if (!kernelcore_remaining)
3421 * If there is still required_kernelcore, we do another pass with one
3422 * less node in the count. This will push zone_movable_pfn[nid] further
3423 * along on the nodes that still have memory until kernelcore is
3427 if (usable_nodes && required_kernelcore > usable_nodes)
3430 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
3431 for (nid = 0; nid < MAX_NUMNODES; nid++)
3432 zone_movable_pfn[nid] =
3433 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
3437 * free_area_init_nodes - Initialise all pg_data_t and zone data
3438 * @max_zone_pfn: an array of max PFNs for each zone
3440 * This will call free_area_init_node() for each active node in the system.
3441 * Using the page ranges provided by add_active_range(), the size of each
3442 * zone in each node and their holes is calculated. If the maximum PFN
3443 * between two adjacent zones match, it is assumed that the zone is empty.
3444 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
3445 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
3446 * starts where the previous one ended. For example, ZONE_DMA32 starts
3447 * at arch_max_dma_pfn.
3449 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
3454 /* Sort early_node_map as initialisation assumes it is sorted */
3457 /* Record where the zone boundaries are */
3458 memset(arch_zone_lowest_possible_pfn, 0,
3459 sizeof(arch_zone_lowest_possible_pfn));
3460 memset(arch_zone_highest_possible_pfn, 0,
3461 sizeof(arch_zone_highest_possible_pfn));
3462 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
3463 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
3464 for (i = 1; i < MAX_NR_ZONES; i++) {
3465 if (i == ZONE_MOVABLE)
3467 arch_zone_lowest_possible_pfn[i] =
3468 arch_zone_highest_possible_pfn[i-1];
3469 arch_zone_highest_possible_pfn[i] =
3470 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
3472 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
3473 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
3475 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
3476 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
3477 find_zone_movable_pfns_for_nodes(zone_movable_pfn);
3479 /* Print out the zone ranges */
3480 printk("Zone PFN ranges:\n");
3481 for (i = 0; i < MAX_NR_ZONES; i++) {
3482 if (i == ZONE_MOVABLE)
3484 printk(" %-8s %8lu -> %8lu\n",
3486 arch_zone_lowest_possible_pfn[i],
3487 arch_zone_highest_possible_pfn[i]);
3490 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
3491 printk("Movable zone start PFN for each node\n");
3492 for (i = 0; i < MAX_NUMNODES; i++) {
3493 if (zone_movable_pfn[i])
3494 printk(" Node %d: %lu\n", i, zone_movable_pfn[i]);
3497 /* Print out the early_node_map[] */
3498 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries);
3499 for (i = 0; i < nr_nodemap_entries; i++)
3500 printk(" %3d: %8lu -> %8lu\n", early_node_map[i].nid,
3501 early_node_map[i].start_pfn,
3502 early_node_map[i].end_pfn);
3504 /* Initialise every node */
3505 setup_nr_node_ids();
3506 for_each_online_node(nid) {
3507 pg_data_t *pgdat = NODE_DATA(nid);
3508 free_area_init_node(nid, pgdat, NULL,
3509 find_min_pfn_for_node(nid), NULL);
3513 static int __init cmdline_parse_core(char *p, unsigned long *core)
3515 unsigned long long coremem;
3519 coremem = memparse(p, &p);
3520 *core = coremem >> PAGE_SHIFT;
3522 /* Paranoid check that UL is enough for the coremem value */
3523 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
3529 * kernelcore=size sets the amount of memory for use for allocations that
3530 * cannot be reclaimed or migrated.
3532 static int __init cmdline_parse_kernelcore(char *p)
3534 return cmdline_parse_core(p, &required_kernelcore);
3538 * movablecore=size sets the amount of memory for use for allocations that
3539 * can be reclaimed or migrated.
3541 static int __init cmdline_parse_movablecore(char *p)
3543 return cmdline_parse_core(p, &required_movablecore);
3546 early_param("kernelcore", cmdline_parse_kernelcore);
3547 early_param("movablecore", cmdline_parse_movablecore);
3549 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
3552 * set_dma_reserve - set the specified number of pages reserved in the first zone
3553 * @new_dma_reserve: The number of pages to mark reserved
3555 * The per-cpu batchsize and zone watermarks are determined by present_pages.
3556 * In the DMA zone, a significant percentage may be consumed by kernel image
3557 * and other unfreeable allocations which can skew the watermarks badly. This
3558 * function may optionally be used to account for unfreeable pages in the
3559 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
3560 * smaller per-cpu batchsize.
3562 void __init set_dma_reserve(unsigned long new_dma_reserve)
3564 dma_reserve = new_dma_reserve;
3567 #ifndef CONFIG_NEED_MULTIPLE_NODES
3568 static bootmem_data_t contig_bootmem_data;
3569 struct pglist_data contig_page_data = { .bdata = &contig_bootmem_data };
3571 EXPORT_SYMBOL(contig_page_data);
3574 void __init free_area_init(unsigned long *zones_size)
3576 free_area_init_node(0, NODE_DATA(0), zones_size,
3577 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
3580 static int page_alloc_cpu_notify(struct notifier_block *self,
3581 unsigned long action, void *hcpu)
3583 int cpu = (unsigned long)hcpu;
3585 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
3586 local_irq_disable();
3588 vm_events_fold_cpu(cpu);
3590 refresh_cpu_vm_stats(cpu);
3595 void __init page_alloc_init(void)
3597 hotcpu_notifier(page_alloc_cpu_notify, 0);
3601 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
3602 * or min_free_kbytes changes.
3604 static void calculate_totalreserve_pages(void)
3606 struct pglist_data *pgdat;
3607 unsigned long reserve_pages = 0;
3608 enum zone_type i, j;
3610 for_each_online_pgdat(pgdat) {
3611 for (i = 0; i < MAX_NR_ZONES; i++) {
3612 struct zone *zone = pgdat->node_zones + i;
3613 unsigned long max = 0;
3615 /* Find valid and maximum lowmem_reserve in the zone */
3616 for (j = i; j < MAX_NR_ZONES; j++) {
3617 if (zone->lowmem_reserve[j] > max)
3618 max = zone->lowmem_reserve[j];
3621 /* we treat pages_high as reserved pages. */
3622 max += zone->pages_high;
3624 if (max > zone->present_pages)
3625 max = zone->present_pages;
3626 reserve_pages += max;
3629 totalreserve_pages = reserve_pages;
3633 * setup_per_zone_lowmem_reserve - called whenever
3634 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
3635 * has a correct pages reserved value, so an adequate number of
3636 * pages are left in the zone after a successful __alloc_pages().
3638 static void setup_per_zone_lowmem_reserve(void)
3640 struct pglist_data *pgdat;
3641 enum zone_type j, idx;
3643 for_each_online_pgdat(pgdat) {
3644 for (j = 0; j < MAX_NR_ZONES; j++) {
3645 struct zone *zone = pgdat->node_zones + j;
3646 unsigned long present_pages = zone->present_pages;
3648 zone->lowmem_reserve[j] = 0;
3652 struct zone *lower_zone;
3656 if (sysctl_lowmem_reserve_ratio[idx] < 1)
3657 sysctl_lowmem_reserve_ratio[idx] = 1;
3659 lower_zone = pgdat->node_zones + idx;
3660 lower_zone->lowmem_reserve[j] = present_pages /
3661 sysctl_lowmem_reserve_ratio[idx];
3662 present_pages += lower_zone->present_pages;
3667 /* update totalreserve_pages */
3668 calculate_totalreserve_pages();
3672 * setup_per_zone_pages_min - called when min_free_kbytes changes.
3674 * Ensures that the pages_{min,low,high} values for each zone are set correctly
3675 * with respect to min_free_kbytes.
3677 void setup_per_zone_pages_min(void)
3679 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
3680 unsigned long lowmem_pages = 0;
3682 unsigned long flags;
3684 /* Calculate total number of !ZONE_HIGHMEM pages */
3685 for_each_zone(zone) {
3686 if (!is_highmem(zone))
3687 lowmem_pages += zone->present_pages;
3690 for_each_zone(zone) {
3693 spin_lock_irqsave(&zone->lru_lock, flags);
3694 tmp = (u64)pages_min * zone->present_pages;
3695 do_div(tmp, lowmem_pages);
3696 if (is_highmem(zone)) {
3698 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
3699 * need highmem pages, so cap pages_min to a small
3702 * The (pages_high-pages_low) and (pages_low-pages_min)
3703 * deltas controls asynch page reclaim, and so should
3704 * not be capped for highmem.
3708 min_pages = zone->present_pages / 1024;
3709 if (min_pages < SWAP_CLUSTER_MAX)
3710 min_pages = SWAP_CLUSTER_MAX;
3711 if (min_pages > 128)
3713 zone->pages_min = min_pages;
3716 * If it's a lowmem zone, reserve a number of pages
3717 * proportionate to the zone's size.
3719 zone->pages_min = tmp;
3722 zone->pages_low = zone->pages_min + (tmp >> 2);
3723 zone->pages_high = zone->pages_min + (tmp >> 1);
3724 spin_unlock_irqrestore(&zone->lru_lock, flags);
3727 /* update totalreserve_pages */
3728 calculate_totalreserve_pages();
3732 * Initialise min_free_kbytes.
3734 * For small machines we want it small (128k min). For large machines
3735 * we want it large (64MB max). But it is not linear, because network
3736 * bandwidth does not increase linearly with machine size. We use
3738 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
3739 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
3755 static int __init init_per_zone_pages_min(void)
3757 unsigned long lowmem_kbytes;
3759 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
3761 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
3762 if (min_free_kbytes < 128)
3763 min_free_kbytes = 128;
3764 if (min_free_kbytes > 65536)
3765 min_free_kbytes = 65536;
3766 setup_per_zone_pages_min();
3767 setup_per_zone_lowmem_reserve();
3770 module_init(init_per_zone_pages_min)
3773 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
3774 * that we can call two helper functions whenever min_free_kbytes
3777 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
3778 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
3780 proc_dointvec(table, write, file, buffer, length, ppos);
3782 setup_per_zone_pages_min();
3787 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
3788 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
3793 rc = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
3798 zone->min_unmapped_pages = (zone->present_pages *
3799 sysctl_min_unmapped_ratio) / 100;
3803 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
3804 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
3809 rc = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
3814 zone->min_slab_pages = (zone->present_pages *
3815 sysctl_min_slab_ratio) / 100;
3821 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
3822 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
3823 * whenever sysctl_lowmem_reserve_ratio changes.
3825 * The reserve ratio obviously has absolutely no relation with the
3826 * pages_min watermarks. The lowmem reserve ratio can only make sense
3827 * if in function of the boot time zone sizes.
3829 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
3830 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
3832 proc_dointvec_minmax(table, write, file, buffer, length, ppos);
3833 setup_per_zone_lowmem_reserve();
3838 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
3839 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
3840 * can have before it gets flushed back to buddy allocator.
3843 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
3844 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
3850 ret = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
3851 if (!write || (ret == -EINVAL))
3853 for_each_zone(zone) {
3854 for_each_online_cpu(cpu) {
3856 high = zone->present_pages / percpu_pagelist_fraction;
3857 setup_pagelist_highmark(zone_pcp(zone, cpu), high);
3863 int hashdist = HASHDIST_DEFAULT;
3866 static int __init set_hashdist(char *str)
3870 hashdist = simple_strtoul(str, &str, 0);
3873 __setup("hashdist=", set_hashdist);
3877 * allocate a large system hash table from bootmem
3878 * - it is assumed that the hash table must contain an exact power-of-2
3879 * quantity of entries
3880 * - limit is the number of hash buckets, not the total allocation size
3882 void *__init alloc_large_system_hash(const char *tablename,
3883 unsigned long bucketsize,
3884 unsigned long numentries,
3887 unsigned int *_hash_shift,
3888 unsigned int *_hash_mask,
3889 unsigned long limit)
3891 unsigned long long max = limit;
3892 unsigned long log2qty, size;
3895 /* allow the kernel cmdline to have a say */
3897 /* round applicable memory size up to nearest megabyte */
3898 numentries = nr_kernel_pages;
3899 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
3900 numentries >>= 20 - PAGE_SHIFT;
3901 numentries <<= 20 - PAGE_SHIFT;
3903 /* limit to 1 bucket per 2^scale bytes of low memory */
3904 if (scale > PAGE_SHIFT)
3905 numentries >>= (scale - PAGE_SHIFT);
3907 numentries <<= (PAGE_SHIFT - scale);
3909 /* Make sure we've got at least a 0-order allocation.. */
3910 if (unlikely((numentries * bucketsize) < PAGE_SIZE))
3911 numentries = PAGE_SIZE / bucketsize;
3913 numentries = roundup_pow_of_two(numentries);
3915 /* limit allocation size to 1/16 total memory by default */
3917 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
3918 do_div(max, bucketsize);
3921 if (numentries > max)
3924 log2qty = ilog2(numentries);
3927 size = bucketsize << log2qty;
3928 if (flags & HASH_EARLY)
3929 table = alloc_bootmem(size);
3931 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
3933 unsigned long order;
3934 for (order = 0; ((1UL << order) << PAGE_SHIFT) < size; order++)
3936 table = (void*) __get_free_pages(GFP_ATOMIC, order);
3938 * If bucketsize is not a power-of-two, we may free
3939 * some pages at the end of hash table.
3942 unsigned long alloc_end = (unsigned long)table +
3943 (PAGE_SIZE << order);
3944 unsigned long used = (unsigned long)table +
3946 split_page(virt_to_page(table), order);
3947 while (used < alloc_end) {
3953 } while (!table && size > PAGE_SIZE && --log2qty);
3956 panic("Failed to allocate %s hash table\n", tablename);
3958 printk(KERN_INFO "%s hash table entries: %d (order: %d, %lu bytes)\n",
3961 ilog2(size) - PAGE_SHIFT,
3965 *_hash_shift = log2qty;
3967 *_hash_mask = (1 << log2qty) - 1;
3972 #ifdef CONFIG_OUT_OF_LINE_PFN_TO_PAGE
3973 struct page *pfn_to_page(unsigned long pfn)
3975 return __pfn_to_page(pfn);
3977 unsigned long page_to_pfn(struct page *page)
3979 return __page_to_pfn(page);
3981 EXPORT_SYMBOL(pfn_to_page);
3982 EXPORT_SYMBOL(page_to_pfn);
3983 #endif /* CONFIG_OUT_OF_LINE_PFN_TO_PAGE */