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 * MCD - HACK: Find somewhere to initialize this EARLY, or make this
53 nodemask_t node_online_map __read_mostly = { { [0] = 1UL } };
54 EXPORT_SYMBOL(node_online_map);
55 nodemask_t node_possible_map __read_mostly = NODE_MASK_ALL;
56 EXPORT_SYMBOL(node_possible_map);
57 unsigned long totalram_pages __read_mostly;
58 unsigned long totalreserve_pages __read_mostly;
60 int percpu_pagelist_fraction;
62 static void __free_pages_ok(struct page *page, unsigned int order);
65 * results with 256, 32 in the lowmem_reserve sysctl:
66 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
67 * 1G machine -> (16M dma, 784M normal, 224M high)
68 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
69 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
70 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
72 * TBD: should special case ZONE_DMA32 machines here - in those we normally
73 * don't need any ZONE_NORMAL reservation
75 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
77 #ifdef CONFIG_ZONE_DMA32
85 EXPORT_SYMBOL(totalram_pages);
87 static char * const zone_names[MAX_NR_ZONES] = {
89 #ifdef CONFIG_ZONE_DMA32
98 int min_free_kbytes = 1024;
100 unsigned long __meminitdata nr_kernel_pages;
101 unsigned long __meminitdata nr_all_pages;
102 static unsigned long __initdata dma_reserve;
104 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
106 * MAX_ACTIVE_REGIONS determines the maxmimum number of distinct
107 * ranges of memory (RAM) that may be registered with add_active_range().
108 * Ranges passed to add_active_range() will be merged if possible
109 * so the number of times add_active_range() can be called is
110 * related to the number of nodes and the number of holes
112 #ifdef CONFIG_MAX_ACTIVE_REGIONS
113 /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */
114 #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS
116 #if MAX_NUMNODES >= 32
117 /* If there can be many nodes, allow up to 50 holes per node */
118 #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50)
120 /* By default, allow up to 256 distinct regions */
121 #define MAX_ACTIVE_REGIONS 256
125 struct node_active_region __initdata early_node_map[MAX_ACTIVE_REGIONS];
126 int __initdata nr_nodemap_entries;
127 unsigned long __initdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
128 unsigned long __initdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
129 #ifdef CONFIG_MEMORY_HOTPLUG_RESERVE
130 unsigned long __initdata node_boundary_start_pfn[MAX_NUMNODES];
131 unsigned long __initdata node_boundary_end_pfn[MAX_NUMNODES];
132 #endif /* CONFIG_MEMORY_HOTPLUG_RESERVE */
133 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
135 #ifdef CONFIG_DEBUG_VM
136 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
140 unsigned long pfn = page_to_pfn(page);
143 seq = zone_span_seqbegin(zone);
144 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
146 else if (pfn < zone->zone_start_pfn)
148 } while (zone_span_seqretry(zone, seq));
153 static int page_is_consistent(struct zone *zone, struct page *page)
155 #ifdef CONFIG_HOLES_IN_ZONE
156 if (!pfn_valid(page_to_pfn(page)))
159 if (zone != page_zone(page))
165 * Temporary debugging check for pages not lying within a given zone.
167 static int bad_range(struct zone *zone, struct page *page)
169 if (page_outside_zone_boundaries(zone, page))
171 if (!page_is_consistent(zone, page))
177 static inline int bad_range(struct zone *zone, struct page *page)
183 static void bad_page(struct page *page)
185 printk(KERN_EMERG "Bad page state in process '%s'\n"
186 KERN_EMERG "page:%p flags:0x%0*lx mapping:%p mapcount:%d count:%d\n"
187 KERN_EMERG "Trying to fix it up, but a reboot is needed\n"
188 KERN_EMERG "Backtrace:\n",
189 current->comm, page, (int)(2*sizeof(unsigned long)),
190 (unsigned long)page->flags, page->mapping,
191 page_mapcount(page), page_count(page));
193 page->flags &= ~(1 << PG_lru |
203 set_page_count(page, 0);
204 reset_page_mapcount(page);
205 page->mapping = NULL;
206 add_taint(TAINT_BAD_PAGE);
210 * Higher-order pages are called "compound pages". They are structured thusly:
212 * The first PAGE_SIZE page is called the "head page".
214 * The remaining PAGE_SIZE pages are called "tail pages".
216 * All pages have PG_compound set. All pages have their ->private pointing at
217 * the head page (even the head page has this).
219 * The first tail page's ->lru.next holds the address of the compound page's
220 * put_page() function. Its ->lru.prev holds the order of allocation.
221 * This usage means that zero-order pages may not be compound.
224 static void free_compound_page(struct page *page)
226 __free_pages_ok(page, (unsigned long)page[1].lru.prev);
229 static void prep_compound_page(struct page *page, unsigned long order)
232 int nr_pages = 1 << order;
234 set_compound_page_dtor(page, free_compound_page);
235 page[1].lru.prev = (void *)order;
236 for (i = 0; i < nr_pages; i++) {
237 struct page *p = page + i;
239 __SetPageCompound(p);
240 set_page_private(p, (unsigned long)page);
244 static void destroy_compound_page(struct page *page, unsigned long order)
247 int nr_pages = 1 << order;
249 if (unlikely((unsigned long)page[1].lru.prev != order))
252 for (i = 0; i < nr_pages; i++) {
253 struct page *p = page + i;
255 if (unlikely(!PageCompound(p) |
256 (page_private(p) != (unsigned long)page)))
258 __ClearPageCompound(p);
262 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
266 VM_BUG_ON((gfp_flags & (__GFP_WAIT | __GFP_HIGHMEM)) == __GFP_HIGHMEM);
268 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
269 * and __GFP_HIGHMEM from hard or soft interrupt context.
271 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
272 for (i = 0; i < (1 << order); i++)
273 clear_highpage(page + i);
277 * function for dealing with page's order in buddy system.
278 * zone->lock is already acquired when we use these.
279 * So, we don't need atomic page->flags operations here.
281 static inline unsigned long page_order(struct page *page)
283 return page_private(page);
286 static inline void set_page_order(struct page *page, int order)
288 set_page_private(page, order);
289 __SetPageBuddy(page);
292 static inline void rmv_page_order(struct page *page)
294 __ClearPageBuddy(page);
295 set_page_private(page, 0);
299 * Locate the struct page for both the matching buddy in our
300 * pair (buddy1) and the combined O(n+1) page they form (page).
302 * 1) Any buddy B1 will have an order O twin B2 which satisfies
303 * the following equation:
305 * For example, if the starting buddy (buddy2) is #8 its order
307 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
309 * 2) Any buddy B will have an order O+1 parent P which
310 * satisfies the following equation:
313 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
315 static inline struct page *
316 __page_find_buddy(struct page *page, unsigned long page_idx, unsigned int order)
318 unsigned long buddy_idx = page_idx ^ (1 << order);
320 return page + (buddy_idx - page_idx);
323 static inline unsigned long
324 __find_combined_index(unsigned long page_idx, unsigned int order)
326 return (page_idx & ~(1 << order));
330 * This function checks whether a page is free && is the buddy
331 * we can do coalesce a page and its buddy if
332 * (a) the buddy is not in a hole &&
333 * (b) the buddy is in the buddy system &&
334 * (c) a page and its buddy have the same order &&
335 * (d) a page and its buddy are in the same zone.
337 * For recording whether a page is in the buddy system, we use PG_buddy.
338 * Setting, clearing, and testing PG_buddy is serialized by zone->lock.
340 * For recording page's order, we use page_private(page).
342 static inline int page_is_buddy(struct page *page, struct page *buddy,
345 #ifdef CONFIG_HOLES_IN_ZONE
346 if (!pfn_valid(page_to_pfn(buddy)))
350 if (page_zone_id(page) != page_zone_id(buddy))
353 if (PageBuddy(buddy) && page_order(buddy) == order) {
354 BUG_ON(page_count(buddy) != 0);
361 * Freeing function for a buddy system allocator.
363 * The concept of a buddy system is to maintain direct-mapped table
364 * (containing bit values) for memory blocks of various "orders".
365 * The bottom level table contains the map for the smallest allocatable
366 * units of memory (here, pages), and each level above it describes
367 * pairs of units from the levels below, hence, "buddies".
368 * At a high level, all that happens here is marking the table entry
369 * at the bottom level available, and propagating the changes upward
370 * as necessary, plus some accounting needed to play nicely with other
371 * parts of the VM system.
372 * At each level, we keep a list of pages, which are heads of continuous
373 * free pages of length of (1 << order) and marked with PG_buddy. Page's
374 * order is recorded in page_private(page) field.
375 * So when we are allocating or freeing one, we can derive the state of the
376 * other. That is, if we allocate a small block, and both were
377 * free, the remainder of the region must be split into blocks.
378 * If a block is freed, and its buddy is also free, then this
379 * triggers coalescing into a block of larger size.
384 static inline void __free_one_page(struct page *page,
385 struct zone *zone, unsigned int order)
387 unsigned long page_idx;
388 int order_size = 1 << order;
390 if (unlikely(PageCompound(page)))
391 destroy_compound_page(page, order);
393 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
395 VM_BUG_ON(page_idx & (order_size - 1));
396 VM_BUG_ON(bad_range(zone, page));
398 __mod_zone_page_state(zone, NR_FREE_PAGES, order_size);
399 while (order < MAX_ORDER-1) {
400 unsigned long combined_idx;
401 struct free_area *area;
404 buddy = __page_find_buddy(page, page_idx, order);
405 if (!page_is_buddy(page, buddy, order))
406 break; /* Move the buddy up one level. */
408 list_del(&buddy->lru);
409 area = zone->free_area + order;
411 rmv_page_order(buddy);
412 combined_idx = __find_combined_index(page_idx, order);
413 page = page + (combined_idx - page_idx);
414 page_idx = combined_idx;
417 set_page_order(page, order);
418 list_add(&page->lru, &zone->free_area[order].free_list);
419 zone->free_area[order].nr_free++;
422 static inline int free_pages_check(struct page *page)
424 if (unlikely(page_mapcount(page) |
425 (page->mapping != NULL) |
426 (page_count(page) != 0) |
440 __ClearPageDirty(page);
442 * For now, we report if PG_reserved was found set, but do not
443 * clear it, and do not free the page. But we shall soon need
444 * to do more, for when the ZERO_PAGE count wraps negative.
446 return PageReserved(page);
450 * Frees a list of pages.
451 * Assumes all pages on list are in same zone, and of same order.
452 * count is the number of pages to free.
454 * If the zone was previously in an "all pages pinned" state then look to
455 * see if this freeing clears that state.
457 * And clear the zone's pages_scanned counter, to hold off the "all pages are
458 * pinned" detection logic.
460 static void free_pages_bulk(struct zone *zone, int count,
461 struct list_head *list, int order)
463 spin_lock(&zone->lock);
464 zone->all_unreclaimable = 0;
465 zone->pages_scanned = 0;
469 VM_BUG_ON(list_empty(list));
470 page = list_entry(list->prev, struct page, lru);
471 /* have to delete it as __free_one_page list manipulates */
472 list_del(&page->lru);
473 __free_one_page(page, zone, order);
475 spin_unlock(&zone->lock);
478 static void free_one_page(struct zone *zone, struct page *page, int order)
480 spin_lock(&zone->lock);
481 zone->all_unreclaimable = 0;
482 zone->pages_scanned = 0;
483 __free_one_page(page, zone, order);
484 spin_unlock(&zone->lock);
487 static void __free_pages_ok(struct page *page, unsigned int order)
493 for (i = 0 ; i < (1 << order) ; ++i)
494 reserved += free_pages_check(page + i);
498 if (!PageHighMem(page))
499 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
500 arch_free_page(page, order);
501 kernel_map_pages(page, 1 << order, 0);
503 local_irq_save(flags);
504 __count_vm_events(PGFREE, 1 << order);
505 free_one_page(page_zone(page), page, order);
506 local_irq_restore(flags);
510 * permit the bootmem allocator to evade page validation on high-order frees
512 void fastcall __init __free_pages_bootmem(struct page *page, unsigned int order)
515 __ClearPageReserved(page);
516 set_page_count(page, 0);
517 set_page_refcounted(page);
523 for (loop = 0; loop < BITS_PER_LONG; loop++) {
524 struct page *p = &page[loop];
526 if (loop + 1 < BITS_PER_LONG)
528 __ClearPageReserved(p);
529 set_page_count(p, 0);
532 set_page_refcounted(page);
533 __free_pages(page, order);
539 * The order of subdivision here is critical for the IO subsystem.
540 * Please do not alter this order without good reasons and regression
541 * testing. Specifically, as large blocks of memory are subdivided,
542 * the order in which smaller blocks are delivered depends on the order
543 * they're subdivided in this function. This is the primary factor
544 * influencing the order in which pages are delivered to the IO
545 * subsystem according to empirical testing, and this is also justified
546 * by considering the behavior of a buddy system containing a single
547 * large block of memory acted on by a series of small allocations.
548 * This behavior is a critical factor in sglist merging's success.
552 static inline void expand(struct zone *zone, struct page *page,
553 int low, int high, struct free_area *area)
555 unsigned long size = 1 << high;
561 VM_BUG_ON(bad_range(zone, &page[size]));
562 list_add(&page[size].lru, &area->free_list);
564 set_page_order(&page[size], high);
569 * This page is about to be returned from the page allocator
571 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
573 if (unlikely(page_mapcount(page) |
574 (page->mapping != NULL) |
575 (page_count(page) != 0) |
591 * For now, we report if PG_reserved was found set, but do not
592 * clear it, and do not allocate the page: as a safety net.
594 if (PageReserved(page))
597 page->flags &= ~(1 << PG_uptodate | 1 << PG_error |
598 1 << PG_referenced | 1 << PG_arch_1 |
599 1 << PG_checked | 1 << PG_mappedtodisk);
600 set_page_private(page, 0);
601 set_page_refcounted(page);
603 arch_alloc_page(page, order);
604 kernel_map_pages(page, 1 << order, 1);
606 if (gfp_flags & __GFP_ZERO)
607 prep_zero_page(page, order, gfp_flags);
609 if (order && (gfp_flags & __GFP_COMP))
610 prep_compound_page(page, order);
616 * Do the hard work of removing an element from the buddy allocator.
617 * Call me with the zone->lock already held.
619 static struct page *__rmqueue(struct zone *zone, unsigned int order)
621 struct free_area * area;
622 unsigned int current_order;
625 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
626 area = zone->free_area + current_order;
627 if (list_empty(&area->free_list))
630 page = list_entry(area->free_list.next, struct page, lru);
631 list_del(&page->lru);
632 rmv_page_order(page);
634 __mod_zone_page_state(zone, NR_FREE_PAGES, - (1UL << order));
635 expand(zone, page, order, current_order, area);
643 * Obtain a specified number of elements from the buddy allocator, all under
644 * a single hold of the lock, for efficiency. Add them to the supplied list.
645 * Returns the number of new pages which were placed at *list.
647 static int rmqueue_bulk(struct zone *zone, unsigned int order,
648 unsigned long count, struct list_head *list)
652 spin_lock(&zone->lock);
653 for (i = 0; i < count; ++i) {
654 struct page *page = __rmqueue(zone, order);
655 if (unlikely(page == NULL))
657 list_add_tail(&page->lru, list);
659 spin_unlock(&zone->lock);
665 * Called from the slab reaper to drain pagesets on a particular node that
666 * belongs to the currently executing processor.
667 * Note that this function must be called with the thread pinned to
668 * a single processor.
670 void drain_node_pages(int nodeid)
676 for (z = 0; z < MAX_NR_ZONES; z++) {
677 struct zone *zone = NODE_DATA(nodeid)->node_zones + z;
678 struct per_cpu_pageset *pset;
680 if (!populated_zone(zone))
683 pset = zone_pcp(zone, smp_processor_id());
684 for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
685 struct per_cpu_pages *pcp;
691 local_irq_save(flags);
692 if (pcp->count >= pcp->batch)
693 to_drain = pcp->batch;
695 to_drain = pcp->count;
696 free_pages_bulk(zone, to_drain, &pcp->list, 0);
697 pcp->count -= to_drain;
698 local_irq_restore(flags);
705 static void __drain_pages(unsigned int cpu)
711 for_each_zone(zone) {
712 struct per_cpu_pageset *pset;
714 if (!populated_zone(zone))
717 pset = zone_pcp(zone, cpu);
718 for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
719 struct per_cpu_pages *pcp;
722 local_irq_save(flags);
723 free_pages_bulk(zone, pcp->count, &pcp->list, 0);
725 local_irq_restore(flags);
732 void mark_free_pages(struct zone *zone)
734 unsigned long pfn, max_zone_pfn;
737 struct list_head *curr;
739 if (!zone->spanned_pages)
742 spin_lock_irqsave(&zone->lock, flags);
744 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
745 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
746 if (pfn_valid(pfn)) {
747 struct page *page = pfn_to_page(pfn);
749 if (!PageNosave(page))
750 ClearPageNosaveFree(page);
753 for (order = MAX_ORDER - 1; order >= 0; --order)
754 list_for_each(curr, &zone->free_area[order].free_list) {
757 pfn = page_to_pfn(list_entry(curr, struct page, lru));
758 for (i = 0; i < (1UL << order); i++)
759 SetPageNosaveFree(pfn_to_page(pfn + i));
762 spin_unlock_irqrestore(&zone->lock, flags);
766 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
768 void drain_local_pages(void)
772 local_irq_save(flags);
773 __drain_pages(smp_processor_id());
774 local_irq_restore(flags);
776 #endif /* CONFIG_PM */
779 * Free a 0-order page
781 static void fastcall free_hot_cold_page(struct page *page, int cold)
783 struct zone *zone = page_zone(page);
784 struct per_cpu_pages *pcp;
788 page->mapping = NULL;
789 if (free_pages_check(page))
792 if (!PageHighMem(page))
793 debug_check_no_locks_freed(page_address(page), PAGE_SIZE);
794 arch_free_page(page, 0);
795 kernel_map_pages(page, 1, 0);
797 pcp = &zone_pcp(zone, get_cpu())->pcp[cold];
798 local_irq_save(flags);
799 __count_vm_event(PGFREE);
800 list_add(&page->lru, &pcp->list);
802 if (pcp->count >= pcp->high) {
803 free_pages_bulk(zone, pcp->batch, &pcp->list, 0);
804 pcp->count -= pcp->batch;
806 local_irq_restore(flags);
810 void fastcall free_hot_page(struct page *page)
812 free_hot_cold_page(page, 0);
815 void fastcall free_cold_page(struct page *page)
817 free_hot_cold_page(page, 1);
821 * split_page takes a non-compound higher-order page, and splits it into
822 * n (1<<order) sub-pages: page[0..n]
823 * Each sub-page must be freed individually.
825 * Note: this is probably too low level an operation for use in drivers.
826 * Please consult with lkml before using this in your driver.
828 void split_page(struct page *page, unsigned int order)
832 VM_BUG_ON(PageCompound(page));
833 VM_BUG_ON(!page_count(page));
834 for (i = 1; i < (1 << order); i++)
835 set_page_refcounted(page + i);
839 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
840 * we cheat by calling it from here, in the order > 0 path. Saves a branch
843 static struct page *buffered_rmqueue(struct zonelist *zonelist,
844 struct zone *zone, int order, gfp_t gfp_flags)
848 int cold = !!(gfp_flags & __GFP_COLD);
853 if (likely(order == 0)) {
854 struct per_cpu_pages *pcp;
856 pcp = &zone_pcp(zone, cpu)->pcp[cold];
857 local_irq_save(flags);
859 pcp->count = rmqueue_bulk(zone, 0,
860 pcp->batch, &pcp->list);
861 if (unlikely(!pcp->count))
864 page = list_entry(pcp->list.next, struct page, lru);
865 list_del(&page->lru);
868 spin_lock_irqsave(&zone->lock, flags);
869 page = __rmqueue(zone, order);
870 spin_unlock(&zone->lock);
875 __count_zone_vm_events(PGALLOC, zone, 1 << order);
876 zone_statistics(zonelist, zone);
877 local_irq_restore(flags);
880 VM_BUG_ON(bad_range(zone, page));
881 if (prep_new_page(page, order, gfp_flags))
886 local_irq_restore(flags);
891 #define ALLOC_NO_WATERMARKS 0x01 /* don't check watermarks at all */
892 #define ALLOC_WMARK_MIN 0x02 /* use pages_min watermark */
893 #define ALLOC_WMARK_LOW 0x04 /* use pages_low watermark */
894 #define ALLOC_WMARK_HIGH 0x08 /* use pages_high watermark */
895 #define ALLOC_HARDER 0x10 /* try to alloc harder */
896 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
897 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
899 #ifdef CONFIG_FAIL_PAGE_ALLOC
901 static struct fail_page_alloc_attr {
902 struct fault_attr attr;
904 u32 ignore_gfp_highmem;
907 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
909 struct dentry *ignore_gfp_highmem_file;
910 struct dentry *ignore_gfp_wait_file;
912 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
914 } fail_page_alloc = {
915 .attr = FAULT_ATTR_INITIALIZER,
916 .ignore_gfp_wait = 1,
917 .ignore_gfp_highmem = 1,
920 static int __init setup_fail_page_alloc(char *str)
922 return setup_fault_attr(&fail_page_alloc.attr, str);
924 __setup("fail_page_alloc=", setup_fail_page_alloc);
926 static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
928 if (gfp_mask & __GFP_NOFAIL)
930 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
932 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
935 return should_fail(&fail_page_alloc.attr, 1 << order);
938 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
940 static int __init fail_page_alloc_debugfs(void)
942 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
946 err = init_fault_attr_dentries(&fail_page_alloc.attr,
950 dir = fail_page_alloc.attr.dentries.dir;
952 fail_page_alloc.ignore_gfp_wait_file =
953 debugfs_create_bool("ignore-gfp-wait", mode, dir,
954 &fail_page_alloc.ignore_gfp_wait);
956 fail_page_alloc.ignore_gfp_highmem_file =
957 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
958 &fail_page_alloc.ignore_gfp_highmem);
960 if (!fail_page_alloc.ignore_gfp_wait_file ||
961 !fail_page_alloc.ignore_gfp_highmem_file) {
963 debugfs_remove(fail_page_alloc.ignore_gfp_wait_file);
964 debugfs_remove(fail_page_alloc.ignore_gfp_highmem_file);
965 cleanup_fault_attr_dentries(&fail_page_alloc.attr);
971 late_initcall(fail_page_alloc_debugfs);
973 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
975 #else /* CONFIG_FAIL_PAGE_ALLOC */
977 static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
982 #endif /* CONFIG_FAIL_PAGE_ALLOC */
985 * Return 1 if free pages are above 'mark'. This takes into account the order
988 int zone_watermark_ok(struct zone *z, int order, unsigned long mark,
989 int classzone_idx, int alloc_flags)
991 /* free_pages my go negative - that's OK */
993 long free_pages = zone_page_state(z, NR_FREE_PAGES) - (1 << order) + 1;
996 if (alloc_flags & ALLOC_HIGH)
998 if (alloc_flags & ALLOC_HARDER)
1001 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
1003 for (o = 0; o < order; o++) {
1004 /* At the next order, this order's pages become unavailable */
1005 free_pages -= z->free_area[o].nr_free << o;
1007 /* Require fewer higher order pages to be free */
1010 if (free_pages <= min)
1018 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1019 * skip over zones that are not allowed by the cpuset, or that have
1020 * been recently (in last second) found to be nearly full. See further
1021 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1022 * that have to skip over alot of full or unallowed zones.
1024 * If the zonelist cache is present in the passed in zonelist, then
1025 * returns a pointer to the allowed node mask (either the current
1026 * tasks mems_allowed, or node_online_map.)
1028 * If the zonelist cache is not available for this zonelist, does
1029 * nothing and returns NULL.
1031 * If the fullzones BITMAP in the zonelist cache is stale (more than
1032 * a second since last zap'd) then we zap it out (clear its bits.)
1034 * We hold off even calling zlc_setup, until after we've checked the
1035 * first zone in the zonelist, on the theory that most allocations will
1036 * be satisfied from that first zone, so best to examine that zone as
1037 * quickly as we can.
1039 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1041 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1042 nodemask_t *allowednodes; /* zonelist_cache approximation */
1044 zlc = zonelist->zlcache_ptr;
1048 if (jiffies - zlc->last_full_zap > 1 * HZ) {
1049 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1050 zlc->last_full_zap = jiffies;
1053 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1054 &cpuset_current_mems_allowed :
1056 return allowednodes;
1060 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1061 * if it is worth looking at further for free memory:
1062 * 1) Check that the zone isn't thought to be full (doesn't have its
1063 * bit set in the zonelist_cache fullzones BITMAP).
1064 * 2) Check that the zones node (obtained from the zonelist_cache
1065 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1066 * Return true (non-zero) if zone is worth looking at further, or
1067 * else return false (zero) if it is not.
1069 * This check -ignores- the distinction between various watermarks,
1070 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1071 * found to be full for any variation of these watermarks, it will
1072 * be considered full for up to one second by all requests, unless
1073 * we are so low on memory on all allowed nodes that we are forced
1074 * into the second scan of the zonelist.
1076 * In the second scan we ignore this zonelist cache and exactly
1077 * apply the watermarks to all zones, even it is slower to do so.
1078 * We are low on memory in the second scan, and should leave no stone
1079 * unturned looking for a free page.
1081 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zone **z,
1082 nodemask_t *allowednodes)
1084 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1085 int i; /* index of *z in zonelist zones */
1086 int n; /* node that zone *z is on */
1088 zlc = zonelist->zlcache_ptr;
1092 i = z - zonelist->zones;
1095 /* This zone is worth trying if it is allowed but not full */
1096 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1100 * Given 'z' scanning a zonelist, set the corresponding bit in
1101 * zlc->fullzones, so that subsequent attempts to allocate a page
1102 * from that zone don't waste time re-examining it.
1104 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zone **z)
1106 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1107 int i; /* index of *z in zonelist zones */
1109 zlc = zonelist->zlcache_ptr;
1113 i = z - zonelist->zones;
1115 set_bit(i, zlc->fullzones);
1118 #else /* CONFIG_NUMA */
1120 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1125 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zone **z,
1126 nodemask_t *allowednodes)
1131 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zone **z)
1134 #endif /* CONFIG_NUMA */
1137 * get_page_from_freelist goes through the zonelist trying to allocate
1140 static struct page *
1141 get_page_from_freelist(gfp_t gfp_mask, unsigned int order,
1142 struct zonelist *zonelist, int alloc_flags)
1145 struct page *page = NULL;
1146 int classzone_idx = zone_idx(zonelist->zones[0]);
1148 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1149 int zlc_active = 0; /* set if using zonelist_cache */
1150 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1154 * Scan zonelist, looking for a zone with enough free.
1155 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1157 z = zonelist->zones;
1160 if (NUMA_BUILD && zlc_active &&
1161 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1164 if (unlikely(NUMA_BUILD && (gfp_mask & __GFP_THISNODE) &&
1165 zone->zone_pgdat != zonelist->zones[0]->zone_pgdat))
1167 if ((alloc_flags & ALLOC_CPUSET) &&
1168 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1171 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1173 if (alloc_flags & ALLOC_WMARK_MIN)
1174 mark = zone->pages_min;
1175 else if (alloc_flags & ALLOC_WMARK_LOW)
1176 mark = zone->pages_low;
1178 mark = zone->pages_high;
1179 if (!zone_watermark_ok(zone, order, mark,
1180 classzone_idx, alloc_flags)) {
1181 if (!zone_reclaim_mode ||
1182 !zone_reclaim(zone, gfp_mask, order))
1183 goto this_zone_full;
1187 page = buffered_rmqueue(zonelist, zone, order, gfp_mask);
1192 zlc_mark_zone_full(zonelist, z);
1194 if (NUMA_BUILD && !did_zlc_setup) {
1195 /* we do zlc_setup after the first zone is tried */
1196 allowednodes = zlc_setup(zonelist, alloc_flags);
1200 } while (*(++z) != NULL);
1202 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) {
1203 /* Disable zlc cache for second zonelist scan */
1211 * This is the 'heart' of the zoned buddy allocator.
1213 struct page * fastcall
1214 __alloc_pages(gfp_t gfp_mask, unsigned int order,
1215 struct zonelist *zonelist)
1217 const gfp_t wait = gfp_mask & __GFP_WAIT;
1220 struct reclaim_state reclaim_state;
1221 struct task_struct *p = current;
1224 int did_some_progress;
1226 might_sleep_if(wait);
1228 if (should_fail_alloc_page(gfp_mask, order))
1232 z = zonelist->zones; /* the list of zones suitable for gfp_mask */
1234 if (unlikely(*z == NULL)) {
1235 /* Should this ever happen?? */
1239 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, order,
1240 zonelist, ALLOC_WMARK_LOW|ALLOC_CPUSET);
1245 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
1246 * __GFP_NOWARN set) should not cause reclaim since the subsystem
1247 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
1248 * using a larger set of nodes after it has established that the
1249 * allowed per node queues are empty and that nodes are
1252 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
1255 for (z = zonelist->zones; *z; z++)
1256 wakeup_kswapd(*z, order);
1259 * OK, we're below the kswapd watermark and have kicked background
1260 * reclaim. Now things get more complex, so set up alloc_flags according
1261 * to how we want to proceed.
1263 * The caller may dip into page reserves a bit more if the caller
1264 * cannot run direct reclaim, or if the caller has realtime scheduling
1265 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
1266 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
1268 alloc_flags = ALLOC_WMARK_MIN;
1269 if ((unlikely(rt_task(p)) && !in_interrupt()) || !wait)
1270 alloc_flags |= ALLOC_HARDER;
1271 if (gfp_mask & __GFP_HIGH)
1272 alloc_flags |= ALLOC_HIGH;
1274 alloc_flags |= ALLOC_CPUSET;
1277 * Go through the zonelist again. Let __GFP_HIGH and allocations
1278 * coming from realtime tasks go deeper into reserves.
1280 * This is the last chance, in general, before the goto nopage.
1281 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
1282 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1284 page = get_page_from_freelist(gfp_mask, order, zonelist, alloc_flags);
1288 /* This allocation should allow future memory freeing. */
1291 if (((p->flags & PF_MEMALLOC) || unlikely(test_thread_flag(TIF_MEMDIE)))
1292 && !in_interrupt()) {
1293 if (!(gfp_mask & __GFP_NOMEMALLOC)) {
1295 /* go through the zonelist yet again, ignoring mins */
1296 page = get_page_from_freelist(gfp_mask, order,
1297 zonelist, ALLOC_NO_WATERMARKS);
1300 if (gfp_mask & __GFP_NOFAIL) {
1301 congestion_wait(WRITE, HZ/50);
1308 /* Atomic allocations - we can't balance anything */
1314 /* We now go into synchronous reclaim */
1315 cpuset_memory_pressure_bump();
1316 p->flags |= PF_MEMALLOC;
1317 reclaim_state.reclaimed_slab = 0;
1318 p->reclaim_state = &reclaim_state;
1320 did_some_progress = try_to_free_pages(zonelist->zones, gfp_mask);
1322 p->reclaim_state = NULL;
1323 p->flags &= ~PF_MEMALLOC;
1327 if (likely(did_some_progress)) {
1328 page = get_page_from_freelist(gfp_mask, order,
1329 zonelist, alloc_flags);
1332 } else if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
1334 * Go through the zonelist yet one more time, keep
1335 * very high watermark here, this is only to catch
1336 * a parallel oom killing, we must fail if we're still
1337 * under heavy pressure.
1339 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, order,
1340 zonelist, ALLOC_WMARK_HIGH|ALLOC_CPUSET);
1344 out_of_memory(zonelist, gfp_mask, order);
1349 * Don't let big-order allocations loop unless the caller explicitly
1350 * requests that. Wait for some write requests to complete then retry.
1352 * In this implementation, __GFP_REPEAT means __GFP_NOFAIL for order
1353 * <= 3, but that may not be true in other implementations.
1356 if (!(gfp_mask & __GFP_NORETRY)) {
1357 if ((order <= 3) || (gfp_mask & __GFP_REPEAT))
1359 if (gfp_mask & __GFP_NOFAIL)
1363 congestion_wait(WRITE, HZ/50);
1368 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
1369 printk(KERN_WARNING "%s: page allocation failure."
1370 " order:%d, mode:0x%x\n",
1371 p->comm, order, gfp_mask);
1379 EXPORT_SYMBOL(__alloc_pages);
1382 * Common helper functions.
1384 fastcall unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
1387 page = alloc_pages(gfp_mask, order);
1390 return (unsigned long) page_address(page);
1393 EXPORT_SYMBOL(__get_free_pages);
1395 fastcall unsigned long get_zeroed_page(gfp_t gfp_mask)
1400 * get_zeroed_page() returns a 32-bit address, which cannot represent
1403 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
1405 page = alloc_pages(gfp_mask | __GFP_ZERO, 0);
1407 return (unsigned long) page_address(page);
1411 EXPORT_SYMBOL(get_zeroed_page);
1413 void __pagevec_free(struct pagevec *pvec)
1415 int i = pagevec_count(pvec);
1418 free_hot_cold_page(pvec->pages[i], pvec->cold);
1421 fastcall void __free_pages(struct page *page, unsigned int order)
1423 if (put_page_testzero(page)) {
1425 free_hot_page(page);
1427 __free_pages_ok(page, order);
1431 EXPORT_SYMBOL(__free_pages);
1433 fastcall void free_pages(unsigned long addr, unsigned int order)
1436 VM_BUG_ON(!virt_addr_valid((void *)addr));
1437 __free_pages(virt_to_page((void *)addr), order);
1441 EXPORT_SYMBOL(free_pages);
1443 static unsigned int nr_free_zone_pages(int offset)
1445 /* Just pick one node, since fallback list is circular */
1446 pg_data_t *pgdat = NODE_DATA(numa_node_id());
1447 unsigned int sum = 0;
1449 struct zonelist *zonelist = pgdat->node_zonelists + offset;
1450 struct zone **zonep = zonelist->zones;
1453 for (zone = *zonep++; zone; zone = *zonep++) {
1454 unsigned long size = zone->present_pages;
1455 unsigned long high = zone->pages_high;
1464 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
1466 unsigned int nr_free_buffer_pages(void)
1468 return nr_free_zone_pages(gfp_zone(GFP_USER));
1472 * Amount of free RAM allocatable within all zones
1474 unsigned int nr_free_pagecache_pages(void)
1476 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER));
1479 static inline void show_node(struct zone *zone)
1482 printk("Node %d ", zone_to_nid(zone));
1485 void si_meminfo(struct sysinfo *val)
1487 val->totalram = totalram_pages;
1489 val->freeram = global_page_state(NR_FREE_PAGES);
1490 val->bufferram = nr_blockdev_pages();
1491 val->totalhigh = totalhigh_pages;
1492 val->freehigh = nr_free_highpages();
1493 val->mem_unit = PAGE_SIZE;
1496 EXPORT_SYMBOL(si_meminfo);
1499 void si_meminfo_node(struct sysinfo *val, int nid)
1501 pg_data_t *pgdat = NODE_DATA(nid);
1503 val->totalram = pgdat->node_present_pages;
1504 val->freeram = node_page_state(nid, NR_FREE_PAGES);
1505 #ifdef CONFIG_HIGHMEM
1506 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
1507 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
1513 val->mem_unit = PAGE_SIZE;
1517 #define K(x) ((x) << (PAGE_SHIFT-10))
1520 * Show free area list (used inside shift_scroll-lock stuff)
1521 * We also calculate the percentage fragmentation. We do this by counting the
1522 * memory on each free list with the exception of the first item on the list.
1524 void show_free_areas(void)
1527 unsigned long active;
1528 unsigned long inactive;
1532 for_each_zone(zone) {
1533 if (!populated_zone(zone))
1537 printk("%s per-cpu:\n", zone->name);
1539 for_each_online_cpu(cpu) {
1540 struct per_cpu_pageset *pageset;
1542 pageset = zone_pcp(zone, cpu);
1544 printk("CPU %4d: Hot: hi:%5d, btch:%4d usd:%4d "
1545 "Cold: hi:%5d, btch:%4d usd:%4d\n",
1546 cpu, pageset->pcp[0].high,
1547 pageset->pcp[0].batch, pageset->pcp[0].count,
1548 pageset->pcp[1].high, pageset->pcp[1].batch,
1549 pageset->pcp[1].count);
1553 get_zone_counts(&active, &inactive, &free);
1555 printk("Active:%lu inactive:%lu dirty:%lu writeback:%lu unstable:%lu\n"
1556 " free:%lu slab:%lu mapped:%lu pagetables:%lu bounce:%lu\n",
1559 global_page_state(NR_FILE_DIRTY),
1560 global_page_state(NR_WRITEBACK),
1561 global_page_state(NR_UNSTABLE_NFS),
1562 global_page_state(NR_FREE_PAGES),
1563 global_page_state(NR_SLAB_RECLAIMABLE) +
1564 global_page_state(NR_SLAB_UNRECLAIMABLE),
1565 global_page_state(NR_FILE_MAPPED),
1566 global_page_state(NR_PAGETABLE),
1567 global_page_state(NR_BOUNCE));
1569 for_each_zone(zone) {
1572 if (!populated_zone(zone))
1584 " pages_scanned:%lu"
1585 " all_unreclaimable? %s"
1588 K(zone_page_state(zone, NR_FREE_PAGES)),
1591 K(zone->pages_high),
1592 K(zone_page_state(zone, NR_ACTIVE)),
1593 K(zone_page_state(zone, NR_INACTIVE)),
1594 K(zone->present_pages),
1595 zone->pages_scanned,
1596 (zone->all_unreclaimable ? "yes" : "no")
1598 printk("lowmem_reserve[]:");
1599 for (i = 0; i < MAX_NR_ZONES; i++)
1600 printk(" %lu", zone->lowmem_reserve[i]);
1604 for_each_zone(zone) {
1605 unsigned long nr[MAX_ORDER], flags, order, total = 0;
1607 if (!populated_zone(zone))
1611 printk("%s: ", zone->name);
1613 spin_lock_irqsave(&zone->lock, flags);
1614 for (order = 0; order < MAX_ORDER; order++) {
1615 nr[order] = zone->free_area[order].nr_free;
1616 total += nr[order] << order;
1618 spin_unlock_irqrestore(&zone->lock, flags);
1619 for (order = 0; order < MAX_ORDER; order++)
1620 printk("%lu*%lukB ", nr[order], K(1UL) << order);
1621 printk("= %lukB\n", K(total));
1624 show_swap_cache_info();
1628 * Builds allocation fallback zone lists.
1630 * Add all populated zones of a node to the zonelist.
1632 static int __meminit build_zonelists_node(pg_data_t *pgdat,
1633 struct zonelist *zonelist, int nr_zones, enum zone_type zone_type)
1637 BUG_ON(zone_type >= MAX_NR_ZONES);
1642 zone = pgdat->node_zones + zone_type;
1643 if (populated_zone(zone)) {
1644 zonelist->zones[nr_zones++] = zone;
1645 check_highest_zone(zone_type);
1648 } while (zone_type);
1653 #define MAX_NODE_LOAD (num_online_nodes())
1654 static int __meminitdata node_load[MAX_NUMNODES];
1656 * find_next_best_node - find the next node that should appear in a given node's fallback list
1657 * @node: node whose fallback list we're appending
1658 * @used_node_mask: nodemask_t of already used nodes
1660 * We use a number of factors to determine which is the next node that should
1661 * appear on a given node's fallback list. The node should not have appeared
1662 * already in @node's fallback list, and it should be the next closest node
1663 * according to the distance array (which contains arbitrary distance values
1664 * from each node to each node in the system), and should also prefer nodes
1665 * with no CPUs, since presumably they'll have very little allocation pressure
1666 * on them otherwise.
1667 * It returns -1 if no node is found.
1669 static int __meminit find_next_best_node(int node, nodemask_t *used_node_mask)
1672 int min_val = INT_MAX;
1675 /* Use the local node if we haven't already */
1676 if (!node_isset(node, *used_node_mask)) {
1677 node_set(node, *used_node_mask);
1681 for_each_online_node(n) {
1684 /* Don't want a node to appear more than once */
1685 if (node_isset(n, *used_node_mask))
1688 /* Use the distance array to find the distance */
1689 val = node_distance(node, n);
1691 /* Penalize nodes under us ("prefer the next node") */
1694 /* Give preference to headless and unused nodes */
1695 tmp = node_to_cpumask(n);
1696 if (!cpus_empty(tmp))
1697 val += PENALTY_FOR_NODE_WITH_CPUS;
1699 /* Slight preference for less loaded node */
1700 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
1701 val += node_load[n];
1703 if (val < min_val) {
1710 node_set(best_node, *used_node_mask);
1715 static void __meminit build_zonelists(pg_data_t *pgdat)
1717 int j, node, local_node;
1719 int prev_node, load;
1720 struct zonelist *zonelist;
1721 nodemask_t used_mask;
1723 /* initialize zonelists */
1724 for (i = 0; i < MAX_NR_ZONES; i++) {
1725 zonelist = pgdat->node_zonelists + i;
1726 zonelist->zones[0] = NULL;
1729 /* NUMA-aware ordering of nodes */
1730 local_node = pgdat->node_id;
1731 load = num_online_nodes();
1732 prev_node = local_node;
1733 nodes_clear(used_mask);
1734 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
1735 int distance = node_distance(local_node, node);
1738 * If another node is sufficiently far away then it is better
1739 * to reclaim pages in a zone before going off node.
1741 if (distance > RECLAIM_DISTANCE)
1742 zone_reclaim_mode = 1;
1745 * We don't want to pressure a particular node.
1746 * So adding penalty to the first node in same
1747 * distance group to make it round-robin.
1750 if (distance != node_distance(local_node, prev_node))
1751 node_load[node] += load;
1754 for (i = 0; i < MAX_NR_ZONES; i++) {
1755 zonelist = pgdat->node_zonelists + i;
1756 for (j = 0; zonelist->zones[j] != NULL; j++);
1758 j = build_zonelists_node(NODE_DATA(node), zonelist, j, i);
1759 zonelist->zones[j] = NULL;
1764 /* Construct the zonelist performance cache - see further mmzone.h */
1765 static void __meminit build_zonelist_cache(pg_data_t *pgdat)
1769 for (i = 0; i < MAX_NR_ZONES; i++) {
1770 struct zonelist *zonelist;
1771 struct zonelist_cache *zlc;
1774 zonelist = pgdat->node_zonelists + i;
1775 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
1776 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1777 for (z = zonelist->zones; *z; z++)
1778 zlc->z_to_n[z - zonelist->zones] = zone_to_nid(*z);
1782 #else /* CONFIG_NUMA */
1784 static void __meminit build_zonelists(pg_data_t *pgdat)
1786 int node, local_node;
1789 local_node = pgdat->node_id;
1790 for (i = 0; i < MAX_NR_ZONES; i++) {
1791 struct zonelist *zonelist;
1793 zonelist = pgdat->node_zonelists + i;
1795 j = build_zonelists_node(pgdat, zonelist, 0, i);
1797 * Now we build the zonelist so that it contains the zones
1798 * of all the other nodes.
1799 * We don't want to pressure a particular node, so when
1800 * building the zones for node N, we make sure that the
1801 * zones coming right after the local ones are those from
1802 * node N+1 (modulo N)
1804 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
1805 if (!node_online(node))
1807 j = build_zonelists_node(NODE_DATA(node), zonelist, j, i);
1809 for (node = 0; node < local_node; node++) {
1810 if (!node_online(node))
1812 j = build_zonelists_node(NODE_DATA(node), zonelist, j, i);
1815 zonelist->zones[j] = NULL;
1819 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
1820 static void __meminit build_zonelist_cache(pg_data_t *pgdat)
1824 for (i = 0; i < MAX_NR_ZONES; i++)
1825 pgdat->node_zonelists[i].zlcache_ptr = NULL;
1828 #endif /* CONFIG_NUMA */
1830 /* return values int ....just for stop_machine_run() */
1831 static int __meminit __build_all_zonelists(void *dummy)
1835 for_each_online_node(nid) {
1836 build_zonelists(NODE_DATA(nid));
1837 build_zonelist_cache(NODE_DATA(nid));
1842 void __meminit build_all_zonelists(void)
1844 if (system_state == SYSTEM_BOOTING) {
1845 __build_all_zonelists(NULL);
1846 cpuset_init_current_mems_allowed();
1848 /* we have to stop all cpus to guaranntee there is no user
1850 stop_machine_run(__build_all_zonelists, NULL, NR_CPUS);
1851 /* cpuset refresh routine should be here */
1853 vm_total_pages = nr_free_pagecache_pages();
1854 printk("Built %i zonelists. Total pages: %ld\n",
1855 num_online_nodes(), vm_total_pages);
1859 * Helper functions to size the waitqueue hash table.
1860 * Essentially these want to choose hash table sizes sufficiently
1861 * large so that collisions trying to wait on pages are rare.
1862 * But in fact, the number of active page waitqueues on typical
1863 * systems is ridiculously low, less than 200. So this is even
1864 * conservative, even though it seems large.
1866 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
1867 * waitqueues, i.e. the size of the waitq table given the number of pages.
1869 #define PAGES_PER_WAITQUEUE 256
1871 #ifndef CONFIG_MEMORY_HOTPLUG
1872 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
1874 unsigned long size = 1;
1876 pages /= PAGES_PER_WAITQUEUE;
1878 while (size < pages)
1882 * Once we have dozens or even hundreds of threads sleeping
1883 * on IO we've got bigger problems than wait queue collision.
1884 * Limit the size of the wait table to a reasonable size.
1886 size = min(size, 4096UL);
1888 return max(size, 4UL);
1892 * A zone's size might be changed by hot-add, so it is not possible to determine
1893 * a suitable size for its wait_table. So we use the maximum size now.
1895 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
1897 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
1898 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
1899 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
1901 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
1902 * or more by the traditional way. (See above). It equals:
1904 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
1905 * ia64(16K page size) : = ( 8G + 4M)byte.
1906 * powerpc (64K page size) : = (32G +16M)byte.
1908 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
1915 * This is an integer logarithm so that shifts can be used later
1916 * to extract the more random high bits from the multiplicative
1917 * hash function before the remainder is taken.
1919 static inline unsigned long wait_table_bits(unsigned long size)
1924 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
1927 * Initially all pages are reserved - free ones are freed
1928 * up by free_all_bootmem() once the early boot process is
1929 * done. Non-atomic initialization, single-pass.
1931 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
1932 unsigned long start_pfn, enum memmap_context context)
1935 unsigned long end_pfn = start_pfn + size;
1938 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
1940 * There can be holes in boot-time mem_map[]s
1941 * handed to this function. They do not
1942 * exist on hotplugged memory.
1944 if (context == MEMMAP_EARLY) {
1945 if (!early_pfn_valid(pfn))
1947 if (!early_pfn_in_nid(pfn, nid))
1950 page = pfn_to_page(pfn);
1951 set_page_links(page, zone, nid, pfn);
1952 init_page_count(page);
1953 reset_page_mapcount(page);
1954 SetPageReserved(page);
1955 INIT_LIST_HEAD(&page->lru);
1956 #ifdef WANT_PAGE_VIRTUAL
1957 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1958 if (!is_highmem_idx(zone))
1959 set_page_address(page, __va(pfn << PAGE_SHIFT));
1964 void zone_init_free_lists(struct pglist_data *pgdat, struct zone *zone,
1968 for (order = 0; order < MAX_ORDER ; order++) {
1969 INIT_LIST_HEAD(&zone->free_area[order].free_list);
1970 zone->free_area[order].nr_free = 0;
1974 #ifndef __HAVE_ARCH_MEMMAP_INIT
1975 #define memmap_init(size, nid, zone, start_pfn) \
1976 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
1979 static int __cpuinit zone_batchsize(struct zone *zone)
1984 * The per-cpu-pages pools are set to around 1000th of the
1985 * size of the zone. But no more than 1/2 of a meg.
1987 * OK, so we don't know how big the cache is. So guess.
1989 batch = zone->present_pages / 1024;
1990 if (batch * PAGE_SIZE > 512 * 1024)
1991 batch = (512 * 1024) / PAGE_SIZE;
1992 batch /= 4; /* We effectively *= 4 below */
1997 * Clamp the batch to a 2^n - 1 value. Having a power
1998 * of 2 value was found to be more likely to have
1999 * suboptimal cache aliasing properties in some cases.
2001 * For example if 2 tasks are alternately allocating
2002 * batches of pages, one task can end up with a lot
2003 * of pages of one half of the possible page colors
2004 * and the other with pages of the other colors.
2006 batch = (1 << (fls(batch + batch/2)-1)) - 1;
2011 inline void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
2013 struct per_cpu_pages *pcp;
2015 memset(p, 0, sizeof(*p));
2017 pcp = &p->pcp[0]; /* hot */
2019 pcp->high = 6 * batch;
2020 pcp->batch = max(1UL, 1 * batch);
2021 INIT_LIST_HEAD(&pcp->list);
2023 pcp = &p->pcp[1]; /* cold*/
2025 pcp->high = 2 * batch;
2026 pcp->batch = max(1UL, batch/2);
2027 INIT_LIST_HEAD(&pcp->list);
2031 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
2032 * to the value high for the pageset p.
2035 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
2038 struct per_cpu_pages *pcp;
2040 pcp = &p->pcp[0]; /* hot list */
2042 pcp->batch = max(1UL, high/4);
2043 if ((high/4) > (PAGE_SHIFT * 8))
2044 pcp->batch = PAGE_SHIFT * 8;
2050 * Boot pageset table. One per cpu which is going to be used for all
2051 * zones and all nodes. The parameters will be set in such a way
2052 * that an item put on a list will immediately be handed over to
2053 * the buddy list. This is safe since pageset manipulation is done
2054 * with interrupts disabled.
2056 * Some NUMA counter updates may also be caught by the boot pagesets.
2058 * The boot_pagesets must be kept even after bootup is complete for
2059 * unused processors and/or zones. They do play a role for bootstrapping
2060 * hotplugged processors.
2062 * zoneinfo_show() and maybe other functions do
2063 * not check if the processor is online before following the pageset pointer.
2064 * Other parts of the kernel may not check if the zone is available.
2066 static struct per_cpu_pageset boot_pageset[NR_CPUS];
2069 * Dynamically allocate memory for the
2070 * per cpu pageset array in struct zone.
2072 static int __cpuinit process_zones(int cpu)
2074 struct zone *zone, *dzone;
2076 for_each_zone(zone) {
2078 if (!populated_zone(zone))
2081 zone_pcp(zone, cpu) = kmalloc_node(sizeof(struct per_cpu_pageset),
2082 GFP_KERNEL, cpu_to_node(cpu));
2083 if (!zone_pcp(zone, cpu))
2086 setup_pageset(zone_pcp(zone, cpu), zone_batchsize(zone));
2088 if (percpu_pagelist_fraction)
2089 setup_pagelist_highmark(zone_pcp(zone, cpu),
2090 (zone->present_pages / percpu_pagelist_fraction));
2095 for_each_zone(dzone) {
2098 kfree(zone_pcp(dzone, cpu));
2099 zone_pcp(dzone, cpu) = NULL;
2104 static inline void free_zone_pagesets(int cpu)
2108 for_each_zone(zone) {
2109 struct per_cpu_pageset *pset = zone_pcp(zone, cpu);
2111 /* Free per_cpu_pageset if it is slab allocated */
2112 if (pset != &boot_pageset[cpu])
2114 zone_pcp(zone, cpu) = NULL;
2118 static int __cpuinit pageset_cpuup_callback(struct notifier_block *nfb,
2119 unsigned long action,
2122 int cpu = (long)hcpu;
2123 int ret = NOTIFY_OK;
2126 case CPU_UP_PREPARE:
2127 if (process_zones(cpu))
2130 case CPU_UP_CANCELED:
2132 free_zone_pagesets(cpu);
2140 static struct notifier_block __cpuinitdata pageset_notifier =
2141 { &pageset_cpuup_callback, NULL, 0 };
2143 void __init setup_per_cpu_pageset(void)
2147 /* Initialize per_cpu_pageset for cpu 0.
2148 * A cpuup callback will do this for every cpu
2149 * as it comes online
2151 err = process_zones(smp_processor_id());
2153 register_cpu_notifier(&pageset_notifier);
2159 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
2162 struct pglist_data *pgdat = zone->zone_pgdat;
2166 * The per-page waitqueue mechanism uses hashed waitqueues
2169 zone->wait_table_hash_nr_entries =
2170 wait_table_hash_nr_entries(zone_size_pages);
2171 zone->wait_table_bits =
2172 wait_table_bits(zone->wait_table_hash_nr_entries);
2173 alloc_size = zone->wait_table_hash_nr_entries
2174 * sizeof(wait_queue_head_t);
2176 if (system_state == SYSTEM_BOOTING) {
2177 zone->wait_table = (wait_queue_head_t *)
2178 alloc_bootmem_node(pgdat, alloc_size);
2181 * This case means that a zone whose size was 0 gets new memory
2182 * via memory hot-add.
2183 * But it may be the case that a new node was hot-added. In
2184 * this case vmalloc() will not be able to use this new node's
2185 * memory - this wait_table must be initialized to use this new
2186 * node itself as well.
2187 * To use this new node's memory, further consideration will be
2190 zone->wait_table = (wait_queue_head_t *)vmalloc(alloc_size);
2192 if (!zone->wait_table)
2195 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
2196 init_waitqueue_head(zone->wait_table + i);
2201 static __meminit void zone_pcp_init(struct zone *zone)
2204 unsigned long batch = zone_batchsize(zone);
2206 for (cpu = 0; cpu < NR_CPUS; cpu++) {
2208 /* Early boot. Slab allocator not functional yet */
2209 zone_pcp(zone, cpu) = &boot_pageset[cpu];
2210 setup_pageset(&boot_pageset[cpu],0);
2212 setup_pageset(zone_pcp(zone,cpu), batch);
2215 if (zone->present_pages)
2216 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%lu\n",
2217 zone->name, zone->present_pages, batch);
2220 __meminit int init_currently_empty_zone(struct zone *zone,
2221 unsigned long zone_start_pfn,
2223 enum memmap_context context)
2225 struct pglist_data *pgdat = zone->zone_pgdat;
2227 ret = zone_wait_table_init(zone, size);
2230 pgdat->nr_zones = zone_idx(zone) + 1;
2232 zone->zone_start_pfn = zone_start_pfn;
2234 memmap_init(size, pgdat->node_id, zone_idx(zone), zone_start_pfn);
2236 zone_init_free_lists(pgdat, zone, zone->spanned_pages);
2241 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
2243 * Basic iterator support. Return the first range of PFNs for a node
2244 * Note: nid == MAX_NUMNODES returns first region regardless of node
2246 static int __init first_active_region_index_in_nid(int nid)
2250 for (i = 0; i < nr_nodemap_entries; i++)
2251 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
2258 * Basic iterator support. Return the next active range of PFNs for a node
2259 * Note: nid == MAX_NUMNODES returns next region regardles of node
2261 static int __init next_active_region_index_in_nid(int index, int nid)
2263 for (index = index + 1; index < nr_nodemap_entries; index++)
2264 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
2270 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
2272 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
2273 * Architectures may implement their own version but if add_active_range()
2274 * was used and there are no special requirements, this is a convenient
2277 int __init early_pfn_to_nid(unsigned long pfn)
2281 for (i = 0; i < nr_nodemap_entries; i++) {
2282 unsigned long start_pfn = early_node_map[i].start_pfn;
2283 unsigned long end_pfn = early_node_map[i].end_pfn;
2285 if (start_pfn <= pfn && pfn < end_pfn)
2286 return early_node_map[i].nid;
2291 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
2293 /* Basic iterator support to walk early_node_map[] */
2294 #define for_each_active_range_index_in_nid(i, nid) \
2295 for (i = first_active_region_index_in_nid(nid); i != -1; \
2296 i = next_active_region_index_in_nid(i, nid))
2299 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
2300 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
2301 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
2303 * If an architecture guarantees that all ranges registered with
2304 * add_active_ranges() contain no holes and may be freed, this
2305 * this function may be used instead of calling free_bootmem() manually.
2307 void __init free_bootmem_with_active_regions(int nid,
2308 unsigned long max_low_pfn)
2312 for_each_active_range_index_in_nid(i, nid) {
2313 unsigned long size_pages = 0;
2314 unsigned long end_pfn = early_node_map[i].end_pfn;
2316 if (early_node_map[i].start_pfn >= max_low_pfn)
2319 if (end_pfn > max_low_pfn)
2320 end_pfn = max_low_pfn;
2322 size_pages = end_pfn - early_node_map[i].start_pfn;
2323 free_bootmem_node(NODE_DATA(early_node_map[i].nid),
2324 PFN_PHYS(early_node_map[i].start_pfn),
2325 size_pages << PAGE_SHIFT);
2330 * sparse_memory_present_with_active_regions - Call memory_present for each active range
2331 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
2333 * If an architecture guarantees that all ranges registered with
2334 * add_active_ranges() contain no holes and may be freed, this
2335 * function may be used instead of calling memory_present() manually.
2337 void __init sparse_memory_present_with_active_regions(int nid)
2341 for_each_active_range_index_in_nid(i, nid)
2342 memory_present(early_node_map[i].nid,
2343 early_node_map[i].start_pfn,
2344 early_node_map[i].end_pfn);
2348 * push_node_boundaries - Push node boundaries to at least the requested boundary
2349 * @nid: The nid of the node to push the boundary for
2350 * @start_pfn: The start pfn of the node
2351 * @end_pfn: The end pfn of the node
2353 * In reserve-based hot-add, mem_map is allocated that is unused until hotadd
2354 * time. Specifically, on x86_64, SRAT will report ranges that can potentially
2355 * be hotplugged even though no physical memory exists. This function allows
2356 * an arch to push out the node boundaries so mem_map is allocated that can
2359 #ifdef CONFIG_MEMORY_HOTPLUG_RESERVE
2360 void __init push_node_boundaries(unsigned int nid,
2361 unsigned long start_pfn, unsigned long end_pfn)
2363 printk(KERN_DEBUG "Entering push_node_boundaries(%u, %lu, %lu)\n",
2364 nid, start_pfn, end_pfn);
2366 /* Initialise the boundary for this node if necessary */
2367 if (node_boundary_end_pfn[nid] == 0)
2368 node_boundary_start_pfn[nid] = -1UL;
2370 /* Update the boundaries */
2371 if (node_boundary_start_pfn[nid] > start_pfn)
2372 node_boundary_start_pfn[nid] = start_pfn;
2373 if (node_boundary_end_pfn[nid] < end_pfn)
2374 node_boundary_end_pfn[nid] = end_pfn;
2377 /* If necessary, push the node boundary out for reserve hotadd */
2378 static void __init account_node_boundary(unsigned int nid,
2379 unsigned long *start_pfn, unsigned long *end_pfn)
2381 printk(KERN_DEBUG "Entering account_node_boundary(%u, %lu, %lu)\n",
2382 nid, *start_pfn, *end_pfn);
2384 /* Return if boundary information has not been provided */
2385 if (node_boundary_end_pfn[nid] == 0)
2388 /* Check the boundaries and update if necessary */
2389 if (node_boundary_start_pfn[nid] < *start_pfn)
2390 *start_pfn = node_boundary_start_pfn[nid];
2391 if (node_boundary_end_pfn[nid] > *end_pfn)
2392 *end_pfn = node_boundary_end_pfn[nid];
2395 void __init push_node_boundaries(unsigned int nid,
2396 unsigned long start_pfn, unsigned long end_pfn) {}
2398 static void __init account_node_boundary(unsigned int nid,
2399 unsigned long *start_pfn, unsigned long *end_pfn) {}
2404 * get_pfn_range_for_nid - Return the start and end page frames for a node
2405 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
2406 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
2407 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
2409 * It returns the start and end page frame of a node based on information
2410 * provided by an arch calling add_active_range(). If called for a node
2411 * with no available memory, a warning is printed and the start and end
2414 void __init get_pfn_range_for_nid(unsigned int nid,
2415 unsigned long *start_pfn, unsigned long *end_pfn)
2421 for_each_active_range_index_in_nid(i, nid) {
2422 *start_pfn = min(*start_pfn, early_node_map[i].start_pfn);
2423 *end_pfn = max(*end_pfn, early_node_map[i].end_pfn);
2426 if (*start_pfn == -1UL) {
2427 printk(KERN_WARNING "Node %u active with no memory\n", nid);
2431 /* Push the node boundaries out if requested */
2432 account_node_boundary(nid, start_pfn, end_pfn);
2436 * Return the number of pages a zone spans in a node, including holes
2437 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
2439 unsigned long __init zone_spanned_pages_in_node(int nid,
2440 unsigned long zone_type,
2441 unsigned long *ignored)
2443 unsigned long node_start_pfn, node_end_pfn;
2444 unsigned long zone_start_pfn, zone_end_pfn;
2446 /* Get the start and end of the node and zone */
2447 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
2448 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
2449 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
2451 /* Check that this node has pages within the zone's required range */
2452 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
2455 /* Move the zone boundaries inside the node if necessary */
2456 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
2457 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
2459 /* Return the spanned pages */
2460 return zone_end_pfn - zone_start_pfn;
2464 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
2465 * then all holes in the requested range will be accounted for.
2467 unsigned long __init __absent_pages_in_range(int nid,
2468 unsigned long range_start_pfn,
2469 unsigned long range_end_pfn)
2472 unsigned long prev_end_pfn = 0, hole_pages = 0;
2473 unsigned long start_pfn;
2475 /* Find the end_pfn of the first active range of pfns in the node */
2476 i = first_active_region_index_in_nid(nid);
2480 /* Account for ranges before physical memory on this node */
2481 if (early_node_map[i].start_pfn > range_start_pfn)
2482 hole_pages = early_node_map[i].start_pfn - range_start_pfn;
2484 prev_end_pfn = early_node_map[i].start_pfn;
2486 /* Find all holes for the zone within the node */
2487 for (; i != -1; i = next_active_region_index_in_nid(i, nid)) {
2489 /* No need to continue if prev_end_pfn is outside the zone */
2490 if (prev_end_pfn >= range_end_pfn)
2493 /* Make sure the end of the zone is not within the hole */
2494 start_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
2495 prev_end_pfn = max(prev_end_pfn, range_start_pfn);
2497 /* Update the hole size cound and move on */
2498 if (start_pfn > range_start_pfn) {
2499 BUG_ON(prev_end_pfn > start_pfn);
2500 hole_pages += start_pfn - prev_end_pfn;
2502 prev_end_pfn = early_node_map[i].end_pfn;
2505 /* Account for ranges past physical memory on this node */
2506 if (range_end_pfn > prev_end_pfn)
2507 hole_pages += range_end_pfn -
2508 max(range_start_pfn, prev_end_pfn);
2514 * absent_pages_in_range - Return number of page frames in holes within a range
2515 * @start_pfn: The start PFN to start searching for holes
2516 * @end_pfn: The end PFN to stop searching for holes
2518 * It returns the number of pages frames in memory holes within a range.
2520 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
2521 unsigned long end_pfn)
2523 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
2526 /* Return the number of page frames in holes in a zone on a node */
2527 unsigned long __init zone_absent_pages_in_node(int nid,
2528 unsigned long zone_type,
2529 unsigned long *ignored)
2531 unsigned long node_start_pfn, node_end_pfn;
2532 unsigned long zone_start_pfn, zone_end_pfn;
2534 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
2535 zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type],
2537 zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type],
2540 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
2544 static inline unsigned long zone_spanned_pages_in_node(int nid,
2545 unsigned long zone_type,
2546 unsigned long *zones_size)
2548 return zones_size[zone_type];
2551 static inline unsigned long zone_absent_pages_in_node(int nid,
2552 unsigned long zone_type,
2553 unsigned long *zholes_size)
2558 return zholes_size[zone_type];
2563 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
2564 unsigned long *zones_size, unsigned long *zholes_size)
2566 unsigned long realtotalpages, totalpages = 0;
2569 for (i = 0; i < MAX_NR_ZONES; i++)
2570 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
2572 pgdat->node_spanned_pages = totalpages;
2574 realtotalpages = totalpages;
2575 for (i = 0; i < MAX_NR_ZONES; i++)
2577 zone_absent_pages_in_node(pgdat->node_id, i,
2579 pgdat->node_present_pages = realtotalpages;
2580 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
2585 * Set up the zone data structures:
2586 * - mark all pages reserved
2587 * - mark all memory queues empty
2588 * - clear the memory bitmaps
2590 static void __meminit free_area_init_core(struct pglist_data *pgdat,
2591 unsigned long *zones_size, unsigned long *zholes_size)
2594 int nid = pgdat->node_id;
2595 unsigned long zone_start_pfn = pgdat->node_start_pfn;
2598 pgdat_resize_init(pgdat);
2599 pgdat->nr_zones = 0;
2600 init_waitqueue_head(&pgdat->kswapd_wait);
2601 pgdat->kswapd_max_order = 0;
2603 for (j = 0; j < MAX_NR_ZONES; j++) {
2604 struct zone *zone = pgdat->node_zones + j;
2605 unsigned long size, realsize, memmap_pages;
2607 size = zone_spanned_pages_in_node(nid, j, zones_size);
2608 realsize = size - zone_absent_pages_in_node(nid, j,
2612 * Adjust realsize so that it accounts for how much memory
2613 * is used by this zone for memmap. This affects the watermark
2614 * and per-cpu initialisations
2616 memmap_pages = (size * sizeof(struct page)) >> PAGE_SHIFT;
2617 if (realsize >= memmap_pages) {
2618 realsize -= memmap_pages;
2620 " %s zone: %lu pages used for memmap\n",
2621 zone_names[j], memmap_pages);
2624 " %s zone: %lu pages exceeds realsize %lu\n",
2625 zone_names[j], memmap_pages, realsize);
2627 /* Account for reserved DMA pages */
2628 if (j == ZONE_DMA && realsize > dma_reserve) {
2629 realsize -= dma_reserve;
2630 printk(KERN_DEBUG " DMA zone: %lu pages reserved\n",
2634 if (!is_highmem_idx(j))
2635 nr_kernel_pages += realsize;
2636 nr_all_pages += realsize;
2638 zone->spanned_pages = size;
2639 zone->present_pages = realsize;
2642 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
2644 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
2646 zone->name = zone_names[j];
2647 spin_lock_init(&zone->lock);
2648 spin_lock_init(&zone->lru_lock);
2649 zone_seqlock_init(zone);
2650 zone->zone_pgdat = pgdat;
2652 zone->prev_priority = DEF_PRIORITY;
2654 zone_pcp_init(zone);
2655 INIT_LIST_HEAD(&zone->active_list);
2656 INIT_LIST_HEAD(&zone->inactive_list);
2657 zone->nr_scan_active = 0;
2658 zone->nr_scan_inactive = 0;
2659 zap_zone_vm_stats(zone);
2660 atomic_set(&zone->reclaim_in_progress, 0);
2664 ret = init_currently_empty_zone(zone, zone_start_pfn,
2665 size, MEMMAP_EARLY);
2667 zone_start_pfn += size;
2671 static void __init alloc_node_mem_map(struct pglist_data *pgdat)
2673 /* Skip empty nodes */
2674 if (!pgdat->node_spanned_pages)
2677 #ifdef CONFIG_FLAT_NODE_MEM_MAP
2678 /* ia64 gets its own node_mem_map, before this, without bootmem */
2679 if (!pgdat->node_mem_map) {
2680 unsigned long size, start, end;
2684 * The zone's endpoints aren't required to be MAX_ORDER
2685 * aligned but the node_mem_map endpoints must be in order
2686 * for the buddy allocator to function correctly.
2688 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
2689 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
2690 end = ALIGN(end, MAX_ORDER_NR_PAGES);
2691 size = (end - start) * sizeof(struct page);
2692 map = alloc_remap(pgdat->node_id, size);
2694 map = alloc_bootmem_node(pgdat, size);
2695 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
2697 #ifdef CONFIG_FLATMEM
2699 * With no DISCONTIG, the global mem_map is just set as node 0's
2701 if (pgdat == NODE_DATA(0)) {
2702 mem_map = NODE_DATA(0)->node_mem_map;
2703 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
2704 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
2705 mem_map -= pgdat->node_start_pfn;
2706 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
2709 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
2712 void __meminit free_area_init_node(int nid, struct pglist_data *pgdat,
2713 unsigned long *zones_size, unsigned long node_start_pfn,
2714 unsigned long *zholes_size)
2716 pgdat->node_id = nid;
2717 pgdat->node_start_pfn = node_start_pfn;
2718 calculate_node_totalpages(pgdat, zones_size, zholes_size);
2720 alloc_node_mem_map(pgdat);
2722 free_area_init_core(pgdat, zones_size, zholes_size);
2725 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
2727 * add_active_range - Register a range of PFNs backed by physical memory
2728 * @nid: The node ID the range resides on
2729 * @start_pfn: The start PFN of the available physical memory
2730 * @end_pfn: The end PFN of the available physical memory
2732 * These ranges are stored in an early_node_map[] and later used by
2733 * free_area_init_nodes() to calculate zone sizes and holes. If the
2734 * range spans a memory hole, it is up to the architecture to ensure
2735 * the memory is not freed by the bootmem allocator. If possible
2736 * the range being registered will be merged with existing ranges.
2738 void __init add_active_range(unsigned int nid, unsigned long start_pfn,
2739 unsigned long end_pfn)
2743 printk(KERN_DEBUG "Entering add_active_range(%d, %lu, %lu) "
2744 "%d entries of %d used\n",
2745 nid, start_pfn, end_pfn,
2746 nr_nodemap_entries, MAX_ACTIVE_REGIONS);
2748 /* Merge with existing active regions if possible */
2749 for (i = 0; i < nr_nodemap_entries; i++) {
2750 if (early_node_map[i].nid != nid)
2753 /* Skip if an existing region covers this new one */
2754 if (start_pfn >= early_node_map[i].start_pfn &&
2755 end_pfn <= early_node_map[i].end_pfn)
2758 /* Merge forward if suitable */
2759 if (start_pfn <= early_node_map[i].end_pfn &&
2760 end_pfn > early_node_map[i].end_pfn) {
2761 early_node_map[i].end_pfn = end_pfn;
2765 /* Merge backward if suitable */
2766 if (start_pfn < early_node_map[i].end_pfn &&
2767 end_pfn >= early_node_map[i].start_pfn) {
2768 early_node_map[i].start_pfn = start_pfn;
2773 /* Check that early_node_map is large enough */
2774 if (i >= MAX_ACTIVE_REGIONS) {
2775 printk(KERN_CRIT "More than %d memory regions, truncating\n",
2776 MAX_ACTIVE_REGIONS);
2780 early_node_map[i].nid = nid;
2781 early_node_map[i].start_pfn = start_pfn;
2782 early_node_map[i].end_pfn = end_pfn;
2783 nr_nodemap_entries = i + 1;
2787 * shrink_active_range - Shrink an existing registered range of PFNs
2788 * @nid: The node id the range is on that should be shrunk
2789 * @old_end_pfn: The old end PFN of the range
2790 * @new_end_pfn: The new PFN of the range
2792 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
2793 * The map is kept at the end physical page range that has already been
2794 * registered with add_active_range(). This function allows an arch to shrink
2795 * an existing registered range.
2797 void __init shrink_active_range(unsigned int nid, unsigned long old_end_pfn,
2798 unsigned long new_end_pfn)
2802 /* Find the old active region end and shrink */
2803 for_each_active_range_index_in_nid(i, nid)
2804 if (early_node_map[i].end_pfn == old_end_pfn) {
2805 early_node_map[i].end_pfn = new_end_pfn;
2811 * remove_all_active_ranges - Remove all currently registered regions
2813 * During discovery, it may be found that a table like SRAT is invalid
2814 * and an alternative discovery method must be used. This function removes
2815 * all currently registered regions.
2817 void __init remove_all_active_ranges(void)
2819 memset(early_node_map, 0, sizeof(early_node_map));
2820 nr_nodemap_entries = 0;
2821 #ifdef CONFIG_MEMORY_HOTPLUG_RESERVE
2822 memset(node_boundary_start_pfn, 0, sizeof(node_boundary_start_pfn));
2823 memset(node_boundary_end_pfn, 0, sizeof(node_boundary_end_pfn));
2824 #endif /* CONFIG_MEMORY_HOTPLUG_RESERVE */
2827 /* Compare two active node_active_regions */
2828 static int __init cmp_node_active_region(const void *a, const void *b)
2830 struct node_active_region *arange = (struct node_active_region *)a;
2831 struct node_active_region *brange = (struct node_active_region *)b;
2833 /* Done this way to avoid overflows */
2834 if (arange->start_pfn > brange->start_pfn)
2836 if (arange->start_pfn < brange->start_pfn)
2842 /* sort the node_map by start_pfn */
2843 static void __init sort_node_map(void)
2845 sort(early_node_map, (size_t)nr_nodemap_entries,
2846 sizeof(struct node_active_region),
2847 cmp_node_active_region, NULL);
2850 /* Find the lowest pfn for a node */
2851 unsigned long __init find_min_pfn_for_node(unsigned long nid)
2854 unsigned long min_pfn = ULONG_MAX;
2856 /* Assuming a sorted map, the first range found has the starting pfn */
2857 for_each_active_range_index_in_nid(i, nid)
2858 min_pfn = min(min_pfn, early_node_map[i].start_pfn);
2860 if (min_pfn == ULONG_MAX) {
2862 "Could not find start_pfn for node %lu\n", nid);
2870 * find_min_pfn_with_active_regions - Find the minimum PFN registered
2872 * It returns the minimum PFN based on information provided via
2873 * add_active_range().
2875 unsigned long __init find_min_pfn_with_active_regions(void)
2877 return find_min_pfn_for_node(MAX_NUMNODES);
2881 * find_max_pfn_with_active_regions - Find the maximum PFN registered
2883 * It returns the maximum PFN based on information provided via
2884 * add_active_range().
2886 unsigned long __init find_max_pfn_with_active_regions(void)
2889 unsigned long max_pfn = 0;
2891 for (i = 0; i < nr_nodemap_entries; i++)
2892 max_pfn = max(max_pfn, early_node_map[i].end_pfn);
2898 * free_area_init_nodes - Initialise all pg_data_t and zone data
2899 * @max_zone_pfn: an array of max PFNs for each zone
2901 * This will call free_area_init_node() for each active node in the system.
2902 * Using the page ranges provided by add_active_range(), the size of each
2903 * zone in each node and their holes is calculated. If the maximum PFN
2904 * between two adjacent zones match, it is assumed that the zone is empty.
2905 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
2906 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
2907 * starts where the previous one ended. For example, ZONE_DMA32 starts
2908 * at arch_max_dma_pfn.
2910 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
2915 /* Sort early_node_map as initialisation assumes it is sorted */
2918 /* Record where the zone boundaries are */
2919 memset(arch_zone_lowest_possible_pfn, 0,
2920 sizeof(arch_zone_lowest_possible_pfn));
2921 memset(arch_zone_highest_possible_pfn, 0,
2922 sizeof(arch_zone_highest_possible_pfn));
2923 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
2924 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
2925 for (i = 1; i < MAX_NR_ZONES; i++) {
2926 arch_zone_lowest_possible_pfn[i] =
2927 arch_zone_highest_possible_pfn[i-1];
2928 arch_zone_highest_possible_pfn[i] =
2929 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
2932 /* Print out the zone ranges */
2933 printk("Zone PFN ranges:\n");
2934 for (i = 0; i < MAX_NR_ZONES; i++)
2935 printk(" %-8s %8lu -> %8lu\n",
2937 arch_zone_lowest_possible_pfn[i],
2938 arch_zone_highest_possible_pfn[i]);
2940 /* Print out the early_node_map[] */
2941 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries);
2942 for (i = 0; i < nr_nodemap_entries; i++)
2943 printk(" %3d: %8lu -> %8lu\n", early_node_map[i].nid,
2944 early_node_map[i].start_pfn,
2945 early_node_map[i].end_pfn);
2947 /* Initialise every node */
2948 for_each_online_node(nid) {
2949 pg_data_t *pgdat = NODE_DATA(nid);
2950 free_area_init_node(nid, pgdat, NULL,
2951 find_min_pfn_for_node(nid), NULL);
2954 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
2957 * set_dma_reserve - set the specified number of pages reserved in the first zone
2958 * @new_dma_reserve: The number of pages to mark reserved
2960 * The per-cpu batchsize and zone watermarks are determined by present_pages.
2961 * In the DMA zone, a significant percentage may be consumed by kernel image
2962 * and other unfreeable allocations which can skew the watermarks badly. This
2963 * function may optionally be used to account for unfreeable pages in the
2964 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
2965 * smaller per-cpu batchsize.
2967 void __init set_dma_reserve(unsigned long new_dma_reserve)
2969 dma_reserve = new_dma_reserve;
2972 #ifndef CONFIG_NEED_MULTIPLE_NODES
2973 static bootmem_data_t contig_bootmem_data;
2974 struct pglist_data contig_page_data = { .bdata = &contig_bootmem_data };
2976 EXPORT_SYMBOL(contig_page_data);
2979 void __init free_area_init(unsigned long *zones_size)
2981 free_area_init_node(0, NODE_DATA(0), zones_size,
2982 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
2985 static int page_alloc_cpu_notify(struct notifier_block *self,
2986 unsigned long action, void *hcpu)
2988 int cpu = (unsigned long)hcpu;
2990 if (action == CPU_DEAD) {
2991 local_irq_disable();
2993 vm_events_fold_cpu(cpu);
2995 refresh_cpu_vm_stats(cpu);
3000 void __init page_alloc_init(void)
3002 hotcpu_notifier(page_alloc_cpu_notify, 0);
3006 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
3007 * or min_free_kbytes changes.
3009 static void calculate_totalreserve_pages(void)
3011 struct pglist_data *pgdat;
3012 unsigned long reserve_pages = 0;
3013 enum zone_type i, j;
3015 for_each_online_pgdat(pgdat) {
3016 for (i = 0; i < MAX_NR_ZONES; i++) {
3017 struct zone *zone = pgdat->node_zones + i;
3018 unsigned long max = 0;
3020 /* Find valid and maximum lowmem_reserve in the zone */
3021 for (j = i; j < MAX_NR_ZONES; j++) {
3022 if (zone->lowmem_reserve[j] > max)
3023 max = zone->lowmem_reserve[j];
3026 /* we treat pages_high as reserved pages. */
3027 max += zone->pages_high;
3029 if (max > zone->present_pages)
3030 max = zone->present_pages;
3031 reserve_pages += max;
3034 totalreserve_pages = reserve_pages;
3038 * setup_per_zone_lowmem_reserve - called whenever
3039 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
3040 * has a correct pages reserved value, so an adequate number of
3041 * pages are left in the zone after a successful __alloc_pages().
3043 static void setup_per_zone_lowmem_reserve(void)
3045 struct pglist_data *pgdat;
3046 enum zone_type j, idx;
3048 for_each_online_pgdat(pgdat) {
3049 for (j = 0; j < MAX_NR_ZONES; j++) {
3050 struct zone *zone = pgdat->node_zones + j;
3051 unsigned long present_pages = zone->present_pages;
3053 zone->lowmem_reserve[j] = 0;
3057 struct zone *lower_zone;
3061 if (sysctl_lowmem_reserve_ratio[idx] < 1)
3062 sysctl_lowmem_reserve_ratio[idx] = 1;
3064 lower_zone = pgdat->node_zones + idx;
3065 lower_zone->lowmem_reserve[j] = present_pages /
3066 sysctl_lowmem_reserve_ratio[idx];
3067 present_pages += lower_zone->present_pages;
3072 /* update totalreserve_pages */
3073 calculate_totalreserve_pages();
3077 * setup_per_zone_pages_min - called when min_free_kbytes changes.
3079 * Ensures that the pages_{min,low,high} values for each zone are set correctly
3080 * with respect to min_free_kbytes.
3082 void setup_per_zone_pages_min(void)
3084 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
3085 unsigned long lowmem_pages = 0;
3087 unsigned long flags;
3089 /* Calculate total number of !ZONE_HIGHMEM pages */
3090 for_each_zone(zone) {
3091 if (!is_highmem(zone))
3092 lowmem_pages += zone->present_pages;
3095 for_each_zone(zone) {
3098 spin_lock_irqsave(&zone->lru_lock, flags);
3099 tmp = (u64)pages_min * zone->present_pages;
3100 do_div(tmp, lowmem_pages);
3101 if (is_highmem(zone)) {
3103 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
3104 * need highmem pages, so cap pages_min to a small
3107 * The (pages_high-pages_low) and (pages_low-pages_min)
3108 * deltas controls asynch page reclaim, and so should
3109 * not be capped for highmem.
3113 min_pages = zone->present_pages / 1024;
3114 if (min_pages < SWAP_CLUSTER_MAX)
3115 min_pages = SWAP_CLUSTER_MAX;
3116 if (min_pages > 128)
3118 zone->pages_min = min_pages;
3121 * If it's a lowmem zone, reserve a number of pages
3122 * proportionate to the zone's size.
3124 zone->pages_min = tmp;
3127 zone->pages_low = zone->pages_min + (tmp >> 2);
3128 zone->pages_high = zone->pages_min + (tmp >> 1);
3129 spin_unlock_irqrestore(&zone->lru_lock, flags);
3132 /* update totalreserve_pages */
3133 calculate_totalreserve_pages();
3137 * Initialise min_free_kbytes.
3139 * For small machines we want it small (128k min). For large machines
3140 * we want it large (64MB max). But it is not linear, because network
3141 * bandwidth does not increase linearly with machine size. We use
3143 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
3144 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
3160 static int __init init_per_zone_pages_min(void)
3162 unsigned long lowmem_kbytes;
3164 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
3166 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
3167 if (min_free_kbytes < 128)
3168 min_free_kbytes = 128;
3169 if (min_free_kbytes > 65536)
3170 min_free_kbytes = 65536;
3171 setup_per_zone_pages_min();
3172 setup_per_zone_lowmem_reserve();
3175 module_init(init_per_zone_pages_min)
3178 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
3179 * that we can call two helper functions whenever min_free_kbytes
3182 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
3183 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
3185 proc_dointvec(table, write, file, buffer, length, ppos);
3186 setup_per_zone_pages_min();
3191 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
3192 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
3197 rc = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
3202 zone->min_unmapped_pages = (zone->present_pages *
3203 sysctl_min_unmapped_ratio) / 100;
3207 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
3208 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
3213 rc = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
3218 zone->min_slab_pages = (zone->present_pages *
3219 sysctl_min_slab_ratio) / 100;
3225 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
3226 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
3227 * whenever sysctl_lowmem_reserve_ratio changes.
3229 * The reserve ratio obviously has absolutely no relation with the
3230 * pages_min watermarks. The lowmem reserve ratio can only make sense
3231 * if in function of the boot time zone sizes.
3233 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
3234 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
3236 proc_dointvec_minmax(table, write, file, buffer, length, ppos);
3237 setup_per_zone_lowmem_reserve();
3242 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
3243 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
3244 * can have before it gets flushed back to buddy allocator.
3247 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
3248 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
3254 ret = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
3255 if (!write || (ret == -EINVAL))
3257 for_each_zone(zone) {
3258 for_each_online_cpu(cpu) {
3260 high = zone->present_pages / percpu_pagelist_fraction;
3261 setup_pagelist_highmark(zone_pcp(zone, cpu), high);
3267 int hashdist = HASHDIST_DEFAULT;
3270 static int __init set_hashdist(char *str)
3274 hashdist = simple_strtoul(str, &str, 0);
3277 __setup("hashdist=", set_hashdist);
3281 * allocate a large system hash table from bootmem
3282 * - it is assumed that the hash table must contain an exact power-of-2
3283 * quantity of entries
3284 * - limit is the number of hash buckets, not the total allocation size
3286 void *__init alloc_large_system_hash(const char *tablename,
3287 unsigned long bucketsize,
3288 unsigned long numentries,
3291 unsigned int *_hash_shift,
3292 unsigned int *_hash_mask,
3293 unsigned long limit)
3295 unsigned long long max = limit;
3296 unsigned long log2qty, size;
3299 /* allow the kernel cmdline to have a say */
3301 /* round applicable memory size up to nearest megabyte */
3302 numentries = nr_kernel_pages;
3303 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
3304 numentries >>= 20 - PAGE_SHIFT;
3305 numentries <<= 20 - PAGE_SHIFT;
3307 /* limit to 1 bucket per 2^scale bytes of low memory */
3308 if (scale > PAGE_SHIFT)
3309 numentries >>= (scale - PAGE_SHIFT);
3311 numentries <<= (PAGE_SHIFT - scale);
3313 /* Make sure we've got at least a 0-order allocation.. */
3314 if (unlikely((numentries * bucketsize) < PAGE_SIZE))
3315 numentries = PAGE_SIZE / bucketsize;
3317 numentries = roundup_pow_of_two(numentries);
3319 /* limit allocation size to 1/16 total memory by default */
3321 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
3322 do_div(max, bucketsize);
3325 if (numentries > max)
3328 log2qty = ilog2(numentries);
3331 size = bucketsize << log2qty;
3332 if (flags & HASH_EARLY)
3333 table = alloc_bootmem(size);
3335 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
3337 unsigned long order;
3338 for (order = 0; ((1UL << order) << PAGE_SHIFT) < size; order++)
3340 table = (void*) __get_free_pages(GFP_ATOMIC, order);
3342 } while (!table && size > PAGE_SIZE && --log2qty);
3345 panic("Failed to allocate %s hash table\n", tablename);
3347 printk("%s hash table entries: %d (order: %d, %lu bytes)\n",
3350 ilog2(size) - PAGE_SHIFT,
3354 *_hash_shift = log2qty;
3356 *_hash_mask = (1 << log2qty) - 1;
3361 #ifdef CONFIG_OUT_OF_LINE_PFN_TO_PAGE
3362 struct page *pfn_to_page(unsigned long pfn)
3364 return __pfn_to_page(pfn);
3366 unsigned long page_to_pfn(struct page *page)
3368 return __page_to_pfn(page);
3370 EXPORT_SYMBOL(pfn_to_page);
3371 EXPORT_SYMBOL(page_to_pfn);
3372 #endif /* CONFIG_OUT_OF_LINE_PFN_TO_PAGE */
3374 #if MAX_NUMNODES > 1
3376 * Find the highest possible node id.
3378 int highest_possible_node_id(void)
3381 unsigned int highest = 0;
3383 for_each_node_mask(node, node_possible_map)
3387 EXPORT_SYMBOL(highest_possible_node_id);