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>
41 #include <asm/tlbflush.h>
42 #include <asm/div64.h>
46 * MCD - HACK: Find somewhere to initialize this EARLY, or make this
49 nodemask_t node_online_map __read_mostly = { { [0] = 1UL } };
50 EXPORT_SYMBOL(node_online_map);
51 nodemask_t node_possible_map __read_mostly = NODE_MASK_ALL;
52 EXPORT_SYMBOL(node_possible_map);
53 unsigned long totalram_pages __read_mostly;
54 unsigned long totalhigh_pages __read_mostly;
55 unsigned long totalreserve_pages __read_mostly;
57 int percpu_pagelist_fraction;
59 static void __free_pages_ok(struct page *page, unsigned int order);
62 * results with 256, 32 in the lowmem_reserve sysctl:
63 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
64 * 1G machine -> (16M dma, 784M normal, 224M high)
65 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
66 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
67 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
69 * TBD: should special case ZONE_DMA32 machines here - in those we normally
70 * don't need any ZONE_NORMAL reservation
72 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = { 256, 256, 32 };
74 EXPORT_SYMBOL(totalram_pages);
77 * Used by page_zone() to look up the address of the struct zone whose
78 * id is encoded in the upper bits of page->flags
80 struct zone *zone_table[1 << ZONETABLE_SHIFT] __read_mostly;
81 EXPORT_SYMBOL(zone_table);
83 static char *zone_names[MAX_NR_ZONES] = { "DMA", "DMA32", "Normal", "HighMem" };
84 int min_free_kbytes = 1024;
86 unsigned long __meminitdata nr_kernel_pages;
87 unsigned long __meminitdata nr_all_pages;
89 #ifdef CONFIG_DEBUG_VM
90 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
94 unsigned long pfn = page_to_pfn(page);
97 seq = zone_span_seqbegin(zone);
98 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
100 else if (pfn < zone->zone_start_pfn)
102 } while (zone_span_seqretry(zone, seq));
107 static int page_is_consistent(struct zone *zone, struct page *page)
109 #ifdef CONFIG_HOLES_IN_ZONE
110 if (!pfn_valid(page_to_pfn(page)))
113 if (zone != page_zone(page))
119 * Temporary debugging check for pages not lying within a given zone.
121 static int bad_range(struct zone *zone, struct page *page)
123 if (page_outside_zone_boundaries(zone, page))
125 if (!page_is_consistent(zone, page))
132 static inline int bad_range(struct zone *zone, struct page *page)
138 static void bad_page(struct page *page)
140 printk(KERN_EMERG "Bad page state in process '%s'\n"
141 KERN_EMERG "page:%p flags:0x%0*lx mapping:%p mapcount:%d count:%d (%s)\n"
142 KERN_EMERG "Trying to fix it up, but a reboot is needed\n"
143 KERN_EMERG "Backtrace:\n",
144 current->comm, page, (int)(2*sizeof(unsigned long)),
145 (unsigned long)page->flags, page->mapping,
146 page_mapcount(page), page_count(page), print_tainted());
148 page->flags &= ~(1 << PG_lru |
158 set_page_count(page, 0);
159 reset_page_mapcount(page);
160 page->mapping = NULL;
161 add_taint(TAINT_BAD_PAGE);
165 * Higher-order pages are called "compound pages". They are structured thusly:
167 * The first PAGE_SIZE page is called the "head page".
169 * The remaining PAGE_SIZE pages are called "tail pages".
171 * All pages have PG_compound set. All pages have their ->private pointing at
172 * the head page (even the head page has this).
174 * The first tail page's ->lru.next holds the address of the compound page's
175 * put_page() function. Its ->lru.prev holds the order of allocation.
176 * This usage means that zero-order pages may not be compound.
179 static void free_compound_page(struct page *page)
181 __free_pages_ok(page, (unsigned long)page[1].lru.prev);
184 static void prep_compound_page(struct page *page, unsigned long order)
187 int nr_pages = 1 << order;
189 page[1].lru.next = (void *)free_compound_page; /* set dtor */
190 page[1].lru.prev = (void *)order;
191 for (i = 0; i < nr_pages; i++) {
192 struct page *p = page + i;
194 __SetPageCompound(p);
195 set_page_private(p, (unsigned long)page);
199 static void destroy_compound_page(struct page *page, unsigned long order)
202 int nr_pages = 1 << order;
204 if (unlikely((unsigned long)page[1].lru.prev != order))
207 for (i = 0; i < nr_pages; i++) {
208 struct page *p = page + i;
210 if (unlikely(!PageCompound(p) |
211 (page_private(p) != (unsigned long)page)))
213 __ClearPageCompound(p);
217 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
221 BUG_ON((gfp_flags & (__GFP_WAIT | __GFP_HIGHMEM)) == __GFP_HIGHMEM);
223 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
224 * and __GFP_HIGHMEM from hard or soft interrupt context.
226 BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
227 for (i = 0; i < (1 << order); i++)
228 clear_highpage(page + i);
232 * function for dealing with page's order in buddy system.
233 * zone->lock is already acquired when we use these.
234 * So, we don't need atomic page->flags operations here.
236 static inline unsigned long page_order(struct page *page)
238 return page_private(page);
241 static inline void set_page_order(struct page *page, int order)
243 set_page_private(page, order);
244 __SetPageBuddy(page);
247 static inline void rmv_page_order(struct page *page)
249 __ClearPageBuddy(page);
250 set_page_private(page, 0);
254 * Locate the struct page for both the matching buddy in our
255 * pair (buddy1) and the combined O(n+1) page they form (page).
257 * 1) Any buddy B1 will have an order O twin B2 which satisfies
258 * the following equation:
260 * For example, if the starting buddy (buddy2) is #8 its order
262 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
264 * 2) Any buddy B will have an order O+1 parent P which
265 * satisfies the following equation:
268 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
270 static inline struct page *
271 __page_find_buddy(struct page *page, unsigned long page_idx, unsigned int order)
273 unsigned long buddy_idx = page_idx ^ (1 << order);
275 return page + (buddy_idx - page_idx);
278 static inline unsigned long
279 __find_combined_index(unsigned long page_idx, unsigned int order)
281 return (page_idx & ~(1 << order));
285 * This function checks whether a page is free && is the buddy
286 * we can do coalesce a page and its buddy if
287 * (a) the buddy is not in a hole &&
288 * (b) the buddy is in the buddy system &&
289 * (c) a page and its buddy have the same order &&
290 * (d) a page and its buddy are in the same zone.
292 * For recording whether a page is in the buddy system, we use PG_buddy.
293 * Setting, clearing, and testing PG_buddy is serialized by zone->lock.
295 * For recording page's order, we use page_private(page).
297 static inline int page_is_buddy(struct page *page, struct page *buddy,
300 #ifdef CONFIG_HOLES_IN_ZONE
301 if (!pfn_valid(page_to_pfn(buddy)))
305 if (page_zone_id(page) != page_zone_id(buddy))
308 if (PageBuddy(buddy) && page_order(buddy) == order) {
309 BUG_ON(page_count(buddy) != 0);
316 * Freeing function for a buddy system allocator.
318 * The concept of a buddy system is to maintain direct-mapped table
319 * (containing bit values) for memory blocks of various "orders".
320 * The bottom level table contains the map for the smallest allocatable
321 * units of memory (here, pages), and each level above it describes
322 * pairs of units from the levels below, hence, "buddies".
323 * At a high level, all that happens here is marking the table entry
324 * at the bottom level available, and propagating the changes upward
325 * as necessary, plus some accounting needed to play nicely with other
326 * parts of the VM system.
327 * At each level, we keep a list of pages, which are heads of continuous
328 * free pages of length of (1 << order) and marked with PG_buddy. Page's
329 * order is recorded in page_private(page) field.
330 * So when we are allocating or freeing one, we can derive the state of the
331 * other. That is, if we allocate a small block, and both were
332 * free, the remainder of the region must be split into blocks.
333 * If a block is freed, and its buddy is also free, then this
334 * triggers coalescing into a block of larger size.
339 static inline void __free_one_page(struct page *page,
340 struct zone *zone, unsigned int order)
342 unsigned long page_idx;
343 int order_size = 1 << order;
345 if (unlikely(PageCompound(page)))
346 destroy_compound_page(page, order);
348 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
350 BUG_ON(page_idx & (order_size - 1));
351 BUG_ON(bad_range(zone, page));
353 zone->free_pages += order_size;
354 while (order < MAX_ORDER-1) {
355 unsigned long combined_idx;
356 struct free_area *area;
359 buddy = __page_find_buddy(page, page_idx, order);
360 if (!page_is_buddy(page, buddy, order))
361 break; /* Move the buddy up one level. */
363 list_del(&buddy->lru);
364 area = zone->free_area + order;
366 rmv_page_order(buddy);
367 combined_idx = __find_combined_index(page_idx, order);
368 page = page + (combined_idx - page_idx);
369 page_idx = combined_idx;
372 set_page_order(page, order);
373 list_add(&page->lru, &zone->free_area[order].free_list);
374 zone->free_area[order].nr_free++;
377 static inline int free_pages_check(struct page *page)
379 if (unlikely(page_mapcount(page) |
380 (page->mapping != NULL) |
381 (page_count(page) != 0) |
395 __ClearPageDirty(page);
397 * For now, we report if PG_reserved was found set, but do not
398 * clear it, and do not free the page. But we shall soon need
399 * to do more, for when the ZERO_PAGE count wraps negative.
401 return PageReserved(page);
405 * Frees a list of pages.
406 * Assumes all pages on list are in same zone, and of same order.
407 * count is the number of pages to free.
409 * If the zone was previously in an "all pages pinned" state then look to
410 * see if this freeing clears that state.
412 * And clear the zone's pages_scanned counter, to hold off the "all pages are
413 * pinned" detection logic.
415 static void free_pages_bulk(struct zone *zone, int count,
416 struct list_head *list, int order)
418 spin_lock(&zone->lock);
419 zone->all_unreclaimable = 0;
420 zone->pages_scanned = 0;
424 BUG_ON(list_empty(list));
425 page = list_entry(list->prev, struct page, lru);
426 /* have to delete it as __free_one_page list manipulates */
427 list_del(&page->lru);
428 __free_one_page(page, zone, order);
430 spin_unlock(&zone->lock);
433 static void free_one_page(struct zone *zone, struct page *page, int order)
436 list_add(&page->lru, &list);
437 free_pages_bulk(zone, 1, &list, order);
440 static void __free_pages_ok(struct page *page, unsigned int order)
446 if (arch_free_page(page, order))
448 if (!PageHighMem(page))
449 debug_check_no_locks_freed(page_address(page),
452 for (i = 0 ; i < (1 << order) ; ++i)
453 reserved += free_pages_check(page + i);
457 kernel_map_pages(page, 1 << order, 0);
458 local_irq_save(flags);
459 __count_vm_events(PGFREE, 1 << order);
460 free_one_page(page_zone(page), page, order);
461 local_irq_restore(flags);
465 * permit the bootmem allocator to evade page validation on high-order frees
467 void fastcall __init __free_pages_bootmem(struct page *page, unsigned int order)
470 __ClearPageReserved(page);
471 set_page_count(page, 0);
472 set_page_refcounted(page);
478 for (loop = 0; loop < BITS_PER_LONG; loop++) {
479 struct page *p = &page[loop];
481 if (loop + 1 < BITS_PER_LONG)
483 __ClearPageReserved(p);
484 set_page_count(p, 0);
487 set_page_refcounted(page);
488 __free_pages(page, order);
494 * The order of subdivision here is critical for the IO subsystem.
495 * Please do not alter this order without good reasons and regression
496 * testing. Specifically, as large blocks of memory are subdivided,
497 * the order in which smaller blocks are delivered depends on the order
498 * they're subdivided in this function. This is the primary factor
499 * influencing the order in which pages are delivered to the IO
500 * subsystem according to empirical testing, and this is also justified
501 * by considering the behavior of a buddy system containing a single
502 * large block of memory acted on by a series of small allocations.
503 * This behavior is a critical factor in sglist merging's success.
507 static inline void expand(struct zone *zone, struct page *page,
508 int low, int high, struct free_area *area)
510 unsigned long size = 1 << high;
516 BUG_ON(bad_range(zone, &page[size]));
517 list_add(&page[size].lru, &area->free_list);
519 set_page_order(&page[size], high);
524 * This page is about to be returned from the page allocator
526 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
528 if (unlikely(page_mapcount(page) |
529 (page->mapping != NULL) |
530 (page_count(page) != 0) |
546 * For now, we report if PG_reserved was found set, but do not
547 * clear it, and do not allocate the page: as a safety net.
549 if (PageReserved(page))
552 page->flags &= ~(1 << PG_uptodate | 1 << PG_error |
553 1 << PG_referenced | 1 << PG_arch_1 |
554 1 << PG_fs_misc | 1 << PG_mappedtodisk);
555 set_page_private(page, 0);
556 set_page_refcounted(page);
557 kernel_map_pages(page, 1 << order, 1);
559 if (gfp_flags & __GFP_ZERO)
560 prep_zero_page(page, order, gfp_flags);
562 if (order && (gfp_flags & __GFP_COMP))
563 prep_compound_page(page, order);
569 * Do the hard work of removing an element from the buddy allocator.
570 * Call me with the zone->lock already held.
572 static struct page *__rmqueue(struct zone *zone, unsigned int order)
574 struct free_area * area;
575 unsigned int current_order;
578 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
579 area = zone->free_area + current_order;
580 if (list_empty(&area->free_list))
583 page = list_entry(area->free_list.next, struct page, lru);
584 list_del(&page->lru);
585 rmv_page_order(page);
587 zone->free_pages -= 1UL << order;
588 expand(zone, page, order, current_order, area);
596 * Obtain a specified number of elements from the buddy allocator, all under
597 * a single hold of the lock, for efficiency. Add them to the supplied list.
598 * Returns the number of new pages which were placed at *list.
600 static int rmqueue_bulk(struct zone *zone, unsigned int order,
601 unsigned long count, struct list_head *list)
605 spin_lock(&zone->lock);
606 for (i = 0; i < count; ++i) {
607 struct page *page = __rmqueue(zone, order);
608 if (unlikely(page == NULL))
610 list_add_tail(&page->lru, list);
612 spin_unlock(&zone->lock);
618 * Called from the slab reaper to drain pagesets on a particular node that
619 * belong to the currently executing processor.
620 * Note that this function must be called with the thread pinned to
621 * a single processor.
623 void drain_node_pages(int nodeid)
628 for (z = 0; z < MAX_NR_ZONES; z++) {
629 struct zone *zone = NODE_DATA(nodeid)->node_zones + z;
630 struct per_cpu_pageset *pset;
632 pset = zone_pcp(zone, smp_processor_id());
633 for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
634 struct per_cpu_pages *pcp;
638 local_irq_save(flags);
639 free_pages_bulk(zone, pcp->count, &pcp->list, 0);
641 local_irq_restore(flags);
648 #if defined(CONFIG_PM) || defined(CONFIG_HOTPLUG_CPU)
649 static void __drain_pages(unsigned int cpu)
655 for_each_zone(zone) {
656 struct per_cpu_pageset *pset;
658 pset = zone_pcp(zone, cpu);
659 for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
660 struct per_cpu_pages *pcp;
663 local_irq_save(flags);
664 free_pages_bulk(zone, pcp->count, &pcp->list, 0);
666 local_irq_restore(flags);
670 #endif /* CONFIG_PM || CONFIG_HOTPLUG_CPU */
674 void mark_free_pages(struct zone *zone)
676 unsigned long zone_pfn, flags;
678 struct list_head *curr;
680 if (!zone->spanned_pages)
683 spin_lock_irqsave(&zone->lock, flags);
684 for (zone_pfn = 0; zone_pfn < zone->spanned_pages; ++zone_pfn)
685 ClearPageNosaveFree(pfn_to_page(zone_pfn + zone->zone_start_pfn));
687 for (order = MAX_ORDER - 1; order >= 0; --order)
688 list_for_each(curr, &zone->free_area[order].free_list) {
689 unsigned long start_pfn, i;
691 start_pfn = page_to_pfn(list_entry(curr, struct page, lru));
693 for (i=0; i < (1<<order); i++)
694 SetPageNosaveFree(pfn_to_page(start_pfn+i));
696 spin_unlock_irqrestore(&zone->lock, flags);
700 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
702 void drain_local_pages(void)
706 local_irq_save(flags);
707 __drain_pages(smp_processor_id());
708 local_irq_restore(flags);
710 #endif /* CONFIG_PM */
713 * Free a 0-order page
715 static void fastcall free_hot_cold_page(struct page *page, int cold)
717 struct zone *zone = page_zone(page);
718 struct per_cpu_pages *pcp;
721 if (arch_free_page(page, 0))
725 page->mapping = NULL;
726 if (free_pages_check(page))
729 kernel_map_pages(page, 1, 0);
731 pcp = &zone_pcp(zone, get_cpu())->pcp[cold];
732 local_irq_save(flags);
733 __count_vm_event(PGFREE);
734 list_add(&page->lru, &pcp->list);
736 if (pcp->count >= pcp->high) {
737 free_pages_bulk(zone, pcp->batch, &pcp->list, 0);
738 pcp->count -= pcp->batch;
740 local_irq_restore(flags);
744 void fastcall free_hot_page(struct page *page)
746 free_hot_cold_page(page, 0);
749 void fastcall free_cold_page(struct page *page)
751 free_hot_cold_page(page, 1);
755 * split_page takes a non-compound higher-order page, and splits it into
756 * n (1<<order) sub-pages: page[0..n]
757 * Each sub-page must be freed individually.
759 * Note: this is probably too low level an operation for use in drivers.
760 * Please consult with lkml before using this in your driver.
762 void split_page(struct page *page, unsigned int order)
766 BUG_ON(PageCompound(page));
767 BUG_ON(!page_count(page));
768 for (i = 1; i < (1 << order); i++)
769 set_page_refcounted(page + i);
773 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
774 * we cheat by calling it from here, in the order > 0 path. Saves a branch
777 static struct page *buffered_rmqueue(struct zonelist *zonelist,
778 struct zone *zone, int order, gfp_t gfp_flags)
782 int cold = !!(gfp_flags & __GFP_COLD);
787 if (likely(order == 0)) {
788 struct per_cpu_pages *pcp;
790 pcp = &zone_pcp(zone, cpu)->pcp[cold];
791 local_irq_save(flags);
793 pcp->count += rmqueue_bulk(zone, 0,
794 pcp->batch, &pcp->list);
795 if (unlikely(!pcp->count))
798 page = list_entry(pcp->list.next, struct page, lru);
799 list_del(&page->lru);
802 spin_lock_irqsave(&zone->lock, flags);
803 page = __rmqueue(zone, order);
804 spin_unlock(&zone->lock);
809 __count_zone_vm_events(PGALLOC, zone, 1 << order);
810 zone_statistics(zonelist, zone);
811 local_irq_restore(flags);
814 BUG_ON(bad_range(zone, page));
815 if (prep_new_page(page, order, gfp_flags))
820 local_irq_restore(flags);
825 #define ALLOC_NO_WATERMARKS 0x01 /* don't check watermarks at all */
826 #define ALLOC_WMARK_MIN 0x02 /* use pages_min watermark */
827 #define ALLOC_WMARK_LOW 0x04 /* use pages_low watermark */
828 #define ALLOC_WMARK_HIGH 0x08 /* use pages_high watermark */
829 #define ALLOC_HARDER 0x10 /* try to alloc harder */
830 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
831 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
834 * Return 1 if free pages are above 'mark'. This takes into account the order
837 int zone_watermark_ok(struct zone *z, int order, unsigned long mark,
838 int classzone_idx, int alloc_flags)
840 /* free_pages my go negative - that's OK */
841 long min = mark, free_pages = z->free_pages - (1 << order) + 1;
844 if (alloc_flags & ALLOC_HIGH)
846 if (alloc_flags & ALLOC_HARDER)
849 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
851 for (o = 0; o < order; o++) {
852 /* At the next order, this order's pages become unavailable */
853 free_pages -= z->free_area[o].nr_free << o;
855 /* Require fewer higher order pages to be free */
858 if (free_pages <= min)
865 * get_page_from_freeliest goes through the zonelist trying to allocate
869 get_page_from_freelist(gfp_t gfp_mask, unsigned int order,
870 struct zonelist *zonelist, int alloc_flags)
872 struct zone **z = zonelist->zones;
873 struct page *page = NULL;
874 int classzone_idx = zone_idx(*z);
877 * Go through the zonelist once, looking for a zone with enough free.
878 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
881 if ((alloc_flags & ALLOC_CPUSET) &&
882 !cpuset_zone_allowed(*z, gfp_mask))
885 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
887 if (alloc_flags & ALLOC_WMARK_MIN)
888 mark = (*z)->pages_min;
889 else if (alloc_flags & ALLOC_WMARK_LOW)
890 mark = (*z)->pages_low;
892 mark = (*z)->pages_high;
893 if (!zone_watermark_ok(*z, order, mark,
894 classzone_idx, alloc_flags))
895 if (!zone_reclaim_mode ||
896 !zone_reclaim(*z, gfp_mask, order))
900 page = buffered_rmqueue(zonelist, *z, order, gfp_mask);
904 } while (*(++z) != NULL);
909 * This is the 'heart' of the zoned buddy allocator.
911 struct page * fastcall
912 __alloc_pages(gfp_t gfp_mask, unsigned int order,
913 struct zonelist *zonelist)
915 const gfp_t wait = gfp_mask & __GFP_WAIT;
918 struct reclaim_state reclaim_state;
919 struct task_struct *p = current;
922 int did_some_progress;
924 might_sleep_if(wait);
927 z = zonelist->zones; /* the list of zones suitable for gfp_mask */
929 if (unlikely(*z == NULL)) {
930 /* Should this ever happen?? */
934 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, order,
935 zonelist, ALLOC_WMARK_LOW|ALLOC_CPUSET);
940 wakeup_kswapd(*z, order);
944 * OK, we're below the kswapd watermark and have kicked background
945 * reclaim. Now things get more complex, so set up alloc_flags according
946 * to how we want to proceed.
948 * The caller may dip into page reserves a bit more if the caller
949 * cannot run direct reclaim, or if the caller has realtime scheduling
950 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
951 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
953 alloc_flags = ALLOC_WMARK_MIN;
954 if ((unlikely(rt_task(p)) && !in_interrupt()) || !wait)
955 alloc_flags |= ALLOC_HARDER;
956 if (gfp_mask & __GFP_HIGH)
957 alloc_flags |= ALLOC_HIGH;
959 alloc_flags |= ALLOC_CPUSET;
962 * Go through the zonelist again. Let __GFP_HIGH and allocations
963 * coming from realtime tasks go deeper into reserves.
965 * This is the last chance, in general, before the goto nopage.
966 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
967 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
969 page = get_page_from_freelist(gfp_mask, order, zonelist, alloc_flags);
973 /* This allocation should allow future memory freeing. */
975 if (((p->flags & PF_MEMALLOC) || unlikely(test_thread_flag(TIF_MEMDIE)))
976 && !in_interrupt()) {
977 if (!(gfp_mask & __GFP_NOMEMALLOC)) {
979 /* go through the zonelist yet again, ignoring mins */
980 page = get_page_from_freelist(gfp_mask, order,
981 zonelist, ALLOC_NO_WATERMARKS);
984 if (gfp_mask & __GFP_NOFAIL) {
985 blk_congestion_wait(WRITE, HZ/50);
992 /* Atomic allocations - we can't balance anything */
999 /* We now go into synchronous reclaim */
1000 cpuset_memory_pressure_bump();
1001 p->flags |= PF_MEMALLOC;
1002 reclaim_state.reclaimed_slab = 0;
1003 p->reclaim_state = &reclaim_state;
1005 did_some_progress = try_to_free_pages(zonelist->zones, gfp_mask);
1007 p->reclaim_state = NULL;
1008 p->flags &= ~PF_MEMALLOC;
1012 if (likely(did_some_progress)) {
1013 page = get_page_from_freelist(gfp_mask, order,
1014 zonelist, alloc_flags);
1017 } else if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
1019 * Go through the zonelist yet one more time, keep
1020 * very high watermark here, this is only to catch
1021 * a parallel oom killing, we must fail if we're still
1022 * under heavy pressure.
1024 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, order,
1025 zonelist, ALLOC_WMARK_HIGH|ALLOC_CPUSET);
1029 out_of_memory(zonelist, gfp_mask, order);
1034 * Don't let big-order allocations loop unless the caller explicitly
1035 * requests that. Wait for some write requests to complete then retry.
1037 * In this implementation, __GFP_REPEAT means __GFP_NOFAIL for order
1038 * <= 3, but that may not be true in other implementations.
1041 if (!(gfp_mask & __GFP_NORETRY)) {
1042 if ((order <= 3) || (gfp_mask & __GFP_REPEAT))
1044 if (gfp_mask & __GFP_NOFAIL)
1048 blk_congestion_wait(WRITE, HZ/50);
1053 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
1054 printk(KERN_WARNING "%s: page allocation failure."
1055 " order:%d, mode:0x%x\n",
1056 p->comm, order, gfp_mask);
1064 EXPORT_SYMBOL(__alloc_pages);
1067 * Common helper functions.
1069 fastcall unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
1072 page = alloc_pages(gfp_mask, order);
1075 return (unsigned long) page_address(page);
1078 EXPORT_SYMBOL(__get_free_pages);
1080 fastcall unsigned long get_zeroed_page(gfp_t gfp_mask)
1085 * get_zeroed_page() returns a 32-bit address, which cannot represent
1088 BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
1090 page = alloc_pages(gfp_mask | __GFP_ZERO, 0);
1092 return (unsigned long) page_address(page);
1096 EXPORT_SYMBOL(get_zeroed_page);
1098 void __pagevec_free(struct pagevec *pvec)
1100 int i = pagevec_count(pvec);
1103 free_hot_cold_page(pvec->pages[i], pvec->cold);
1106 fastcall void __free_pages(struct page *page, unsigned int order)
1108 if (put_page_testzero(page)) {
1110 free_hot_page(page);
1112 __free_pages_ok(page, order);
1116 EXPORT_SYMBOL(__free_pages);
1118 fastcall void free_pages(unsigned long addr, unsigned int order)
1121 BUG_ON(!virt_addr_valid((void *)addr));
1122 __free_pages(virt_to_page((void *)addr), order);
1126 EXPORT_SYMBOL(free_pages);
1129 * Total amount of free (allocatable) RAM:
1131 unsigned int nr_free_pages(void)
1133 unsigned int sum = 0;
1137 sum += zone->free_pages;
1142 EXPORT_SYMBOL(nr_free_pages);
1145 unsigned int nr_free_pages_pgdat(pg_data_t *pgdat)
1147 unsigned int i, sum = 0;
1149 for (i = 0; i < MAX_NR_ZONES; i++)
1150 sum += pgdat->node_zones[i].free_pages;
1156 static unsigned int nr_free_zone_pages(int offset)
1158 /* Just pick one node, since fallback list is circular */
1159 pg_data_t *pgdat = NODE_DATA(numa_node_id());
1160 unsigned int sum = 0;
1162 struct zonelist *zonelist = pgdat->node_zonelists + offset;
1163 struct zone **zonep = zonelist->zones;
1166 for (zone = *zonep++; zone; zone = *zonep++) {
1167 unsigned long size = zone->present_pages;
1168 unsigned long high = zone->pages_high;
1177 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
1179 unsigned int nr_free_buffer_pages(void)
1181 return nr_free_zone_pages(gfp_zone(GFP_USER));
1185 * Amount of free RAM allocatable within all zones
1187 unsigned int nr_free_pagecache_pages(void)
1189 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER));
1192 #ifdef CONFIG_HIGHMEM
1193 unsigned int nr_free_highpages (void)
1196 unsigned int pages = 0;
1198 for_each_online_pgdat(pgdat)
1199 pages += pgdat->node_zones[ZONE_HIGHMEM].free_pages;
1206 static void show_node(struct zone *zone)
1208 printk("Node %d ", zone->zone_pgdat->node_id);
1211 #define show_node(zone) do { } while (0)
1214 void si_meminfo(struct sysinfo *val)
1216 val->totalram = totalram_pages;
1218 val->freeram = nr_free_pages();
1219 val->bufferram = nr_blockdev_pages();
1220 #ifdef CONFIG_HIGHMEM
1221 val->totalhigh = totalhigh_pages;
1222 val->freehigh = nr_free_highpages();
1227 val->mem_unit = PAGE_SIZE;
1228 if (vx_flags(VXF_VIRT_MEM, 0))
1229 vx_vsi_meminfo(val);
1232 EXPORT_SYMBOL(si_meminfo);
1235 void si_meminfo_node(struct sysinfo *val, int nid)
1237 pg_data_t *pgdat = NODE_DATA(nid);
1239 val->totalram = pgdat->node_present_pages;
1240 val->freeram = nr_free_pages_pgdat(pgdat);
1241 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
1242 val->freehigh = pgdat->node_zones[ZONE_HIGHMEM].free_pages;
1243 val->mem_unit = PAGE_SIZE;
1244 if (vx_flags(VXF_VIRT_MEM, 0))
1245 vx_vsi_meminfo(val);
1249 #define K(x) ((x) << (PAGE_SHIFT-10))
1252 * Show free area list (used inside shift_scroll-lock stuff)
1253 * We also calculate the percentage fragmentation. We do this by counting the
1254 * memory on each free list with the exception of the first item on the list.
1256 void show_free_areas(void)
1258 int cpu, temperature;
1259 unsigned long active;
1260 unsigned long inactive;
1264 for_each_zone(zone) {
1266 printk("%s per-cpu:", zone->name);
1268 if (!populated_zone(zone)) {
1274 for_each_online_cpu(cpu) {
1275 struct per_cpu_pageset *pageset;
1277 pageset = zone_pcp(zone, cpu);
1279 for (temperature = 0; temperature < 2; temperature++)
1280 printk("cpu %d %s: high %d, batch %d used:%d\n",
1282 temperature ? "cold" : "hot",
1283 pageset->pcp[temperature].high,
1284 pageset->pcp[temperature].batch,
1285 pageset->pcp[temperature].count);
1289 get_zone_counts(&active, &inactive, &free);
1291 printk("Free pages: %11ukB (%ukB HighMem)\n",
1293 K(nr_free_highpages()));
1295 printk("Active:%lu inactive:%lu dirty:%lu writeback:%lu "
1296 "unstable:%lu free:%u slab:%lu mapped:%lu pagetables:%lu\n",
1299 global_page_state(NR_FILE_DIRTY),
1300 global_page_state(NR_WRITEBACK),
1301 global_page_state(NR_UNSTABLE_NFS),
1303 global_page_state(NR_SLAB),
1304 global_page_state(NR_FILE_MAPPED),
1305 global_page_state(NR_PAGETABLE));
1307 for_each_zone(zone) {
1319 " pages_scanned:%lu"
1320 " all_unreclaimable? %s"
1323 K(zone->free_pages),
1326 K(zone->pages_high),
1328 K(zone->nr_inactive),
1329 K(zone->present_pages),
1330 zone->pages_scanned,
1331 (zone->all_unreclaimable ? "yes" : "no")
1333 printk("lowmem_reserve[]:");
1334 for (i = 0; i < MAX_NR_ZONES; i++)
1335 printk(" %lu", zone->lowmem_reserve[i]);
1339 for_each_zone(zone) {
1340 unsigned long nr[MAX_ORDER], flags, order, total = 0;
1343 printk("%s: ", zone->name);
1344 if (!populated_zone(zone)) {
1349 spin_lock_irqsave(&zone->lock, flags);
1350 for (order = 0; order < MAX_ORDER; order++) {
1351 nr[order] = zone->free_area[order].nr_free;
1352 total += nr[order] << order;
1354 spin_unlock_irqrestore(&zone->lock, flags);
1355 for (order = 0; order < MAX_ORDER; order++)
1356 printk("%lu*%lukB ", nr[order], K(1UL) << order);
1357 printk("= %lukB\n", K(total));
1360 show_swap_cache_info();
1364 * Builds allocation fallback zone lists.
1366 * Add all populated zones of a node to the zonelist.
1368 static int __meminit build_zonelists_node(pg_data_t *pgdat,
1369 struct zonelist *zonelist, int nr_zones, int zone_type)
1373 BUG_ON(zone_type > ZONE_HIGHMEM);
1376 zone = pgdat->node_zones + zone_type;
1377 if (populated_zone(zone)) {
1378 #ifndef CONFIG_HIGHMEM
1379 BUG_ON(zone_type > ZONE_NORMAL);
1381 zonelist->zones[nr_zones++] = zone;
1382 check_highest_zone(zone_type);
1386 } while (zone_type >= 0);
1390 static inline int highest_zone(int zone_bits)
1392 int res = ZONE_NORMAL;
1393 if (zone_bits & (__force int)__GFP_HIGHMEM)
1395 if (zone_bits & (__force int)__GFP_DMA32)
1397 if (zone_bits & (__force int)__GFP_DMA)
1403 #define MAX_NODE_LOAD (num_online_nodes())
1404 static int __meminitdata node_load[MAX_NUMNODES];
1406 * find_next_best_node - find the next node that should appear in a given node's fallback list
1407 * @node: node whose fallback list we're appending
1408 * @used_node_mask: nodemask_t of already used nodes
1410 * We use a number of factors to determine which is the next node that should
1411 * appear on a given node's fallback list. The node should not have appeared
1412 * already in @node's fallback list, and it should be the next closest node
1413 * according to the distance array (which contains arbitrary distance values
1414 * from each node to each node in the system), and should also prefer nodes
1415 * with no CPUs, since presumably they'll have very little allocation pressure
1416 * on them otherwise.
1417 * It returns -1 if no node is found.
1419 static int __meminit find_next_best_node(int node, nodemask_t *used_node_mask)
1422 int min_val = INT_MAX;
1425 /* Use the local node if we haven't already */
1426 if (!node_isset(node, *used_node_mask)) {
1427 node_set(node, *used_node_mask);
1431 for_each_online_node(n) {
1434 /* Don't want a node to appear more than once */
1435 if (node_isset(n, *used_node_mask))
1438 /* Use the distance array to find the distance */
1439 val = node_distance(node, n);
1441 /* Penalize nodes under us ("prefer the next node") */
1444 /* Give preference to headless and unused nodes */
1445 tmp = node_to_cpumask(n);
1446 if (!cpus_empty(tmp))
1447 val += PENALTY_FOR_NODE_WITH_CPUS;
1449 /* Slight preference for less loaded node */
1450 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
1451 val += node_load[n];
1453 if (val < min_val) {
1460 node_set(best_node, *used_node_mask);
1465 static void __meminit build_zonelists(pg_data_t *pgdat)
1467 int i, j, k, node, local_node;
1468 int prev_node, load;
1469 struct zonelist *zonelist;
1470 nodemask_t used_mask;
1472 /* initialize zonelists */
1473 for (i = 0; i < GFP_ZONETYPES; i++) {
1474 zonelist = pgdat->node_zonelists + i;
1475 zonelist->zones[0] = NULL;
1478 /* NUMA-aware ordering of nodes */
1479 local_node = pgdat->node_id;
1480 load = num_online_nodes();
1481 prev_node = local_node;
1482 nodes_clear(used_mask);
1483 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
1484 int distance = node_distance(local_node, node);
1487 * If another node is sufficiently far away then it is better
1488 * to reclaim pages in a zone before going off node.
1490 if (distance > RECLAIM_DISTANCE)
1491 zone_reclaim_mode = 1;
1494 * We don't want to pressure a particular node.
1495 * So adding penalty to the first node in same
1496 * distance group to make it round-robin.
1499 if (distance != node_distance(local_node, prev_node))
1500 node_load[node] += load;
1503 for (i = 0; i < GFP_ZONETYPES; i++) {
1504 zonelist = pgdat->node_zonelists + i;
1505 for (j = 0; zonelist->zones[j] != NULL; j++);
1507 k = highest_zone(i);
1509 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1510 zonelist->zones[j] = NULL;
1515 #else /* CONFIG_NUMA */
1517 static void __meminit build_zonelists(pg_data_t *pgdat)
1519 int i, j, k, node, local_node;
1521 local_node = pgdat->node_id;
1522 for (i = 0; i < GFP_ZONETYPES; i++) {
1523 struct zonelist *zonelist;
1525 zonelist = pgdat->node_zonelists + i;
1528 k = highest_zone(i);
1529 j = build_zonelists_node(pgdat, zonelist, j, k);
1531 * Now we build the zonelist so that it contains the zones
1532 * of all the other nodes.
1533 * We don't want to pressure a particular node, so when
1534 * building the zones for node N, we make sure that the
1535 * zones coming right after the local ones are those from
1536 * node N+1 (modulo N)
1538 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
1539 if (!node_online(node))
1541 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1543 for (node = 0; node < local_node; node++) {
1544 if (!node_online(node))
1546 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1549 zonelist->zones[j] = NULL;
1553 #endif /* CONFIG_NUMA */
1555 /* return values int ....just for stop_machine_run() */
1556 static int __meminit __build_all_zonelists(void *dummy)
1559 for_each_online_node(nid)
1560 build_zonelists(NODE_DATA(nid));
1564 void __meminit build_all_zonelists(void)
1566 if (system_state == SYSTEM_BOOTING) {
1567 __build_all_zonelists(0);
1568 cpuset_init_current_mems_allowed();
1570 /* we have to stop all cpus to guaranntee there is no user
1572 stop_machine_run(__build_all_zonelists, NULL, NR_CPUS);
1573 /* cpuset refresh routine should be here */
1575 vm_total_pages = nr_free_pagecache_pages();
1576 printk("Built %i zonelists. Total pages: %ld\n",
1577 num_online_nodes(), vm_total_pages);
1581 * Helper functions to size the waitqueue hash table.
1582 * Essentially these want to choose hash table sizes sufficiently
1583 * large so that collisions trying to wait on pages are rare.
1584 * But in fact, the number of active page waitqueues on typical
1585 * systems is ridiculously low, less than 200. So this is even
1586 * conservative, even though it seems large.
1588 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
1589 * waitqueues, i.e. the size of the waitq table given the number of pages.
1591 #define PAGES_PER_WAITQUEUE 256
1593 #ifndef CONFIG_MEMORY_HOTPLUG
1594 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
1596 unsigned long size = 1;
1598 pages /= PAGES_PER_WAITQUEUE;
1600 while (size < pages)
1604 * Once we have dozens or even hundreds of threads sleeping
1605 * on IO we've got bigger problems than wait queue collision.
1606 * Limit the size of the wait table to a reasonable size.
1608 size = min(size, 4096UL);
1610 return max(size, 4UL);
1614 * A zone's size might be changed by hot-add, so it is not possible to determine
1615 * a suitable size for its wait_table. So we use the maximum size now.
1617 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
1619 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
1620 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
1621 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
1623 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
1624 * or more by the traditional way. (See above). It equals:
1626 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
1627 * ia64(16K page size) : = ( 8G + 4M)byte.
1628 * powerpc (64K page size) : = (32G +16M)byte.
1630 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
1637 * This is an integer logarithm so that shifts can be used later
1638 * to extract the more random high bits from the multiplicative
1639 * hash function before the remainder is taken.
1641 static inline unsigned long wait_table_bits(unsigned long size)
1646 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
1648 static void __init calculate_zone_totalpages(struct pglist_data *pgdat,
1649 unsigned long *zones_size, unsigned long *zholes_size)
1651 unsigned long realtotalpages, totalpages = 0;
1654 for (i = 0; i < MAX_NR_ZONES; i++)
1655 totalpages += zones_size[i];
1656 pgdat->node_spanned_pages = totalpages;
1658 realtotalpages = totalpages;
1660 for (i = 0; i < MAX_NR_ZONES; i++)
1661 realtotalpages -= zholes_size[i];
1662 pgdat->node_present_pages = realtotalpages;
1663 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id, realtotalpages);
1668 * Initially all pages are reserved - free ones are freed
1669 * up by free_all_bootmem() once the early boot process is
1670 * done. Non-atomic initialization, single-pass.
1672 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
1673 unsigned long start_pfn)
1676 unsigned long end_pfn = start_pfn + size;
1679 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
1680 if (!early_pfn_valid(pfn))
1682 if (!early_pfn_in_nid(pfn, nid))
1684 page = pfn_to_page(pfn);
1685 set_page_links(page, zone, nid, pfn);
1686 init_page_count(page);
1687 reset_page_mapcount(page);
1688 SetPageReserved(page);
1689 INIT_LIST_HEAD(&page->lru);
1690 #ifdef WANT_PAGE_VIRTUAL
1691 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1692 if (!is_highmem_idx(zone))
1693 set_page_address(page, __va(pfn << PAGE_SHIFT));
1698 void zone_init_free_lists(struct pglist_data *pgdat, struct zone *zone,
1702 for (order = 0; order < MAX_ORDER ; order++) {
1703 INIT_LIST_HEAD(&zone->free_area[order].free_list);
1704 zone->free_area[order].nr_free = 0;
1708 #define ZONETABLE_INDEX(x, zone_nr) ((x << ZONES_SHIFT) | zone_nr)
1709 void zonetable_add(struct zone *zone, int nid, int zid, unsigned long pfn,
1712 unsigned long snum = pfn_to_section_nr(pfn);
1713 unsigned long end = pfn_to_section_nr(pfn + size);
1716 zone_table[ZONETABLE_INDEX(nid, zid)] = zone;
1718 for (; snum <= end; snum++)
1719 zone_table[ZONETABLE_INDEX(snum, zid)] = zone;
1722 #ifndef __HAVE_ARCH_MEMMAP_INIT
1723 #define memmap_init(size, nid, zone, start_pfn) \
1724 memmap_init_zone((size), (nid), (zone), (start_pfn))
1727 static int __cpuinit zone_batchsize(struct zone *zone)
1732 * The per-cpu-pages pools are set to around 1000th of the
1733 * size of the zone. But no more than 1/2 of a meg.
1735 * OK, so we don't know how big the cache is. So guess.
1737 batch = zone->present_pages / 1024;
1738 if (batch * PAGE_SIZE > 512 * 1024)
1739 batch = (512 * 1024) / PAGE_SIZE;
1740 batch /= 4; /* We effectively *= 4 below */
1745 * Clamp the batch to a 2^n - 1 value. Having a power
1746 * of 2 value was found to be more likely to have
1747 * suboptimal cache aliasing properties in some cases.
1749 * For example if 2 tasks are alternately allocating
1750 * batches of pages, one task can end up with a lot
1751 * of pages of one half of the possible page colors
1752 * and the other with pages of the other colors.
1754 batch = (1 << (fls(batch + batch/2)-1)) - 1;
1759 inline void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
1761 struct per_cpu_pages *pcp;
1763 memset(p, 0, sizeof(*p));
1765 pcp = &p->pcp[0]; /* hot */
1767 pcp->high = 6 * batch;
1768 pcp->batch = max(1UL, 1 * batch);
1769 INIT_LIST_HEAD(&pcp->list);
1771 pcp = &p->pcp[1]; /* cold*/
1773 pcp->high = 2 * batch;
1774 pcp->batch = max(1UL, batch/2);
1775 INIT_LIST_HEAD(&pcp->list);
1779 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
1780 * to the value high for the pageset p.
1783 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
1786 struct per_cpu_pages *pcp;
1788 pcp = &p->pcp[0]; /* hot list */
1790 pcp->batch = max(1UL, high/4);
1791 if ((high/4) > (PAGE_SHIFT * 8))
1792 pcp->batch = PAGE_SHIFT * 8;
1798 * Boot pageset table. One per cpu which is going to be used for all
1799 * zones and all nodes. The parameters will be set in such a way
1800 * that an item put on a list will immediately be handed over to
1801 * the buddy list. This is safe since pageset manipulation is done
1802 * with interrupts disabled.
1804 * Some NUMA counter updates may also be caught by the boot pagesets.
1806 * The boot_pagesets must be kept even after bootup is complete for
1807 * unused processors and/or zones. They do play a role for bootstrapping
1808 * hotplugged processors.
1810 * zoneinfo_show() and maybe other functions do
1811 * not check if the processor is online before following the pageset pointer.
1812 * Other parts of the kernel may not check if the zone is available.
1814 static struct per_cpu_pageset boot_pageset[NR_CPUS];
1817 * Dynamically allocate memory for the
1818 * per cpu pageset array in struct zone.
1820 static int __cpuinit process_zones(int cpu)
1822 struct zone *zone, *dzone;
1824 for_each_zone(zone) {
1826 zone_pcp(zone, cpu) = kmalloc_node(sizeof(struct per_cpu_pageset),
1827 GFP_KERNEL, cpu_to_node(cpu));
1828 if (!zone_pcp(zone, cpu))
1831 setup_pageset(zone_pcp(zone, cpu), zone_batchsize(zone));
1833 if (percpu_pagelist_fraction)
1834 setup_pagelist_highmark(zone_pcp(zone, cpu),
1835 (zone->present_pages / percpu_pagelist_fraction));
1840 for_each_zone(dzone) {
1843 kfree(zone_pcp(dzone, cpu));
1844 zone_pcp(dzone, cpu) = NULL;
1849 static inline void free_zone_pagesets(int cpu)
1853 for_each_zone(zone) {
1854 struct per_cpu_pageset *pset = zone_pcp(zone, cpu);
1856 /* Free per_cpu_pageset if it is slab allocated */
1857 if (pset != &boot_pageset[cpu])
1859 zone_pcp(zone, cpu) = NULL;
1863 static int __cpuinit pageset_cpuup_callback(struct notifier_block *nfb,
1864 unsigned long action,
1867 int cpu = (long)hcpu;
1868 int ret = NOTIFY_OK;
1871 case CPU_UP_PREPARE:
1872 if (process_zones(cpu))
1875 case CPU_UP_CANCELED:
1877 free_zone_pagesets(cpu);
1885 static struct notifier_block __cpuinitdata pageset_notifier =
1886 { &pageset_cpuup_callback, NULL, 0 };
1888 void __init setup_per_cpu_pageset(void)
1892 /* Initialize per_cpu_pageset for cpu 0.
1893 * A cpuup callback will do this for every cpu
1894 * as it comes online
1896 err = process_zones(smp_processor_id());
1898 register_cpu_notifier(&pageset_notifier);
1904 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
1907 struct pglist_data *pgdat = zone->zone_pgdat;
1911 * The per-page waitqueue mechanism uses hashed waitqueues
1914 zone->wait_table_hash_nr_entries =
1915 wait_table_hash_nr_entries(zone_size_pages);
1916 zone->wait_table_bits =
1917 wait_table_bits(zone->wait_table_hash_nr_entries);
1918 alloc_size = zone->wait_table_hash_nr_entries
1919 * sizeof(wait_queue_head_t);
1921 if (system_state == SYSTEM_BOOTING) {
1922 zone->wait_table = (wait_queue_head_t *)
1923 alloc_bootmem_node(pgdat, alloc_size);
1926 * This case means that a zone whose size was 0 gets new memory
1927 * via memory hot-add.
1928 * But it may be the case that a new node was hot-added. In
1929 * this case vmalloc() will not be able to use this new node's
1930 * memory - this wait_table must be initialized to use this new
1931 * node itself as well.
1932 * To use this new node's memory, further consideration will be
1935 zone->wait_table = (wait_queue_head_t *)vmalloc(alloc_size);
1937 if (!zone->wait_table)
1940 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
1941 init_waitqueue_head(zone->wait_table + i);
1946 static __meminit void zone_pcp_init(struct zone *zone)
1949 unsigned long batch = zone_batchsize(zone);
1951 for (cpu = 0; cpu < NR_CPUS; cpu++) {
1953 /* Early boot. Slab allocator not functional yet */
1954 zone_pcp(zone, cpu) = &boot_pageset[cpu];
1955 setup_pageset(&boot_pageset[cpu],0);
1957 setup_pageset(zone_pcp(zone,cpu), batch);
1960 if (zone->present_pages)
1961 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%lu\n",
1962 zone->name, zone->present_pages, batch);
1965 __meminit int init_currently_empty_zone(struct zone *zone,
1966 unsigned long zone_start_pfn,
1969 struct pglist_data *pgdat = zone->zone_pgdat;
1971 ret = zone_wait_table_init(zone, size);
1974 pgdat->nr_zones = zone_idx(zone) + 1;
1976 zone->zone_start_pfn = zone_start_pfn;
1978 memmap_init(size, pgdat->node_id, zone_idx(zone), zone_start_pfn);
1980 zone_init_free_lists(pgdat, zone, zone->spanned_pages);
1986 * Set up the zone data structures:
1987 * - mark all pages reserved
1988 * - mark all memory queues empty
1989 * - clear the memory bitmaps
1991 static void __meminit free_area_init_core(struct pglist_data *pgdat,
1992 unsigned long *zones_size, unsigned long *zholes_size)
1995 int nid = pgdat->node_id;
1996 unsigned long zone_start_pfn = pgdat->node_start_pfn;
1999 pgdat_resize_init(pgdat);
2000 pgdat->nr_zones = 0;
2001 init_waitqueue_head(&pgdat->kswapd_wait);
2002 pgdat->kswapd_max_order = 0;
2004 for (j = 0; j < MAX_NR_ZONES; j++) {
2005 struct zone *zone = pgdat->node_zones + j;
2006 unsigned long size, realsize;
2008 realsize = size = zones_size[j];
2010 realsize -= zholes_size[j];
2012 if (j < ZONE_HIGHMEM)
2013 nr_kernel_pages += realsize;
2014 nr_all_pages += realsize;
2016 zone->spanned_pages = size;
2017 zone->present_pages = realsize;
2019 zone->min_unmapped_ratio = (realsize*sysctl_min_unmapped_ratio)
2021 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
2023 zone->name = zone_names[j];
2024 spin_lock_init(&zone->lock);
2025 spin_lock_init(&zone->lru_lock);
2026 zone_seqlock_init(zone);
2027 zone->zone_pgdat = pgdat;
2028 zone->free_pages = 0;
2030 zone->prev_priority = DEF_PRIORITY;
2032 zone_pcp_init(zone);
2033 INIT_LIST_HEAD(&zone->active_list);
2034 INIT_LIST_HEAD(&zone->inactive_list);
2035 zone->nr_scan_active = 0;
2036 zone->nr_scan_inactive = 0;
2037 zone->nr_active = 0;
2038 zone->nr_inactive = 0;
2039 zap_zone_vm_stats(zone);
2040 atomic_set(&zone->reclaim_in_progress, 0);
2044 zonetable_add(zone, nid, j, zone_start_pfn, size);
2045 ret = init_currently_empty_zone(zone, zone_start_pfn, size);
2047 zone_start_pfn += size;
2051 static void __init alloc_node_mem_map(struct pglist_data *pgdat)
2053 /* Skip empty nodes */
2054 if (!pgdat->node_spanned_pages)
2057 #ifdef CONFIG_FLAT_NODE_MEM_MAP
2058 /* ia64 gets its own node_mem_map, before this, without bootmem */
2059 if (!pgdat->node_mem_map) {
2060 unsigned long size, start, end;
2064 * The zone's endpoints aren't required to be MAX_ORDER
2065 * aligned but the node_mem_map endpoints must be in order
2066 * for the buddy allocator to function correctly.
2068 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
2069 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
2070 end = ALIGN(end, MAX_ORDER_NR_PAGES);
2071 size = (end - start) * sizeof(struct page);
2072 map = alloc_remap(pgdat->node_id, size);
2074 map = alloc_bootmem_node(pgdat, size);
2075 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
2077 #ifdef CONFIG_FLATMEM
2079 * With no DISCONTIG, the global mem_map is just set as node 0's
2081 if (pgdat == NODE_DATA(0))
2082 mem_map = NODE_DATA(0)->node_mem_map;
2084 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
2087 void __meminit free_area_init_node(int nid, struct pglist_data *pgdat,
2088 unsigned long *zones_size, unsigned long node_start_pfn,
2089 unsigned long *zholes_size)
2091 pgdat->node_id = nid;
2092 pgdat->node_start_pfn = node_start_pfn;
2093 calculate_zone_totalpages(pgdat, zones_size, zholes_size);
2095 alloc_node_mem_map(pgdat);
2097 free_area_init_core(pgdat, zones_size, zholes_size);
2100 #ifndef CONFIG_NEED_MULTIPLE_NODES
2101 static bootmem_data_t contig_bootmem_data;
2102 struct pglist_data contig_page_data = { .bdata = &contig_bootmem_data };
2104 EXPORT_SYMBOL(contig_page_data);
2107 void __init free_area_init(unsigned long *zones_size)
2109 free_area_init_node(0, NODE_DATA(0), zones_size,
2110 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
2113 #ifdef CONFIG_HOTPLUG_CPU
2114 static int page_alloc_cpu_notify(struct notifier_block *self,
2115 unsigned long action, void *hcpu)
2117 int cpu = (unsigned long)hcpu;
2119 if (action == CPU_DEAD) {
2120 local_irq_disable();
2122 vm_events_fold_cpu(cpu);
2124 refresh_cpu_vm_stats(cpu);
2128 #endif /* CONFIG_HOTPLUG_CPU */
2130 void __init page_alloc_init(void)
2132 hotcpu_notifier(page_alloc_cpu_notify, 0);
2136 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
2137 * or min_free_kbytes changes.
2139 static void calculate_totalreserve_pages(void)
2141 struct pglist_data *pgdat;
2142 unsigned long reserve_pages = 0;
2145 for_each_online_pgdat(pgdat) {
2146 for (i = 0; i < MAX_NR_ZONES; i++) {
2147 struct zone *zone = pgdat->node_zones + i;
2148 unsigned long max = 0;
2150 /* Find valid and maximum lowmem_reserve in the zone */
2151 for (j = i; j < MAX_NR_ZONES; j++) {
2152 if (zone->lowmem_reserve[j] > max)
2153 max = zone->lowmem_reserve[j];
2156 /* we treat pages_high as reserved pages. */
2157 max += zone->pages_high;
2159 if (max > zone->present_pages)
2160 max = zone->present_pages;
2161 reserve_pages += max;
2164 totalreserve_pages = reserve_pages;
2168 * setup_per_zone_lowmem_reserve - called whenever
2169 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
2170 * has a correct pages reserved value, so an adequate number of
2171 * pages are left in the zone after a successful __alloc_pages().
2173 static void setup_per_zone_lowmem_reserve(void)
2175 struct pglist_data *pgdat;
2178 for_each_online_pgdat(pgdat) {
2179 for (j = 0; j < MAX_NR_ZONES; j++) {
2180 struct zone *zone = pgdat->node_zones + j;
2181 unsigned long present_pages = zone->present_pages;
2183 zone->lowmem_reserve[j] = 0;
2185 for (idx = j-1; idx >= 0; idx--) {
2186 struct zone *lower_zone;
2188 if (sysctl_lowmem_reserve_ratio[idx] < 1)
2189 sysctl_lowmem_reserve_ratio[idx] = 1;
2191 lower_zone = pgdat->node_zones + idx;
2192 lower_zone->lowmem_reserve[j] = present_pages /
2193 sysctl_lowmem_reserve_ratio[idx];
2194 present_pages += lower_zone->present_pages;
2199 /* update totalreserve_pages */
2200 calculate_totalreserve_pages();
2204 * setup_per_zone_pages_min - called when min_free_kbytes changes. Ensures
2205 * that the pages_{min,low,high} values for each zone are set correctly
2206 * with respect to min_free_kbytes.
2208 void setup_per_zone_pages_min(void)
2210 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
2211 unsigned long lowmem_pages = 0;
2213 unsigned long flags;
2215 /* Calculate total number of !ZONE_HIGHMEM pages */
2216 for_each_zone(zone) {
2217 if (!is_highmem(zone))
2218 lowmem_pages += zone->present_pages;
2221 for_each_zone(zone) {
2224 spin_lock_irqsave(&zone->lru_lock, flags);
2225 tmp = (u64)pages_min * zone->present_pages;
2226 do_div(tmp, lowmem_pages);
2227 if (is_highmem(zone)) {
2229 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
2230 * need highmem pages, so cap pages_min to a small
2233 * The (pages_high-pages_low) and (pages_low-pages_min)
2234 * deltas controls asynch page reclaim, and so should
2235 * not be capped for highmem.
2239 min_pages = zone->present_pages / 1024;
2240 if (min_pages < SWAP_CLUSTER_MAX)
2241 min_pages = SWAP_CLUSTER_MAX;
2242 if (min_pages > 128)
2244 zone->pages_min = min_pages;
2247 * If it's a lowmem zone, reserve a number of pages
2248 * proportionate to the zone's size.
2250 zone->pages_min = tmp;
2253 zone->pages_low = zone->pages_min + (tmp >> 2);
2254 zone->pages_high = zone->pages_min + (tmp >> 1);
2255 spin_unlock_irqrestore(&zone->lru_lock, flags);
2258 /* update totalreserve_pages */
2259 calculate_totalreserve_pages();
2263 * Initialise min_free_kbytes.
2265 * For small machines we want it small (128k min). For large machines
2266 * we want it large (64MB max). But it is not linear, because network
2267 * bandwidth does not increase linearly with machine size. We use
2269 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
2270 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
2286 static int __init init_per_zone_pages_min(void)
2288 unsigned long lowmem_kbytes;
2290 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
2292 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
2293 if (min_free_kbytes < 128)
2294 min_free_kbytes = 128;
2295 if (min_free_kbytes > 65536)
2296 min_free_kbytes = 65536;
2297 setup_per_zone_pages_min();
2298 setup_per_zone_lowmem_reserve();
2301 module_init(init_per_zone_pages_min)
2304 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
2305 * that we can call two helper functions whenever min_free_kbytes
2308 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
2309 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2311 proc_dointvec(table, write, file, buffer, length, ppos);
2312 setup_per_zone_pages_min();
2317 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
2318 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2323 rc = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
2328 zone->min_unmapped_ratio = (zone->present_pages *
2329 sysctl_min_unmapped_ratio) / 100;
2333 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
2334 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2339 rc = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
2344 zone->min_slab_pages = (zone->present_pages *
2345 sysctl_min_slab_ratio) / 100;
2351 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
2352 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
2353 * whenever sysctl_lowmem_reserve_ratio changes.
2355 * The reserve ratio obviously has absolutely no relation with the
2356 * pages_min watermarks. The lowmem reserve ratio can only make sense
2357 * if in function of the boot time zone sizes.
2359 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
2360 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2362 proc_dointvec_minmax(table, write, file, buffer, length, ppos);
2363 setup_per_zone_lowmem_reserve();
2368 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
2369 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
2370 * can have before it gets flushed back to buddy allocator.
2373 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
2374 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2380 ret = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
2381 if (!write || (ret == -EINVAL))
2383 for_each_zone(zone) {
2384 for_each_online_cpu(cpu) {
2386 high = zone->present_pages / percpu_pagelist_fraction;
2387 setup_pagelist_highmark(zone_pcp(zone, cpu), high);
2393 __initdata int hashdist = HASHDIST_DEFAULT;
2396 static int __init set_hashdist(char *str)
2400 hashdist = simple_strtoul(str, &str, 0);
2403 __setup("hashdist=", set_hashdist);
2407 * allocate a large system hash table from bootmem
2408 * - it is assumed that the hash table must contain an exact power-of-2
2409 * quantity of entries
2410 * - limit is the number of hash buckets, not the total allocation size
2412 void *__init alloc_large_system_hash(const char *tablename,
2413 unsigned long bucketsize,
2414 unsigned long numentries,
2417 unsigned int *_hash_shift,
2418 unsigned int *_hash_mask,
2419 unsigned long limit)
2421 unsigned long long max = limit;
2422 unsigned long log2qty, size;
2425 /* allow the kernel cmdline to have a say */
2427 /* round applicable memory size up to nearest megabyte */
2428 numentries = (flags & HASH_HIGHMEM) ? nr_all_pages : nr_kernel_pages;
2429 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
2430 numentries >>= 20 - PAGE_SHIFT;
2431 numentries <<= 20 - PAGE_SHIFT;
2433 /* limit to 1 bucket per 2^scale bytes of low memory */
2434 if (scale > PAGE_SHIFT)
2435 numentries >>= (scale - PAGE_SHIFT);
2437 numentries <<= (PAGE_SHIFT - scale);
2439 numentries = roundup_pow_of_two(numentries);
2441 /* limit allocation size to 1/16 total memory by default */
2443 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
2444 do_div(max, bucketsize);
2447 if (numentries > max)
2450 log2qty = long_log2(numentries);
2453 size = bucketsize << log2qty;
2454 if (flags & HASH_EARLY)
2455 table = alloc_bootmem(size);
2457 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
2459 unsigned long order;
2460 for (order = 0; ((1UL << order) << PAGE_SHIFT) < size; order++)
2462 table = (void*) __get_free_pages(GFP_ATOMIC, order);
2464 } while (!table && size > PAGE_SIZE && --log2qty);
2467 panic("Failed to allocate %s hash table\n", tablename);
2469 printk("%s hash table entries: %d (order: %d, %lu bytes)\n",
2472 long_log2(size) - PAGE_SHIFT,
2476 *_hash_shift = log2qty;
2478 *_hash_mask = (1 << log2qty) - 1;
2483 #ifdef CONFIG_OUT_OF_LINE_PFN_TO_PAGE
2484 struct page *pfn_to_page(unsigned long pfn)
2486 return __pfn_to_page(pfn);
2488 unsigned long page_to_pfn(struct page *page)
2490 return __page_to_pfn(page);
2492 EXPORT_SYMBOL(pfn_to_page);
2493 EXPORT_SYMBOL(page_to_pfn);
2494 #endif /* CONFIG_OUT_OF_LINE_PFN_TO_PAGE */