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/config.h>
18 #include <linux/stddef.h>
20 #include <linux/swap.h>
21 #include <linux/interrupt.h>
22 #include <linux/pagemap.h>
23 #include <linux/bootmem.h>
24 #include <linux/compiler.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>
35 #include <asm/tlbflush.h>
37 DECLARE_BITMAP(node_online_map, MAX_NUMNODES);
38 struct pglist_data *pgdat_list;
39 unsigned long totalram_pages;
40 unsigned long totalhigh_pages;
43 int sysctl_lower_zone_protection = 0;
45 EXPORT_SYMBOL(totalram_pages);
46 EXPORT_SYMBOL(nr_swap_pages);
49 * Used by page_zone() to look up the address of the struct zone whose
50 * id is encoded in the upper bits of page->flags
52 struct zone *zone_table[1 << (ZONES_SHIFT + NODES_SHIFT)];
53 EXPORT_SYMBOL(zone_table);
55 static char *zone_names[MAX_NR_ZONES] = { "DMA", "Normal", "HighMem" };
56 int min_free_kbytes = 1024;
58 static unsigned long __initdata nr_kernel_pages;
59 static unsigned long __initdata nr_all_pages;
62 * Temporary debugging check for pages not lying within a given zone.
64 static int bad_range(struct zone *zone, struct page *page)
66 if (page_to_pfn(page) >= zone->zone_start_pfn + zone->spanned_pages)
68 if (page_to_pfn(page) < zone->zone_start_pfn)
70 if (zone != page_zone(page))
75 static void bad_page(const char *function, struct page *page)
77 printk(KERN_EMERG "Bad page state at %s (in process '%s', page %p)\n",
78 function, current->comm, page);
79 printk(KERN_EMERG "flags:0x%08lx mapping:%p mapcount:%d count:%d\n",
80 (unsigned long)page->flags, page->mapping,
81 (int)page->mapcount, page_count(page));
82 printk(KERN_EMERG "Backtrace:\n");
84 printk(KERN_EMERG "Trying to fix it up, but a reboot is needed\n");
85 page->flags &= ~(1 << PG_private |
94 set_page_count(page, 0);
99 #ifndef CONFIG_HUGETLB_PAGE
100 #define prep_compound_page(page, order) do { } while (0)
101 #define destroy_compound_page(page, order) do { } while (0)
104 * Higher-order pages are called "compound pages". They are structured thusly:
106 * The first PAGE_SIZE page is called the "head page".
108 * The remaining PAGE_SIZE pages are called "tail pages".
110 * All pages have PG_compound set. All pages have their ->private pointing at
111 * the head page (even the head page has this).
113 * The first tail page's ->mapping, if non-zero, holds the address of the
114 * compound page's put_page() function.
116 * The order of the allocation is stored in the first tail page's ->index
117 * This is only for debug at present. This usage means that zero-order pages
118 * may not be compound.
120 static void prep_compound_page(struct page *page, unsigned long order)
123 int nr_pages = 1 << order;
125 page[1].mapping = NULL;
126 page[1].index = order;
127 for (i = 0; i < nr_pages; i++) {
128 struct page *p = page + i;
131 p->private = (unsigned long)page;
135 static void destroy_compound_page(struct page *page, unsigned long order)
138 int nr_pages = 1 << order;
140 if (!PageCompound(page))
143 if (page[1].index != order)
144 bad_page(__FUNCTION__, page);
146 for (i = 0; i < nr_pages; i++) {
147 struct page *p = page + i;
149 if (!PageCompound(p))
150 bad_page(__FUNCTION__, page);
151 if (p->private != (unsigned long)page)
152 bad_page(__FUNCTION__, page);
153 ClearPageCompound(p);
156 #endif /* CONFIG_HUGETLB_PAGE */
159 * Freeing function for a buddy system allocator.
161 * The concept of a buddy system is to maintain direct-mapped table
162 * (containing bit values) for memory blocks of various "orders".
163 * The bottom level table contains the map for the smallest allocatable
164 * units of memory (here, pages), and each level above it describes
165 * pairs of units from the levels below, hence, "buddies".
166 * At a high level, all that happens here is marking the table entry
167 * at the bottom level available, and propagating the changes upward
168 * as necessary, plus some accounting needed to play nicely with other
169 * parts of the VM system.
170 * At each level, we keep one bit for each pair of blocks, which
171 * is set to 1 iff only one of the pair is allocated. So when we
172 * are allocating or freeing one, we can derive the state of the
173 * other. That is, if we allocate a small block, and both were
174 * free, the remainder of the region must be split into blocks.
175 * If a block is freed, and its buddy is also free, then this
176 * triggers coalescing into a block of larger size.
181 static inline void __free_pages_bulk (struct page *page, struct page *base,
182 struct zone *zone, struct free_area *area, unsigned int order)
184 unsigned long page_idx, index, mask;
187 destroy_compound_page(page, order);
188 mask = (~0UL) << order;
189 page_idx = page - base;
190 if (page_idx & ~mask)
192 index = page_idx >> (1 + order);
194 zone->free_pages += 1 << order;
195 while (order < MAX_ORDER-1) {
196 struct page *buddy1, *buddy2;
198 BUG_ON(area >= zone->free_area + MAX_ORDER);
199 if (!__test_and_change_bit(index, area->map))
201 * the buddy page is still allocated.
205 /* Move the buddy up one level. */
206 buddy1 = base + (page_idx ^ (1 << order));
207 buddy2 = base + page_idx;
208 BUG_ON(bad_range(zone, buddy1));
209 BUG_ON(bad_range(zone, buddy2));
210 list_del(&buddy1->lru);
217 list_add(&(base + page_idx)->lru, &area->free_list);
220 static inline void free_pages_check(const char *function, struct page *page)
222 if ( page_mapped(page) ||
223 page->mapping != NULL ||
224 page_count(page) != 0 ||
235 1 << PG_writeback )))
236 bad_page(function, page);
238 ClearPageDirty(page);
242 * Frees a list of pages.
243 * Assumes all pages on list are in same zone, and of same order.
244 * count is the number of pages to free, or 0 for all on the list.
246 * If the zone was previously in an "all pages pinned" state then look to
247 * see if this freeing clears that state.
249 * And clear the zone's pages_scanned counter, to hold off the "all pages are
250 * pinned" detection logic.
253 free_pages_bulk(struct zone *zone, int count,
254 struct list_head *list, unsigned int order)
257 struct free_area *area;
258 struct page *base, *page = NULL;
261 base = zone->zone_mem_map;
262 area = zone->free_area + order;
263 spin_lock_irqsave(&zone->lock, flags);
264 zone->all_unreclaimable = 0;
265 zone->pages_scanned = 0;
266 while (!list_empty(list) && count--) {
267 page = list_entry(list->prev, struct page, lru);
268 /* have to delete it as __free_pages_bulk list manipulates */
269 list_del(&page->lru);
270 __free_pages_bulk(page, base, zone, area, order);
273 spin_unlock_irqrestore(&zone->lock, flags);
277 void __free_pages_ok(struct page *page, unsigned int order)
282 mod_page_state(pgfree, 1 << order);
283 for (i = 0 ; i < (1 << order) ; ++i)
284 free_pages_check(__FUNCTION__, page + i);
285 list_add(&page->lru, &list);
286 kernel_map_pages(page, 1<<order, 0);
287 free_pages_bulk(page_zone(page), 1, &list, order);
290 #define MARK_USED(index, order, area) \
291 __change_bit((index) >> (1+(order)), (area)->map)
294 * The order of subdivision here is critical for the IO subsystem.
295 * Please do not alter this order without good reasons and regression
296 * testing. Specifically, as large blocks of memory are subdivided,
297 * the order in which smaller blocks are delivered depends on the order
298 * they're subdivided in this function. This is the primary factor
299 * influencing the order in which pages are delivered to the IO
300 * subsystem according to empirical testing, and this is also justified
301 * by considering the behavior of a buddy system containing a single
302 * large block of memory acted on by a series of small allocations.
303 * This behavior is a critical factor in sglist merging's success.
307 static inline struct page *
308 expand(struct zone *zone, struct page *page,
309 unsigned long index, int low, int high, struct free_area *area)
311 unsigned long size = 1 << high;
317 BUG_ON(bad_range(zone, &page[size]));
318 list_add(&page[size].lru, &area->free_list);
319 MARK_USED(index + size, high, area);
324 static inline void set_page_refs(struct page *page, int order)
327 set_page_count(page, 1);
332 * We need to reference all the pages for this order, otherwise if
333 * anyone accesses one of the pages with (get/put) it will be freed.
335 for (i = 0; i < (1 << order); i++)
336 set_page_count(page+i, 1);
337 #endif /* CONFIG_MMU */
341 * This page is about to be returned from the page allocator
343 static void prep_new_page(struct page *page, int order)
345 if (page->mapping || page_mapped(page) ||
356 1 << PG_writeback )))
357 bad_page(__FUNCTION__, page);
359 page->flags &= ~(1 << PG_uptodate | 1 << PG_error |
360 1 << PG_referenced | 1 << PG_arch_1 |
361 1 << PG_checked | 1 << PG_mappedtodisk);
363 set_page_refs(page, order);
367 * Do the hard work of removing an element from the buddy allocator.
368 * Call me with the zone->lock already held.
370 static struct page *__rmqueue(struct zone *zone, unsigned int order)
372 struct free_area * area;
373 unsigned int current_order;
377 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
378 area = zone->free_area + current_order;
379 if (list_empty(&area->free_list))
382 page = list_entry(area->free_list.next, struct page, lru);
383 list_del(&page->lru);
384 index = page - zone->zone_mem_map;
385 if (current_order != MAX_ORDER-1)
386 MARK_USED(index, current_order, area);
387 zone->free_pages -= 1UL << order;
388 return expand(zone, page, index, order, current_order, area);
395 * Obtain a specified number of elements from the buddy allocator, all under
396 * a single hold of the lock, for efficiency. Add them to the supplied list.
397 * Returns the number of new pages which were placed at *list.
399 static int rmqueue_bulk(struct zone *zone, unsigned int order,
400 unsigned long count, struct list_head *list)
407 spin_lock_irqsave(&zone->lock, flags);
408 for (i = 0; i < count; ++i) {
409 page = __rmqueue(zone, order);
413 list_add_tail(&page->lru, list);
415 spin_unlock_irqrestore(&zone->lock, flags);
419 #if defined(CONFIG_PM) || defined(CONFIG_HOTPLUG_CPU)
420 static void __drain_pages(unsigned int cpu)
425 for_each_zone(zone) {
426 struct per_cpu_pageset *pset;
428 pset = &zone->pageset[cpu];
429 for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
430 struct per_cpu_pages *pcp;
433 pcp->count -= free_pages_bulk(zone, pcp->count,
438 #endif /* CONFIG_PM || CONFIG_HOTPLUG_CPU */
441 int is_head_of_free_region(struct page *page)
443 struct zone *zone = page_zone(page);
446 struct list_head *curr;
449 * Should not matter as we need quiescent system for
450 * suspend anyway, but...
452 spin_lock_irqsave(&zone->lock, flags);
453 for (order = MAX_ORDER - 1; order >= 0; --order)
454 list_for_each(curr, &zone->free_area[order].free_list)
455 if (page == list_entry(curr, struct page, lru)) {
456 spin_unlock_irqrestore(&zone->lock, flags);
459 spin_unlock_irqrestore(&zone->lock, flags);
464 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
466 void drain_local_pages(void)
470 local_irq_save(flags);
471 __drain_pages(smp_processor_id());
472 local_irq_restore(flags);
474 #endif /* CONFIG_PM */
476 static void zone_statistics(struct zonelist *zonelist, struct zone *z)
481 pg_data_t *pg = z->zone_pgdat;
482 pg_data_t *orig = zonelist->zones[0]->zone_pgdat;
483 struct per_cpu_pageset *p;
485 local_irq_save(flags);
486 cpu = smp_processor_id();
487 p = &z->pageset[cpu];
489 z->pageset[cpu].numa_hit++;
492 zonelist->zones[0]->pageset[cpu].numa_foreign++;
494 if (pg == NODE_DATA(numa_node_id()))
498 local_irq_restore(flags);
503 * Free a 0-order page
505 static void FASTCALL(free_hot_cold_page(struct page *page, int cold));
506 static void fastcall free_hot_cold_page(struct page *page, int cold)
508 struct zone *zone = page_zone(page);
509 struct per_cpu_pages *pcp;
512 kernel_map_pages(page, 1, 0);
513 inc_page_state(pgfree);
514 free_pages_check(__FUNCTION__, page);
515 pcp = &zone->pageset[get_cpu()].pcp[cold];
516 local_irq_save(flags);
517 if (pcp->count >= pcp->high)
518 pcp->count -= free_pages_bulk(zone, pcp->batch, &pcp->list, 0);
519 list_add(&page->lru, &pcp->list);
521 local_irq_restore(flags);
525 void fastcall free_hot_page(struct page *page)
527 free_hot_cold_page(page, 0);
530 void fastcall free_cold_page(struct page *page)
532 free_hot_cold_page(page, 1);
536 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
537 * we cheat by calling it from here, in the order > 0 path. Saves a branch
542 buffered_rmqueue(struct zone *zone, int order, int gfp_flags)
545 struct page *page = NULL;
546 int cold = !!(gfp_flags & __GFP_COLD);
549 struct per_cpu_pages *pcp;
551 pcp = &zone->pageset[get_cpu()].pcp[cold];
552 local_irq_save(flags);
553 if (pcp->count <= pcp->low)
554 pcp->count += rmqueue_bulk(zone, 0,
555 pcp->batch, &pcp->list);
557 page = list_entry(pcp->list.next, struct page, lru);
558 list_del(&page->lru);
561 local_irq_restore(flags);
566 spin_lock_irqsave(&zone->lock, flags);
567 page = __rmqueue(zone, order);
568 spin_unlock_irqrestore(&zone->lock, flags);
572 BUG_ON(bad_range(zone, page));
573 mod_page_state_zone(zone, pgalloc, 1 << order);
574 prep_new_page(page, order);
575 if (order && (gfp_flags & __GFP_COMP))
576 prep_compound_page(page, order);
582 * This is the 'heart' of the zoned buddy allocator.
584 * Herein lies the mysterious "incremental min". That's the
586 * local_low = z->pages_low;
589 * thing. The intent here is to provide additional protection to low zones for
590 * allocation requests which _could_ use higher zones. So a GFP_HIGHMEM
591 * request is not allowed to dip as deeply into the normal zone as a GFP_KERNEL
592 * request. This preserves additional space in those lower zones for requests
593 * which really do need memory from those zones. It means that on a decent
594 * sized machine, GFP_HIGHMEM and GFP_KERNEL requests basically leave the DMA
597 struct page * fastcall
598 __alloc_pages(unsigned int gfp_mask, unsigned int order,
599 struct zonelist *zonelist)
601 const int wait = gfp_mask & __GFP_WAIT;
605 struct reclaim_state reclaim_state;
606 struct task_struct *p = current;
611 #if defined(CONFIG_MIPS_BRCM)
616 might_sleep_if(wait);
618 zones = zonelist->zones; /* the list of zones suitable for gfp_mask */
619 if (zones[0] == NULL) /* no zones in the zonelist */
622 alloc_type = zone_idx(zones[0]);
624 /* Go through the zonelist once, looking for a zone with enough free */
625 for (i = 0; zones[i] != NULL; i++) {
626 struct zone *z = zones[i];
628 min = (1<<order) + z->protection[alloc_type];
631 * We let real-time tasks dip their real-time paws a little
632 * deeper into reserves.
635 min -= z->pages_low >> 1;
637 if (z->free_pages >= min ||
638 (!wait && z->free_pages >= z->pages_high)) {
639 page = buffered_rmqueue(z, order, gfp_mask);
641 zone_statistics(zonelist, z);
647 /* we're somewhat low on memory, failed to find what we needed */
648 for (i = 0; zones[i] != NULL; i++)
649 wakeup_kswapd(zones[i]);
651 /* Go through the zonelist again, taking __GFP_HIGH into account */
652 for (i = 0; zones[i] != NULL; i++) {
653 struct zone *z = zones[i];
655 min = (1<<order) + z->protection[alloc_type];
657 if (gfp_mask & __GFP_HIGH)
658 min -= z->pages_low >> 2;
660 min -= z->pages_low >> 1;
662 if (z->free_pages >= min ||
663 (!wait && z->free_pages >= z->pages_high)) {
664 page = buffered_rmqueue(z, order, gfp_mask);
666 zone_statistics(zonelist, z);
670 #if defined(CONFIG_MIPS_BRCM)
671 tpages += z->nr_active + z->nr_inactive;
675 /* here we're in the low on memory slow path */
677 #if defined(CONFIG_MIPS_BRCM)
678 nretry = tpages/SWAP_CLUSTER_MAX;
679 //printk("============to rebalance, tpages %d nretry %d ", tpages, nretry);
683 if ((p->flags & (PF_MEMALLOC | PF_MEMDIE)) && !in_interrupt()) {
684 /* go through the zonelist yet again, ignoring mins */
685 for (i = 0; zones[i] != NULL; i++) {
686 struct zone *z = zones[i];
688 page = buffered_rmqueue(z, order, gfp_mask);
690 zone_statistics(zonelist, z);
697 /* Atomic allocations - we can't balance anything */
701 p->flags |= PF_MEMALLOC;
702 reclaim_state.reclaimed_slab = 0;
703 p->reclaim_state = &reclaim_state;
705 try_to_free_pages(zones, gfp_mask, order);
707 p->reclaim_state = NULL;
708 p->flags &= ~PF_MEMALLOC;
710 /* go through the zonelist yet one more time */
711 for (i = 0; zones[i] != NULL; i++) {
712 struct zone *z = zones[i];
714 min = (1UL << order) + z->protection[alloc_type];
716 if (z->free_pages >= min ||
717 (!wait && z->free_pages >= z->pages_high)) {
718 page = buffered_rmqueue(z, order, gfp_mask);
720 zone_statistics(zonelist, z);
725 #if defined(CONFIG_MIPS_BRCM)
727 * For embedded system, we have small memory and we assume well-behaved software
728 * Let it scan all pages if necessary
729 * This is similar to including __GFP_NOFAIL/__GFP_REPEAT by default, except we use nretry as a safety mechanism
732 if (!(gfp_mask & __GFP_NORETRY)) {
733 if ((gfp_mask & __GFP_NOFAIL) || (gfp_mask & __GFP_REPEAT)) {
736 else if (nretry > 0) {
744 * Don't let big-order allocations loop unless the caller explicitly
745 * requests that. Wait for some write requests to complete then retry.
747 * In this implementation, __GFP_REPEAT means __GFP_NOFAIL, but that
748 * may not be true in other implementations.
751 if (!(gfp_mask & __GFP_NORETRY)) {
752 if ((order <= 3) || (gfp_mask & __GFP_REPEAT))
754 if (gfp_mask & __GFP_NOFAIL)
760 blk_congestion_wait(WRITE, HZ/50);
765 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
766 printk(KERN_WARNING "%s: page allocation failure."
767 " order:%d, mode:0x%x\n",
768 p->comm, order, gfp_mask);
773 kernel_map_pages(page, 1 << order, 1);
777 EXPORT_SYMBOL(__alloc_pages);
780 * Common helper functions.
782 fastcall unsigned long __get_free_pages(unsigned int gfp_mask, unsigned int order)
785 page = alloc_pages(gfp_mask, order);
788 return (unsigned long) page_address(page);
791 EXPORT_SYMBOL(__get_free_pages);
793 fastcall unsigned long get_zeroed_page(unsigned int gfp_mask)
798 * get_zeroed_page() returns a 32-bit address, which cannot represent
801 BUG_ON(gfp_mask & __GFP_HIGHMEM);
803 page = alloc_pages(gfp_mask, 0);
805 void *address = page_address(page);
807 return (unsigned long) address;
812 EXPORT_SYMBOL(get_zeroed_page);
814 void __pagevec_free(struct pagevec *pvec)
816 int i = pagevec_count(pvec);
819 free_hot_cold_page(pvec->pages[i], pvec->cold);
822 fastcall void __free_pages(struct page *page, unsigned int order)
824 if (!PageReserved(page) && put_page_testzero(page)) {
828 __free_pages_ok(page, order);
832 EXPORT_SYMBOL(__free_pages);
834 fastcall void free_pages(unsigned long addr, unsigned int order)
837 BUG_ON(!virt_addr_valid(addr));
838 __free_pages(virt_to_page(addr), order);
842 EXPORT_SYMBOL(free_pages);
845 * Total amount of free (allocatable) RAM:
847 unsigned int nr_free_pages(void)
849 unsigned int sum = 0;
853 sum += zone->free_pages;
858 EXPORT_SYMBOL(nr_free_pages);
861 unsigned int nr_free_pages_pgdat(pg_data_t *pgdat)
863 unsigned int i, sum = 0;
865 for (i = 0; i < MAX_NR_ZONES; i++)
866 sum += pgdat->node_zones[i].free_pages;
872 static unsigned int nr_free_zone_pages(int offset)
875 unsigned int sum = 0;
877 for_each_pgdat(pgdat) {
878 struct zonelist *zonelist = pgdat->node_zonelists + offset;
879 struct zone **zonep = zonelist->zones;
882 for (zone = *zonep++; zone; zone = *zonep++) {
883 unsigned long size = zone->present_pages;
884 unsigned long high = zone->pages_high;
894 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
896 unsigned int nr_free_buffer_pages(void)
898 return nr_free_zone_pages(GFP_USER & GFP_ZONEMASK);
902 * Amount of free RAM allocatable within all zones
904 unsigned int nr_free_pagecache_pages(void)
906 return nr_free_zone_pages(GFP_HIGHUSER & GFP_ZONEMASK);
909 #ifdef CONFIG_HIGHMEM
910 unsigned int nr_free_highpages (void)
913 unsigned int pages = 0;
915 for_each_pgdat(pgdat)
916 pages += pgdat->node_zones[ZONE_HIGHMEM].free_pages;
923 static void show_node(struct zone *zone)
925 printk("Node %d ", zone->zone_pgdat->node_id);
928 #define show_node(zone) do { } while (0)
932 * Accumulate the page_state information across all CPUs.
933 * The result is unavoidably approximate - it can change
934 * during and after execution of this function.
936 DEFINE_PER_CPU(struct page_state, page_states) = {0};
937 EXPORT_PER_CPU_SYMBOL(page_states);
939 atomic_t nr_pagecache = ATOMIC_INIT(0);
940 EXPORT_SYMBOL(nr_pagecache);
942 DEFINE_PER_CPU(long, nr_pagecache_local) = 0;
945 void __get_page_state(struct page_state *ret, int nr)
949 memset(ret, 0, sizeof(*ret));
950 while (cpu < NR_CPUS) {
951 unsigned long *in, *out, off;
953 if (!cpu_possible(cpu)) {
958 in = (unsigned long *)&per_cpu(page_states, cpu);
960 if (cpu < NR_CPUS && cpu_possible(cpu))
961 prefetch(&per_cpu(page_states, cpu));
962 out = (unsigned long *)ret;
963 for (off = 0; off < nr; off++)
968 void get_page_state(struct page_state *ret)
972 nr = offsetof(struct page_state, GET_PAGE_STATE_LAST);
973 nr /= sizeof(unsigned long);
975 __get_page_state(ret, nr + 1);
978 void get_full_page_state(struct page_state *ret)
980 __get_page_state(ret, sizeof(*ret) / sizeof(unsigned long));
983 unsigned long __read_page_state(unsigned offset)
985 unsigned long ret = 0;
988 for (cpu = 0; cpu < NR_CPUS; cpu++) {
991 if (!cpu_possible(cpu))
994 in = (unsigned long)&per_cpu(page_states, cpu) + offset;
995 ret += *((unsigned long *)in);
1000 void get_zone_counts(unsigned long *active,
1001 unsigned long *inactive, unsigned long *free)
1008 for_each_zone(zone) {
1009 *active += zone->nr_active;
1010 *inactive += zone->nr_inactive;
1011 *free += zone->free_pages;
1015 void si_meminfo(struct sysinfo *val)
1017 val->totalram = totalram_pages;
1019 val->freeram = nr_free_pages();
1020 val->bufferram = nr_blockdev_pages();
1021 #ifdef CONFIG_HIGHMEM
1022 val->totalhigh = totalhigh_pages;
1023 val->freehigh = nr_free_highpages();
1028 val->mem_unit = PAGE_SIZE;
1031 EXPORT_SYMBOL(si_meminfo);
1034 void si_meminfo_node(struct sysinfo *val, int nid)
1036 pg_data_t *pgdat = NODE_DATA(nid);
1038 val->totalram = pgdat->node_present_pages;
1039 val->freeram = nr_free_pages_pgdat(pgdat);
1040 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
1041 val->freehigh = pgdat->node_zones[ZONE_HIGHMEM].free_pages;
1042 val->mem_unit = PAGE_SIZE;
1046 #define K(x) ((x) << (PAGE_SHIFT-10))
1049 * Show free area list (used inside shift_scroll-lock stuff)
1050 * We also calculate the percentage fragmentation. We do this by counting the
1051 * memory on each free list with the exception of the first item on the list.
1053 void show_free_areas(void)
1055 struct page_state ps;
1056 int cpu, temperature;
1057 unsigned long active;
1058 unsigned long inactive;
1062 for_each_zone(zone) {
1064 printk("%s per-cpu:", zone->name);
1066 if (!zone->present_pages) {
1072 for (cpu = 0; cpu < NR_CPUS; ++cpu) {
1073 struct per_cpu_pageset *pageset;
1075 if (!cpu_possible(cpu))
1078 pageset = zone->pageset + cpu;
1080 for (temperature = 0; temperature < 2; temperature++)
1081 printk("cpu %d %s: low %d, high %d, batch %d\n",
1083 temperature ? "cold" : "hot",
1084 pageset->pcp[temperature].low,
1085 pageset->pcp[temperature].high,
1086 pageset->pcp[temperature].batch);
1090 get_page_state(&ps);
1091 get_zone_counts(&active, &inactive, &free);
1093 printk("\nFree pages: %11ukB (%ukB HighMem)\n",
1095 K(nr_free_highpages()));
1097 printk("Active:%lu inactive:%lu dirty:%lu writeback:%lu "
1098 "unstable:%lu free:%u slab:%lu mapped:%lu pagetables:%lu\n",
1107 ps.nr_page_table_pages);
1109 for_each_zone(zone) {
1123 K(zone->free_pages),
1126 K(zone->pages_high),
1128 K(zone->nr_inactive),
1129 K(zone->present_pages)
1131 printk("protections[]:");
1132 for (i = 0; i < MAX_NR_ZONES; i++)
1133 printk(" %lu", zone->protection[i]);
1137 for_each_zone(zone) {
1138 struct list_head *elem;
1139 unsigned long nr, flags, order, total = 0;
1142 printk("%s: ", zone->name);
1143 if (!zone->present_pages) {
1148 spin_lock_irqsave(&zone->lock, flags);
1149 for (order = 0; order < MAX_ORDER; order++) {
1151 list_for_each(elem, &zone->free_area[order].free_list)
1153 total += nr << order;
1154 printk("%lu*%lukB ", nr, K(1UL) << order);
1156 spin_unlock_irqrestore(&zone->lock, flags);
1157 printk("= %lukB\n", K(total));
1160 show_swap_cache_info();
1164 * Builds allocation fallback zone lists.
1166 static int __init build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist, int j, int k)
1173 zone = pgdat->node_zones + ZONE_HIGHMEM;
1174 if (zone->present_pages) {
1175 #ifndef CONFIG_HIGHMEM
1178 zonelist->zones[j++] = zone;
1181 zone = pgdat->node_zones + ZONE_NORMAL;
1182 if (zone->present_pages)
1183 zonelist->zones[j++] = zone;
1185 zone = pgdat->node_zones + ZONE_DMA;
1186 if (zone->present_pages)
1187 zonelist->zones[j++] = zone;
1194 #define MAX_NODE_LOAD (numnodes)
1195 static int __initdata node_load[MAX_NUMNODES];
1197 * find_next_best_node - find the next node that should appear in a given
1198 * node's fallback list
1199 * @node: node whose fallback list we're appending
1200 * @used_node_mask: pointer to the bitmap of already used nodes
1202 * We use a number of factors to determine which is the next node that should
1203 * appear on a given node's fallback list. The node should not have appeared
1204 * already in @node's fallback list, and it should be the next closest node
1205 * according to the distance array (which contains arbitrary distance values
1206 * from each node to each node in the system), and should also prefer nodes
1207 * with no CPUs, since presumably they'll have very little allocation pressure
1208 * on them otherwise.
1209 * It returns -1 if no node is found.
1211 static int __init find_next_best_node(int node, void *used_node_mask)
1214 int min_val = INT_MAX;
1217 for (i = 0; i < numnodes; i++) {
1220 /* Start from local node */
1221 n = (node+i)%numnodes;
1223 /* Don't want a node to appear more than once */
1224 if (test_bit(n, used_node_mask))
1227 /* Use the distance array to find the distance */
1228 val = node_distance(node, n);
1230 /* Give preference to headless and unused nodes */
1231 tmp = node_to_cpumask(n);
1232 if (!cpus_empty(tmp))
1233 val += PENALTY_FOR_NODE_WITH_CPUS;
1235 /* Slight preference for less loaded node */
1236 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
1237 val += node_load[n];
1239 if (val < min_val) {
1246 set_bit(best_node, used_node_mask);
1251 static void __init build_zonelists(pg_data_t *pgdat)
1253 int i, j, k, node, local_node;
1254 int prev_node, load;
1255 struct zonelist *zonelist;
1256 DECLARE_BITMAP(used_mask, MAX_NUMNODES);
1258 /* initialize zonelists */
1259 for (i = 0; i < GFP_ZONETYPES; i++) {
1260 zonelist = pgdat->node_zonelists + i;
1261 memset(zonelist, 0, sizeof(*zonelist));
1262 zonelist->zones[0] = NULL;
1265 /* NUMA-aware ordering of nodes */
1266 local_node = pgdat->node_id;
1268 prev_node = local_node;
1269 bitmap_zero(used_mask, MAX_NUMNODES);
1270 while ((node = find_next_best_node(local_node, used_mask)) >= 0) {
1272 * We don't want to pressure a particular node.
1273 * So adding penalty to the first node in same
1274 * distance group to make it round-robin.
1276 if (node_distance(local_node, node) !=
1277 node_distance(local_node, prev_node))
1278 node_load[node] += load;
1281 for (i = 0; i < GFP_ZONETYPES; i++) {
1282 zonelist = pgdat->node_zonelists + i;
1283 for (j = 0; zonelist->zones[j] != NULL; j++);
1286 if (i & __GFP_HIGHMEM)
1291 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1292 zonelist->zones[j] = NULL;
1297 #else /* CONFIG_NUMA */
1299 static void __init build_zonelists(pg_data_t *pgdat)
1301 int i, j, k, node, local_node;
1303 local_node = pgdat->node_id;
1304 for (i = 0; i < GFP_ZONETYPES; i++) {
1305 struct zonelist *zonelist;
1307 zonelist = pgdat->node_zonelists + i;
1308 memset(zonelist, 0, sizeof(*zonelist));
1312 if (i & __GFP_HIGHMEM)
1317 j = build_zonelists_node(pgdat, zonelist, j, k);
1319 * Now we build the zonelist so that it contains the zones
1320 * of all the other nodes.
1321 * We don't want to pressure a particular node, so when
1322 * building the zones for node N, we make sure that the
1323 * zones coming right after the local ones are those from
1324 * node N+1 (modulo N)
1326 for (node = local_node + 1; node < numnodes; node++)
1327 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1328 for (node = 0; node < local_node; node++)
1329 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1331 zonelist->zones[j] = NULL;
1335 #endif /* CONFIG_NUMA */
1337 void __init build_all_zonelists(void)
1341 for(i = 0 ; i < numnodes ; i++)
1342 build_zonelists(NODE_DATA(i));
1343 printk("Built %i zonelists\n", numnodes);
1347 * Helper functions to size the waitqueue hash table.
1348 * Essentially these want to choose hash table sizes sufficiently
1349 * large so that collisions trying to wait on pages are rare.
1350 * But in fact, the number of active page waitqueues on typical
1351 * systems is ridiculously low, less than 200. So this is even
1352 * conservative, even though it seems large.
1354 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
1355 * waitqueues, i.e. the size of the waitq table given the number of pages.
1357 #define PAGES_PER_WAITQUEUE 256
1359 static inline unsigned long wait_table_size(unsigned long pages)
1361 unsigned long size = 1;
1363 pages /= PAGES_PER_WAITQUEUE;
1365 while (size < pages)
1369 * Once we have dozens or even hundreds of threads sleeping
1370 * on IO we've got bigger problems than wait queue collision.
1371 * Limit the size of the wait table to a reasonable size.
1373 size = min(size, 4096UL);
1375 return max(size, 4UL);
1379 * This is an integer logarithm so that shifts can be used later
1380 * to extract the more random high bits from the multiplicative
1381 * hash function before the remainder is taken.
1383 static inline unsigned long wait_table_bits(unsigned long size)
1388 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
1390 static void __init calculate_zone_totalpages(struct pglist_data *pgdat,
1391 unsigned long *zones_size, unsigned long *zholes_size)
1393 unsigned long realtotalpages, totalpages = 0;
1396 for (i = 0; i < MAX_NR_ZONES; i++)
1397 totalpages += zones_size[i];
1398 pgdat->node_spanned_pages = totalpages;
1400 realtotalpages = totalpages;
1402 for (i = 0; i < MAX_NR_ZONES; i++)
1403 realtotalpages -= zholes_size[i];
1404 pgdat->node_present_pages = realtotalpages;
1405 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id, realtotalpages);
1410 * Initially all pages are reserved - free ones are freed
1411 * up by free_all_bootmem() once the early boot process is
1412 * done. Non-atomic initialization, single-pass.
1414 void __init memmap_init_zone(struct page *start, unsigned long size, int nid,
1415 unsigned long zone, unsigned long start_pfn)
1419 for (page = start; page < (start + size); page++) {
1420 set_page_zone(page, NODEZONE(nid, zone));
1421 set_page_count(page, 0);
1422 SetPageReserved(page);
1423 INIT_LIST_HEAD(&page->lru);
1424 #ifdef WANT_PAGE_VIRTUAL
1425 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1426 if (!is_highmem_idx(zone))
1427 set_page_address(page, __va(start_pfn << PAGE_SHIFT));
1433 #ifndef __HAVE_ARCH_MEMMAP_INIT
1434 #define memmap_init(start, size, nid, zone, start_pfn) \
1435 memmap_init_zone((start), (size), (nid), (zone), (start_pfn))
1439 * Set up the zone data structures:
1440 * - mark all pages reserved
1441 * - mark all memory queues empty
1442 * - clear the memory bitmaps
1444 static void __init free_area_init_core(struct pglist_data *pgdat,
1445 unsigned long *zones_size, unsigned long *zholes_size)
1448 const unsigned long zone_required_alignment = 1UL << (MAX_ORDER-1);
1449 int cpu, nid = pgdat->node_id;
1450 struct page *lmem_map = pgdat->node_mem_map;
1451 unsigned long zone_start_pfn = pgdat->node_start_pfn;
1453 pgdat->nr_zones = 0;
1454 init_waitqueue_head(&pgdat->kswapd_wait);
1456 for (j = 0; j < MAX_NR_ZONES; j++) {
1457 struct zone *zone = pgdat->node_zones + j;
1458 unsigned long size, realsize;
1459 unsigned long batch;
1461 zone_table[NODEZONE(nid, j)] = zone;
1462 realsize = size = zones_size[j];
1464 realsize -= zholes_size[j];
1466 if (j == ZONE_DMA || j == ZONE_NORMAL)
1467 nr_kernel_pages += realsize;
1468 nr_all_pages += realsize;
1470 zone->spanned_pages = size;
1471 zone->present_pages = realsize;
1472 zone->name = zone_names[j];
1473 spin_lock_init(&zone->lock);
1474 spin_lock_init(&zone->lru_lock);
1475 zone->zone_pgdat = pgdat;
1476 zone->free_pages = 0;
1478 zone->temp_priority = zone->prev_priority = DEF_PRIORITY;
1481 * The per-cpu-pages pools are set to around 1000th of the
1482 * size of the zone. But no more than 1/4 of a meg - there's
1483 * no point in going beyond the size of L2 cache.
1485 * OK, so we don't know how big the cache is. So guess.
1487 batch = zone->present_pages / 1024;
1488 if (batch * PAGE_SIZE > 256 * 1024)
1489 batch = (256 * 1024) / PAGE_SIZE;
1490 batch /= 4; /* We effectively *= 4 below */
1494 for (cpu = 0; cpu < NR_CPUS; cpu++) {
1495 struct per_cpu_pages *pcp;
1497 pcp = &zone->pageset[cpu].pcp[0]; /* hot */
1499 pcp->low = 2 * batch;
1500 pcp->high = 6 * batch;
1501 pcp->batch = 1 * batch;
1502 INIT_LIST_HEAD(&pcp->list);
1504 pcp = &zone->pageset[cpu].pcp[1]; /* cold */
1507 pcp->high = 2 * batch;
1508 pcp->batch = 1 * batch;
1509 INIT_LIST_HEAD(&pcp->list);
1511 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%lu\n",
1512 zone_names[j], realsize, batch);
1513 INIT_LIST_HEAD(&zone->active_list);
1514 INIT_LIST_HEAD(&zone->inactive_list);
1515 zone->nr_scan_active = 0;
1516 zone->nr_scan_inactive = 0;
1517 zone->nr_active = 0;
1518 zone->nr_inactive = 0;
1523 * The per-page waitqueue mechanism uses hashed waitqueues
1526 zone->wait_table_size = wait_table_size(size);
1527 zone->wait_table_bits =
1528 wait_table_bits(zone->wait_table_size);
1529 zone->wait_table = (wait_queue_head_t *)
1530 alloc_bootmem_node(pgdat, zone->wait_table_size
1531 * sizeof(wait_queue_head_t));
1533 for(i = 0; i < zone->wait_table_size; ++i)
1534 init_waitqueue_head(zone->wait_table + i);
1536 pgdat->nr_zones = j+1;
1538 zone->zone_mem_map = lmem_map;
1539 zone->zone_start_pfn = zone_start_pfn;
1541 if ((zone_start_pfn) & (zone_required_alignment-1))
1542 printk("BUG: wrong zone alignment, it will crash\n");
1544 memmap_init(lmem_map, size, nid, j, zone_start_pfn);
1546 zone_start_pfn += size;
1549 for (i = 0; ; i++) {
1550 unsigned long bitmap_size;
1552 INIT_LIST_HEAD(&zone->free_area[i].free_list);
1553 if (i == MAX_ORDER-1) {
1554 zone->free_area[i].map = NULL;
1559 * Page buddy system uses "index >> (i+1)",
1560 * where "index" is at most "size-1".
1562 * The extra "+3" is to round down to byte
1563 * size (8 bits per byte assumption). Thus
1564 * we get "(size-1) >> (i+4)" as the last byte
1567 * The "+1" is because we want to round the
1568 * byte allocation up rather than down. So
1569 * we should have had a "+7" before we shifted
1570 * down by three. Also, we have to add one as
1571 * we actually _use_ the last bit (it's [0,n]
1572 * inclusive, not [0,n[).
1574 * So we actually had +7+1 before we shift
1575 * down by 3. But (n+8) >> 3 == (n >> 3) + 1
1576 * (modulo overflows, which we do not have).
1578 * Finally, we LONG_ALIGN because all bitmap
1579 * operations are on longs.
1581 bitmap_size = (size-1) >> (i+4);
1582 bitmap_size = LONG_ALIGN(bitmap_size+1);
1583 zone->free_area[i].map =
1584 (unsigned long *) alloc_bootmem_node(pgdat, bitmap_size);
1589 void __init free_area_init_node(int nid, struct pglist_data *pgdat,
1590 struct page *node_mem_map, unsigned long *zones_size,
1591 unsigned long node_start_pfn, unsigned long *zholes_size)
1595 pgdat->node_id = nid;
1596 pgdat->node_start_pfn = node_start_pfn;
1597 calculate_zone_totalpages(pgdat, zones_size, zholes_size);
1598 if (!node_mem_map) {
1599 size = (pgdat->node_spanned_pages + 1) * sizeof(struct page);
1600 node_mem_map = alloc_bootmem_node(pgdat, size);
1602 pgdat->node_mem_map = node_mem_map;
1604 free_area_init_core(pgdat, zones_size, zholes_size);
1607 #ifndef CONFIG_DISCONTIGMEM
1608 static bootmem_data_t contig_bootmem_data;
1609 struct pglist_data contig_page_data = { .bdata = &contig_bootmem_data };
1611 EXPORT_SYMBOL(contig_page_data);
1613 void __init free_area_init(unsigned long *zones_size)
1615 free_area_init_node(0, &contig_page_data, NULL, zones_size,
1616 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
1617 mem_map = contig_page_data.node_mem_map;
1621 #ifdef CONFIG_PROC_FS
1623 #include <linux/seq_file.h>
1625 static void *frag_start(struct seq_file *m, loff_t *pos)
1630 for (pgdat = pgdat_list; pgdat && node; pgdat = pgdat->pgdat_next)
1636 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
1638 pg_data_t *pgdat = (pg_data_t *)arg;
1641 return pgdat->pgdat_next;
1644 static void frag_stop(struct seq_file *m, void *arg)
1649 * This walks the freelist for each zone. Whilst this is slow, I'd rather
1650 * be slow here than slow down the fast path by keeping stats - mjbligh
1652 static int frag_show(struct seq_file *m, void *arg)
1654 pg_data_t *pgdat = (pg_data_t *)arg;
1656 struct zone *node_zones = pgdat->node_zones;
1657 unsigned long flags;
1660 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
1661 if (!zone->present_pages)
1664 spin_lock_irqsave(&zone->lock, flags);
1665 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1666 for (order = 0; order < MAX_ORDER; ++order) {
1667 unsigned long nr_bufs = 0;
1668 struct list_head *elem;
1670 list_for_each(elem, &(zone->free_area[order].free_list))
1672 seq_printf(m, "%6lu ", nr_bufs);
1674 spin_unlock_irqrestore(&zone->lock, flags);
1680 struct seq_operations fragmentation_op = {
1681 .start = frag_start,
1687 static char *vmstat_text[] = {
1691 "nr_page_table_pages",
1716 "pgscan_kswapd_high",
1717 "pgscan_kswapd_normal",
1719 "pgscan_kswapd_dma",
1720 "pgscan_direct_high",
1721 "pgscan_direct_normal",
1722 "pgscan_direct_dma",
1727 "kswapd_inodesteal",
1734 static void *vmstat_start(struct seq_file *m, loff_t *pos)
1736 struct page_state *ps;
1738 if (*pos >= ARRAY_SIZE(vmstat_text))
1741 ps = kmalloc(sizeof(*ps), GFP_KERNEL);
1744 return ERR_PTR(-ENOMEM);
1745 get_full_page_state(ps);
1746 ps->pgpgin /= 2; /* sectors -> kbytes */
1748 return (unsigned long *)ps + *pos;
1751 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
1754 if (*pos >= ARRAY_SIZE(vmstat_text))
1756 return (unsigned long *)m->private + *pos;
1759 static int vmstat_show(struct seq_file *m, void *arg)
1761 unsigned long *l = arg;
1762 unsigned long off = l - (unsigned long *)m->private;
1764 seq_printf(m, "%s %lu\n", vmstat_text[off], *l);
1768 static void vmstat_stop(struct seq_file *m, void *arg)
1774 struct seq_operations vmstat_op = {
1775 .start = vmstat_start,
1776 .next = vmstat_next,
1777 .stop = vmstat_stop,
1778 .show = vmstat_show,
1781 #endif /* CONFIG_PROC_FS */
1783 #ifdef CONFIG_HOTPLUG_CPU
1784 static int page_alloc_cpu_notify(struct notifier_block *self,
1785 unsigned long action, void *hcpu)
1787 int cpu = (unsigned long)hcpu;
1790 if (action == CPU_DEAD) {
1791 /* Drain local pagecache count. */
1792 count = &per_cpu(nr_pagecache_local, cpu);
1793 atomic_add(*count, &nr_pagecache);
1795 local_irq_disable();
1801 #endif /* CONFIG_HOTPLUG_CPU */
1803 void __init page_alloc_init(void)
1805 hotcpu_notifier(page_alloc_cpu_notify, 0);
1808 static unsigned long higherzone_val(struct zone *z, int max_zone,
1811 int z_idx = zone_idx(z);
1812 struct zone *higherzone;
1813 unsigned long pages;
1815 /* there is no higher zone to get a contribution from */
1816 if (z_idx == MAX_NR_ZONES-1)
1819 higherzone = &z->zone_pgdat->node_zones[z_idx+1];
1821 /* We always start with the higher zone's protection value */
1822 pages = higherzone->protection[alloc_type];
1825 * We get a lower-zone-protection contribution only if there are
1826 * pages in the higher zone and if we're not the highest zone
1827 * in the current zonelist. e.g., never happens for GFP_DMA. Happens
1828 * only for ZONE_DMA in a GFP_KERNEL allocation and happens for ZONE_DMA
1829 * and ZONE_NORMAL for a GFP_HIGHMEM allocation.
1831 if (higherzone->present_pages && z_idx < alloc_type)
1832 pages += higherzone->pages_low * sysctl_lower_zone_protection;
1838 * setup_per_zone_protection - called whenver min_free_kbytes or
1839 * sysctl_lower_zone_protection changes. Ensures that each zone
1840 * has a correct pages_protected value, so an adequate number of
1841 * pages are left in the zone after a successful __alloc_pages().
1843 * This algorithm is way confusing. I tries to keep the same behavior
1844 * as we had with the incremental min iterative algorithm.
1846 static void setup_per_zone_protection(void)
1848 struct pglist_data *pgdat;
1849 struct zone *zones, *zone;
1853 for_each_pgdat(pgdat) {
1854 zones = pgdat->node_zones;
1856 for (i = 0, max_zone = 0; i < MAX_NR_ZONES; i++)
1857 if (zones[i].present_pages)
1861 * For each of the different allocation types:
1862 * GFP_DMA -> GFP_KERNEL -> GFP_HIGHMEM
1864 for (i = 0; i < GFP_ZONETYPES; i++) {
1866 * For each of the zones:
1867 * ZONE_HIGHMEM -> ZONE_NORMAL -> ZONE_DMA
1869 for (j = MAX_NR_ZONES-1; j >= 0; j--) {
1873 * We never protect zones that don't have memory
1874 * in them (j>max_zone) or zones that aren't in
1875 * the zonelists for a certain type of
1876 * allocation (j>i). We have to assign these to
1877 * zero because the lower zones take
1878 * contributions from the higher zones.
1880 if (j > max_zone || j > i) {
1881 zone->protection[i] = 0;
1885 * The contribution of the next higher zone
1887 zone->protection[i] = higherzone_val(zone,
1889 zone->protection[i] += zone->pages_low;
1896 * setup_per_zone_pages_min - called when min_free_kbytes changes. Ensures
1897 * that the pages_{min,low,high} values for each zone are set correctly
1898 * with respect to min_free_kbytes.
1900 static void setup_per_zone_pages_min(void)
1902 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
1903 unsigned long lowmem_pages = 0;
1905 unsigned long flags;
1907 /* Calculate total number of !ZONE_HIGHMEM pages */
1908 for_each_zone(zone) {
1909 if (!is_highmem(zone))
1910 lowmem_pages += zone->present_pages;
1913 for_each_zone(zone) {
1914 spin_lock_irqsave(&zone->lru_lock, flags);
1915 #if !defined(CONFIG_MIPS_BRCM)
1916 if (is_highmem(zone)) {
1918 * Often, highmem doesn't need to reserve any pages.
1919 * But the pages_min/low/high values are also used for
1920 * batching up page reclaim activity so we need a
1921 * decent value here.
1925 min_pages = zone->present_pages / 1024;
1926 if (min_pages < SWAP_CLUSTER_MAX)
1927 min_pages = SWAP_CLUSTER_MAX;
1928 if (min_pages > 128)
1930 zone->pages_min = min_pages;
1932 /* if it's a lowmem zone, reserve a number of pages
1933 * proportionate to the zone's size.
1935 zone->pages_min = (pages_min * zone->present_pages) /
1938 zone->pages_low = zone->pages_min * 2;
1939 zone->pages_high = zone->pages_min * 3;
1941 /* Tuned watermarks for better out of memory performance on our swapless system */
1942 /* As we don't have a disc to swap out dirty pages, there is not much we can do if memory is really low */
1943 /* We want to lower watermarks to prevent excessive scanning, which could slow down or lock up the system */
1944 zone->pages_min = 0; /* this is actually not used elsewhere except for calculating the other two */
1945 zone->pages_low = 48;
1946 zone->pages_high = 50;
1949 spin_unlock_irqrestore(&zone->lru_lock, flags);
1954 * Initialise min_free_kbytes.
1956 * For small machines we want it small (128k min). For large machines
1957 * we want it large (16MB max). But it is not linear, because network
1958 * bandwidth does not increase linearly with machine size. We use
1960 * min_free_kbytes = sqrt(lowmem_kbytes)
1976 static int __init init_per_zone_pages_min(void)
1978 unsigned long lowmem_kbytes;
1980 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
1982 min_free_kbytes = int_sqrt(lowmem_kbytes);
1983 if (min_free_kbytes < 128)
1984 min_free_kbytes = 128;
1985 if (min_free_kbytes > 16384)
1986 min_free_kbytes = 16384;
1987 setup_per_zone_pages_min();
1988 setup_per_zone_protection();
1991 module_init(init_per_zone_pages_min)
1994 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
1995 * that we can call two helper functions whenever min_free_kbytes
1998 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
1999 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2001 proc_dointvec(table, write, file, buffer, length, ppos);
2002 setup_per_zone_pages_min();
2003 setup_per_zone_protection();
2008 * lower_zone_protection_sysctl_handler - just a wrapper around
2009 * proc_dointvec() so that we can call setup_per_zone_protection()
2010 * whenever sysctl_lower_zone_protection changes.
2012 int lower_zone_protection_sysctl_handler(ctl_table *table, int write,
2013 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2015 proc_dointvec_minmax(table, write, file, buffer, length, ppos);
2016 setup_per_zone_protection();
2021 * allocate a large system hash table from bootmem
2022 * - it is assumed that the hash table must contain an exact power-of-2
2023 * quantity of entries
2025 void *__init alloc_large_system_hash(const char *tablename,
2026 unsigned long bucketsize,
2027 unsigned long numentries,
2029 int consider_highmem,
2030 unsigned int *_hash_shift,
2031 unsigned int *_hash_mask)
2033 unsigned long mem, max, log2qty, size;
2036 /* round applicable memory size up to nearest megabyte */
2037 mem = consider_highmem ? nr_all_pages : nr_kernel_pages;
2038 mem += (1UL << (20 - PAGE_SHIFT)) - 1;
2039 mem >>= 20 - PAGE_SHIFT;
2040 mem <<= 20 - PAGE_SHIFT;
2042 /* limit to 1 bucket per 2^scale bytes of low memory (rounded up to
2043 * nearest power of 2 in size) */
2044 if (scale > PAGE_SHIFT)
2045 mem >>= (scale - PAGE_SHIFT);
2047 mem <<= (PAGE_SHIFT - scale);
2049 mem = 1UL << (long_log2(mem) + 1);
2051 /* limit allocation size */
2052 max = (1UL << (PAGE_SHIFT + MAX_SYS_HASH_TABLE_ORDER)) / bucketsize;
2056 /* allow the kernel cmdline to have a say */
2057 if (!numentries || numentries > max)
2060 log2qty = long_log2(numentries);
2063 size = bucketsize << log2qty;
2065 table = (void *) alloc_bootmem(size);
2067 } while (!table && size > PAGE_SIZE);
2070 panic("Failed to allocate %s hash table\n", tablename);
2072 printk("%s hash table entries: %d (order: %d, %lu bytes)\n",
2075 long_log2(size) - PAGE_SHIFT,
2079 *_hash_shift = log2qty;
2081 *_hash_mask = (1 << log2qty) - 1;