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;
59 * Temporary debugging check for pages not lying within a given zone.
61 static int bad_range(struct zone *zone, struct page *page)
63 if (page_to_pfn(page) >= zone->zone_start_pfn + zone->spanned_pages)
65 if (page_to_pfn(page) < zone->zone_start_pfn)
67 if (zone != page_zone(page))
72 static void bad_page(const char *function, struct page *page)
74 printk(KERN_EMERG "Bad page state at %s (in process '%s', page %p)\n",
75 function, current->comm, page);
76 printk(KERN_EMERG "flags:0x%08lx mapping:%p mapped:%d count:%d\n",
77 (unsigned long)page->flags, page->mapping,
78 page_mapped(page), page_count(page));
79 printk(KERN_EMERG "Backtrace:\n");
81 printk(KERN_EMERG "Trying to fix it up, but a reboot is needed\n");
82 page->flags &= ~(1 << PG_private |
91 set_page_count(page, 0);
95 #ifndef CONFIG_HUGETLB_PAGE
96 #define prep_compound_page(page, order) do { } while (0)
97 #define destroy_compound_page(page, order) do { } while (0)
100 * Higher-order pages are called "compound pages". They are structured thusly:
102 * The first PAGE_SIZE page is called the "head page".
104 * The remaining PAGE_SIZE pages are called "tail pages".
106 * All pages have PG_compound set. All pages have their ->private pointing at
107 * the head page (even the head page has this).
109 * The first tail page's ->mapping, if non-zero, holds the address of the
110 * compound page's put_page() function.
112 * The order of the allocation is stored in the first tail page's ->index
113 * This is only for debug at present. This usage means that zero-order pages
114 * may not be compound.
116 static void prep_compound_page(struct page *page, unsigned long order)
119 int nr_pages = 1 << order;
122 page[1].index = order;
123 for (i = 0; i < nr_pages; i++) {
124 struct page *p = page + i;
127 p->private = (unsigned long)page;
131 static void destroy_compound_page(struct page *page, unsigned long order)
134 int nr_pages = 1 << order;
136 if (!PageCompound(page))
139 if (page[1].index != order)
140 bad_page(__FUNCTION__, page);
142 for (i = 0; i < nr_pages; i++) {
143 struct page *p = page + i;
145 if (!PageCompound(p))
146 bad_page(__FUNCTION__, page);
147 if (p->private != (unsigned long)page)
148 bad_page(__FUNCTION__, page);
149 ClearPageCompound(p);
152 #endif /* CONFIG_HUGETLB_PAGE */
155 * Freeing function for a buddy system allocator.
157 * The concept of a buddy system is to maintain direct-mapped table
158 * (containing bit values) for memory blocks of various "orders".
159 * The bottom level table contains the map for the smallest allocatable
160 * units of memory (here, pages), and each level above it describes
161 * pairs of units from the levels below, hence, "buddies".
162 * At a high level, all that happens here is marking the table entry
163 * at the bottom level available, and propagating the changes upward
164 * as necessary, plus some accounting needed to play nicely with other
165 * parts of the VM system.
166 * At each level, we keep one bit for each pair of blocks, which
167 * is set to 1 iff only one of the pair is allocated. So when we
168 * are allocating or freeing one, we can derive the state of the
169 * other. That is, if we allocate a small block, and both were
170 * free, the remainder of the region must be split into blocks.
171 * If a block is freed, and its buddy is also free, then this
172 * triggers coalescing into a block of larger size.
177 static inline void __free_pages_bulk (struct page *page, struct page *base,
178 struct zone *zone, struct free_area *area, unsigned long mask,
181 unsigned long page_idx, index;
184 destroy_compound_page(page, order);
185 page_idx = page - base;
186 if (page_idx & ~mask)
188 index = page_idx >> (1 + order);
190 zone->free_pages -= mask;
191 while (mask + (1 << (MAX_ORDER-1))) {
192 struct page *buddy1, *buddy2;
194 BUG_ON(area >= zone->free_area + MAX_ORDER);
195 if (!__test_and_change_bit(index, area->map))
197 * the buddy page is still allocated.
201 * Move the buddy up one level.
202 * This code is taking advantage of the identity:
205 buddy1 = base + (page_idx ^ -mask);
206 buddy2 = base + page_idx;
207 BUG_ON(bad_range(zone, buddy1));
208 BUG_ON(bad_range(zone, buddy2));
209 list_del(&buddy1->lru);
215 list_add(&(base + page_idx)->lru, &area->free_list);
218 static inline void free_pages_check(const char *function, struct page *page)
220 if ( page_mapped(page) ||
221 page->mapping != NULL ||
222 page_count(page) != 0 ||
233 1 << PG_writeback )))
234 bad_page(function, page);
236 ClearPageDirty(page);
240 * Frees a list of pages.
241 * Assumes all pages on list are in same zone, and of same order.
242 * count is the number of pages to free, or 0 for all on the list.
244 * If the zone was previously in an "all pages pinned" state then look to
245 * see if this freeing clears that state.
247 * And clear the zone's pages_scanned counter, to hold off the "all pages are
248 * pinned" detection logic.
251 free_pages_bulk(struct zone *zone, int count,
252 struct list_head *list, unsigned int order)
254 unsigned long mask, flags;
255 struct free_area *area;
256 struct page *base, *page = NULL;
259 mask = (~0UL) << order;
260 base = zone->zone_mem_map;
261 area = zone->free_area + order;
262 spin_lock_irqsave(&zone->lock, flags);
263 zone->all_unreclaimable = 0;
264 zone->pages_scanned = 0;
265 while (!list_empty(list) && count--) {
266 page = list_entry(list->prev, struct page, lru);
267 /* have to delete it as __free_pages_bulk list manipulates */
268 list_del(&page->lru);
269 __free_pages_bulk(page, base, zone, area, mask, order);
272 spin_unlock_irqrestore(&zone->lock, flags);
276 void __free_pages_ok(struct page *page, unsigned int order)
281 mod_page_state(pgfree, 1 << order);
282 for (i = 0 ; i < (1 << order) ; ++i)
283 free_pages_check(__FUNCTION__, page + i);
284 list_add(&page->lru, &list);
285 kernel_map_pages(page, 1<<order, 0);
286 free_pages_bulk(page_zone(page), 1, &list, order);
289 #define MARK_USED(index, order, area) \
290 __change_bit((index) >> (1+(order)), (area)->map)
292 static inline struct page *
293 expand(struct zone *zone, struct page *page,
294 unsigned long index, int low, int high, struct free_area *area)
296 unsigned long size = 1 << high;
299 BUG_ON(bad_range(zone, page));
303 list_add(&page->lru, &area->free_list);
304 MARK_USED(index, high, area);
311 static inline void set_page_refs(struct page *page, int order)
314 set_page_count(page, 1);
319 * We need to reference all the pages for this order, otherwise if
320 * anyone accesses one of the pages with (get/put) it will be freed.
322 for (i = 0; i < (1 << order); i++)
323 set_page_count(page+i, 1);
324 #endif /* CONFIG_MMU */
328 * This page is about to be returned from the page allocator
330 static void prep_new_page(struct page *page, int order)
332 if (page->mapping || page_mapped(page) ||
343 1 << PG_writeback )))
344 bad_page(__FUNCTION__, page);
346 page->flags &= ~(1 << PG_uptodate | 1 << PG_error |
347 1 << PG_referenced | 1 << PG_arch_1 |
348 1 << PG_checked | 1 << PG_mappedtodisk);
350 set_page_refs(page, order);
354 * Do the hard work of removing an element from the buddy allocator.
355 * Call me with the zone->lock already held.
357 static struct page *__rmqueue(struct zone *zone, unsigned int order)
359 struct free_area * area;
360 unsigned int current_order;
364 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
365 area = zone->free_area + current_order;
366 if (list_empty(&area->free_list))
369 page = list_entry(area->free_list.next, struct page, lru);
370 list_del(&page->lru);
371 index = page - zone->zone_mem_map;
372 if (current_order != MAX_ORDER-1)
373 MARK_USED(index, current_order, area);
374 zone->free_pages -= 1UL << order;
375 return expand(zone, page, index, order, current_order, area);
382 * Obtain a specified number of elements from the buddy allocator, all under
383 * a single hold of the lock, for efficiency. Add them to the supplied list.
384 * Returns the number of new pages which were placed at *list.
386 static int rmqueue_bulk(struct zone *zone, unsigned int order,
387 unsigned long count, struct list_head *list)
394 spin_lock_irqsave(&zone->lock, flags);
395 for (i = 0; i < count; ++i) {
396 page = __rmqueue(zone, order);
400 list_add_tail(&page->lru, list);
402 spin_unlock_irqrestore(&zone->lock, flags);
406 #if defined(CONFIG_PM) || defined(CONFIG_HOTPLUG_CPU)
407 static void __drain_pages(unsigned int cpu)
412 for_each_zone(zone) {
413 struct per_cpu_pageset *pset;
415 pset = &zone->pageset[cpu];
416 for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
417 struct per_cpu_pages *pcp;
420 pcp->count -= free_pages_bulk(zone, pcp->count,
425 #endif /* CONFIG_PM || CONFIG_HOTPLUG_CPU */
428 int is_head_of_free_region(struct page *page)
430 struct zone *zone = page_zone(page);
433 struct list_head *curr;
436 * Should not matter as we need quiescent system for
437 * suspend anyway, but...
439 spin_lock_irqsave(&zone->lock, flags);
440 for (order = MAX_ORDER - 1; order >= 0; --order)
441 list_for_each(curr, &zone->free_area[order].free_list)
442 if (page == list_entry(curr, struct page, lru)) {
443 spin_unlock_irqrestore(&zone->lock, flags);
446 spin_unlock_irqrestore(&zone->lock, flags);
451 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
453 void drain_local_pages(void)
457 local_irq_save(flags);
458 __drain_pages(smp_processor_id());
459 local_irq_restore(flags);
461 #endif /* CONFIG_PM */
464 * Free a 0-order page
466 static void FASTCALL(free_hot_cold_page(struct page *page, int cold));
467 static void fastcall free_hot_cold_page(struct page *page, int cold)
469 struct zone *zone = page_zone(page);
470 struct per_cpu_pages *pcp;
473 kernel_map_pages(page, 1, 0);
474 inc_page_state(pgfree);
475 free_pages_check(__FUNCTION__, page);
476 pcp = &zone->pageset[get_cpu()].pcp[cold];
477 local_irq_save(flags);
478 if (pcp->count >= pcp->high)
479 pcp->count -= free_pages_bulk(zone, pcp->batch, &pcp->list, 0);
480 list_add(&page->lru, &pcp->list);
482 local_irq_restore(flags);
486 void fastcall free_hot_page(struct page *page)
488 free_hot_cold_page(page, 0);
491 void fastcall free_cold_page(struct page *page)
493 free_hot_cold_page(page, 1);
497 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
498 * we cheat by calling it from here, in the order > 0 path. Saves a branch
503 buffered_rmqueue(struct zone *zone, int order, int gfp_flags)
506 struct page *page = NULL;
507 int cold = !!(gfp_flags & __GFP_COLD);
510 struct per_cpu_pages *pcp;
512 pcp = &zone->pageset[get_cpu()].pcp[cold];
513 local_irq_save(flags);
514 if (pcp->count <= pcp->low)
515 pcp->count += rmqueue_bulk(zone, 0,
516 pcp->batch, &pcp->list);
518 page = list_entry(pcp->list.next, struct page, lru);
519 list_del(&page->lru);
522 local_irq_restore(flags);
527 spin_lock_irqsave(&zone->lock, flags);
528 page = __rmqueue(zone, order);
529 spin_unlock_irqrestore(&zone->lock, flags);
533 BUG_ON(bad_range(zone, page));
534 mod_page_state_zone(zone, pgalloc, 1 << order);
535 prep_new_page(page, order);
536 if (order && (gfp_flags & __GFP_COMP))
537 prep_compound_page(page, order);
543 * This is the 'heart' of the zoned buddy allocator.
545 * Herein lies the mysterious "incremental min". That's the
547 * local_low = z->pages_low;
550 * thing. The intent here is to provide additional protection to low zones for
551 * allocation requests which _could_ use higher zones. So a GFP_HIGHMEM
552 * request is not allowed to dip as deeply into the normal zone as a GFP_KERNEL
553 * request. This preserves additional space in those lower zones for requests
554 * which really do need memory from those zones. It means that on a decent
555 * sized machine, GFP_HIGHMEM and GFP_KERNEL requests basically leave the DMA
558 struct page * fastcall
559 __alloc_pages(unsigned int gfp_mask, unsigned int order,
560 struct zonelist *zonelist)
562 const int wait = gfp_mask & __GFP_WAIT;
566 struct reclaim_state reclaim_state;
567 struct task_struct *p = current;
572 might_sleep_if(wait);
574 zones = zonelist->zones; /* the list of zones suitable for gfp_mask */
575 if (zones[0] == NULL) /* no zones in the zonelist */
578 alloc_type = zone_idx(zones[0]);
580 /* Go through the zonelist once, looking for a zone with enough free */
581 for (i = 0; zones[i] != NULL; i++) {
582 struct zone *z = zones[i];
584 min = (1<<order) + z->protection[alloc_type];
587 * We let real-time tasks dip their real-time paws a little
588 * deeper into reserves.
591 min -= z->pages_low >> 1;
593 if (z->free_pages >= min ||
594 (!wait && z->free_pages >= z->pages_high)) {
595 page = buffered_rmqueue(z, order, gfp_mask);
601 /* we're somewhat low on memory, failed to find what we needed */
602 for (i = 0; zones[i] != NULL; i++)
603 wakeup_kswapd(zones[i]);
605 /* Go through the zonelist again, taking __GFP_HIGH into account */
606 for (i = 0; zones[i] != NULL; i++) {
607 struct zone *z = zones[i];
609 min = (1<<order) + z->protection[alloc_type];
611 if (gfp_mask & __GFP_HIGH)
612 min -= z->pages_low >> 2;
614 min -= z->pages_low >> 1;
616 if (z->free_pages >= min ||
617 (!wait && z->free_pages >= z->pages_high)) {
618 page = buffered_rmqueue(z, order, gfp_mask);
624 /* here we're in the low on memory slow path */
627 if ((p->flags & (PF_MEMALLOC | PF_MEMDIE)) && !in_interrupt()) {
628 /* go through the zonelist yet again, ignoring mins */
629 for (i = 0; zones[i] != NULL; i++) {
630 struct zone *z = zones[i];
632 page = buffered_rmqueue(z, order, gfp_mask);
639 /* Atomic allocations - we can't balance anything */
643 p->flags |= PF_MEMALLOC;
644 reclaim_state.reclaimed_slab = 0;
645 p->reclaim_state = &reclaim_state;
647 try_to_free_pages(zones, gfp_mask, order);
649 p->reclaim_state = NULL;
650 p->flags &= ~PF_MEMALLOC;
652 /* go through the zonelist yet one more time */
653 for (i = 0; zones[i] != NULL; i++) {
654 struct zone *z = zones[i];
656 min = (1UL << order) + z->protection[alloc_type];
658 if (z->free_pages >= min ||
659 (!wait && z->free_pages >= z->pages_high)) {
660 page = buffered_rmqueue(z, order, gfp_mask);
667 * Don't let big-order allocations loop unless the caller explicitly
668 * requests that. Wait for some write requests to complete then retry.
670 * In this implementation, __GFP_REPEAT means __GFP_NOFAIL, but that
671 * may not be true in other implementations.
674 if (!(gfp_mask & __GFP_NORETRY)) {
675 if ((order <= 3) || (gfp_mask & __GFP_REPEAT))
677 if (gfp_mask & __GFP_NOFAIL)
681 blk_congestion_wait(WRITE, HZ/50);
686 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
687 printk(KERN_WARNING "%s: page allocation failure."
688 " order:%d, mode:0x%x\n",
689 p->comm, order, gfp_mask);
694 kernel_map_pages(page, 1 << order, 1);
698 EXPORT_SYMBOL(__alloc_pages);
701 /* Early boot: Everything is done by one cpu, but the data structures will be
702 * used by all cpus - spread them on all nodes.
704 static __init unsigned long get_boot_pages(unsigned int gfp_mask, unsigned int order)
711 if (i > nodenr + numnodes)
713 if (node_present_pages(i%numnodes)) {
715 /* The node contains memory. Check that there is
716 * memory in the intended zonelist.
718 z = NODE_DATA(i%numnodes)->node_zonelists[gfp_mask & GFP_ZONEMASK].zones;
720 if ( (*z)->free_pages > (1UL<<order))
729 page = alloc_pages_node(i%numnodes, gfp_mask, order);
732 return (unsigned long) page_address(page);
737 * Common helper functions.
739 fastcall unsigned long __get_free_pages(unsigned int gfp_mask, unsigned int order)
744 if (unlikely(system_state == SYSTEM_BOOTING))
745 return get_boot_pages(gfp_mask, order);
747 page = alloc_pages(gfp_mask, order);
750 return (unsigned long) page_address(page);
753 EXPORT_SYMBOL(__get_free_pages);
755 fastcall unsigned long get_zeroed_page(unsigned int gfp_mask)
760 * get_zeroed_page() returns a 32-bit address, which cannot represent
763 BUG_ON(gfp_mask & __GFP_HIGHMEM);
765 page = alloc_pages(gfp_mask, 0);
767 void *address = page_address(page);
769 return (unsigned long) address;
774 EXPORT_SYMBOL(get_zeroed_page);
776 void __pagevec_free(struct pagevec *pvec)
778 int i = pagevec_count(pvec);
781 free_hot_cold_page(pvec->pages[i], pvec->cold);
784 fastcall void __free_pages(struct page *page, unsigned int order)
786 if (!PageReserved(page) && put_page_testzero(page)) {
790 __free_pages_ok(page, order);
794 EXPORT_SYMBOL(__free_pages);
796 fastcall void free_pages(unsigned long addr, unsigned int order)
799 BUG_ON(!virt_addr_valid(addr));
800 __free_pages(virt_to_page(addr), order);
804 EXPORT_SYMBOL(free_pages);
807 * Total amount of free (allocatable) RAM:
809 unsigned int nr_free_pages(void)
811 unsigned int sum = 0;
815 sum += zone->free_pages;
820 EXPORT_SYMBOL(nr_free_pages);
822 unsigned int nr_used_zone_pages(void)
824 unsigned int pages = 0;
828 pages += zone->nr_active + zone->nr_inactive;
834 unsigned int nr_free_pages_pgdat(pg_data_t *pgdat)
836 unsigned int i, sum = 0;
838 for (i = 0; i < MAX_NR_ZONES; i++)
839 sum += pgdat->node_zones[i].free_pages;
845 static unsigned int nr_free_zone_pages(int offset)
848 unsigned int sum = 0;
850 for_each_pgdat(pgdat) {
851 struct zonelist *zonelist = pgdat->node_zonelists + offset;
852 struct zone **zonep = zonelist->zones;
855 for (zone = *zonep++; zone; zone = *zonep++) {
856 unsigned long size = zone->present_pages;
857 unsigned long high = zone->pages_high;
867 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
869 unsigned int nr_free_buffer_pages(void)
871 return nr_free_zone_pages(GFP_USER & GFP_ZONEMASK);
875 * Amount of free RAM allocatable within all zones
877 unsigned int nr_free_pagecache_pages(void)
879 return nr_free_zone_pages(GFP_HIGHUSER & GFP_ZONEMASK);
882 #ifdef CONFIG_HIGHMEM
883 unsigned int nr_free_highpages (void)
886 unsigned int pages = 0;
888 for_each_pgdat(pgdat)
889 pages += pgdat->node_zones[ZONE_HIGHMEM].free_pages;
896 static void show_node(struct zone *zone)
898 printk("Node %d ", zone->zone_pgdat->node_id);
901 #define show_node(zone) do { } while (0)
905 * Accumulate the page_state information across all CPUs.
906 * The result is unavoidably approximate - it can change
907 * during and after execution of this function.
909 DEFINE_PER_CPU(struct page_state, page_states) = {0};
910 EXPORT_PER_CPU_SYMBOL(page_states);
912 atomic_t nr_pagecache = ATOMIC_INIT(0);
913 EXPORT_SYMBOL(nr_pagecache);
915 DEFINE_PER_CPU(long, nr_pagecache_local) = 0;
918 void __get_page_state(struct page_state *ret, int nr)
922 memset(ret, 0, sizeof(*ret));
923 while (cpu < NR_CPUS) {
924 unsigned long *in, *out, off;
926 if (!cpu_possible(cpu)) {
931 in = (unsigned long *)&per_cpu(page_states, cpu);
933 if (cpu < NR_CPUS && cpu_possible(cpu))
934 prefetch(&per_cpu(page_states, cpu));
935 out = (unsigned long *)ret;
936 for (off = 0; off < nr; off++)
941 void get_page_state(struct page_state *ret)
945 nr = offsetof(struct page_state, GET_PAGE_STATE_LAST);
946 nr /= sizeof(unsigned long);
948 __get_page_state(ret, nr + 1);
951 void get_full_page_state(struct page_state *ret)
953 __get_page_state(ret, sizeof(*ret) / sizeof(unsigned long));
956 void get_zone_counts(unsigned long *active,
957 unsigned long *inactive, unsigned long *free)
964 for_each_zone(zone) {
965 *active += zone->nr_active;
966 *inactive += zone->nr_inactive;
967 *free += zone->free_pages;
971 void si_meminfo(struct sysinfo *val)
973 val->totalram = totalram_pages;
975 val->freeram = nr_free_pages();
976 val->bufferram = nr_blockdev_pages();
977 #ifdef CONFIG_HIGHMEM
978 val->totalhigh = totalhigh_pages;
979 val->freehigh = nr_free_highpages();
984 val->mem_unit = PAGE_SIZE;
987 EXPORT_SYMBOL(si_meminfo);
990 void si_meminfo_node(struct sysinfo *val, int nid)
992 pg_data_t *pgdat = NODE_DATA(nid);
994 val->totalram = pgdat->node_present_pages;
995 val->freeram = nr_free_pages_pgdat(pgdat);
996 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
997 val->freehigh = pgdat->node_zones[ZONE_HIGHMEM].free_pages;
998 val->mem_unit = PAGE_SIZE;
1002 #define K(x) ((x) << (PAGE_SHIFT-10))
1005 * Show free area list (used inside shift_scroll-lock stuff)
1006 * We also calculate the percentage fragmentation. We do this by counting the
1007 * memory on each free list with the exception of the first item on the list.
1009 void show_free_areas(void)
1011 struct page_state ps;
1012 int cpu, temperature;
1013 unsigned long active;
1014 unsigned long inactive;
1018 for_each_zone(zone) {
1020 printk("%s per-cpu:", zone->name);
1022 if (!zone->present_pages) {
1028 for (cpu = 0; cpu < NR_CPUS; ++cpu) {
1029 struct per_cpu_pageset *pageset;
1031 if (!cpu_possible(cpu))
1034 pageset = zone->pageset + cpu;
1036 for (temperature = 0; temperature < 2; temperature++)
1037 printk("cpu %d %s: low %d, high %d, batch %d\n",
1039 temperature ? "cold" : "hot",
1040 pageset->pcp[temperature].low,
1041 pageset->pcp[temperature].high,
1042 pageset->pcp[temperature].batch);
1046 get_page_state(&ps);
1047 get_zone_counts(&active, &inactive, &free);
1049 printk("\nFree pages: %11ukB (%ukB HighMem)\n",
1051 K(nr_free_highpages()));
1053 printk("Active:%lu inactive:%lu dirty:%lu writeback:%lu "
1054 "unstable:%lu free:%u slab:%lu mapped:%lu pagetables:%lu\n",
1063 ps.nr_page_table_pages);
1065 for_each_zone(zone) {
1079 K(zone->free_pages),
1082 K(zone->pages_high),
1084 K(zone->nr_inactive),
1085 K(zone->present_pages)
1087 printk("protections[]:");
1088 for (i = 0; i < MAX_NR_ZONES; i++)
1089 printk(" %lu", zone->protection[i]);
1093 for_each_zone(zone) {
1094 struct list_head *elem;
1095 unsigned long nr, flags, order, total = 0;
1098 printk("%s: ", zone->name);
1099 if (!zone->present_pages) {
1104 spin_lock_irqsave(&zone->lock, flags);
1105 for (order = 0; order < MAX_ORDER; order++) {
1107 list_for_each(elem, &zone->free_area[order].free_list)
1109 total += nr << order;
1110 printk("%lu*%lukB ", nr, K(1UL) << order);
1112 spin_unlock_irqrestore(&zone->lock, flags);
1113 printk("= %lukB\n", K(total));
1116 show_swap_cache_info();
1120 * Builds allocation fallback zone lists.
1122 static int __init build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist, int j, int k)
1129 zone = pgdat->node_zones + ZONE_HIGHMEM;
1130 if (zone->present_pages) {
1131 #ifndef CONFIG_HIGHMEM
1134 zonelist->zones[j++] = zone;
1137 zone = pgdat->node_zones + ZONE_NORMAL;
1138 if (zone->present_pages)
1139 zonelist->zones[j++] = zone;
1141 zone = pgdat->node_zones + ZONE_DMA;
1142 if (zone->present_pages)
1143 zonelist->zones[j++] = zone;
1150 #define MAX_NODE_LOAD (numnodes)
1151 static int __initdata node_load[MAX_NUMNODES];
1153 * find_next_best_node - find the next node that should appear in a given
1154 * node's fallback list
1155 * @node: node whose fallback list we're appending
1156 * @used_node_mask: pointer to the bitmap of already used nodes
1158 * We use a number of factors to determine which is the next node that should
1159 * appear on a given node's fallback list. The node should not have appeared
1160 * already in @node's fallback list, and it should be the next closest node
1161 * according to the distance array (which contains arbitrary distance values
1162 * from each node to each node in the system), and should also prefer nodes
1163 * with no CPUs, since presumably they'll have very little allocation pressure
1164 * on them otherwise.
1165 * It returns -1 if no node is found.
1167 static int __init find_next_best_node(int node, void *used_node_mask)
1170 int min_val = INT_MAX;
1173 for (i = 0; i < numnodes; i++) {
1176 /* Start from local node */
1177 n = (node+i)%numnodes;
1179 /* Don't want a node to appear more than once */
1180 if (test_bit(n, used_node_mask))
1183 /* Use the distance array to find the distance */
1184 val = node_distance(node, n);
1186 /* Give preference to headless and unused nodes */
1187 tmp = node_to_cpumask(n);
1188 if (!cpus_empty(tmp))
1189 val += PENALTY_FOR_NODE_WITH_CPUS;
1191 /* Slight preference for less loaded node */
1192 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
1193 val += node_load[n];
1195 if (val < min_val) {
1202 set_bit(best_node, used_node_mask);
1207 static void __init build_zonelists(pg_data_t *pgdat)
1209 int i, j, k, node, local_node;
1210 int prev_node, load;
1211 struct zonelist *zonelist;
1212 DECLARE_BITMAP(used_mask, MAX_NUMNODES);
1214 /* initialize zonelists */
1215 for (i = 0; i < MAX_NR_ZONES; i++) {
1216 zonelist = pgdat->node_zonelists + i;
1217 memset(zonelist, 0, sizeof(*zonelist));
1218 zonelist->zones[0] = NULL;
1221 /* NUMA-aware ordering of nodes */
1222 local_node = pgdat->node_id;
1224 prev_node = local_node;
1225 bitmap_zero(used_mask, MAX_NUMNODES);
1226 while ((node = find_next_best_node(local_node, used_mask)) >= 0) {
1228 * We don't want to pressure a particular node.
1229 * So adding penalty to the first node in same
1230 * distance group to make it round-robin.
1232 if (node_distance(local_node, node) !=
1233 node_distance(local_node, prev_node))
1234 node_load[node] += load;
1237 for (i = 0; i < MAX_NR_ZONES; i++) {
1238 zonelist = pgdat->node_zonelists + i;
1239 for (j = 0; zonelist->zones[j] != NULL; j++);
1242 if (i & __GFP_HIGHMEM)
1247 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1248 zonelist->zones[j] = NULL;
1253 #else /* CONFIG_NUMA */
1255 static void __init build_zonelists(pg_data_t *pgdat)
1257 int i, j, k, node, local_node;
1259 local_node = pgdat->node_id;
1260 for (i = 0; i < MAX_NR_ZONES; i++) {
1261 struct zonelist *zonelist;
1263 zonelist = pgdat->node_zonelists + i;
1264 memset(zonelist, 0, sizeof(*zonelist));
1268 if (i & __GFP_HIGHMEM)
1273 j = build_zonelists_node(pgdat, zonelist, j, k);
1275 * Now we build the zonelist so that it contains the zones
1276 * of all the other nodes.
1277 * We don't want to pressure a particular node, so when
1278 * building the zones for node N, we make sure that the
1279 * zones coming right after the local ones are those from
1280 * node N+1 (modulo N)
1282 for (node = local_node + 1; node < numnodes; node++)
1283 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1284 for (node = 0; node < local_node; node++)
1285 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1287 zonelist->zones[j++] = NULL;
1291 #endif /* CONFIG_NUMA */
1293 void __init build_all_zonelists(void)
1297 for(i = 0 ; i < numnodes ; i++)
1298 build_zonelists(NODE_DATA(i));
1299 printk("Built %i zonelists\n", numnodes);
1303 * Helper functions to size the waitqueue hash table.
1304 * Essentially these want to choose hash table sizes sufficiently
1305 * large so that collisions trying to wait on pages are rare.
1306 * But in fact, the number of active page waitqueues on typical
1307 * systems is ridiculously low, less than 200. So this is even
1308 * conservative, even though it seems large.
1310 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
1311 * waitqueues, i.e. the size of the waitq table given the number of pages.
1313 #define PAGES_PER_WAITQUEUE 256
1315 static inline unsigned long wait_table_size(unsigned long pages)
1317 unsigned long size = 1;
1319 pages /= PAGES_PER_WAITQUEUE;
1321 while (size < pages)
1325 * Once we have dozens or even hundreds of threads sleeping
1326 * on IO we've got bigger problems than wait queue collision.
1327 * Limit the size of the wait table to a reasonable size.
1329 size = min(size, 4096UL);
1331 return max(size, 4UL);
1335 * This is an integer logarithm so that shifts can be used later
1336 * to extract the more random high bits from the multiplicative
1337 * hash function before the remainder is taken.
1339 static inline unsigned long wait_table_bits(unsigned long size)
1344 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
1346 static void __init calculate_zone_totalpages(struct pglist_data *pgdat,
1347 unsigned long *zones_size, unsigned long *zholes_size)
1349 unsigned long realtotalpages, totalpages = 0;
1352 for (i = 0; i < MAX_NR_ZONES; i++)
1353 totalpages += zones_size[i];
1354 pgdat->node_spanned_pages = totalpages;
1356 realtotalpages = totalpages;
1358 for (i = 0; i < MAX_NR_ZONES; i++)
1359 realtotalpages -= zholes_size[i];
1360 pgdat->node_present_pages = realtotalpages;
1361 printk("On node %d totalpages: %lu\n", pgdat->node_id, realtotalpages);
1366 * Initially all pages are reserved - free ones are freed
1367 * up by free_all_bootmem() once the early boot process is
1368 * done. Non-atomic initialization, single-pass.
1370 void __init memmap_init_zone(struct page *start, unsigned long size, int nid,
1371 unsigned long zone, unsigned long start_pfn)
1375 for (page = start; page < (start + size); page++) {
1376 set_page_zone(page, NODEZONE(nid, zone));
1377 set_page_count(page, 0);
1378 SetPageReserved(page);
1379 INIT_LIST_HEAD(&page->lru);
1380 #ifdef WANT_PAGE_VIRTUAL
1381 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1382 if (zone != ZONE_HIGHMEM)
1383 set_page_address(page, __va(start_pfn << PAGE_SHIFT));
1389 #ifndef __HAVE_ARCH_MEMMAP_INIT
1390 #define memmap_init(start, size, nid, zone, start_pfn) \
1391 memmap_init_zone((start), (size), (nid), (zone), (start_pfn))
1395 * Set up the zone data structures:
1396 * - mark all pages reserved
1397 * - mark all memory queues empty
1398 * - clear the memory bitmaps
1400 static void __init free_area_init_core(struct pglist_data *pgdat,
1401 unsigned long *zones_size, unsigned long *zholes_size)
1404 const unsigned long zone_required_alignment = 1UL << (MAX_ORDER-1);
1405 int cpu, nid = pgdat->node_id;
1406 struct page *lmem_map = pgdat->node_mem_map;
1407 unsigned long zone_start_pfn = pgdat->node_start_pfn;
1409 pgdat->nr_zones = 0;
1410 init_waitqueue_head(&pgdat->kswapd_wait);
1412 for (j = 0; j < MAX_NR_ZONES; j++) {
1413 struct zone *zone = pgdat->node_zones + j;
1414 unsigned long size, realsize;
1415 unsigned long batch;
1417 zone_table[NODEZONE(nid, j)] = zone;
1418 realsize = size = zones_size[j];
1420 realsize -= zholes_size[j];
1422 zone->spanned_pages = size;
1423 zone->present_pages = realsize;
1424 zone->name = zone_names[j];
1425 spin_lock_init(&zone->lock);
1426 spin_lock_init(&zone->lru_lock);
1427 zone->zone_pgdat = pgdat;
1428 zone->free_pages = 0;
1430 zone->temp_priority = zone->prev_priority = DEF_PRIORITY;
1433 * The per-cpu-pages pools are set to around 1000th of the
1434 * size of the zone. But no more than 1/4 of a meg - there's
1435 * no point in going beyond the size of L2 cache.
1437 * OK, so we don't know how big the cache is. So guess.
1439 batch = zone->present_pages / 1024;
1440 if (batch * PAGE_SIZE > 256 * 1024)
1441 batch = (256 * 1024) / PAGE_SIZE;
1442 batch /= 4; /* We effectively *= 4 below */
1446 for (cpu = 0; cpu < NR_CPUS; cpu++) {
1447 struct per_cpu_pages *pcp;
1449 pcp = &zone->pageset[cpu].pcp[0]; /* hot */
1451 pcp->low = 2 * batch;
1452 pcp->high = 6 * batch;
1453 pcp->batch = 1 * batch;
1454 INIT_LIST_HEAD(&pcp->list);
1456 pcp = &zone->pageset[cpu].pcp[1]; /* cold */
1459 pcp->high = 2 * batch;
1460 pcp->batch = 1 * batch;
1461 INIT_LIST_HEAD(&pcp->list);
1463 printk(" %s zone: %lu pages, LIFO batch:%lu\n",
1464 zone_names[j], realsize, batch);
1465 INIT_LIST_HEAD(&zone->active_list);
1466 INIT_LIST_HEAD(&zone->inactive_list);
1467 atomic_set(&zone->nr_scan_active, 0);
1468 atomic_set(&zone->nr_scan_inactive, 0);
1469 zone->nr_active = 0;
1470 zone->nr_inactive = 0;
1475 * The per-page waitqueue mechanism uses hashed waitqueues
1478 zone->wait_table_size = wait_table_size(size);
1479 zone->wait_table_bits =
1480 wait_table_bits(zone->wait_table_size);
1481 zone->wait_table = (wait_queue_head_t *)
1482 alloc_bootmem_node(pgdat, zone->wait_table_size
1483 * sizeof(wait_queue_head_t));
1485 for(i = 0; i < zone->wait_table_size; ++i)
1486 init_waitqueue_head(zone->wait_table + i);
1488 pgdat->nr_zones = j+1;
1490 zone->zone_mem_map = lmem_map;
1491 zone->zone_start_pfn = zone_start_pfn;
1493 if ((zone_start_pfn) & (zone_required_alignment-1))
1494 printk("BUG: wrong zone alignment, it will crash\n");
1496 memmap_init(lmem_map, size, nid, j, zone_start_pfn);
1498 zone_start_pfn += size;
1501 for (i = 0; ; i++) {
1502 unsigned long bitmap_size;
1504 INIT_LIST_HEAD(&zone->free_area[i].free_list);
1505 if (i == MAX_ORDER-1) {
1506 zone->free_area[i].map = NULL;
1511 * Page buddy system uses "index >> (i+1)",
1512 * where "index" is at most "size-1".
1514 * The extra "+3" is to round down to byte
1515 * size (8 bits per byte assumption). Thus
1516 * we get "(size-1) >> (i+4)" as the last byte
1519 * The "+1" is because we want to round the
1520 * byte allocation up rather than down. So
1521 * we should have had a "+7" before we shifted
1522 * down by three. Also, we have to add one as
1523 * we actually _use_ the last bit (it's [0,n]
1524 * inclusive, not [0,n[).
1526 * So we actually had +7+1 before we shift
1527 * down by 3. But (n+8) >> 3 == (n >> 3) + 1
1528 * (modulo overflows, which we do not have).
1530 * Finally, we LONG_ALIGN because all bitmap
1531 * operations are on longs.
1533 bitmap_size = (size-1) >> (i+4);
1534 bitmap_size = LONG_ALIGN(bitmap_size+1);
1535 zone->free_area[i].map =
1536 (unsigned long *) alloc_bootmem_node(pgdat, bitmap_size);
1541 void __init free_area_init_node(int nid, struct pglist_data *pgdat,
1542 struct page *node_mem_map, unsigned long *zones_size,
1543 unsigned long node_start_pfn, unsigned long *zholes_size)
1547 pgdat->node_id = nid;
1548 pgdat->node_start_pfn = node_start_pfn;
1549 calculate_zone_totalpages(pgdat, zones_size, zholes_size);
1550 if (!node_mem_map) {
1551 size = (pgdat->node_spanned_pages + 1) * sizeof(struct page);
1552 node_mem_map = alloc_bootmem_node(pgdat, size);
1554 pgdat->node_mem_map = node_mem_map;
1556 free_area_init_core(pgdat, zones_size, zholes_size);
1559 #ifndef CONFIG_DISCONTIGMEM
1560 static bootmem_data_t contig_bootmem_data;
1561 struct pglist_data contig_page_data = { .bdata = &contig_bootmem_data };
1563 EXPORT_SYMBOL(contig_page_data);
1565 void __init free_area_init(unsigned long *zones_size)
1567 free_area_init_node(0, &contig_page_data, NULL, zones_size,
1568 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
1569 mem_map = contig_page_data.node_mem_map;
1573 #ifdef CONFIG_PROC_FS
1575 #include <linux/seq_file.h>
1577 static void *frag_start(struct seq_file *m, loff_t *pos)
1582 for (pgdat = pgdat_list; pgdat && node; pgdat = pgdat->pgdat_next)
1588 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
1590 pg_data_t *pgdat = (pg_data_t *)arg;
1593 return pgdat->pgdat_next;
1596 static void frag_stop(struct seq_file *m, void *arg)
1601 * This walks the freelist for each zone. Whilst this is slow, I'd rather
1602 * be slow here than slow down the fast path by keeping stats - mjbligh
1604 static int frag_show(struct seq_file *m, void *arg)
1606 pg_data_t *pgdat = (pg_data_t *)arg;
1608 struct zone *node_zones = pgdat->node_zones;
1609 unsigned long flags;
1612 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
1613 if (!zone->present_pages)
1616 spin_lock_irqsave(&zone->lock, flags);
1617 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1618 for (order = 0; order < MAX_ORDER; ++order) {
1619 unsigned long nr_bufs = 0;
1620 struct list_head *elem;
1622 list_for_each(elem, &(zone->free_area[order].free_list))
1624 seq_printf(m, "%6lu ", nr_bufs);
1626 spin_unlock_irqrestore(&zone->lock, flags);
1632 struct seq_operations fragmentation_op = {
1633 .start = frag_start,
1639 static char *vmstat_text[] = {
1643 "nr_page_table_pages",
1668 "pgscan_kswapd_high",
1669 "pgscan_kswapd_normal",
1671 "pgscan_kswapd_dma",
1672 "pgscan_direct_high",
1673 "pgscan_direct_normal",
1674 "pgscan_direct_dma",
1679 "kswapd_inodesteal",
1686 static void *vmstat_start(struct seq_file *m, loff_t *pos)
1688 struct page_state *ps;
1690 if (*pos >= ARRAY_SIZE(vmstat_text))
1693 ps = kmalloc(sizeof(*ps), GFP_KERNEL);
1696 return ERR_PTR(-ENOMEM);
1697 get_full_page_state(ps);
1698 ps->pgpgin /= 2; /* sectors -> kbytes */
1700 return (unsigned long *)ps + *pos;
1703 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
1706 if (*pos >= ARRAY_SIZE(vmstat_text))
1708 return (unsigned long *)m->private + *pos;
1711 static int vmstat_show(struct seq_file *m, void *arg)
1713 unsigned long *l = arg;
1714 unsigned long off = l - (unsigned long *)m->private;
1716 seq_printf(m, "%s %lu\n", vmstat_text[off], *l);
1720 static void vmstat_stop(struct seq_file *m, void *arg)
1726 struct seq_operations vmstat_op = {
1727 .start = vmstat_start,
1728 .next = vmstat_next,
1729 .stop = vmstat_stop,
1730 .show = vmstat_show,
1733 #endif /* CONFIG_PROC_FS */
1735 #ifdef CONFIG_HOTPLUG_CPU
1736 static int page_alloc_cpu_notify(struct notifier_block *self,
1737 unsigned long action, void *hcpu)
1739 int cpu = (unsigned long)hcpu;
1742 if (action == CPU_DEAD) {
1743 /* Drain local pagecache count. */
1744 count = &per_cpu(nr_pagecache_local, cpu);
1745 atomic_add(*count, &nr_pagecache);
1747 local_irq_disable();
1753 #endif /* CONFIG_HOTPLUG_CPU */
1755 void __init page_alloc_init(void)
1757 hotcpu_notifier(page_alloc_cpu_notify, 0);
1760 static unsigned long higherzone_val(struct zone *z, int max_zone,
1763 int z_idx = zone_idx(z);
1764 struct zone *higherzone;
1765 unsigned long pages;
1767 /* there is no higher zone to get a contribution from */
1768 if (z_idx == MAX_NR_ZONES-1)
1771 higherzone = &z->zone_pgdat->node_zones[z_idx+1];
1773 /* We always start with the higher zone's protection value */
1774 pages = higherzone->protection[alloc_type];
1777 * We get a lower-zone-protection contribution only if there are
1778 * pages in the higher zone and if we're not the highest zone
1779 * in the current zonelist. e.g., never happens for GFP_DMA. Happens
1780 * only for ZONE_DMA in a GFP_KERNEL allocation and happens for ZONE_DMA
1781 * and ZONE_NORMAL for a GFP_HIGHMEM allocation.
1783 if (higherzone->present_pages && z_idx < alloc_type)
1784 pages += higherzone->pages_low * sysctl_lower_zone_protection;
1790 * setup_per_zone_protection - called whenver min_free_kbytes or
1791 * sysctl_lower_zone_protection changes. Ensures that each zone
1792 * has a correct pages_protected value, so an adequate number of
1793 * pages are left in the zone after a successful __alloc_pages().
1795 * This algorithm is way confusing. I tries to keep the same behavior
1796 * as we had with the incremental min iterative algorithm.
1798 static void setup_per_zone_protection(void)
1800 struct pglist_data *pgdat;
1801 struct zone *zones, *zone;
1805 for_each_pgdat(pgdat) {
1806 zones = pgdat->node_zones;
1808 for (i = 0, max_zone = 0; i < MAX_NR_ZONES; i++)
1809 if (zones[i].present_pages)
1813 * For each of the different allocation types:
1814 * GFP_DMA -> GFP_KERNEL -> GFP_HIGHMEM
1816 for (i = 0; i < MAX_NR_ZONES; i++) {
1818 * For each of the zones:
1819 * ZONE_HIGHMEM -> ZONE_NORMAL -> ZONE_DMA
1821 for (j = MAX_NR_ZONES-1; j >= 0; j--) {
1825 * We never protect zones that don't have memory
1826 * in them (j>max_zone) or zones that aren't in
1827 * the zonelists for a certain type of
1828 * allocation (j>i). We have to assign these to
1829 * zero because the lower zones take
1830 * contributions from the higher zones.
1832 if (j > max_zone || j > i) {
1833 zone->protection[i] = 0;
1837 * The contribution of the next higher zone
1839 zone->protection[i] = higherzone_val(zone,
1841 zone->protection[i] += zone->pages_low;
1848 * setup_per_zone_pages_min - called when min_free_kbytes changes. Ensures
1849 * that the pages_{min,low,high} values for each zone are set correctly
1850 * with respect to min_free_kbytes.
1852 static void setup_per_zone_pages_min(void)
1854 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
1855 unsigned long lowmem_pages = 0;
1857 unsigned long flags;
1859 /* Calculate total number of !ZONE_HIGHMEM pages */
1860 for_each_zone(zone) {
1861 if (!is_highmem(zone))
1862 lowmem_pages += zone->present_pages;
1865 for_each_zone(zone) {
1866 spin_lock_irqsave(&zone->lru_lock, flags);
1867 if (is_highmem(zone)) {
1869 * Often, highmem doesn't need to reserve any pages.
1870 * But the pages_min/low/high values are also used for
1871 * batching up page reclaim activity so we need a
1872 * decent value here.
1876 min_pages = zone->present_pages / 1024;
1877 if (min_pages < SWAP_CLUSTER_MAX)
1878 min_pages = SWAP_CLUSTER_MAX;
1879 if (min_pages > 128)
1881 zone->pages_min = min_pages;
1883 /* if it's a lowmem zone, reserve a number of pages
1884 * proportionate to the zone's size.
1886 zone->pages_min = (pages_min * zone->present_pages) /
1890 zone->pages_low = zone->pages_min * 2;
1891 zone->pages_high = zone->pages_min * 3;
1892 spin_unlock_irqrestore(&zone->lru_lock, flags);
1897 * Initialise min_free_kbytes.
1899 * For small machines we want it small (128k min). For large machines
1900 * we want it large (16MB max). But it is not linear, because network
1901 * bandwidth does not increase linearly with machine size. We use
1903 * min_free_kbytes = sqrt(lowmem_kbytes)
1919 static int __init init_per_zone_pages_min(void)
1921 unsigned long lowmem_kbytes;
1923 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
1925 min_free_kbytes = int_sqrt(lowmem_kbytes);
1926 if (min_free_kbytes < 128)
1927 min_free_kbytes = 128;
1928 if (min_free_kbytes > 16384)
1929 min_free_kbytes = 16384;
1930 setup_per_zone_pages_min();
1931 setup_per_zone_protection();
1934 module_init(init_per_zone_pages_min)
1937 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
1938 * that we can call two helper functions whenever min_free_kbytes
1941 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
1942 struct file *file, void __user *buffer, size_t *length)
1944 proc_dointvec(table, write, file, buffer, length);
1945 setup_per_zone_pages_min();
1946 setup_per_zone_protection();
1951 * lower_zone_protection_sysctl_handler - just a wrapper around
1952 * proc_dointvec() so that we can call setup_per_zone_protection()
1953 * whenever sysctl_lower_zone_protection changes.
1955 int lower_zone_protection_sysctl_handler(ctl_table *table, int write,
1956 struct file *file, void __user *buffer, size_t *length)
1958 proc_dointvec_minmax(table, write, file, buffer, length);
1959 setup_per_zone_protection();