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>
34 #include <linux/vs_base.h>
35 #include <linux/vs_limit.h>
36 #include <linux/ckrm_mem_inline.h>
37 #include <linux/nodemask.h>
38 #include <linux/vs_limit.h>
40 #include <asm/tlbflush.h>
42 nodemask_t node_online_map = NODE_MASK_NONE;
43 nodemask_t node_possible_map = NODE_MASK_ALL;
44 struct pglist_data *pgdat_list;
45 unsigned long totalram_pages;
46 unsigned long totalhigh_pages;
49 int sysctl_lower_zone_protection = 0;
51 EXPORT_SYMBOL(totalram_pages);
52 EXPORT_SYMBOL(nr_swap_pages);
54 #ifdef CONFIG_CRASH_DUMP_MODULE
55 /* This symbol has to be exported to use 'for_each_pgdat' macro by modules. */
56 EXPORT_SYMBOL(pgdat_list);
60 * Used by page_zone() to look up the address of the struct zone whose
61 * id is encoded in the upper bits of page->flags
63 struct zone *zone_table[1 << (ZONES_SHIFT + NODES_SHIFT)];
64 EXPORT_SYMBOL(zone_table);
66 static char *zone_names[MAX_NR_ZONES] = { "DMA", "Normal", "HighMem" };
67 int min_free_kbytes = 1024;
69 unsigned long __initdata nr_kernel_pages;
70 unsigned long __initdata nr_all_pages;
73 * Temporary debugging check for pages not lying within a given zone.
75 static int bad_range(struct zone *zone, struct page *page)
77 if (page_to_pfn(page) >= zone->zone_start_pfn + zone->spanned_pages)
79 if (page_to_pfn(page) < zone->zone_start_pfn)
81 if (zone != page_zone(page))
86 static void bad_page(const char *function, struct page *page)
88 printk(KERN_EMERG "Bad page state at %s (in process '%s', page %p)\n",
89 function, current->comm, page);
90 printk(KERN_EMERG "flags:0x%0*lx mapping:%p mapcount:%d count:%d (%s)\n",
91 (int)(2*sizeof(page_flags_t)), (unsigned long)page->flags,
92 page->mapping, page_mapcount(page), page_count(page), print_tainted());
93 printk(KERN_EMERG "Backtrace:\n");
95 printk(KERN_EMERG "Trying to fix it up, but a reboot is needed\n");
96 page->flags &= ~(1 << PG_private |
103 set_page_count(page, 0);
104 reset_page_mapcount(page);
105 page->mapping = NULL;
106 tainted |= TAINT_BAD_PAGE;
109 #if !defined(CONFIG_HUGETLB_PAGE) && !defined(CONFIG_CRASH_DUMP) \
110 && !defined(CONFIG_CRASH_DUMP_MODULE)
111 #define prep_compound_page(page, order) do { } while (0)
112 #define destroy_compound_page(page, order) do { } while (0)
115 * Higher-order pages are called "compound pages". They are structured thusly:
117 * The first PAGE_SIZE page is called the "head page".
119 * The remaining PAGE_SIZE pages are called "tail pages".
121 * All pages have PG_compound set. All pages have their ->private pointing at
122 * the head page (even the head page has this).
124 * The first tail page's ->mapping, if non-zero, holds the address of the
125 * compound page's put_page() function.
127 * The order of the allocation is stored in the first tail page's ->index
128 * This is only for debug at present. This usage means that zero-order pages
129 * may not be compound.
131 static void prep_compound_page(struct page *page, unsigned long order)
134 int nr_pages = 1 << order;
136 page[1].mapping = NULL;
137 page[1].index = order;
138 for (i = 0; i < nr_pages; i++) {
139 struct page *p = page + i;
142 p->private = (unsigned long)page;
146 static void destroy_compound_page(struct page *page, unsigned long order)
149 int nr_pages = 1 << order;
151 if (!PageCompound(page))
154 if (page[1].index != order)
155 bad_page(__FUNCTION__, page);
157 for (i = 0; i < nr_pages; i++) {
158 struct page *p = page + i;
160 if (!PageCompound(p))
161 bad_page(__FUNCTION__, page);
162 if (p->private != (unsigned long)page)
163 bad_page(__FUNCTION__, page);
164 ClearPageCompound(p);
167 #endif /* CONFIG_HUGETLB_PAGE */
170 * Freeing function for a buddy system allocator.
172 * The concept of a buddy system is to maintain direct-mapped table
173 * (containing bit values) for memory blocks of various "orders".
174 * The bottom level table contains the map for the smallest allocatable
175 * units of memory (here, pages), and each level above it describes
176 * pairs of units from the levels below, hence, "buddies".
177 * At a high level, all that happens here is marking the table entry
178 * at the bottom level available, and propagating the changes upward
179 * as necessary, plus some accounting needed to play nicely with other
180 * parts of the VM system.
181 * At each level, we keep one bit for each pair of blocks, which
182 * is set to 1 iff only one of the pair is allocated. So when we
183 * are allocating or freeing one, we can derive the state of the
184 * other. That is, if we allocate a small block, and both were
185 * free, the remainder of the region must be split into blocks.
186 * If a block is freed, and its buddy is also free, then this
187 * triggers coalescing into a block of larger size.
192 static inline void __free_pages_bulk (struct page *page, struct page *base,
193 struct zone *zone, struct free_area *area, unsigned int order)
195 unsigned long page_idx, index, mask;
198 destroy_compound_page(page, order);
199 mask = (~0UL) << order;
200 page_idx = page - base;
201 if (page_idx & ~mask)
203 index = page_idx >> (1 + order);
205 zone->free_pages += 1 << order;
206 while (order < MAX_ORDER-1) {
207 struct page *buddy1, *buddy2;
209 BUG_ON(area >= zone->free_area + MAX_ORDER);
210 if (!__test_and_change_bit(index, area->map))
212 * the buddy page is still allocated.
216 /* Move the buddy up one level. */
217 buddy1 = base + (page_idx ^ (1 << order));
218 buddy2 = base + page_idx;
219 BUG_ON(bad_range(zone, buddy1));
220 BUG_ON(bad_range(zone, buddy2));
221 list_del(&buddy1->lru);
228 list_add(&(base + page_idx)->lru, &area->free_list);
231 static inline void free_pages_check(const char *function, struct page *page)
233 if ( page_mapped(page) ||
234 page->mapping != NULL ||
235 page_count(page) != 0 ||
244 1 << PG_writeback )))
245 bad_page(function, page);
247 ClearPageDirty(page);
251 * Frees a list of pages.
252 * Assumes all pages on list are in same zone, and of same order.
253 * count is the number of pages to free, or 0 for all on the list.
255 * If the zone was previously in an "all pages pinned" state then look to
256 * see if this freeing clears that state.
258 * And clear the zone's pages_scanned counter, to hold off the "all pages are
259 * pinned" detection logic.
262 free_pages_bulk(struct zone *zone, int count,
263 struct list_head *list, unsigned int order)
266 struct free_area *area;
267 struct page *base, *page = NULL;
270 base = zone->zone_mem_map;
271 area = zone->free_area + order;
272 spin_lock_irqsave(&zone->lock, flags);
273 zone->all_unreclaimable = 0;
274 zone->pages_scanned = 0;
275 while (!list_empty(list) && count--) {
276 page = list_entry(list->prev, struct page, lru);
277 /* have to delete it as __free_pages_bulk list manipulates */
278 list_del(&page->lru);
279 __free_pages_bulk(page, base, zone, area, order);
280 ckrm_clear_page_class(page);
283 spin_unlock_irqrestore(&zone->lock, flags);
287 void __free_pages_ok(struct page *page, unsigned int order)
292 arch_free_page(page, order);
294 mod_page_state(pgfree, 1 << order);
295 for (i = 0 ; i < (1 << order) ; ++i)
296 free_pages_check(__FUNCTION__, page + i);
297 list_add(&page->lru, &list);
298 kernel_map_pages(page, 1<<order, 0);
299 free_pages_bulk(page_zone(page), 1, &list, order);
302 #define MARK_USED(index, order, area) \
303 __change_bit((index) >> (1+(order)), (area)->map)
306 * The order of subdivision here is critical for the IO subsystem.
307 * Please do not alter this order without good reasons and regression
308 * testing. Specifically, as large blocks of memory are subdivided,
309 * the order in which smaller blocks are delivered depends on the order
310 * they're subdivided in this function. This is the primary factor
311 * influencing the order in which pages are delivered to the IO
312 * subsystem according to empirical testing, and this is also justified
313 * by considering the behavior of a buddy system containing a single
314 * large block of memory acted on by a series of small allocations.
315 * This behavior is a critical factor in sglist merging's success.
319 static inline struct page *
320 expand(struct zone *zone, struct page *page,
321 unsigned long index, int low, int high, struct free_area *area)
323 unsigned long size = 1 << high;
329 BUG_ON(bad_range(zone, &page[size]));
330 list_add(&page[size].lru, &area->free_list);
331 MARK_USED(index + size, high, area);
336 static inline void set_page_refs(struct page *page, int order)
339 set_page_count(page, 1);
344 * We need to reference all the pages for this order, otherwise if
345 * anyone accesses one of the pages with (get/put) it will be freed.
347 for (i = 0; i < (1 << order); i++)
348 set_page_count(page+i, 1);
349 #endif /* CONFIG_MMU */
353 * This page is about to be returned from the page allocator
355 static void prep_new_page(struct page *page, int order)
357 if (page->mapping || page_mapped(page) ||
366 1 << PG_writeback )))
367 bad_page(__FUNCTION__, page);
369 page->flags &= ~(1 << PG_uptodate | 1 << PG_error |
370 1 << PG_referenced | 1 << PG_arch_1 |
371 1 << PG_checked | 1 << PG_mappedtodisk);
373 set_page_refs(page, order);
377 * Do the hard work of removing an element from the buddy allocator.
378 * Call me with the zone->lock already held.
380 static struct page *__rmqueue(struct zone *zone, unsigned int order)
382 struct free_area * area;
383 unsigned int current_order;
387 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
388 area = zone->free_area + current_order;
389 if (list_empty(&area->free_list))
392 page = list_entry(area->free_list.next, struct page, lru);
393 list_del(&page->lru);
394 index = page - zone->zone_mem_map;
395 if (current_order != MAX_ORDER-1)
396 MARK_USED(index, current_order, area);
397 zone->free_pages -= 1UL << order;
398 return expand(zone, page, index, order, current_order, area);
405 * Obtain a specified number of elements from the buddy allocator, all under
406 * a single hold of the lock, for efficiency. Add them to the supplied list.
407 * Returns the number of new pages which were placed at *list.
409 static int rmqueue_bulk(struct zone *zone, unsigned int order,
410 unsigned long count, struct list_head *list)
417 spin_lock_irqsave(&zone->lock, flags);
418 for (i = 0; i < count; ++i) {
419 page = __rmqueue(zone, order);
423 list_add_tail(&page->lru, list);
425 spin_unlock_irqrestore(&zone->lock, flags);
429 #if defined(CONFIG_PM) || defined(CONFIG_HOTPLUG_CPU)
430 static void __drain_pages(unsigned int cpu)
435 for_each_zone(zone) {
436 struct per_cpu_pageset *pset;
438 pset = &zone->pageset[cpu];
439 for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
440 struct per_cpu_pages *pcp;
443 pcp->count -= free_pages_bulk(zone, pcp->count,
448 #endif /* CONFIG_PM || CONFIG_HOTPLUG_CPU */
451 int is_head_of_free_region(struct page *page)
453 struct zone *zone = page_zone(page);
456 struct list_head *curr;
459 * Should not matter as we need quiescent system for
460 * suspend anyway, but...
462 spin_lock_irqsave(&zone->lock, flags);
463 for (order = MAX_ORDER - 1; order >= 0; --order)
464 list_for_each(curr, &zone->free_area[order].free_list)
465 if (page == list_entry(curr, struct page, lru)) {
466 spin_unlock_irqrestore(&zone->lock, flags);
469 spin_unlock_irqrestore(&zone->lock, flags);
474 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
476 void drain_local_pages(void)
480 local_irq_save(flags);
481 __drain_pages(smp_processor_id());
482 local_irq_restore(flags);
484 #endif /* CONFIG_PM */
486 static void zone_statistics(struct zonelist *zonelist, struct zone *z)
491 pg_data_t *pg = z->zone_pgdat;
492 pg_data_t *orig = zonelist->zones[0]->zone_pgdat;
493 struct per_cpu_pageset *p;
495 local_irq_save(flags);
496 cpu = smp_processor_id();
497 p = &z->pageset[cpu];
499 z->pageset[cpu].numa_hit++;
502 zonelist->zones[0]->pageset[cpu].numa_foreign++;
504 if (pg == NODE_DATA(numa_node_id()))
508 local_irq_restore(flags);
513 * Free a 0-order page
515 static void FASTCALL(free_hot_cold_page(struct page *page, int cold));
516 static void fastcall free_hot_cold_page(struct page *page, int cold)
518 struct zone *zone = page_zone(page);
519 struct per_cpu_pages *pcp;
522 arch_free_page(page, 0);
524 kernel_map_pages(page, 1, 0);
525 inc_page_state(pgfree);
527 page->mapping = NULL;
528 free_pages_check(__FUNCTION__, page);
529 pcp = &zone->pageset[get_cpu()].pcp[cold];
530 local_irq_save(flags);
531 if (pcp->count >= pcp->high)
532 pcp->count -= free_pages_bulk(zone, pcp->batch, &pcp->list, 0);
533 list_add(&page->lru, &pcp->list);
535 local_irq_restore(flags);
539 void fastcall free_hot_page(struct page *page)
541 free_hot_cold_page(page, 0);
544 void fastcall free_cold_page(struct page *page)
546 free_hot_cold_page(page, 1);
550 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
551 * we cheat by calling it from here, in the order > 0 path. Saves a branch
556 buffered_rmqueue(struct zone *zone, int order, int gfp_flags)
559 struct page *page = NULL;
560 int cold = !!(gfp_flags & __GFP_COLD);
563 struct per_cpu_pages *pcp;
565 pcp = &zone->pageset[get_cpu()].pcp[cold];
566 local_irq_save(flags);
567 if (pcp->count <= pcp->low)
568 pcp->count += rmqueue_bulk(zone, 0,
569 pcp->batch, &pcp->list);
571 page = list_entry(pcp->list.next, struct page, lru);
572 list_del(&page->lru);
575 local_irq_restore(flags);
580 spin_lock_irqsave(&zone->lock, flags);
581 page = __rmqueue(zone, order);
582 spin_unlock_irqrestore(&zone->lock, flags);
586 BUG_ON(bad_range(zone, page));
587 mod_page_state_zone(zone, pgalloc, 1 << order);
588 prep_new_page(page, order);
589 if (order && (gfp_flags & __GFP_COMP))
590 prep_compound_page(page, order);
596 * This is the 'heart' of the zoned buddy allocator.
598 * Herein lies the mysterious "incremental min". That's the
600 * local_low = z->pages_low;
603 * thing. The intent here is to provide additional protection to low zones for
604 * allocation requests which _could_ use higher zones. So a GFP_HIGHMEM
605 * request is not allowed to dip as deeply into the normal zone as a GFP_KERNEL
606 * request. This preserves additional space in those lower zones for requests
607 * which really do need memory from those zones. It means that on a decent
608 * sized machine, GFP_HIGHMEM and GFP_KERNEL requests basically leave the DMA
611 struct page * fastcall
612 __alloc_pages(unsigned int gfp_mask, unsigned int order,
613 struct zonelist *zonelist)
615 const int wait = gfp_mask & __GFP_WAIT;
617 struct zone **zones, *z;
619 struct reclaim_state reclaim_state;
620 struct task_struct *p = current;
626 might_sleep_if(wait);
628 if (!ckrm_class_limit_ok((GET_MEM_CLASS(current)))) {
633 * The caller may dip into page reserves a bit more if the caller
634 * cannot run direct reclaim, or is the caller has realtime scheduling
637 can_try_harder = (unlikely(rt_task(p)) && !in_interrupt()) || !wait;
639 zones = zonelist->zones; /* the list of zones suitable for gfp_mask */
641 if (unlikely(zones[0] == NULL)) {
642 /* Should this ever happen?? */
646 alloc_type = zone_idx(zones[0]);
648 /* Go through the zonelist once, looking for a zone with enough free */
649 for (i = 0; (z = zones[i]) != NULL; i++) {
650 min = z->pages_low + (1<<order) + z->protection[alloc_type];
652 if (z->free_pages < min)
655 page = buffered_rmqueue(z, order, gfp_mask);
660 for (i = 0; (z = zones[i]) != NULL; i++)
664 * Go through the zonelist again. Let __GFP_HIGH and allocations
665 * coming from realtime tasks to go deeper into reserves
667 for (i = 0; (z = zones[i]) != NULL; i++) {
669 if (gfp_mask & __GFP_HIGH)
673 min += (1<<order) + z->protection[alloc_type];
675 if (z->free_pages < min)
678 page = buffered_rmqueue(z, order, gfp_mask);
683 /* This allocation should allow future memory freeing. */
684 if ((p->flags & (PF_MEMALLOC | PF_MEMDIE)) && !in_interrupt()) {
685 /* go through the zonelist yet again, ignoring mins */
686 for (i = 0; (z = zones[i]) != NULL; i++) {
687 page = buffered_rmqueue(z, order, gfp_mask);
694 /* Atomic allocations - we can't balance anything */
699 /* We now go into synchronous reclaim */
700 p->flags |= PF_MEMALLOC;
701 reclaim_state.reclaimed_slab = 0;
702 p->reclaim_state = &reclaim_state;
704 try_to_free_pages(zones, gfp_mask, order);
706 p->reclaim_state = NULL;
707 p->flags &= ~PF_MEMALLOC;
709 /* go through the zonelist yet one more time */
710 for (i = 0; (z = zones[i]) != NULL; i++) {
712 if (gfp_mask & __GFP_HIGH)
716 min += (1<<order) + z->protection[alloc_type];
718 if (z->free_pages < min)
721 page = buffered_rmqueue(z, order, gfp_mask);
727 * Don't let big-order allocations loop unless the caller explicitly
728 * requests that. Wait for some write requests to complete then retry.
730 * In this implementation, __GFP_REPEAT means __GFP_NOFAIL for order
731 * <= 3, but that may not be true in other implementations.
734 if (!(gfp_mask & __GFP_NORETRY)) {
735 if ((order <= 3) || (gfp_mask & __GFP_REPEAT))
737 if (gfp_mask & __GFP_NOFAIL)
741 blk_congestion_wait(WRITE, HZ/50);
746 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
747 printk(KERN_WARNING "%s: page allocation failure."
748 " order:%d, mode:0x%x\n",
749 p->comm, order, gfp_mask);
754 zone_statistics(zonelist, z);
755 kernel_map_pages(page, 1 << order, 1);
756 ckrm_set_pages_class(page, 1 << order, GET_MEM_CLASS(current));
760 EXPORT_SYMBOL(__alloc_pages);
763 * Common helper functions.
765 fastcall unsigned long __get_free_pages(unsigned int gfp_mask, unsigned int order)
768 page = alloc_pages(gfp_mask, order);
771 return (unsigned long) page_address(page);
774 EXPORT_SYMBOL(__get_free_pages);
776 fastcall unsigned long get_zeroed_page(unsigned int gfp_mask)
781 * get_zeroed_page() returns a 32-bit address, which cannot represent
784 BUG_ON(gfp_mask & __GFP_HIGHMEM);
786 page = alloc_pages(gfp_mask, 0);
788 void *address = page_address(page);
790 return (unsigned long) address;
795 EXPORT_SYMBOL(get_zeroed_page);
797 void __pagevec_free(struct pagevec *pvec)
799 int i = pagevec_count(pvec);
802 free_hot_cold_page(pvec->pages[i], pvec->cold);
805 fastcall void __free_pages(struct page *page, unsigned int order)
807 if (!PageReserved(page) && put_page_testzero(page)) {
811 __free_pages_ok(page, order);
815 EXPORT_SYMBOL(__free_pages);
817 fastcall void free_pages(unsigned long addr, unsigned int order)
820 BUG_ON(!virt_addr_valid((void *)addr));
821 __free_pages(virt_to_page((void *)addr), order);
825 EXPORT_SYMBOL(free_pages);
828 * Total amount of free (allocatable) RAM:
830 unsigned int nr_free_pages(void)
832 unsigned int sum = 0;
836 sum += zone->free_pages;
841 EXPORT_SYMBOL(nr_free_pages);
844 unsigned int nr_free_pages_pgdat(pg_data_t *pgdat)
846 unsigned int i, sum = 0;
848 for (i = 0; i < MAX_NR_ZONES; i++)
849 sum += pgdat->node_zones[i].free_pages;
855 static unsigned int nr_free_zone_pages(int offset)
858 unsigned int sum = 0;
860 for_each_pgdat(pgdat) {
861 struct zonelist *zonelist = pgdat->node_zonelists + offset;
862 struct zone **zonep = zonelist->zones;
865 for (zone = *zonep++; zone; zone = *zonep++) {
866 unsigned long size = zone->present_pages;
867 unsigned long high = zone->pages_high;
877 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
879 unsigned int nr_free_buffer_pages(void)
881 return nr_free_zone_pages(GFP_USER & GFP_ZONEMASK);
885 * Amount of free RAM allocatable within all zones
887 unsigned int nr_free_pagecache_pages(void)
889 return nr_free_zone_pages(GFP_HIGHUSER & GFP_ZONEMASK);
892 #ifdef CONFIG_HIGHMEM
893 unsigned int nr_free_highpages (void)
896 unsigned int pages = 0;
898 for_each_pgdat(pgdat)
899 pages += pgdat->node_zones[ZONE_HIGHMEM].free_pages;
906 static void show_node(struct zone *zone)
908 printk("Node %d ", zone->zone_pgdat->node_id);
911 #define show_node(zone) do { } while (0)
915 * Accumulate the page_state information across all CPUs.
916 * The result is unavoidably approximate - it can change
917 * during and after execution of this function.
919 DEFINE_PER_CPU(struct page_state, page_states) = {0};
920 EXPORT_PER_CPU_SYMBOL(page_states);
922 atomic_t nr_pagecache = ATOMIC_INIT(0);
923 EXPORT_SYMBOL(nr_pagecache);
925 DEFINE_PER_CPU(long, nr_pagecache_local) = 0;
928 void __get_page_state(struct page_state *ret, int nr)
932 memset(ret, 0, sizeof(*ret));
933 while (cpu < NR_CPUS) {
934 unsigned long *in, *out, off;
936 if (!cpu_possible(cpu)) {
941 in = (unsigned long *)&per_cpu(page_states, cpu);
943 if (cpu < NR_CPUS && cpu_possible(cpu))
944 prefetch(&per_cpu(page_states, cpu));
945 out = (unsigned long *)ret;
946 for (off = 0; off < nr; off++)
951 void get_page_state(struct page_state *ret)
955 nr = offsetof(struct page_state, GET_PAGE_STATE_LAST);
956 nr /= sizeof(unsigned long);
958 __get_page_state(ret, nr + 1);
961 void get_full_page_state(struct page_state *ret)
963 __get_page_state(ret, sizeof(*ret) / sizeof(unsigned long));
966 unsigned long __read_page_state(unsigned offset)
968 unsigned long ret = 0;
971 for (cpu = 0; cpu < NR_CPUS; cpu++) {
974 if (!cpu_possible(cpu))
977 in = (unsigned long)&per_cpu(page_states, cpu) + offset;
978 ret += *((unsigned long *)in);
983 void __get_zone_counts(unsigned long *active, unsigned long *inactive,
984 unsigned long *free, struct pglist_data *pgdat)
986 struct zone *zones = pgdat->node_zones;
992 for (i = 0; i < MAX_NR_ZONES; i++) {
993 *active += zones[i].nr_active;
994 *inactive += zones[i].nr_inactive;
995 *free += zones[i].free_pages;
999 void get_zone_counts(unsigned long *active,
1000 unsigned long *inactive, unsigned long *free)
1002 struct pglist_data *pgdat;
1007 for_each_pgdat(pgdat) {
1008 unsigned long l, m, n;
1009 __get_zone_counts(&l, &m, &n, pgdat);
1016 void si_meminfo(struct sysinfo *val)
1018 val->totalram = totalram_pages;
1020 val->freeram = nr_free_pages();
1021 val->bufferram = nr_blockdev_pages();
1022 #ifdef CONFIG_HIGHMEM
1023 val->totalhigh = totalhigh_pages;
1024 val->freehigh = nr_free_highpages();
1029 val->mem_unit = PAGE_SIZE;
1030 if (vx_flags(VXF_VIRT_MEM, 0))
1031 vx_vsi_meminfo(val);
1034 EXPORT_SYMBOL(si_meminfo);
1037 void si_meminfo_node(struct sysinfo *val, int nid)
1039 pg_data_t *pgdat = NODE_DATA(nid);
1041 val->totalram = pgdat->node_present_pages;
1042 val->freeram = nr_free_pages_pgdat(pgdat);
1043 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
1044 val->freehigh = pgdat->node_zones[ZONE_HIGHMEM].free_pages;
1045 val->mem_unit = PAGE_SIZE;
1049 #define K(x) ((x) << (PAGE_SHIFT-10))
1052 * Show free area list (used inside shift_scroll-lock stuff)
1053 * We also calculate the percentage fragmentation. We do this by counting the
1054 * memory on each free list with the exception of the first item on the list.
1056 void show_free_areas(void)
1058 struct page_state ps;
1059 int cpu, temperature;
1060 unsigned long active;
1061 unsigned long inactive;
1065 for_each_zone(zone) {
1067 printk("%s per-cpu:", zone->name);
1069 if (!zone->present_pages) {
1075 for (cpu = 0; cpu < NR_CPUS; ++cpu) {
1076 struct per_cpu_pageset *pageset;
1078 if (!cpu_possible(cpu))
1081 pageset = zone->pageset + cpu;
1083 for (temperature = 0; temperature < 2; temperature++)
1084 printk("cpu %d %s: low %d, high %d, batch %d\n",
1086 temperature ? "cold" : "hot",
1087 pageset->pcp[temperature].low,
1088 pageset->pcp[temperature].high,
1089 pageset->pcp[temperature].batch);
1093 get_page_state(&ps);
1094 get_zone_counts(&active, &inactive, &free);
1096 printk("\nFree pages: %11ukB (%ukB HighMem)\n",
1098 K(nr_free_highpages()));
1100 printk("Active:%lu inactive:%lu dirty:%lu writeback:%lu "
1101 "unstable:%lu free:%u slab:%lu mapped:%lu pagetables:%lu\n",
1110 ps.nr_page_table_pages);
1112 for_each_zone(zone) {
1124 " pages_scanned:%lu"
1125 " all_unreclaimable? %s"
1128 K(zone->free_pages),
1131 K(zone->pages_high),
1133 K(zone->nr_inactive),
1134 K(zone->present_pages),
1135 zone->pages_scanned,
1136 (zone->all_unreclaimable ? "yes" : "no")
1138 printk("protections[]:");
1139 for (i = 0; i < MAX_NR_ZONES; i++)
1140 printk(" %lu", zone->protection[i]);
1144 for_each_zone(zone) {
1145 struct list_head *elem;
1146 unsigned long nr, flags, order, total = 0;
1149 printk("%s: ", zone->name);
1150 if (!zone->present_pages) {
1155 spin_lock_irqsave(&zone->lock, flags);
1156 for (order = 0; order < MAX_ORDER; order++) {
1158 list_for_each(elem, &zone->free_area[order].free_list)
1160 total += nr << order;
1161 printk("%lu*%lukB ", nr, K(1UL) << order);
1163 spin_unlock_irqrestore(&zone->lock, flags);
1164 printk("= %lukB\n", K(total));
1167 show_swap_cache_info();
1171 * Builds allocation fallback zone lists.
1173 static int __init build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist, int j, int k)
1180 zone = pgdat->node_zones + ZONE_HIGHMEM;
1181 if (zone->present_pages) {
1182 #ifndef CONFIG_HIGHMEM
1185 zonelist->zones[j++] = zone;
1188 zone = pgdat->node_zones + ZONE_NORMAL;
1189 if (zone->present_pages)
1190 zonelist->zones[j++] = zone;
1192 zone = pgdat->node_zones + ZONE_DMA;
1193 if (zone->present_pages)
1194 zonelist->zones[j++] = zone;
1201 #define MAX_NODE_LOAD (numnodes)
1202 static int __initdata node_load[MAX_NUMNODES];
1204 * find_next_best_node - find the next node that should appear in a given
1205 * node's fallback list
1206 * @node: node whose fallback list we're appending
1207 * @used_node_mask: pointer to the bitmap of already used nodes
1209 * We use a number of factors to determine which is the next node that should
1210 * appear on a given node's fallback list. The node should not have appeared
1211 * already in @node's fallback list, and it should be the next closest node
1212 * according to the distance array (which contains arbitrary distance values
1213 * from each node to each node in the system), and should also prefer nodes
1214 * with no CPUs, since presumably they'll have very little allocation pressure
1215 * on them otherwise.
1216 * It returns -1 if no node is found.
1218 static int __init find_next_best_node(int node, void *used_node_mask)
1221 int min_val = INT_MAX;
1224 for (i = 0; i < numnodes; i++) {
1227 /* Start from local node */
1228 n = (node+i)%numnodes;
1230 /* Don't want a node to appear more than once */
1231 if (test_bit(n, used_node_mask))
1234 /* Use the local node if we haven't already */
1235 if (!test_bit(node, used_node_mask)) {
1240 /* Use the distance array to find the distance */
1241 val = node_distance(node, n);
1243 /* Give preference to headless and unused nodes */
1244 tmp = node_to_cpumask(n);
1245 if (!cpus_empty(tmp))
1246 val += PENALTY_FOR_NODE_WITH_CPUS;
1248 /* Slight preference for less loaded node */
1249 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
1250 val += node_load[n];
1252 if (val < min_val) {
1259 set_bit(best_node, used_node_mask);
1264 static void __init build_zonelists(pg_data_t *pgdat)
1266 int i, j, k, node, local_node;
1267 int prev_node, load;
1268 struct zonelist *zonelist;
1269 DECLARE_BITMAP(used_mask, MAX_NUMNODES);
1271 /* initialize zonelists */
1272 for (i = 0; i < GFP_ZONETYPES; i++) {
1273 zonelist = pgdat->node_zonelists + i;
1274 memset(zonelist, 0, sizeof(*zonelist));
1275 zonelist->zones[0] = NULL;
1278 /* NUMA-aware ordering of nodes */
1279 local_node = pgdat->node_id;
1281 prev_node = local_node;
1282 bitmap_zero(used_mask, MAX_NUMNODES);
1283 while ((node = find_next_best_node(local_node, used_mask)) >= 0) {
1285 * We don't want to pressure a particular node.
1286 * So adding penalty to the first node in same
1287 * distance group to make it round-robin.
1289 if (node_distance(local_node, node) !=
1290 node_distance(local_node, prev_node))
1291 node_load[node] += load;
1294 for (i = 0; i < GFP_ZONETYPES; i++) {
1295 zonelist = pgdat->node_zonelists + i;
1296 for (j = 0; zonelist->zones[j] != NULL; j++);
1299 if (i & __GFP_HIGHMEM)
1304 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1305 zonelist->zones[j] = NULL;
1310 #else /* CONFIG_NUMA */
1312 static void __init build_zonelists(pg_data_t *pgdat)
1314 int i, j, k, node, local_node;
1316 local_node = pgdat->node_id;
1317 for (i = 0; i < GFP_ZONETYPES; i++) {
1318 struct zonelist *zonelist;
1320 zonelist = pgdat->node_zonelists + i;
1321 memset(zonelist, 0, sizeof(*zonelist));
1325 if (i & __GFP_HIGHMEM)
1330 j = build_zonelists_node(pgdat, zonelist, j, k);
1332 * Now we build the zonelist so that it contains the zones
1333 * of all the other nodes.
1334 * We don't want to pressure a particular node, so when
1335 * building the zones for node N, we make sure that the
1336 * zones coming right after the local ones are those from
1337 * node N+1 (modulo N)
1339 for (node = local_node + 1; node < numnodes; node++)
1340 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1341 for (node = 0; node < local_node; node++)
1342 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1344 zonelist->zones[j] = NULL;
1348 #endif /* CONFIG_NUMA */
1350 void __init build_all_zonelists(void)
1354 for(i = 0 ; i < numnodes ; i++)
1355 build_zonelists(NODE_DATA(i));
1356 printk("Built %i zonelists\n", numnodes);
1360 * Helper functions to size the waitqueue hash table.
1361 * Essentially these want to choose hash table sizes sufficiently
1362 * large so that collisions trying to wait on pages are rare.
1363 * But in fact, the number of active page waitqueues on typical
1364 * systems is ridiculously low, less than 200. So this is even
1365 * conservative, even though it seems large.
1367 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
1368 * waitqueues, i.e. the size of the waitq table given the number of pages.
1370 #define PAGES_PER_WAITQUEUE 256
1372 static inline unsigned long wait_table_size(unsigned long pages)
1374 unsigned long size = 1;
1376 pages /= PAGES_PER_WAITQUEUE;
1378 while (size < pages)
1382 * Once we have dozens or even hundreds of threads sleeping
1383 * on IO we've got bigger problems than wait queue collision.
1384 * Limit the size of the wait table to a reasonable size.
1386 size = min(size, 4096UL);
1388 return max(size, 4UL);
1392 * This is an integer logarithm so that shifts can be used later
1393 * to extract the more random high bits from the multiplicative
1394 * hash function before the remainder is taken.
1396 static inline unsigned long wait_table_bits(unsigned long size)
1401 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
1403 static void __init calculate_zone_totalpages(struct pglist_data *pgdat,
1404 unsigned long *zones_size, unsigned long *zholes_size)
1406 unsigned long realtotalpages, totalpages = 0;
1409 for (i = 0; i < MAX_NR_ZONES; i++)
1410 totalpages += zones_size[i];
1411 pgdat->node_spanned_pages = totalpages;
1413 realtotalpages = totalpages;
1415 for (i = 0; i < MAX_NR_ZONES; i++)
1416 realtotalpages -= zholes_size[i];
1417 pgdat->node_present_pages = realtotalpages;
1418 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id, realtotalpages);
1423 * Initially all pages are reserved - free ones are freed
1424 * up by free_all_bootmem() once the early boot process is
1425 * done. Non-atomic initialization, single-pass.
1427 void __init memmap_init_zone(unsigned long size, int nid, unsigned long zone,
1428 unsigned long start_pfn)
1430 struct page *start = pfn_to_page(start_pfn);
1433 for (page = start; page < (start + size); page++) {
1434 set_page_zone(page, NODEZONE(nid, zone));
1435 set_page_count(page, 0);
1436 reset_page_mapcount(page);
1437 SetPageReserved(page);
1438 INIT_LIST_HEAD(&page->lru);
1439 #ifdef WANT_PAGE_VIRTUAL
1440 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1441 if (!is_highmem_idx(zone))
1442 set_page_address(page, __va(start_pfn << PAGE_SHIFT));
1449 * Page buddy system uses "index >> (i+1)", where "index" is
1452 * The extra "+3" is to round down to byte size (8 bits per byte
1453 * assumption). Thus we get "(size-1) >> (i+4)" as the last byte
1456 * The "+1" is because we want to round the byte allocation up
1457 * rather than down. So we should have had a "+7" before we shifted
1458 * down by three. Also, we have to add one as we actually _use_ the
1459 * last bit (it's [0,n] inclusive, not [0,n[).
1461 * So we actually had +7+1 before we shift down by 3. But
1462 * (n+8) >> 3 == (n >> 3) + 1 (modulo overflows, which we do not have).
1464 * Finally, we LONG_ALIGN because all bitmap operations are on longs.
1466 unsigned long pages_to_bitmap_size(unsigned long order, unsigned long nr_pages)
1468 unsigned long bitmap_size;
1470 bitmap_size = (nr_pages-1) >> (order+4);
1471 bitmap_size = LONG_ALIGN(bitmap_size+1);
1476 void zone_init_free_lists(struct pglist_data *pgdat, struct zone *zone, unsigned long size)
1479 for (order = 0; ; order++) {
1480 unsigned long bitmap_size;
1482 INIT_LIST_HEAD(&zone->free_area[order].free_list);
1483 if (order == MAX_ORDER-1) {
1484 zone->free_area[order].map = NULL;
1488 bitmap_size = pages_to_bitmap_size(order, size);
1489 zone->free_area[order].map =
1490 (unsigned long *) alloc_bootmem_node(pgdat, bitmap_size);
1494 #ifndef __HAVE_ARCH_MEMMAP_INIT
1495 #define memmap_init(size, nid, zone, start_pfn) \
1496 memmap_init_zone((size), (nid), (zone), (start_pfn))
1500 * Set up the zone data structures:
1501 * - mark all pages reserved
1502 * - mark all memory queues empty
1503 * - clear the memory bitmaps
1505 static void __init free_area_init_core(struct pglist_data *pgdat,
1506 unsigned long *zones_size, unsigned long *zholes_size)
1509 const unsigned long zone_required_alignment = 1UL << (MAX_ORDER-1);
1510 int cpu, nid = pgdat->node_id;
1511 unsigned long zone_start_pfn = pgdat->node_start_pfn;
1513 pgdat->nr_zones = 0;
1514 init_waitqueue_head(&pgdat->kswapd_wait);
1516 for (j = 0; j < MAX_NR_ZONES; j++) {
1517 struct zone *zone = pgdat->node_zones + j;
1518 unsigned long size, realsize;
1519 unsigned long batch;
1521 zone_table[NODEZONE(nid, j)] = zone;
1522 realsize = size = zones_size[j];
1524 realsize -= zholes_size[j];
1526 if (j == ZONE_DMA || j == ZONE_NORMAL)
1527 nr_kernel_pages += realsize;
1528 nr_all_pages += realsize;
1530 zone->spanned_pages = size;
1531 zone->present_pages = realsize;
1532 zone->name = zone_names[j];
1533 spin_lock_init(&zone->lock);
1534 spin_lock_init(&zone->lru_lock);
1535 zone->zone_pgdat = pgdat;
1536 zone->free_pages = 0;
1538 zone->temp_priority = zone->prev_priority = DEF_PRIORITY;
1541 * The per-cpu-pages pools are set to around 1000th of the
1542 * size of the zone. But no more than 1/4 of a meg - there's
1543 * no point in going beyond the size of L2 cache.
1545 * OK, so we don't know how big the cache is. So guess.
1547 batch = zone->present_pages / 1024;
1548 if (batch * PAGE_SIZE > 256 * 1024)
1549 batch = (256 * 1024) / PAGE_SIZE;
1550 batch /= 4; /* We effectively *= 4 below */
1554 for (cpu = 0; cpu < NR_CPUS; cpu++) {
1555 struct per_cpu_pages *pcp;
1557 pcp = &zone->pageset[cpu].pcp[0]; /* hot */
1559 pcp->low = 2 * batch;
1560 pcp->high = 6 * batch;
1561 pcp->batch = 1 * batch;
1562 INIT_LIST_HEAD(&pcp->list);
1564 pcp = &zone->pageset[cpu].pcp[1]; /* cold */
1567 pcp->high = 2 * batch;
1568 pcp->batch = 1 * batch;
1569 INIT_LIST_HEAD(&pcp->list);
1571 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%lu\n",
1572 zone_names[j], realsize, batch);
1573 INIT_LIST_HEAD(&zone->active_list);
1574 INIT_LIST_HEAD(&zone->inactive_list);
1575 zone->nr_scan_active = 0;
1576 zone->nr_scan_inactive = 0;
1577 zone->nr_active = 0;
1578 zone->nr_inactive = 0;
1583 * The per-page waitqueue mechanism uses hashed waitqueues
1586 zone->wait_table_size = wait_table_size(size);
1587 zone->wait_table_bits =
1588 wait_table_bits(zone->wait_table_size);
1589 zone->wait_table = (wait_queue_head_t *)
1590 alloc_bootmem_node(pgdat, zone->wait_table_size
1591 * sizeof(wait_queue_head_t));
1593 for(i = 0; i < zone->wait_table_size; ++i)
1594 init_waitqueue_head(zone->wait_table + i);
1596 pgdat->nr_zones = j+1;
1598 zone->zone_mem_map = pfn_to_page(zone_start_pfn);
1599 zone->zone_start_pfn = zone_start_pfn;
1601 if ((zone_start_pfn) & (zone_required_alignment-1))
1602 printk("BUG: wrong zone alignment, it will crash\n");
1604 memmap_init(size, nid, j, zone_start_pfn);
1606 zone_start_pfn += size;
1608 zone_init_free_lists(pgdat, zone, zone->spanned_pages);
1612 void __init node_alloc_mem_map(struct pglist_data *pgdat)
1616 size = (pgdat->node_spanned_pages + 1) * sizeof(struct page);
1617 pgdat->node_mem_map = alloc_bootmem_node(pgdat, size);
1618 #ifndef CONFIG_DISCONTIGMEM
1619 mem_map = contig_page_data.node_mem_map;
1623 void __init free_area_init_node(int nid, struct pglist_data *pgdat,
1624 unsigned long *zones_size, unsigned long node_start_pfn,
1625 unsigned long *zholes_size)
1627 pgdat->node_id = nid;
1628 pgdat->node_start_pfn = node_start_pfn;
1629 calculate_zone_totalpages(pgdat, zones_size, zholes_size);
1631 if (!pfn_to_page(node_start_pfn))
1632 node_alloc_mem_map(pgdat);
1634 free_area_init_core(pgdat, zones_size, zholes_size);
1637 #ifndef CONFIG_DISCONTIGMEM
1638 static bootmem_data_t contig_bootmem_data;
1639 struct pglist_data contig_page_data = { .bdata = &contig_bootmem_data };
1641 EXPORT_SYMBOL(contig_page_data);
1643 void __init free_area_init(unsigned long *zones_size)
1645 free_area_init_node(0, &contig_page_data, zones_size,
1646 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
1650 #ifdef CONFIG_PROC_FS
1652 #include <linux/seq_file.h>
1654 static void *frag_start(struct seq_file *m, loff_t *pos)
1659 for (pgdat = pgdat_list; pgdat && node; pgdat = pgdat->pgdat_next)
1665 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
1667 pg_data_t *pgdat = (pg_data_t *)arg;
1670 return pgdat->pgdat_next;
1673 static void frag_stop(struct seq_file *m, void *arg)
1678 * This walks the freelist for each zone. Whilst this is slow, I'd rather
1679 * be slow here than slow down the fast path by keeping stats - mjbligh
1681 static int frag_show(struct seq_file *m, void *arg)
1683 pg_data_t *pgdat = (pg_data_t *)arg;
1685 struct zone *node_zones = pgdat->node_zones;
1686 unsigned long flags;
1689 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
1690 if (!zone->present_pages)
1693 spin_lock_irqsave(&zone->lock, flags);
1694 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1695 for (order = 0; order < MAX_ORDER; ++order) {
1696 unsigned long nr_bufs = 0;
1697 struct list_head *elem;
1699 list_for_each(elem, &(zone->free_area[order].free_list))
1701 seq_printf(m, "%6lu ", nr_bufs);
1703 spin_unlock_irqrestore(&zone->lock, flags);
1709 struct seq_operations fragmentation_op = {
1710 .start = frag_start,
1716 static char *vmstat_text[] = {
1720 "nr_page_table_pages",
1745 "pgscan_kswapd_high",
1746 "pgscan_kswapd_normal",
1748 "pgscan_kswapd_dma",
1749 "pgscan_direct_high",
1750 "pgscan_direct_normal",
1751 "pgscan_direct_dma",
1756 "kswapd_inodesteal",
1763 static void *vmstat_start(struct seq_file *m, loff_t *pos)
1765 struct page_state *ps;
1767 if (*pos >= ARRAY_SIZE(vmstat_text))
1770 ps = kmalloc(sizeof(*ps), GFP_KERNEL);
1773 return ERR_PTR(-ENOMEM);
1774 get_full_page_state(ps);
1775 ps->pgpgin /= 2; /* sectors -> kbytes */
1777 return (unsigned long *)ps + *pos;
1780 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
1783 if (*pos >= ARRAY_SIZE(vmstat_text))
1785 return (unsigned long *)m->private + *pos;
1788 static int vmstat_show(struct seq_file *m, void *arg)
1790 unsigned long *l = arg;
1791 unsigned long off = l - (unsigned long *)m->private;
1793 seq_printf(m, "%s %lu\n", vmstat_text[off], *l);
1797 static void vmstat_stop(struct seq_file *m, void *arg)
1803 struct seq_operations vmstat_op = {
1804 .start = vmstat_start,
1805 .next = vmstat_next,
1806 .stop = vmstat_stop,
1807 .show = vmstat_show,
1810 #endif /* CONFIG_PROC_FS */
1812 #ifdef CONFIG_HOTPLUG_CPU
1813 static int page_alloc_cpu_notify(struct notifier_block *self,
1814 unsigned long action, void *hcpu)
1816 int cpu = (unsigned long)hcpu;
1819 if (action == CPU_DEAD) {
1820 /* Drain local pagecache count. */
1821 count = &per_cpu(nr_pagecache_local, cpu);
1822 atomic_add(*count, &nr_pagecache);
1824 local_irq_disable();
1830 #endif /* CONFIG_HOTPLUG_CPU */
1832 void __init page_alloc_init(void)
1834 hotcpu_notifier(page_alloc_cpu_notify, 0);
1837 static unsigned long higherzone_val(struct zone *z, int max_zone,
1840 int z_idx = zone_idx(z);
1841 struct zone *higherzone;
1842 unsigned long pages;
1844 /* there is no higher zone to get a contribution from */
1845 if (z_idx == MAX_NR_ZONES-1)
1848 higherzone = &z->zone_pgdat->node_zones[z_idx+1];
1850 /* We always start with the higher zone's protection value */
1851 pages = higherzone->protection[alloc_type];
1854 * We get a lower-zone-protection contribution only if there are
1855 * pages in the higher zone and if we're not the highest zone
1856 * in the current zonelist. e.g., never happens for GFP_DMA. Happens
1857 * only for ZONE_DMA in a GFP_KERNEL allocation and happens for ZONE_DMA
1858 * and ZONE_NORMAL for a GFP_HIGHMEM allocation.
1860 if (higherzone->present_pages && z_idx < alloc_type)
1861 pages += higherzone->pages_low * sysctl_lower_zone_protection;
1867 * setup_per_zone_protection - called whenver min_free_kbytes or
1868 * sysctl_lower_zone_protection changes. Ensures that each zone
1869 * has a correct pages_protected value, so an adequate number of
1870 * pages are left in the zone after a successful __alloc_pages().
1872 * This algorithm is way confusing. I tries to keep the same behavior
1873 * as we had with the incremental min iterative algorithm.
1875 static void setup_per_zone_protection(void)
1877 struct pglist_data *pgdat;
1878 struct zone *zones, *zone;
1882 for_each_pgdat(pgdat) {
1883 zones = pgdat->node_zones;
1885 for (i = 0, max_zone = 0; i < MAX_NR_ZONES; i++)
1886 if (zones[i].present_pages)
1890 * For each of the different allocation types:
1891 * GFP_DMA -> GFP_KERNEL -> GFP_HIGHMEM
1893 for (i = 0; i < GFP_ZONETYPES; i++) {
1895 * For each of the zones:
1896 * ZONE_HIGHMEM -> ZONE_NORMAL -> ZONE_DMA
1898 for (j = MAX_NR_ZONES-1; j >= 0; j--) {
1902 * We never protect zones that don't have memory
1903 * in them (j>max_zone) or zones that aren't in
1904 * the zonelists for a certain type of
1905 * allocation (j>=i). We have to assign these
1906 * to zero because the lower zones take
1907 * contributions from the higher zones.
1909 if (j > max_zone || j >= i) {
1910 zone->protection[i] = 0;
1914 * The contribution of the next higher zone
1916 zone->protection[i] = higherzone_val(zone,
1924 * setup_per_zone_pages_min - called when min_free_kbytes changes. Ensures
1925 * that the pages_{min,low,high} values for each zone are set correctly
1926 * with respect to min_free_kbytes.
1928 static void setup_per_zone_pages_min(void)
1930 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
1931 unsigned long lowmem_pages = 0;
1933 unsigned long flags;
1935 /* Calculate total number of !ZONE_HIGHMEM pages */
1936 for_each_zone(zone) {
1937 if (!is_highmem(zone))
1938 lowmem_pages += zone->present_pages;
1941 for_each_zone(zone) {
1942 spin_lock_irqsave(&zone->lru_lock, flags);
1943 if (is_highmem(zone)) {
1945 * Often, highmem doesn't need to reserve any pages.
1946 * But the pages_min/low/high values are also used for
1947 * batching up page reclaim activity so we need a
1948 * decent value here.
1952 min_pages = zone->present_pages / 1024;
1953 if (min_pages < SWAP_CLUSTER_MAX)
1954 min_pages = SWAP_CLUSTER_MAX;
1955 if (min_pages > 128)
1957 zone->pages_min = min_pages;
1959 /* if it's a lowmem zone, reserve a number of pages
1960 * proportionate to the zone's size.
1962 zone->pages_min = (pages_min * zone->present_pages) /
1967 * When interpreting these watermarks, just keep in mind that:
1968 * zone->pages_min == (zone->pages_min * 4) / 4;
1970 zone->pages_low = (zone->pages_min * 5) / 4;
1971 zone->pages_high = (zone->pages_min * 6) / 4;
1972 spin_unlock_irqrestore(&zone->lru_lock, flags);
1977 * Initialise min_free_kbytes.
1979 * For small machines we want it small (128k min). For large machines
1980 * we want it large (64MB max). But it is not linear, because network
1981 * bandwidth does not increase linearly with machine size. We use
1983 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
1984 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
2000 static int __init init_per_zone_pages_min(void)
2002 unsigned long lowmem_kbytes;
2004 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
2006 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
2007 if (min_free_kbytes < 128)
2008 min_free_kbytes = 128;
2009 if (min_free_kbytes > 65536)
2010 min_free_kbytes = 65536;
2011 setup_per_zone_pages_min();
2012 setup_per_zone_protection();
2015 module_init(init_per_zone_pages_min)
2018 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
2019 * that we can call two helper functions whenever min_free_kbytes
2022 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
2023 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2025 proc_dointvec(table, write, file, buffer, length, ppos);
2026 setup_per_zone_pages_min();
2027 setup_per_zone_protection();
2032 * lower_zone_protection_sysctl_handler - just a wrapper around
2033 * proc_dointvec() so that we can call setup_per_zone_protection()
2034 * whenever sysctl_lower_zone_protection changes.
2036 int lower_zone_protection_sysctl_handler(ctl_table *table, int write,
2037 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2039 proc_dointvec_minmax(table, write, file, buffer, length, ppos);
2040 setup_per_zone_protection();
2045 * allocate a large system hash table from bootmem
2046 * - it is assumed that the hash table must contain an exact power-of-2
2047 * quantity of entries
2049 void *__init alloc_large_system_hash(const char *tablename,
2050 unsigned long bucketsize,
2051 unsigned long numentries,
2053 int consider_highmem,
2054 unsigned int *_hash_shift,
2055 unsigned int *_hash_mask)
2057 unsigned long long max;
2058 unsigned long log2qty, size;
2061 /* allow the kernel cmdline to have a say */
2063 /* round applicable memory size up to nearest megabyte */
2064 numentries = consider_highmem ? nr_all_pages : nr_kernel_pages;
2065 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
2066 numentries >>= 20 - PAGE_SHIFT;
2067 numentries <<= 20 - PAGE_SHIFT;
2069 /* limit to 1 bucket per 2^scale bytes of low memory */
2070 if (scale > PAGE_SHIFT)
2071 numentries >>= (scale - PAGE_SHIFT);
2073 numentries <<= (PAGE_SHIFT - scale);
2075 /* rounded up to nearest power of 2 in size */
2076 numentries = 1UL << (long_log2(numentries) + 1);
2078 /* limit allocation size to 1/16 total memory */
2079 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
2080 do_div(max, bucketsize);
2082 if (numentries > max)
2085 log2qty = long_log2(numentries);
2088 size = bucketsize << log2qty;
2089 table = alloc_bootmem(size);
2090 } while (!table && size > PAGE_SIZE && --log2qty);
2093 panic("Failed to allocate %s hash table\n", tablename);
2095 printk("%s hash table entries: %d (order: %d, %lu bytes)\n",
2098 long_log2(size) - PAGE_SHIFT,
2102 *_hash_shift = log2qty;
2104 *_hash_mask = (1 << log2qty) - 1;