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>
39 #include <asm/tlbflush.h>
41 nodemask_t node_online_map = NODE_MASK_NONE;
42 nodemask_t node_possible_map = NODE_MASK_ALL;
43 struct pglist_data *pgdat_list;
44 unsigned long totalram_pages;
45 unsigned long totalhigh_pages;
48 int sysctl_lower_zone_protection = 0;
50 EXPORT_SYMBOL(totalram_pages);
51 EXPORT_SYMBOL(nr_swap_pages);
53 #ifdef CONFIG_CRASH_DUMP_MODULE
54 /* This symbol has to be exported to use 'for_each_pgdat' macro by modules. */
55 EXPORT_SYMBOL(pgdat_list);
59 * Used by page_zone() to look up the address of the struct zone whose
60 * id is encoded in the upper bits of page->flags
62 struct zone *zone_table[1 << (ZONES_SHIFT + NODES_SHIFT)];
63 EXPORT_SYMBOL(zone_table);
65 static char *zone_names[MAX_NR_ZONES] = { "DMA", "Normal", "HighMem" };
66 int min_free_kbytes = 1024;
68 unsigned long __initdata nr_kernel_pages;
69 unsigned long __initdata nr_all_pages;
72 * Temporary debugging check for pages not lying within a given zone.
74 static int bad_range(struct zone *zone, struct page *page)
76 if (page_to_pfn(page) >= zone->zone_start_pfn + zone->spanned_pages)
78 if (page_to_pfn(page) < zone->zone_start_pfn)
80 if (zone != page_zone(page))
85 static void bad_page(const char *function, struct page *page)
87 printk(KERN_EMERG "Bad page state at %s (in process '%s', page %p)\n",
88 function, current->comm, page);
89 printk(KERN_EMERG "flags:0x%0*lx mapping:%p mapcount:%d count:%d (%s)\n",
90 (int)(2*sizeof(page_flags_t)), (unsigned long)page->flags,
91 page->mapping, page_mapcount(page), page_count(page), print_tainted());
92 printk(KERN_EMERG "Backtrace:\n");
94 printk(KERN_EMERG "Trying to fix it up, but a reboot is needed\n");
95 page->flags &= ~(1 << PG_private |
102 set_page_count(page, 0);
103 reset_page_mapcount(page);
104 page->mapping = NULL;
105 tainted |= TAINT_BAD_PAGE;
108 #if !defined(CONFIG_HUGETLB_PAGE) && !defined(CONFIG_CRASH_DUMP) \
109 && !defined(CONFIG_CRASH_DUMP_MODULE)
110 #define prep_compound_page(page, order) do { } while (0)
111 #define destroy_compound_page(page, order) do { } while (0)
114 * Higher-order pages are called "compound pages". They are structured thusly:
116 * The first PAGE_SIZE page is called the "head page".
118 * The remaining PAGE_SIZE pages are called "tail pages".
120 * All pages have PG_compound set. All pages have their ->private pointing at
121 * the head page (even the head page has this).
123 * The first tail page's ->mapping, if non-zero, holds the address of the
124 * compound page's put_page() function.
126 * The order of the allocation is stored in the first tail page's ->index
127 * This is only for debug at present. This usage means that zero-order pages
128 * may not be compound.
130 static void prep_compound_page(struct page *page, unsigned long order)
133 int nr_pages = 1 << order;
135 page[1].mapping = NULL;
136 page[1].index = order;
137 for (i = 0; i < nr_pages; i++) {
138 struct page *p = page + i;
141 p->private = (unsigned long)page;
145 static void destroy_compound_page(struct page *page, unsigned long order)
148 int nr_pages = 1 << order;
150 if (!PageCompound(page))
153 if (page[1].index != order)
154 bad_page(__FUNCTION__, page);
156 for (i = 0; i < nr_pages; i++) {
157 struct page *p = page + i;
159 if (!PageCompound(p))
160 bad_page(__FUNCTION__, page);
161 if (p->private != (unsigned long)page)
162 bad_page(__FUNCTION__, page);
163 ClearPageCompound(p);
166 #endif /* CONFIG_HUGETLB_PAGE */
169 * Freeing function for a buddy system allocator.
171 * The concept of a buddy system is to maintain direct-mapped table
172 * (containing bit values) for memory blocks of various "orders".
173 * The bottom level table contains the map for the smallest allocatable
174 * units of memory (here, pages), and each level above it describes
175 * pairs of units from the levels below, hence, "buddies".
176 * At a high level, all that happens here is marking the table entry
177 * at the bottom level available, and propagating the changes upward
178 * as necessary, plus some accounting needed to play nicely with other
179 * parts of the VM system.
180 * At each level, we keep one bit for each pair of blocks, which
181 * is set to 1 iff only one of the pair is allocated. So when we
182 * are allocating or freeing one, we can derive the state of the
183 * other. That is, if we allocate a small block, and both were
184 * free, the remainder of the region must be split into blocks.
185 * If a block is freed, and its buddy is also free, then this
186 * triggers coalescing into a block of larger size.
191 static inline void __free_pages_bulk (struct page *page, struct page *base,
192 struct zone *zone, struct free_area *area, unsigned int order)
194 unsigned long page_idx, index, mask;
197 destroy_compound_page(page, order);
198 mask = (~0UL) << order;
199 page_idx = page - base;
200 if (page_idx & ~mask)
202 index = page_idx >> (1 + order);
204 zone->free_pages += 1 << order;
205 while (order < MAX_ORDER-1) {
206 struct page *buddy1, *buddy2;
208 BUG_ON(area >= zone->free_area + MAX_ORDER);
209 if (!__test_and_change_bit(index, area->map))
211 * the buddy page is still allocated.
215 /* Move the buddy up one level. */
216 buddy1 = base + (page_idx ^ (1 << order));
217 buddy2 = base + page_idx;
218 BUG_ON(bad_range(zone, buddy1));
219 BUG_ON(bad_range(zone, buddy2));
220 list_del(&buddy1->lru);
227 list_add(&(base + page_idx)->lru, &area->free_list);
230 static inline void free_pages_check(const char *function, struct page *page)
232 if ( page_mapped(page) ||
233 page->mapping != NULL ||
234 page_count(page) != 0 ||
243 1 << PG_writeback )))
244 bad_page(function, page);
246 ClearPageDirty(page);
250 * Frees a list of pages.
251 * Assumes all pages on list are in same zone, and of same order.
252 * count is the number of pages to free, or 0 for all on the list.
254 * If the zone was previously in an "all pages pinned" state then look to
255 * see if this freeing clears that state.
257 * And clear the zone's pages_scanned counter, to hold off the "all pages are
258 * pinned" detection logic.
261 free_pages_bulk(struct zone *zone, int count,
262 struct list_head *list, unsigned int order)
265 struct free_area *area;
266 struct page *base, *page = NULL;
269 base = zone->zone_mem_map;
270 area = zone->free_area + order;
271 spin_lock_irqsave(&zone->lock, flags);
272 zone->all_unreclaimable = 0;
273 zone->pages_scanned = 0;
274 while (!list_empty(list) && count--) {
275 page = list_entry(list->prev, struct page, lru);
276 /* have to delete it as __free_pages_bulk list manipulates */
277 list_del(&page->lru);
278 __free_pages_bulk(page, base, zone, area, order);
279 ckrm_clear_page_class(page);
282 spin_unlock_irqrestore(&zone->lock, flags);
286 void __free_pages_ok(struct page *page, unsigned int order)
291 arch_free_page(page, order);
293 mod_page_state(pgfree, 1 << order);
294 for (i = 0 ; i < (1 << order) ; ++i)
295 free_pages_check(__FUNCTION__, page + i);
296 list_add(&page->lru, &list);
297 kernel_map_pages(page, 1<<order, 0);
298 free_pages_bulk(page_zone(page), 1, &list, order);
301 #define MARK_USED(index, order, area) \
302 __change_bit((index) >> (1+(order)), (area)->map)
305 * The order of subdivision here is critical for the IO subsystem.
306 * Please do not alter this order without good reasons and regression
307 * testing. Specifically, as large blocks of memory are subdivided,
308 * the order in which smaller blocks are delivered depends on the order
309 * they're subdivided in this function. This is the primary factor
310 * influencing the order in which pages are delivered to the IO
311 * subsystem according to empirical testing, and this is also justified
312 * by considering the behavior of a buddy system containing a single
313 * large block of memory acted on by a series of small allocations.
314 * This behavior is a critical factor in sglist merging's success.
318 static inline struct page *
319 expand(struct zone *zone, struct page *page,
320 unsigned long index, int low, int high, struct free_area *area)
322 unsigned long size = 1 << high;
328 BUG_ON(bad_range(zone, &page[size]));
329 list_add(&page[size].lru, &area->free_list);
330 MARK_USED(index + size, high, area);
335 static inline void set_page_refs(struct page *page, int order)
338 set_page_count(page, 1);
343 * We need to reference all the pages for this order, otherwise if
344 * anyone accesses one of the pages with (get/put) it will be freed.
346 for (i = 0; i < (1 << order); i++)
347 set_page_count(page+i, 1);
348 #endif /* CONFIG_MMU */
352 * This page is about to be returned from the page allocator
354 static void prep_new_page(struct page *page, int order)
356 if (page->mapping || page_mapped(page) ||
365 1 << PG_writeback )))
366 bad_page(__FUNCTION__, page);
368 page->flags &= ~(1 << PG_uptodate | 1 << PG_error |
369 1 << PG_referenced | 1 << PG_arch_1 |
370 1 << PG_checked | 1 << PG_mappedtodisk);
372 set_page_refs(page, order);
376 * Do the hard work of removing an element from the buddy allocator.
377 * Call me with the zone->lock already held.
379 static struct page *__rmqueue(struct zone *zone, unsigned int order)
381 struct free_area * area;
382 unsigned int current_order;
386 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
387 area = zone->free_area + current_order;
388 if (list_empty(&area->free_list))
391 page = list_entry(area->free_list.next, struct page, lru);
392 list_del(&page->lru);
393 index = page - zone->zone_mem_map;
394 if (current_order != MAX_ORDER-1)
395 MARK_USED(index, current_order, area);
396 zone->free_pages -= 1UL << order;
397 return expand(zone, page, index, order, current_order, area);
404 * Obtain a specified number of elements from the buddy allocator, all under
405 * a single hold of the lock, for efficiency. Add them to the supplied list.
406 * Returns the number of new pages which were placed at *list.
408 static int rmqueue_bulk(struct zone *zone, unsigned int order,
409 unsigned long count, struct list_head *list)
416 spin_lock_irqsave(&zone->lock, flags);
417 for (i = 0; i < count; ++i) {
418 page = __rmqueue(zone, order);
422 list_add_tail(&page->lru, list);
424 spin_unlock_irqrestore(&zone->lock, flags);
428 #if defined(CONFIG_PM) || defined(CONFIG_HOTPLUG_CPU)
429 static void __drain_pages(unsigned int cpu)
434 for_each_zone(zone) {
435 struct per_cpu_pageset *pset;
437 pset = &zone->pageset[cpu];
438 for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
439 struct per_cpu_pages *pcp;
442 pcp->count -= free_pages_bulk(zone, pcp->count,
447 #endif /* CONFIG_PM || CONFIG_HOTPLUG_CPU */
450 int is_head_of_free_region(struct page *page)
452 struct zone *zone = page_zone(page);
455 struct list_head *curr;
458 * Should not matter as we need quiescent system for
459 * suspend anyway, but...
461 spin_lock_irqsave(&zone->lock, flags);
462 for (order = MAX_ORDER - 1; order >= 0; --order)
463 list_for_each(curr, &zone->free_area[order].free_list)
464 if (page == list_entry(curr, struct page, lru)) {
465 spin_unlock_irqrestore(&zone->lock, flags);
468 spin_unlock_irqrestore(&zone->lock, flags);
473 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
475 void drain_local_pages(void)
479 local_irq_save(flags);
480 __drain_pages(smp_processor_id());
481 local_irq_restore(flags);
483 #endif /* CONFIG_PM */
485 static void zone_statistics(struct zonelist *zonelist, struct zone *z)
490 pg_data_t *pg = z->zone_pgdat;
491 pg_data_t *orig = zonelist->zones[0]->zone_pgdat;
492 struct per_cpu_pageset *p;
494 local_irq_save(flags);
495 cpu = smp_processor_id();
496 p = &z->pageset[cpu];
498 z->pageset[cpu].numa_hit++;
501 zonelist->zones[0]->pageset[cpu].numa_foreign++;
503 if (pg == NODE_DATA(numa_node_id()))
507 local_irq_restore(flags);
512 * Free a 0-order page
514 static void FASTCALL(free_hot_cold_page(struct page *page, int cold));
515 static void fastcall free_hot_cold_page(struct page *page, int cold)
517 struct zone *zone = page_zone(page);
518 struct per_cpu_pages *pcp;
521 arch_free_page(page, 0);
523 kernel_map_pages(page, 1, 0);
524 inc_page_state(pgfree);
526 page->mapping = NULL;
527 free_pages_check(__FUNCTION__, page);
528 pcp = &zone->pageset[get_cpu()].pcp[cold];
529 local_irq_save(flags);
530 if (pcp->count >= pcp->high)
531 pcp->count -= free_pages_bulk(zone, pcp->batch, &pcp->list, 0);
532 list_add(&page->lru, &pcp->list);
534 local_irq_restore(flags);
538 void fastcall free_hot_page(struct page *page)
540 free_hot_cold_page(page, 0);
543 void fastcall free_cold_page(struct page *page)
545 free_hot_cold_page(page, 1);
549 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
550 * we cheat by calling it from here, in the order > 0 path. Saves a branch
555 buffered_rmqueue(struct zone *zone, int order, int gfp_flags)
558 struct page *page = NULL;
559 int cold = !!(gfp_flags & __GFP_COLD);
562 struct per_cpu_pages *pcp;
564 pcp = &zone->pageset[get_cpu()].pcp[cold];
565 local_irq_save(flags);
566 if (pcp->count <= pcp->low)
567 pcp->count += rmqueue_bulk(zone, 0,
568 pcp->batch, &pcp->list);
570 page = list_entry(pcp->list.next, struct page, lru);
571 list_del(&page->lru);
574 local_irq_restore(flags);
579 spin_lock_irqsave(&zone->lock, flags);
580 page = __rmqueue(zone, order);
581 spin_unlock_irqrestore(&zone->lock, flags);
585 BUG_ON(bad_range(zone, page));
586 mod_page_state_zone(zone, pgalloc, 1 << order);
587 prep_new_page(page, order);
588 if (order && (gfp_flags & __GFP_COMP))
589 prep_compound_page(page, order);
595 * This is the 'heart' of the zoned buddy allocator.
597 * Herein lies the mysterious "incremental min". That's the
599 * local_low = z->pages_low;
602 * thing. The intent here is to provide additional protection to low zones for
603 * allocation requests which _could_ use higher zones. So a GFP_HIGHMEM
604 * request is not allowed to dip as deeply into the normal zone as a GFP_KERNEL
605 * request. This preserves additional space in those lower zones for requests
606 * which really do need memory from those zones. It means that on a decent
607 * sized machine, GFP_HIGHMEM and GFP_KERNEL requests basically leave the DMA
610 struct page * fastcall
611 __alloc_pages(unsigned int gfp_mask, unsigned int order,
612 struct zonelist *zonelist)
614 const int wait = gfp_mask & __GFP_WAIT;
616 struct zone **zones, *z;
618 struct reclaim_state reclaim_state;
619 struct task_struct *p = current;
625 might_sleep_if(wait);
627 if (!ckrm_class_limit_ok((GET_MEM_CLASS(current)))) {
632 * The caller may dip into page reserves a bit more if the caller
633 * cannot run direct reclaim, or is the caller has realtime scheduling
636 can_try_harder = (unlikely(rt_task(p)) && !in_interrupt()) || !wait;
638 zones = zonelist->zones; /* the list of zones suitable for gfp_mask */
640 if (unlikely(zones[0] == NULL)) {
641 /* Should this ever happen?? */
645 alloc_type = zone_idx(zones[0]);
647 /* Go through the zonelist once, looking for a zone with enough free */
648 for (i = 0; (z = zones[i]) != NULL; i++) {
649 min = z->pages_low + (1<<order) + z->protection[alloc_type];
651 if (z->free_pages < min)
654 page = buffered_rmqueue(z, order, gfp_mask);
659 for (i = 0; (z = zones[i]) != NULL; i++)
663 * Go through the zonelist again. Let __GFP_HIGH and allocations
664 * coming from realtime tasks to go deeper into reserves
666 for (i = 0; (z = zones[i]) != NULL; i++) {
668 if (gfp_mask & __GFP_HIGH)
672 min += (1<<order) + z->protection[alloc_type];
674 if (z->free_pages < min)
677 page = buffered_rmqueue(z, order, gfp_mask);
682 /* This allocation should allow future memory freeing. */
683 if ((p->flags & (PF_MEMALLOC | PF_MEMDIE)) && !in_interrupt()) {
684 /* go through the zonelist yet again, ignoring mins */
685 for (i = 0; (z = zones[i]) != NULL; i++) {
686 page = buffered_rmqueue(z, order, gfp_mask);
693 /* Atomic allocations - we can't balance anything */
698 /* We now go into synchronous reclaim */
699 p->flags |= PF_MEMALLOC;
700 reclaim_state.reclaimed_slab = 0;
701 p->reclaim_state = &reclaim_state;
703 try_to_free_pages(zones, gfp_mask, order);
705 p->reclaim_state = NULL;
706 p->flags &= ~PF_MEMALLOC;
708 /* go through the zonelist yet one more time */
709 for (i = 0; (z = zones[i]) != NULL; i++) {
711 if (gfp_mask & __GFP_HIGH)
715 min += (1<<order) + z->protection[alloc_type];
717 if (z->free_pages < min)
720 page = buffered_rmqueue(z, order, gfp_mask);
726 * Don't let big-order allocations loop unless the caller explicitly
727 * requests that. Wait for some write requests to complete then retry.
729 * In this implementation, __GFP_REPEAT means __GFP_NOFAIL for order
730 * <= 3, but that may not be true in other implementations.
733 if (!(gfp_mask & __GFP_NORETRY)) {
734 if ((order <= 3) || (gfp_mask & __GFP_REPEAT))
736 if (gfp_mask & __GFP_NOFAIL)
740 blk_congestion_wait(WRITE, HZ/50);
745 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
746 printk(KERN_WARNING "%s: page allocation failure."
747 " order:%d, mode:0x%x\n",
748 p->comm, order, gfp_mask);
753 zone_statistics(zonelist, z);
754 kernel_map_pages(page, 1 << order, 1);
755 ckrm_set_pages_class(page, 1 << order, GET_MEM_CLASS(current));
759 EXPORT_SYMBOL(__alloc_pages);
762 * Common helper functions.
764 fastcall unsigned long __get_free_pages(unsigned int gfp_mask, unsigned int order)
767 page = alloc_pages(gfp_mask, order);
770 return (unsigned long) page_address(page);
773 EXPORT_SYMBOL(__get_free_pages);
775 fastcall unsigned long get_zeroed_page(unsigned int gfp_mask)
780 * get_zeroed_page() returns a 32-bit address, which cannot represent
783 BUG_ON(gfp_mask & __GFP_HIGHMEM);
785 page = alloc_pages(gfp_mask, 0);
787 void *address = page_address(page);
789 return (unsigned long) address;
794 EXPORT_SYMBOL(get_zeroed_page);
796 void __pagevec_free(struct pagevec *pvec)
798 int i = pagevec_count(pvec);
801 free_hot_cold_page(pvec->pages[i], pvec->cold);
804 fastcall void __free_pages(struct page *page, unsigned int order)
806 if (!PageReserved(page) && put_page_testzero(page)) {
810 __free_pages_ok(page, order);
814 EXPORT_SYMBOL(__free_pages);
816 fastcall void free_pages(unsigned long addr, unsigned int order)
819 BUG_ON(!virt_addr_valid((void *)addr));
820 __free_pages(virt_to_page((void *)addr), order);
824 EXPORT_SYMBOL(free_pages);
827 * Total amount of free (allocatable) RAM:
829 unsigned int nr_free_pages(void)
831 unsigned int sum = 0;
835 sum += zone->free_pages;
840 EXPORT_SYMBOL(nr_free_pages);
843 unsigned int nr_free_pages_pgdat(pg_data_t *pgdat)
845 unsigned int i, sum = 0;
847 for (i = 0; i < MAX_NR_ZONES; i++)
848 sum += pgdat->node_zones[i].free_pages;
854 static unsigned int nr_free_zone_pages(int offset)
857 unsigned int sum = 0;
859 for_each_pgdat(pgdat) {
860 struct zonelist *zonelist = pgdat->node_zonelists + offset;
861 struct zone **zonep = zonelist->zones;
864 for (zone = *zonep++; zone; zone = *zonep++) {
865 unsigned long size = zone->present_pages;
866 unsigned long high = zone->pages_high;
876 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
878 unsigned int nr_free_buffer_pages(void)
880 return nr_free_zone_pages(GFP_USER & GFP_ZONEMASK);
884 * Amount of free RAM allocatable within all zones
886 unsigned int nr_free_pagecache_pages(void)
888 return nr_free_zone_pages(GFP_HIGHUSER & GFP_ZONEMASK);
891 #ifdef CONFIG_HIGHMEM
892 unsigned int nr_free_highpages (void)
895 unsigned int pages = 0;
897 for_each_pgdat(pgdat)
898 pages += pgdat->node_zones[ZONE_HIGHMEM].free_pages;
905 static void show_node(struct zone *zone)
907 printk("Node %d ", zone->zone_pgdat->node_id);
910 #define show_node(zone) do { } while (0)
914 * Accumulate the page_state information across all CPUs.
915 * The result is unavoidably approximate - it can change
916 * during and after execution of this function.
918 DEFINE_PER_CPU(struct page_state, page_states) = {0};
919 EXPORT_PER_CPU_SYMBOL(page_states);
921 atomic_t nr_pagecache = ATOMIC_INIT(0);
922 EXPORT_SYMBOL(nr_pagecache);
924 DEFINE_PER_CPU(long, nr_pagecache_local) = 0;
927 void __get_page_state(struct page_state *ret, int nr)
931 memset(ret, 0, sizeof(*ret));
932 while (cpu < NR_CPUS) {
933 unsigned long *in, *out, off;
935 if (!cpu_possible(cpu)) {
940 in = (unsigned long *)&per_cpu(page_states, cpu);
942 if (cpu < NR_CPUS && cpu_possible(cpu))
943 prefetch(&per_cpu(page_states, cpu));
944 out = (unsigned long *)ret;
945 for (off = 0; off < nr; off++)
950 void get_page_state(struct page_state *ret)
954 nr = offsetof(struct page_state, GET_PAGE_STATE_LAST);
955 nr /= sizeof(unsigned long);
957 __get_page_state(ret, nr + 1);
960 void get_full_page_state(struct page_state *ret)
962 __get_page_state(ret, sizeof(*ret) / sizeof(unsigned long));
965 unsigned long __read_page_state(unsigned offset)
967 unsigned long ret = 0;
970 for (cpu = 0; cpu < NR_CPUS; cpu++) {
973 if (!cpu_possible(cpu))
976 in = (unsigned long)&per_cpu(page_states, cpu) + offset;
977 ret += *((unsigned long *)in);
982 void __get_zone_counts(unsigned long *active, unsigned long *inactive,
983 unsigned long *free, struct pglist_data *pgdat)
985 struct zone *zones = pgdat->node_zones;
991 for (i = 0; i < MAX_NR_ZONES; i++) {
992 *active += zones[i].nr_active;
993 *inactive += zones[i].nr_inactive;
994 *free += zones[i].free_pages;
998 void get_zone_counts(unsigned long *active,
999 unsigned long *inactive, unsigned long *free)
1001 struct pglist_data *pgdat;
1006 for_each_pgdat(pgdat) {
1007 unsigned long l, m, n;
1008 __get_zone_counts(&l, &m, &n, pgdat);
1015 void si_meminfo(struct sysinfo *val)
1017 val->totalram = totalram_pages;
1019 val->freeram = nr_free_pages();
1020 val->bufferram = nr_blockdev_pages();
1021 #ifdef CONFIG_HIGHMEM
1022 val->totalhigh = totalhigh_pages;
1023 val->freehigh = nr_free_highpages();
1028 val->mem_unit = PAGE_SIZE;
1029 if (vx_flags(VXF_VIRT_MEM, 0))
1030 vx_vsi_meminfo(val);
1033 EXPORT_SYMBOL(si_meminfo);
1036 void si_meminfo_node(struct sysinfo *val, int nid)
1038 pg_data_t *pgdat = NODE_DATA(nid);
1040 val->totalram = pgdat->node_present_pages;
1041 val->freeram = nr_free_pages_pgdat(pgdat);
1042 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
1043 val->freehigh = pgdat->node_zones[ZONE_HIGHMEM].free_pages;
1044 val->mem_unit = PAGE_SIZE;
1048 #define K(x) ((x) << (PAGE_SHIFT-10))
1051 * Show free area list (used inside shift_scroll-lock stuff)
1052 * We also calculate the percentage fragmentation. We do this by counting the
1053 * memory on each free list with the exception of the first item on the list.
1055 void show_free_areas(void)
1057 struct page_state ps;
1058 int cpu, temperature;
1059 unsigned long active;
1060 unsigned long inactive;
1064 for_each_zone(zone) {
1066 printk("%s per-cpu:", zone->name);
1068 if (!zone->present_pages) {
1074 for (cpu = 0; cpu < NR_CPUS; ++cpu) {
1075 struct per_cpu_pageset *pageset;
1077 if (!cpu_possible(cpu))
1080 pageset = zone->pageset + cpu;
1082 for (temperature = 0; temperature < 2; temperature++)
1083 printk("cpu %d %s: low %d, high %d, batch %d\n",
1085 temperature ? "cold" : "hot",
1086 pageset->pcp[temperature].low,
1087 pageset->pcp[temperature].high,
1088 pageset->pcp[temperature].batch);
1092 get_page_state(&ps);
1093 get_zone_counts(&active, &inactive, &free);
1095 printk("\nFree pages: %11ukB (%ukB HighMem)\n",
1097 K(nr_free_highpages()));
1099 printk("Active:%lu inactive:%lu dirty:%lu writeback:%lu "
1100 "unstable:%lu free:%u slab:%lu mapped:%lu pagetables:%lu\n",
1109 ps.nr_page_table_pages);
1111 for_each_zone(zone) {
1123 " pages_scanned:%lu"
1124 " all_unreclaimable? %s"
1127 K(zone->free_pages),
1130 K(zone->pages_high),
1132 K(zone->nr_inactive),
1133 K(zone->present_pages),
1134 zone->pages_scanned,
1135 (zone->all_unreclaimable ? "yes" : "no")
1137 printk("protections[]:");
1138 for (i = 0; i < MAX_NR_ZONES; i++)
1139 printk(" %lu", zone->protection[i]);
1143 for_each_zone(zone) {
1144 struct list_head *elem;
1145 unsigned long nr, flags, order, total = 0;
1148 printk("%s: ", zone->name);
1149 if (!zone->present_pages) {
1154 spin_lock_irqsave(&zone->lock, flags);
1155 for (order = 0; order < MAX_ORDER; order++) {
1157 list_for_each(elem, &zone->free_area[order].free_list)
1159 total += nr << order;
1160 printk("%lu*%lukB ", nr, K(1UL) << order);
1162 spin_unlock_irqrestore(&zone->lock, flags);
1163 printk("= %lukB\n", K(total));
1166 show_swap_cache_info();
1170 * Builds allocation fallback zone lists.
1172 static int __init build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist, int j, int k)
1179 zone = pgdat->node_zones + ZONE_HIGHMEM;
1180 if (zone->present_pages) {
1181 #ifndef CONFIG_HIGHMEM
1184 zonelist->zones[j++] = zone;
1187 zone = pgdat->node_zones + ZONE_NORMAL;
1188 if (zone->present_pages)
1189 zonelist->zones[j++] = zone;
1191 zone = pgdat->node_zones + ZONE_DMA;
1192 if (zone->present_pages)
1193 zonelist->zones[j++] = zone;
1200 #define MAX_NODE_LOAD (numnodes)
1201 static int __initdata node_load[MAX_NUMNODES];
1203 * find_next_best_node - find the next node that should appear in a given
1204 * node's fallback list
1205 * @node: node whose fallback list we're appending
1206 * @used_node_mask: pointer to the bitmap of already used nodes
1208 * We use a number of factors to determine which is the next node that should
1209 * appear on a given node's fallback list. The node should not have appeared
1210 * already in @node's fallback list, and it should be the next closest node
1211 * according to the distance array (which contains arbitrary distance values
1212 * from each node to each node in the system), and should also prefer nodes
1213 * with no CPUs, since presumably they'll have very little allocation pressure
1214 * on them otherwise.
1215 * It returns -1 if no node is found.
1217 static int __init find_next_best_node(int node, void *used_node_mask)
1220 int min_val = INT_MAX;
1223 for (i = 0; i < numnodes; i++) {
1226 /* Start from local node */
1227 n = (node+i)%numnodes;
1229 /* Don't want a node to appear more than once */
1230 if (test_bit(n, used_node_mask))
1233 /* Use the local node if we haven't already */
1234 if (!test_bit(node, used_node_mask)) {
1239 /* Use the distance array to find the distance */
1240 val = node_distance(node, n);
1242 /* Give preference to headless and unused nodes */
1243 tmp = node_to_cpumask(n);
1244 if (!cpus_empty(tmp))
1245 val += PENALTY_FOR_NODE_WITH_CPUS;
1247 /* Slight preference for less loaded node */
1248 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
1249 val += node_load[n];
1251 if (val < min_val) {
1258 set_bit(best_node, used_node_mask);
1263 static void __init build_zonelists(pg_data_t *pgdat)
1265 int i, j, k, node, local_node;
1266 int prev_node, load;
1267 struct zonelist *zonelist;
1268 DECLARE_BITMAP(used_mask, MAX_NUMNODES);
1270 /* initialize zonelists */
1271 for (i = 0; i < GFP_ZONETYPES; i++) {
1272 zonelist = pgdat->node_zonelists + i;
1273 memset(zonelist, 0, sizeof(*zonelist));
1274 zonelist->zones[0] = NULL;
1277 /* NUMA-aware ordering of nodes */
1278 local_node = pgdat->node_id;
1280 prev_node = local_node;
1281 bitmap_zero(used_mask, MAX_NUMNODES);
1282 while ((node = find_next_best_node(local_node, used_mask)) >= 0) {
1284 * We don't want to pressure a particular node.
1285 * So adding penalty to the first node in same
1286 * distance group to make it round-robin.
1288 if (node_distance(local_node, node) !=
1289 node_distance(local_node, prev_node))
1290 node_load[node] += load;
1293 for (i = 0; i < GFP_ZONETYPES; i++) {
1294 zonelist = pgdat->node_zonelists + i;
1295 for (j = 0; zonelist->zones[j] != NULL; j++);
1298 if (i & __GFP_HIGHMEM)
1303 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1304 zonelist->zones[j] = NULL;
1309 #else /* CONFIG_NUMA */
1311 static void __init build_zonelists(pg_data_t *pgdat)
1313 int i, j, k, node, local_node;
1315 local_node = pgdat->node_id;
1316 for (i = 0; i < GFP_ZONETYPES; i++) {
1317 struct zonelist *zonelist;
1319 zonelist = pgdat->node_zonelists + i;
1320 memset(zonelist, 0, sizeof(*zonelist));
1324 if (i & __GFP_HIGHMEM)
1329 j = build_zonelists_node(pgdat, zonelist, j, k);
1331 * Now we build the zonelist so that it contains the zones
1332 * of all the other nodes.
1333 * We don't want to pressure a particular node, so when
1334 * building the zones for node N, we make sure that the
1335 * zones coming right after the local ones are those from
1336 * node N+1 (modulo N)
1338 for (node = local_node + 1; node < numnodes; node++)
1339 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1340 for (node = 0; node < local_node; node++)
1341 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1343 zonelist->zones[j] = NULL;
1347 #endif /* CONFIG_NUMA */
1349 void __init build_all_zonelists(void)
1353 for(i = 0 ; i < numnodes ; i++)
1354 build_zonelists(NODE_DATA(i));
1355 printk("Built %i zonelists\n", numnodes);
1359 * Helper functions to size the waitqueue hash table.
1360 * Essentially these want to choose hash table sizes sufficiently
1361 * large so that collisions trying to wait on pages are rare.
1362 * But in fact, the number of active page waitqueues on typical
1363 * systems is ridiculously low, less than 200. So this is even
1364 * conservative, even though it seems large.
1366 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
1367 * waitqueues, i.e. the size of the waitq table given the number of pages.
1369 #define PAGES_PER_WAITQUEUE 256
1371 static inline unsigned long wait_table_size(unsigned long pages)
1373 unsigned long size = 1;
1375 pages /= PAGES_PER_WAITQUEUE;
1377 while (size < pages)
1381 * Once we have dozens or even hundreds of threads sleeping
1382 * on IO we've got bigger problems than wait queue collision.
1383 * Limit the size of the wait table to a reasonable size.
1385 size = min(size, 4096UL);
1387 return max(size, 4UL);
1391 * This is an integer logarithm so that shifts can be used later
1392 * to extract the more random high bits from the multiplicative
1393 * hash function before the remainder is taken.
1395 static inline unsigned long wait_table_bits(unsigned long size)
1400 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
1402 static void __init calculate_zone_totalpages(struct pglist_data *pgdat,
1403 unsigned long *zones_size, unsigned long *zholes_size)
1405 unsigned long realtotalpages, totalpages = 0;
1408 for (i = 0; i < MAX_NR_ZONES; i++)
1409 totalpages += zones_size[i];
1410 pgdat->node_spanned_pages = totalpages;
1412 realtotalpages = totalpages;
1414 for (i = 0; i < MAX_NR_ZONES; i++)
1415 realtotalpages -= zholes_size[i];
1416 pgdat->node_present_pages = realtotalpages;
1417 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id, realtotalpages);
1422 * Initially all pages are reserved - free ones are freed
1423 * up by free_all_bootmem() once the early boot process is
1424 * done. Non-atomic initialization, single-pass.
1426 void __init memmap_init_zone(unsigned long size, int nid, unsigned long zone,
1427 unsigned long start_pfn)
1429 struct page *start = pfn_to_page(start_pfn);
1432 for (page = start; page < (start + size); page++) {
1433 set_page_zone(page, NODEZONE(nid, zone));
1434 set_page_count(page, 0);
1435 reset_page_mapcount(page);
1436 SetPageReserved(page);
1437 INIT_LIST_HEAD(&page->lru);
1438 #ifdef WANT_PAGE_VIRTUAL
1439 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1440 if (!is_highmem_idx(zone))
1441 set_page_address(page, __va(start_pfn << PAGE_SHIFT));
1448 * Page buddy system uses "index >> (i+1)", where "index" is
1451 * The extra "+3" is to round down to byte size (8 bits per byte
1452 * assumption). Thus we get "(size-1) >> (i+4)" as the last byte
1455 * The "+1" is because we want to round the byte allocation up
1456 * rather than down. So we should have had a "+7" before we shifted
1457 * down by three. Also, we have to add one as we actually _use_ the
1458 * last bit (it's [0,n] inclusive, not [0,n[).
1460 * So we actually had +7+1 before we shift down by 3. But
1461 * (n+8) >> 3 == (n >> 3) + 1 (modulo overflows, which we do not have).
1463 * Finally, we LONG_ALIGN because all bitmap operations are on longs.
1465 unsigned long pages_to_bitmap_size(unsigned long order, unsigned long nr_pages)
1467 unsigned long bitmap_size;
1469 bitmap_size = (nr_pages-1) >> (order+4);
1470 bitmap_size = LONG_ALIGN(bitmap_size+1);
1475 void zone_init_free_lists(struct pglist_data *pgdat, struct zone *zone, unsigned long size)
1478 for (order = 0; ; order++) {
1479 unsigned long bitmap_size;
1481 INIT_LIST_HEAD(&zone->free_area[order].free_list);
1482 if (order == MAX_ORDER-1) {
1483 zone->free_area[order].map = NULL;
1487 bitmap_size = pages_to_bitmap_size(order, size);
1488 zone->free_area[order].map =
1489 (unsigned long *) alloc_bootmem_node(pgdat, bitmap_size);
1493 #ifndef __HAVE_ARCH_MEMMAP_INIT
1494 #define memmap_init(size, nid, zone, start_pfn) \
1495 memmap_init_zone((size), (nid), (zone), (start_pfn))
1499 * Set up the zone data structures:
1500 * - mark all pages reserved
1501 * - mark all memory queues empty
1502 * - clear the memory bitmaps
1504 static void __init free_area_init_core(struct pglist_data *pgdat,
1505 unsigned long *zones_size, unsigned long *zholes_size)
1508 const unsigned long zone_required_alignment = 1UL << (MAX_ORDER-1);
1509 int cpu, nid = pgdat->node_id;
1510 unsigned long zone_start_pfn = pgdat->node_start_pfn;
1512 pgdat->nr_zones = 0;
1513 init_waitqueue_head(&pgdat->kswapd_wait);
1515 for (j = 0; j < MAX_NR_ZONES; j++) {
1516 struct zone *zone = pgdat->node_zones + j;
1517 unsigned long size, realsize;
1518 unsigned long batch;
1520 zone_table[NODEZONE(nid, j)] = zone;
1521 realsize = size = zones_size[j];
1523 realsize -= zholes_size[j];
1525 if (j == ZONE_DMA || j == ZONE_NORMAL)
1526 nr_kernel_pages += realsize;
1527 nr_all_pages += realsize;
1529 zone->spanned_pages = size;
1530 zone->present_pages = realsize;
1531 zone->name = zone_names[j];
1532 spin_lock_init(&zone->lock);
1533 spin_lock_init(&zone->lru_lock);
1534 zone->zone_pgdat = pgdat;
1535 zone->free_pages = 0;
1537 zone->temp_priority = zone->prev_priority = DEF_PRIORITY;
1540 * The per-cpu-pages pools are set to around 1000th of the
1541 * size of the zone. But no more than 1/4 of a meg - there's
1542 * no point in going beyond the size of L2 cache.
1544 * OK, so we don't know how big the cache is. So guess.
1546 batch = zone->present_pages / 1024;
1547 if (batch * PAGE_SIZE > 256 * 1024)
1548 batch = (256 * 1024) / PAGE_SIZE;
1549 batch /= 4; /* We effectively *= 4 below */
1553 for (cpu = 0; cpu < NR_CPUS; cpu++) {
1554 struct per_cpu_pages *pcp;
1556 pcp = &zone->pageset[cpu].pcp[0]; /* hot */
1558 pcp->low = 2 * batch;
1559 pcp->high = 6 * batch;
1560 pcp->batch = 1 * batch;
1561 INIT_LIST_HEAD(&pcp->list);
1563 pcp = &zone->pageset[cpu].pcp[1]; /* cold */
1566 pcp->high = 2 * batch;
1567 pcp->batch = 1 * batch;
1568 INIT_LIST_HEAD(&pcp->list);
1570 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%lu\n",
1571 zone_names[j], realsize, batch);
1572 INIT_LIST_HEAD(&zone->active_list);
1573 INIT_LIST_HEAD(&zone->inactive_list);
1574 zone->nr_scan_active = 0;
1575 zone->nr_scan_inactive = 0;
1576 zone->nr_active = 0;
1577 zone->nr_inactive = 0;
1582 * The per-page waitqueue mechanism uses hashed waitqueues
1585 zone->wait_table_size = wait_table_size(size);
1586 zone->wait_table_bits =
1587 wait_table_bits(zone->wait_table_size);
1588 zone->wait_table = (wait_queue_head_t *)
1589 alloc_bootmem_node(pgdat, zone->wait_table_size
1590 * sizeof(wait_queue_head_t));
1592 for(i = 0; i < zone->wait_table_size; ++i)
1593 init_waitqueue_head(zone->wait_table + i);
1595 pgdat->nr_zones = j+1;
1597 zone->zone_mem_map = pfn_to_page(zone_start_pfn);
1598 zone->zone_start_pfn = zone_start_pfn;
1600 if ((zone_start_pfn) & (zone_required_alignment-1))
1601 printk("BUG: wrong zone alignment, it will crash\n");
1603 memmap_init(size, nid, j, zone_start_pfn);
1605 zone_start_pfn += size;
1607 zone_init_free_lists(pgdat, zone, zone->spanned_pages);
1611 void __init node_alloc_mem_map(struct pglist_data *pgdat)
1615 size = (pgdat->node_spanned_pages + 1) * sizeof(struct page);
1616 pgdat->node_mem_map = alloc_bootmem_node(pgdat, size);
1617 #ifndef CONFIG_DISCONTIGMEM
1618 mem_map = contig_page_data.node_mem_map;
1622 void __init free_area_init_node(int nid, struct pglist_data *pgdat,
1623 unsigned long *zones_size, unsigned long node_start_pfn,
1624 unsigned long *zholes_size)
1626 pgdat->node_id = nid;
1627 pgdat->node_start_pfn = node_start_pfn;
1628 calculate_zone_totalpages(pgdat, zones_size, zholes_size);
1630 if (!pfn_to_page(node_start_pfn))
1631 node_alloc_mem_map(pgdat);
1633 free_area_init_core(pgdat, zones_size, zholes_size);
1636 #ifndef CONFIG_DISCONTIGMEM
1637 static bootmem_data_t contig_bootmem_data;
1638 struct pglist_data contig_page_data = { .bdata = &contig_bootmem_data };
1640 EXPORT_SYMBOL(contig_page_data);
1642 void __init free_area_init(unsigned long *zones_size)
1644 free_area_init_node(0, &contig_page_data, zones_size,
1645 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
1649 #ifdef CONFIG_PROC_FS
1651 #include <linux/seq_file.h>
1653 static void *frag_start(struct seq_file *m, loff_t *pos)
1658 for (pgdat = pgdat_list; pgdat && node; pgdat = pgdat->pgdat_next)
1664 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
1666 pg_data_t *pgdat = (pg_data_t *)arg;
1669 return pgdat->pgdat_next;
1672 static void frag_stop(struct seq_file *m, void *arg)
1677 * This walks the freelist for each zone. Whilst this is slow, I'd rather
1678 * be slow here than slow down the fast path by keeping stats - mjbligh
1680 static int frag_show(struct seq_file *m, void *arg)
1682 pg_data_t *pgdat = (pg_data_t *)arg;
1684 struct zone *node_zones = pgdat->node_zones;
1685 unsigned long flags;
1688 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
1689 if (!zone->present_pages)
1692 spin_lock_irqsave(&zone->lock, flags);
1693 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1694 for (order = 0; order < MAX_ORDER; ++order) {
1695 unsigned long nr_bufs = 0;
1696 struct list_head *elem;
1698 list_for_each(elem, &(zone->free_area[order].free_list))
1700 seq_printf(m, "%6lu ", nr_bufs);
1702 spin_unlock_irqrestore(&zone->lock, flags);
1708 struct seq_operations fragmentation_op = {
1709 .start = frag_start,
1715 static char *vmstat_text[] = {
1719 "nr_page_table_pages",
1744 "pgscan_kswapd_high",
1745 "pgscan_kswapd_normal",
1747 "pgscan_kswapd_dma",
1748 "pgscan_direct_high",
1749 "pgscan_direct_normal",
1750 "pgscan_direct_dma",
1755 "kswapd_inodesteal",
1762 static void *vmstat_start(struct seq_file *m, loff_t *pos)
1764 struct page_state *ps;
1766 if (*pos >= ARRAY_SIZE(vmstat_text))
1769 ps = kmalloc(sizeof(*ps), GFP_KERNEL);
1772 return ERR_PTR(-ENOMEM);
1773 get_full_page_state(ps);
1774 ps->pgpgin /= 2; /* sectors -> kbytes */
1776 return (unsigned long *)ps + *pos;
1779 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
1782 if (*pos >= ARRAY_SIZE(vmstat_text))
1784 return (unsigned long *)m->private + *pos;
1787 static int vmstat_show(struct seq_file *m, void *arg)
1789 unsigned long *l = arg;
1790 unsigned long off = l - (unsigned long *)m->private;
1792 seq_printf(m, "%s %lu\n", vmstat_text[off], *l);
1796 static void vmstat_stop(struct seq_file *m, void *arg)
1802 struct seq_operations vmstat_op = {
1803 .start = vmstat_start,
1804 .next = vmstat_next,
1805 .stop = vmstat_stop,
1806 .show = vmstat_show,
1809 #endif /* CONFIG_PROC_FS */
1811 #ifdef CONFIG_HOTPLUG_CPU
1812 static int page_alloc_cpu_notify(struct notifier_block *self,
1813 unsigned long action, void *hcpu)
1815 int cpu = (unsigned long)hcpu;
1818 if (action == CPU_DEAD) {
1819 /* Drain local pagecache count. */
1820 count = &per_cpu(nr_pagecache_local, cpu);
1821 atomic_add(*count, &nr_pagecache);
1823 local_irq_disable();
1829 #endif /* CONFIG_HOTPLUG_CPU */
1831 void __init page_alloc_init(void)
1833 hotcpu_notifier(page_alloc_cpu_notify, 0);
1836 static unsigned long higherzone_val(struct zone *z, int max_zone,
1839 int z_idx = zone_idx(z);
1840 struct zone *higherzone;
1841 unsigned long pages;
1843 /* there is no higher zone to get a contribution from */
1844 if (z_idx == MAX_NR_ZONES-1)
1847 higherzone = &z->zone_pgdat->node_zones[z_idx+1];
1849 /* We always start with the higher zone's protection value */
1850 pages = higherzone->protection[alloc_type];
1853 * We get a lower-zone-protection contribution only if there are
1854 * pages in the higher zone and if we're not the highest zone
1855 * in the current zonelist. e.g., never happens for GFP_DMA. Happens
1856 * only for ZONE_DMA in a GFP_KERNEL allocation and happens for ZONE_DMA
1857 * and ZONE_NORMAL for a GFP_HIGHMEM allocation.
1859 if (higherzone->present_pages && z_idx < alloc_type)
1860 pages += higherzone->pages_low * sysctl_lower_zone_protection;
1866 * setup_per_zone_protection - called whenver min_free_kbytes or
1867 * sysctl_lower_zone_protection changes. Ensures that each zone
1868 * has a correct pages_protected value, so an adequate number of
1869 * pages are left in the zone after a successful __alloc_pages().
1871 * This algorithm is way confusing. I tries to keep the same behavior
1872 * as we had with the incremental min iterative algorithm.
1874 static void setup_per_zone_protection(void)
1876 struct pglist_data *pgdat;
1877 struct zone *zones, *zone;
1881 for_each_pgdat(pgdat) {
1882 zones = pgdat->node_zones;
1884 for (i = 0, max_zone = 0; i < MAX_NR_ZONES; i++)
1885 if (zones[i].present_pages)
1889 * For each of the different allocation types:
1890 * GFP_DMA -> GFP_KERNEL -> GFP_HIGHMEM
1892 for (i = 0; i < GFP_ZONETYPES; i++) {
1894 * For each of the zones:
1895 * ZONE_HIGHMEM -> ZONE_NORMAL -> ZONE_DMA
1897 for (j = MAX_NR_ZONES-1; j >= 0; j--) {
1901 * We never protect zones that don't have memory
1902 * in them (j>max_zone) or zones that aren't in
1903 * the zonelists for a certain type of
1904 * allocation (j>=i). We have to assign these
1905 * to zero because the lower zones take
1906 * contributions from the higher zones.
1908 if (j > max_zone || j >= i) {
1909 zone->protection[i] = 0;
1913 * The contribution of the next higher zone
1915 zone->protection[i] = higherzone_val(zone,
1923 * setup_per_zone_pages_min - called when min_free_kbytes changes. Ensures
1924 * that the pages_{min,low,high} values for each zone are set correctly
1925 * with respect to min_free_kbytes.
1927 static void setup_per_zone_pages_min(void)
1929 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
1930 unsigned long lowmem_pages = 0;
1932 unsigned long flags;
1934 /* Calculate total number of !ZONE_HIGHMEM pages */
1935 for_each_zone(zone) {
1936 if (!is_highmem(zone))
1937 lowmem_pages += zone->present_pages;
1940 for_each_zone(zone) {
1941 spin_lock_irqsave(&zone->lru_lock, flags);
1942 if (is_highmem(zone)) {
1944 * Often, highmem doesn't need to reserve any pages.
1945 * But the pages_min/low/high values are also used for
1946 * batching up page reclaim activity so we need a
1947 * decent value here.
1951 min_pages = zone->present_pages / 1024;
1952 if (min_pages < SWAP_CLUSTER_MAX)
1953 min_pages = SWAP_CLUSTER_MAX;
1954 if (min_pages > 128)
1956 zone->pages_min = min_pages;
1958 /* if it's a lowmem zone, reserve a number of pages
1959 * proportionate to the zone's size.
1961 zone->pages_min = (pages_min * zone->present_pages) /
1966 * When interpreting these watermarks, just keep in mind that:
1967 * zone->pages_min == (zone->pages_min * 4) / 4;
1969 zone->pages_low = (zone->pages_min * 5) / 4;
1970 zone->pages_high = (zone->pages_min * 6) / 4;
1971 spin_unlock_irqrestore(&zone->lru_lock, flags);
1976 * Initialise min_free_kbytes.
1978 * For small machines we want it small (128k min). For large machines
1979 * we want it large (64MB max). But it is not linear, because network
1980 * bandwidth does not increase linearly with machine size. We use
1982 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
1983 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
1999 static int __init init_per_zone_pages_min(void)
2001 unsigned long lowmem_kbytes;
2003 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
2005 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
2006 if (min_free_kbytes < 128)
2007 min_free_kbytes = 128;
2008 if (min_free_kbytes > 65536)
2009 min_free_kbytes = 65536;
2010 setup_per_zone_pages_min();
2011 setup_per_zone_protection();
2014 module_init(init_per_zone_pages_min)
2017 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
2018 * that we can call two helper functions whenever min_free_kbytes
2021 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
2022 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2024 proc_dointvec(table, write, file, buffer, length, ppos);
2025 setup_per_zone_pages_min();
2026 setup_per_zone_protection();
2031 * lower_zone_protection_sysctl_handler - just a wrapper around
2032 * proc_dointvec() so that we can call setup_per_zone_protection()
2033 * whenever sysctl_lower_zone_protection changes.
2035 int lower_zone_protection_sysctl_handler(ctl_table *table, int write,
2036 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2038 proc_dointvec_minmax(table, write, file, buffer, length, ppos);
2039 setup_per_zone_protection();
2044 * allocate a large system hash table from bootmem
2045 * - it is assumed that the hash table must contain an exact power-of-2
2046 * quantity of entries
2048 void *__init alloc_large_system_hash(const char *tablename,
2049 unsigned long bucketsize,
2050 unsigned long numentries,
2052 int consider_highmem,
2053 unsigned int *_hash_shift,
2054 unsigned int *_hash_mask)
2056 unsigned long long max;
2057 unsigned long log2qty, size;
2060 /* allow the kernel cmdline to have a say */
2062 /* round applicable memory size up to nearest megabyte */
2063 numentries = consider_highmem ? nr_all_pages : nr_kernel_pages;
2064 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
2065 numentries >>= 20 - PAGE_SHIFT;
2066 numentries <<= 20 - PAGE_SHIFT;
2068 /* limit to 1 bucket per 2^scale bytes of low memory */
2069 if (scale > PAGE_SHIFT)
2070 numentries >>= (scale - PAGE_SHIFT);
2072 numentries <<= (PAGE_SHIFT - scale);
2074 /* rounded up to nearest power of 2 in size */
2075 numentries = 1UL << (long_log2(numentries) + 1);
2077 /* limit allocation size to 1/16 total memory */
2078 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
2079 do_div(max, bucketsize);
2081 if (numentries > max)
2084 log2qty = long_log2(numentries);
2087 size = bucketsize << log2qty;
2088 table = alloc_bootmem(size);
2089 } while (!table && size > PAGE_SIZE && --log2qty);
2092 panic("Failed to allocate %s hash table\n", tablename);
2094 printk("%s hash table entries: %d (order: %d, %lu bytes)\n",
2097 long_log2(size) - PAGE_SHIFT,
2101 *_hash_shift = log2qty;
2103 *_hash_mask = (1 << log2qty) - 1;