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/ckrm_mem_inline.h>
35 #include <linux/vs_base.h>
36 #include <linux/vs_limit.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 #ifdef CONFIG_CKRM_RES_MEM
371 1 << PG_ckrm_account |
373 1 << PG_checked | 1 << PG_mappedtodisk);
375 #ifdef CONFIG_CKRM_RES_MEM
376 page->ckrm_zone = NULL;
378 set_page_refs(page, order);
382 * Do the hard work of removing an element from the buddy allocator.
383 * Call me with the zone->lock already held.
385 static struct page *__rmqueue(struct zone *zone, unsigned int order)
387 struct free_area * area;
388 unsigned int current_order;
392 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
393 area = zone->free_area + current_order;
394 if (list_empty(&area->free_list))
397 page = list_entry(area->free_list.next, struct page, lru);
398 list_del(&page->lru);
399 index = page - zone->zone_mem_map;
400 if (current_order != MAX_ORDER-1)
401 MARK_USED(index, current_order, area);
402 zone->free_pages -= 1UL << order;
403 return expand(zone, page, index, order, current_order, area);
410 * Obtain a specified number of elements from the buddy allocator, all under
411 * a single hold of the lock, for efficiency. Add them to the supplied list.
412 * Returns the number of new pages which were placed at *list.
414 static int rmqueue_bulk(struct zone *zone, unsigned int order,
415 unsigned long count, struct list_head *list)
422 spin_lock_irqsave(&zone->lock, flags);
423 for (i = 0; i < count; ++i) {
424 page = __rmqueue(zone, order);
428 list_add_tail(&page->lru, list);
430 spin_unlock_irqrestore(&zone->lock, flags);
434 #if defined(CONFIG_PM) || defined(CONFIG_HOTPLUG_CPU)
435 static void __drain_pages(unsigned int cpu)
440 for_each_zone(zone) {
441 struct per_cpu_pageset *pset;
443 pset = &zone->pageset[cpu];
444 for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
445 struct per_cpu_pages *pcp;
448 pcp->count -= free_pages_bulk(zone, pcp->count,
453 #endif /* CONFIG_PM || CONFIG_HOTPLUG_CPU */
456 int is_head_of_free_region(struct page *page)
458 struct zone *zone = page_zone(page);
461 struct list_head *curr;
464 * Should not matter as we need quiescent system for
465 * suspend anyway, but...
467 spin_lock_irqsave(&zone->lock, flags);
468 for (order = MAX_ORDER - 1; order >= 0; --order)
469 list_for_each(curr, &zone->free_area[order].free_list)
470 if (page == list_entry(curr, struct page, lru)) {
471 spin_unlock_irqrestore(&zone->lock, flags);
474 spin_unlock_irqrestore(&zone->lock, flags);
479 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
481 void drain_local_pages(void)
485 local_irq_save(flags);
486 __drain_pages(smp_processor_id());
487 local_irq_restore(flags);
489 #endif /* CONFIG_PM */
491 static void zone_statistics(struct zonelist *zonelist, struct zone *z)
496 pg_data_t *pg = z->zone_pgdat;
497 pg_data_t *orig = zonelist->zones[0]->zone_pgdat;
498 struct per_cpu_pageset *p;
500 local_irq_save(flags);
501 cpu = smp_processor_id();
502 p = &z->pageset[cpu];
504 z->pageset[cpu].numa_hit++;
507 zonelist->zones[0]->pageset[cpu].numa_foreign++;
509 if (pg == NODE_DATA(numa_node_id()))
513 local_irq_restore(flags);
518 * Free a 0-order page
520 static void FASTCALL(free_hot_cold_page(struct page *page, int cold));
521 static void fastcall free_hot_cold_page(struct page *page, int cold)
523 struct zone *zone = page_zone(page);
524 struct per_cpu_pages *pcp;
527 arch_free_page(page, 0);
529 kernel_map_pages(page, 1, 0);
530 inc_page_state(pgfree);
532 page->mapping = NULL;
533 free_pages_check(__FUNCTION__, page);
534 pcp = &zone->pageset[get_cpu()].pcp[cold];
535 local_irq_save(flags);
536 if (pcp->count >= pcp->high)
537 pcp->count -= free_pages_bulk(zone, pcp->batch, &pcp->list, 0);
538 list_add(&page->lru, &pcp->list);
540 local_irq_restore(flags);
544 void fastcall free_hot_page(struct page *page)
546 free_hot_cold_page(page, 0);
549 void fastcall free_cold_page(struct page *page)
551 free_hot_cold_page(page, 1);
555 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
556 * we cheat by calling it from here, in the order > 0 path. Saves a branch
561 buffered_rmqueue(struct zone *zone, int order, int gfp_flags)
564 struct page *page = NULL;
565 int cold = !!(gfp_flags & __GFP_COLD);
568 struct per_cpu_pages *pcp;
570 pcp = &zone->pageset[get_cpu()].pcp[cold];
571 local_irq_save(flags);
572 if (pcp->count <= pcp->low)
573 pcp->count += rmqueue_bulk(zone, 0,
574 pcp->batch, &pcp->list);
576 page = list_entry(pcp->list.next, struct page, lru);
577 list_del(&page->lru);
580 local_irq_restore(flags);
585 spin_lock_irqsave(&zone->lock, flags);
586 page = __rmqueue(zone, order);
587 spin_unlock_irqrestore(&zone->lock, flags);
591 BUG_ON(bad_range(zone, page));
592 mod_page_state_zone(zone, pgalloc, 1 << order);
593 prep_new_page(page, order);
594 if (order && (gfp_flags & __GFP_COMP))
595 prep_compound_page(page, order);
601 * This is the 'heart' of the zoned buddy allocator.
603 * Herein lies the mysterious "incremental min". That's the
605 * local_low = z->pages_low;
608 * thing. The intent here is to provide additional protection to low zones for
609 * allocation requests which _could_ use higher zones. So a GFP_HIGHMEM
610 * request is not allowed to dip as deeply into the normal zone as a GFP_KERNEL
611 * request. This preserves additional space in those lower zones for requests
612 * which really do need memory from those zones. It means that on a decent
613 * sized machine, GFP_HIGHMEM and GFP_KERNEL requests basically leave the DMA
616 struct page * fastcall
617 __alloc_pages(unsigned int gfp_mask, unsigned int order,
618 struct zonelist *zonelist)
620 const int wait = gfp_mask & __GFP_WAIT;
622 struct zone **zones, *z;
624 struct reclaim_state reclaim_state;
625 struct task_struct *p = current;
631 might_sleep_if(wait);
634 * The caller may dip into page reserves a bit more if the caller
635 * cannot run direct reclaim, or is the caller has realtime scheduling
638 can_try_harder = (unlikely(rt_task(p)) && !in_interrupt()) || !wait;
640 if (!ckrm_class_limit_ok((ckrm_get_mem_class(current)))) {
644 zones = zonelist->zones; /* the list of zones suitable for gfp_mask */
646 if (unlikely(zones[0] == NULL)) {
647 /* Should this ever happen?? */
651 alloc_type = zone_idx(zones[0]);
653 /* Go through the zonelist once, looking for a zone with enough free */
654 for (i = 0; (z = zones[i]) != NULL; i++) {
655 min = z->pages_low + (1<<order) + z->protection[alloc_type];
657 if (z->free_pages < min)
660 page = buffered_rmqueue(z, order, gfp_mask);
665 for (i = 0; (z = zones[i]) != NULL; i++)
669 * Go through the zonelist again. Let __GFP_HIGH and allocations
670 * coming from realtime tasks to go deeper into reserves
672 for (i = 0; (z = zones[i]) != NULL; i++) {
674 if (gfp_mask & __GFP_HIGH)
678 min += (1<<order) + z->protection[alloc_type];
680 if (z->free_pages < min)
683 page = buffered_rmqueue(z, order, gfp_mask);
688 /* This allocation should allow future memory freeing. */
689 if ((p->flags & (PF_MEMALLOC | PF_MEMDIE)) && !in_interrupt()) {
690 /* go through the zonelist yet again, ignoring mins */
691 for (i = 0; (z = zones[i]) != NULL; i++) {
692 page = buffered_rmqueue(z, order, gfp_mask);
699 /* Atomic allocations - we can't balance anything */
704 /* We now go into synchronous reclaim */
705 p->flags |= PF_MEMALLOC;
706 reclaim_state.reclaimed_slab = 0;
707 p->reclaim_state = &reclaim_state;
709 try_to_free_pages(zones, gfp_mask, order);
711 p->reclaim_state = NULL;
712 p->flags &= ~PF_MEMALLOC;
714 /* go through the zonelist yet one more time */
715 for (i = 0; (z = zones[i]) != NULL; i++) {
717 if (gfp_mask & __GFP_HIGH)
721 min += (1<<order) + z->protection[alloc_type];
723 if (z->free_pages < min)
726 page = buffered_rmqueue(z, order, gfp_mask);
732 * Don't let big-order allocations loop unless the caller explicitly
733 * requests that. Wait for some write requests to complete then retry.
735 * In this implementation, __GFP_REPEAT means __GFP_NOFAIL for order
736 * <= 3, but that may not be true in other implementations.
739 if (!(gfp_mask & __GFP_NORETRY)) {
740 if ((order <= 3) || (gfp_mask & __GFP_REPEAT))
742 if (gfp_mask & __GFP_NOFAIL)
746 blk_congestion_wait(WRITE, HZ/50);
751 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
752 printk(KERN_WARNING "%s: page allocation failure."
753 " order:%d, mode:0x%x\n",
754 p->comm, order, gfp_mask);
759 zone_statistics(zonelist, z);
760 kernel_map_pages(page, 1 << order, 1);
764 EXPORT_SYMBOL(__alloc_pages);
767 * Common helper functions.
769 fastcall unsigned long __get_free_pages(unsigned int gfp_mask, unsigned int order)
772 page = alloc_pages(gfp_mask, order);
775 return (unsigned long) page_address(page);
778 EXPORT_SYMBOL(__get_free_pages);
780 fastcall unsigned long get_zeroed_page(unsigned int gfp_mask)
785 * get_zeroed_page() returns a 32-bit address, which cannot represent
788 BUG_ON(gfp_mask & __GFP_HIGHMEM);
790 page = alloc_pages(gfp_mask, 0);
792 void *address = page_address(page);
794 return (unsigned long) address;
799 EXPORT_SYMBOL(get_zeroed_page);
801 void __pagevec_free(struct pagevec *pvec)
803 int i = pagevec_count(pvec);
806 free_hot_cold_page(pvec->pages[i], pvec->cold);
809 fastcall void __free_pages(struct page *page, unsigned int order)
811 if (!PageReserved(page) && put_page_testzero(page)) {
815 __free_pages_ok(page, order);
819 EXPORT_SYMBOL(__free_pages);
821 fastcall void free_pages(unsigned long addr, unsigned int order)
824 BUG_ON(!virt_addr_valid((void *)addr));
825 __free_pages(virt_to_page((void *)addr), order);
829 EXPORT_SYMBOL(free_pages);
832 * Total amount of free (allocatable) RAM:
834 unsigned int nr_free_pages(void)
836 unsigned int sum = 0;
840 sum += zone->free_pages;
845 EXPORT_SYMBOL(nr_free_pages);
848 unsigned int nr_free_pages_pgdat(pg_data_t *pgdat)
850 unsigned int i, sum = 0;
852 for (i = 0; i < MAX_NR_ZONES; i++)
853 sum += pgdat->node_zones[i].free_pages;
859 static unsigned int nr_free_zone_pages(int offset)
862 unsigned int sum = 0;
864 for_each_pgdat(pgdat) {
865 struct zonelist *zonelist = pgdat->node_zonelists + offset;
866 struct zone **zonep = zonelist->zones;
869 for (zone = *zonep++; zone; zone = *zonep++) {
870 unsigned long size = zone->present_pages;
871 unsigned long high = zone->pages_high;
881 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
883 unsigned int nr_free_buffer_pages(void)
885 return nr_free_zone_pages(GFP_USER & GFP_ZONEMASK);
889 * Amount of free RAM allocatable within all zones
891 unsigned int nr_free_pagecache_pages(void)
893 return nr_free_zone_pages(GFP_HIGHUSER & GFP_ZONEMASK);
896 #ifdef CONFIG_HIGHMEM
897 unsigned int nr_free_highpages (void)
900 unsigned int pages = 0;
902 for_each_pgdat(pgdat)
903 pages += pgdat->node_zones[ZONE_HIGHMEM].free_pages;
910 static void show_node(struct zone *zone)
912 printk("Node %d ", zone->zone_pgdat->node_id);
915 #define show_node(zone) do { } while (0)
919 * Accumulate the page_state information across all CPUs.
920 * The result is unavoidably approximate - it can change
921 * during and after execution of this function.
923 DEFINE_PER_CPU(struct page_state, page_states) = {0};
924 EXPORT_PER_CPU_SYMBOL(page_states);
926 atomic_t nr_pagecache = ATOMIC_INIT(0);
927 EXPORT_SYMBOL(nr_pagecache);
929 DEFINE_PER_CPU(long, nr_pagecache_local) = 0;
932 void __get_page_state(struct page_state *ret, int nr)
936 memset(ret, 0, sizeof(*ret));
937 while (cpu < NR_CPUS) {
938 unsigned long *in, *out, off;
940 if (!cpu_possible(cpu)) {
945 in = (unsigned long *)&per_cpu(page_states, cpu);
947 if (cpu < NR_CPUS && cpu_possible(cpu))
948 prefetch(&per_cpu(page_states, cpu));
949 out = (unsigned long *)ret;
950 for (off = 0; off < nr; off++)
955 void get_page_state(struct page_state *ret)
959 nr = offsetof(struct page_state, GET_PAGE_STATE_LAST);
960 nr /= sizeof(unsigned long);
962 __get_page_state(ret, nr + 1);
965 void get_full_page_state(struct page_state *ret)
967 __get_page_state(ret, sizeof(*ret) / sizeof(unsigned long));
970 unsigned long __read_page_state(unsigned offset)
972 unsigned long ret = 0;
975 for (cpu = 0; cpu < NR_CPUS; cpu++) {
978 if (!cpu_possible(cpu))
981 in = (unsigned long)&per_cpu(page_states, cpu) + offset;
982 ret += *((unsigned long *)in);
987 void __get_zone_counts(unsigned long *active, unsigned long *inactive,
988 unsigned long *free, struct pglist_data *pgdat)
990 struct zone *zones = pgdat->node_zones;
996 for (i = 0; i < MAX_NR_ZONES; i++) {
997 *active += zones[i].nr_active;
998 *inactive += zones[i].nr_inactive;
999 *free += zones[i].free_pages;
1003 void get_zone_counts(unsigned long *active,
1004 unsigned long *inactive, unsigned long *free)
1006 struct pglist_data *pgdat;
1011 for_each_pgdat(pgdat) {
1012 unsigned long l, m, n;
1013 __get_zone_counts(&l, &m, &n, pgdat);
1020 void si_meminfo(struct sysinfo *val)
1022 val->totalram = totalram_pages;
1024 val->freeram = nr_free_pages();
1025 val->bufferram = nr_blockdev_pages();
1026 #ifdef CONFIG_HIGHMEM
1027 val->totalhigh = totalhigh_pages;
1028 val->freehigh = nr_free_highpages();
1033 val->mem_unit = PAGE_SIZE;
1034 if (vx_flags(VXF_VIRT_MEM, 0))
1035 vx_vsi_meminfo(val);
1038 EXPORT_SYMBOL(si_meminfo);
1041 void si_meminfo_node(struct sysinfo *val, int nid)
1043 pg_data_t *pgdat = NODE_DATA(nid);
1045 val->totalram = pgdat->node_present_pages;
1046 val->freeram = nr_free_pages_pgdat(pgdat);
1047 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
1048 val->freehigh = pgdat->node_zones[ZONE_HIGHMEM].free_pages;
1049 val->mem_unit = PAGE_SIZE;
1053 #define K(x) ((x) << (PAGE_SHIFT-10))
1056 * Show free area list (used inside shift_scroll-lock stuff)
1057 * We also calculate the percentage fragmentation. We do this by counting the
1058 * memory on each free list with the exception of the first item on the list.
1060 void show_free_areas(void)
1062 struct page_state ps;
1063 int cpu, temperature;
1064 unsigned long active;
1065 unsigned long inactive;
1069 for_each_zone(zone) {
1071 printk("%s per-cpu:", zone->name);
1073 if (!zone->present_pages) {
1079 for (cpu = 0; cpu < NR_CPUS; ++cpu) {
1080 struct per_cpu_pageset *pageset;
1082 if (!cpu_possible(cpu))
1085 pageset = zone->pageset + cpu;
1087 for (temperature = 0; temperature < 2; temperature++)
1088 printk("cpu %d %s: low %d, high %d, batch %d\n",
1090 temperature ? "cold" : "hot",
1091 pageset->pcp[temperature].low,
1092 pageset->pcp[temperature].high,
1093 pageset->pcp[temperature].batch);
1097 get_page_state(&ps);
1098 get_zone_counts(&active, &inactive, &free);
1100 printk("\nFree pages: %11ukB (%ukB HighMem)\n",
1102 K(nr_free_highpages()));
1104 printk("Active:%lu inactive:%lu dirty:%lu writeback:%lu "
1105 "unstable:%lu free:%u slab:%lu mapped:%lu pagetables:%lu\n",
1114 ps.nr_page_table_pages);
1116 for_each_zone(zone) {
1128 " pages_scanned:%lu"
1129 " all_unreclaimable? %s"
1132 K(zone->free_pages),
1135 K(zone->pages_high),
1137 K(zone->nr_inactive),
1138 K(zone->present_pages),
1139 zone->pages_scanned,
1140 (zone->all_unreclaimable ? "yes" : "no")
1142 printk("protections[]:");
1143 for (i = 0; i < MAX_NR_ZONES; i++)
1144 printk(" %lu", zone->protection[i]);
1148 for_each_zone(zone) {
1149 struct list_head *elem;
1150 unsigned long nr, flags, order, total = 0;
1153 printk("%s: ", zone->name);
1154 if (!zone->present_pages) {
1159 spin_lock_irqsave(&zone->lock, flags);
1160 for (order = 0; order < MAX_ORDER; order++) {
1162 list_for_each(elem, &zone->free_area[order].free_list)
1164 total += nr << order;
1165 printk("%lu*%lukB ", nr, K(1UL) << order);
1167 spin_unlock_irqrestore(&zone->lock, flags);
1168 printk("= %lukB\n", K(total));
1171 show_swap_cache_info();
1175 * Builds allocation fallback zone lists.
1177 static int __init build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist, int j, int k)
1184 zone = pgdat->node_zones + ZONE_HIGHMEM;
1185 if (zone->present_pages) {
1186 #ifndef CONFIG_HIGHMEM
1189 zonelist->zones[j++] = zone;
1192 zone = pgdat->node_zones + ZONE_NORMAL;
1193 if (zone->present_pages)
1194 zonelist->zones[j++] = zone;
1196 zone = pgdat->node_zones + ZONE_DMA;
1197 if (zone->present_pages)
1198 zonelist->zones[j++] = zone;
1205 #define MAX_NODE_LOAD (numnodes)
1206 static int __initdata node_load[MAX_NUMNODES];
1208 * find_next_best_node - find the next node that should appear in a given
1209 * node's fallback list
1210 * @node: node whose fallback list we're appending
1211 * @used_node_mask: pointer to the bitmap of already used nodes
1213 * We use a number of factors to determine which is the next node that should
1214 * appear on a given node's fallback list. The node should not have appeared
1215 * already in @node's fallback list, and it should be the next closest node
1216 * according to the distance array (which contains arbitrary distance values
1217 * from each node to each node in the system), and should also prefer nodes
1218 * with no CPUs, since presumably they'll have very little allocation pressure
1219 * on them otherwise.
1220 * It returns -1 if no node is found.
1222 static int __init find_next_best_node(int node, void *used_node_mask)
1225 int min_val = INT_MAX;
1228 for (i = 0; i < numnodes; i++) {
1231 /* Start from local node */
1232 n = (node+i)%numnodes;
1234 /* Don't want a node to appear more than once */
1235 if (test_bit(n, used_node_mask))
1238 /* Use the local node if we haven't already */
1239 if (!test_bit(node, used_node_mask)) {
1244 /* Use the distance array to find the distance */
1245 val = node_distance(node, n);
1247 /* Give preference to headless and unused nodes */
1248 tmp = node_to_cpumask(n);
1249 if (!cpus_empty(tmp))
1250 val += PENALTY_FOR_NODE_WITH_CPUS;
1252 /* Slight preference for less loaded node */
1253 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
1254 val += node_load[n];
1256 if (val < min_val) {
1263 set_bit(best_node, used_node_mask);
1268 static void __init build_zonelists(pg_data_t *pgdat)
1270 int i, j, k, node, local_node;
1271 int prev_node, load;
1272 struct zonelist *zonelist;
1273 DECLARE_BITMAP(used_mask, MAX_NUMNODES);
1275 /* initialize zonelists */
1276 for (i = 0; i < GFP_ZONETYPES; i++) {
1277 zonelist = pgdat->node_zonelists + i;
1278 memset(zonelist, 0, sizeof(*zonelist));
1279 zonelist->zones[0] = NULL;
1282 /* NUMA-aware ordering of nodes */
1283 local_node = pgdat->node_id;
1285 prev_node = local_node;
1286 bitmap_zero(used_mask, MAX_NUMNODES);
1287 while ((node = find_next_best_node(local_node, used_mask)) >= 0) {
1289 * We don't want to pressure a particular node.
1290 * So adding penalty to the first node in same
1291 * distance group to make it round-robin.
1293 if (node_distance(local_node, node) !=
1294 node_distance(local_node, prev_node))
1295 node_load[node] += load;
1298 for (i = 0; i < GFP_ZONETYPES; i++) {
1299 zonelist = pgdat->node_zonelists + i;
1300 for (j = 0; zonelist->zones[j] != NULL; j++);
1303 if (i & __GFP_HIGHMEM)
1308 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1309 zonelist->zones[j] = NULL;
1314 #else /* CONFIG_NUMA */
1316 static void __init build_zonelists(pg_data_t *pgdat)
1318 int i, j, k, node, local_node;
1320 local_node = pgdat->node_id;
1321 for (i = 0; i < GFP_ZONETYPES; i++) {
1322 struct zonelist *zonelist;
1324 zonelist = pgdat->node_zonelists + i;
1325 memset(zonelist, 0, sizeof(*zonelist));
1329 if (i & __GFP_HIGHMEM)
1334 j = build_zonelists_node(pgdat, zonelist, j, k);
1336 * Now we build the zonelist so that it contains the zones
1337 * of all the other nodes.
1338 * We don't want to pressure a particular node, so when
1339 * building the zones for node N, we make sure that the
1340 * zones coming right after the local ones are those from
1341 * node N+1 (modulo N)
1343 for (node = local_node + 1; node < numnodes; node++)
1344 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1345 for (node = 0; node < local_node; node++)
1346 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1348 zonelist->zones[j] = NULL;
1352 #endif /* CONFIG_NUMA */
1354 void __init build_all_zonelists(void)
1358 for(i = 0 ; i < numnodes ; i++)
1359 build_zonelists(NODE_DATA(i));
1360 printk("Built %i zonelists\n", numnodes);
1364 * Helper functions to size the waitqueue hash table.
1365 * Essentially these want to choose hash table sizes sufficiently
1366 * large so that collisions trying to wait on pages are rare.
1367 * But in fact, the number of active page waitqueues on typical
1368 * systems is ridiculously low, less than 200. So this is even
1369 * conservative, even though it seems large.
1371 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
1372 * waitqueues, i.e. the size of the waitq table given the number of pages.
1374 #define PAGES_PER_WAITQUEUE 256
1376 static inline unsigned long wait_table_size(unsigned long pages)
1378 unsigned long size = 1;
1380 pages /= PAGES_PER_WAITQUEUE;
1382 while (size < pages)
1386 * Once we have dozens or even hundreds of threads sleeping
1387 * on IO we've got bigger problems than wait queue collision.
1388 * Limit the size of the wait table to a reasonable size.
1390 size = min(size, 4096UL);
1392 return max(size, 4UL);
1396 * This is an integer logarithm so that shifts can be used later
1397 * to extract the more random high bits from the multiplicative
1398 * hash function before the remainder is taken.
1400 static inline unsigned long wait_table_bits(unsigned long size)
1405 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
1407 static void __init calculate_zone_totalpages(struct pglist_data *pgdat,
1408 unsigned long *zones_size, unsigned long *zholes_size)
1410 unsigned long realtotalpages, totalpages = 0;
1413 for (i = 0; i < MAX_NR_ZONES; i++)
1414 totalpages += zones_size[i];
1415 pgdat->node_spanned_pages = totalpages;
1417 realtotalpages = totalpages;
1419 for (i = 0; i < MAX_NR_ZONES; i++)
1420 realtotalpages -= zholes_size[i];
1421 pgdat->node_present_pages = realtotalpages;
1422 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id, realtotalpages);
1427 * Initially all pages are reserved - free ones are freed
1428 * up by free_all_bootmem() once the early boot process is
1429 * done. Non-atomic initialization, single-pass.
1431 void __init memmap_init_zone(unsigned long size, int nid, unsigned long zone,
1432 unsigned long start_pfn)
1434 struct page *start = pfn_to_page(start_pfn);
1437 for (page = start; page < (start + size); page++) {
1438 set_page_zone(page, NODEZONE(nid, zone));
1439 set_page_count(page, 0);
1440 reset_page_mapcount(page);
1441 SetPageReserved(page);
1442 INIT_LIST_HEAD(&page->lru);
1443 #ifdef WANT_PAGE_VIRTUAL
1444 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1445 if (!is_highmem_idx(zone))
1446 set_page_address(page, __va(start_pfn << PAGE_SHIFT));
1453 * Page buddy system uses "index >> (i+1)", where "index" is
1456 * The extra "+3" is to round down to byte size (8 bits per byte
1457 * assumption). Thus we get "(size-1) >> (i+4)" as the last byte
1460 * The "+1" is because we want to round the byte allocation up
1461 * rather than down. So we should have had a "+7" before we shifted
1462 * down by three. Also, we have to add one as we actually _use_ the
1463 * last bit (it's [0,n] inclusive, not [0,n[).
1465 * So we actually had +7+1 before we shift down by 3. But
1466 * (n+8) >> 3 == (n >> 3) + 1 (modulo overflows, which we do not have).
1468 * Finally, we LONG_ALIGN because all bitmap operations are on longs.
1470 unsigned long pages_to_bitmap_size(unsigned long order, unsigned long nr_pages)
1472 unsigned long bitmap_size;
1474 bitmap_size = (nr_pages-1) >> (order+4);
1475 bitmap_size = LONG_ALIGN(bitmap_size+1);
1480 void zone_init_free_lists(struct pglist_data *pgdat, struct zone *zone, unsigned long size)
1483 for (order = 0; ; order++) {
1484 unsigned long bitmap_size;
1486 INIT_LIST_HEAD(&zone->free_area[order].free_list);
1487 if (order == MAX_ORDER-1) {
1488 zone->free_area[order].map = NULL;
1492 bitmap_size = pages_to_bitmap_size(order, size);
1493 zone->free_area[order].map =
1494 (unsigned long *) alloc_bootmem_node(pgdat, bitmap_size);
1498 #ifndef __HAVE_ARCH_MEMMAP_INIT
1499 #define memmap_init(size, nid, zone, start_pfn) \
1500 memmap_init_zone((size), (nid), (zone), (start_pfn))
1504 * Set up the zone data structures:
1505 * - mark all pages reserved
1506 * - mark all memory queues empty
1507 * - clear the memory bitmaps
1509 static void __init free_area_init_core(struct pglist_data *pgdat,
1510 unsigned long *zones_size, unsigned long *zholes_size)
1513 const unsigned long zone_required_alignment = 1UL << (MAX_ORDER-1);
1514 int cpu, nid = pgdat->node_id;
1515 unsigned long zone_start_pfn = pgdat->node_start_pfn;
1517 pgdat->nr_zones = 0;
1518 init_waitqueue_head(&pgdat->kswapd_wait);
1520 for (j = 0; j < MAX_NR_ZONES; j++) {
1521 struct zone *zone = pgdat->node_zones + j;
1522 unsigned long size, realsize;
1523 unsigned long batch;
1525 zone_table[NODEZONE(nid, j)] = zone;
1526 realsize = size = zones_size[j];
1528 realsize -= zholes_size[j];
1530 if (j == ZONE_DMA || j == ZONE_NORMAL)
1531 nr_kernel_pages += realsize;
1532 nr_all_pages += realsize;
1534 zone->spanned_pages = size;
1535 zone->present_pages = realsize;
1536 zone->name = zone_names[j];
1537 spin_lock_init(&zone->lock);
1538 spin_lock_init(&zone->lru_lock);
1539 zone->zone_pgdat = pgdat;
1540 zone->free_pages = 0;
1542 zone->temp_priority = zone->prev_priority = DEF_PRIORITY;
1545 * The per-cpu-pages pools are set to around 1000th of the
1546 * size of the zone. But no more than 1/4 of a meg - there's
1547 * no point in going beyond the size of L2 cache.
1549 * OK, so we don't know how big the cache is. So guess.
1551 batch = zone->present_pages / 1024;
1552 if (batch * PAGE_SIZE > 256 * 1024)
1553 batch = (256 * 1024) / PAGE_SIZE;
1554 batch /= 4; /* We effectively *= 4 below */
1558 for (cpu = 0; cpu < NR_CPUS; cpu++) {
1559 struct per_cpu_pages *pcp;
1561 pcp = &zone->pageset[cpu].pcp[0]; /* hot */
1563 pcp->low = 2 * batch;
1564 pcp->high = 6 * batch;
1565 pcp->batch = 1 * batch;
1566 INIT_LIST_HEAD(&pcp->list);
1568 pcp = &zone->pageset[cpu].pcp[1]; /* cold */
1571 pcp->high = 2 * batch;
1572 pcp->batch = 1 * batch;
1573 INIT_LIST_HEAD(&pcp->list);
1575 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%lu\n",
1576 zone_names[j], realsize, batch);
1577 #ifndef CONFIG_CKRM_RES_MEM
1578 INIT_LIST_HEAD(&zone->active_list);
1579 INIT_LIST_HEAD(&zone->inactive_list);
1581 zone->nr_scan_active = 0;
1582 zone->nr_scan_inactive = 0;
1583 zone->nr_active = 0;
1584 zone->nr_inactive = 0;
1589 * The per-page waitqueue mechanism uses hashed waitqueues
1592 zone->wait_table_size = wait_table_size(size);
1593 zone->wait_table_bits =
1594 wait_table_bits(zone->wait_table_size);
1595 zone->wait_table = (wait_queue_head_t *)
1596 alloc_bootmem_node(pgdat, zone->wait_table_size
1597 * sizeof(wait_queue_head_t));
1599 for(i = 0; i < zone->wait_table_size; ++i)
1600 init_waitqueue_head(zone->wait_table + i);
1602 pgdat->nr_zones = j+1;
1604 zone->zone_mem_map = pfn_to_page(zone_start_pfn);
1605 zone->zone_start_pfn = zone_start_pfn;
1607 if ((zone_start_pfn) & (zone_required_alignment-1))
1608 printk("BUG: wrong zone alignment, it will crash\n");
1610 memmap_init(size, nid, j, zone_start_pfn);
1612 zone_start_pfn += size;
1614 zone_init_free_lists(pgdat, zone, zone->spanned_pages);
1618 void __init node_alloc_mem_map(struct pglist_data *pgdat)
1622 size = (pgdat->node_spanned_pages + 1) * sizeof(struct page);
1623 pgdat->node_mem_map = alloc_bootmem_node(pgdat, size);
1624 #ifndef CONFIG_DISCONTIGMEM
1625 mem_map = contig_page_data.node_mem_map;
1629 void __init free_area_init_node(int nid, struct pglist_data *pgdat,
1630 unsigned long *zones_size, unsigned long node_start_pfn,
1631 unsigned long *zholes_size)
1633 pgdat->node_id = nid;
1634 pgdat->node_start_pfn = node_start_pfn;
1635 calculate_zone_totalpages(pgdat, zones_size, zholes_size);
1637 if (!pfn_to_page(node_start_pfn))
1638 node_alloc_mem_map(pgdat);
1640 free_area_init_core(pgdat, zones_size, zholes_size);
1643 #ifndef CONFIG_DISCONTIGMEM
1644 static bootmem_data_t contig_bootmem_data;
1645 struct pglist_data contig_page_data = { .bdata = &contig_bootmem_data };
1647 EXPORT_SYMBOL(contig_page_data);
1649 void __init free_area_init(unsigned long *zones_size)
1651 free_area_init_node(0, &contig_page_data, zones_size,
1652 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
1656 #ifdef CONFIG_PROC_FS
1658 #include <linux/seq_file.h>
1660 static void *frag_start(struct seq_file *m, loff_t *pos)
1665 for (pgdat = pgdat_list; pgdat && node; pgdat = pgdat->pgdat_next)
1671 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
1673 pg_data_t *pgdat = (pg_data_t *)arg;
1676 return pgdat->pgdat_next;
1679 static void frag_stop(struct seq_file *m, void *arg)
1684 * This walks the freelist for each zone. Whilst this is slow, I'd rather
1685 * be slow here than slow down the fast path by keeping stats - mjbligh
1687 static int frag_show(struct seq_file *m, void *arg)
1689 pg_data_t *pgdat = (pg_data_t *)arg;
1691 struct zone *node_zones = pgdat->node_zones;
1692 unsigned long flags;
1695 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
1696 if (!zone->present_pages)
1699 spin_lock_irqsave(&zone->lock, flags);
1700 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1701 for (order = 0; order < MAX_ORDER; ++order) {
1702 unsigned long nr_bufs = 0;
1703 struct list_head *elem;
1705 list_for_each(elem, &(zone->free_area[order].free_list))
1707 seq_printf(m, "%6lu ", nr_bufs);
1709 spin_unlock_irqrestore(&zone->lock, flags);
1715 struct seq_operations fragmentation_op = {
1716 .start = frag_start,
1722 static char *vmstat_text[] = {
1726 "nr_page_table_pages",
1751 "pgscan_kswapd_high",
1752 "pgscan_kswapd_normal",
1754 "pgscan_kswapd_dma",
1755 "pgscan_direct_high",
1756 "pgscan_direct_normal",
1757 "pgscan_direct_dma",
1762 "kswapd_inodesteal",
1769 static void *vmstat_start(struct seq_file *m, loff_t *pos)
1771 struct page_state *ps;
1773 if (*pos >= ARRAY_SIZE(vmstat_text))
1776 ps = kmalloc(sizeof(*ps), GFP_KERNEL);
1779 return ERR_PTR(-ENOMEM);
1780 get_full_page_state(ps);
1781 ps->pgpgin /= 2; /* sectors -> kbytes */
1783 return (unsigned long *)ps + *pos;
1786 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
1789 if (*pos >= ARRAY_SIZE(vmstat_text))
1791 return (unsigned long *)m->private + *pos;
1794 static int vmstat_show(struct seq_file *m, void *arg)
1796 unsigned long *l = arg;
1797 unsigned long off = l - (unsigned long *)m->private;
1799 seq_printf(m, "%s %lu\n", vmstat_text[off], *l);
1803 static void vmstat_stop(struct seq_file *m, void *arg)
1809 struct seq_operations vmstat_op = {
1810 .start = vmstat_start,
1811 .next = vmstat_next,
1812 .stop = vmstat_stop,
1813 .show = vmstat_show,
1816 #endif /* CONFIG_PROC_FS */
1818 #ifdef CONFIG_HOTPLUG_CPU
1819 static int page_alloc_cpu_notify(struct notifier_block *self,
1820 unsigned long action, void *hcpu)
1822 int cpu = (unsigned long)hcpu;
1825 if (action == CPU_DEAD) {
1826 /* Drain local pagecache count. */
1827 count = &per_cpu(nr_pagecache_local, cpu);
1828 atomic_add(*count, &nr_pagecache);
1830 local_irq_disable();
1836 #endif /* CONFIG_HOTPLUG_CPU */
1838 void __init page_alloc_init(void)
1840 hotcpu_notifier(page_alloc_cpu_notify, 0);
1843 static unsigned long higherzone_val(struct zone *z, int max_zone,
1846 int z_idx = zone_idx(z);
1847 struct zone *higherzone;
1848 unsigned long pages;
1850 /* there is no higher zone to get a contribution from */
1851 if (z_idx == MAX_NR_ZONES-1)
1854 higherzone = &z->zone_pgdat->node_zones[z_idx+1];
1856 /* We always start with the higher zone's protection value */
1857 pages = higherzone->protection[alloc_type];
1860 * We get a lower-zone-protection contribution only if there are
1861 * pages in the higher zone and if we're not the highest zone
1862 * in the current zonelist. e.g., never happens for GFP_DMA. Happens
1863 * only for ZONE_DMA in a GFP_KERNEL allocation and happens for ZONE_DMA
1864 * and ZONE_NORMAL for a GFP_HIGHMEM allocation.
1866 if (higherzone->present_pages && z_idx < alloc_type)
1867 pages += higherzone->pages_low * sysctl_lower_zone_protection;
1873 * setup_per_zone_protection - called whenver min_free_kbytes or
1874 * sysctl_lower_zone_protection changes. Ensures that each zone
1875 * has a correct pages_protected value, so an adequate number of
1876 * pages are left in the zone after a successful __alloc_pages().
1878 * This algorithm is way confusing. I tries to keep the same behavior
1879 * as we had with the incremental min iterative algorithm.
1881 static void setup_per_zone_protection(void)
1883 struct pglist_data *pgdat;
1884 struct zone *zones, *zone;
1888 for_each_pgdat(pgdat) {
1889 zones = pgdat->node_zones;
1891 for (i = 0, max_zone = 0; i < MAX_NR_ZONES; i++)
1892 if (zones[i].present_pages)
1896 * For each of the different allocation types:
1897 * GFP_DMA -> GFP_KERNEL -> GFP_HIGHMEM
1899 for (i = 0; i < GFP_ZONETYPES; i++) {
1901 * For each of the zones:
1902 * ZONE_HIGHMEM -> ZONE_NORMAL -> ZONE_DMA
1904 for (j = MAX_NR_ZONES-1; j >= 0; j--) {
1908 * We never protect zones that don't have memory
1909 * in them (j>max_zone) or zones that aren't in
1910 * the zonelists for a certain type of
1911 * allocation (j>=i). We have to assign these
1912 * to zero because the lower zones take
1913 * contributions from the higher zones.
1915 if (j > max_zone || j >= i) {
1916 zone->protection[i] = 0;
1920 * The contribution of the next higher zone
1922 zone->protection[i] = higherzone_val(zone,
1930 * setup_per_zone_pages_min - called when min_free_kbytes changes. Ensures
1931 * that the pages_{min,low,high} values for each zone are set correctly
1932 * with respect to min_free_kbytes.
1934 static void setup_per_zone_pages_min(void)
1936 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
1937 unsigned long lowmem_pages = 0;
1939 unsigned long flags;
1941 /* Calculate total number of !ZONE_HIGHMEM pages */
1942 for_each_zone(zone) {
1943 if (!is_highmem(zone))
1944 lowmem_pages += zone->present_pages;
1947 for_each_zone(zone) {
1948 spin_lock_irqsave(&zone->lru_lock, flags);
1949 if (is_highmem(zone)) {
1951 * Often, highmem doesn't need to reserve any pages.
1952 * But the pages_min/low/high values are also used for
1953 * batching up page reclaim activity so we need a
1954 * decent value here.
1958 min_pages = zone->present_pages / 1024;
1959 if (min_pages < SWAP_CLUSTER_MAX)
1960 min_pages = SWAP_CLUSTER_MAX;
1961 if (min_pages > 128)
1963 zone->pages_min = min_pages;
1965 /* if it's a lowmem zone, reserve a number of pages
1966 * proportionate to the zone's size.
1968 zone->pages_min = (pages_min * zone->present_pages) /
1973 * When interpreting these watermarks, just keep in mind that:
1974 * zone->pages_min == (zone->pages_min * 4) / 4;
1976 zone->pages_low = (zone->pages_min * 5) / 4;
1977 zone->pages_high = (zone->pages_min * 6) / 4;
1978 spin_unlock_irqrestore(&zone->lru_lock, flags);
1983 * Initialise min_free_kbytes.
1985 * For small machines we want it small (128k min). For large machines
1986 * we want it large (64MB max). But it is not linear, because network
1987 * bandwidth does not increase linearly with machine size. We use
1989 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
1990 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
2006 static int __init init_per_zone_pages_min(void)
2008 unsigned long lowmem_kbytes;
2010 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
2012 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
2013 if (min_free_kbytes < 128)
2014 min_free_kbytes = 128;
2015 if (min_free_kbytes > 65536)
2016 min_free_kbytes = 65536;
2017 setup_per_zone_pages_min();
2018 setup_per_zone_protection();
2021 module_init(init_per_zone_pages_min)
2024 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
2025 * that we can call two helper functions whenever min_free_kbytes
2028 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
2029 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2031 proc_dointvec(table, write, file, buffer, length, ppos);
2032 setup_per_zone_pages_min();
2033 setup_per_zone_protection();
2038 * lower_zone_protection_sysctl_handler - just a wrapper around
2039 * proc_dointvec() so that we can call setup_per_zone_protection()
2040 * whenever sysctl_lower_zone_protection changes.
2042 int lower_zone_protection_sysctl_handler(ctl_table *table, int write,
2043 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2045 proc_dointvec_minmax(table, write, file, buffer, length, ppos);
2046 setup_per_zone_protection();
2051 * allocate a large system hash table from bootmem
2052 * - it is assumed that the hash table must contain an exact power-of-2
2053 * quantity of entries
2055 void *__init alloc_large_system_hash(const char *tablename,
2056 unsigned long bucketsize,
2057 unsigned long numentries,
2059 int consider_highmem,
2060 unsigned int *_hash_shift,
2061 unsigned int *_hash_mask)
2063 unsigned long long max;
2064 unsigned long log2qty, size;
2067 /* allow the kernel cmdline to have a say */
2069 /* round applicable memory size up to nearest megabyte */
2070 numentries = consider_highmem ? nr_all_pages : nr_kernel_pages;
2071 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
2072 numentries >>= 20 - PAGE_SHIFT;
2073 numentries <<= 20 - PAGE_SHIFT;
2075 /* limit to 1 bucket per 2^scale bytes of low memory */
2076 if (scale > PAGE_SHIFT)
2077 numentries >>= (scale - PAGE_SHIFT);
2079 numentries <<= (PAGE_SHIFT - scale);
2081 /* rounded up to nearest power of 2 in size */
2082 numentries = 1UL << (long_log2(numentries) + 1);
2084 /* limit allocation size to 1/16 total memory */
2085 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
2086 do_div(max, bucketsize);
2088 if (numentries > max)
2091 log2qty = long_log2(numentries);
2094 size = bucketsize << log2qty;
2095 table = alloc_bootmem(size);
2096 } while (!table && size > PAGE_SIZE && --log2qty);
2099 panic("Failed to allocate %s hash table\n", tablename);
2101 printk("%s hash table entries: %d (order: %d, %lu bytes)\n",
2104 long_log2(size) - PAGE_SHIFT,
2108 *_hash_shift = log2qty;
2110 *_hash_mask = (1 << log2qty) - 1;