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
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 #define prep_compound_page(page, order) do { } while (0)
110 #define destroy_compound_page(page, order) do { } while (0)
113 * Higher-order pages are called "compound pages". They are structured thusly:
115 * The first PAGE_SIZE page is called the "head page".
117 * The remaining PAGE_SIZE pages are called "tail pages".
119 * All pages have PG_compound set. All pages have their ->private pointing at
120 * the head page (even the head page has this).
122 * The first tail page's ->mapping, if non-zero, holds the address of the
123 * compound page's put_page() function.
125 * The order of the allocation is stored in the first tail page's ->index
126 * This is only for debug at present. This usage means that zero-order pages
127 * may not be compound.
129 static void prep_compound_page(struct page *page, unsigned long order)
132 int nr_pages = 1 << order;
134 page[1].mapping = NULL;
135 page[1].index = order;
136 for (i = 0; i < nr_pages; i++) {
137 struct page *p = page + i;
140 p->private = (unsigned long)page;
144 static void destroy_compound_page(struct page *page, unsigned long order)
147 int nr_pages = 1 << order;
149 if (!PageCompound(page))
152 if (page[1].index != order)
153 bad_page(__FUNCTION__, page);
155 for (i = 0; i < nr_pages; i++) {
156 struct page *p = page + i;
158 if (!PageCompound(p))
159 bad_page(__FUNCTION__, page);
160 if (p->private != (unsigned long)page)
161 bad_page(__FUNCTION__, page);
162 ClearPageCompound(p);
165 #endif /* CONFIG_HUGETLB_PAGE */
168 * Freeing function for a buddy system allocator.
170 * The concept of a buddy system is to maintain direct-mapped table
171 * (containing bit values) for memory blocks of various "orders".
172 * The bottom level table contains the map for the smallest allocatable
173 * units of memory (here, pages), and each level above it describes
174 * pairs of units from the levels below, hence, "buddies".
175 * At a high level, all that happens here is marking the table entry
176 * at the bottom level available, and propagating the changes upward
177 * as necessary, plus some accounting needed to play nicely with other
178 * parts of the VM system.
179 * At each level, we keep one bit for each pair of blocks, which
180 * is set to 1 iff only one of the pair is allocated. So when we
181 * are allocating or freeing one, we can derive the state of the
182 * other. That is, if we allocate a small block, and both were
183 * free, the remainder of the region must be split into blocks.
184 * If a block is freed, and its buddy is also free, then this
185 * triggers coalescing into a block of larger size.
190 static inline void __free_pages_bulk (struct page *page, struct page *base,
191 struct zone *zone, struct free_area *area, unsigned int order)
193 unsigned long page_idx, index, mask;
196 destroy_compound_page(page, order);
197 mask = (~0UL) << order;
198 page_idx = page - base;
199 if (page_idx & ~mask)
201 index = page_idx >> (1 + order);
203 zone->free_pages += 1 << order;
204 while (order < MAX_ORDER-1) {
205 struct page *buddy1, *buddy2;
207 BUG_ON(area >= zone->free_area + MAX_ORDER);
208 if (!__test_and_change_bit(index, area->map))
210 * the buddy page is still allocated.
214 /* Move the buddy up one level. */
215 buddy1 = base + (page_idx ^ (1 << order));
216 buddy2 = base + page_idx;
217 BUG_ON(bad_range(zone, buddy1));
218 BUG_ON(bad_range(zone, buddy2));
219 list_del(&buddy1->lru);
226 list_add(&(base + page_idx)->lru, &area->free_list);
229 static inline void free_pages_check(const char *function, struct page *page)
231 if ( page_mapped(page) ||
232 page->mapping != NULL ||
233 page_count(page) != 0 ||
242 1 << PG_writeback )))
243 bad_page(function, page);
245 ClearPageDirty(page);
249 * Frees a list of pages.
250 * Assumes all pages on list are in same zone, and of same order.
251 * count is the number of pages to free, or 0 for all on the list.
253 * If the zone was previously in an "all pages pinned" state then look to
254 * see if this freeing clears that state.
256 * And clear the zone's pages_scanned counter, to hold off the "all pages are
257 * pinned" detection logic.
260 free_pages_bulk(struct zone *zone, int count,
261 struct list_head *list, unsigned int order)
264 struct free_area *area;
265 struct page *base, *page = NULL;
268 base = zone->zone_mem_map;
269 area = zone->free_area + order;
270 spin_lock_irqsave(&zone->lock, flags);
271 zone->all_unreclaimable = 0;
272 zone->pages_scanned = 0;
273 while (!list_empty(list) && count--) {
274 page = list_entry(list->prev, struct page, lru);
275 /* have to delete it as __free_pages_bulk list manipulates */
276 list_del(&page->lru);
277 __free_pages_bulk(page, base, zone, area, order);
278 ckrm_clear_page_class(page);
281 spin_unlock_irqrestore(&zone->lock, flags);
285 void __free_pages_ok(struct page *page, unsigned int order)
290 arch_free_page(page, order);
292 mod_page_state(pgfree, 1 << order);
293 for (i = 0 ; i < (1 << order) ; ++i)
294 free_pages_check(__FUNCTION__, page + i);
295 list_add(&page->lru, &list);
296 kernel_map_pages(page, 1<<order, 0);
297 free_pages_bulk(page_zone(page), 1, &list, order);
300 #define MARK_USED(index, order, area) \
301 __change_bit((index) >> (1+(order)), (area)->map)
304 * The order of subdivision here is critical for the IO subsystem.
305 * Please do not alter this order without good reasons and regression
306 * testing. Specifically, as large blocks of memory are subdivided,
307 * the order in which smaller blocks are delivered depends on the order
308 * they're subdivided in this function. This is the primary factor
309 * influencing the order in which pages are delivered to the IO
310 * subsystem according to empirical testing, and this is also justified
311 * by considering the behavior of a buddy system containing a single
312 * large block of memory acted on by a series of small allocations.
313 * This behavior is a critical factor in sglist merging's success.
317 static inline struct page *
318 expand(struct zone *zone, struct page *page,
319 unsigned long index, int low, int high, struct free_area *area)
321 unsigned long size = 1 << high;
327 BUG_ON(bad_range(zone, &page[size]));
328 list_add(&page[size].lru, &area->free_list);
329 MARK_USED(index + size, high, area);
334 static inline void set_page_refs(struct page *page, int order)
337 set_page_count(page, 1);
342 * We need to reference all the pages for this order, otherwise if
343 * anyone accesses one of the pages with (get/put) it will be freed.
345 for (i = 0; i < (1 << order); i++)
346 set_page_count(page+i, 1);
347 #endif /* CONFIG_MMU */
351 * This page is about to be returned from the page allocator
353 static void prep_new_page(struct page *page, int order)
355 if (page->mapping || page_mapped(page) ||
364 1 << PG_writeback )))
365 bad_page(__FUNCTION__, page);
367 page->flags &= ~(1 << PG_uptodate | 1 << PG_error |
368 1 << PG_referenced | 1 << PG_arch_1 |
369 #ifdef CONFIG_CKRM_RES_MEM
370 1 << PG_ckrm_account |
372 1 << PG_checked | 1 << PG_mappedtodisk);
374 #ifdef CONFIG_CKRM_RES_MEM
375 page->ckrm_zone = NULL;
377 set_page_refs(page, order);
381 * Do the hard work of removing an element from the buddy allocator.
382 * Call me with the zone->lock already held.
384 static struct page *__rmqueue(struct zone *zone, unsigned int order)
386 struct free_area * area;
387 unsigned int current_order;
391 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
392 area = zone->free_area + current_order;
393 if (list_empty(&area->free_list))
396 page = list_entry(area->free_list.next, struct page, lru);
397 list_del(&page->lru);
398 index = page - zone->zone_mem_map;
399 if (current_order != MAX_ORDER-1)
400 MARK_USED(index, current_order, area);
401 zone->free_pages -= 1UL << order;
402 return expand(zone, page, index, order, current_order, area);
409 * Obtain a specified number of elements from the buddy allocator, all under
410 * a single hold of the lock, for efficiency. Add them to the supplied list.
411 * Returns the number of new pages which were placed at *list.
413 static int rmqueue_bulk(struct zone *zone, unsigned int order,
414 unsigned long count, struct list_head *list)
421 spin_lock_irqsave(&zone->lock, flags);
422 for (i = 0; i < count; ++i) {
423 page = __rmqueue(zone, order);
427 list_add_tail(&page->lru, list);
429 spin_unlock_irqrestore(&zone->lock, flags);
433 #if defined(CONFIG_PM) || defined(CONFIG_HOTPLUG_CPU)
434 static void __drain_pages(unsigned int cpu)
439 for_each_zone(zone) {
440 struct per_cpu_pageset *pset;
442 pset = &zone->pageset[cpu];
443 for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
444 struct per_cpu_pages *pcp;
447 pcp->count -= free_pages_bulk(zone, pcp->count,
452 #endif /* CONFIG_PM || CONFIG_HOTPLUG_CPU */
455 int is_head_of_free_region(struct page *page)
457 struct zone *zone = page_zone(page);
460 struct list_head *curr;
463 * Should not matter as we need quiescent system for
464 * suspend anyway, but...
466 spin_lock_irqsave(&zone->lock, flags);
467 for (order = MAX_ORDER - 1; order >= 0; --order)
468 list_for_each(curr, &zone->free_area[order].free_list)
469 if (page == list_entry(curr, struct page, lru)) {
470 spin_unlock_irqrestore(&zone->lock, flags);
473 spin_unlock_irqrestore(&zone->lock, flags);
478 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
480 void drain_local_pages(void)
484 local_irq_save(flags);
485 __drain_pages(smp_processor_id());
486 local_irq_restore(flags);
488 #endif /* CONFIG_PM */
490 static void zone_statistics(struct zonelist *zonelist, struct zone *z)
495 pg_data_t *pg = z->zone_pgdat;
496 pg_data_t *orig = zonelist->zones[0]->zone_pgdat;
497 struct per_cpu_pageset *p;
499 local_irq_save(flags);
500 cpu = smp_processor_id();
501 p = &z->pageset[cpu];
503 z->pageset[cpu].numa_hit++;
506 zonelist->zones[0]->pageset[cpu].numa_foreign++;
508 if (pg == NODE_DATA(numa_node_id()))
512 local_irq_restore(flags);
517 * Free a 0-order page
519 static void FASTCALL(free_hot_cold_page(struct page *page, int cold));
520 static void fastcall free_hot_cold_page(struct page *page, int cold)
522 struct zone *zone = page_zone(page);
523 struct per_cpu_pages *pcp;
526 arch_free_page(page, 0);
528 kernel_map_pages(page, 1, 0);
529 inc_page_state(pgfree);
531 page->mapping = NULL;
532 free_pages_check(__FUNCTION__, page);
533 pcp = &zone->pageset[get_cpu()].pcp[cold];
534 local_irq_save(flags);
535 if (pcp->count >= pcp->high)
536 pcp->count -= free_pages_bulk(zone, pcp->batch, &pcp->list, 0);
537 list_add(&page->lru, &pcp->list);
539 local_irq_restore(flags);
543 void fastcall free_hot_page(struct page *page)
545 free_hot_cold_page(page, 0);
548 void fastcall free_cold_page(struct page *page)
550 free_hot_cold_page(page, 1);
554 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
555 * we cheat by calling it from here, in the order > 0 path. Saves a branch
560 buffered_rmqueue(struct zone *zone, int order, int gfp_flags)
563 struct page *page = NULL;
564 int cold = !!(gfp_flags & __GFP_COLD);
567 struct per_cpu_pages *pcp;
569 pcp = &zone->pageset[get_cpu()].pcp[cold];
570 local_irq_save(flags);
571 if (pcp->count <= pcp->low)
572 pcp->count += rmqueue_bulk(zone, 0,
573 pcp->batch, &pcp->list);
575 page = list_entry(pcp->list.next, struct page, lru);
576 list_del(&page->lru);
579 local_irq_restore(flags);
584 spin_lock_irqsave(&zone->lock, flags);
585 page = __rmqueue(zone, order);
586 spin_unlock_irqrestore(&zone->lock, flags);
590 BUG_ON(bad_range(zone, page));
591 mod_page_state_zone(zone, pgalloc, 1 << order);
592 prep_new_page(page, order);
593 if (order && (gfp_flags & __GFP_COMP))
594 prep_compound_page(page, order);
600 * This is the 'heart' of the zoned buddy allocator.
602 * Herein lies the mysterious "incremental min". That's the
604 * local_low = z->pages_low;
607 * thing. The intent here is to provide additional protection to low zones for
608 * allocation requests which _could_ use higher zones. So a GFP_HIGHMEM
609 * request is not allowed to dip as deeply into the normal zone as a GFP_KERNEL
610 * request. This preserves additional space in those lower zones for requests
611 * which really do need memory from those zones. It means that on a decent
612 * sized machine, GFP_HIGHMEM and GFP_KERNEL requests basically leave the DMA
615 struct page * fastcall
616 __alloc_pages(unsigned int gfp_mask, unsigned int order,
617 struct zonelist *zonelist)
619 const int wait = gfp_mask & __GFP_WAIT;
621 struct zone **zones, *z;
623 struct reclaim_state reclaim_state;
624 struct task_struct *p = current;
630 might_sleep_if(wait);
633 * The caller may dip into page reserves a bit more if the caller
634 * cannot run direct reclaim, or is the caller has realtime scheduling
637 can_try_harder = (unlikely(rt_task(p)) && !in_interrupt()) || !wait;
639 if (!ckrm_class_limit_ok((ckrm_get_mem_class(current)))) {
643 zones = zonelist->zones; /* the list of zones suitable for gfp_mask */
645 if (unlikely(zones[0] == NULL)) {
646 /* Should this ever happen?? */
650 alloc_type = zone_idx(zones[0]);
652 /* Go through the zonelist once, looking for a zone with enough free */
653 for (i = 0; (z = zones[i]) != NULL; i++) {
654 min = z->pages_low + (1<<order) + z->protection[alloc_type];
656 if (z->free_pages < min)
659 page = buffered_rmqueue(z, order, gfp_mask);
664 for (i = 0; (z = zones[i]) != NULL; i++)
668 * Go through the zonelist again. Let __GFP_HIGH and allocations
669 * coming from realtime tasks to go deeper into reserves
671 for (i = 0; (z = zones[i]) != NULL; i++) {
673 if (gfp_mask & __GFP_HIGH)
677 min += (1<<order) + z->protection[alloc_type];
679 if (z->free_pages < min)
682 page = buffered_rmqueue(z, order, gfp_mask);
687 /* This allocation should allow future memory freeing. */
688 if ((p->flags & (PF_MEMALLOC | PF_MEMDIE)) && !in_interrupt()) {
689 /* go through the zonelist yet again, ignoring mins */
690 for (i = 0; (z = zones[i]) != NULL; i++) {
691 page = buffered_rmqueue(z, order, gfp_mask);
698 /* Atomic allocations - we can't balance anything */
703 /* We now go into synchronous reclaim */
704 p->flags |= PF_MEMALLOC;
705 reclaim_state.reclaimed_slab = 0;
706 p->reclaim_state = &reclaim_state;
708 try_to_free_pages(zones, gfp_mask, order);
710 p->reclaim_state = NULL;
711 p->flags &= ~PF_MEMALLOC;
713 /* go through the zonelist yet one more time */
714 for (i = 0; (z = zones[i]) != NULL; i++) {
716 if (gfp_mask & __GFP_HIGH)
720 min += (1<<order) + z->protection[alloc_type];
722 if (z->free_pages < min)
725 page = buffered_rmqueue(z, order, gfp_mask);
731 * Don't let big-order allocations loop unless the caller explicitly
732 * requests that. Wait for some write requests to complete then retry.
734 * In this implementation, __GFP_REPEAT means __GFP_NOFAIL for order
735 * <= 3, but that may not be true in other implementations.
738 if (!(gfp_mask & __GFP_NORETRY)) {
739 if ((order <= 3) || (gfp_mask & __GFP_REPEAT))
741 if (gfp_mask & __GFP_NOFAIL)
745 blk_congestion_wait(WRITE, HZ/50);
750 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
751 printk(KERN_WARNING "%s: page allocation failure."
752 " order:%d, mode:0x%x\n",
753 p->comm, order, gfp_mask);
758 zone_statistics(zonelist, z);
759 kernel_map_pages(page, 1 << order, 1);
763 EXPORT_SYMBOL(__alloc_pages);
766 * Common helper functions.
768 fastcall unsigned long __get_free_pages(unsigned int gfp_mask, unsigned int order)
771 page = alloc_pages(gfp_mask, order);
774 return (unsigned long) page_address(page);
777 EXPORT_SYMBOL(__get_free_pages);
779 fastcall unsigned long get_zeroed_page(unsigned int gfp_mask)
784 * get_zeroed_page() returns a 32-bit address, which cannot represent
787 BUG_ON(gfp_mask & __GFP_HIGHMEM);
789 page = alloc_pages(gfp_mask, 0);
791 void *address = page_address(page);
793 return (unsigned long) address;
798 EXPORT_SYMBOL(get_zeroed_page);
800 void __pagevec_free(struct pagevec *pvec)
802 int i = pagevec_count(pvec);
805 free_hot_cold_page(pvec->pages[i], pvec->cold);
808 fastcall void __free_pages(struct page *page, unsigned int order)
810 if (!PageReserved(page) && put_page_testzero(page)) {
814 __free_pages_ok(page, order);
818 EXPORT_SYMBOL(__free_pages);
820 fastcall void free_pages(unsigned long addr, unsigned int order)
823 BUG_ON(!virt_addr_valid((void *)addr));
824 __free_pages(virt_to_page((void *)addr), order);
828 EXPORT_SYMBOL(free_pages);
831 * Total amount of free (allocatable) RAM:
833 unsigned int nr_free_pages(void)
835 unsigned int sum = 0;
839 sum += zone->free_pages;
844 EXPORT_SYMBOL(nr_free_pages);
847 unsigned int nr_free_pages_pgdat(pg_data_t *pgdat)
849 unsigned int i, sum = 0;
851 for (i = 0; i < MAX_NR_ZONES; i++)
852 sum += pgdat->node_zones[i].free_pages;
858 static unsigned int nr_free_zone_pages(int offset)
861 unsigned int sum = 0;
863 for_each_pgdat(pgdat) {
864 struct zonelist *zonelist = pgdat->node_zonelists + offset;
865 struct zone **zonep = zonelist->zones;
868 for (zone = *zonep++; zone; zone = *zonep++) {
869 unsigned long size = zone->present_pages;
870 unsigned long high = zone->pages_high;
880 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
882 unsigned int nr_free_buffer_pages(void)
884 return nr_free_zone_pages(GFP_USER & GFP_ZONEMASK);
888 * Amount of free RAM allocatable within all zones
890 unsigned int nr_free_pagecache_pages(void)
892 return nr_free_zone_pages(GFP_HIGHUSER & GFP_ZONEMASK);
895 #ifdef CONFIG_HIGHMEM
896 unsigned int nr_free_highpages (void)
899 unsigned int pages = 0;
901 for_each_pgdat(pgdat)
902 pages += pgdat->node_zones[ZONE_HIGHMEM].free_pages;
909 static void show_node(struct zone *zone)
911 printk("Node %d ", zone->zone_pgdat->node_id);
914 #define show_node(zone) do { } while (0)
918 * Accumulate the page_state information across all CPUs.
919 * The result is unavoidably approximate - it can change
920 * during and after execution of this function.
922 DEFINE_PER_CPU(struct page_state, page_states) = {0};
923 EXPORT_PER_CPU_SYMBOL(page_states);
925 atomic_t nr_pagecache = ATOMIC_INIT(0);
926 EXPORT_SYMBOL(nr_pagecache);
928 DEFINE_PER_CPU(long, nr_pagecache_local) = 0;
931 void __get_page_state(struct page_state *ret, int nr)
935 memset(ret, 0, sizeof(*ret));
936 while (cpu < NR_CPUS) {
937 unsigned long *in, *out, off;
939 if (!cpu_possible(cpu)) {
944 in = (unsigned long *)&per_cpu(page_states, cpu);
946 if (cpu < NR_CPUS && cpu_possible(cpu))
947 prefetch(&per_cpu(page_states, cpu));
948 out = (unsigned long *)ret;
949 for (off = 0; off < nr; off++)
954 void get_page_state(struct page_state *ret)
958 nr = offsetof(struct page_state, GET_PAGE_STATE_LAST);
959 nr /= sizeof(unsigned long);
961 __get_page_state(ret, nr + 1);
964 void get_full_page_state(struct page_state *ret)
966 __get_page_state(ret, sizeof(*ret) / sizeof(unsigned long));
969 unsigned long __read_page_state(unsigned offset)
971 unsigned long ret = 0;
974 for (cpu = 0; cpu < NR_CPUS; cpu++) {
977 if (!cpu_possible(cpu))
980 in = (unsigned long)&per_cpu(page_states, cpu) + offset;
981 ret += *((unsigned long *)in);
986 void __get_zone_counts(unsigned long *active, unsigned long *inactive,
987 unsigned long *free, struct pglist_data *pgdat)
989 struct zone *zones = pgdat->node_zones;
995 for (i = 0; i < MAX_NR_ZONES; i++) {
996 *active += zones[i].nr_active;
997 *inactive += zones[i].nr_inactive;
998 *free += zones[i].free_pages;
1002 void get_zone_counts(unsigned long *active,
1003 unsigned long *inactive, unsigned long *free)
1005 struct pglist_data *pgdat;
1010 for_each_pgdat(pgdat) {
1011 unsigned long l, m, n;
1012 __get_zone_counts(&l, &m, &n, pgdat);
1019 void si_meminfo(struct sysinfo *val)
1021 val->totalram = totalram_pages;
1023 val->freeram = nr_free_pages();
1024 val->bufferram = nr_blockdev_pages();
1025 #ifdef CONFIG_HIGHMEM
1026 val->totalhigh = totalhigh_pages;
1027 val->freehigh = nr_free_highpages();
1032 val->mem_unit = PAGE_SIZE;
1033 if (vx_flags(VXF_VIRT_MEM, 0))
1034 vx_vsi_meminfo(val);
1037 EXPORT_SYMBOL(si_meminfo);
1040 void si_meminfo_node(struct sysinfo *val, int nid)
1042 pg_data_t *pgdat = NODE_DATA(nid);
1044 val->totalram = pgdat->node_present_pages;
1045 val->freeram = nr_free_pages_pgdat(pgdat);
1046 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
1047 val->freehigh = pgdat->node_zones[ZONE_HIGHMEM].free_pages;
1048 val->mem_unit = PAGE_SIZE;
1052 #define K(x) ((x) << (PAGE_SHIFT-10))
1055 * Show free area list (used inside shift_scroll-lock stuff)
1056 * We also calculate the percentage fragmentation. We do this by counting the
1057 * memory on each free list with the exception of the first item on the list.
1059 void show_free_areas(void)
1061 struct page_state ps;
1062 int cpu, temperature;
1063 unsigned long active;
1064 unsigned long inactive;
1068 for_each_zone(zone) {
1070 printk("%s per-cpu:", zone->name);
1072 if (!zone->present_pages) {
1078 for (cpu = 0; cpu < NR_CPUS; ++cpu) {
1079 struct per_cpu_pageset *pageset;
1081 if (!cpu_possible(cpu))
1084 pageset = zone->pageset + cpu;
1086 for (temperature = 0; temperature < 2; temperature++)
1087 printk("cpu %d %s: low %d, high %d, batch %d\n",
1089 temperature ? "cold" : "hot",
1090 pageset->pcp[temperature].low,
1091 pageset->pcp[temperature].high,
1092 pageset->pcp[temperature].batch);
1096 get_page_state(&ps);
1097 get_zone_counts(&active, &inactive, &free);
1099 printk("\nFree pages: %11ukB (%ukB HighMem)\n",
1101 K(nr_free_highpages()));
1103 printk("Active:%lu inactive:%lu dirty:%lu writeback:%lu "
1104 "unstable:%lu free:%u slab:%lu mapped:%lu pagetables:%lu\n",
1113 ps.nr_page_table_pages);
1115 for_each_zone(zone) {
1127 " pages_scanned:%lu"
1128 " all_unreclaimable? %s"
1131 K(zone->free_pages),
1134 K(zone->pages_high),
1136 K(zone->nr_inactive),
1137 K(zone->present_pages),
1138 zone->pages_scanned,
1139 (zone->all_unreclaimable ? "yes" : "no")
1141 printk("protections[]:");
1142 for (i = 0; i < MAX_NR_ZONES; i++)
1143 printk(" %lu", zone->protection[i]);
1147 for_each_zone(zone) {
1148 struct list_head *elem;
1149 unsigned long nr, flags, order, total = 0;
1152 printk("%s: ", zone->name);
1153 if (!zone->present_pages) {
1158 spin_lock_irqsave(&zone->lock, flags);
1159 for (order = 0; order < MAX_ORDER; order++) {
1161 list_for_each(elem, &zone->free_area[order].free_list)
1163 total += nr << order;
1164 printk("%lu*%lukB ", nr, K(1UL) << order);
1166 spin_unlock_irqrestore(&zone->lock, flags);
1167 printk("= %lukB\n", K(total));
1170 show_swap_cache_info();
1174 * Builds allocation fallback zone lists.
1176 static int __init build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist, int j, int k)
1183 zone = pgdat->node_zones + ZONE_HIGHMEM;
1184 if (zone->present_pages) {
1185 #ifndef CONFIG_HIGHMEM
1188 zonelist->zones[j++] = zone;
1191 zone = pgdat->node_zones + ZONE_NORMAL;
1192 if (zone->present_pages)
1193 zonelist->zones[j++] = zone;
1195 zone = pgdat->node_zones + ZONE_DMA;
1196 if (zone->present_pages)
1197 zonelist->zones[j++] = zone;
1204 #define MAX_NODE_LOAD (numnodes)
1205 static int __initdata node_load[MAX_NUMNODES];
1207 * find_next_best_node - find the next node that should appear in a given
1208 * node's fallback list
1209 * @node: node whose fallback list we're appending
1210 * @used_node_mask: pointer to the bitmap of already used nodes
1212 * We use a number of factors to determine which is the next node that should
1213 * appear on a given node's fallback list. The node should not have appeared
1214 * already in @node's fallback list, and it should be the next closest node
1215 * according to the distance array (which contains arbitrary distance values
1216 * from each node to each node in the system), and should also prefer nodes
1217 * with no CPUs, since presumably they'll have very little allocation pressure
1218 * on them otherwise.
1219 * It returns -1 if no node is found.
1221 static int __init find_next_best_node(int node, void *used_node_mask)
1224 int min_val = INT_MAX;
1227 for (i = 0; i < numnodes; i++) {
1230 /* Start from local node */
1231 n = (node+i)%numnodes;
1233 /* Don't want a node to appear more than once */
1234 if (test_bit(n, used_node_mask))
1237 /* Use the local node if we haven't already */
1238 if (!test_bit(node, used_node_mask)) {
1243 /* Use the distance array to find the distance */
1244 val = node_distance(node, n);
1246 /* Give preference to headless and unused nodes */
1247 tmp = node_to_cpumask(n);
1248 if (!cpus_empty(tmp))
1249 val += PENALTY_FOR_NODE_WITH_CPUS;
1251 /* Slight preference for less loaded node */
1252 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
1253 val += node_load[n];
1255 if (val < min_val) {
1262 set_bit(best_node, used_node_mask);
1267 static void __init build_zonelists(pg_data_t *pgdat)
1269 int i, j, k, node, local_node;
1270 int prev_node, load;
1271 struct zonelist *zonelist;
1272 DECLARE_BITMAP(used_mask, MAX_NUMNODES);
1274 /* initialize zonelists */
1275 for (i = 0; i < GFP_ZONETYPES; i++) {
1276 zonelist = pgdat->node_zonelists + i;
1277 memset(zonelist, 0, sizeof(*zonelist));
1278 zonelist->zones[0] = NULL;
1281 /* NUMA-aware ordering of nodes */
1282 local_node = pgdat->node_id;
1284 prev_node = local_node;
1285 bitmap_zero(used_mask, MAX_NUMNODES);
1286 while ((node = find_next_best_node(local_node, used_mask)) >= 0) {
1288 * We don't want to pressure a particular node.
1289 * So adding penalty to the first node in same
1290 * distance group to make it round-robin.
1292 if (node_distance(local_node, node) !=
1293 node_distance(local_node, prev_node))
1294 node_load[node] += load;
1297 for (i = 0; i < GFP_ZONETYPES; i++) {
1298 zonelist = pgdat->node_zonelists + i;
1299 for (j = 0; zonelist->zones[j] != NULL; j++);
1302 if (i & __GFP_HIGHMEM)
1307 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1308 zonelist->zones[j] = NULL;
1313 #else /* CONFIG_NUMA */
1315 static void __init build_zonelists(pg_data_t *pgdat)
1317 int i, j, k, node, local_node;
1319 local_node = pgdat->node_id;
1320 for (i = 0; i < GFP_ZONETYPES; i++) {
1321 struct zonelist *zonelist;
1323 zonelist = pgdat->node_zonelists + i;
1324 memset(zonelist, 0, sizeof(*zonelist));
1328 if (i & __GFP_HIGHMEM)
1333 j = build_zonelists_node(pgdat, zonelist, j, k);
1335 * Now we build the zonelist so that it contains the zones
1336 * of all the other nodes.
1337 * We don't want to pressure a particular node, so when
1338 * building the zones for node N, we make sure that the
1339 * zones coming right after the local ones are those from
1340 * node N+1 (modulo N)
1342 for (node = local_node + 1; node < numnodes; node++)
1343 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1344 for (node = 0; node < local_node; node++)
1345 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1347 zonelist->zones[j] = NULL;
1351 #endif /* CONFIG_NUMA */
1353 void __init build_all_zonelists(void)
1357 for(i = 0 ; i < numnodes ; i++)
1358 build_zonelists(NODE_DATA(i));
1359 printk("Built %i zonelists\n", numnodes);
1363 * Helper functions to size the waitqueue hash table.
1364 * Essentially these want to choose hash table sizes sufficiently
1365 * large so that collisions trying to wait on pages are rare.
1366 * But in fact, the number of active page waitqueues on typical
1367 * systems is ridiculously low, less than 200. So this is even
1368 * conservative, even though it seems large.
1370 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
1371 * waitqueues, i.e. the size of the waitq table given the number of pages.
1373 #define PAGES_PER_WAITQUEUE 256
1375 static inline unsigned long wait_table_size(unsigned long pages)
1377 unsigned long size = 1;
1379 pages /= PAGES_PER_WAITQUEUE;
1381 while (size < pages)
1385 * Once we have dozens or even hundreds of threads sleeping
1386 * on IO we've got bigger problems than wait queue collision.
1387 * Limit the size of the wait table to a reasonable size.
1389 size = min(size, 4096UL);
1391 return max(size, 4UL);
1395 * This is an integer logarithm so that shifts can be used later
1396 * to extract the more random high bits from the multiplicative
1397 * hash function before the remainder is taken.
1399 static inline unsigned long wait_table_bits(unsigned long size)
1404 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
1406 static void __init calculate_zone_totalpages(struct pglist_data *pgdat,
1407 unsigned long *zones_size, unsigned long *zholes_size)
1409 unsigned long realtotalpages, totalpages = 0;
1412 for (i = 0; i < MAX_NR_ZONES; i++)
1413 totalpages += zones_size[i];
1414 pgdat->node_spanned_pages = totalpages;
1416 realtotalpages = totalpages;
1418 for (i = 0; i < MAX_NR_ZONES; i++)
1419 realtotalpages -= zholes_size[i];
1420 pgdat->node_present_pages = realtotalpages;
1421 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id, realtotalpages);
1426 * Initially all pages are reserved - free ones are freed
1427 * up by free_all_bootmem() once the early boot process is
1428 * done. Non-atomic initialization, single-pass.
1430 void __init memmap_init_zone(unsigned long size, int nid, unsigned long zone,
1431 unsigned long start_pfn)
1433 struct page *start = pfn_to_page(start_pfn);
1436 for (page = start; page < (start + size); page++) {
1437 set_page_zone(page, NODEZONE(nid, zone));
1438 set_page_count(page, 0);
1439 reset_page_mapcount(page);
1440 SetPageReserved(page);
1441 INIT_LIST_HEAD(&page->lru);
1442 #ifdef WANT_PAGE_VIRTUAL
1443 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1444 if (!is_highmem_idx(zone))
1445 set_page_address(page, __va(start_pfn << PAGE_SHIFT));
1452 * Page buddy system uses "index >> (i+1)", where "index" is
1455 * The extra "+3" is to round down to byte size (8 bits per byte
1456 * assumption). Thus we get "(size-1) >> (i+4)" as the last byte
1459 * The "+1" is because we want to round the byte allocation up
1460 * rather than down. So we should have had a "+7" before we shifted
1461 * down by three. Also, we have to add one as we actually _use_ the
1462 * last bit (it's [0,n] inclusive, not [0,n[).
1464 * So we actually had +7+1 before we shift down by 3. But
1465 * (n+8) >> 3 == (n >> 3) + 1 (modulo overflows, which we do not have).
1467 * Finally, we LONG_ALIGN because all bitmap operations are on longs.
1469 unsigned long pages_to_bitmap_size(unsigned long order, unsigned long nr_pages)
1471 unsigned long bitmap_size;
1473 bitmap_size = (nr_pages-1) >> (order+4);
1474 bitmap_size = LONG_ALIGN(bitmap_size+1);
1479 void zone_init_free_lists(struct pglist_data *pgdat, struct zone *zone, unsigned long size)
1482 for (order = 0; ; order++) {
1483 unsigned long bitmap_size;
1485 INIT_LIST_HEAD(&zone->free_area[order].free_list);
1486 if (order == MAX_ORDER-1) {
1487 zone->free_area[order].map = NULL;
1491 bitmap_size = pages_to_bitmap_size(order, size);
1492 zone->free_area[order].map =
1493 (unsigned long *) alloc_bootmem_node(pgdat, bitmap_size);
1497 #ifndef __HAVE_ARCH_MEMMAP_INIT
1498 #define memmap_init(size, nid, zone, start_pfn) \
1499 memmap_init_zone((size), (nid), (zone), (start_pfn))
1503 * Set up the zone data structures:
1504 * - mark all pages reserved
1505 * - mark all memory queues empty
1506 * - clear the memory bitmaps
1508 static void __init free_area_init_core(struct pglist_data *pgdat,
1509 unsigned long *zones_size, unsigned long *zholes_size)
1512 const unsigned long zone_required_alignment = 1UL << (MAX_ORDER-1);
1513 int cpu, nid = pgdat->node_id;
1514 unsigned long zone_start_pfn = pgdat->node_start_pfn;
1516 pgdat->nr_zones = 0;
1517 init_waitqueue_head(&pgdat->kswapd_wait);
1519 for (j = 0; j < MAX_NR_ZONES; j++) {
1520 struct zone *zone = pgdat->node_zones + j;
1521 unsigned long size, realsize;
1522 unsigned long batch;
1524 zone_table[NODEZONE(nid, j)] = zone;
1525 realsize = size = zones_size[j];
1527 realsize -= zholes_size[j];
1529 if (j == ZONE_DMA || j == ZONE_NORMAL)
1530 nr_kernel_pages += realsize;
1531 nr_all_pages += realsize;
1533 zone->spanned_pages = size;
1534 zone->present_pages = realsize;
1535 zone->name = zone_names[j];
1536 spin_lock_init(&zone->lock);
1537 spin_lock_init(&zone->lru_lock);
1538 zone->zone_pgdat = pgdat;
1539 zone->free_pages = 0;
1541 zone->temp_priority = zone->prev_priority = DEF_PRIORITY;
1544 * The per-cpu-pages pools are set to around 1000th of the
1545 * size of the zone. But no more than 1/4 of a meg - there's
1546 * no point in going beyond the size of L2 cache.
1548 * OK, so we don't know how big the cache is. So guess.
1550 batch = zone->present_pages / 1024;
1551 if (batch * PAGE_SIZE > 256 * 1024)
1552 batch = (256 * 1024) / PAGE_SIZE;
1553 batch /= 4; /* We effectively *= 4 below */
1557 for (cpu = 0; cpu < NR_CPUS; cpu++) {
1558 struct per_cpu_pages *pcp;
1560 pcp = &zone->pageset[cpu].pcp[0]; /* hot */
1562 pcp->low = 2 * batch;
1563 pcp->high = 6 * batch;
1564 pcp->batch = 1 * batch;
1565 INIT_LIST_HEAD(&pcp->list);
1567 pcp = &zone->pageset[cpu].pcp[1]; /* cold */
1570 pcp->high = 2 * batch;
1571 pcp->batch = 1 * batch;
1572 INIT_LIST_HEAD(&pcp->list);
1574 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%lu\n",
1575 zone_names[j], realsize, batch);
1576 #ifndef CONFIG_CKRM_RES_MEM
1577 INIT_LIST_HEAD(&zone->active_list);
1578 INIT_LIST_HEAD(&zone->inactive_list);
1580 zone->nr_scan_active = 0;
1581 zone->nr_scan_inactive = 0;
1582 zone->nr_active = 0;
1583 zone->nr_inactive = 0;
1588 * The per-page waitqueue mechanism uses hashed waitqueues
1591 zone->wait_table_size = wait_table_size(size);
1592 zone->wait_table_bits =
1593 wait_table_bits(zone->wait_table_size);
1594 zone->wait_table = (wait_queue_head_t *)
1595 alloc_bootmem_node(pgdat, zone->wait_table_size
1596 * sizeof(wait_queue_head_t));
1598 for(i = 0; i < zone->wait_table_size; ++i)
1599 init_waitqueue_head(zone->wait_table + i);
1601 pgdat->nr_zones = j+1;
1603 zone->zone_mem_map = pfn_to_page(zone_start_pfn);
1604 zone->zone_start_pfn = zone_start_pfn;
1606 if ((zone_start_pfn) & (zone_required_alignment-1))
1607 printk("BUG: wrong zone alignment, it will crash\n");
1609 memmap_init(size, nid, j, zone_start_pfn);
1611 zone_start_pfn += size;
1613 zone_init_free_lists(pgdat, zone, zone->spanned_pages);
1617 void __init node_alloc_mem_map(struct pglist_data *pgdat)
1621 size = (pgdat->node_spanned_pages + 1) * sizeof(struct page);
1622 pgdat->node_mem_map = alloc_bootmem_node(pgdat, size);
1623 #ifndef CONFIG_DISCONTIGMEM
1624 mem_map = contig_page_data.node_mem_map;
1628 void __init free_area_init_node(int nid, struct pglist_data *pgdat,
1629 unsigned long *zones_size, unsigned long node_start_pfn,
1630 unsigned long *zholes_size)
1632 pgdat->node_id = nid;
1633 pgdat->node_start_pfn = node_start_pfn;
1634 calculate_zone_totalpages(pgdat, zones_size, zholes_size);
1636 if (!pfn_to_page(node_start_pfn))
1637 node_alloc_mem_map(pgdat);
1639 free_area_init_core(pgdat, zones_size, zholes_size);
1642 #ifndef CONFIG_DISCONTIGMEM
1643 static bootmem_data_t contig_bootmem_data;
1644 struct pglist_data contig_page_data = { .bdata = &contig_bootmem_data };
1646 EXPORT_SYMBOL(contig_page_data);
1648 void __init free_area_init(unsigned long *zones_size)
1650 free_area_init_node(0, &contig_page_data, zones_size,
1651 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
1655 #ifdef CONFIG_PROC_FS
1657 #include <linux/seq_file.h>
1659 static void *frag_start(struct seq_file *m, loff_t *pos)
1664 for (pgdat = pgdat_list; pgdat && node; pgdat = pgdat->pgdat_next)
1670 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
1672 pg_data_t *pgdat = (pg_data_t *)arg;
1675 return pgdat->pgdat_next;
1678 static void frag_stop(struct seq_file *m, void *arg)
1683 * This walks the freelist for each zone. Whilst this is slow, I'd rather
1684 * be slow here than slow down the fast path by keeping stats - mjbligh
1686 static int frag_show(struct seq_file *m, void *arg)
1688 pg_data_t *pgdat = (pg_data_t *)arg;
1690 struct zone *node_zones = pgdat->node_zones;
1691 unsigned long flags;
1694 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
1695 if (!zone->present_pages)
1698 spin_lock_irqsave(&zone->lock, flags);
1699 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1700 for (order = 0; order < MAX_ORDER; ++order) {
1701 unsigned long nr_bufs = 0;
1702 struct list_head *elem;
1704 list_for_each(elem, &(zone->free_area[order].free_list))
1706 seq_printf(m, "%6lu ", nr_bufs);
1708 spin_unlock_irqrestore(&zone->lock, flags);
1714 struct seq_operations fragmentation_op = {
1715 .start = frag_start,
1721 static char *vmstat_text[] = {
1725 "nr_page_table_pages",
1750 "pgscan_kswapd_high",
1751 "pgscan_kswapd_normal",
1753 "pgscan_kswapd_dma",
1754 "pgscan_direct_high",
1755 "pgscan_direct_normal",
1756 "pgscan_direct_dma",
1761 "kswapd_inodesteal",
1768 static void *vmstat_start(struct seq_file *m, loff_t *pos)
1770 struct page_state *ps;
1772 if (*pos >= ARRAY_SIZE(vmstat_text))
1775 ps = kmalloc(sizeof(*ps), GFP_KERNEL);
1778 return ERR_PTR(-ENOMEM);
1779 get_full_page_state(ps);
1780 ps->pgpgin /= 2; /* sectors -> kbytes */
1782 return (unsigned long *)ps + *pos;
1785 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
1788 if (*pos >= ARRAY_SIZE(vmstat_text))
1790 return (unsigned long *)m->private + *pos;
1793 static int vmstat_show(struct seq_file *m, void *arg)
1795 unsigned long *l = arg;
1796 unsigned long off = l - (unsigned long *)m->private;
1798 seq_printf(m, "%s %lu\n", vmstat_text[off], *l);
1802 static void vmstat_stop(struct seq_file *m, void *arg)
1808 struct seq_operations vmstat_op = {
1809 .start = vmstat_start,
1810 .next = vmstat_next,
1811 .stop = vmstat_stop,
1812 .show = vmstat_show,
1815 #endif /* CONFIG_PROC_FS */
1817 #ifdef CONFIG_HOTPLUG_CPU
1818 static int page_alloc_cpu_notify(struct notifier_block *self,
1819 unsigned long action, void *hcpu)
1821 int cpu = (unsigned long)hcpu;
1824 if (action == CPU_DEAD) {
1825 /* Drain local pagecache count. */
1826 count = &per_cpu(nr_pagecache_local, cpu);
1827 atomic_add(*count, &nr_pagecache);
1829 local_irq_disable();
1835 #endif /* CONFIG_HOTPLUG_CPU */
1837 void __init page_alloc_init(void)
1839 hotcpu_notifier(page_alloc_cpu_notify, 0);
1842 static unsigned long higherzone_val(struct zone *z, int max_zone,
1845 int z_idx = zone_idx(z);
1846 struct zone *higherzone;
1847 unsigned long pages;
1849 /* there is no higher zone to get a contribution from */
1850 if (z_idx == MAX_NR_ZONES-1)
1853 higherzone = &z->zone_pgdat->node_zones[z_idx+1];
1855 /* We always start with the higher zone's protection value */
1856 pages = higherzone->protection[alloc_type];
1859 * We get a lower-zone-protection contribution only if there are
1860 * pages in the higher zone and if we're not the highest zone
1861 * in the current zonelist. e.g., never happens for GFP_DMA. Happens
1862 * only for ZONE_DMA in a GFP_KERNEL allocation and happens for ZONE_DMA
1863 * and ZONE_NORMAL for a GFP_HIGHMEM allocation.
1865 if (higherzone->present_pages && z_idx < alloc_type)
1866 pages += higherzone->pages_low * sysctl_lower_zone_protection;
1872 * setup_per_zone_protection - called whenver min_free_kbytes or
1873 * sysctl_lower_zone_protection changes. Ensures that each zone
1874 * has a correct pages_protected value, so an adequate number of
1875 * pages are left in the zone after a successful __alloc_pages().
1877 * This algorithm is way confusing. I tries to keep the same behavior
1878 * as we had with the incremental min iterative algorithm.
1880 static void setup_per_zone_protection(void)
1882 struct pglist_data *pgdat;
1883 struct zone *zones, *zone;
1887 for_each_pgdat(pgdat) {
1888 zones = pgdat->node_zones;
1890 for (i = 0, max_zone = 0; i < MAX_NR_ZONES; i++)
1891 if (zones[i].present_pages)
1895 * For each of the different allocation types:
1896 * GFP_DMA -> GFP_KERNEL -> GFP_HIGHMEM
1898 for (i = 0; i < GFP_ZONETYPES; i++) {
1900 * For each of the zones:
1901 * ZONE_HIGHMEM -> ZONE_NORMAL -> ZONE_DMA
1903 for (j = MAX_NR_ZONES-1; j >= 0; j--) {
1907 * We never protect zones that don't have memory
1908 * in them (j>max_zone) or zones that aren't in
1909 * the zonelists for a certain type of
1910 * allocation (j>=i). We have to assign these
1911 * to zero because the lower zones take
1912 * contributions from the higher zones.
1914 if (j > max_zone || j >= i) {
1915 zone->protection[i] = 0;
1919 * The contribution of the next higher zone
1921 zone->protection[i] = higherzone_val(zone,
1929 * setup_per_zone_pages_min - called when min_free_kbytes changes. Ensures
1930 * that the pages_{min,low,high} values for each zone are set correctly
1931 * with respect to min_free_kbytes.
1933 static void setup_per_zone_pages_min(void)
1935 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
1936 unsigned long lowmem_pages = 0;
1938 unsigned long flags;
1940 /* Calculate total number of !ZONE_HIGHMEM pages */
1941 for_each_zone(zone) {
1942 if (!is_highmem(zone))
1943 lowmem_pages += zone->present_pages;
1946 for_each_zone(zone) {
1947 spin_lock_irqsave(&zone->lru_lock, flags);
1948 if (is_highmem(zone)) {
1950 * Often, highmem doesn't need to reserve any pages.
1951 * But the pages_min/low/high values are also used for
1952 * batching up page reclaim activity so we need a
1953 * decent value here.
1957 min_pages = zone->present_pages / 1024;
1958 if (min_pages < SWAP_CLUSTER_MAX)
1959 min_pages = SWAP_CLUSTER_MAX;
1960 if (min_pages > 128)
1962 zone->pages_min = min_pages;
1964 /* if it's a lowmem zone, reserve a number of pages
1965 * proportionate to the zone's size.
1967 zone->pages_min = (pages_min * zone->present_pages) /
1972 * When interpreting these watermarks, just keep in mind that:
1973 * zone->pages_min == (zone->pages_min * 4) / 4;
1975 zone->pages_low = (zone->pages_min * 5) / 4;
1976 zone->pages_high = (zone->pages_min * 6) / 4;
1977 spin_unlock_irqrestore(&zone->lru_lock, flags);
1982 * Initialise min_free_kbytes.
1984 * For small machines we want it small (128k min). For large machines
1985 * we want it large (64MB max). But it is not linear, because network
1986 * bandwidth does not increase linearly with machine size. We use
1988 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
1989 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
2005 static int __init init_per_zone_pages_min(void)
2007 unsigned long lowmem_kbytes;
2009 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
2011 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
2012 if (min_free_kbytes < 128)
2013 min_free_kbytes = 128;
2014 if (min_free_kbytes > 65536)
2015 min_free_kbytes = 65536;
2016 setup_per_zone_pages_min();
2017 setup_per_zone_protection();
2020 module_init(init_per_zone_pages_min)
2023 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
2024 * that we can call two helper functions whenever min_free_kbytes
2027 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
2028 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2030 proc_dointvec(table, write, file, buffer, length, ppos);
2031 setup_per_zone_pages_min();
2032 setup_per_zone_protection();
2037 * lower_zone_protection_sysctl_handler - just a wrapper around
2038 * proc_dointvec() so that we can call setup_per_zone_protection()
2039 * whenever sysctl_lower_zone_protection changes.
2041 int lower_zone_protection_sysctl_handler(ctl_table *table, int write,
2042 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2044 proc_dointvec_minmax(table, write, file, buffer, length, ppos);
2045 setup_per_zone_protection();
2050 * allocate a large system hash table from bootmem
2051 * - it is assumed that the hash table must contain an exact power-of-2
2052 * quantity of entries
2054 void *__init alloc_large_system_hash(const char *tablename,
2055 unsigned long bucketsize,
2056 unsigned long numentries,
2058 int consider_highmem,
2059 unsigned int *_hash_shift,
2060 unsigned int *_hash_mask)
2062 unsigned long long max;
2063 unsigned long log2qty, size;
2066 /* allow the kernel cmdline to have a say */
2068 /* round applicable memory size up to nearest megabyte */
2069 numentries = consider_highmem ? nr_all_pages : nr_kernel_pages;
2070 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
2071 numentries >>= 20 - PAGE_SHIFT;
2072 numentries <<= 20 - PAGE_SHIFT;
2074 /* limit to 1 bucket per 2^scale bytes of low memory */
2075 if (scale > PAGE_SHIFT)
2076 numentries >>= (scale - PAGE_SHIFT);
2078 numentries <<= (PAGE_SHIFT - scale);
2080 /* rounded up to nearest power of 2 in size */
2081 numentries = 1UL << (long_log2(numentries) + 1);
2083 /* limit allocation size to 1/16 total memory */
2084 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
2085 do_div(max, bucketsize);
2087 if (numentries > max)
2090 log2qty = long_log2(numentries);
2093 size = bucketsize << log2qty;
2094 table = alloc_bootmem(size);
2095 } while (!table && size > PAGE_SIZE && --log2qty);
2098 panic("Failed to allocate %s hash table\n", tablename);
2100 printk("%s hash table entries: %d (order: %d, %lu bytes)\n",
2103 long_log2(size) - PAGE_SHIFT,
2107 *_hash_shift = log2qty;
2109 *_hash_mask = (1 << log2qty) - 1;