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/nodemask.h>
35 #include <linux/ckrm_mem_inline.h>
37 #include <asm/tlbflush.h>
39 nodemask_t node_online_map = NODE_MASK_NONE;
40 nodemask_t node_possible_map = NODE_MASK_ALL;
41 struct pglist_data *pgdat_list;
42 unsigned long totalram_pages;
43 unsigned long totalhigh_pages;
46 int sysctl_lower_zone_protection = 0;
48 EXPORT_SYMBOL(totalram_pages);
49 EXPORT_SYMBOL(nr_swap_pages);
52 * Used by page_zone() to look up the address of the struct zone whose
53 * id is encoded in the upper bits of page->flags
55 struct zone *zone_table[1 << (ZONES_SHIFT + NODES_SHIFT)];
56 EXPORT_SYMBOL(zone_table);
58 static char *zone_names[MAX_NR_ZONES] = { "DMA", "Normal", "HighMem" };
59 int min_free_kbytes = 1024;
61 unsigned long __initdata nr_kernel_pages;
62 unsigned long __initdata nr_all_pages;
65 * Temporary debugging check for pages not lying within a given zone.
67 static int bad_range(struct zone *zone, struct page *page)
69 if (page_to_pfn(page) >= zone->zone_start_pfn + zone->spanned_pages)
71 if (page_to_pfn(page) < zone->zone_start_pfn)
73 if (zone != page_zone(page))
78 static void bad_page(const char *function, struct page *page)
80 printk(KERN_EMERG "Bad page state at %s (in process '%s', page %p)\n",
81 function, current->comm, page);
82 printk(KERN_EMERG "flags:0x%0*lx mapping:%p mapcount:%d count:%d\n",
83 (int)(2*sizeof(page_flags_t)), (unsigned long)page->flags,
84 page->mapping, page_mapcount(page), page_count(page));
85 printk(KERN_EMERG "Backtrace:\n");
87 printk(KERN_EMERG "Trying to fix it up, but a reboot is needed\n");
88 page->flags &= ~(1 << PG_private |
95 set_page_count(page, 0);
96 reset_page_mapcount(page);
98 tainted |= TAINT_BAD_PAGE;
101 #ifndef CONFIG_HUGETLB_PAGE
102 #define prep_compound_page(page, order) do { } while (0)
103 #define destroy_compound_page(page, order) do { } while (0)
106 * Higher-order pages are called "compound pages". They are structured thusly:
108 * The first PAGE_SIZE page is called the "head page".
110 * The remaining PAGE_SIZE pages are called "tail pages".
112 * All pages have PG_compound set. All pages have their ->private pointing at
113 * the head page (even the head page has this).
115 * The first tail page's ->mapping, if non-zero, holds the address of the
116 * compound page's put_page() function.
118 * The order of the allocation is stored in the first tail page's ->index
119 * This is only for debug at present. This usage means that zero-order pages
120 * may not be compound.
122 static void prep_compound_page(struct page *page, unsigned long order)
125 int nr_pages = 1 << order;
127 page[1].mapping = NULL;
128 page[1].index = order;
129 for (i = 0; i < nr_pages; i++) {
130 struct page *p = page + i;
133 p->private = (unsigned long)page;
137 static void destroy_compound_page(struct page *page, unsigned long order)
140 int nr_pages = 1 << order;
142 if (!PageCompound(page))
145 if (page[1].index != order)
146 bad_page(__FUNCTION__, page);
148 for (i = 0; i < nr_pages; i++) {
149 struct page *p = page + i;
151 if (!PageCompound(p))
152 bad_page(__FUNCTION__, page);
153 if (p->private != (unsigned long)page)
154 bad_page(__FUNCTION__, page);
155 ClearPageCompound(p);
158 #endif /* CONFIG_HUGETLB_PAGE */
161 * Freeing function for a buddy system allocator.
163 * The concept of a buddy system is to maintain direct-mapped table
164 * (containing bit values) for memory blocks of various "orders".
165 * The bottom level table contains the map for the smallest allocatable
166 * units of memory (here, pages), and each level above it describes
167 * pairs of units from the levels below, hence, "buddies".
168 * At a high level, all that happens here is marking the table entry
169 * at the bottom level available, and propagating the changes upward
170 * as necessary, plus some accounting needed to play nicely with other
171 * parts of the VM system.
172 * At each level, we keep one bit for each pair of blocks, which
173 * is set to 1 iff only one of the pair is allocated. So when we
174 * are allocating or freeing one, we can derive the state of the
175 * other. That is, if we allocate a small block, and both were
176 * free, the remainder of the region must be split into blocks.
177 * If a block is freed, and its buddy is also free, then this
178 * triggers coalescing into a block of larger size.
183 static inline void __free_pages_bulk (struct page *page, struct page *base,
184 struct zone *zone, struct free_area *area, unsigned int order)
186 unsigned long page_idx, index, mask;
189 destroy_compound_page(page, order);
190 mask = (~0UL) << order;
191 page_idx = page - base;
192 if (page_idx & ~mask)
194 index = page_idx >> (1 + order);
196 zone->free_pages += 1 << order;
197 while (order < MAX_ORDER-1) {
198 struct page *buddy1, *buddy2;
200 BUG_ON(area >= zone->free_area + MAX_ORDER);
201 if (!__test_and_change_bit(index, area->map))
203 * the buddy page is still allocated.
207 /* Move the buddy up one level. */
208 buddy1 = base + (page_idx ^ (1 << order));
209 buddy2 = base + page_idx;
210 BUG_ON(bad_range(zone, buddy1));
211 BUG_ON(bad_range(zone, buddy2));
212 list_del(&buddy1->lru);
219 list_add(&(base + page_idx)->lru, &area->free_list);
222 static inline void free_pages_check(const char *function, struct page *page)
224 if ( page_mapped(page) ||
225 page->mapping != NULL ||
226 page_count(page) != 0 ||
235 1 << PG_writeback )))
236 bad_page(function, page);
238 ClearPageDirty(page);
242 * Frees a list of pages.
243 * Assumes all pages on list are in same zone, and of same order.
244 * count is the number of pages to free, or 0 for all on the list.
246 * If the zone was previously in an "all pages pinned" state then look to
247 * see if this freeing clears that state.
249 * And clear the zone's pages_scanned counter, to hold off the "all pages are
250 * pinned" detection logic.
253 free_pages_bulk(struct zone *zone, int count,
254 struct list_head *list, unsigned int order)
257 struct free_area *area;
258 struct page *base, *page = NULL;
261 base = zone->zone_mem_map;
262 area = zone->free_area + order;
263 spin_lock_irqsave(&zone->lock, flags);
264 zone->all_unreclaimable = 0;
265 zone->pages_scanned = 0;
266 while (!list_empty(list) && count--) {
267 page = list_entry(list->prev, struct page, lru);
268 /* have to delete it as __free_pages_bulk list manipulates */
269 list_del(&page->lru);
270 __free_pages_bulk(page, base, zone, area, order);
271 ckrm_clear_page_class(page);
274 spin_unlock_irqrestore(&zone->lock, flags);
278 void __free_pages_ok(struct page *page, unsigned int order)
283 arch_free_page(page, order);
285 mod_page_state(pgfree, 1 << order);
286 for (i = 0 ; i < (1 << order) ; ++i)
287 free_pages_check(__FUNCTION__, page + i);
288 list_add(&page->lru, &list);
289 kernel_map_pages(page, 1<<order, 0);
290 free_pages_bulk(page_zone(page), 1, &list, order);
293 #define MARK_USED(index, order, area) \
294 __change_bit((index) >> (1+(order)), (area)->map)
297 * The order of subdivision here is critical for the IO subsystem.
298 * Please do not alter this order without good reasons and regression
299 * testing. Specifically, as large blocks of memory are subdivided,
300 * the order in which smaller blocks are delivered depends on the order
301 * they're subdivided in this function. This is the primary factor
302 * influencing the order in which pages are delivered to the IO
303 * subsystem according to empirical testing, and this is also justified
304 * by considering the behavior of a buddy system containing a single
305 * large block of memory acted on by a series of small allocations.
306 * This behavior is a critical factor in sglist merging's success.
310 static inline struct page *
311 expand(struct zone *zone, struct page *page,
312 unsigned long index, int low, int high, struct free_area *area)
314 unsigned long size = 1 << high;
320 BUG_ON(bad_range(zone, &page[size]));
321 list_add(&page[size].lru, &area->free_list);
322 MARK_USED(index + size, high, area);
327 static inline void set_page_refs(struct page *page, int order)
330 set_page_count(page, 1);
335 * We need to reference all the pages for this order, otherwise if
336 * anyone accesses one of the pages with (get/put) it will be freed.
338 for (i = 0; i < (1 << order); i++)
339 set_page_count(page+i, 1);
340 #endif /* CONFIG_MMU */
344 * This page is about to be returned from the page allocator
346 static void prep_new_page(struct page *page, int order)
348 if (page->mapping || page_mapped(page) ||
357 1 << PG_writeback )))
358 bad_page(__FUNCTION__, page);
360 page->flags &= ~(1 << PG_uptodate | 1 << PG_error |
361 1 << PG_referenced | 1 << PG_arch_1 |
362 1 << PG_checked | 1 << PG_mappedtodisk);
364 ckrm_page_init(page);
365 set_page_refs(page, order);
369 * Do the hard work of removing an element from the buddy allocator.
370 * Call me with the zone->lock already held.
372 static struct page *__rmqueue(struct zone *zone, unsigned int order)
374 struct free_area * area;
375 unsigned int current_order;
379 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
380 area = zone->free_area + current_order;
381 if (list_empty(&area->free_list))
384 page = list_entry(area->free_list.next, struct page, lru);
385 list_del(&page->lru);
386 index = page - zone->zone_mem_map;
387 if (current_order != MAX_ORDER-1)
388 MARK_USED(index, current_order, area);
389 zone->free_pages -= 1UL << order;
390 return expand(zone, page, index, order, current_order, area);
397 * Obtain a specified number of elements from the buddy allocator, all under
398 * a single hold of the lock, for efficiency. Add them to the supplied list.
399 * Returns the number of new pages which were placed at *list.
401 static int rmqueue_bulk(struct zone *zone, unsigned int order,
402 unsigned long count, struct list_head *list)
409 spin_lock_irqsave(&zone->lock, flags);
410 for (i = 0; i < count; ++i) {
411 page = __rmqueue(zone, order);
415 list_add_tail(&page->lru, list);
417 spin_unlock_irqrestore(&zone->lock, flags);
421 #if defined(CONFIG_PM) || defined(CONFIG_HOTPLUG_CPU)
422 static void __drain_pages(unsigned int cpu)
427 for_each_zone(zone) {
428 struct per_cpu_pageset *pset;
430 pset = &zone->pageset[cpu];
431 for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
432 struct per_cpu_pages *pcp;
435 pcp->count -= free_pages_bulk(zone, pcp->count,
440 #endif /* CONFIG_PM || CONFIG_HOTPLUG_CPU */
443 int is_head_of_free_region(struct page *page)
445 struct zone *zone = page_zone(page);
448 struct list_head *curr;
451 * Should not matter as we need quiescent system for
452 * suspend anyway, but...
454 spin_lock_irqsave(&zone->lock, flags);
455 for (order = MAX_ORDER - 1; order >= 0; --order)
456 list_for_each(curr, &zone->free_area[order].free_list)
457 if (page == list_entry(curr, struct page, lru)) {
458 spin_unlock_irqrestore(&zone->lock, flags);
461 spin_unlock_irqrestore(&zone->lock, flags);
466 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
468 void drain_local_pages(void)
472 local_irq_save(flags);
473 __drain_pages(smp_processor_id());
474 local_irq_restore(flags);
476 #endif /* CONFIG_PM */
478 static void zone_statistics(struct zonelist *zonelist, struct zone *z)
483 pg_data_t *pg = z->zone_pgdat;
484 pg_data_t *orig = zonelist->zones[0]->zone_pgdat;
485 struct per_cpu_pageset *p;
487 local_irq_save(flags);
488 cpu = smp_processor_id();
489 p = &z->pageset[cpu];
491 z->pageset[cpu].numa_hit++;
494 zonelist->zones[0]->pageset[cpu].numa_foreign++;
496 if (pg == NODE_DATA(numa_node_id()))
500 local_irq_restore(flags);
505 * Free a 0-order page
507 static void FASTCALL(free_hot_cold_page(struct page *page, int cold));
508 static void fastcall free_hot_cold_page(struct page *page, int cold)
510 struct zone *zone = page_zone(page);
511 struct per_cpu_pages *pcp;
514 arch_free_page(page, 0);
516 kernel_map_pages(page, 1, 0);
517 inc_page_state(pgfree);
519 page->mapping = NULL;
520 free_pages_check(__FUNCTION__, page);
521 pcp = &zone->pageset[get_cpu()].pcp[cold];
522 local_irq_save(flags);
523 if (pcp->count >= pcp->high)
524 pcp->count -= free_pages_bulk(zone, pcp->batch, &pcp->list, 0);
525 list_add(&page->lru, &pcp->list);
527 local_irq_restore(flags);
531 void fastcall free_hot_page(struct page *page)
533 free_hot_cold_page(page, 0);
536 void fastcall free_cold_page(struct page *page)
538 free_hot_cold_page(page, 1);
542 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
543 * we cheat by calling it from here, in the order > 0 path. Saves a branch
548 buffered_rmqueue(struct zone *zone, int order, int gfp_flags)
551 struct page *page = NULL;
552 int cold = !!(gfp_flags & __GFP_COLD);
555 struct per_cpu_pages *pcp;
557 pcp = &zone->pageset[get_cpu()].pcp[cold];
558 local_irq_save(flags);
559 if (pcp->count <= pcp->low)
560 pcp->count += rmqueue_bulk(zone, 0,
561 pcp->batch, &pcp->list);
563 page = list_entry(pcp->list.next, struct page, lru);
564 list_del(&page->lru);
567 local_irq_restore(flags);
572 spin_lock_irqsave(&zone->lock, flags);
573 page = __rmqueue(zone, order);
574 spin_unlock_irqrestore(&zone->lock, flags);
578 BUG_ON(bad_range(zone, page));
579 mod_page_state_zone(zone, pgalloc, 1 << order);
580 prep_new_page(page, order);
581 if (order && (gfp_flags & __GFP_COMP))
582 prep_compound_page(page, order);
588 * This is the 'heart' of the zoned buddy allocator.
590 * Herein lies the mysterious "incremental min". That's the
592 * local_low = z->pages_low;
595 * thing. The intent here is to provide additional protection to low zones for
596 * allocation requests which _could_ use higher zones. So a GFP_HIGHMEM
597 * request is not allowed to dip as deeply into the normal zone as a GFP_KERNEL
598 * request. This preserves additional space in those lower zones for requests
599 * which really do need memory from those zones. It means that on a decent
600 * sized machine, GFP_HIGHMEM and GFP_KERNEL requests basically leave the DMA
603 struct page * fastcall
604 __alloc_pages(unsigned int gfp_mask, unsigned int order,
605 struct zonelist *zonelist)
607 const int wait = gfp_mask & __GFP_WAIT;
609 struct zone **zones, *z;
611 struct reclaim_state reclaim_state;
612 struct task_struct *p = current;
618 might_sleep_if(wait);
621 * The caller may dip into page reserves a bit more if the caller
622 * cannot run direct reclaim, or is the caller has realtime scheduling
625 can_try_harder = (unlikely(rt_task(p)) && !in_interrupt()) || !wait;
627 if (!in_interrupt() && !ckrm_class_limit_ok(ckrm_get_mem_class(p)))
630 zones = zonelist->zones; /* the list of zones suitable for gfp_mask */
632 if (unlikely(zones[0] == NULL)) {
633 /* Should this ever happen?? */
637 alloc_type = zone_idx(zones[0]);
639 /* Go through the zonelist once, looking for a zone with enough free */
640 for (i = 0; (z = zones[i]) != NULL; i++) {
641 min = z->pages_low + (1<<order) + z->protection[alloc_type];
643 if (z->free_pages < min)
646 page = buffered_rmqueue(z, order, gfp_mask);
651 for (i = 0; (z = zones[i]) != NULL; i++)
655 * Go through the zonelist again. Let __GFP_HIGH and allocations
656 * coming from realtime tasks to go deeper into reserves
658 for (i = 0; (z = zones[i]) != NULL; i++) {
660 if (gfp_mask & __GFP_HIGH)
664 min += (1<<order) + z->protection[alloc_type];
666 if (z->free_pages < min)
669 page = buffered_rmqueue(z, order, gfp_mask);
674 /* This allocation should allow future memory freeing. */
675 if ((p->flags & (PF_MEMALLOC | PF_MEMDIE)) && !in_interrupt()) {
676 /* go through the zonelist yet again, ignoring mins */
677 for (i = 0; (z = zones[i]) != NULL; i++) {
678 page = buffered_rmqueue(z, order, gfp_mask);
685 /* Atomic allocations - we can't balance anything */
690 /* We now go into synchronous reclaim */
691 p->flags |= PF_MEMALLOC;
692 reclaim_state.reclaimed_slab = 0;
693 p->reclaim_state = &reclaim_state;
695 try_to_free_pages(zones, gfp_mask, order);
697 p->reclaim_state = NULL;
698 p->flags &= ~PF_MEMALLOC;
700 /* go through the zonelist yet one more time */
701 for (i = 0; (z = zones[i]) != NULL; i++) {
703 if (gfp_mask & __GFP_HIGH)
707 min += (1<<order) + z->protection[alloc_type];
709 if (z->free_pages < min)
712 page = buffered_rmqueue(z, order, gfp_mask);
718 * Don't let big-order allocations loop unless the caller explicitly
719 * requests that. Wait for some write requests to complete then retry.
721 * In this implementation, __GFP_REPEAT means __GFP_NOFAIL for order
722 * <= 3, but that may not be true in other implementations.
725 if (!(gfp_mask & __GFP_NORETRY)) {
726 if ((order <= 3) || (gfp_mask & __GFP_REPEAT))
728 if (gfp_mask & __GFP_NOFAIL)
732 blk_congestion_wait(WRITE, HZ/50);
737 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
738 printk(KERN_WARNING "%s: page allocation failure."
739 " order:%d, mode:0x%x\n",
740 p->comm, order, gfp_mask);
745 zone_statistics(zonelist, z);
746 kernel_map_pages(page, 1 << order, 1);
750 EXPORT_SYMBOL(__alloc_pages);
753 * Common helper functions.
755 fastcall unsigned long __get_free_pages(unsigned int gfp_mask, unsigned int order)
758 page = alloc_pages(gfp_mask, order);
761 return (unsigned long) page_address(page);
764 EXPORT_SYMBOL(__get_free_pages);
766 fastcall unsigned long get_zeroed_page(unsigned int gfp_mask)
771 * get_zeroed_page() returns a 32-bit address, which cannot represent
774 BUG_ON(gfp_mask & __GFP_HIGHMEM);
776 page = alloc_pages(gfp_mask, 0);
778 void *address = page_address(page);
780 return (unsigned long) address;
785 EXPORT_SYMBOL(get_zeroed_page);
787 void __pagevec_free(struct pagevec *pvec)
789 int i = pagevec_count(pvec);
792 free_hot_cold_page(pvec->pages[i], pvec->cold);
795 fastcall void __free_pages(struct page *page, unsigned int order)
797 if (!PageReserved(page) && put_page_testzero(page)) {
801 __free_pages_ok(page, order);
805 EXPORT_SYMBOL(__free_pages);
807 fastcall void free_pages(unsigned long addr, unsigned int order)
810 BUG_ON(!virt_addr_valid((void *)addr));
811 __free_pages(virt_to_page((void *)addr), order);
815 EXPORT_SYMBOL(free_pages);
818 * Total amount of free (allocatable) RAM:
820 unsigned int nr_free_pages(void)
822 unsigned int sum = 0;
826 sum += zone->free_pages;
831 EXPORT_SYMBOL(nr_free_pages);
834 unsigned int nr_free_pages_pgdat(pg_data_t *pgdat)
836 unsigned int i, sum = 0;
838 for (i = 0; i < MAX_NR_ZONES; i++)
839 sum += pgdat->node_zones[i].free_pages;
845 static unsigned int nr_free_zone_pages(int offset)
848 unsigned int sum = 0;
850 for_each_pgdat(pgdat) {
851 struct zonelist *zonelist = pgdat->node_zonelists + offset;
852 struct zone **zonep = zonelist->zones;
855 for (zone = *zonep++; zone; zone = *zonep++) {
856 unsigned long size = zone->present_pages;
857 unsigned long high = zone->pages_high;
867 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
869 unsigned int nr_free_buffer_pages(void)
871 return nr_free_zone_pages(GFP_USER & GFP_ZONEMASK);
875 * Amount of free RAM allocatable within all zones
877 unsigned int nr_free_pagecache_pages(void)
879 return nr_free_zone_pages(GFP_HIGHUSER & GFP_ZONEMASK);
882 #ifdef CONFIG_HIGHMEM
883 unsigned int nr_free_highpages (void)
886 unsigned int pages = 0;
888 for_each_pgdat(pgdat)
889 pages += pgdat->node_zones[ZONE_HIGHMEM].free_pages;
896 static void show_node(struct zone *zone)
898 printk("Node %d ", zone->zone_pgdat->node_id);
901 #define show_node(zone) do { } while (0)
905 * Accumulate the page_state information across all CPUs.
906 * The result is unavoidably approximate - it can change
907 * during and after execution of this function.
909 DEFINE_PER_CPU(struct page_state, page_states) = {0};
910 EXPORT_PER_CPU_SYMBOL(page_states);
912 atomic_t nr_pagecache = ATOMIC_INIT(0);
913 EXPORT_SYMBOL(nr_pagecache);
915 DEFINE_PER_CPU(long, nr_pagecache_local) = 0;
918 void __get_page_state(struct page_state *ret, int nr)
922 memset(ret, 0, sizeof(*ret));
923 while (cpu < NR_CPUS) {
924 unsigned long *in, *out, off;
926 if (!cpu_possible(cpu)) {
931 in = (unsigned long *)&per_cpu(page_states, cpu);
933 if (cpu < NR_CPUS && cpu_possible(cpu))
934 prefetch(&per_cpu(page_states, cpu));
935 out = (unsigned long *)ret;
936 for (off = 0; off < nr; off++)
941 void get_page_state(struct page_state *ret)
945 nr = offsetof(struct page_state, GET_PAGE_STATE_LAST);
946 nr /= sizeof(unsigned long);
948 __get_page_state(ret, nr + 1);
951 void get_full_page_state(struct page_state *ret)
953 __get_page_state(ret, sizeof(*ret) / sizeof(unsigned long));
956 unsigned long __read_page_state(unsigned offset)
958 unsigned long ret = 0;
961 for (cpu = 0; cpu < NR_CPUS; cpu++) {
964 if (!cpu_possible(cpu))
967 in = (unsigned long)&per_cpu(page_states, cpu) + offset;
968 ret += *((unsigned long *)in);
973 void __get_zone_counts(unsigned long *active, unsigned long *inactive,
974 unsigned long *free, struct pglist_data *pgdat)
976 struct zone *zones = pgdat->node_zones;
982 for (i = 0; i < MAX_NR_ZONES; i++) {
983 *active += zones[i].nr_active;
984 *inactive += zones[i].nr_inactive;
985 *free += zones[i].free_pages;
989 void get_zone_counts(unsigned long *active,
990 unsigned long *inactive, unsigned long *free)
992 struct pglist_data *pgdat;
997 for_each_pgdat(pgdat) {
998 unsigned long l, m, n;
999 __get_zone_counts(&l, &m, &n, pgdat);
1006 void si_meminfo(struct sysinfo *val)
1008 val->totalram = totalram_pages;
1010 val->freeram = nr_free_pages();
1011 val->bufferram = nr_blockdev_pages();
1012 #ifdef CONFIG_HIGHMEM
1013 val->totalhigh = totalhigh_pages;
1014 val->freehigh = nr_free_highpages();
1019 val->mem_unit = PAGE_SIZE;
1022 EXPORT_SYMBOL(si_meminfo);
1025 void si_meminfo_node(struct sysinfo *val, int nid)
1027 pg_data_t *pgdat = NODE_DATA(nid);
1029 val->totalram = pgdat->node_present_pages;
1030 val->freeram = nr_free_pages_pgdat(pgdat);
1031 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
1032 val->freehigh = pgdat->node_zones[ZONE_HIGHMEM].free_pages;
1033 val->mem_unit = PAGE_SIZE;
1037 #define K(x) ((x) << (PAGE_SHIFT-10))
1040 * Show free area list (used inside shift_scroll-lock stuff)
1041 * We also calculate the percentage fragmentation. We do this by counting the
1042 * memory on each free list with the exception of the first item on the list.
1044 void show_free_areas(void)
1046 struct page_state ps;
1047 int cpu, temperature;
1048 unsigned long active;
1049 unsigned long inactive;
1053 for_each_zone(zone) {
1055 printk("%s per-cpu:", zone->name);
1057 if (!zone->present_pages) {
1063 for (cpu = 0; cpu < NR_CPUS; ++cpu) {
1064 struct per_cpu_pageset *pageset;
1066 if (!cpu_possible(cpu))
1069 pageset = zone->pageset + cpu;
1071 for (temperature = 0; temperature < 2; temperature++)
1072 printk("cpu %d %s: low %d, high %d, batch %d\n",
1074 temperature ? "cold" : "hot",
1075 pageset->pcp[temperature].low,
1076 pageset->pcp[temperature].high,
1077 pageset->pcp[temperature].batch);
1081 get_page_state(&ps);
1082 get_zone_counts(&active, &inactive, &free);
1084 printk("\nFree pages: %11ukB (%ukB HighMem)\n",
1086 K(nr_free_highpages()));
1088 printk("Active:%lu inactive:%lu dirty:%lu writeback:%lu "
1089 "unstable:%lu free:%u slab:%lu mapped:%lu pagetables:%lu\n",
1098 ps.nr_page_table_pages);
1100 for_each_zone(zone) {
1112 " pages_scanned:%lu"
1113 " all_unreclaimable? %s"
1116 K(zone->free_pages),
1119 K(zone->pages_high),
1121 K(zone->nr_inactive),
1122 K(zone->present_pages),
1123 zone->pages_scanned,
1124 (zone->all_unreclaimable ? "yes" : "no")
1126 printk("protections[]:");
1127 for (i = 0; i < MAX_NR_ZONES; i++)
1128 printk(" %lu", zone->protection[i]);
1132 for_each_zone(zone) {
1133 struct list_head *elem;
1134 unsigned long nr, flags, order, total = 0;
1137 printk("%s: ", zone->name);
1138 if (!zone->present_pages) {
1143 spin_lock_irqsave(&zone->lock, flags);
1144 for (order = 0; order < MAX_ORDER; order++) {
1146 list_for_each(elem, &zone->free_area[order].free_list)
1148 total += nr << order;
1149 printk("%lu*%lukB ", nr, K(1UL) << order);
1151 spin_unlock_irqrestore(&zone->lock, flags);
1152 printk("= %lukB\n", K(total));
1155 show_swap_cache_info();
1159 * Builds allocation fallback zone lists.
1161 static int __init build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist, int j, int k)
1168 zone = pgdat->node_zones + ZONE_HIGHMEM;
1169 if (zone->present_pages) {
1170 #ifndef CONFIG_HIGHMEM
1173 zonelist->zones[j++] = zone;
1176 zone = pgdat->node_zones + ZONE_NORMAL;
1177 if (zone->present_pages)
1178 zonelist->zones[j++] = zone;
1180 zone = pgdat->node_zones + ZONE_DMA;
1181 if (zone->present_pages)
1182 zonelist->zones[j++] = zone;
1189 #define MAX_NODE_LOAD (numnodes)
1190 static int __initdata node_load[MAX_NUMNODES];
1192 * find_next_best_node - find the next node that should appear in a given
1193 * node's fallback list
1194 * @node: node whose fallback list we're appending
1195 * @used_node_mask: pointer to the bitmap of already used nodes
1197 * We use a number of factors to determine which is the next node that should
1198 * appear on a given node's fallback list. The node should not have appeared
1199 * already in @node's fallback list, and it should be the next closest node
1200 * according to the distance array (which contains arbitrary distance values
1201 * from each node to each node in the system), and should also prefer nodes
1202 * with no CPUs, since presumably they'll have very little allocation pressure
1203 * on them otherwise.
1204 * It returns -1 if no node is found.
1206 static int __init find_next_best_node(int node, void *used_node_mask)
1209 int min_val = INT_MAX;
1212 for (i = 0; i < numnodes; i++) {
1215 /* Start from local node */
1216 n = (node+i)%numnodes;
1218 /* Don't want a node to appear more than once */
1219 if (test_bit(n, used_node_mask))
1222 /* Use the local node if we haven't already */
1223 if (!test_bit(node, used_node_mask)) {
1228 /* Use the distance array to find the distance */
1229 val = node_distance(node, n);
1231 /* Give preference to headless and unused nodes */
1232 tmp = node_to_cpumask(n);
1233 if (!cpus_empty(tmp))
1234 val += PENALTY_FOR_NODE_WITH_CPUS;
1236 /* Slight preference for less loaded node */
1237 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
1238 val += node_load[n];
1240 if (val < min_val) {
1247 set_bit(best_node, used_node_mask);
1252 static void __init build_zonelists(pg_data_t *pgdat)
1254 int i, j, k, node, local_node;
1255 int prev_node, load;
1256 struct zonelist *zonelist;
1257 DECLARE_BITMAP(used_mask, MAX_NUMNODES);
1259 /* initialize zonelists */
1260 for (i = 0; i < GFP_ZONETYPES; i++) {
1261 zonelist = pgdat->node_zonelists + i;
1262 memset(zonelist, 0, sizeof(*zonelist));
1263 zonelist->zones[0] = NULL;
1266 /* NUMA-aware ordering of nodes */
1267 local_node = pgdat->node_id;
1269 prev_node = local_node;
1270 bitmap_zero(used_mask, MAX_NUMNODES);
1271 while ((node = find_next_best_node(local_node, used_mask)) >= 0) {
1273 * We don't want to pressure a particular node.
1274 * So adding penalty to the first node in same
1275 * distance group to make it round-robin.
1277 if (node_distance(local_node, node) !=
1278 node_distance(local_node, prev_node))
1279 node_load[node] += load;
1282 for (i = 0; i < GFP_ZONETYPES; i++) {
1283 zonelist = pgdat->node_zonelists + i;
1284 for (j = 0; zonelist->zones[j] != NULL; j++);
1287 if (i & __GFP_HIGHMEM)
1292 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1293 zonelist->zones[j] = NULL;
1298 #else /* CONFIG_NUMA */
1300 static void __init build_zonelists(pg_data_t *pgdat)
1302 int i, j, k, node, local_node;
1304 local_node = pgdat->node_id;
1305 for (i = 0; i < GFP_ZONETYPES; i++) {
1306 struct zonelist *zonelist;
1308 zonelist = pgdat->node_zonelists + i;
1309 memset(zonelist, 0, sizeof(*zonelist));
1313 if (i & __GFP_HIGHMEM)
1318 j = build_zonelists_node(pgdat, zonelist, j, k);
1320 * Now we build the zonelist so that it contains the zones
1321 * of all the other nodes.
1322 * We don't want to pressure a particular node, so when
1323 * building the zones for node N, we make sure that the
1324 * zones coming right after the local ones are those from
1325 * node N+1 (modulo N)
1327 for (node = local_node + 1; node < numnodes; node++)
1328 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1329 for (node = 0; node < local_node; node++)
1330 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1332 zonelist->zones[j] = NULL;
1336 #endif /* CONFIG_NUMA */
1338 void __init build_all_zonelists(void)
1342 for(i = 0 ; i < numnodes ; i++)
1343 build_zonelists(NODE_DATA(i));
1344 printk("Built %i zonelists\n", numnodes);
1348 * Helper functions to size the waitqueue hash table.
1349 * Essentially these want to choose hash table sizes sufficiently
1350 * large so that collisions trying to wait on pages are rare.
1351 * But in fact, the number of active page waitqueues on typical
1352 * systems is ridiculously low, less than 200. So this is even
1353 * conservative, even though it seems large.
1355 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
1356 * waitqueues, i.e. the size of the waitq table given the number of pages.
1358 #define PAGES_PER_WAITQUEUE 256
1360 static inline unsigned long wait_table_size(unsigned long pages)
1362 unsigned long size = 1;
1364 pages /= PAGES_PER_WAITQUEUE;
1366 while (size < pages)
1370 * Once we have dozens or even hundreds of threads sleeping
1371 * on IO we've got bigger problems than wait queue collision.
1372 * Limit the size of the wait table to a reasonable size.
1374 size = min(size, 4096UL);
1376 return max(size, 4UL);
1380 * This is an integer logarithm so that shifts can be used later
1381 * to extract the more random high bits from the multiplicative
1382 * hash function before the remainder is taken.
1384 static inline unsigned long wait_table_bits(unsigned long size)
1389 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
1391 static void __init calculate_zone_totalpages(struct pglist_data *pgdat,
1392 unsigned long *zones_size, unsigned long *zholes_size)
1394 unsigned long realtotalpages, totalpages = 0;
1397 for (i = 0; i < MAX_NR_ZONES; i++)
1398 totalpages += zones_size[i];
1399 pgdat->node_spanned_pages = totalpages;
1401 realtotalpages = totalpages;
1403 for (i = 0; i < MAX_NR_ZONES; i++)
1404 realtotalpages -= zholes_size[i];
1405 pgdat->node_present_pages = realtotalpages;
1406 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id, realtotalpages);
1411 * Initially all pages are reserved - free ones are freed
1412 * up by free_all_bootmem() once the early boot process is
1413 * done. Non-atomic initialization, single-pass.
1415 void __init memmap_init_zone(unsigned long size, int nid, unsigned long zone,
1416 unsigned long start_pfn)
1418 struct page *start = pfn_to_page(start_pfn);
1421 for (page = start; page < (start + size); page++) {
1422 set_page_zone(page, NODEZONE(nid, zone));
1423 set_page_count(page, 0);
1424 reset_page_mapcount(page);
1425 SetPageReserved(page);
1426 INIT_LIST_HEAD(&page->lru);
1427 #ifdef WANT_PAGE_VIRTUAL
1428 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1429 if (!is_highmem_idx(zone))
1430 set_page_address(page, __va(start_pfn << PAGE_SHIFT));
1437 * Page buddy system uses "index >> (i+1)", where "index" is
1440 * The extra "+3" is to round down to byte size (8 bits per byte
1441 * assumption). Thus we get "(size-1) >> (i+4)" as the last byte
1444 * The "+1" is because we want to round the byte allocation up
1445 * rather than down. So we should have had a "+7" before we shifted
1446 * down by three. Also, we have to add one as we actually _use_ the
1447 * last bit (it's [0,n] inclusive, not [0,n[).
1449 * So we actually had +7+1 before we shift down by 3. But
1450 * (n+8) >> 3 == (n >> 3) + 1 (modulo overflows, which we do not have).
1452 * Finally, we LONG_ALIGN because all bitmap operations are on longs.
1454 unsigned long pages_to_bitmap_size(unsigned long order, unsigned long nr_pages)
1456 unsigned long bitmap_size;
1458 bitmap_size = (nr_pages-1) >> (order+4);
1459 bitmap_size = LONG_ALIGN(bitmap_size+1);
1464 void zone_init_free_lists(struct pglist_data *pgdat, struct zone *zone, unsigned long size)
1467 for (order = 0; ; order++) {
1468 unsigned long bitmap_size;
1470 INIT_LIST_HEAD(&zone->free_area[order].free_list);
1471 if (order == MAX_ORDER-1) {
1472 zone->free_area[order].map = NULL;
1476 bitmap_size = pages_to_bitmap_size(order, size);
1477 zone->free_area[order].map =
1478 (unsigned long *) alloc_bootmem_node(pgdat, bitmap_size);
1482 #ifndef __HAVE_ARCH_MEMMAP_INIT
1483 #define memmap_init(size, nid, zone, start_pfn) \
1484 memmap_init_zone((size), (nid), (zone), (start_pfn))
1488 * Set up the zone data structures:
1489 * - mark all pages reserved
1490 * - mark all memory queues empty
1491 * - clear the memory bitmaps
1493 static void __init free_area_init_core(struct pglist_data *pgdat,
1494 unsigned long *zones_size, unsigned long *zholes_size)
1497 const unsigned long zone_required_alignment = 1UL << (MAX_ORDER-1);
1498 int cpu, nid = pgdat->node_id;
1499 unsigned long zone_start_pfn = pgdat->node_start_pfn;
1501 pgdat->nr_zones = 0;
1502 init_waitqueue_head(&pgdat->kswapd_wait);
1504 for (j = 0; j < MAX_NR_ZONES; j++) {
1505 struct zone *zone = pgdat->node_zones + j;
1506 unsigned long size, realsize;
1507 unsigned long batch;
1509 zone_table[NODEZONE(nid, j)] = zone;
1510 realsize = size = zones_size[j];
1512 realsize -= zholes_size[j];
1514 if (j == ZONE_DMA || j == ZONE_NORMAL)
1515 nr_kernel_pages += realsize;
1516 nr_all_pages += realsize;
1518 zone->spanned_pages = size;
1519 zone->present_pages = realsize;
1520 zone->name = zone_names[j];
1521 spin_lock_init(&zone->lock);
1522 spin_lock_init(&zone->lru_lock);
1523 zone->zone_pgdat = pgdat;
1524 zone->free_pages = 0;
1526 zone->temp_priority = zone->prev_priority = DEF_PRIORITY;
1529 * The per-cpu-pages pools are set to around 1000th of the
1530 * size of the zone. But no more than 1/4 of a meg - there's
1531 * no point in going beyond the size of L2 cache.
1533 * OK, so we don't know how big the cache is. So guess.
1535 batch = zone->present_pages / 1024;
1536 if (batch * PAGE_SIZE > 256 * 1024)
1537 batch = (256 * 1024) / PAGE_SIZE;
1538 batch /= 4; /* We effectively *= 4 below */
1542 for (cpu = 0; cpu < NR_CPUS; cpu++) {
1543 struct per_cpu_pages *pcp;
1545 pcp = &zone->pageset[cpu].pcp[0]; /* hot */
1547 pcp->low = 2 * batch;
1548 pcp->high = 6 * batch;
1549 pcp->batch = 1 * batch;
1550 INIT_LIST_HEAD(&pcp->list);
1552 pcp = &zone->pageset[cpu].pcp[1]; /* cold */
1555 pcp->high = 2 * batch;
1556 pcp->batch = 1 * batch;
1557 INIT_LIST_HEAD(&pcp->list);
1559 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%lu\n",
1560 zone_names[j], realsize, batch);
1561 ckrm_init_lists(zone);
1562 zone->nr_scan_active = 0;
1563 zone->nr_scan_inactive = 0;
1564 zone->nr_active = 0;
1565 zone->nr_inactive = 0;
1570 * The per-page waitqueue mechanism uses hashed waitqueues
1573 zone->wait_table_size = wait_table_size(size);
1574 zone->wait_table_bits =
1575 wait_table_bits(zone->wait_table_size);
1576 zone->wait_table = (wait_queue_head_t *)
1577 alloc_bootmem_node(pgdat, zone->wait_table_size
1578 * sizeof(wait_queue_head_t));
1580 for(i = 0; i < zone->wait_table_size; ++i)
1581 init_waitqueue_head(zone->wait_table + i);
1583 pgdat->nr_zones = j+1;
1585 zone->zone_mem_map = pfn_to_page(zone_start_pfn);
1586 zone->zone_start_pfn = zone_start_pfn;
1588 if ((zone_start_pfn) & (zone_required_alignment-1))
1589 printk("BUG: wrong zone alignment, it will crash\n");
1591 memmap_init(size, nid, j, zone_start_pfn);
1593 zone_start_pfn += size;
1595 zone_init_free_lists(pgdat, zone, zone->spanned_pages);
1599 void __init node_alloc_mem_map(struct pglist_data *pgdat)
1603 size = (pgdat->node_spanned_pages + 1) * sizeof(struct page);
1604 pgdat->node_mem_map = alloc_bootmem_node(pgdat, size);
1605 #ifndef CONFIG_DISCONTIGMEM
1606 mem_map = contig_page_data.node_mem_map;
1610 void __init free_area_init_node(int nid, struct pglist_data *pgdat,
1611 unsigned long *zones_size, unsigned long node_start_pfn,
1612 unsigned long *zholes_size)
1614 pgdat->node_id = nid;
1615 pgdat->node_start_pfn = node_start_pfn;
1616 calculate_zone_totalpages(pgdat, zones_size, zholes_size);
1618 if (!pfn_to_page(node_start_pfn))
1619 node_alloc_mem_map(pgdat);
1621 free_area_init_core(pgdat, zones_size, zholes_size);
1624 #ifndef CONFIG_DISCONTIGMEM
1625 static bootmem_data_t contig_bootmem_data;
1626 struct pglist_data contig_page_data = { .bdata = &contig_bootmem_data };
1628 EXPORT_SYMBOL(contig_page_data);
1630 void __init free_area_init(unsigned long *zones_size)
1632 free_area_init_node(0, &contig_page_data, zones_size,
1633 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
1637 #ifdef CONFIG_PROC_FS
1639 #include <linux/seq_file.h>
1641 static void *frag_start(struct seq_file *m, loff_t *pos)
1646 for (pgdat = pgdat_list; pgdat && node; pgdat = pgdat->pgdat_next)
1652 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
1654 pg_data_t *pgdat = (pg_data_t *)arg;
1657 return pgdat->pgdat_next;
1660 static void frag_stop(struct seq_file *m, void *arg)
1665 * This walks the freelist for each zone. Whilst this is slow, I'd rather
1666 * be slow here than slow down the fast path by keeping stats - mjbligh
1668 static int frag_show(struct seq_file *m, void *arg)
1670 pg_data_t *pgdat = (pg_data_t *)arg;
1672 struct zone *node_zones = pgdat->node_zones;
1673 unsigned long flags;
1676 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
1677 if (!zone->present_pages)
1680 spin_lock_irqsave(&zone->lock, flags);
1681 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1682 for (order = 0; order < MAX_ORDER; ++order) {
1683 unsigned long nr_bufs = 0;
1684 struct list_head *elem;
1686 list_for_each(elem, &(zone->free_area[order].free_list))
1688 seq_printf(m, "%6lu ", nr_bufs);
1690 spin_unlock_irqrestore(&zone->lock, flags);
1696 struct seq_operations fragmentation_op = {
1697 .start = frag_start,
1703 static char *vmstat_text[] = {
1707 "nr_page_table_pages",
1732 "pgscan_kswapd_high",
1733 "pgscan_kswapd_normal",
1735 "pgscan_kswapd_dma",
1736 "pgscan_direct_high",
1737 "pgscan_direct_normal",
1738 "pgscan_direct_dma",
1743 "kswapd_inodesteal",
1750 static void *vmstat_start(struct seq_file *m, loff_t *pos)
1752 struct page_state *ps;
1754 if (*pos >= ARRAY_SIZE(vmstat_text))
1757 ps = kmalloc(sizeof(*ps), GFP_KERNEL);
1760 return ERR_PTR(-ENOMEM);
1761 get_full_page_state(ps);
1762 ps->pgpgin /= 2; /* sectors -> kbytes */
1764 return (unsigned long *)ps + *pos;
1767 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
1770 if (*pos >= ARRAY_SIZE(vmstat_text))
1772 return (unsigned long *)m->private + *pos;
1775 static int vmstat_show(struct seq_file *m, void *arg)
1777 unsigned long *l = arg;
1778 unsigned long off = l - (unsigned long *)m->private;
1780 seq_printf(m, "%s %lu\n", vmstat_text[off], *l);
1784 static void vmstat_stop(struct seq_file *m, void *arg)
1790 struct seq_operations vmstat_op = {
1791 .start = vmstat_start,
1792 .next = vmstat_next,
1793 .stop = vmstat_stop,
1794 .show = vmstat_show,
1797 #endif /* CONFIG_PROC_FS */
1799 #ifdef CONFIG_HOTPLUG_CPU
1800 static int page_alloc_cpu_notify(struct notifier_block *self,
1801 unsigned long action, void *hcpu)
1803 int cpu = (unsigned long)hcpu;
1806 if (action == CPU_DEAD) {
1807 /* Drain local pagecache count. */
1808 count = &per_cpu(nr_pagecache_local, cpu);
1809 atomic_add(*count, &nr_pagecache);
1811 local_irq_disable();
1817 #endif /* CONFIG_HOTPLUG_CPU */
1819 void __init page_alloc_init(void)
1821 hotcpu_notifier(page_alloc_cpu_notify, 0);
1824 static unsigned long higherzone_val(struct zone *z, int max_zone,
1827 int z_idx = zone_idx(z);
1828 struct zone *higherzone;
1829 unsigned long pages;
1831 /* there is no higher zone to get a contribution from */
1832 if (z_idx == MAX_NR_ZONES-1)
1835 higherzone = &z->zone_pgdat->node_zones[z_idx+1];
1837 /* We always start with the higher zone's protection value */
1838 pages = higherzone->protection[alloc_type];
1841 * We get a lower-zone-protection contribution only if there are
1842 * pages in the higher zone and if we're not the highest zone
1843 * in the current zonelist. e.g., never happens for GFP_DMA. Happens
1844 * only for ZONE_DMA in a GFP_KERNEL allocation and happens for ZONE_DMA
1845 * and ZONE_NORMAL for a GFP_HIGHMEM allocation.
1847 if (higherzone->present_pages && z_idx < alloc_type)
1848 pages += higherzone->pages_low * sysctl_lower_zone_protection;
1854 * setup_per_zone_protection - called whenver min_free_kbytes or
1855 * sysctl_lower_zone_protection changes. Ensures that each zone
1856 * has a correct pages_protected value, so an adequate number of
1857 * pages are left in the zone after a successful __alloc_pages().
1859 * This algorithm is way confusing. I tries to keep the same behavior
1860 * as we had with the incremental min iterative algorithm.
1862 static void setup_per_zone_protection(void)
1864 struct pglist_data *pgdat;
1865 struct zone *zones, *zone;
1869 for_each_pgdat(pgdat) {
1870 zones = pgdat->node_zones;
1872 for (i = 0, max_zone = 0; i < MAX_NR_ZONES; i++)
1873 if (zones[i].present_pages)
1877 * For each of the different allocation types:
1878 * GFP_DMA -> GFP_KERNEL -> GFP_HIGHMEM
1880 for (i = 0; i < GFP_ZONETYPES; i++) {
1882 * For each of the zones:
1883 * ZONE_HIGHMEM -> ZONE_NORMAL -> ZONE_DMA
1885 for (j = MAX_NR_ZONES-1; j >= 0; j--) {
1889 * We never protect zones that don't have memory
1890 * in them (j>max_zone) or zones that aren't in
1891 * the zonelists for a certain type of
1892 * allocation (j>=i). We have to assign these
1893 * to zero because the lower zones take
1894 * contributions from the higher zones.
1896 if (j > max_zone || j >= i) {
1897 zone->protection[i] = 0;
1901 * The contribution of the next higher zone
1903 zone->protection[i] = higherzone_val(zone,
1911 * setup_per_zone_pages_min - called when min_free_kbytes changes. Ensures
1912 * that the pages_{min,low,high} values for each zone are set correctly
1913 * with respect to min_free_kbytes.
1915 static void setup_per_zone_pages_min(void)
1917 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
1918 unsigned long lowmem_pages = 0;
1920 unsigned long flags;
1922 /* Calculate total number of !ZONE_HIGHMEM pages */
1923 for_each_zone(zone) {
1924 if (!is_highmem(zone))
1925 lowmem_pages += zone->present_pages;
1928 for_each_zone(zone) {
1929 spin_lock_irqsave(&zone->lru_lock, flags);
1930 if (is_highmem(zone)) {
1932 * Often, highmem doesn't need to reserve any pages.
1933 * But the pages_min/low/high values are also used for
1934 * batching up page reclaim activity so we need a
1935 * decent value here.
1939 min_pages = zone->present_pages / 1024;
1940 if (min_pages < SWAP_CLUSTER_MAX)
1941 min_pages = SWAP_CLUSTER_MAX;
1942 if (min_pages > 128)
1944 zone->pages_min = min_pages;
1946 /* if it's a lowmem zone, reserve a number of pages
1947 * proportionate to the zone's size.
1949 zone->pages_min = (pages_min * zone->present_pages) /
1954 * When interpreting these watermarks, just keep in mind that:
1955 * zone->pages_min == (zone->pages_min * 4) / 4;
1957 zone->pages_low = (zone->pages_min * 5) / 4;
1958 zone->pages_high = (zone->pages_min * 6) / 4;
1959 spin_unlock_irqrestore(&zone->lru_lock, flags);
1964 * Initialise min_free_kbytes.
1966 * For small machines we want it small (128k min). For large machines
1967 * we want it large (64MB max). But it is not linear, because network
1968 * bandwidth does not increase linearly with machine size. We use
1970 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
1971 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
1987 static int __init init_per_zone_pages_min(void)
1989 unsigned long lowmem_kbytes;
1991 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
1993 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
1994 if (min_free_kbytes < 128)
1995 min_free_kbytes = 128;
1996 if (min_free_kbytes > 65536)
1997 min_free_kbytes = 65536;
1998 setup_per_zone_pages_min();
1999 setup_per_zone_protection();
2002 module_init(init_per_zone_pages_min)
2005 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
2006 * that we can call two helper functions whenever min_free_kbytes
2009 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
2010 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2012 proc_dointvec(table, write, file, buffer, length, ppos);
2013 setup_per_zone_pages_min();
2014 setup_per_zone_protection();
2019 * lower_zone_protection_sysctl_handler - just a wrapper around
2020 * proc_dointvec() so that we can call setup_per_zone_protection()
2021 * whenever sysctl_lower_zone_protection changes.
2023 int lower_zone_protection_sysctl_handler(ctl_table *table, int write,
2024 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2026 proc_dointvec_minmax(table, write, file, buffer, length, ppos);
2027 setup_per_zone_protection();
2032 * allocate a large system hash table from bootmem
2033 * - it is assumed that the hash table must contain an exact power-of-2
2034 * quantity of entries
2036 void *__init alloc_large_system_hash(const char *tablename,
2037 unsigned long bucketsize,
2038 unsigned long numentries,
2040 int consider_highmem,
2041 unsigned int *_hash_shift,
2042 unsigned int *_hash_mask)
2044 unsigned long long max;
2045 unsigned long log2qty, size;
2048 /* allow the kernel cmdline to have a say */
2050 /* round applicable memory size up to nearest megabyte */
2051 numentries = consider_highmem ? nr_all_pages : nr_kernel_pages;
2052 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
2053 numentries >>= 20 - PAGE_SHIFT;
2054 numentries <<= 20 - PAGE_SHIFT;
2056 /* limit to 1 bucket per 2^scale bytes of low memory */
2057 if (scale > PAGE_SHIFT)
2058 numentries >>= (scale - PAGE_SHIFT);
2060 numentries <<= (PAGE_SHIFT - scale);
2062 /* rounded up to nearest power of 2 in size */
2063 numentries = 1UL << (long_log2(numentries) + 1);
2065 /* limit allocation size to 1/16 total memory */
2066 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
2067 do_div(max, bucketsize);
2069 if (numentries > max)
2072 log2qty = long_log2(numentries);
2075 size = bucketsize << log2qty;
2076 table = alloc_bootmem(size);
2077 } while (!table && size > PAGE_SIZE && --log2qty);
2080 panic("Failed to allocate %s hash table\n", tablename);
2082 printk("%s hash table entries: %d (order: %d, %lu bytes)\n",
2085 long_log2(size) - PAGE_SHIFT,
2089 *_hash_shift = log2qty;
2091 *_hash_mask = (1 << log2qty) - 1;