semaphore: __down_common: use signal_pending_state()
[linux-2.6/mini2440.git] / mm / page_alloc.c
blob401d104d2bb6e12c08e70700eebad915e72c7c2a
1 /*
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/stddef.h>
18 #include <linux/mm.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/compiler.h>
25 #include <linux/kernel.h>
26 #include <linux/module.h>
27 #include <linux/suspend.h>
28 #include <linux/pagevec.h>
29 #include <linux/blkdev.h>
30 #include <linux/slab.h>
31 #include <linux/oom.h>
32 #include <linux/notifier.h>
33 #include <linux/topology.h>
34 #include <linux/sysctl.h>
35 #include <linux/cpu.h>
36 #include <linux/cpuset.h>
37 #include <linux/memory_hotplug.h>
38 #include <linux/nodemask.h>
39 #include <linux/vmalloc.h>
40 #include <linux/mempolicy.h>
41 #include <linux/stop_machine.h>
42 #include <linux/sort.h>
43 #include <linux/pfn.h>
44 #include <linux/backing-dev.h>
45 #include <linux/fault-inject.h>
46 #include <linux/page-isolation.h>
47 #include <linux/memcontrol.h>
48 #include <linux/debugobjects.h>
50 #include <asm/tlbflush.h>
51 #include <asm/div64.h>
52 #include "internal.h"
55 * Array of node states.
57 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
58 [N_POSSIBLE] = NODE_MASK_ALL,
59 [N_ONLINE] = { { [0] = 1UL } },
60 #ifndef CONFIG_NUMA
61 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
62 #ifdef CONFIG_HIGHMEM
63 [N_HIGH_MEMORY] = { { [0] = 1UL } },
64 #endif
65 [N_CPU] = { { [0] = 1UL } },
66 #endif /* NUMA */
68 EXPORT_SYMBOL(node_states);
70 unsigned long totalram_pages __read_mostly;
71 unsigned long totalreserve_pages __read_mostly;
72 long nr_swap_pages;
73 int percpu_pagelist_fraction;
75 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
76 int pageblock_order __read_mostly;
77 #endif
79 static void __free_pages_ok(struct page *page, unsigned int order);
82 * results with 256, 32 in the lowmem_reserve sysctl:
83 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
84 * 1G machine -> (16M dma, 784M normal, 224M high)
85 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
86 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
87 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
89 * TBD: should special case ZONE_DMA32 machines here - in those we normally
90 * don't need any ZONE_NORMAL reservation
92 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
93 #ifdef CONFIG_ZONE_DMA
94 256,
95 #endif
96 #ifdef CONFIG_ZONE_DMA32
97 256,
98 #endif
99 #ifdef CONFIG_HIGHMEM
101 #endif
105 EXPORT_SYMBOL(totalram_pages);
107 static char * const zone_names[MAX_NR_ZONES] = {
108 #ifdef CONFIG_ZONE_DMA
109 "DMA",
110 #endif
111 #ifdef CONFIG_ZONE_DMA32
112 "DMA32",
113 #endif
114 "Normal",
115 #ifdef CONFIG_HIGHMEM
116 "HighMem",
117 #endif
118 "Movable",
121 int min_free_kbytes = 1024;
123 unsigned long __meminitdata nr_kernel_pages;
124 unsigned long __meminitdata nr_all_pages;
125 static unsigned long __meminitdata dma_reserve;
127 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
129 * MAX_ACTIVE_REGIONS determines the maximum number of distinct
130 * ranges of memory (RAM) that may be registered with add_active_range().
131 * Ranges passed to add_active_range() will be merged if possible
132 * so the number of times add_active_range() can be called is
133 * related to the number of nodes and the number of holes
135 #ifdef CONFIG_MAX_ACTIVE_REGIONS
136 /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */
137 #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS
138 #else
139 #if MAX_NUMNODES >= 32
140 /* If there can be many nodes, allow up to 50 holes per node */
141 #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50)
142 #else
143 /* By default, allow up to 256 distinct regions */
144 #define MAX_ACTIVE_REGIONS 256
145 #endif
146 #endif
148 static struct node_active_region __meminitdata early_node_map[MAX_ACTIVE_REGIONS];
149 static int __meminitdata nr_nodemap_entries;
150 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
151 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
152 #ifdef CONFIG_MEMORY_HOTPLUG_RESERVE
153 static unsigned long __meminitdata node_boundary_start_pfn[MAX_NUMNODES];
154 static unsigned long __meminitdata node_boundary_end_pfn[MAX_NUMNODES];
155 #endif /* CONFIG_MEMORY_HOTPLUG_RESERVE */
156 static unsigned long __initdata required_kernelcore;
157 static unsigned long __initdata required_movablecore;
158 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
160 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
161 int movable_zone;
162 EXPORT_SYMBOL(movable_zone);
163 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
165 #if MAX_NUMNODES > 1
166 int nr_node_ids __read_mostly = MAX_NUMNODES;
167 EXPORT_SYMBOL(nr_node_ids);
168 #endif
170 int page_group_by_mobility_disabled __read_mostly;
172 static void set_pageblock_migratetype(struct page *page, int migratetype)
174 set_pageblock_flags_group(page, (unsigned long)migratetype,
175 PB_migrate, PB_migrate_end);
178 #ifdef CONFIG_DEBUG_VM
179 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
181 int ret = 0;
182 unsigned seq;
183 unsigned long pfn = page_to_pfn(page);
185 do {
186 seq = zone_span_seqbegin(zone);
187 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
188 ret = 1;
189 else if (pfn < zone->zone_start_pfn)
190 ret = 1;
191 } while (zone_span_seqretry(zone, seq));
193 return ret;
196 static int page_is_consistent(struct zone *zone, struct page *page)
198 if (!pfn_valid_within(page_to_pfn(page)))
199 return 0;
200 if (zone != page_zone(page))
201 return 0;
203 return 1;
206 * Temporary debugging check for pages not lying within a given zone.
208 static int bad_range(struct zone *zone, struct page *page)
210 if (page_outside_zone_boundaries(zone, page))
211 return 1;
212 if (!page_is_consistent(zone, page))
213 return 1;
215 return 0;
217 #else
218 static inline int bad_range(struct zone *zone, struct page *page)
220 return 0;
222 #endif
224 static void bad_page(struct page *page)
226 void *pc = page_get_page_cgroup(page);
228 printk(KERN_EMERG "Bad page state in process '%s'\n" KERN_EMERG
229 "page:%p flags:0x%0*lx mapping:%p mapcount:%d count:%d\n",
230 current->comm, page, (int)(2*sizeof(unsigned long)),
231 (unsigned long)page->flags, page->mapping,
232 page_mapcount(page), page_count(page));
233 if (pc) {
234 printk(KERN_EMERG "cgroup:%p\n", pc);
235 page_reset_bad_cgroup(page);
237 printk(KERN_EMERG "Trying to fix it up, but a reboot is needed\n"
238 KERN_EMERG "Backtrace:\n");
239 dump_stack();
240 page->flags &= ~PAGE_FLAGS_CLEAR_WHEN_BAD;
241 set_page_count(page, 0);
242 reset_page_mapcount(page);
243 page->mapping = NULL;
244 add_taint(TAINT_BAD_PAGE);
248 * Higher-order pages are called "compound pages". They are structured thusly:
250 * The first PAGE_SIZE page is called the "head page".
252 * The remaining PAGE_SIZE pages are called "tail pages".
254 * All pages have PG_compound set. All pages have their ->private pointing at
255 * the head page (even the head page has this).
257 * The first tail page's ->lru.next holds the address of the compound page's
258 * put_page() function. Its ->lru.prev holds the order of allocation.
259 * This usage means that zero-order pages may not be compound.
262 static void free_compound_page(struct page *page)
264 __free_pages_ok(page, compound_order(page));
267 void prep_compound_page(struct page *page, unsigned long order)
269 int i;
270 int nr_pages = 1 << order;
272 set_compound_page_dtor(page, free_compound_page);
273 set_compound_order(page, order);
274 __SetPageHead(page);
275 for (i = 1; i < nr_pages; i++) {
276 struct page *p = page + i;
278 __SetPageTail(p);
279 p->first_page = page;
283 static void destroy_compound_page(struct page *page, unsigned long order)
285 int i;
286 int nr_pages = 1 << order;
288 if (unlikely(compound_order(page) != order))
289 bad_page(page);
291 if (unlikely(!PageHead(page)))
292 bad_page(page);
293 __ClearPageHead(page);
294 for (i = 1; i < nr_pages; i++) {
295 struct page *p = page + i;
297 if (unlikely(!PageTail(p) |
298 (p->first_page != page)))
299 bad_page(page);
300 __ClearPageTail(p);
304 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
306 int i;
309 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
310 * and __GFP_HIGHMEM from hard or soft interrupt context.
312 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
313 for (i = 0; i < (1 << order); i++)
314 clear_highpage(page + i);
317 static inline void set_page_order(struct page *page, int order)
319 set_page_private(page, order);
320 __SetPageBuddy(page);
323 static inline void rmv_page_order(struct page *page)
325 __ClearPageBuddy(page);
326 set_page_private(page, 0);
330 * Locate the struct page for both the matching buddy in our
331 * pair (buddy1) and the combined O(n+1) page they form (page).
333 * 1) Any buddy B1 will have an order O twin B2 which satisfies
334 * the following equation:
335 * B2 = B1 ^ (1 << O)
336 * For example, if the starting buddy (buddy2) is #8 its order
337 * 1 buddy is #10:
338 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
340 * 2) Any buddy B will have an order O+1 parent P which
341 * satisfies the following equation:
342 * P = B & ~(1 << O)
344 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
346 static inline struct page *
347 __page_find_buddy(struct page *page, unsigned long page_idx, unsigned int order)
349 unsigned long buddy_idx = page_idx ^ (1 << order);
351 return page + (buddy_idx - page_idx);
354 static inline unsigned long
355 __find_combined_index(unsigned long page_idx, unsigned int order)
357 return (page_idx & ~(1 << order));
361 * This function checks whether a page is free && is the buddy
362 * we can do coalesce a page and its buddy if
363 * (a) the buddy is not in a hole &&
364 * (b) the buddy is in the buddy system &&
365 * (c) a page and its buddy have the same order &&
366 * (d) a page and its buddy are in the same zone.
368 * For recording whether a page is in the buddy system, we use PG_buddy.
369 * Setting, clearing, and testing PG_buddy is serialized by zone->lock.
371 * For recording page's order, we use page_private(page).
373 static inline int page_is_buddy(struct page *page, struct page *buddy,
374 int order)
376 if (!pfn_valid_within(page_to_pfn(buddy)))
377 return 0;
379 if (page_zone_id(page) != page_zone_id(buddy))
380 return 0;
382 if (PageBuddy(buddy) && page_order(buddy) == order) {
383 BUG_ON(page_count(buddy) != 0);
384 return 1;
386 return 0;
390 * Freeing function for a buddy system allocator.
392 * The concept of a buddy system is to maintain direct-mapped table
393 * (containing bit values) for memory blocks of various "orders".
394 * The bottom level table contains the map for the smallest allocatable
395 * units of memory (here, pages), and each level above it describes
396 * pairs of units from the levels below, hence, "buddies".
397 * At a high level, all that happens here is marking the table entry
398 * at the bottom level available, and propagating the changes upward
399 * as necessary, plus some accounting needed to play nicely with other
400 * parts of the VM system.
401 * At each level, we keep a list of pages, which are heads of continuous
402 * free pages of length of (1 << order) and marked with PG_buddy. Page's
403 * order is recorded in page_private(page) field.
404 * So when we are allocating or freeing one, we can derive the state of the
405 * other. That is, if we allocate a small block, and both were
406 * free, the remainder of the region must be split into blocks.
407 * If a block is freed, and its buddy is also free, then this
408 * triggers coalescing into a block of larger size.
410 * -- wli
413 static inline void __free_one_page(struct page *page,
414 struct zone *zone, unsigned int order)
416 unsigned long page_idx;
417 int order_size = 1 << order;
418 int migratetype = get_pageblock_migratetype(page);
420 if (unlikely(PageCompound(page)))
421 destroy_compound_page(page, order);
423 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
425 VM_BUG_ON(page_idx & (order_size - 1));
426 VM_BUG_ON(bad_range(zone, page));
428 __mod_zone_page_state(zone, NR_FREE_PAGES, order_size);
429 while (order < MAX_ORDER-1) {
430 unsigned long combined_idx;
431 struct page *buddy;
433 buddy = __page_find_buddy(page, page_idx, order);
434 if (!page_is_buddy(page, buddy, order))
435 break;
437 /* Our buddy is free, merge with it and move up one order. */
438 list_del(&buddy->lru);
439 zone->free_area[order].nr_free--;
440 rmv_page_order(buddy);
441 combined_idx = __find_combined_index(page_idx, order);
442 page = page + (combined_idx - page_idx);
443 page_idx = combined_idx;
444 order++;
446 set_page_order(page, order);
447 list_add(&page->lru,
448 &zone->free_area[order].free_list[migratetype]);
449 zone->free_area[order].nr_free++;
452 static inline int free_pages_check(struct page *page)
454 if (unlikely(page_mapcount(page) |
455 (page->mapping != NULL) |
456 (page_get_page_cgroup(page) != NULL) |
457 (page_count(page) != 0) |
458 (page->flags & PAGE_FLAGS_CHECK_AT_FREE)))
459 bad_page(page);
460 if (PageDirty(page))
461 __ClearPageDirty(page);
463 * For now, we report if PG_reserved was found set, but do not
464 * clear it, and do not free the page. But we shall soon need
465 * to do more, for when the ZERO_PAGE count wraps negative.
467 return PageReserved(page);
471 * Frees a list of pages.
472 * Assumes all pages on list are in same zone, and of same order.
473 * count is the number of pages to free.
475 * If the zone was previously in an "all pages pinned" state then look to
476 * see if this freeing clears that state.
478 * And clear the zone's pages_scanned counter, to hold off the "all pages are
479 * pinned" detection logic.
481 static void free_pages_bulk(struct zone *zone, int count,
482 struct list_head *list, int order)
484 spin_lock(&zone->lock);
485 zone_clear_flag(zone, ZONE_ALL_UNRECLAIMABLE);
486 zone->pages_scanned = 0;
487 while (count--) {
488 struct page *page;
490 VM_BUG_ON(list_empty(list));
491 page = list_entry(list->prev, struct page, lru);
492 /* have to delete it as __free_one_page list manipulates */
493 list_del(&page->lru);
494 __free_one_page(page, zone, order);
496 spin_unlock(&zone->lock);
499 static void free_one_page(struct zone *zone, struct page *page, int order)
501 spin_lock(&zone->lock);
502 zone_clear_flag(zone, ZONE_ALL_UNRECLAIMABLE);
503 zone->pages_scanned = 0;
504 __free_one_page(page, zone, order);
505 spin_unlock(&zone->lock);
508 static void __free_pages_ok(struct page *page, unsigned int order)
510 unsigned long flags;
511 int i;
512 int reserved = 0;
514 for (i = 0 ; i < (1 << order) ; ++i)
515 reserved += free_pages_check(page + i);
516 if (reserved)
517 return;
519 if (!PageHighMem(page)) {
520 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
521 debug_check_no_obj_freed(page_address(page),
522 PAGE_SIZE << order);
524 arch_free_page(page, order);
525 kernel_map_pages(page, 1 << order, 0);
527 local_irq_save(flags);
528 __count_vm_events(PGFREE, 1 << order);
529 free_one_page(page_zone(page), page, order);
530 local_irq_restore(flags);
534 * permit the bootmem allocator to evade page validation on high-order frees
536 void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
538 if (order == 0) {
539 __ClearPageReserved(page);
540 set_page_count(page, 0);
541 set_page_refcounted(page);
542 __free_page(page);
543 } else {
544 int loop;
546 prefetchw(page);
547 for (loop = 0; loop < BITS_PER_LONG; loop++) {
548 struct page *p = &page[loop];
550 if (loop + 1 < BITS_PER_LONG)
551 prefetchw(p + 1);
552 __ClearPageReserved(p);
553 set_page_count(p, 0);
556 set_page_refcounted(page);
557 __free_pages(page, order);
563 * The order of subdivision here is critical for the IO subsystem.
564 * Please do not alter this order without good reasons and regression
565 * testing. Specifically, as large blocks of memory are subdivided,
566 * the order in which smaller blocks are delivered depends on the order
567 * they're subdivided in this function. This is the primary factor
568 * influencing the order in which pages are delivered to the IO
569 * subsystem according to empirical testing, and this is also justified
570 * by considering the behavior of a buddy system containing a single
571 * large block of memory acted on by a series of small allocations.
572 * This behavior is a critical factor in sglist merging's success.
574 * -- wli
576 static inline void expand(struct zone *zone, struct page *page,
577 int low, int high, struct free_area *area,
578 int migratetype)
580 unsigned long size = 1 << high;
582 while (high > low) {
583 area--;
584 high--;
585 size >>= 1;
586 VM_BUG_ON(bad_range(zone, &page[size]));
587 list_add(&page[size].lru, &area->free_list[migratetype]);
588 area->nr_free++;
589 set_page_order(&page[size], high);
594 * This page is about to be returned from the page allocator
596 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
598 if (unlikely(page_mapcount(page) |
599 (page->mapping != NULL) |
600 (page_get_page_cgroup(page) != NULL) |
601 (page_count(page) != 0) |
602 (page->flags & PAGE_FLAGS_CHECK_AT_PREP)))
603 bad_page(page);
606 * For now, we report if PG_reserved was found set, but do not
607 * clear it, and do not allocate the page: as a safety net.
609 if (PageReserved(page))
610 return 1;
612 page->flags &= ~(1 << PG_uptodate | 1 << PG_error | 1 << PG_reclaim |
613 1 << PG_referenced | 1 << PG_arch_1 |
614 1 << PG_owner_priv_1 | 1 << PG_mappedtodisk);
615 set_page_private(page, 0);
616 set_page_refcounted(page);
618 arch_alloc_page(page, order);
619 kernel_map_pages(page, 1 << order, 1);
621 if (gfp_flags & __GFP_ZERO)
622 prep_zero_page(page, order, gfp_flags);
624 if (order && (gfp_flags & __GFP_COMP))
625 prep_compound_page(page, order);
627 return 0;
631 * Go through the free lists for the given migratetype and remove
632 * the smallest available page from the freelists
634 static struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
635 int migratetype)
637 unsigned int current_order;
638 struct free_area * area;
639 struct page *page;
641 /* Find a page of the appropriate size in the preferred list */
642 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
643 area = &(zone->free_area[current_order]);
644 if (list_empty(&area->free_list[migratetype]))
645 continue;
647 page = list_entry(area->free_list[migratetype].next,
648 struct page, lru);
649 list_del(&page->lru);
650 rmv_page_order(page);
651 area->nr_free--;
652 __mod_zone_page_state(zone, NR_FREE_PAGES, - (1UL << order));
653 expand(zone, page, order, current_order, area, migratetype);
654 return page;
657 return NULL;
662 * This array describes the order lists are fallen back to when
663 * the free lists for the desirable migrate type are depleted
665 static int fallbacks[MIGRATE_TYPES][MIGRATE_TYPES-1] = {
666 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
667 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
668 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
669 [MIGRATE_RESERVE] = { MIGRATE_RESERVE, MIGRATE_RESERVE, MIGRATE_RESERVE }, /* Never used */
673 * Move the free pages in a range to the free lists of the requested type.
674 * Note that start_page and end_pages are not aligned on a pageblock
675 * boundary. If alignment is required, use move_freepages_block()
677 static int move_freepages(struct zone *zone,
678 struct page *start_page, struct page *end_page,
679 int migratetype)
681 struct page *page;
682 unsigned long order;
683 int pages_moved = 0;
685 #ifndef CONFIG_HOLES_IN_ZONE
687 * page_zone is not safe to call in this context when
688 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
689 * anyway as we check zone boundaries in move_freepages_block().
690 * Remove at a later date when no bug reports exist related to
691 * grouping pages by mobility
693 BUG_ON(page_zone(start_page) != page_zone(end_page));
694 #endif
696 for (page = start_page; page <= end_page;) {
697 if (!pfn_valid_within(page_to_pfn(page))) {
698 page++;
699 continue;
702 if (!PageBuddy(page)) {
703 page++;
704 continue;
707 order = page_order(page);
708 list_del(&page->lru);
709 list_add(&page->lru,
710 &zone->free_area[order].free_list[migratetype]);
711 page += 1 << order;
712 pages_moved += 1 << order;
715 return pages_moved;
718 static int move_freepages_block(struct zone *zone, struct page *page,
719 int migratetype)
721 unsigned long start_pfn, end_pfn;
722 struct page *start_page, *end_page;
724 start_pfn = page_to_pfn(page);
725 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
726 start_page = pfn_to_page(start_pfn);
727 end_page = start_page + pageblock_nr_pages - 1;
728 end_pfn = start_pfn + pageblock_nr_pages - 1;
730 /* Do not cross zone boundaries */
731 if (start_pfn < zone->zone_start_pfn)
732 start_page = page;
733 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
734 return 0;
736 return move_freepages(zone, start_page, end_page, migratetype);
739 /* Remove an element from the buddy allocator from the fallback list */
740 static struct page *__rmqueue_fallback(struct zone *zone, int order,
741 int start_migratetype)
743 struct free_area * area;
744 int current_order;
745 struct page *page;
746 int migratetype, i;
748 /* Find the largest possible block of pages in the other list */
749 for (current_order = MAX_ORDER-1; current_order >= order;
750 --current_order) {
751 for (i = 0; i < MIGRATE_TYPES - 1; i++) {
752 migratetype = fallbacks[start_migratetype][i];
754 /* MIGRATE_RESERVE handled later if necessary */
755 if (migratetype == MIGRATE_RESERVE)
756 continue;
758 area = &(zone->free_area[current_order]);
759 if (list_empty(&area->free_list[migratetype]))
760 continue;
762 page = list_entry(area->free_list[migratetype].next,
763 struct page, lru);
764 area->nr_free--;
767 * If breaking a large block of pages, move all free
768 * pages to the preferred allocation list. If falling
769 * back for a reclaimable kernel allocation, be more
770 * agressive about taking ownership of free pages
772 if (unlikely(current_order >= (pageblock_order >> 1)) ||
773 start_migratetype == MIGRATE_RECLAIMABLE) {
774 unsigned long pages;
775 pages = move_freepages_block(zone, page,
776 start_migratetype);
778 /* Claim the whole block if over half of it is free */
779 if (pages >= (1 << (pageblock_order-1)))
780 set_pageblock_migratetype(page,
781 start_migratetype);
783 migratetype = start_migratetype;
786 /* Remove the page from the freelists */
787 list_del(&page->lru);
788 rmv_page_order(page);
789 __mod_zone_page_state(zone, NR_FREE_PAGES,
790 -(1UL << order));
792 if (current_order == pageblock_order)
793 set_pageblock_migratetype(page,
794 start_migratetype);
796 expand(zone, page, order, current_order, area, migratetype);
797 return page;
801 /* Use MIGRATE_RESERVE rather than fail an allocation */
802 return __rmqueue_smallest(zone, order, MIGRATE_RESERVE);
806 * Do the hard work of removing an element from the buddy allocator.
807 * Call me with the zone->lock already held.
809 static struct page *__rmqueue(struct zone *zone, unsigned int order,
810 int migratetype)
812 struct page *page;
814 page = __rmqueue_smallest(zone, order, migratetype);
816 if (unlikely(!page))
817 page = __rmqueue_fallback(zone, order, migratetype);
819 return page;
823 * Obtain a specified number of elements from the buddy allocator, all under
824 * a single hold of the lock, for efficiency. Add them to the supplied list.
825 * Returns the number of new pages which were placed at *list.
827 static int rmqueue_bulk(struct zone *zone, unsigned int order,
828 unsigned long count, struct list_head *list,
829 int migratetype)
831 int i;
833 spin_lock(&zone->lock);
834 for (i = 0; i < count; ++i) {
835 struct page *page = __rmqueue(zone, order, migratetype);
836 if (unlikely(page == NULL))
837 break;
840 * Split buddy pages returned by expand() are received here
841 * in physical page order. The page is added to the callers and
842 * list and the list head then moves forward. From the callers
843 * perspective, the linked list is ordered by page number in
844 * some conditions. This is useful for IO devices that can
845 * merge IO requests if the physical pages are ordered
846 * properly.
848 list_add(&page->lru, list);
849 set_page_private(page, migratetype);
850 list = &page->lru;
852 spin_unlock(&zone->lock);
853 return i;
856 #ifdef CONFIG_NUMA
858 * Called from the vmstat counter updater to drain pagesets of this
859 * currently executing processor on remote nodes after they have
860 * expired.
862 * Note that this function must be called with the thread pinned to
863 * a single processor.
865 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
867 unsigned long flags;
868 int to_drain;
870 local_irq_save(flags);
871 if (pcp->count >= pcp->batch)
872 to_drain = pcp->batch;
873 else
874 to_drain = pcp->count;
875 free_pages_bulk(zone, to_drain, &pcp->list, 0);
876 pcp->count -= to_drain;
877 local_irq_restore(flags);
879 #endif
882 * Drain pages of the indicated processor.
884 * The processor must either be the current processor and the
885 * thread pinned to the current processor or a processor that
886 * is not online.
888 static void drain_pages(unsigned int cpu)
890 unsigned long flags;
891 struct zone *zone;
893 for_each_zone(zone) {
894 struct per_cpu_pageset *pset;
895 struct per_cpu_pages *pcp;
897 if (!populated_zone(zone))
898 continue;
900 pset = zone_pcp(zone, cpu);
902 pcp = &pset->pcp;
903 local_irq_save(flags);
904 free_pages_bulk(zone, pcp->count, &pcp->list, 0);
905 pcp->count = 0;
906 local_irq_restore(flags);
911 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
913 void drain_local_pages(void *arg)
915 drain_pages(smp_processor_id());
919 * Spill all the per-cpu pages from all CPUs back into the buddy allocator
921 void drain_all_pages(void)
923 on_each_cpu(drain_local_pages, NULL, 1);
926 #ifdef CONFIG_HIBERNATION
928 void mark_free_pages(struct zone *zone)
930 unsigned long pfn, max_zone_pfn;
931 unsigned long flags;
932 int order, t;
933 struct list_head *curr;
935 if (!zone->spanned_pages)
936 return;
938 spin_lock_irqsave(&zone->lock, flags);
940 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
941 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
942 if (pfn_valid(pfn)) {
943 struct page *page = pfn_to_page(pfn);
945 if (!swsusp_page_is_forbidden(page))
946 swsusp_unset_page_free(page);
949 for_each_migratetype_order(order, t) {
950 list_for_each(curr, &zone->free_area[order].free_list[t]) {
951 unsigned long i;
953 pfn = page_to_pfn(list_entry(curr, struct page, lru));
954 for (i = 0; i < (1UL << order); i++)
955 swsusp_set_page_free(pfn_to_page(pfn + i));
958 spin_unlock_irqrestore(&zone->lock, flags);
960 #endif /* CONFIG_PM */
963 * Free a 0-order page
965 static void free_hot_cold_page(struct page *page, int cold)
967 struct zone *zone = page_zone(page);
968 struct per_cpu_pages *pcp;
969 unsigned long flags;
971 if (PageAnon(page))
972 page->mapping = NULL;
973 if (free_pages_check(page))
974 return;
976 if (!PageHighMem(page)) {
977 debug_check_no_locks_freed(page_address(page), PAGE_SIZE);
978 debug_check_no_obj_freed(page_address(page), PAGE_SIZE);
980 arch_free_page(page, 0);
981 kernel_map_pages(page, 1, 0);
983 pcp = &zone_pcp(zone, get_cpu())->pcp;
984 local_irq_save(flags);
985 __count_vm_event(PGFREE);
986 if (cold)
987 list_add_tail(&page->lru, &pcp->list);
988 else
989 list_add(&page->lru, &pcp->list);
990 set_page_private(page, get_pageblock_migratetype(page));
991 pcp->count++;
992 if (pcp->count >= pcp->high) {
993 free_pages_bulk(zone, pcp->batch, &pcp->list, 0);
994 pcp->count -= pcp->batch;
996 local_irq_restore(flags);
997 put_cpu();
1000 void free_hot_page(struct page *page)
1002 free_hot_cold_page(page, 0);
1005 void free_cold_page(struct page *page)
1007 free_hot_cold_page(page, 1);
1011 * split_page takes a non-compound higher-order page, and splits it into
1012 * n (1<<order) sub-pages: page[0..n]
1013 * Each sub-page must be freed individually.
1015 * Note: this is probably too low level an operation for use in drivers.
1016 * Please consult with lkml before using this in your driver.
1018 void split_page(struct page *page, unsigned int order)
1020 int i;
1022 VM_BUG_ON(PageCompound(page));
1023 VM_BUG_ON(!page_count(page));
1024 for (i = 1; i < (1 << order); i++)
1025 set_page_refcounted(page + i);
1029 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1030 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1031 * or two.
1033 static struct page *buffered_rmqueue(struct zone *preferred_zone,
1034 struct zone *zone, int order, gfp_t gfp_flags)
1036 unsigned long flags;
1037 struct page *page;
1038 int cold = !!(gfp_flags & __GFP_COLD);
1039 int cpu;
1040 int migratetype = allocflags_to_migratetype(gfp_flags);
1042 again:
1043 cpu = get_cpu();
1044 if (likely(order == 0)) {
1045 struct per_cpu_pages *pcp;
1047 pcp = &zone_pcp(zone, cpu)->pcp;
1048 local_irq_save(flags);
1049 if (!pcp->count) {
1050 pcp->count = rmqueue_bulk(zone, 0,
1051 pcp->batch, &pcp->list, migratetype);
1052 if (unlikely(!pcp->count))
1053 goto failed;
1056 /* Find a page of the appropriate migrate type */
1057 if (cold) {
1058 list_for_each_entry_reverse(page, &pcp->list, lru)
1059 if (page_private(page) == migratetype)
1060 break;
1061 } else {
1062 list_for_each_entry(page, &pcp->list, lru)
1063 if (page_private(page) == migratetype)
1064 break;
1067 /* Allocate more to the pcp list if necessary */
1068 if (unlikely(&page->lru == &pcp->list)) {
1069 pcp->count += rmqueue_bulk(zone, 0,
1070 pcp->batch, &pcp->list, migratetype);
1071 page = list_entry(pcp->list.next, struct page, lru);
1074 list_del(&page->lru);
1075 pcp->count--;
1076 } else {
1077 spin_lock_irqsave(&zone->lock, flags);
1078 page = __rmqueue(zone, order, migratetype);
1079 spin_unlock(&zone->lock);
1080 if (!page)
1081 goto failed;
1084 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1085 zone_statistics(preferred_zone, zone);
1086 local_irq_restore(flags);
1087 put_cpu();
1089 VM_BUG_ON(bad_range(zone, page));
1090 if (prep_new_page(page, order, gfp_flags))
1091 goto again;
1092 return page;
1094 failed:
1095 local_irq_restore(flags);
1096 put_cpu();
1097 return NULL;
1100 #define ALLOC_NO_WATERMARKS 0x01 /* don't check watermarks at all */
1101 #define ALLOC_WMARK_MIN 0x02 /* use pages_min watermark */
1102 #define ALLOC_WMARK_LOW 0x04 /* use pages_low watermark */
1103 #define ALLOC_WMARK_HIGH 0x08 /* use pages_high watermark */
1104 #define ALLOC_HARDER 0x10 /* try to alloc harder */
1105 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
1106 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
1108 #ifdef CONFIG_FAIL_PAGE_ALLOC
1110 static struct fail_page_alloc_attr {
1111 struct fault_attr attr;
1113 u32 ignore_gfp_highmem;
1114 u32 ignore_gfp_wait;
1115 u32 min_order;
1117 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1119 struct dentry *ignore_gfp_highmem_file;
1120 struct dentry *ignore_gfp_wait_file;
1121 struct dentry *min_order_file;
1123 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1125 } fail_page_alloc = {
1126 .attr = FAULT_ATTR_INITIALIZER,
1127 .ignore_gfp_wait = 1,
1128 .ignore_gfp_highmem = 1,
1129 .min_order = 1,
1132 static int __init setup_fail_page_alloc(char *str)
1134 return setup_fault_attr(&fail_page_alloc.attr, str);
1136 __setup("fail_page_alloc=", setup_fail_page_alloc);
1138 static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1140 if (order < fail_page_alloc.min_order)
1141 return 0;
1142 if (gfp_mask & __GFP_NOFAIL)
1143 return 0;
1144 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1145 return 0;
1146 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1147 return 0;
1149 return should_fail(&fail_page_alloc.attr, 1 << order);
1152 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1154 static int __init fail_page_alloc_debugfs(void)
1156 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1157 struct dentry *dir;
1158 int err;
1160 err = init_fault_attr_dentries(&fail_page_alloc.attr,
1161 "fail_page_alloc");
1162 if (err)
1163 return err;
1164 dir = fail_page_alloc.attr.dentries.dir;
1166 fail_page_alloc.ignore_gfp_wait_file =
1167 debugfs_create_bool("ignore-gfp-wait", mode, dir,
1168 &fail_page_alloc.ignore_gfp_wait);
1170 fail_page_alloc.ignore_gfp_highmem_file =
1171 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1172 &fail_page_alloc.ignore_gfp_highmem);
1173 fail_page_alloc.min_order_file =
1174 debugfs_create_u32("min-order", mode, dir,
1175 &fail_page_alloc.min_order);
1177 if (!fail_page_alloc.ignore_gfp_wait_file ||
1178 !fail_page_alloc.ignore_gfp_highmem_file ||
1179 !fail_page_alloc.min_order_file) {
1180 err = -ENOMEM;
1181 debugfs_remove(fail_page_alloc.ignore_gfp_wait_file);
1182 debugfs_remove(fail_page_alloc.ignore_gfp_highmem_file);
1183 debugfs_remove(fail_page_alloc.min_order_file);
1184 cleanup_fault_attr_dentries(&fail_page_alloc.attr);
1187 return err;
1190 late_initcall(fail_page_alloc_debugfs);
1192 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1194 #else /* CONFIG_FAIL_PAGE_ALLOC */
1196 static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1198 return 0;
1201 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1204 * Return 1 if free pages are above 'mark'. This takes into account the order
1205 * of the allocation.
1207 int zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1208 int classzone_idx, int alloc_flags)
1210 /* free_pages my go negative - that's OK */
1211 long min = mark;
1212 long free_pages = zone_page_state(z, NR_FREE_PAGES) - (1 << order) + 1;
1213 int o;
1215 if (alloc_flags & ALLOC_HIGH)
1216 min -= min / 2;
1217 if (alloc_flags & ALLOC_HARDER)
1218 min -= min / 4;
1220 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
1221 return 0;
1222 for (o = 0; o < order; o++) {
1223 /* At the next order, this order's pages become unavailable */
1224 free_pages -= z->free_area[o].nr_free << o;
1226 /* Require fewer higher order pages to be free */
1227 min >>= 1;
1229 if (free_pages <= min)
1230 return 0;
1232 return 1;
1235 #ifdef CONFIG_NUMA
1237 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1238 * skip over zones that are not allowed by the cpuset, or that have
1239 * been recently (in last second) found to be nearly full. See further
1240 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1241 * that have to skip over a lot of full or unallowed zones.
1243 * If the zonelist cache is present in the passed in zonelist, then
1244 * returns a pointer to the allowed node mask (either the current
1245 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1247 * If the zonelist cache is not available for this zonelist, does
1248 * nothing and returns NULL.
1250 * If the fullzones BITMAP in the zonelist cache is stale (more than
1251 * a second since last zap'd) then we zap it out (clear its bits.)
1253 * We hold off even calling zlc_setup, until after we've checked the
1254 * first zone in the zonelist, on the theory that most allocations will
1255 * be satisfied from that first zone, so best to examine that zone as
1256 * quickly as we can.
1258 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1260 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1261 nodemask_t *allowednodes; /* zonelist_cache approximation */
1263 zlc = zonelist->zlcache_ptr;
1264 if (!zlc)
1265 return NULL;
1267 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1268 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1269 zlc->last_full_zap = jiffies;
1272 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1273 &cpuset_current_mems_allowed :
1274 &node_states[N_HIGH_MEMORY];
1275 return allowednodes;
1279 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1280 * if it is worth looking at further for free memory:
1281 * 1) Check that the zone isn't thought to be full (doesn't have its
1282 * bit set in the zonelist_cache fullzones BITMAP).
1283 * 2) Check that the zones node (obtained from the zonelist_cache
1284 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1285 * Return true (non-zero) if zone is worth looking at further, or
1286 * else return false (zero) if it is not.
1288 * This check -ignores- the distinction between various watermarks,
1289 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1290 * found to be full for any variation of these watermarks, it will
1291 * be considered full for up to one second by all requests, unless
1292 * we are so low on memory on all allowed nodes that we are forced
1293 * into the second scan of the zonelist.
1295 * In the second scan we ignore this zonelist cache and exactly
1296 * apply the watermarks to all zones, even it is slower to do so.
1297 * We are low on memory in the second scan, and should leave no stone
1298 * unturned looking for a free page.
1300 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1301 nodemask_t *allowednodes)
1303 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1304 int i; /* index of *z in zonelist zones */
1305 int n; /* node that zone *z is on */
1307 zlc = zonelist->zlcache_ptr;
1308 if (!zlc)
1309 return 1;
1311 i = z - zonelist->_zonerefs;
1312 n = zlc->z_to_n[i];
1314 /* This zone is worth trying if it is allowed but not full */
1315 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1319 * Given 'z' scanning a zonelist, set the corresponding bit in
1320 * zlc->fullzones, so that subsequent attempts to allocate a page
1321 * from that zone don't waste time re-examining it.
1323 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1325 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1326 int i; /* index of *z in zonelist zones */
1328 zlc = zonelist->zlcache_ptr;
1329 if (!zlc)
1330 return;
1332 i = z - zonelist->_zonerefs;
1334 set_bit(i, zlc->fullzones);
1337 #else /* CONFIG_NUMA */
1339 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1341 return NULL;
1344 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1345 nodemask_t *allowednodes)
1347 return 1;
1350 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1353 #endif /* CONFIG_NUMA */
1356 * get_page_from_freelist goes through the zonelist trying to allocate
1357 * a page.
1359 static struct page *
1360 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1361 struct zonelist *zonelist, int high_zoneidx, int alloc_flags)
1363 struct zoneref *z;
1364 struct page *page = NULL;
1365 int classzone_idx;
1366 struct zone *zone, *preferred_zone;
1367 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1368 int zlc_active = 0; /* set if using zonelist_cache */
1369 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1371 (void)first_zones_zonelist(zonelist, high_zoneidx, nodemask,
1372 &preferred_zone);
1373 if (!preferred_zone)
1374 return NULL;
1376 classzone_idx = zone_idx(preferred_zone);
1378 zonelist_scan:
1380 * Scan zonelist, looking for a zone with enough free.
1381 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1383 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1384 high_zoneidx, nodemask) {
1385 if (NUMA_BUILD && zlc_active &&
1386 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1387 continue;
1388 if ((alloc_flags & ALLOC_CPUSET) &&
1389 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1390 goto try_next_zone;
1392 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1393 unsigned long mark;
1394 if (alloc_flags & ALLOC_WMARK_MIN)
1395 mark = zone->pages_min;
1396 else if (alloc_flags & ALLOC_WMARK_LOW)
1397 mark = zone->pages_low;
1398 else
1399 mark = zone->pages_high;
1400 if (!zone_watermark_ok(zone, order, mark,
1401 classzone_idx, alloc_flags)) {
1402 if (!zone_reclaim_mode ||
1403 !zone_reclaim(zone, gfp_mask, order))
1404 goto this_zone_full;
1408 page = buffered_rmqueue(preferred_zone, zone, order, gfp_mask);
1409 if (page)
1410 break;
1411 this_zone_full:
1412 if (NUMA_BUILD)
1413 zlc_mark_zone_full(zonelist, z);
1414 try_next_zone:
1415 if (NUMA_BUILD && !did_zlc_setup) {
1416 /* we do zlc_setup after the first zone is tried */
1417 allowednodes = zlc_setup(zonelist, alloc_flags);
1418 zlc_active = 1;
1419 did_zlc_setup = 1;
1423 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) {
1424 /* Disable zlc cache for second zonelist scan */
1425 zlc_active = 0;
1426 goto zonelist_scan;
1428 return page;
1432 * This is the 'heart' of the zoned buddy allocator.
1434 struct page *
1435 __alloc_pages_internal(gfp_t gfp_mask, unsigned int order,
1436 struct zonelist *zonelist, nodemask_t *nodemask)
1438 const gfp_t wait = gfp_mask & __GFP_WAIT;
1439 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
1440 struct zoneref *z;
1441 struct zone *zone;
1442 struct page *page;
1443 struct reclaim_state reclaim_state;
1444 struct task_struct *p = current;
1445 int do_retry;
1446 int alloc_flags;
1447 unsigned long did_some_progress;
1448 unsigned long pages_reclaimed = 0;
1450 might_sleep_if(wait);
1452 if (should_fail_alloc_page(gfp_mask, order))
1453 return NULL;
1455 restart:
1456 z = zonelist->_zonerefs; /* the list of zones suitable for gfp_mask */
1458 if (unlikely(!z->zone)) {
1460 * Happens if we have an empty zonelist as a result of
1461 * GFP_THISNODE being used on a memoryless node
1463 return NULL;
1466 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
1467 zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET);
1468 if (page)
1469 goto got_pg;
1472 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
1473 * __GFP_NOWARN set) should not cause reclaim since the subsystem
1474 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
1475 * using a larger set of nodes after it has established that the
1476 * allowed per node queues are empty and that nodes are
1477 * over allocated.
1479 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
1480 goto nopage;
1482 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
1483 wakeup_kswapd(zone, order);
1486 * OK, we're below the kswapd watermark and have kicked background
1487 * reclaim. Now things get more complex, so set up alloc_flags according
1488 * to how we want to proceed.
1490 * The caller may dip into page reserves a bit more if the caller
1491 * cannot run direct reclaim, or if the caller has realtime scheduling
1492 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
1493 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
1495 alloc_flags = ALLOC_WMARK_MIN;
1496 if ((unlikely(rt_task(p)) && !in_interrupt()) || !wait)
1497 alloc_flags |= ALLOC_HARDER;
1498 if (gfp_mask & __GFP_HIGH)
1499 alloc_flags |= ALLOC_HIGH;
1500 if (wait)
1501 alloc_flags |= ALLOC_CPUSET;
1504 * Go through the zonelist again. Let __GFP_HIGH and allocations
1505 * coming from realtime tasks go deeper into reserves.
1507 * This is the last chance, in general, before the goto nopage.
1508 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
1509 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1511 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
1512 high_zoneidx, alloc_flags);
1513 if (page)
1514 goto got_pg;
1516 /* This allocation should allow future memory freeing. */
1518 rebalance:
1519 if (((p->flags & PF_MEMALLOC) || unlikely(test_thread_flag(TIF_MEMDIE)))
1520 && !in_interrupt()) {
1521 if (!(gfp_mask & __GFP_NOMEMALLOC)) {
1522 nofail_alloc:
1523 /* go through the zonelist yet again, ignoring mins */
1524 page = get_page_from_freelist(gfp_mask, nodemask, order,
1525 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS);
1526 if (page)
1527 goto got_pg;
1528 if (gfp_mask & __GFP_NOFAIL) {
1529 congestion_wait(WRITE, HZ/50);
1530 goto nofail_alloc;
1533 goto nopage;
1536 /* Atomic allocations - we can't balance anything */
1537 if (!wait)
1538 goto nopage;
1540 cond_resched();
1542 /* We now go into synchronous reclaim */
1543 cpuset_memory_pressure_bump();
1544 p->flags |= PF_MEMALLOC;
1545 reclaim_state.reclaimed_slab = 0;
1546 p->reclaim_state = &reclaim_state;
1548 did_some_progress = try_to_free_pages(zonelist, order, gfp_mask);
1550 p->reclaim_state = NULL;
1551 p->flags &= ~PF_MEMALLOC;
1553 cond_resched();
1555 if (order != 0)
1556 drain_all_pages();
1558 if (likely(did_some_progress)) {
1559 page = get_page_from_freelist(gfp_mask, nodemask, order,
1560 zonelist, high_zoneidx, alloc_flags);
1561 if (page)
1562 goto got_pg;
1563 } else if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
1564 if (!try_set_zone_oom(zonelist, gfp_mask)) {
1565 schedule_timeout_uninterruptible(1);
1566 goto restart;
1570 * Go through the zonelist yet one more time, keep
1571 * very high watermark here, this is only to catch
1572 * a parallel oom killing, we must fail if we're still
1573 * under heavy pressure.
1575 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
1576 order, zonelist, high_zoneidx,
1577 ALLOC_WMARK_HIGH|ALLOC_CPUSET);
1578 if (page) {
1579 clear_zonelist_oom(zonelist, gfp_mask);
1580 goto got_pg;
1583 /* The OOM killer will not help higher order allocs so fail */
1584 if (order > PAGE_ALLOC_COSTLY_ORDER) {
1585 clear_zonelist_oom(zonelist, gfp_mask);
1586 goto nopage;
1589 out_of_memory(zonelist, gfp_mask, order);
1590 clear_zonelist_oom(zonelist, gfp_mask);
1591 goto restart;
1595 * Don't let big-order allocations loop unless the caller explicitly
1596 * requests that. Wait for some write requests to complete then retry.
1598 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
1599 * means __GFP_NOFAIL, but that may not be true in other
1600 * implementations.
1602 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
1603 * specified, then we retry until we no longer reclaim any pages
1604 * (above), or we've reclaimed an order of pages at least as
1605 * large as the allocation's order. In both cases, if the
1606 * allocation still fails, we stop retrying.
1608 pages_reclaimed += did_some_progress;
1609 do_retry = 0;
1610 if (!(gfp_mask & __GFP_NORETRY)) {
1611 if (order <= PAGE_ALLOC_COSTLY_ORDER) {
1612 do_retry = 1;
1613 } else {
1614 if (gfp_mask & __GFP_REPEAT &&
1615 pages_reclaimed < (1 << order))
1616 do_retry = 1;
1618 if (gfp_mask & __GFP_NOFAIL)
1619 do_retry = 1;
1621 if (do_retry) {
1622 congestion_wait(WRITE, HZ/50);
1623 goto rebalance;
1626 nopage:
1627 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
1628 printk(KERN_WARNING "%s: page allocation failure."
1629 " order:%d, mode:0x%x\n",
1630 p->comm, order, gfp_mask);
1631 dump_stack();
1632 show_mem();
1634 got_pg:
1635 return page;
1637 EXPORT_SYMBOL(__alloc_pages_internal);
1640 * Common helper functions.
1642 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
1644 struct page * page;
1645 page = alloc_pages(gfp_mask, order);
1646 if (!page)
1647 return 0;
1648 return (unsigned long) page_address(page);
1651 EXPORT_SYMBOL(__get_free_pages);
1653 unsigned long get_zeroed_page(gfp_t gfp_mask)
1655 struct page * page;
1658 * get_zeroed_page() returns a 32-bit address, which cannot represent
1659 * a highmem page
1661 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
1663 page = alloc_pages(gfp_mask | __GFP_ZERO, 0);
1664 if (page)
1665 return (unsigned long) page_address(page);
1666 return 0;
1669 EXPORT_SYMBOL(get_zeroed_page);
1671 void __pagevec_free(struct pagevec *pvec)
1673 int i = pagevec_count(pvec);
1675 while (--i >= 0)
1676 free_hot_cold_page(pvec->pages[i], pvec->cold);
1679 void __free_pages(struct page *page, unsigned int order)
1681 if (put_page_testzero(page)) {
1682 if (order == 0)
1683 free_hot_page(page);
1684 else
1685 __free_pages_ok(page, order);
1689 EXPORT_SYMBOL(__free_pages);
1691 void free_pages(unsigned long addr, unsigned int order)
1693 if (addr != 0) {
1694 VM_BUG_ON(!virt_addr_valid((void *)addr));
1695 __free_pages(virt_to_page((void *)addr), order);
1699 EXPORT_SYMBOL(free_pages);
1702 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
1703 * @size: the number of bytes to allocate
1704 * @gfp_mask: GFP flags for the allocation
1706 * This function is similar to alloc_pages(), except that it allocates the
1707 * minimum number of pages to satisfy the request. alloc_pages() can only
1708 * allocate memory in power-of-two pages.
1710 * This function is also limited by MAX_ORDER.
1712 * Memory allocated by this function must be released by free_pages_exact().
1714 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
1716 unsigned int order = get_order(size);
1717 unsigned long addr;
1719 addr = __get_free_pages(gfp_mask, order);
1720 if (addr) {
1721 unsigned long alloc_end = addr + (PAGE_SIZE << order);
1722 unsigned long used = addr + PAGE_ALIGN(size);
1724 split_page(virt_to_page(addr), order);
1725 while (used < alloc_end) {
1726 free_page(used);
1727 used += PAGE_SIZE;
1731 return (void *)addr;
1733 EXPORT_SYMBOL(alloc_pages_exact);
1736 * free_pages_exact - release memory allocated via alloc_pages_exact()
1737 * @virt: the value returned by alloc_pages_exact.
1738 * @size: size of allocation, same value as passed to alloc_pages_exact().
1740 * Release the memory allocated by a previous call to alloc_pages_exact.
1742 void free_pages_exact(void *virt, size_t size)
1744 unsigned long addr = (unsigned long)virt;
1745 unsigned long end = addr + PAGE_ALIGN(size);
1747 while (addr < end) {
1748 free_page(addr);
1749 addr += PAGE_SIZE;
1752 EXPORT_SYMBOL(free_pages_exact);
1754 static unsigned int nr_free_zone_pages(int offset)
1756 struct zoneref *z;
1757 struct zone *zone;
1759 /* Just pick one node, since fallback list is circular */
1760 unsigned int sum = 0;
1762 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
1764 for_each_zone_zonelist(zone, z, zonelist, offset) {
1765 unsigned long size = zone->present_pages;
1766 unsigned long high = zone->pages_high;
1767 if (size > high)
1768 sum += size - high;
1771 return sum;
1775 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
1777 unsigned int nr_free_buffer_pages(void)
1779 return nr_free_zone_pages(gfp_zone(GFP_USER));
1781 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
1784 * Amount of free RAM allocatable within all zones
1786 unsigned int nr_free_pagecache_pages(void)
1788 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
1791 static inline void show_node(struct zone *zone)
1793 if (NUMA_BUILD)
1794 printk("Node %d ", zone_to_nid(zone));
1797 void si_meminfo(struct sysinfo *val)
1799 val->totalram = totalram_pages;
1800 val->sharedram = 0;
1801 val->freeram = global_page_state(NR_FREE_PAGES);
1802 val->bufferram = nr_blockdev_pages();
1803 val->totalhigh = totalhigh_pages;
1804 val->freehigh = nr_free_highpages();
1805 val->mem_unit = PAGE_SIZE;
1808 EXPORT_SYMBOL(si_meminfo);
1810 #ifdef CONFIG_NUMA
1811 void si_meminfo_node(struct sysinfo *val, int nid)
1813 pg_data_t *pgdat = NODE_DATA(nid);
1815 val->totalram = pgdat->node_present_pages;
1816 val->freeram = node_page_state(nid, NR_FREE_PAGES);
1817 #ifdef CONFIG_HIGHMEM
1818 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
1819 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
1820 NR_FREE_PAGES);
1821 #else
1822 val->totalhigh = 0;
1823 val->freehigh = 0;
1824 #endif
1825 val->mem_unit = PAGE_SIZE;
1827 #endif
1829 #define K(x) ((x) << (PAGE_SHIFT-10))
1832 * Show free area list (used inside shift_scroll-lock stuff)
1833 * We also calculate the percentage fragmentation. We do this by counting the
1834 * memory on each free list with the exception of the first item on the list.
1836 void show_free_areas(void)
1838 int cpu;
1839 struct zone *zone;
1841 for_each_zone(zone) {
1842 if (!populated_zone(zone))
1843 continue;
1845 show_node(zone);
1846 printk("%s per-cpu:\n", zone->name);
1848 for_each_online_cpu(cpu) {
1849 struct per_cpu_pageset *pageset;
1851 pageset = zone_pcp(zone, cpu);
1853 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
1854 cpu, pageset->pcp.high,
1855 pageset->pcp.batch, pageset->pcp.count);
1859 printk("Active:%lu inactive:%lu dirty:%lu writeback:%lu unstable:%lu\n"
1860 " free:%lu slab:%lu mapped:%lu pagetables:%lu bounce:%lu\n",
1861 global_page_state(NR_ACTIVE),
1862 global_page_state(NR_INACTIVE),
1863 global_page_state(NR_FILE_DIRTY),
1864 global_page_state(NR_WRITEBACK),
1865 global_page_state(NR_UNSTABLE_NFS),
1866 global_page_state(NR_FREE_PAGES),
1867 global_page_state(NR_SLAB_RECLAIMABLE) +
1868 global_page_state(NR_SLAB_UNRECLAIMABLE),
1869 global_page_state(NR_FILE_MAPPED),
1870 global_page_state(NR_PAGETABLE),
1871 global_page_state(NR_BOUNCE));
1873 for_each_zone(zone) {
1874 int i;
1876 if (!populated_zone(zone))
1877 continue;
1879 show_node(zone);
1880 printk("%s"
1881 " free:%lukB"
1882 " min:%lukB"
1883 " low:%lukB"
1884 " high:%lukB"
1885 " active:%lukB"
1886 " inactive:%lukB"
1887 " present:%lukB"
1888 " pages_scanned:%lu"
1889 " all_unreclaimable? %s"
1890 "\n",
1891 zone->name,
1892 K(zone_page_state(zone, NR_FREE_PAGES)),
1893 K(zone->pages_min),
1894 K(zone->pages_low),
1895 K(zone->pages_high),
1896 K(zone_page_state(zone, NR_ACTIVE)),
1897 K(zone_page_state(zone, NR_INACTIVE)),
1898 K(zone->present_pages),
1899 zone->pages_scanned,
1900 (zone_is_all_unreclaimable(zone) ? "yes" : "no")
1902 printk("lowmem_reserve[]:");
1903 for (i = 0; i < MAX_NR_ZONES; i++)
1904 printk(" %lu", zone->lowmem_reserve[i]);
1905 printk("\n");
1908 for_each_zone(zone) {
1909 unsigned long nr[MAX_ORDER], flags, order, total = 0;
1911 if (!populated_zone(zone))
1912 continue;
1914 show_node(zone);
1915 printk("%s: ", zone->name);
1917 spin_lock_irqsave(&zone->lock, flags);
1918 for (order = 0; order < MAX_ORDER; order++) {
1919 nr[order] = zone->free_area[order].nr_free;
1920 total += nr[order] << order;
1922 spin_unlock_irqrestore(&zone->lock, flags);
1923 for (order = 0; order < MAX_ORDER; order++)
1924 printk("%lu*%lukB ", nr[order], K(1UL) << order);
1925 printk("= %lukB\n", K(total));
1928 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
1930 show_swap_cache_info();
1933 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
1935 zoneref->zone = zone;
1936 zoneref->zone_idx = zone_idx(zone);
1940 * Builds allocation fallback zone lists.
1942 * Add all populated zones of a node to the zonelist.
1944 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
1945 int nr_zones, enum zone_type zone_type)
1947 struct zone *zone;
1949 BUG_ON(zone_type >= MAX_NR_ZONES);
1950 zone_type++;
1952 do {
1953 zone_type--;
1954 zone = pgdat->node_zones + zone_type;
1955 if (populated_zone(zone)) {
1956 zoneref_set_zone(zone,
1957 &zonelist->_zonerefs[nr_zones++]);
1958 check_highest_zone(zone_type);
1961 } while (zone_type);
1962 return nr_zones;
1967 * zonelist_order:
1968 * 0 = automatic detection of better ordering.
1969 * 1 = order by ([node] distance, -zonetype)
1970 * 2 = order by (-zonetype, [node] distance)
1972 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
1973 * the same zonelist. So only NUMA can configure this param.
1975 #define ZONELIST_ORDER_DEFAULT 0
1976 #define ZONELIST_ORDER_NODE 1
1977 #define ZONELIST_ORDER_ZONE 2
1979 /* zonelist order in the kernel.
1980 * set_zonelist_order() will set this to NODE or ZONE.
1982 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
1983 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
1986 #ifdef CONFIG_NUMA
1987 /* The value user specified ....changed by config */
1988 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
1989 /* string for sysctl */
1990 #define NUMA_ZONELIST_ORDER_LEN 16
1991 char numa_zonelist_order[16] = "default";
1994 * interface for configure zonelist ordering.
1995 * command line option "numa_zonelist_order"
1996 * = "[dD]efault - default, automatic configuration.
1997 * = "[nN]ode - order by node locality, then by zone within node
1998 * = "[zZ]one - order by zone, then by locality within zone
2001 static int __parse_numa_zonelist_order(char *s)
2003 if (*s == 'd' || *s == 'D') {
2004 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2005 } else if (*s == 'n' || *s == 'N') {
2006 user_zonelist_order = ZONELIST_ORDER_NODE;
2007 } else if (*s == 'z' || *s == 'Z') {
2008 user_zonelist_order = ZONELIST_ORDER_ZONE;
2009 } else {
2010 printk(KERN_WARNING
2011 "Ignoring invalid numa_zonelist_order value: "
2012 "%s\n", s);
2013 return -EINVAL;
2015 return 0;
2018 static __init int setup_numa_zonelist_order(char *s)
2020 if (s)
2021 return __parse_numa_zonelist_order(s);
2022 return 0;
2024 early_param("numa_zonelist_order", setup_numa_zonelist_order);
2027 * sysctl handler for numa_zonelist_order
2029 int numa_zonelist_order_handler(ctl_table *table, int write,
2030 struct file *file, void __user *buffer, size_t *length,
2031 loff_t *ppos)
2033 char saved_string[NUMA_ZONELIST_ORDER_LEN];
2034 int ret;
2036 if (write)
2037 strncpy(saved_string, (char*)table->data,
2038 NUMA_ZONELIST_ORDER_LEN);
2039 ret = proc_dostring(table, write, file, buffer, length, ppos);
2040 if (ret)
2041 return ret;
2042 if (write) {
2043 int oldval = user_zonelist_order;
2044 if (__parse_numa_zonelist_order((char*)table->data)) {
2046 * bogus value. restore saved string
2048 strncpy((char*)table->data, saved_string,
2049 NUMA_ZONELIST_ORDER_LEN);
2050 user_zonelist_order = oldval;
2051 } else if (oldval != user_zonelist_order)
2052 build_all_zonelists();
2054 return 0;
2058 #define MAX_NODE_LOAD (num_online_nodes())
2059 static int node_load[MAX_NUMNODES];
2062 * find_next_best_node - find the next node that should appear in a given node's fallback list
2063 * @node: node whose fallback list we're appending
2064 * @used_node_mask: nodemask_t of already used nodes
2066 * We use a number of factors to determine which is the next node that should
2067 * appear on a given node's fallback list. The node should not have appeared
2068 * already in @node's fallback list, and it should be the next closest node
2069 * according to the distance array (which contains arbitrary distance values
2070 * from each node to each node in the system), and should also prefer nodes
2071 * with no CPUs, since presumably they'll have very little allocation pressure
2072 * on them otherwise.
2073 * It returns -1 if no node is found.
2075 static int find_next_best_node(int node, nodemask_t *used_node_mask)
2077 int n, val;
2078 int min_val = INT_MAX;
2079 int best_node = -1;
2080 node_to_cpumask_ptr(tmp, 0);
2082 /* Use the local node if we haven't already */
2083 if (!node_isset(node, *used_node_mask)) {
2084 node_set(node, *used_node_mask);
2085 return node;
2088 for_each_node_state(n, N_HIGH_MEMORY) {
2090 /* Don't want a node to appear more than once */
2091 if (node_isset(n, *used_node_mask))
2092 continue;
2094 /* Use the distance array to find the distance */
2095 val = node_distance(node, n);
2097 /* Penalize nodes under us ("prefer the next node") */
2098 val += (n < node);
2100 /* Give preference to headless and unused nodes */
2101 node_to_cpumask_ptr_next(tmp, n);
2102 if (!cpus_empty(*tmp))
2103 val += PENALTY_FOR_NODE_WITH_CPUS;
2105 /* Slight preference for less loaded node */
2106 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
2107 val += node_load[n];
2109 if (val < min_val) {
2110 min_val = val;
2111 best_node = n;
2115 if (best_node >= 0)
2116 node_set(best_node, *used_node_mask);
2118 return best_node;
2123 * Build zonelists ordered by node and zones within node.
2124 * This results in maximum locality--normal zone overflows into local
2125 * DMA zone, if any--but risks exhausting DMA zone.
2127 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
2129 int j;
2130 struct zonelist *zonelist;
2132 zonelist = &pgdat->node_zonelists[0];
2133 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
2135 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2136 MAX_NR_ZONES - 1);
2137 zonelist->_zonerefs[j].zone = NULL;
2138 zonelist->_zonerefs[j].zone_idx = 0;
2142 * Build gfp_thisnode zonelists
2144 static void build_thisnode_zonelists(pg_data_t *pgdat)
2146 int j;
2147 struct zonelist *zonelist;
2149 zonelist = &pgdat->node_zonelists[1];
2150 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2151 zonelist->_zonerefs[j].zone = NULL;
2152 zonelist->_zonerefs[j].zone_idx = 0;
2156 * Build zonelists ordered by zone and nodes within zones.
2157 * This results in conserving DMA zone[s] until all Normal memory is
2158 * exhausted, but results in overflowing to remote node while memory
2159 * may still exist in local DMA zone.
2161 static int node_order[MAX_NUMNODES];
2163 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
2165 int pos, j, node;
2166 int zone_type; /* needs to be signed */
2167 struct zone *z;
2168 struct zonelist *zonelist;
2170 zonelist = &pgdat->node_zonelists[0];
2171 pos = 0;
2172 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
2173 for (j = 0; j < nr_nodes; j++) {
2174 node = node_order[j];
2175 z = &NODE_DATA(node)->node_zones[zone_type];
2176 if (populated_zone(z)) {
2177 zoneref_set_zone(z,
2178 &zonelist->_zonerefs[pos++]);
2179 check_highest_zone(zone_type);
2183 zonelist->_zonerefs[pos].zone = NULL;
2184 zonelist->_zonerefs[pos].zone_idx = 0;
2187 static int default_zonelist_order(void)
2189 int nid, zone_type;
2190 unsigned long low_kmem_size,total_size;
2191 struct zone *z;
2192 int average_size;
2194 * ZONE_DMA and ZONE_DMA32 can be very small area in the sytem.
2195 * If they are really small and used heavily, the system can fall
2196 * into OOM very easily.
2197 * This function detect ZONE_DMA/DMA32 size and confgigures zone order.
2199 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
2200 low_kmem_size = 0;
2201 total_size = 0;
2202 for_each_online_node(nid) {
2203 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2204 z = &NODE_DATA(nid)->node_zones[zone_type];
2205 if (populated_zone(z)) {
2206 if (zone_type < ZONE_NORMAL)
2207 low_kmem_size += z->present_pages;
2208 total_size += z->present_pages;
2212 if (!low_kmem_size || /* there are no DMA area. */
2213 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
2214 return ZONELIST_ORDER_NODE;
2216 * look into each node's config.
2217 * If there is a node whose DMA/DMA32 memory is very big area on
2218 * local memory, NODE_ORDER may be suitable.
2220 average_size = total_size /
2221 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1);
2222 for_each_online_node(nid) {
2223 low_kmem_size = 0;
2224 total_size = 0;
2225 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2226 z = &NODE_DATA(nid)->node_zones[zone_type];
2227 if (populated_zone(z)) {
2228 if (zone_type < ZONE_NORMAL)
2229 low_kmem_size += z->present_pages;
2230 total_size += z->present_pages;
2233 if (low_kmem_size &&
2234 total_size > average_size && /* ignore small node */
2235 low_kmem_size > total_size * 70/100)
2236 return ZONELIST_ORDER_NODE;
2238 return ZONELIST_ORDER_ZONE;
2241 static void set_zonelist_order(void)
2243 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
2244 current_zonelist_order = default_zonelist_order();
2245 else
2246 current_zonelist_order = user_zonelist_order;
2249 static void build_zonelists(pg_data_t *pgdat)
2251 int j, node, load;
2252 enum zone_type i;
2253 nodemask_t used_mask;
2254 int local_node, prev_node;
2255 struct zonelist *zonelist;
2256 int order = current_zonelist_order;
2258 /* initialize zonelists */
2259 for (i = 0; i < MAX_ZONELISTS; i++) {
2260 zonelist = pgdat->node_zonelists + i;
2261 zonelist->_zonerefs[0].zone = NULL;
2262 zonelist->_zonerefs[0].zone_idx = 0;
2265 /* NUMA-aware ordering of nodes */
2266 local_node = pgdat->node_id;
2267 load = num_online_nodes();
2268 prev_node = local_node;
2269 nodes_clear(used_mask);
2271 memset(node_load, 0, sizeof(node_load));
2272 memset(node_order, 0, sizeof(node_order));
2273 j = 0;
2275 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
2276 int distance = node_distance(local_node, node);
2279 * If another node is sufficiently far away then it is better
2280 * to reclaim pages in a zone before going off node.
2282 if (distance > RECLAIM_DISTANCE)
2283 zone_reclaim_mode = 1;
2286 * We don't want to pressure a particular node.
2287 * So adding penalty to the first node in same
2288 * distance group to make it round-robin.
2290 if (distance != node_distance(local_node, prev_node))
2291 node_load[node] = load;
2293 prev_node = node;
2294 load--;
2295 if (order == ZONELIST_ORDER_NODE)
2296 build_zonelists_in_node_order(pgdat, node);
2297 else
2298 node_order[j++] = node; /* remember order */
2301 if (order == ZONELIST_ORDER_ZONE) {
2302 /* calculate node order -- i.e., DMA last! */
2303 build_zonelists_in_zone_order(pgdat, j);
2306 build_thisnode_zonelists(pgdat);
2309 /* Construct the zonelist performance cache - see further mmzone.h */
2310 static void build_zonelist_cache(pg_data_t *pgdat)
2312 struct zonelist *zonelist;
2313 struct zonelist_cache *zlc;
2314 struct zoneref *z;
2316 zonelist = &pgdat->node_zonelists[0];
2317 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
2318 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
2319 for (z = zonelist->_zonerefs; z->zone; z++)
2320 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
2324 #else /* CONFIG_NUMA */
2326 static void set_zonelist_order(void)
2328 current_zonelist_order = ZONELIST_ORDER_ZONE;
2331 static void build_zonelists(pg_data_t *pgdat)
2333 int node, local_node;
2334 enum zone_type j;
2335 struct zonelist *zonelist;
2337 local_node = pgdat->node_id;
2339 zonelist = &pgdat->node_zonelists[0];
2340 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2343 * Now we build the zonelist so that it contains the zones
2344 * of all the other nodes.
2345 * We don't want to pressure a particular node, so when
2346 * building the zones for node N, we make sure that the
2347 * zones coming right after the local ones are those from
2348 * node N+1 (modulo N)
2350 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
2351 if (!node_online(node))
2352 continue;
2353 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2354 MAX_NR_ZONES - 1);
2356 for (node = 0; node < local_node; node++) {
2357 if (!node_online(node))
2358 continue;
2359 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2360 MAX_NR_ZONES - 1);
2363 zonelist->_zonerefs[j].zone = NULL;
2364 zonelist->_zonerefs[j].zone_idx = 0;
2367 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
2368 static void build_zonelist_cache(pg_data_t *pgdat)
2370 pgdat->node_zonelists[0].zlcache_ptr = NULL;
2373 #endif /* CONFIG_NUMA */
2375 /* return values int ....just for stop_machine() */
2376 static int __build_all_zonelists(void *dummy)
2378 int nid;
2380 for_each_online_node(nid) {
2381 pg_data_t *pgdat = NODE_DATA(nid);
2383 build_zonelists(pgdat);
2384 build_zonelist_cache(pgdat);
2386 return 0;
2389 void build_all_zonelists(void)
2391 set_zonelist_order();
2393 if (system_state == SYSTEM_BOOTING) {
2394 __build_all_zonelists(NULL);
2395 mminit_verify_zonelist();
2396 cpuset_init_current_mems_allowed();
2397 } else {
2398 /* we have to stop all cpus to guarantee there is no user
2399 of zonelist */
2400 stop_machine(__build_all_zonelists, NULL, NULL);
2401 /* cpuset refresh routine should be here */
2403 vm_total_pages = nr_free_pagecache_pages();
2405 * Disable grouping by mobility if the number of pages in the
2406 * system is too low to allow the mechanism to work. It would be
2407 * more accurate, but expensive to check per-zone. This check is
2408 * made on memory-hotadd so a system can start with mobility
2409 * disabled and enable it later
2411 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
2412 page_group_by_mobility_disabled = 1;
2413 else
2414 page_group_by_mobility_disabled = 0;
2416 printk("Built %i zonelists in %s order, mobility grouping %s. "
2417 "Total pages: %ld\n",
2418 num_online_nodes(),
2419 zonelist_order_name[current_zonelist_order],
2420 page_group_by_mobility_disabled ? "off" : "on",
2421 vm_total_pages);
2422 #ifdef CONFIG_NUMA
2423 printk("Policy zone: %s\n", zone_names[policy_zone]);
2424 #endif
2428 * Helper functions to size the waitqueue hash table.
2429 * Essentially these want to choose hash table sizes sufficiently
2430 * large so that collisions trying to wait on pages are rare.
2431 * But in fact, the number of active page waitqueues on typical
2432 * systems is ridiculously low, less than 200. So this is even
2433 * conservative, even though it seems large.
2435 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
2436 * waitqueues, i.e. the size of the waitq table given the number of pages.
2438 #define PAGES_PER_WAITQUEUE 256
2440 #ifndef CONFIG_MEMORY_HOTPLUG
2441 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
2443 unsigned long size = 1;
2445 pages /= PAGES_PER_WAITQUEUE;
2447 while (size < pages)
2448 size <<= 1;
2451 * Once we have dozens or even hundreds of threads sleeping
2452 * on IO we've got bigger problems than wait queue collision.
2453 * Limit the size of the wait table to a reasonable size.
2455 size = min(size, 4096UL);
2457 return max(size, 4UL);
2459 #else
2461 * A zone's size might be changed by hot-add, so it is not possible to determine
2462 * a suitable size for its wait_table. So we use the maximum size now.
2464 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
2466 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
2467 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
2468 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
2470 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
2471 * or more by the traditional way. (See above). It equals:
2473 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
2474 * ia64(16K page size) : = ( 8G + 4M)byte.
2475 * powerpc (64K page size) : = (32G +16M)byte.
2477 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
2479 return 4096UL;
2481 #endif
2484 * This is an integer logarithm so that shifts can be used later
2485 * to extract the more random high bits from the multiplicative
2486 * hash function before the remainder is taken.
2488 static inline unsigned long wait_table_bits(unsigned long size)
2490 return ffz(~size);
2493 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
2496 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
2497 * of blocks reserved is based on zone->pages_min. The memory within the
2498 * reserve will tend to store contiguous free pages. Setting min_free_kbytes
2499 * higher will lead to a bigger reserve which will get freed as contiguous
2500 * blocks as reclaim kicks in
2502 static void setup_zone_migrate_reserve(struct zone *zone)
2504 unsigned long start_pfn, pfn, end_pfn;
2505 struct page *page;
2506 unsigned long reserve, block_migratetype;
2508 /* Get the start pfn, end pfn and the number of blocks to reserve */
2509 start_pfn = zone->zone_start_pfn;
2510 end_pfn = start_pfn + zone->spanned_pages;
2511 reserve = roundup(zone->pages_min, pageblock_nr_pages) >>
2512 pageblock_order;
2514 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
2515 if (!pfn_valid(pfn))
2516 continue;
2517 page = pfn_to_page(pfn);
2519 /* Blocks with reserved pages will never free, skip them. */
2520 if (PageReserved(page))
2521 continue;
2523 block_migratetype = get_pageblock_migratetype(page);
2525 /* If this block is reserved, account for it */
2526 if (reserve > 0 && block_migratetype == MIGRATE_RESERVE) {
2527 reserve--;
2528 continue;
2531 /* Suitable for reserving if this block is movable */
2532 if (reserve > 0 && block_migratetype == MIGRATE_MOVABLE) {
2533 set_pageblock_migratetype(page, MIGRATE_RESERVE);
2534 move_freepages_block(zone, page, MIGRATE_RESERVE);
2535 reserve--;
2536 continue;
2540 * If the reserve is met and this is a previous reserved block,
2541 * take it back
2543 if (block_migratetype == MIGRATE_RESERVE) {
2544 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
2545 move_freepages_block(zone, page, MIGRATE_MOVABLE);
2551 * Initially all pages are reserved - free ones are freed
2552 * up by free_all_bootmem() once the early boot process is
2553 * done. Non-atomic initialization, single-pass.
2555 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
2556 unsigned long start_pfn, enum memmap_context context)
2558 struct page *page;
2559 unsigned long end_pfn = start_pfn + size;
2560 unsigned long pfn;
2561 struct zone *z;
2563 z = &NODE_DATA(nid)->node_zones[zone];
2564 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
2566 * There can be holes in boot-time mem_map[]s
2567 * handed to this function. They do not
2568 * exist on hotplugged memory.
2570 if (context == MEMMAP_EARLY) {
2571 if (!early_pfn_valid(pfn))
2572 continue;
2573 if (!early_pfn_in_nid(pfn, nid))
2574 continue;
2576 page = pfn_to_page(pfn);
2577 set_page_links(page, zone, nid, pfn);
2578 mminit_verify_page_links(page, zone, nid, pfn);
2579 init_page_count(page);
2580 reset_page_mapcount(page);
2581 SetPageReserved(page);
2583 * Mark the block movable so that blocks are reserved for
2584 * movable at startup. This will force kernel allocations
2585 * to reserve their blocks rather than leaking throughout
2586 * the address space during boot when many long-lived
2587 * kernel allocations are made. Later some blocks near
2588 * the start are marked MIGRATE_RESERVE by
2589 * setup_zone_migrate_reserve()
2591 * bitmap is created for zone's valid pfn range. but memmap
2592 * can be created for invalid pages (for alignment)
2593 * check here not to call set_pageblock_migratetype() against
2594 * pfn out of zone.
2596 if ((z->zone_start_pfn <= pfn)
2597 && (pfn < z->zone_start_pfn + z->spanned_pages)
2598 && !(pfn & (pageblock_nr_pages - 1)))
2599 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
2601 INIT_LIST_HEAD(&page->lru);
2602 #ifdef WANT_PAGE_VIRTUAL
2603 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
2604 if (!is_highmem_idx(zone))
2605 set_page_address(page, __va(pfn << PAGE_SHIFT));
2606 #endif
2610 static void __meminit zone_init_free_lists(struct zone *zone)
2612 int order, t;
2613 for_each_migratetype_order(order, t) {
2614 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
2615 zone->free_area[order].nr_free = 0;
2619 #ifndef __HAVE_ARCH_MEMMAP_INIT
2620 #define memmap_init(size, nid, zone, start_pfn) \
2621 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
2622 #endif
2624 static int zone_batchsize(struct zone *zone)
2626 int batch;
2629 * The per-cpu-pages pools are set to around 1000th of the
2630 * size of the zone. But no more than 1/2 of a meg.
2632 * OK, so we don't know how big the cache is. So guess.
2634 batch = zone->present_pages / 1024;
2635 if (batch * PAGE_SIZE > 512 * 1024)
2636 batch = (512 * 1024) / PAGE_SIZE;
2637 batch /= 4; /* We effectively *= 4 below */
2638 if (batch < 1)
2639 batch = 1;
2642 * Clamp the batch to a 2^n - 1 value. Having a power
2643 * of 2 value was found to be more likely to have
2644 * suboptimal cache aliasing properties in some cases.
2646 * For example if 2 tasks are alternately allocating
2647 * batches of pages, one task can end up with a lot
2648 * of pages of one half of the possible page colors
2649 * and the other with pages of the other colors.
2651 batch = (1 << (fls(batch + batch/2)-1)) - 1;
2653 return batch;
2656 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
2658 struct per_cpu_pages *pcp;
2660 memset(p, 0, sizeof(*p));
2662 pcp = &p->pcp;
2663 pcp->count = 0;
2664 pcp->high = 6 * batch;
2665 pcp->batch = max(1UL, 1 * batch);
2666 INIT_LIST_HEAD(&pcp->list);
2670 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
2671 * to the value high for the pageset p.
2674 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
2675 unsigned long high)
2677 struct per_cpu_pages *pcp;
2679 pcp = &p->pcp;
2680 pcp->high = high;
2681 pcp->batch = max(1UL, high/4);
2682 if ((high/4) > (PAGE_SHIFT * 8))
2683 pcp->batch = PAGE_SHIFT * 8;
2687 #ifdef CONFIG_NUMA
2689 * Boot pageset table. One per cpu which is going to be used for all
2690 * zones and all nodes. The parameters will be set in such a way
2691 * that an item put on a list will immediately be handed over to
2692 * the buddy list. This is safe since pageset manipulation is done
2693 * with interrupts disabled.
2695 * Some NUMA counter updates may also be caught by the boot pagesets.
2697 * The boot_pagesets must be kept even after bootup is complete for
2698 * unused processors and/or zones. They do play a role for bootstrapping
2699 * hotplugged processors.
2701 * zoneinfo_show() and maybe other functions do
2702 * not check if the processor is online before following the pageset pointer.
2703 * Other parts of the kernel may not check if the zone is available.
2705 static struct per_cpu_pageset boot_pageset[NR_CPUS];
2708 * Dynamically allocate memory for the
2709 * per cpu pageset array in struct zone.
2711 static int __cpuinit process_zones(int cpu)
2713 struct zone *zone, *dzone;
2714 int node = cpu_to_node(cpu);
2716 node_set_state(node, N_CPU); /* this node has a cpu */
2718 for_each_zone(zone) {
2720 if (!populated_zone(zone))
2721 continue;
2723 zone_pcp(zone, cpu) = kmalloc_node(sizeof(struct per_cpu_pageset),
2724 GFP_KERNEL, node);
2725 if (!zone_pcp(zone, cpu))
2726 goto bad;
2728 setup_pageset(zone_pcp(zone, cpu), zone_batchsize(zone));
2730 if (percpu_pagelist_fraction)
2731 setup_pagelist_highmark(zone_pcp(zone, cpu),
2732 (zone->present_pages / percpu_pagelist_fraction));
2735 return 0;
2736 bad:
2737 for_each_zone(dzone) {
2738 if (!populated_zone(dzone))
2739 continue;
2740 if (dzone == zone)
2741 break;
2742 kfree(zone_pcp(dzone, cpu));
2743 zone_pcp(dzone, cpu) = NULL;
2745 return -ENOMEM;
2748 static inline void free_zone_pagesets(int cpu)
2750 struct zone *zone;
2752 for_each_zone(zone) {
2753 struct per_cpu_pageset *pset = zone_pcp(zone, cpu);
2755 /* Free per_cpu_pageset if it is slab allocated */
2756 if (pset != &boot_pageset[cpu])
2757 kfree(pset);
2758 zone_pcp(zone, cpu) = NULL;
2762 static int __cpuinit pageset_cpuup_callback(struct notifier_block *nfb,
2763 unsigned long action,
2764 void *hcpu)
2766 int cpu = (long)hcpu;
2767 int ret = NOTIFY_OK;
2769 switch (action) {
2770 case CPU_UP_PREPARE:
2771 case CPU_UP_PREPARE_FROZEN:
2772 if (process_zones(cpu))
2773 ret = NOTIFY_BAD;
2774 break;
2775 case CPU_UP_CANCELED:
2776 case CPU_UP_CANCELED_FROZEN:
2777 case CPU_DEAD:
2778 case CPU_DEAD_FROZEN:
2779 free_zone_pagesets(cpu);
2780 break;
2781 default:
2782 break;
2784 return ret;
2787 static struct notifier_block __cpuinitdata pageset_notifier =
2788 { &pageset_cpuup_callback, NULL, 0 };
2790 void __init setup_per_cpu_pageset(void)
2792 int err;
2794 /* Initialize per_cpu_pageset for cpu 0.
2795 * A cpuup callback will do this for every cpu
2796 * as it comes online
2798 err = process_zones(smp_processor_id());
2799 BUG_ON(err);
2800 register_cpu_notifier(&pageset_notifier);
2803 #endif
2805 static noinline __init_refok
2806 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
2808 int i;
2809 struct pglist_data *pgdat = zone->zone_pgdat;
2810 size_t alloc_size;
2813 * The per-page waitqueue mechanism uses hashed waitqueues
2814 * per zone.
2816 zone->wait_table_hash_nr_entries =
2817 wait_table_hash_nr_entries(zone_size_pages);
2818 zone->wait_table_bits =
2819 wait_table_bits(zone->wait_table_hash_nr_entries);
2820 alloc_size = zone->wait_table_hash_nr_entries
2821 * sizeof(wait_queue_head_t);
2823 if (!slab_is_available()) {
2824 zone->wait_table = (wait_queue_head_t *)
2825 alloc_bootmem_node(pgdat, alloc_size);
2826 } else {
2828 * This case means that a zone whose size was 0 gets new memory
2829 * via memory hot-add.
2830 * But it may be the case that a new node was hot-added. In
2831 * this case vmalloc() will not be able to use this new node's
2832 * memory - this wait_table must be initialized to use this new
2833 * node itself as well.
2834 * To use this new node's memory, further consideration will be
2835 * necessary.
2837 zone->wait_table = vmalloc(alloc_size);
2839 if (!zone->wait_table)
2840 return -ENOMEM;
2842 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
2843 init_waitqueue_head(zone->wait_table + i);
2845 return 0;
2848 static __meminit void zone_pcp_init(struct zone *zone)
2850 int cpu;
2851 unsigned long batch = zone_batchsize(zone);
2853 for (cpu = 0; cpu < NR_CPUS; cpu++) {
2854 #ifdef CONFIG_NUMA
2855 /* Early boot. Slab allocator not functional yet */
2856 zone_pcp(zone, cpu) = &boot_pageset[cpu];
2857 setup_pageset(&boot_pageset[cpu],0);
2858 #else
2859 setup_pageset(zone_pcp(zone,cpu), batch);
2860 #endif
2862 if (zone->present_pages)
2863 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%lu\n",
2864 zone->name, zone->present_pages, batch);
2867 __meminit int init_currently_empty_zone(struct zone *zone,
2868 unsigned long zone_start_pfn,
2869 unsigned long size,
2870 enum memmap_context context)
2872 struct pglist_data *pgdat = zone->zone_pgdat;
2873 int ret;
2874 ret = zone_wait_table_init(zone, size);
2875 if (ret)
2876 return ret;
2877 pgdat->nr_zones = zone_idx(zone) + 1;
2879 zone->zone_start_pfn = zone_start_pfn;
2881 mminit_dprintk(MMINIT_TRACE, "memmap_init",
2882 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
2883 pgdat->node_id,
2884 (unsigned long)zone_idx(zone),
2885 zone_start_pfn, (zone_start_pfn + size));
2887 zone_init_free_lists(zone);
2889 return 0;
2892 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
2894 * Basic iterator support. Return the first range of PFNs for a node
2895 * Note: nid == MAX_NUMNODES returns first region regardless of node
2897 static int __meminit first_active_region_index_in_nid(int nid)
2899 int i;
2901 for (i = 0; i < nr_nodemap_entries; i++)
2902 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
2903 return i;
2905 return -1;
2909 * Basic iterator support. Return the next active range of PFNs for a node
2910 * Note: nid == MAX_NUMNODES returns next region regardless of node
2912 static int __meminit next_active_region_index_in_nid(int index, int nid)
2914 for (index = index + 1; index < nr_nodemap_entries; index++)
2915 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
2916 return index;
2918 return -1;
2921 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
2923 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
2924 * Architectures may implement their own version but if add_active_range()
2925 * was used and there are no special requirements, this is a convenient
2926 * alternative
2928 int __meminit early_pfn_to_nid(unsigned long pfn)
2930 int i;
2932 for (i = 0; i < nr_nodemap_entries; i++) {
2933 unsigned long start_pfn = early_node_map[i].start_pfn;
2934 unsigned long end_pfn = early_node_map[i].end_pfn;
2936 if (start_pfn <= pfn && pfn < end_pfn)
2937 return early_node_map[i].nid;
2940 return 0;
2942 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
2944 /* Basic iterator support to walk early_node_map[] */
2945 #define for_each_active_range_index_in_nid(i, nid) \
2946 for (i = first_active_region_index_in_nid(nid); i != -1; \
2947 i = next_active_region_index_in_nid(i, nid))
2950 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
2951 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
2952 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
2954 * If an architecture guarantees that all ranges registered with
2955 * add_active_ranges() contain no holes and may be freed, this
2956 * this function may be used instead of calling free_bootmem() manually.
2958 void __init free_bootmem_with_active_regions(int nid,
2959 unsigned long max_low_pfn)
2961 int i;
2963 for_each_active_range_index_in_nid(i, nid) {
2964 unsigned long size_pages = 0;
2965 unsigned long end_pfn = early_node_map[i].end_pfn;
2967 if (early_node_map[i].start_pfn >= max_low_pfn)
2968 continue;
2970 if (end_pfn > max_low_pfn)
2971 end_pfn = max_low_pfn;
2973 size_pages = end_pfn - early_node_map[i].start_pfn;
2974 free_bootmem_node(NODE_DATA(early_node_map[i].nid),
2975 PFN_PHYS(early_node_map[i].start_pfn),
2976 size_pages << PAGE_SHIFT);
2980 void __init work_with_active_regions(int nid, work_fn_t work_fn, void *data)
2982 int i;
2983 int ret;
2985 for_each_active_range_index_in_nid(i, nid) {
2986 ret = work_fn(early_node_map[i].start_pfn,
2987 early_node_map[i].end_pfn, data);
2988 if (ret)
2989 break;
2993 * sparse_memory_present_with_active_regions - Call memory_present for each active range
2994 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
2996 * If an architecture guarantees that all ranges registered with
2997 * add_active_ranges() contain no holes and may be freed, this
2998 * function may be used instead of calling memory_present() manually.
3000 void __init sparse_memory_present_with_active_regions(int nid)
3002 int i;
3004 for_each_active_range_index_in_nid(i, nid)
3005 memory_present(early_node_map[i].nid,
3006 early_node_map[i].start_pfn,
3007 early_node_map[i].end_pfn);
3011 * push_node_boundaries - Push node boundaries to at least the requested boundary
3012 * @nid: The nid of the node to push the boundary for
3013 * @start_pfn: The start pfn of the node
3014 * @end_pfn: The end pfn of the node
3016 * In reserve-based hot-add, mem_map is allocated that is unused until hotadd
3017 * time. Specifically, on x86_64, SRAT will report ranges that can potentially
3018 * be hotplugged even though no physical memory exists. This function allows
3019 * an arch to push out the node boundaries so mem_map is allocated that can
3020 * be used later.
3022 #ifdef CONFIG_MEMORY_HOTPLUG_RESERVE
3023 void __init push_node_boundaries(unsigned int nid,
3024 unsigned long start_pfn, unsigned long end_pfn)
3026 mminit_dprintk(MMINIT_TRACE, "zoneboundary",
3027 "Entering push_node_boundaries(%u, %lu, %lu)\n",
3028 nid, start_pfn, end_pfn);
3030 /* Initialise the boundary for this node if necessary */
3031 if (node_boundary_end_pfn[nid] == 0)
3032 node_boundary_start_pfn[nid] = -1UL;
3034 /* Update the boundaries */
3035 if (node_boundary_start_pfn[nid] > start_pfn)
3036 node_boundary_start_pfn[nid] = start_pfn;
3037 if (node_boundary_end_pfn[nid] < end_pfn)
3038 node_boundary_end_pfn[nid] = end_pfn;
3041 /* If necessary, push the node boundary out for reserve hotadd */
3042 static void __meminit account_node_boundary(unsigned int nid,
3043 unsigned long *start_pfn, unsigned long *end_pfn)
3045 mminit_dprintk(MMINIT_TRACE, "zoneboundary",
3046 "Entering account_node_boundary(%u, %lu, %lu)\n",
3047 nid, *start_pfn, *end_pfn);
3049 /* Return if boundary information has not been provided */
3050 if (node_boundary_end_pfn[nid] == 0)
3051 return;
3053 /* Check the boundaries and update if necessary */
3054 if (node_boundary_start_pfn[nid] < *start_pfn)
3055 *start_pfn = node_boundary_start_pfn[nid];
3056 if (node_boundary_end_pfn[nid] > *end_pfn)
3057 *end_pfn = node_boundary_end_pfn[nid];
3059 #else
3060 void __init push_node_boundaries(unsigned int nid,
3061 unsigned long start_pfn, unsigned long end_pfn) {}
3063 static void __meminit account_node_boundary(unsigned int nid,
3064 unsigned long *start_pfn, unsigned long *end_pfn) {}
3065 #endif
3069 * get_pfn_range_for_nid - Return the start and end page frames for a node
3070 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
3071 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
3072 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
3074 * It returns the start and end page frame of a node based on information
3075 * provided by an arch calling add_active_range(). If called for a node
3076 * with no available memory, a warning is printed and the start and end
3077 * PFNs will be 0.
3079 void __meminit get_pfn_range_for_nid(unsigned int nid,
3080 unsigned long *start_pfn, unsigned long *end_pfn)
3082 int i;
3083 *start_pfn = -1UL;
3084 *end_pfn = 0;
3086 for_each_active_range_index_in_nid(i, nid) {
3087 *start_pfn = min(*start_pfn, early_node_map[i].start_pfn);
3088 *end_pfn = max(*end_pfn, early_node_map[i].end_pfn);
3091 if (*start_pfn == -1UL)
3092 *start_pfn = 0;
3094 /* Push the node boundaries out if requested */
3095 account_node_boundary(nid, start_pfn, end_pfn);
3099 * This finds a zone that can be used for ZONE_MOVABLE pages. The
3100 * assumption is made that zones within a node are ordered in monotonic
3101 * increasing memory addresses so that the "highest" populated zone is used
3103 static void __init find_usable_zone_for_movable(void)
3105 int zone_index;
3106 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
3107 if (zone_index == ZONE_MOVABLE)
3108 continue;
3110 if (arch_zone_highest_possible_pfn[zone_index] >
3111 arch_zone_lowest_possible_pfn[zone_index])
3112 break;
3115 VM_BUG_ON(zone_index == -1);
3116 movable_zone = zone_index;
3120 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
3121 * because it is sized independant of architecture. Unlike the other zones,
3122 * the starting point for ZONE_MOVABLE is not fixed. It may be different
3123 * in each node depending on the size of each node and how evenly kernelcore
3124 * is distributed. This helper function adjusts the zone ranges
3125 * provided by the architecture for a given node by using the end of the
3126 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
3127 * zones within a node are in order of monotonic increases memory addresses
3129 static void __meminit adjust_zone_range_for_zone_movable(int nid,
3130 unsigned long zone_type,
3131 unsigned long node_start_pfn,
3132 unsigned long node_end_pfn,
3133 unsigned long *zone_start_pfn,
3134 unsigned long *zone_end_pfn)
3136 /* Only adjust if ZONE_MOVABLE is on this node */
3137 if (zone_movable_pfn[nid]) {
3138 /* Size ZONE_MOVABLE */
3139 if (zone_type == ZONE_MOVABLE) {
3140 *zone_start_pfn = zone_movable_pfn[nid];
3141 *zone_end_pfn = min(node_end_pfn,
3142 arch_zone_highest_possible_pfn[movable_zone]);
3144 /* Adjust for ZONE_MOVABLE starting within this range */
3145 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
3146 *zone_end_pfn > zone_movable_pfn[nid]) {
3147 *zone_end_pfn = zone_movable_pfn[nid];
3149 /* Check if this whole range is within ZONE_MOVABLE */
3150 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
3151 *zone_start_pfn = *zone_end_pfn;
3156 * Return the number of pages a zone spans in a node, including holes
3157 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
3159 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
3160 unsigned long zone_type,
3161 unsigned long *ignored)
3163 unsigned long node_start_pfn, node_end_pfn;
3164 unsigned long zone_start_pfn, zone_end_pfn;
3166 /* Get the start and end of the node and zone */
3167 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
3168 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
3169 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
3170 adjust_zone_range_for_zone_movable(nid, zone_type,
3171 node_start_pfn, node_end_pfn,
3172 &zone_start_pfn, &zone_end_pfn);
3174 /* Check that this node has pages within the zone's required range */
3175 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
3176 return 0;
3178 /* Move the zone boundaries inside the node if necessary */
3179 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
3180 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
3182 /* Return the spanned pages */
3183 return zone_end_pfn - zone_start_pfn;
3187 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
3188 * then all holes in the requested range will be accounted for.
3190 static unsigned long __meminit __absent_pages_in_range(int nid,
3191 unsigned long range_start_pfn,
3192 unsigned long range_end_pfn)
3194 int i = 0;
3195 unsigned long prev_end_pfn = 0, hole_pages = 0;
3196 unsigned long start_pfn;
3198 /* Find the end_pfn of the first active range of pfns in the node */
3199 i = first_active_region_index_in_nid(nid);
3200 if (i == -1)
3201 return 0;
3203 prev_end_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
3205 /* Account for ranges before physical memory on this node */
3206 if (early_node_map[i].start_pfn > range_start_pfn)
3207 hole_pages = prev_end_pfn - range_start_pfn;
3209 /* Find all holes for the zone within the node */
3210 for (; i != -1; i = next_active_region_index_in_nid(i, nid)) {
3212 /* No need to continue if prev_end_pfn is outside the zone */
3213 if (prev_end_pfn >= range_end_pfn)
3214 break;
3216 /* Make sure the end of the zone is not within the hole */
3217 start_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
3218 prev_end_pfn = max(prev_end_pfn, range_start_pfn);
3220 /* Update the hole size cound and move on */
3221 if (start_pfn > range_start_pfn) {
3222 BUG_ON(prev_end_pfn > start_pfn);
3223 hole_pages += start_pfn - prev_end_pfn;
3225 prev_end_pfn = early_node_map[i].end_pfn;
3228 /* Account for ranges past physical memory on this node */
3229 if (range_end_pfn > prev_end_pfn)
3230 hole_pages += range_end_pfn -
3231 max(range_start_pfn, prev_end_pfn);
3233 return hole_pages;
3237 * absent_pages_in_range - Return number of page frames in holes within a range
3238 * @start_pfn: The start PFN to start searching for holes
3239 * @end_pfn: The end PFN to stop searching for holes
3241 * It returns the number of pages frames in memory holes within a range.
3243 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
3244 unsigned long end_pfn)
3246 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
3249 /* Return the number of page frames in holes in a zone on a node */
3250 static unsigned long __meminit zone_absent_pages_in_node(int nid,
3251 unsigned long zone_type,
3252 unsigned long *ignored)
3254 unsigned long node_start_pfn, node_end_pfn;
3255 unsigned long zone_start_pfn, zone_end_pfn;
3257 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
3258 zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type],
3259 node_start_pfn);
3260 zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type],
3261 node_end_pfn);
3263 adjust_zone_range_for_zone_movable(nid, zone_type,
3264 node_start_pfn, node_end_pfn,
3265 &zone_start_pfn, &zone_end_pfn);
3266 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
3269 #else
3270 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
3271 unsigned long zone_type,
3272 unsigned long *zones_size)
3274 return zones_size[zone_type];
3277 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
3278 unsigned long zone_type,
3279 unsigned long *zholes_size)
3281 if (!zholes_size)
3282 return 0;
3284 return zholes_size[zone_type];
3287 #endif
3289 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
3290 unsigned long *zones_size, unsigned long *zholes_size)
3292 unsigned long realtotalpages, totalpages = 0;
3293 enum zone_type i;
3295 for (i = 0; i < MAX_NR_ZONES; i++)
3296 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
3297 zones_size);
3298 pgdat->node_spanned_pages = totalpages;
3300 realtotalpages = totalpages;
3301 for (i = 0; i < MAX_NR_ZONES; i++)
3302 realtotalpages -=
3303 zone_absent_pages_in_node(pgdat->node_id, i,
3304 zholes_size);
3305 pgdat->node_present_pages = realtotalpages;
3306 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
3307 realtotalpages);
3310 #ifndef CONFIG_SPARSEMEM
3312 * Calculate the size of the zone->blockflags rounded to an unsigned long
3313 * Start by making sure zonesize is a multiple of pageblock_order by rounding
3314 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
3315 * round what is now in bits to nearest long in bits, then return it in
3316 * bytes.
3318 static unsigned long __init usemap_size(unsigned long zonesize)
3320 unsigned long usemapsize;
3322 usemapsize = roundup(zonesize, pageblock_nr_pages);
3323 usemapsize = usemapsize >> pageblock_order;
3324 usemapsize *= NR_PAGEBLOCK_BITS;
3325 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
3327 return usemapsize / 8;
3330 static void __init setup_usemap(struct pglist_data *pgdat,
3331 struct zone *zone, unsigned long zonesize)
3333 unsigned long usemapsize = usemap_size(zonesize);
3334 zone->pageblock_flags = NULL;
3335 if (usemapsize) {
3336 zone->pageblock_flags = alloc_bootmem_node(pgdat, usemapsize);
3337 memset(zone->pageblock_flags, 0, usemapsize);
3340 #else
3341 static void inline setup_usemap(struct pglist_data *pgdat,
3342 struct zone *zone, unsigned long zonesize) {}
3343 #endif /* CONFIG_SPARSEMEM */
3345 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
3347 /* Return a sensible default order for the pageblock size. */
3348 static inline int pageblock_default_order(void)
3350 if (HPAGE_SHIFT > PAGE_SHIFT)
3351 return HUGETLB_PAGE_ORDER;
3353 return MAX_ORDER-1;
3356 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
3357 static inline void __init set_pageblock_order(unsigned int order)
3359 /* Check that pageblock_nr_pages has not already been setup */
3360 if (pageblock_order)
3361 return;
3364 * Assume the largest contiguous order of interest is a huge page.
3365 * This value may be variable depending on boot parameters on IA64
3367 pageblock_order = order;
3369 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
3372 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
3373 * and pageblock_default_order() are unused as pageblock_order is set
3374 * at compile-time. See include/linux/pageblock-flags.h for the values of
3375 * pageblock_order based on the kernel config
3377 static inline int pageblock_default_order(unsigned int order)
3379 return MAX_ORDER-1;
3381 #define set_pageblock_order(x) do {} while (0)
3383 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
3386 * Set up the zone data structures:
3387 * - mark all pages reserved
3388 * - mark all memory queues empty
3389 * - clear the memory bitmaps
3391 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
3392 unsigned long *zones_size, unsigned long *zholes_size)
3394 enum zone_type j;
3395 int nid = pgdat->node_id;
3396 unsigned long zone_start_pfn = pgdat->node_start_pfn;
3397 int ret;
3399 pgdat_resize_init(pgdat);
3400 pgdat->nr_zones = 0;
3401 init_waitqueue_head(&pgdat->kswapd_wait);
3402 pgdat->kswapd_max_order = 0;
3404 for (j = 0; j < MAX_NR_ZONES; j++) {
3405 struct zone *zone = pgdat->node_zones + j;
3406 unsigned long size, realsize, memmap_pages;
3408 size = zone_spanned_pages_in_node(nid, j, zones_size);
3409 realsize = size - zone_absent_pages_in_node(nid, j,
3410 zholes_size);
3413 * Adjust realsize so that it accounts for how much memory
3414 * is used by this zone for memmap. This affects the watermark
3415 * and per-cpu initialisations
3417 memmap_pages =
3418 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT;
3419 if (realsize >= memmap_pages) {
3420 realsize -= memmap_pages;
3421 mminit_dprintk(MMINIT_TRACE, "memmap_init",
3422 "%s zone: %lu pages used for memmap\n",
3423 zone_names[j], memmap_pages);
3424 } else
3425 printk(KERN_WARNING
3426 " %s zone: %lu pages exceeds realsize %lu\n",
3427 zone_names[j], memmap_pages, realsize);
3429 /* Account for reserved pages */
3430 if (j == 0 && realsize > dma_reserve) {
3431 realsize -= dma_reserve;
3432 mminit_dprintk(MMINIT_TRACE, "memmap_init",
3433 "%s zone: %lu pages reserved\n",
3434 zone_names[0], dma_reserve);
3437 if (!is_highmem_idx(j))
3438 nr_kernel_pages += realsize;
3439 nr_all_pages += realsize;
3441 zone->spanned_pages = size;
3442 zone->present_pages = realsize;
3443 #ifdef CONFIG_NUMA
3444 zone->node = nid;
3445 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
3446 / 100;
3447 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
3448 #endif
3449 zone->name = zone_names[j];
3450 spin_lock_init(&zone->lock);
3451 spin_lock_init(&zone->lru_lock);
3452 zone_seqlock_init(zone);
3453 zone->zone_pgdat = pgdat;
3455 zone->prev_priority = DEF_PRIORITY;
3457 zone_pcp_init(zone);
3458 INIT_LIST_HEAD(&zone->active_list);
3459 INIT_LIST_HEAD(&zone->inactive_list);
3460 zone->nr_scan_active = 0;
3461 zone->nr_scan_inactive = 0;
3462 zap_zone_vm_stats(zone);
3463 zone->flags = 0;
3464 if (!size)
3465 continue;
3467 set_pageblock_order(pageblock_default_order());
3468 setup_usemap(pgdat, zone, size);
3469 ret = init_currently_empty_zone(zone, zone_start_pfn,
3470 size, MEMMAP_EARLY);
3471 BUG_ON(ret);
3472 memmap_init(size, nid, j, zone_start_pfn);
3473 zone_start_pfn += size;
3477 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
3479 /* Skip empty nodes */
3480 if (!pgdat->node_spanned_pages)
3481 return;
3483 #ifdef CONFIG_FLAT_NODE_MEM_MAP
3484 /* ia64 gets its own node_mem_map, before this, without bootmem */
3485 if (!pgdat->node_mem_map) {
3486 unsigned long size, start, end;
3487 struct page *map;
3490 * The zone's endpoints aren't required to be MAX_ORDER
3491 * aligned but the node_mem_map endpoints must be in order
3492 * for the buddy allocator to function correctly.
3494 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
3495 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
3496 end = ALIGN(end, MAX_ORDER_NR_PAGES);
3497 size = (end - start) * sizeof(struct page);
3498 map = alloc_remap(pgdat->node_id, size);
3499 if (!map)
3500 map = alloc_bootmem_node(pgdat, size);
3501 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
3503 #ifndef CONFIG_NEED_MULTIPLE_NODES
3505 * With no DISCONTIG, the global mem_map is just set as node 0's
3507 if (pgdat == NODE_DATA(0)) {
3508 mem_map = NODE_DATA(0)->node_mem_map;
3509 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3510 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
3511 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
3512 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
3514 #endif
3515 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
3518 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
3519 unsigned long node_start_pfn, unsigned long *zholes_size)
3521 pg_data_t *pgdat = NODE_DATA(nid);
3523 pgdat->node_id = nid;
3524 pgdat->node_start_pfn = node_start_pfn;
3525 calculate_node_totalpages(pgdat, zones_size, zholes_size);
3527 alloc_node_mem_map(pgdat);
3528 #ifdef CONFIG_FLAT_NODE_MEM_MAP
3529 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
3530 nid, (unsigned long)pgdat,
3531 (unsigned long)pgdat->node_mem_map);
3532 #endif
3534 free_area_init_core(pgdat, zones_size, zholes_size);
3537 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3539 #if MAX_NUMNODES > 1
3541 * Figure out the number of possible node ids.
3543 static void __init setup_nr_node_ids(void)
3545 unsigned int node;
3546 unsigned int highest = 0;
3548 for_each_node_mask(node, node_possible_map)
3549 highest = node;
3550 nr_node_ids = highest + 1;
3552 #else
3553 static inline void setup_nr_node_ids(void)
3556 #endif
3559 * add_active_range - Register a range of PFNs backed by physical memory
3560 * @nid: The node ID the range resides on
3561 * @start_pfn: The start PFN of the available physical memory
3562 * @end_pfn: The end PFN of the available physical memory
3564 * These ranges are stored in an early_node_map[] and later used by
3565 * free_area_init_nodes() to calculate zone sizes and holes. If the
3566 * range spans a memory hole, it is up to the architecture to ensure
3567 * the memory is not freed by the bootmem allocator. If possible
3568 * the range being registered will be merged with existing ranges.
3570 void __init add_active_range(unsigned int nid, unsigned long start_pfn,
3571 unsigned long end_pfn)
3573 int i;
3575 mminit_dprintk(MMINIT_TRACE, "memory_register",
3576 "Entering add_active_range(%d, %#lx, %#lx) "
3577 "%d entries of %d used\n",
3578 nid, start_pfn, end_pfn,
3579 nr_nodemap_entries, MAX_ACTIVE_REGIONS);
3581 mminit_validate_memmodel_limits(&start_pfn, &end_pfn);
3583 /* Merge with existing active regions if possible */
3584 for (i = 0; i < nr_nodemap_entries; i++) {
3585 if (early_node_map[i].nid != nid)
3586 continue;
3588 /* Skip if an existing region covers this new one */
3589 if (start_pfn >= early_node_map[i].start_pfn &&
3590 end_pfn <= early_node_map[i].end_pfn)
3591 return;
3593 /* Merge forward if suitable */
3594 if (start_pfn <= early_node_map[i].end_pfn &&
3595 end_pfn > early_node_map[i].end_pfn) {
3596 early_node_map[i].end_pfn = end_pfn;
3597 return;
3600 /* Merge backward if suitable */
3601 if (start_pfn < early_node_map[i].end_pfn &&
3602 end_pfn >= early_node_map[i].start_pfn) {
3603 early_node_map[i].start_pfn = start_pfn;
3604 return;
3608 /* Check that early_node_map is large enough */
3609 if (i >= MAX_ACTIVE_REGIONS) {
3610 printk(KERN_CRIT "More than %d memory regions, truncating\n",
3611 MAX_ACTIVE_REGIONS);
3612 return;
3615 early_node_map[i].nid = nid;
3616 early_node_map[i].start_pfn = start_pfn;
3617 early_node_map[i].end_pfn = end_pfn;
3618 nr_nodemap_entries = i + 1;
3622 * remove_active_range - Shrink an existing registered range of PFNs
3623 * @nid: The node id the range is on that should be shrunk
3624 * @start_pfn: The new PFN of the range
3625 * @end_pfn: The new PFN of the range
3627 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
3628 * The map is kept near the end physical page range that has already been
3629 * registered. This function allows an arch to shrink an existing registered
3630 * range.
3632 void __init remove_active_range(unsigned int nid, unsigned long start_pfn,
3633 unsigned long end_pfn)
3635 int i, j;
3636 int removed = 0;
3638 printk(KERN_DEBUG "remove_active_range (%d, %lu, %lu)\n",
3639 nid, start_pfn, end_pfn);
3641 /* Find the old active region end and shrink */
3642 for_each_active_range_index_in_nid(i, nid) {
3643 if (early_node_map[i].start_pfn >= start_pfn &&
3644 early_node_map[i].end_pfn <= end_pfn) {
3645 /* clear it */
3646 early_node_map[i].start_pfn = 0;
3647 early_node_map[i].end_pfn = 0;
3648 removed = 1;
3649 continue;
3651 if (early_node_map[i].start_pfn < start_pfn &&
3652 early_node_map[i].end_pfn > start_pfn) {
3653 unsigned long temp_end_pfn = early_node_map[i].end_pfn;
3654 early_node_map[i].end_pfn = start_pfn;
3655 if (temp_end_pfn > end_pfn)
3656 add_active_range(nid, end_pfn, temp_end_pfn);
3657 continue;
3659 if (early_node_map[i].start_pfn >= start_pfn &&
3660 early_node_map[i].end_pfn > end_pfn &&
3661 early_node_map[i].start_pfn < end_pfn) {
3662 early_node_map[i].start_pfn = end_pfn;
3663 continue;
3667 if (!removed)
3668 return;
3670 /* remove the blank ones */
3671 for (i = nr_nodemap_entries - 1; i > 0; i--) {
3672 if (early_node_map[i].nid != nid)
3673 continue;
3674 if (early_node_map[i].end_pfn)
3675 continue;
3676 /* we found it, get rid of it */
3677 for (j = i; j < nr_nodemap_entries - 1; j++)
3678 memcpy(&early_node_map[j], &early_node_map[j+1],
3679 sizeof(early_node_map[j]));
3680 j = nr_nodemap_entries - 1;
3681 memset(&early_node_map[j], 0, sizeof(early_node_map[j]));
3682 nr_nodemap_entries--;
3687 * remove_all_active_ranges - Remove all currently registered regions
3689 * During discovery, it may be found that a table like SRAT is invalid
3690 * and an alternative discovery method must be used. This function removes
3691 * all currently registered regions.
3693 void __init remove_all_active_ranges(void)
3695 memset(early_node_map, 0, sizeof(early_node_map));
3696 nr_nodemap_entries = 0;
3697 #ifdef CONFIG_MEMORY_HOTPLUG_RESERVE
3698 memset(node_boundary_start_pfn, 0, sizeof(node_boundary_start_pfn));
3699 memset(node_boundary_end_pfn, 0, sizeof(node_boundary_end_pfn));
3700 #endif /* CONFIG_MEMORY_HOTPLUG_RESERVE */
3703 /* Compare two active node_active_regions */
3704 static int __init cmp_node_active_region(const void *a, const void *b)
3706 struct node_active_region *arange = (struct node_active_region *)a;
3707 struct node_active_region *brange = (struct node_active_region *)b;
3709 /* Done this way to avoid overflows */
3710 if (arange->start_pfn > brange->start_pfn)
3711 return 1;
3712 if (arange->start_pfn < brange->start_pfn)
3713 return -1;
3715 return 0;
3718 /* sort the node_map by start_pfn */
3719 static void __init sort_node_map(void)
3721 sort(early_node_map, (size_t)nr_nodemap_entries,
3722 sizeof(struct node_active_region),
3723 cmp_node_active_region, NULL);
3726 /* Find the lowest pfn for a node */
3727 static unsigned long __init find_min_pfn_for_node(int nid)
3729 int i;
3730 unsigned long min_pfn = ULONG_MAX;
3732 /* Assuming a sorted map, the first range found has the starting pfn */
3733 for_each_active_range_index_in_nid(i, nid)
3734 min_pfn = min(min_pfn, early_node_map[i].start_pfn);
3736 if (min_pfn == ULONG_MAX) {
3737 printk(KERN_WARNING
3738 "Could not find start_pfn for node %d\n", nid);
3739 return 0;
3742 return min_pfn;
3746 * find_min_pfn_with_active_regions - Find the minimum PFN registered
3748 * It returns the minimum PFN based on information provided via
3749 * add_active_range().
3751 unsigned long __init find_min_pfn_with_active_regions(void)
3753 return find_min_pfn_for_node(MAX_NUMNODES);
3757 * early_calculate_totalpages()
3758 * Sum pages in active regions for movable zone.
3759 * Populate N_HIGH_MEMORY for calculating usable_nodes.
3761 static unsigned long __init early_calculate_totalpages(void)
3763 int i;
3764 unsigned long totalpages = 0;
3766 for (i = 0; i < nr_nodemap_entries; i++) {
3767 unsigned long pages = early_node_map[i].end_pfn -
3768 early_node_map[i].start_pfn;
3769 totalpages += pages;
3770 if (pages)
3771 node_set_state(early_node_map[i].nid, N_HIGH_MEMORY);
3773 return totalpages;
3777 * Find the PFN the Movable zone begins in each node. Kernel memory
3778 * is spread evenly between nodes as long as the nodes have enough
3779 * memory. When they don't, some nodes will have more kernelcore than
3780 * others
3782 static void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn)
3784 int i, nid;
3785 unsigned long usable_startpfn;
3786 unsigned long kernelcore_node, kernelcore_remaining;
3787 unsigned long totalpages = early_calculate_totalpages();
3788 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
3791 * If movablecore was specified, calculate what size of
3792 * kernelcore that corresponds so that memory usable for
3793 * any allocation type is evenly spread. If both kernelcore
3794 * and movablecore are specified, then the value of kernelcore
3795 * will be used for required_kernelcore if it's greater than
3796 * what movablecore would have allowed.
3798 if (required_movablecore) {
3799 unsigned long corepages;
3802 * Round-up so that ZONE_MOVABLE is at least as large as what
3803 * was requested by the user
3805 required_movablecore =
3806 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
3807 corepages = totalpages - required_movablecore;
3809 required_kernelcore = max(required_kernelcore, corepages);
3812 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
3813 if (!required_kernelcore)
3814 return;
3816 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
3817 find_usable_zone_for_movable();
3818 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
3820 restart:
3821 /* Spread kernelcore memory as evenly as possible throughout nodes */
3822 kernelcore_node = required_kernelcore / usable_nodes;
3823 for_each_node_state(nid, N_HIGH_MEMORY) {
3825 * Recalculate kernelcore_node if the division per node
3826 * now exceeds what is necessary to satisfy the requested
3827 * amount of memory for the kernel
3829 if (required_kernelcore < kernelcore_node)
3830 kernelcore_node = required_kernelcore / usable_nodes;
3833 * As the map is walked, we track how much memory is usable
3834 * by the kernel using kernelcore_remaining. When it is
3835 * 0, the rest of the node is usable by ZONE_MOVABLE
3837 kernelcore_remaining = kernelcore_node;
3839 /* Go through each range of PFNs within this node */
3840 for_each_active_range_index_in_nid(i, nid) {
3841 unsigned long start_pfn, end_pfn;
3842 unsigned long size_pages;
3844 start_pfn = max(early_node_map[i].start_pfn,
3845 zone_movable_pfn[nid]);
3846 end_pfn = early_node_map[i].end_pfn;
3847 if (start_pfn >= end_pfn)
3848 continue;
3850 /* Account for what is only usable for kernelcore */
3851 if (start_pfn < usable_startpfn) {
3852 unsigned long kernel_pages;
3853 kernel_pages = min(end_pfn, usable_startpfn)
3854 - start_pfn;
3856 kernelcore_remaining -= min(kernel_pages,
3857 kernelcore_remaining);
3858 required_kernelcore -= min(kernel_pages,
3859 required_kernelcore);
3861 /* Continue if range is now fully accounted */
3862 if (end_pfn <= usable_startpfn) {
3865 * Push zone_movable_pfn to the end so
3866 * that if we have to rebalance
3867 * kernelcore across nodes, we will
3868 * not double account here
3870 zone_movable_pfn[nid] = end_pfn;
3871 continue;
3873 start_pfn = usable_startpfn;
3877 * The usable PFN range for ZONE_MOVABLE is from
3878 * start_pfn->end_pfn. Calculate size_pages as the
3879 * number of pages used as kernelcore
3881 size_pages = end_pfn - start_pfn;
3882 if (size_pages > kernelcore_remaining)
3883 size_pages = kernelcore_remaining;
3884 zone_movable_pfn[nid] = start_pfn + size_pages;
3887 * Some kernelcore has been met, update counts and
3888 * break if the kernelcore for this node has been
3889 * satisified
3891 required_kernelcore -= min(required_kernelcore,
3892 size_pages);
3893 kernelcore_remaining -= size_pages;
3894 if (!kernelcore_remaining)
3895 break;
3900 * If there is still required_kernelcore, we do another pass with one
3901 * less node in the count. This will push zone_movable_pfn[nid] further
3902 * along on the nodes that still have memory until kernelcore is
3903 * satisified
3905 usable_nodes--;
3906 if (usable_nodes && required_kernelcore > usable_nodes)
3907 goto restart;
3909 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
3910 for (nid = 0; nid < MAX_NUMNODES; nid++)
3911 zone_movable_pfn[nid] =
3912 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
3915 /* Any regular memory on that node ? */
3916 static void check_for_regular_memory(pg_data_t *pgdat)
3918 #ifdef CONFIG_HIGHMEM
3919 enum zone_type zone_type;
3921 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) {
3922 struct zone *zone = &pgdat->node_zones[zone_type];
3923 if (zone->present_pages)
3924 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY);
3926 #endif
3930 * free_area_init_nodes - Initialise all pg_data_t and zone data
3931 * @max_zone_pfn: an array of max PFNs for each zone
3933 * This will call free_area_init_node() for each active node in the system.
3934 * Using the page ranges provided by add_active_range(), the size of each
3935 * zone in each node and their holes is calculated. If the maximum PFN
3936 * between two adjacent zones match, it is assumed that the zone is empty.
3937 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
3938 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
3939 * starts where the previous one ended. For example, ZONE_DMA32 starts
3940 * at arch_max_dma_pfn.
3942 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
3944 unsigned long nid;
3945 enum zone_type i;
3947 /* Sort early_node_map as initialisation assumes it is sorted */
3948 sort_node_map();
3950 /* Record where the zone boundaries are */
3951 memset(arch_zone_lowest_possible_pfn, 0,
3952 sizeof(arch_zone_lowest_possible_pfn));
3953 memset(arch_zone_highest_possible_pfn, 0,
3954 sizeof(arch_zone_highest_possible_pfn));
3955 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
3956 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
3957 for (i = 1; i < MAX_NR_ZONES; i++) {
3958 if (i == ZONE_MOVABLE)
3959 continue;
3960 arch_zone_lowest_possible_pfn[i] =
3961 arch_zone_highest_possible_pfn[i-1];
3962 arch_zone_highest_possible_pfn[i] =
3963 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
3965 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
3966 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
3968 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
3969 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
3970 find_zone_movable_pfns_for_nodes(zone_movable_pfn);
3972 /* Print out the zone ranges */
3973 printk("Zone PFN ranges:\n");
3974 for (i = 0; i < MAX_NR_ZONES; i++) {
3975 if (i == ZONE_MOVABLE)
3976 continue;
3977 printk(" %-8s %0#10lx -> %0#10lx\n",
3978 zone_names[i],
3979 arch_zone_lowest_possible_pfn[i],
3980 arch_zone_highest_possible_pfn[i]);
3983 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
3984 printk("Movable zone start PFN for each node\n");
3985 for (i = 0; i < MAX_NUMNODES; i++) {
3986 if (zone_movable_pfn[i])
3987 printk(" Node %d: %lu\n", i, zone_movable_pfn[i]);
3990 /* Print out the early_node_map[] */
3991 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries);
3992 for (i = 0; i < nr_nodemap_entries; i++)
3993 printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map[i].nid,
3994 early_node_map[i].start_pfn,
3995 early_node_map[i].end_pfn);
3997 /* Initialise every node */
3998 mminit_verify_pageflags_layout();
3999 setup_nr_node_ids();
4000 for_each_online_node(nid) {
4001 pg_data_t *pgdat = NODE_DATA(nid);
4002 free_area_init_node(nid, NULL,
4003 find_min_pfn_for_node(nid), NULL);
4005 /* Any memory on that node */
4006 if (pgdat->node_present_pages)
4007 node_set_state(nid, N_HIGH_MEMORY);
4008 check_for_regular_memory(pgdat);
4012 static int __init cmdline_parse_core(char *p, unsigned long *core)
4014 unsigned long long coremem;
4015 if (!p)
4016 return -EINVAL;
4018 coremem = memparse(p, &p);
4019 *core = coremem >> PAGE_SHIFT;
4021 /* Paranoid check that UL is enough for the coremem value */
4022 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
4024 return 0;
4028 * kernelcore=size sets the amount of memory for use for allocations that
4029 * cannot be reclaimed or migrated.
4031 static int __init cmdline_parse_kernelcore(char *p)
4033 return cmdline_parse_core(p, &required_kernelcore);
4037 * movablecore=size sets the amount of memory for use for allocations that
4038 * can be reclaimed or migrated.
4040 static int __init cmdline_parse_movablecore(char *p)
4042 return cmdline_parse_core(p, &required_movablecore);
4045 early_param("kernelcore", cmdline_parse_kernelcore);
4046 early_param("movablecore", cmdline_parse_movablecore);
4048 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4051 * set_dma_reserve - set the specified number of pages reserved in the first zone
4052 * @new_dma_reserve: The number of pages to mark reserved
4054 * The per-cpu batchsize and zone watermarks are determined by present_pages.
4055 * In the DMA zone, a significant percentage may be consumed by kernel image
4056 * and other unfreeable allocations which can skew the watermarks badly. This
4057 * function may optionally be used to account for unfreeable pages in the
4058 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
4059 * smaller per-cpu batchsize.
4061 void __init set_dma_reserve(unsigned long new_dma_reserve)
4063 dma_reserve = new_dma_reserve;
4066 #ifndef CONFIG_NEED_MULTIPLE_NODES
4067 struct pglist_data contig_page_data = { .bdata = &bootmem_node_data[0] };
4068 EXPORT_SYMBOL(contig_page_data);
4069 #endif
4071 void __init free_area_init(unsigned long *zones_size)
4073 free_area_init_node(0, zones_size,
4074 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
4077 static int page_alloc_cpu_notify(struct notifier_block *self,
4078 unsigned long action, void *hcpu)
4080 int cpu = (unsigned long)hcpu;
4082 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
4083 drain_pages(cpu);
4086 * Spill the event counters of the dead processor
4087 * into the current processors event counters.
4088 * This artificially elevates the count of the current
4089 * processor.
4091 vm_events_fold_cpu(cpu);
4094 * Zero the differential counters of the dead processor
4095 * so that the vm statistics are consistent.
4097 * This is only okay since the processor is dead and cannot
4098 * race with what we are doing.
4100 refresh_cpu_vm_stats(cpu);
4102 return NOTIFY_OK;
4105 void __init page_alloc_init(void)
4107 hotcpu_notifier(page_alloc_cpu_notify, 0);
4111 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
4112 * or min_free_kbytes changes.
4114 static void calculate_totalreserve_pages(void)
4116 struct pglist_data *pgdat;
4117 unsigned long reserve_pages = 0;
4118 enum zone_type i, j;
4120 for_each_online_pgdat(pgdat) {
4121 for (i = 0; i < MAX_NR_ZONES; i++) {
4122 struct zone *zone = pgdat->node_zones + i;
4123 unsigned long max = 0;
4125 /* Find valid and maximum lowmem_reserve in the zone */
4126 for (j = i; j < MAX_NR_ZONES; j++) {
4127 if (zone->lowmem_reserve[j] > max)
4128 max = zone->lowmem_reserve[j];
4131 /* we treat pages_high as reserved pages. */
4132 max += zone->pages_high;
4134 if (max > zone->present_pages)
4135 max = zone->present_pages;
4136 reserve_pages += max;
4139 totalreserve_pages = reserve_pages;
4143 * setup_per_zone_lowmem_reserve - called whenever
4144 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
4145 * has a correct pages reserved value, so an adequate number of
4146 * pages are left in the zone after a successful __alloc_pages().
4148 static void setup_per_zone_lowmem_reserve(void)
4150 struct pglist_data *pgdat;
4151 enum zone_type j, idx;
4153 for_each_online_pgdat(pgdat) {
4154 for (j = 0; j < MAX_NR_ZONES; j++) {
4155 struct zone *zone = pgdat->node_zones + j;
4156 unsigned long present_pages = zone->present_pages;
4158 zone->lowmem_reserve[j] = 0;
4160 idx = j;
4161 while (idx) {
4162 struct zone *lower_zone;
4164 idx--;
4166 if (sysctl_lowmem_reserve_ratio[idx] < 1)
4167 sysctl_lowmem_reserve_ratio[idx] = 1;
4169 lower_zone = pgdat->node_zones + idx;
4170 lower_zone->lowmem_reserve[j] = present_pages /
4171 sysctl_lowmem_reserve_ratio[idx];
4172 present_pages += lower_zone->present_pages;
4177 /* update totalreserve_pages */
4178 calculate_totalreserve_pages();
4182 * setup_per_zone_pages_min - called when min_free_kbytes changes.
4184 * Ensures that the pages_{min,low,high} values for each zone are set correctly
4185 * with respect to min_free_kbytes.
4187 void setup_per_zone_pages_min(void)
4189 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
4190 unsigned long lowmem_pages = 0;
4191 struct zone *zone;
4192 unsigned long flags;
4194 /* Calculate total number of !ZONE_HIGHMEM pages */
4195 for_each_zone(zone) {
4196 if (!is_highmem(zone))
4197 lowmem_pages += zone->present_pages;
4200 for_each_zone(zone) {
4201 u64 tmp;
4203 spin_lock_irqsave(&zone->lru_lock, flags);
4204 tmp = (u64)pages_min * zone->present_pages;
4205 do_div(tmp, lowmem_pages);
4206 if (is_highmem(zone)) {
4208 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
4209 * need highmem pages, so cap pages_min to a small
4210 * value here.
4212 * The (pages_high-pages_low) and (pages_low-pages_min)
4213 * deltas controls asynch page reclaim, and so should
4214 * not be capped for highmem.
4216 int min_pages;
4218 min_pages = zone->present_pages / 1024;
4219 if (min_pages < SWAP_CLUSTER_MAX)
4220 min_pages = SWAP_CLUSTER_MAX;
4221 if (min_pages > 128)
4222 min_pages = 128;
4223 zone->pages_min = min_pages;
4224 } else {
4226 * If it's a lowmem zone, reserve a number of pages
4227 * proportionate to the zone's size.
4229 zone->pages_min = tmp;
4232 zone->pages_low = zone->pages_min + (tmp >> 2);
4233 zone->pages_high = zone->pages_min + (tmp >> 1);
4234 setup_zone_migrate_reserve(zone);
4235 spin_unlock_irqrestore(&zone->lru_lock, flags);
4238 /* update totalreserve_pages */
4239 calculate_totalreserve_pages();
4243 * Initialise min_free_kbytes.
4245 * For small machines we want it small (128k min). For large machines
4246 * we want it large (64MB max). But it is not linear, because network
4247 * bandwidth does not increase linearly with machine size. We use
4249 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
4250 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
4252 * which yields
4254 * 16MB: 512k
4255 * 32MB: 724k
4256 * 64MB: 1024k
4257 * 128MB: 1448k
4258 * 256MB: 2048k
4259 * 512MB: 2896k
4260 * 1024MB: 4096k
4261 * 2048MB: 5792k
4262 * 4096MB: 8192k
4263 * 8192MB: 11584k
4264 * 16384MB: 16384k
4266 static int __init init_per_zone_pages_min(void)
4268 unsigned long lowmem_kbytes;
4270 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
4272 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
4273 if (min_free_kbytes < 128)
4274 min_free_kbytes = 128;
4275 if (min_free_kbytes > 65536)
4276 min_free_kbytes = 65536;
4277 setup_per_zone_pages_min();
4278 setup_per_zone_lowmem_reserve();
4279 return 0;
4281 module_init(init_per_zone_pages_min)
4284 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
4285 * that we can call two helper functions whenever min_free_kbytes
4286 * changes.
4288 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
4289 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4291 proc_dointvec(table, write, file, buffer, length, ppos);
4292 if (write)
4293 setup_per_zone_pages_min();
4294 return 0;
4297 #ifdef CONFIG_NUMA
4298 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
4299 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4301 struct zone *zone;
4302 int rc;
4304 rc = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
4305 if (rc)
4306 return rc;
4308 for_each_zone(zone)
4309 zone->min_unmapped_pages = (zone->present_pages *
4310 sysctl_min_unmapped_ratio) / 100;
4311 return 0;
4314 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
4315 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4317 struct zone *zone;
4318 int rc;
4320 rc = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
4321 if (rc)
4322 return rc;
4324 for_each_zone(zone)
4325 zone->min_slab_pages = (zone->present_pages *
4326 sysctl_min_slab_ratio) / 100;
4327 return 0;
4329 #endif
4332 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
4333 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
4334 * whenever sysctl_lowmem_reserve_ratio changes.
4336 * The reserve ratio obviously has absolutely no relation with the
4337 * pages_min watermarks. The lowmem reserve ratio can only make sense
4338 * if in function of the boot time zone sizes.
4340 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
4341 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4343 proc_dointvec_minmax(table, write, file, buffer, length, ppos);
4344 setup_per_zone_lowmem_reserve();
4345 return 0;
4349 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
4350 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
4351 * can have before it gets flushed back to buddy allocator.
4354 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
4355 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4357 struct zone *zone;
4358 unsigned int cpu;
4359 int ret;
4361 ret = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
4362 if (!write || (ret == -EINVAL))
4363 return ret;
4364 for_each_zone(zone) {
4365 for_each_online_cpu(cpu) {
4366 unsigned long high;
4367 high = zone->present_pages / percpu_pagelist_fraction;
4368 setup_pagelist_highmark(zone_pcp(zone, cpu), high);
4371 return 0;
4374 int hashdist = HASHDIST_DEFAULT;
4376 #ifdef CONFIG_NUMA
4377 static int __init set_hashdist(char *str)
4379 if (!str)
4380 return 0;
4381 hashdist = simple_strtoul(str, &str, 0);
4382 return 1;
4384 __setup("hashdist=", set_hashdist);
4385 #endif
4388 * allocate a large system hash table from bootmem
4389 * - it is assumed that the hash table must contain an exact power-of-2
4390 * quantity of entries
4391 * - limit is the number of hash buckets, not the total allocation size
4393 void *__init alloc_large_system_hash(const char *tablename,
4394 unsigned long bucketsize,
4395 unsigned long numentries,
4396 int scale,
4397 int flags,
4398 unsigned int *_hash_shift,
4399 unsigned int *_hash_mask,
4400 unsigned long limit)
4402 unsigned long long max = limit;
4403 unsigned long log2qty, size;
4404 void *table = NULL;
4406 /* allow the kernel cmdline to have a say */
4407 if (!numentries) {
4408 /* round applicable memory size up to nearest megabyte */
4409 numentries = nr_kernel_pages;
4410 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
4411 numentries >>= 20 - PAGE_SHIFT;
4412 numentries <<= 20 - PAGE_SHIFT;
4414 /* limit to 1 bucket per 2^scale bytes of low memory */
4415 if (scale > PAGE_SHIFT)
4416 numentries >>= (scale - PAGE_SHIFT);
4417 else
4418 numentries <<= (PAGE_SHIFT - scale);
4420 /* Make sure we've got at least a 0-order allocation.. */
4421 if (unlikely((numentries * bucketsize) < PAGE_SIZE))
4422 numentries = PAGE_SIZE / bucketsize;
4424 numentries = roundup_pow_of_two(numentries);
4426 /* limit allocation size to 1/16 total memory by default */
4427 if (max == 0) {
4428 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
4429 do_div(max, bucketsize);
4432 if (numentries > max)
4433 numentries = max;
4435 log2qty = ilog2(numentries);
4437 do {
4438 size = bucketsize << log2qty;
4439 if (flags & HASH_EARLY)
4440 table = alloc_bootmem(size);
4441 else if (hashdist)
4442 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
4443 else {
4444 unsigned long order = get_order(size);
4445 table = (void*) __get_free_pages(GFP_ATOMIC, order);
4447 * If bucketsize is not a power-of-two, we may free
4448 * some pages at the end of hash table.
4450 if (table) {
4451 unsigned long alloc_end = (unsigned long)table +
4452 (PAGE_SIZE << order);
4453 unsigned long used = (unsigned long)table +
4454 PAGE_ALIGN(size);
4455 split_page(virt_to_page(table), order);
4456 while (used < alloc_end) {
4457 free_page(used);
4458 used += PAGE_SIZE;
4462 } while (!table && size > PAGE_SIZE && --log2qty);
4464 if (!table)
4465 panic("Failed to allocate %s hash table\n", tablename);
4467 printk(KERN_INFO "%s hash table entries: %d (order: %d, %lu bytes)\n",
4468 tablename,
4469 (1U << log2qty),
4470 ilog2(size) - PAGE_SHIFT,
4471 size);
4473 if (_hash_shift)
4474 *_hash_shift = log2qty;
4475 if (_hash_mask)
4476 *_hash_mask = (1 << log2qty) - 1;
4478 return table;
4481 #ifdef CONFIG_OUT_OF_LINE_PFN_TO_PAGE
4482 struct page *pfn_to_page(unsigned long pfn)
4484 return __pfn_to_page(pfn);
4486 unsigned long page_to_pfn(struct page *page)
4488 return __page_to_pfn(page);
4490 EXPORT_SYMBOL(pfn_to_page);
4491 EXPORT_SYMBOL(page_to_pfn);
4492 #endif /* CONFIG_OUT_OF_LINE_PFN_TO_PAGE */
4494 /* Return a pointer to the bitmap storing bits affecting a block of pages */
4495 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
4496 unsigned long pfn)
4498 #ifdef CONFIG_SPARSEMEM
4499 return __pfn_to_section(pfn)->pageblock_flags;
4500 #else
4501 return zone->pageblock_flags;
4502 #endif /* CONFIG_SPARSEMEM */
4505 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
4507 #ifdef CONFIG_SPARSEMEM
4508 pfn &= (PAGES_PER_SECTION-1);
4509 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
4510 #else
4511 pfn = pfn - zone->zone_start_pfn;
4512 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
4513 #endif /* CONFIG_SPARSEMEM */
4517 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
4518 * @page: The page within the block of interest
4519 * @start_bitidx: The first bit of interest to retrieve
4520 * @end_bitidx: The last bit of interest
4521 * returns pageblock_bits flags
4523 unsigned long get_pageblock_flags_group(struct page *page,
4524 int start_bitidx, int end_bitidx)
4526 struct zone *zone;
4527 unsigned long *bitmap;
4528 unsigned long pfn, bitidx;
4529 unsigned long flags = 0;
4530 unsigned long value = 1;
4532 zone = page_zone(page);
4533 pfn = page_to_pfn(page);
4534 bitmap = get_pageblock_bitmap(zone, pfn);
4535 bitidx = pfn_to_bitidx(zone, pfn);
4537 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
4538 if (test_bit(bitidx + start_bitidx, bitmap))
4539 flags |= value;
4541 return flags;
4545 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
4546 * @page: The page within the block of interest
4547 * @start_bitidx: The first bit of interest
4548 * @end_bitidx: The last bit of interest
4549 * @flags: The flags to set
4551 void set_pageblock_flags_group(struct page *page, unsigned long flags,
4552 int start_bitidx, int end_bitidx)
4554 struct zone *zone;
4555 unsigned long *bitmap;
4556 unsigned long pfn, bitidx;
4557 unsigned long value = 1;
4559 zone = page_zone(page);
4560 pfn = page_to_pfn(page);
4561 bitmap = get_pageblock_bitmap(zone, pfn);
4562 bitidx = pfn_to_bitidx(zone, pfn);
4563 VM_BUG_ON(pfn < zone->zone_start_pfn);
4564 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages);
4566 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
4567 if (flags & value)
4568 __set_bit(bitidx + start_bitidx, bitmap);
4569 else
4570 __clear_bit(bitidx + start_bitidx, bitmap);
4574 * This is designed as sub function...plz see page_isolation.c also.
4575 * set/clear page block's type to be ISOLATE.
4576 * page allocater never alloc memory from ISOLATE block.
4579 int set_migratetype_isolate(struct page *page)
4581 struct zone *zone;
4582 unsigned long flags;
4583 int ret = -EBUSY;
4585 zone = page_zone(page);
4586 spin_lock_irqsave(&zone->lock, flags);
4588 * In future, more migrate types will be able to be isolation target.
4590 if (get_pageblock_migratetype(page) != MIGRATE_MOVABLE)
4591 goto out;
4592 set_pageblock_migratetype(page, MIGRATE_ISOLATE);
4593 move_freepages_block(zone, page, MIGRATE_ISOLATE);
4594 ret = 0;
4595 out:
4596 spin_unlock_irqrestore(&zone->lock, flags);
4597 if (!ret)
4598 drain_all_pages();
4599 return ret;
4602 void unset_migratetype_isolate(struct page *page)
4604 struct zone *zone;
4605 unsigned long flags;
4606 zone = page_zone(page);
4607 spin_lock_irqsave(&zone->lock, flags);
4608 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
4609 goto out;
4610 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4611 move_freepages_block(zone, page, MIGRATE_MOVABLE);
4612 out:
4613 spin_unlock_irqrestore(&zone->lock, flags);
4616 #ifdef CONFIG_MEMORY_HOTREMOVE
4618 * All pages in the range must be isolated before calling this.
4620 void
4621 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
4623 struct page *page;
4624 struct zone *zone;
4625 int order, i;
4626 unsigned long pfn;
4627 unsigned long flags;
4628 /* find the first valid pfn */
4629 for (pfn = start_pfn; pfn < end_pfn; pfn++)
4630 if (pfn_valid(pfn))
4631 break;
4632 if (pfn == end_pfn)
4633 return;
4634 zone = page_zone(pfn_to_page(pfn));
4635 spin_lock_irqsave(&zone->lock, flags);
4636 pfn = start_pfn;
4637 while (pfn < end_pfn) {
4638 if (!pfn_valid(pfn)) {
4639 pfn++;
4640 continue;
4642 page = pfn_to_page(pfn);
4643 BUG_ON(page_count(page));
4644 BUG_ON(!PageBuddy(page));
4645 order = page_order(page);
4646 #ifdef CONFIG_DEBUG_VM
4647 printk(KERN_INFO "remove from free list %lx %d %lx\n",
4648 pfn, 1 << order, end_pfn);
4649 #endif
4650 list_del(&page->lru);
4651 rmv_page_order(page);
4652 zone->free_area[order].nr_free--;
4653 __mod_zone_page_state(zone, NR_FREE_PAGES,
4654 - (1UL << order));
4655 for (i = 0; i < (1 << order); i++)
4656 SetPageReserved((page+i));
4657 pfn += (1 << order);
4659 spin_unlock_irqrestore(&zone->lock, flags);
4661 #endif