ext4: Replace hackish ext4_mb_poll_new_transaction with commit callback
[linux-2.6/mini2440.git] / mm / page_alloc.c
blob9eb9eb92828510efbc7100addea5fa766c8f48ea
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;
271 struct page *p = page + 1;
273 set_compound_page_dtor(page, free_compound_page);
274 set_compound_order(page, order);
275 __SetPageHead(page);
276 for (i = 1; i < nr_pages; i++, p++) {
277 if (unlikely((i & (MAX_ORDER_NR_PAGES - 1)) == 0))
278 p = pfn_to_page(page_to_pfn(page) + i);
279 __SetPageTail(p);
280 p->first_page = page;
284 static void destroy_compound_page(struct page *page, unsigned long order)
286 int i;
287 int nr_pages = 1 << order;
288 struct page *p = page + 1;
290 if (unlikely(compound_order(page) != order))
291 bad_page(page);
293 if (unlikely(!PageHead(page)))
294 bad_page(page);
295 __ClearPageHead(page);
296 for (i = 1; i < nr_pages; i++, p++) {
297 if (unlikely((i & (MAX_ORDER_NR_PAGES - 1)) == 0))
298 p = pfn_to_page(page_to_pfn(page) + i);
300 if (unlikely(!PageTail(p) |
301 (p->first_page != page)))
302 bad_page(page);
303 __ClearPageTail(p);
307 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
309 int i;
312 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
313 * and __GFP_HIGHMEM from hard or soft interrupt context.
315 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
316 for (i = 0; i < (1 << order); i++)
317 clear_highpage(page + i);
320 static inline void set_page_order(struct page *page, int order)
322 set_page_private(page, order);
323 __SetPageBuddy(page);
326 static inline void rmv_page_order(struct page *page)
328 __ClearPageBuddy(page);
329 set_page_private(page, 0);
333 * Locate the struct page for both the matching buddy in our
334 * pair (buddy1) and the combined O(n+1) page they form (page).
336 * 1) Any buddy B1 will have an order O twin B2 which satisfies
337 * the following equation:
338 * B2 = B1 ^ (1 << O)
339 * For example, if the starting buddy (buddy2) is #8 its order
340 * 1 buddy is #10:
341 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
343 * 2) Any buddy B will have an order O+1 parent P which
344 * satisfies the following equation:
345 * P = B & ~(1 << O)
347 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
349 static inline struct page *
350 __page_find_buddy(struct page *page, unsigned long page_idx, unsigned int order)
352 unsigned long buddy_idx = page_idx ^ (1 << order);
354 return page + (buddy_idx - page_idx);
357 static inline unsigned long
358 __find_combined_index(unsigned long page_idx, unsigned int order)
360 return (page_idx & ~(1 << order));
364 * This function checks whether a page is free && is the buddy
365 * we can do coalesce a page and its buddy if
366 * (a) the buddy is not in a hole &&
367 * (b) the buddy is in the buddy system &&
368 * (c) a page and its buddy have the same order &&
369 * (d) a page and its buddy are in the same zone.
371 * For recording whether a page is in the buddy system, we use PG_buddy.
372 * Setting, clearing, and testing PG_buddy is serialized by zone->lock.
374 * For recording page's order, we use page_private(page).
376 static inline int page_is_buddy(struct page *page, struct page *buddy,
377 int order)
379 if (!pfn_valid_within(page_to_pfn(buddy)))
380 return 0;
382 if (page_zone_id(page) != page_zone_id(buddy))
383 return 0;
385 if (PageBuddy(buddy) && page_order(buddy) == order) {
386 BUG_ON(page_count(buddy) != 0);
387 return 1;
389 return 0;
393 * Freeing function for a buddy system allocator.
395 * The concept of a buddy system is to maintain direct-mapped table
396 * (containing bit values) for memory blocks of various "orders".
397 * The bottom level table contains the map for the smallest allocatable
398 * units of memory (here, pages), and each level above it describes
399 * pairs of units from the levels below, hence, "buddies".
400 * At a high level, all that happens here is marking the table entry
401 * at the bottom level available, and propagating the changes upward
402 * as necessary, plus some accounting needed to play nicely with other
403 * parts of the VM system.
404 * At each level, we keep a list of pages, which are heads of continuous
405 * free pages of length of (1 << order) and marked with PG_buddy. Page's
406 * order is recorded in page_private(page) field.
407 * So when we are allocating or freeing one, we can derive the state of the
408 * other. That is, if we allocate a small block, and both were
409 * free, the remainder of the region must be split into blocks.
410 * If a block is freed, and its buddy is also free, then this
411 * triggers coalescing into a block of larger size.
413 * -- wli
416 static inline void __free_one_page(struct page *page,
417 struct zone *zone, unsigned int order)
419 unsigned long page_idx;
420 int order_size = 1 << order;
421 int migratetype = get_pageblock_migratetype(page);
423 if (unlikely(PageCompound(page)))
424 destroy_compound_page(page, order);
426 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
428 VM_BUG_ON(page_idx & (order_size - 1));
429 VM_BUG_ON(bad_range(zone, page));
431 __mod_zone_page_state(zone, NR_FREE_PAGES, order_size);
432 while (order < MAX_ORDER-1) {
433 unsigned long combined_idx;
434 struct page *buddy;
436 buddy = __page_find_buddy(page, page_idx, order);
437 if (!page_is_buddy(page, buddy, order))
438 break;
440 /* Our buddy is free, merge with it and move up one order. */
441 list_del(&buddy->lru);
442 zone->free_area[order].nr_free--;
443 rmv_page_order(buddy);
444 combined_idx = __find_combined_index(page_idx, order);
445 page = page + (combined_idx - page_idx);
446 page_idx = combined_idx;
447 order++;
449 set_page_order(page, order);
450 list_add(&page->lru,
451 &zone->free_area[order].free_list[migratetype]);
452 zone->free_area[order].nr_free++;
455 static inline int free_pages_check(struct page *page)
457 if (unlikely(page_mapcount(page) |
458 (page->mapping != NULL) |
459 (page_get_page_cgroup(page) != NULL) |
460 (page_count(page) != 0) |
461 (page->flags & PAGE_FLAGS_CHECK_AT_FREE)))
462 bad_page(page);
463 if (PageDirty(page))
464 __ClearPageDirty(page);
466 * For now, we report if PG_reserved was found set, but do not
467 * clear it, and do not free the page. But we shall soon need
468 * to do more, for when the ZERO_PAGE count wraps negative.
470 return PageReserved(page);
474 * Frees a list of pages.
475 * Assumes all pages on list are in same zone, and of same order.
476 * count is the number of pages to free.
478 * If the zone was previously in an "all pages pinned" state then look to
479 * see if this freeing clears that state.
481 * And clear the zone's pages_scanned counter, to hold off the "all pages are
482 * pinned" detection logic.
484 static void free_pages_bulk(struct zone *zone, int count,
485 struct list_head *list, int order)
487 spin_lock(&zone->lock);
488 zone_clear_flag(zone, ZONE_ALL_UNRECLAIMABLE);
489 zone->pages_scanned = 0;
490 while (count--) {
491 struct page *page;
493 VM_BUG_ON(list_empty(list));
494 page = list_entry(list->prev, struct page, lru);
495 /* have to delete it as __free_one_page list manipulates */
496 list_del(&page->lru);
497 __free_one_page(page, zone, order);
499 spin_unlock(&zone->lock);
502 static void free_one_page(struct zone *zone, struct page *page, int order)
504 spin_lock(&zone->lock);
505 zone_clear_flag(zone, ZONE_ALL_UNRECLAIMABLE);
506 zone->pages_scanned = 0;
507 __free_one_page(page, zone, order);
508 spin_unlock(&zone->lock);
511 static void __free_pages_ok(struct page *page, unsigned int order)
513 unsigned long flags;
514 int i;
515 int reserved = 0;
517 for (i = 0 ; i < (1 << order) ; ++i)
518 reserved += free_pages_check(page + i);
519 if (reserved)
520 return;
522 if (!PageHighMem(page)) {
523 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
524 debug_check_no_obj_freed(page_address(page),
525 PAGE_SIZE << order);
527 arch_free_page(page, order);
528 kernel_map_pages(page, 1 << order, 0);
530 local_irq_save(flags);
531 __count_vm_events(PGFREE, 1 << order);
532 free_one_page(page_zone(page), page, order);
533 local_irq_restore(flags);
537 * permit the bootmem allocator to evade page validation on high-order frees
539 void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
541 if (order == 0) {
542 __ClearPageReserved(page);
543 set_page_count(page, 0);
544 set_page_refcounted(page);
545 __free_page(page);
546 } else {
547 int loop;
549 prefetchw(page);
550 for (loop = 0; loop < BITS_PER_LONG; loop++) {
551 struct page *p = &page[loop];
553 if (loop + 1 < BITS_PER_LONG)
554 prefetchw(p + 1);
555 __ClearPageReserved(p);
556 set_page_count(p, 0);
559 set_page_refcounted(page);
560 __free_pages(page, order);
566 * The order of subdivision here is critical for the IO subsystem.
567 * Please do not alter this order without good reasons and regression
568 * testing. Specifically, as large blocks of memory are subdivided,
569 * the order in which smaller blocks are delivered depends on the order
570 * they're subdivided in this function. This is the primary factor
571 * influencing the order in which pages are delivered to the IO
572 * subsystem according to empirical testing, and this is also justified
573 * by considering the behavior of a buddy system containing a single
574 * large block of memory acted on by a series of small allocations.
575 * This behavior is a critical factor in sglist merging's success.
577 * -- wli
579 static inline void expand(struct zone *zone, struct page *page,
580 int low, int high, struct free_area *area,
581 int migratetype)
583 unsigned long size = 1 << high;
585 while (high > low) {
586 area--;
587 high--;
588 size >>= 1;
589 VM_BUG_ON(bad_range(zone, &page[size]));
590 list_add(&page[size].lru, &area->free_list[migratetype]);
591 area->nr_free++;
592 set_page_order(&page[size], high);
597 * This page is about to be returned from the page allocator
599 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
601 if (unlikely(page_mapcount(page) |
602 (page->mapping != NULL) |
603 (page_get_page_cgroup(page) != NULL) |
604 (page_count(page) != 0) |
605 (page->flags & PAGE_FLAGS_CHECK_AT_PREP)))
606 bad_page(page);
609 * For now, we report if PG_reserved was found set, but do not
610 * clear it, and do not allocate the page: as a safety net.
612 if (PageReserved(page))
613 return 1;
615 page->flags &= ~(1 << PG_uptodate | 1 << PG_error | 1 << PG_reclaim |
616 1 << PG_referenced | 1 << PG_arch_1 |
617 1 << PG_owner_priv_1 | 1 << PG_mappedtodisk);
618 set_page_private(page, 0);
619 set_page_refcounted(page);
621 arch_alloc_page(page, order);
622 kernel_map_pages(page, 1 << order, 1);
624 if (gfp_flags & __GFP_ZERO)
625 prep_zero_page(page, order, gfp_flags);
627 if (order && (gfp_flags & __GFP_COMP))
628 prep_compound_page(page, order);
630 return 0;
634 * Go through the free lists for the given migratetype and remove
635 * the smallest available page from the freelists
637 static struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
638 int migratetype)
640 unsigned int current_order;
641 struct free_area * area;
642 struct page *page;
644 /* Find a page of the appropriate size in the preferred list */
645 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
646 area = &(zone->free_area[current_order]);
647 if (list_empty(&area->free_list[migratetype]))
648 continue;
650 page = list_entry(area->free_list[migratetype].next,
651 struct page, lru);
652 list_del(&page->lru);
653 rmv_page_order(page);
654 area->nr_free--;
655 __mod_zone_page_state(zone, NR_FREE_PAGES, - (1UL << order));
656 expand(zone, page, order, current_order, area, migratetype);
657 return page;
660 return NULL;
665 * This array describes the order lists are fallen back to when
666 * the free lists for the desirable migrate type are depleted
668 static int fallbacks[MIGRATE_TYPES][MIGRATE_TYPES-1] = {
669 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
670 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
671 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
672 [MIGRATE_RESERVE] = { MIGRATE_RESERVE, MIGRATE_RESERVE, MIGRATE_RESERVE }, /* Never used */
676 * Move the free pages in a range to the free lists of the requested type.
677 * Note that start_page and end_pages are not aligned on a pageblock
678 * boundary. If alignment is required, use move_freepages_block()
680 static int move_freepages(struct zone *zone,
681 struct page *start_page, struct page *end_page,
682 int migratetype)
684 struct page *page;
685 unsigned long order;
686 int pages_moved = 0;
688 #ifndef CONFIG_HOLES_IN_ZONE
690 * page_zone is not safe to call in this context when
691 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
692 * anyway as we check zone boundaries in move_freepages_block().
693 * Remove at a later date when no bug reports exist related to
694 * grouping pages by mobility
696 BUG_ON(page_zone(start_page) != page_zone(end_page));
697 #endif
699 for (page = start_page; page <= end_page;) {
700 /* Make sure we are not inadvertently changing nodes */
701 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
703 if (!pfn_valid_within(page_to_pfn(page))) {
704 page++;
705 continue;
708 if (!PageBuddy(page)) {
709 page++;
710 continue;
713 order = page_order(page);
714 list_del(&page->lru);
715 list_add(&page->lru,
716 &zone->free_area[order].free_list[migratetype]);
717 page += 1 << order;
718 pages_moved += 1 << order;
721 return pages_moved;
724 static int move_freepages_block(struct zone *zone, struct page *page,
725 int migratetype)
727 unsigned long start_pfn, end_pfn;
728 struct page *start_page, *end_page;
730 start_pfn = page_to_pfn(page);
731 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
732 start_page = pfn_to_page(start_pfn);
733 end_page = start_page + pageblock_nr_pages - 1;
734 end_pfn = start_pfn + pageblock_nr_pages - 1;
736 /* Do not cross zone boundaries */
737 if (start_pfn < zone->zone_start_pfn)
738 start_page = page;
739 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
740 return 0;
742 return move_freepages(zone, start_page, end_page, migratetype);
745 /* Remove an element from the buddy allocator from the fallback list */
746 static struct page *__rmqueue_fallback(struct zone *zone, int order,
747 int start_migratetype)
749 struct free_area * area;
750 int current_order;
751 struct page *page;
752 int migratetype, i;
754 /* Find the largest possible block of pages in the other list */
755 for (current_order = MAX_ORDER-1; current_order >= order;
756 --current_order) {
757 for (i = 0; i < MIGRATE_TYPES - 1; i++) {
758 migratetype = fallbacks[start_migratetype][i];
760 /* MIGRATE_RESERVE handled later if necessary */
761 if (migratetype == MIGRATE_RESERVE)
762 continue;
764 area = &(zone->free_area[current_order]);
765 if (list_empty(&area->free_list[migratetype]))
766 continue;
768 page = list_entry(area->free_list[migratetype].next,
769 struct page, lru);
770 area->nr_free--;
773 * If breaking a large block of pages, move all free
774 * pages to the preferred allocation list. If falling
775 * back for a reclaimable kernel allocation, be more
776 * agressive about taking ownership of free pages
778 if (unlikely(current_order >= (pageblock_order >> 1)) ||
779 start_migratetype == MIGRATE_RECLAIMABLE) {
780 unsigned long pages;
781 pages = move_freepages_block(zone, page,
782 start_migratetype);
784 /* Claim the whole block if over half of it is free */
785 if (pages >= (1 << (pageblock_order-1)))
786 set_pageblock_migratetype(page,
787 start_migratetype);
789 migratetype = start_migratetype;
792 /* Remove the page from the freelists */
793 list_del(&page->lru);
794 rmv_page_order(page);
795 __mod_zone_page_state(zone, NR_FREE_PAGES,
796 -(1UL << order));
798 if (current_order == pageblock_order)
799 set_pageblock_migratetype(page,
800 start_migratetype);
802 expand(zone, page, order, current_order, area, migratetype);
803 return page;
807 /* Use MIGRATE_RESERVE rather than fail an allocation */
808 return __rmqueue_smallest(zone, order, MIGRATE_RESERVE);
812 * Do the hard work of removing an element from the buddy allocator.
813 * Call me with the zone->lock already held.
815 static struct page *__rmqueue(struct zone *zone, unsigned int order,
816 int migratetype)
818 struct page *page;
820 page = __rmqueue_smallest(zone, order, migratetype);
822 if (unlikely(!page))
823 page = __rmqueue_fallback(zone, order, migratetype);
825 return page;
829 * Obtain a specified number of elements from the buddy allocator, all under
830 * a single hold of the lock, for efficiency. Add them to the supplied list.
831 * Returns the number of new pages which were placed at *list.
833 static int rmqueue_bulk(struct zone *zone, unsigned int order,
834 unsigned long count, struct list_head *list,
835 int migratetype)
837 int i;
839 spin_lock(&zone->lock);
840 for (i = 0; i < count; ++i) {
841 struct page *page = __rmqueue(zone, order, migratetype);
842 if (unlikely(page == NULL))
843 break;
846 * Split buddy pages returned by expand() are received here
847 * in physical page order. The page is added to the callers and
848 * list and the list head then moves forward. From the callers
849 * perspective, the linked list is ordered by page number in
850 * some conditions. This is useful for IO devices that can
851 * merge IO requests if the physical pages are ordered
852 * properly.
854 list_add(&page->lru, list);
855 set_page_private(page, migratetype);
856 list = &page->lru;
858 spin_unlock(&zone->lock);
859 return i;
862 #ifdef CONFIG_NUMA
864 * Called from the vmstat counter updater to drain pagesets of this
865 * currently executing processor on remote nodes after they have
866 * expired.
868 * Note that this function must be called with the thread pinned to
869 * a single processor.
871 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
873 unsigned long flags;
874 int to_drain;
876 local_irq_save(flags);
877 if (pcp->count >= pcp->batch)
878 to_drain = pcp->batch;
879 else
880 to_drain = pcp->count;
881 free_pages_bulk(zone, to_drain, &pcp->list, 0);
882 pcp->count -= to_drain;
883 local_irq_restore(flags);
885 #endif
888 * Drain pages of the indicated processor.
890 * The processor must either be the current processor and the
891 * thread pinned to the current processor or a processor that
892 * is not online.
894 static void drain_pages(unsigned int cpu)
896 unsigned long flags;
897 struct zone *zone;
899 for_each_zone(zone) {
900 struct per_cpu_pageset *pset;
901 struct per_cpu_pages *pcp;
903 if (!populated_zone(zone))
904 continue;
906 pset = zone_pcp(zone, cpu);
908 pcp = &pset->pcp;
909 local_irq_save(flags);
910 free_pages_bulk(zone, pcp->count, &pcp->list, 0);
911 pcp->count = 0;
912 local_irq_restore(flags);
917 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
919 void drain_local_pages(void *arg)
921 drain_pages(smp_processor_id());
925 * Spill all the per-cpu pages from all CPUs back into the buddy allocator
927 void drain_all_pages(void)
929 on_each_cpu(drain_local_pages, NULL, 1);
932 #ifdef CONFIG_HIBERNATION
934 void mark_free_pages(struct zone *zone)
936 unsigned long pfn, max_zone_pfn;
937 unsigned long flags;
938 int order, t;
939 struct list_head *curr;
941 if (!zone->spanned_pages)
942 return;
944 spin_lock_irqsave(&zone->lock, flags);
946 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
947 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
948 if (pfn_valid(pfn)) {
949 struct page *page = pfn_to_page(pfn);
951 if (!swsusp_page_is_forbidden(page))
952 swsusp_unset_page_free(page);
955 for_each_migratetype_order(order, t) {
956 list_for_each(curr, &zone->free_area[order].free_list[t]) {
957 unsigned long i;
959 pfn = page_to_pfn(list_entry(curr, struct page, lru));
960 for (i = 0; i < (1UL << order); i++)
961 swsusp_set_page_free(pfn_to_page(pfn + i));
964 spin_unlock_irqrestore(&zone->lock, flags);
966 #endif /* CONFIG_PM */
969 * Free a 0-order page
971 static void free_hot_cold_page(struct page *page, int cold)
973 struct zone *zone = page_zone(page);
974 struct per_cpu_pages *pcp;
975 unsigned long flags;
977 if (PageAnon(page))
978 page->mapping = NULL;
979 if (free_pages_check(page))
980 return;
982 if (!PageHighMem(page)) {
983 debug_check_no_locks_freed(page_address(page), PAGE_SIZE);
984 debug_check_no_obj_freed(page_address(page), PAGE_SIZE);
986 arch_free_page(page, 0);
987 kernel_map_pages(page, 1, 0);
989 pcp = &zone_pcp(zone, get_cpu())->pcp;
990 local_irq_save(flags);
991 __count_vm_event(PGFREE);
992 if (cold)
993 list_add_tail(&page->lru, &pcp->list);
994 else
995 list_add(&page->lru, &pcp->list);
996 set_page_private(page, get_pageblock_migratetype(page));
997 pcp->count++;
998 if (pcp->count >= pcp->high) {
999 free_pages_bulk(zone, pcp->batch, &pcp->list, 0);
1000 pcp->count -= pcp->batch;
1002 local_irq_restore(flags);
1003 put_cpu();
1006 void free_hot_page(struct page *page)
1008 free_hot_cold_page(page, 0);
1011 void free_cold_page(struct page *page)
1013 free_hot_cold_page(page, 1);
1017 * split_page takes a non-compound higher-order page, and splits it into
1018 * n (1<<order) sub-pages: page[0..n]
1019 * Each sub-page must be freed individually.
1021 * Note: this is probably too low level an operation for use in drivers.
1022 * Please consult with lkml before using this in your driver.
1024 void split_page(struct page *page, unsigned int order)
1026 int i;
1028 VM_BUG_ON(PageCompound(page));
1029 VM_BUG_ON(!page_count(page));
1030 for (i = 1; i < (1 << order); i++)
1031 set_page_refcounted(page + i);
1035 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1036 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1037 * or two.
1039 static struct page *buffered_rmqueue(struct zone *preferred_zone,
1040 struct zone *zone, int order, gfp_t gfp_flags)
1042 unsigned long flags;
1043 struct page *page;
1044 int cold = !!(gfp_flags & __GFP_COLD);
1045 int cpu;
1046 int migratetype = allocflags_to_migratetype(gfp_flags);
1048 again:
1049 cpu = get_cpu();
1050 if (likely(order == 0)) {
1051 struct per_cpu_pages *pcp;
1053 pcp = &zone_pcp(zone, cpu)->pcp;
1054 local_irq_save(flags);
1055 if (!pcp->count) {
1056 pcp->count = rmqueue_bulk(zone, 0,
1057 pcp->batch, &pcp->list, migratetype);
1058 if (unlikely(!pcp->count))
1059 goto failed;
1062 /* Find a page of the appropriate migrate type */
1063 if (cold) {
1064 list_for_each_entry_reverse(page, &pcp->list, lru)
1065 if (page_private(page) == migratetype)
1066 break;
1067 } else {
1068 list_for_each_entry(page, &pcp->list, lru)
1069 if (page_private(page) == migratetype)
1070 break;
1073 /* Allocate more to the pcp list if necessary */
1074 if (unlikely(&page->lru == &pcp->list)) {
1075 pcp->count += rmqueue_bulk(zone, 0,
1076 pcp->batch, &pcp->list, migratetype);
1077 page = list_entry(pcp->list.next, struct page, lru);
1080 list_del(&page->lru);
1081 pcp->count--;
1082 } else {
1083 spin_lock_irqsave(&zone->lock, flags);
1084 page = __rmqueue(zone, order, migratetype);
1085 spin_unlock(&zone->lock);
1086 if (!page)
1087 goto failed;
1090 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1091 zone_statistics(preferred_zone, zone);
1092 local_irq_restore(flags);
1093 put_cpu();
1095 VM_BUG_ON(bad_range(zone, page));
1096 if (prep_new_page(page, order, gfp_flags))
1097 goto again;
1098 return page;
1100 failed:
1101 local_irq_restore(flags);
1102 put_cpu();
1103 return NULL;
1106 #define ALLOC_NO_WATERMARKS 0x01 /* don't check watermarks at all */
1107 #define ALLOC_WMARK_MIN 0x02 /* use pages_min watermark */
1108 #define ALLOC_WMARK_LOW 0x04 /* use pages_low watermark */
1109 #define ALLOC_WMARK_HIGH 0x08 /* use pages_high watermark */
1110 #define ALLOC_HARDER 0x10 /* try to alloc harder */
1111 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
1112 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
1114 #ifdef CONFIG_FAIL_PAGE_ALLOC
1116 static struct fail_page_alloc_attr {
1117 struct fault_attr attr;
1119 u32 ignore_gfp_highmem;
1120 u32 ignore_gfp_wait;
1121 u32 min_order;
1123 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1125 struct dentry *ignore_gfp_highmem_file;
1126 struct dentry *ignore_gfp_wait_file;
1127 struct dentry *min_order_file;
1129 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1131 } fail_page_alloc = {
1132 .attr = FAULT_ATTR_INITIALIZER,
1133 .ignore_gfp_wait = 1,
1134 .ignore_gfp_highmem = 1,
1135 .min_order = 1,
1138 static int __init setup_fail_page_alloc(char *str)
1140 return setup_fault_attr(&fail_page_alloc.attr, str);
1142 __setup("fail_page_alloc=", setup_fail_page_alloc);
1144 static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1146 if (order < fail_page_alloc.min_order)
1147 return 0;
1148 if (gfp_mask & __GFP_NOFAIL)
1149 return 0;
1150 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1151 return 0;
1152 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1153 return 0;
1155 return should_fail(&fail_page_alloc.attr, 1 << order);
1158 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1160 static int __init fail_page_alloc_debugfs(void)
1162 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1163 struct dentry *dir;
1164 int err;
1166 err = init_fault_attr_dentries(&fail_page_alloc.attr,
1167 "fail_page_alloc");
1168 if (err)
1169 return err;
1170 dir = fail_page_alloc.attr.dentries.dir;
1172 fail_page_alloc.ignore_gfp_wait_file =
1173 debugfs_create_bool("ignore-gfp-wait", mode, dir,
1174 &fail_page_alloc.ignore_gfp_wait);
1176 fail_page_alloc.ignore_gfp_highmem_file =
1177 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1178 &fail_page_alloc.ignore_gfp_highmem);
1179 fail_page_alloc.min_order_file =
1180 debugfs_create_u32("min-order", mode, dir,
1181 &fail_page_alloc.min_order);
1183 if (!fail_page_alloc.ignore_gfp_wait_file ||
1184 !fail_page_alloc.ignore_gfp_highmem_file ||
1185 !fail_page_alloc.min_order_file) {
1186 err = -ENOMEM;
1187 debugfs_remove(fail_page_alloc.ignore_gfp_wait_file);
1188 debugfs_remove(fail_page_alloc.ignore_gfp_highmem_file);
1189 debugfs_remove(fail_page_alloc.min_order_file);
1190 cleanup_fault_attr_dentries(&fail_page_alloc.attr);
1193 return err;
1196 late_initcall(fail_page_alloc_debugfs);
1198 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1200 #else /* CONFIG_FAIL_PAGE_ALLOC */
1202 static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1204 return 0;
1207 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1210 * Return 1 if free pages are above 'mark'. This takes into account the order
1211 * of the allocation.
1213 int zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1214 int classzone_idx, int alloc_flags)
1216 /* free_pages my go negative - that's OK */
1217 long min = mark;
1218 long free_pages = zone_page_state(z, NR_FREE_PAGES) - (1 << order) + 1;
1219 int o;
1221 if (alloc_flags & ALLOC_HIGH)
1222 min -= min / 2;
1223 if (alloc_flags & ALLOC_HARDER)
1224 min -= min / 4;
1226 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
1227 return 0;
1228 for (o = 0; o < order; o++) {
1229 /* At the next order, this order's pages become unavailable */
1230 free_pages -= z->free_area[o].nr_free << o;
1232 /* Require fewer higher order pages to be free */
1233 min >>= 1;
1235 if (free_pages <= min)
1236 return 0;
1238 return 1;
1241 #ifdef CONFIG_NUMA
1243 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1244 * skip over zones that are not allowed by the cpuset, or that have
1245 * been recently (in last second) found to be nearly full. See further
1246 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1247 * that have to skip over a lot of full or unallowed zones.
1249 * If the zonelist cache is present in the passed in zonelist, then
1250 * returns a pointer to the allowed node mask (either the current
1251 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1253 * If the zonelist cache is not available for this zonelist, does
1254 * nothing and returns NULL.
1256 * If the fullzones BITMAP in the zonelist cache is stale (more than
1257 * a second since last zap'd) then we zap it out (clear its bits.)
1259 * We hold off even calling zlc_setup, until after we've checked the
1260 * first zone in the zonelist, on the theory that most allocations will
1261 * be satisfied from that first zone, so best to examine that zone as
1262 * quickly as we can.
1264 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1266 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1267 nodemask_t *allowednodes; /* zonelist_cache approximation */
1269 zlc = zonelist->zlcache_ptr;
1270 if (!zlc)
1271 return NULL;
1273 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1274 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1275 zlc->last_full_zap = jiffies;
1278 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1279 &cpuset_current_mems_allowed :
1280 &node_states[N_HIGH_MEMORY];
1281 return allowednodes;
1285 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1286 * if it is worth looking at further for free memory:
1287 * 1) Check that the zone isn't thought to be full (doesn't have its
1288 * bit set in the zonelist_cache fullzones BITMAP).
1289 * 2) Check that the zones node (obtained from the zonelist_cache
1290 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1291 * Return true (non-zero) if zone is worth looking at further, or
1292 * else return false (zero) if it is not.
1294 * This check -ignores- the distinction between various watermarks,
1295 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1296 * found to be full for any variation of these watermarks, it will
1297 * be considered full for up to one second by all requests, unless
1298 * we are so low on memory on all allowed nodes that we are forced
1299 * into the second scan of the zonelist.
1301 * In the second scan we ignore this zonelist cache and exactly
1302 * apply the watermarks to all zones, even it is slower to do so.
1303 * We are low on memory in the second scan, and should leave no stone
1304 * unturned looking for a free page.
1306 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1307 nodemask_t *allowednodes)
1309 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1310 int i; /* index of *z in zonelist zones */
1311 int n; /* node that zone *z is on */
1313 zlc = zonelist->zlcache_ptr;
1314 if (!zlc)
1315 return 1;
1317 i = z - zonelist->_zonerefs;
1318 n = zlc->z_to_n[i];
1320 /* This zone is worth trying if it is allowed but not full */
1321 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1325 * Given 'z' scanning a zonelist, set the corresponding bit in
1326 * zlc->fullzones, so that subsequent attempts to allocate a page
1327 * from that zone don't waste time re-examining it.
1329 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1331 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1332 int i; /* index of *z in zonelist zones */
1334 zlc = zonelist->zlcache_ptr;
1335 if (!zlc)
1336 return;
1338 i = z - zonelist->_zonerefs;
1340 set_bit(i, zlc->fullzones);
1343 #else /* CONFIG_NUMA */
1345 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1347 return NULL;
1350 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1351 nodemask_t *allowednodes)
1353 return 1;
1356 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1359 #endif /* CONFIG_NUMA */
1362 * get_page_from_freelist goes through the zonelist trying to allocate
1363 * a page.
1365 static struct page *
1366 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1367 struct zonelist *zonelist, int high_zoneidx, int alloc_flags)
1369 struct zoneref *z;
1370 struct page *page = NULL;
1371 int classzone_idx;
1372 struct zone *zone, *preferred_zone;
1373 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1374 int zlc_active = 0; /* set if using zonelist_cache */
1375 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1377 (void)first_zones_zonelist(zonelist, high_zoneidx, nodemask,
1378 &preferred_zone);
1379 if (!preferred_zone)
1380 return NULL;
1382 classzone_idx = zone_idx(preferred_zone);
1384 zonelist_scan:
1386 * Scan zonelist, looking for a zone with enough free.
1387 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1389 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1390 high_zoneidx, nodemask) {
1391 if (NUMA_BUILD && zlc_active &&
1392 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1393 continue;
1394 if ((alloc_flags & ALLOC_CPUSET) &&
1395 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1396 goto try_next_zone;
1398 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1399 unsigned long mark;
1400 if (alloc_flags & ALLOC_WMARK_MIN)
1401 mark = zone->pages_min;
1402 else if (alloc_flags & ALLOC_WMARK_LOW)
1403 mark = zone->pages_low;
1404 else
1405 mark = zone->pages_high;
1406 if (!zone_watermark_ok(zone, order, mark,
1407 classzone_idx, alloc_flags)) {
1408 if (!zone_reclaim_mode ||
1409 !zone_reclaim(zone, gfp_mask, order))
1410 goto this_zone_full;
1414 page = buffered_rmqueue(preferred_zone, zone, order, gfp_mask);
1415 if (page)
1416 break;
1417 this_zone_full:
1418 if (NUMA_BUILD)
1419 zlc_mark_zone_full(zonelist, z);
1420 try_next_zone:
1421 if (NUMA_BUILD && !did_zlc_setup) {
1422 /* we do zlc_setup after the first zone is tried */
1423 allowednodes = zlc_setup(zonelist, alloc_flags);
1424 zlc_active = 1;
1425 did_zlc_setup = 1;
1429 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) {
1430 /* Disable zlc cache for second zonelist scan */
1431 zlc_active = 0;
1432 goto zonelist_scan;
1434 return page;
1438 * This is the 'heart' of the zoned buddy allocator.
1440 struct page *
1441 __alloc_pages_internal(gfp_t gfp_mask, unsigned int order,
1442 struct zonelist *zonelist, nodemask_t *nodemask)
1444 const gfp_t wait = gfp_mask & __GFP_WAIT;
1445 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
1446 struct zoneref *z;
1447 struct zone *zone;
1448 struct page *page;
1449 struct reclaim_state reclaim_state;
1450 struct task_struct *p = current;
1451 int do_retry;
1452 int alloc_flags;
1453 unsigned long did_some_progress;
1454 unsigned long pages_reclaimed = 0;
1456 might_sleep_if(wait);
1458 if (should_fail_alloc_page(gfp_mask, order))
1459 return NULL;
1461 restart:
1462 z = zonelist->_zonerefs; /* the list of zones suitable for gfp_mask */
1464 if (unlikely(!z->zone)) {
1466 * Happens if we have an empty zonelist as a result of
1467 * GFP_THISNODE being used on a memoryless node
1469 return NULL;
1472 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
1473 zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET);
1474 if (page)
1475 goto got_pg;
1478 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
1479 * __GFP_NOWARN set) should not cause reclaim since the subsystem
1480 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
1481 * using a larger set of nodes after it has established that the
1482 * allowed per node queues are empty and that nodes are
1483 * over allocated.
1485 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
1486 goto nopage;
1488 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
1489 wakeup_kswapd(zone, order);
1492 * OK, we're below the kswapd watermark and have kicked background
1493 * reclaim. Now things get more complex, so set up alloc_flags according
1494 * to how we want to proceed.
1496 * The caller may dip into page reserves a bit more if the caller
1497 * cannot run direct reclaim, or if the caller has realtime scheduling
1498 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
1499 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
1501 alloc_flags = ALLOC_WMARK_MIN;
1502 if ((unlikely(rt_task(p)) && !in_interrupt()) || !wait)
1503 alloc_flags |= ALLOC_HARDER;
1504 if (gfp_mask & __GFP_HIGH)
1505 alloc_flags |= ALLOC_HIGH;
1506 if (wait)
1507 alloc_flags |= ALLOC_CPUSET;
1510 * Go through the zonelist again. Let __GFP_HIGH and allocations
1511 * coming from realtime tasks go deeper into reserves.
1513 * This is the last chance, in general, before the goto nopage.
1514 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
1515 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1517 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
1518 high_zoneidx, alloc_flags);
1519 if (page)
1520 goto got_pg;
1522 /* This allocation should allow future memory freeing. */
1524 rebalance:
1525 if (((p->flags & PF_MEMALLOC) || unlikely(test_thread_flag(TIF_MEMDIE)))
1526 && !in_interrupt()) {
1527 if (!(gfp_mask & __GFP_NOMEMALLOC)) {
1528 nofail_alloc:
1529 /* go through the zonelist yet again, ignoring mins */
1530 page = get_page_from_freelist(gfp_mask, nodemask, order,
1531 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS);
1532 if (page)
1533 goto got_pg;
1534 if (gfp_mask & __GFP_NOFAIL) {
1535 congestion_wait(WRITE, HZ/50);
1536 goto nofail_alloc;
1539 goto nopage;
1542 /* Atomic allocations - we can't balance anything */
1543 if (!wait)
1544 goto nopage;
1546 cond_resched();
1548 /* We now go into synchronous reclaim */
1549 cpuset_memory_pressure_bump();
1550 p->flags |= PF_MEMALLOC;
1551 reclaim_state.reclaimed_slab = 0;
1552 p->reclaim_state = &reclaim_state;
1554 did_some_progress = try_to_free_pages(zonelist, order, gfp_mask);
1556 p->reclaim_state = NULL;
1557 p->flags &= ~PF_MEMALLOC;
1559 cond_resched();
1561 if (order != 0)
1562 drain_all_pages();
1564 if (likely(did_some_progress)) {
1565 page = get_page_from_freelist(gfp_mask, nodemask, order,
1566 zonelist, high_zoneidx, alloc_flags);
1567 if (page)
1568 goto got_pg;
1569 } else if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
1570 if (!try_set_zone_oom(zonelist, gfp_mask)) {
1571 schedule_timeout_uninterruptible(1);
1572 goto restart;
1576 * Go through the zonelist yet one more time, keep
1577 * very high watermark here, this is only to catch
1578 * a parallel oom killing, we must fail if we're still
1579 * under heavy pressure.
1581 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
1582 order, zonelist, high_zoneidx,
1583 ALLOC_WMARK_HIGH|ALLOC_CPUSET);
1584 if (page) {
1585 clear_zonelist_oom(zonelist, gfp_mask);
1586 goto got_pg;
1589 /* The OOM killer will not help higher order allocs so fail */
1590 if (order > PAGE_ALLOC_COSTLY_ORDER) {
1591 clear_zonelist_oom(zonelist, gfp_mask);
1592 goto nopage;
1595 out_of_memory(zonelist, gfp_mask, order);
1596 clear_zonelist_oom(zonelist, gfp_mask);
1597 goto restart;
1601 * Don't let big-order allocations loop unless the caller explicitly
1602 * requests that. Wait for some write requests to complete then retry.
1604 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
1605 * means __GFP_NOFAIL, but that may not be true in other
1606 * implementations.
1608 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
1609 * specified, then we retry until we no longer reclaim any pages
1610 * (above), or we've reclaimed an order of pages at least as
1611 * large as the allocation's order. In both cases, if the
1612 * allocation still fails, we stop retrying.
1614 pages_reclaimed += did_some_progress;
1615 do_retry = 0;
1616 if (!(gfp_mask & __GFP_NORETRY)) {
1617 if (order <= PAGE_ALLOC_COSTLY_ORDER) {
1618 do_retry = 1;
1619 } else {
1620 if (gfp_mask & __GFP_REPEAT &&
1621 pages_reclaimed < (1 << order))
1622 do_retry = 1;
1624 if (gfp_mask & __GFP_NOFAIL)
1625 do_retry = 1;
1627 if (do_retry) {
1628 congestion_wait(WRITE, HZ/50);
1629 goto rebalance;
1632 nopage:
1633 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
1634 printk(KERN_WARNING "%s: page allocation failure."
1635 " order:%d, mode:0x%x\n",
1636 p->comm, order, gfp_mask);
1637 dump_stack();
1638 show_mem();
1640 got_pg:
1641 return page;
1643 EXPORT_SYMBOL(__alloc_pages_internal);
1646 * Common helper functions.
1648 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
1650 struct page * page;
1651 page = alloc_pages(gfp_mask, order);
1652 if (!page)
1653 return 0;
1654 return (unsigned long) page_address(page);
1657 EXPORT_SYMBOL(__get_free_pages);
1659 unsigned long get_zeroed_page(gfp_t gfp_mask)
1661 struct page * page;
1664 * get_zeroed_page() returns a 32-bit address, which cannot represent
1665 * a highmem page
1667 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
1669 page = alloc_pages(gfp_mask | __GFP_ZERO, 0);
1670 if (page)
1671 return (unsigned long) page_address(page);
1672 return 0;
1675 EXPORT_SYMBOL(get_zeroed_page);
1677 void __pagevec_free(struct pagevec *pvec)
1679 int i = pagevec_count(pvec);
1681 while (--i >= 0)
1682 free_hot_cold_page(pvec->pages[i], pvec->cold);
1685 void __free_pages(struct page *page, unsigned int order)
1687 if (put_page_testzero(page)) {
1688 if (order == 0)
1689 free_hot_page(page);
1690 else
1691 __free_pages_ok(page, order);
1695 EXPORT_SYMBOL(__free_pages);
1697 void free_pages(unsigned long addr, unsigned int order)
1699 if (addr != 0) {
1700 VM_BUG_ON(!virt_addr_valid((void *)addr));
1701 __free_pages(virt_to_page((void *)addr), order);
1705 EXPORT_SYMBOL(free_pages);
1708 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
1709 * @size: the number of bytes to allocate
1710 * @gfp_mask: GFP flags for the allocation
1712 * This function is similar to alloc_pages(), except that it allocates the
1713 * minimum number of pages to satisfy the request. alloc_pages() can only
1714 * allocate memory in power-of-two pages.
1716 * This function is also limited by MAX_ORDER.
1718 * Memory allocated by this function must be released by free_pages_exact().
1720 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
1722 unsigned int order = get_order(size);
1723 unsigned long addr;
1725 addr = __get_free_pages(gfp_mask, order);
1726 if (addr) {
1727 unsigned long alloc_end = addr + (PAGE_SIZE << order);
1728 unsigned long used = addr + PAGE_ALIGN(size);
1730 split_page(virt_to_page(addr), order);
1731 while (used < alloc_end) {
1732 free_page(used);
1733 used += PAGE_SIZE;
1737 return (void *)addr;
1739 EXPORT_SYMBOL(alloc_pages_exact);
1742 * free_pages_exact - release memory allocated via alloc_pages_exact()
1743 * @virt: the value returned by alloc_pages_exact.
1744 * @size: size of allocation, same value as passed to alloc_pages_exact().
1746 * Release the memory allocated by a previous call to alloc_pages_exact.
1748 void free_pages_exact(void *virt, size_t size)
1750 unsigned long addr = (unsigned long)virt;
1751 unsigned long end = addr + PAGE_ALIGN(size);
1753 while (addr < end) {
1754 free_page(addr);
1755 addr += PAGE_SIZE;
1758 EXPORT_SYMBOL(free_pages_exact);
1760 static unsigned int nr_free_zone_pages(int offset)
1762 struct zoneref *z;
1763 struct zone *zone;
1765 /* Just pick one node, since fallback list is circular */
1766 unsigned int sum = 0;
1768 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
1770 for_each_zone_zonelist(zone, z, zonelist, offset) {
1771 unsigned long size = zone->present_pages;
1772 unsigned long high = zone->pages_high;
1773 if (size > high)
1774 sum += size - high;
1777 return sum;
1781 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
1783 unsigned int nr_free_buffer_pages(void)
1785 return nr_free_zone_pages(gfp_zone(GFP_USER));
1787 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
1790 * Amount of free RAM allocatable within all zones
1792 unsigned int nr_free_pagecache_pages(void)
1794 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
1797 static inline void show_node(struct zone *zone)
1799 if (NUMA_BUILD)
1800 printk("Node %d ", zone_to_nid(zone));
1803 void si_meminfo(struct sysinfo *val)
1805 val->totalram = totalram_pages;
1806 val->sharedram = 0;
1807 val->freeram = global_page_state(NR_FREE_PAGES);
1808 val->bufferram = nr_blockdev_pages();
1809 val->totalhigh = totalhigh_pages;
1810 val->freehigh = nr_free_highpages();
1811 val->mem_unit = PAGE_SIZE;
1814 EXPORT_SYMBOL(si_meminfo);
1816 #ifdef CONFIG_NUMA
1817 void si_meminfo_node(struct sysinfo *val, int nid)
1819 pg_data_t *pgdat = NODE_DATA(nid);
1821 val->totalram = pgdat->node_present_pages;
1822 val->freeram = node_page_state(nid, NR_FREE_PAGES);
1823 #ifdef CONFIG_HIGHMEM
1824 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
1825 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
1826 NR_FREE_PAGES);
1827 #else
1828 val->totalhigh = 0;
1829 val->freehigh = 0;
1830 #endif
1831 val->mem_unit = PAGE_SIZE;
1833 #endif
1835 #define K(x) ((x) << (PAGE_SHIFT-10))
1838 * Show free area list (used inside shift_scroll-lock stuff)
1839 * We also calculate the percentage fragmentation. We do this by counting the
1840 * memory on each free list with the exception of the first item on the list.
1842 void show_free_areas(void)
1844 int cpu;
1845 struct zone *zone;
1847 for_each_zone(zone) {
1848 if (!populated_zone(zone))
1849 continue;
1851 show_node(zone);
1852 printk("%s per-cpu:\n", zone->name);
1854 for_each_online_cpu(cpu) {
1855 struct per_cpu_pageset *pageset;
1857 pageset = zone_pcp(zone, cpu);
1859 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
1860 cpu, pageset->pcp.high,
1861 pageset->pcp.batch, pageset->pcp.count);
1865 printk("Active:%lu inactive:%lu dirty:%lu writeback:%lu unstable:%lu\n"
1866 " free:%lu slab:%lu mapped:%lu pagetables:%lu bounce:%lu\n",
1867 global_page_state(NR_ACTIVE),
1868 global_page_state(NR_INACTIVE),
1869 global_page_state(NR_FILE_DIRTY),
1870 global_page_state(NR_WRITEBACK),
1871 global_page_state(NR_UNSTABLE_NFS),
1872 global_page_state(NR_FREE_PAGES),
1873 global_page_state(NR_SLAB_RECLAIMABLE) +
1874 global_page_state(NR_SLAB_UNRECLAIMABLE),
1875 global_page_state(NR_FILE_MAPPED),
1876 global_page_state(NR_PAGETABLE),
1877 global_page_state(NR_BOUNCE));
1879 for_each_zone(zone) {
1880 int i;
1882 if (!populated_zone(zone))
1883 continue;
1885 show_node(zone);
1886 printk("%s"
1887 " free:%lukB"
1888 " min:%lukB"
1889 " low:%lukB"
1890 " high:%lukB"
1891 " active:%lukB"
1892 " inactive:%lukB"
1893 " present:%lukB"
1894 " pages_scanned:%lu"
1895 " all_unreclaimable? %s"
1896 "\n",
1897 zone->name,
1898 K(zone_page_state(zone, NR_FREE_PAGES)),
1899 K(zone->pages_min),
1900 K(zone->pages_low),
1901 K(zone->pages_high),
1902 K(zone_page_state(zone, NR_ACTIVE)),
1903 K(zone_page_state(zone, NR_INACTIVE)),
1904 K(zone->present_pages),
1905 zone->pages_scanned,
1906 (zone_is_all_unreclaimable(zone) ? "yes" : "no")
1908 printk("lowmem_reserve[]:");
1909 for (i = 0; i < MAX_NR_ZONES; i++)
1910 printk(" %lu", zone->lowmem_reserve[i]);
1911 printk("\n");
1914 for_each_zone(zone) {
1915 unsigned long nr[MAX_ORDER], flags, order, total = 0;
1917 if (!populated_zone(zone))
1918 continue;
1920 show_node(zone);
1921 printk("%s: ", zone->name);
1923 spin_lock_irqsave(&zone->lock, flags);
1924 for (order = 0; order < MAX_ORDER; order++) {
1925 nr[order] = zone->free_area[order].nr_free;
1926 total += nr[order] << order;
1928 spin_unlock_irqrestore(&zone->lock, flags);
1929 for (order = 0; order < MAX_ORDER; order++)
1930 printk("%lu*%lukB ", nr[order], K(1UL) << order);
1931 printk("= %lukB\n", K(total));
1934 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
1936 show_swap_cache_info();
1939 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
1941 zoneref->zone = zone;
1942 zoneref->zone_idx = zone_idx(zone);
1946 * Builds allocation fallback zone lists.
1948 * Add all populated zones of a node to the zonelist.
1950 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
1951 int nr_zones, enum zone_type zone_type)
1953 struct zone *zone;
1955 BUG_ON(zone_type >= MAX_NR_ZONES);
1956 zone_type++;
1958 do {
1959 zone_type--;
1960 zone = pgdat->node_zones + zone_type;
1961 if (populated_zone(zone)) {
1962 zoneref_set_zone(zone,
1963 &zonelist->_zonerefs[nr_zones++]);
1964 check_highest_zone(zone_type);
1967 } while (zone_type);
1968 return nr_zones;
1973 * zonelist_order:
1974 * 0 = automatic detection of better ordering.
1975 * 1 = order by ([node] distance, -zonetype)
1976 * 2 = order by (-zonetype, [node] distance)
1978 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
1979 * the same zonelist. So only NUMA can configure this param.
1981 #define ZONELIST_ORDER_DEFAULT 0
1982 #define ZONELIST_ORDER_NODE 1
1983 #define ZONELIST_ORDER_ZONE 2
1985 /* zonelist order in the kernel.
1986 * set_zonelist_order() will set this to NODE or ZONE.
1988 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
1989 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
1992 #ifdef CONFIG_NUMA
1993 /* The value user specified ....changed by config */
1994 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
1995 /* string for sysctl */
1996 #define NUMA_ZONELIST_ORDER_LEN 16
1997 char numa_zonelist_order[16] = "default";
2000 * interface for configure zonelist ordering.
2001 * command line option "numa_zonelist_order"
2002 * = "[dD]efault - default, automatic configuration.
2003 * = "[nN]ode - order by node locality, then by zone within node
2004 * = "[zZ]one - order by zone, then by locality within zone
2007 static int __parse_numa_zonelist_order(char *s)
2009 if (*s == 'd' || *s == 'D') {
2010 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2011 } else if (*s == 'n' || *s == 'N') {
2012 user_zonelist_order = ZONELIST_ORDER_NODE;
2013 } else if (*s == 'z' || *s == 'Z') {
2014 user_zonelist_order = ZONELIST_ORDER_ZONE;
2015 } else {
2016 printk(KERN_WARNING
2017 "Ignoring invalid numa_zonelist_order value: "
2018 "%s\n", s);
2019 return -EINVAL;
2021 return 0;
2024 static __init int setup_numa_zonelist_order(char *s)
2026 if (s)
2027 return __parse_numa_zonelist_order(s);
2028 return 0;
2030 early_param("numa_zonelist_order", setup_numa_zonelist_order);
2033 * sysctl handler for numa_zonelist_order
2035 int numa_zonelist_order_handler(ctl_table *table, int write,
2036 struct file *file, void __user *buffer, size_t *length,
2037 loff_t *ppos)
2039 char saved_string[NUMA_ZONELIST_ORDER_LEN];
2040 int ret;
2042 if (write)
2043 strncpy(saved_string, (char*)table->data,
2044 NUMA_ZONELIST_ORDER_LEN);
2045 ret = proc_dostring(table, write, file, buffer, length, ppos);
2046 if (ret)
2047 return ret;
2048 if (write) {
2049 int oldval = user_zonelist_order;
2050 if (__parse_numa_zonelist_order((char*)table->data)) {
2052 * bogus value. restore saved string
2054 strncpy((char*)table->data, saved_string,
2055 NUMA_ZONELIST_ORDER_LEN);
2056 user_zonelist_order = oldval;
2057 } else if (oldval != user_zonelist_order)
2058 build_all_zonelists();
2060 return 0;
2064 #define MAX_NODE_LOAD (num_online_nodes())
2065 static int node_load[MAX_NUMNODES];
2068 * find_next_best_node - find the next node that should appear in a given node's fallback list
2069 * @node: node whose fallback list we're appending
2070 * @used_node_mask: nodemask_t of already used nodes
2072 * We use a number of factors to determine which is the next node that should
2073 * appear on a given node's fallback list. The node should not have appeared
2074 * already in @node's fallback list, and it should be the next closest node
2075 * according to the distance array (which contains arbitrary distance values
2076 * from each node to each node in the system), and should also prefer nodes
2077 * with no CPUs, since presumably they'll have very little allocation pressure
2078 * on them otherwise.
2079 * It returns -1 if no node is found.
2081 static int find_next_best_node(int node, nodemask_t *used_node_mask)
2083 int n, val;
2084 int min_val = INT_MAX;
2085 int best_node = -1;
2086 node_to_cpumask_ptr(tmp, 0);
2088 /* Use the local node if we haven't already */
2089 if (!node_isset(node, *used_node_mask)) {
2090 node_set(node, *used_node_mask);
2091 return node;
2094 for_each_node_state(n, N_HIGH_MEMORY) {
2096 /* Don't want a node to appear more than once */
2097 if (node_isset(n, *used_node_mask))
2098 continue;
2100 /* Use the distance array to find the distance */
2101 val = node_distance(node, n);
2103 /* Penalize nodes under us ("prefer the next node") */
2104 val += (n < node);
2106 /* Give preference to headless and unused nodes */
2107 node_to_cpumask_ptr_next(tmp, n);
2108 if (!cpus_empty(*tmp))
2109 val += PENALTY_FOR_NODE_WITH_CPUS;
2111 /* Slight preference for less loaded node */
2112 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
2113 val += node_load[n];
2115 if (val < min_val) {
2116 min_val = val;
2117 best_node = n;
2121 if (best_node >= 0)
2122 node_set(best_node, *used_node_mask);
2124 return best_node;
2129 * Build zonelists ordered by node and zones within node.
2130 * This results in maximum locality--normal zone overflows into local
2131 * DMA zone, if any--but risks exhausting DMA zone.
2133 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
2135 int j;
2136 struct zonelist *zonelist;
2138 zonelist = &pgdat->node_zonelists[0];
2139 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
2141 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2142 MAX_NR_ZONES - 1);
2143 zonelist->_zonerefs[j].zone = NULL;
2144 zonelist->_zonerefs[j].zone_idx = 0;
2148 * Build gfp_thisnode zonelists
2150 static void build_thisnode_zonelists(pg_data_t *pgdat)
2152 int j;
2153 struct zonelist *zonelist;
2155 zonelist = &pgdat->node_zonelists[1];
2156 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2157 zonelist->_zonerefs[j].zone = NULL;
2158 zonelist->_zonerefs[j].zone_idx = 0;
2162 * Build zonelists ordered by zone and nodes within zones.
2163 * This results in conserving DMA zone[s] until all Normal memory is
2164 * exhausted, but results in overflowing to remote node while memory
2165 * may still exist in local DMA zone.
2167 static int node_order[MAX_NUMNODES];
2169 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
2171 int pos, j, node;
2172 int zone_type; /* needs to be signed */
2173 struct zone *z;
2174 struct zonelist *zonelist;
2176 zonelist = &pgdat->node_zonelists[0];
2177 pos = 0;
2178 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
2179 for (j = 0; j < nr_nodes; j++) {
2180 node = node_order[j];
2181 z = &NODE_DATA(node)->node_zones[zone_type];
2182 if (populated_zone(z)) {
2183 zoneref_set_zone(z,
2184 &zonelist->_zonerefs[pos++]);
2185 check_highest_zone(zone_type);
2189 zonelist->_zonerefs[pos].zone = NULL;
2190 zonelist->_zonerefs[pos].zone_idx = 0;
2193 static int default_zonelist_order(void)
2195 int nid, zone_type;
2196 unsigned long low_kmem_size,total_size;
2197 struct zone *z;
2198 int average_size;
2200 * ZONE_DMA and ZONE_DMA32 can be very small area in the sytem.
2201 * If they are really small and used heavily, the system can fall
2202 * into OOM very easily.
2203 * This function detect ZONE_DMA/DMA32 size and confgigures zone order.
2205 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
2206 low_kmem_size = 0;
2207 total_size = 0;
2208 for_each_online_node(nid) {
2209 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2210 z = &NODE_DATA(nid)->node_zones[zone_type];
2211 if (populated_zone(z)) {
2212 if (zone_type < ZONE_NORMAL)
2213 low_kmem_size += z->present_pages;
2214 total_size += z->present_pages;
2218 if (!low_kmem_size || /* there are no DMA area. */
2219 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
2220 return ZONELIST_ORDER_NODE;
2222 * look into each node's config.
2223 * If there is a node whose DMA/DMA32 memory is very big area on
2224 * local memory, NODE_ORDER may be suitable.
2226 average_size = total_size /
2227 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1);
2228 for_each_online_node(nid) {
2229 low_kmem_size = 0;
2230 total_size = 0;
2231 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2232 z = &NODE_DATA(nid)->node_zones[zone_type];
2233 if (populated_zone(z)) {
2234 if (zone_type < ZONE_NORMAL)
2235 low_kmem_size += z->present_pages;
2236 total_size += z->present_pages;
2239 if (low_kmem_size &&
2240 total_size > average_size && /* ignore small node */
2241 low_kmem_size > total_size * 70/100)
2242 return ZONELIST_ORDER_NODE;
2244 return ZONELIST_ORDER_ZONE;
2247 static void set_zonelist_order(void)
2249 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
2250 current_zonelist_order = default_zonelist_order();
2251 else
2252 current_zonelist_order = user_zonelist_order;
2255 static void build_zonelists(pg_data_t *pgdat)
2257 int j, node, load;
2258 enum zone_type i;
2259 nodemask_t used_mask;
2260 int local_node, prev_node;
2261 struct zonelist *zonelist;
2262 int order = current_zonelist_order;
2264 /* initialize zonelists */
2265 for (i = 0; i < MAX_ZONELISTS; i++) {
2266 zonelist = pgdat->node_zonelists + i;
2267 zonelist->_zonerefs[0].zone = NULL;
2268 zonelist->_zonerefs[0].zone_idx = 0;
2271 /* NUMA-aware ordering of nodes */
2272 local_node = pgdat->node_id;
2273 load = num_online_nodes();
2274 prev_node = local_node;
2275 nodes_clear(used_mask);
2277 memset(node_load, 0, sizeof(node_load));
2278 memset(node_order, 0, sizeof(node_order));
2279 j = 0;
2281 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
2282 int distance = node_distance(local_node, node);
2285 * If another node is sufficiently far away then it is better
2286 * to reclaim pages in a zone before going off node.
2288 if (distance > RECLAIM_DISTANCE)
2289 zone_reclaim_mode = 1;
2292 * We don't want to pressure a particular node.
2293 * So adding penalty to the first node in same
2294 * distance group to make it round-robin.
2296 if (distance != node_distance(local_node, prev_node))
2297 node_load[node] = load;
2299 prev_node = node;
2300 load--;
2301 if (order == ZONELIST_ORDER_NODE)
2302 build_zonelists_in_node_order(pgdat, node);
2303 else
2304 node_order[j++] = node; /* remember order */
2307 if (order == ZONELIST_ORDER_ZONE) {
2308 /* calculate node order -- i.e., DMA last! */
2309 build_zonelists_in_zone_order(pgdat, j);
2312 build_thisnode_zonelists(pgdat);
2315 /* Construct the zonelist performance cache - see further mmzone.h */
2316 static void build_zonelist_cache(pg_data_t *pgdat)
2318 struct zonelist *zonelist;
2319 struct zonelist_cache *zlc;
2320 struct zoneref *z;
2322 zonelist = &pgdat->node_zonelists[0];
2323 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
2324 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
2325 for (z = zonelist->_zonerefs; z->zone; z++)
2326 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
2330 #else /* CONFIG_NUMA */
2332 static void set_zonelist_order(void)
2334 current_zonelist_order = ZONELIST_ORDER_ZONE;
2337 static void build_zonelists(pg_data_t *pgdat)
2339 int node, local_node;
2340 enum zone_type j;
2341 struct zonelist *zonelist;
2343 local_node = pgdat->node_id;
2345 zonelist = &pgdat->node_zonelists[0];
2346 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2349 * Now we build the zonelist so that it contains the zones
2350 * of all the other nodes.
2351 * We don't want to pressure a particular node, so when
2352 * building the zones for node N, we make sure that the
2353 * zones coming right after the local ones are those from
2354 * node N+1 (modulo N)
2356 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
2357 if (!node_online(node))
2358 continue;
2359 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2360 MAX_NR_ZONES - 1);
2362 for (node = 0; node < local_node; node++) {
2363 if (!node_online(node))
2364 continue;
2365 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2366 MAX_NR_ZONES - 1);
2369 zonelist->_zonerefs[j].zone = NULL;
2370 zonelist->_zonerefs[j].zone_idx = 0;
2373 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
2374 static void build_zonelist_cache(pg_data_t *pgdat)
2376 pgdat->node_zonelists[0].zlcache_ptr = NULL;
2379 #endif /* CONFIG_NUMA */
2381 /* return values int ....just for stop_machine() */
2382 static int __build_all_zonelists(void *dummy)
2384 int nid;
2386 for_each_online_node(nid) {
2387 pg_data_t *pgdat = NODE_DATA(nid);
2389 build_zonelists(pgdat);
2390 build_zonelist_cache(pgdat);
2392 return 0;
2395 void build_all_zonelists(void)
2397 set_zonelist_order();
2399 if (system_state == SYSTEM_BOOTING) {
2400 __build_all_zonelists(NULL);
2401 mminit_verify_zonelist();
2402 cpuset_init_current_mems_allowed();
2403 } else {
2404 /* we have to stop all cpus to guarantee there is no user
2405 of zonelist */
2406 stop_machine(__build_all_zonelists, NULL, NULL);
2407 /* cpuset refresh routine should be here */
2409 vm_total_pages = nr_free_pagecache_pages();
2411 * Disable grouping by mobility if the number of pages in the
2412 * system is too low to allow the mechanism to work. It would be
2413 * more accurate, but expensive to check per-zone. This check is
2414 * made on memory-hotadd so a system can start with mobility
2415 * disabled and enable it later
2417 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
2418 page_group_by_mobility_disabled = 1;
2419 else
2420 page_group_by_mobility_disabled = 0;
2422 printk("Built %i zonelists in %s order, mobility grouping %s. "
2423 "Total pages: %ld\n",
2424 num_online_nodes(),
2425 zonelist_order_name[current_zonelist_order],
2426 page_group_by_mobility_disabled ? "off" : "on",
2427 vm_total_pages);
2428 #ifdef CONFIG_NUMA
2429 printk("Policy zone: %s\n", zone_names[policy_zone]);
2430 #endif
2434 * Helper functions to size the waitqueue hash table.
2435 * Essentially these want to choose hash table sizes sufficiently
2436 * large so that collisions trying to wait on pages are rare.
2437 * But in fact, the number of active page waitqueues on typical
2438 * systems is ridiculously low, less than 200. So this is even
2439 * conservative, even though it seems large.
2441 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
2442 * waitqueues, i.e. the size of the waitq table given the number of pages.
2444 #define PAGES_PER_WAITQUEUE 256
2446 #ifndef CONFIG_MEMORY_HOTPLUG
2447 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
2449 unsigned long size = 1;
2451 pages /= PAGES_PER_WAITQUEUE;
2453 while (size < pages)
2454 size <<= 1;
2457 * Once we have dozens or even hundreds of threads sleeping
2458 * on IO we've got bigger problems than wait queue collision.
2459 * Limit the size of the wait table to a reasonable size.
2461 size = min(size, 4096UL);
2463 return max(size, 4UL);
2465 #else
2467 * A zone's size might be changed by hot-add, so it is not possible to determine
2468 * a suitable size for its wait_table. So we use the maximum size now.
2470 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
2472 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
2473 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
2474 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
2476 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
2477 * or more by the traditional way. (See above). It equals:
2479 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
2480 * ia64(16K page size) : = ( 8G + 4M)byte.
2481 * powerpc (64K page size) : = (32G +16M)byte.
2483 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
2485 return 4096UL;
2487 #endif
2490 * This is an integer logarithm so that shifts can be used later
2491 * to extract the more random high bits from the multiplicative
2492 * hash function before the remainder is taken.
2494 static inline unsigned long wait_table_bits(unsigned long size)
2496 return ffz(~size);
2499 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
2502 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
2503 * of blocks reserved is based on zone->pages_min. The memory within the
2504 * reserve will tend to store contiguous free pages. Setting min_free_kbytes
2505 * higher will lead to a bigger reserve which will get freed as contiguous
2506 * blocks as reclaim kicks in
2508 static void setup_zone_migrate_reserve(struct zone *zone)
2510 unsigned long start_pfn, pfn, end_pfn;
2511 struct page *page;
2512 unsigned long reserve, block_migratetype;
2514 /* Get the start pfn, end pfn and the number of blocks to reserve */
2515 start_pfn = zone->zone_start_pfn;
2516 end_pfn = start_pfn + zone->spanned_pages;
2517 reserve = roundup(zone->pages_min, pageblock_nr_pages) >>
2518 pageblock_order;
2520 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
2521 if (!pfn_valid(pfn))
2522 continue;
2523 page = pfn_to_page(pfn);
2525 /* Watch out for overlapping nodes */
2526 if (page_to_nid(page) != zone_to_nid(zone))
2527 continue;
2529 /* Blocks with reserved pages will never free, skip them. */
2530 if (PageReserved(page))
2531 continue;
2533 block_migratetype = get_pageblock_migratetype(page);
2535 /* If this block is reserved, account for it */
2536 if (reserve > 0 && block_migratetype == MIGRATE_RESERVE) {
2537 reserve--;
2538 continue;
2541 /* Suitable for reserving if this block is movable */
2542 if (reserve > 0 && block_migratetype == MIGRATE_MOVABLE) {
2543 set_pageblock_migratetype(page, MIGRATE_RESERVE);
2544 move_freepages_block(zone, page, MIGRATE_RESERVE);
2545 reserve--;
2546 continue;
2550 * If the reserve is met and this is a previous reserved block,
2551 * take it back
2553 if (block_migratetype == MIGRATE_RESERVE) {
2554 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
2555 move_freepages_block(zone, page, MIGRATE_MOVABLE);
2561 * Initially all pages are reserved - free ones are freed
2562 * up by free_all_bootmem() once the early boot process is
2563 * done. Non-atomic initialization, single-pass.
2565 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
2566 unsigned long start_pfn, enum memmap_context context)
2568 struct page *page;
2569 unsigned long end_pfn = start_pfn + size;
2570 unsigned long pfn;
2571 struct zone *z;
2573 z = &NODE_DATA(nid)->node_zones[zone];
2574 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
2576 * There can be holes in boot-time mem_map[]s
2577 * handed to this function. They do not
2578 * exist on hotplugged memory.
2580 if (context == MEMMAP_EARLY) {
2581 if (!early_pfn_valid(pfn))
2582 continue;
2583 if (!early_pfn_in_nid(pfn, nid))
2584 continue;
2586 page = pfn_to_page(pfn);
2587 set_page_links(page, zone, nid, pfn);
2588 mminit_verify_page_links(page, zone, nid, pfn);
2589 init_page_count(page);
2590 reset_page_mapcount(page);
2591 SetPageReserved(page);
2593 * Mark the block movable so that blocks are reserved for
2594 * movable at startup. This will force kernel allocations
2595 * to reserve their blocks rather than leaking throughout
2596 * the address space during boot when many long-lived
2597 * kernel allocations are made. Later some blocks near
2598 * the start are marked MIGRATE_RESERVE by
2599 * setup_zone_migrate_reserve()
2601 * bitmap is created for zone's valid pfn range. but memmap
2602 * can be created for invalid pages (for alignment)
2603 * check here not to call set_pageblock_migratetype() against
2604 * pfn out of zone.
2606 if ((z->zone_start_pfn <= pfn)
2607 && (pfn < z->zone_start_pfn + z->spanned_pages)
2608 && !(pfn & (pageblock_nr_pages - 1)))
2609 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
2611 INIT_LIST_HEAD(&page->lru);
2612 #ifdef WANT_PAGE_VIRTUAL
2613 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
2614 if (!is_highmem_idx(zone))
2615 set_page_address(page, __va(pfn << PAGE_SHIFT));
2616 #endif
2620 static void __meminit zone_init_free_lists(struct zone *zone)
2622 int order, t;
2623 for_each_migratetype_order(order, t) {
2624 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
2625 zone->free_area[order].nr_free = 0;
2629 #ifndef __HAVE_ARCH_MEMMAP_INIT
2630 #define memmap_init(size, nid, zone, start_pfn) \
2631 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
2632 #endif
2634 static int zone_batchsize(struct zone *zone)
2636 int batch;
2639 * The per-cpu-pages pools are set to around 1000th of the
2640 * size of the zone. But no more than 1/2 of a meg.
2642 * OK, so we don't know how big the cache is. So guess.
2644 batch = zone->present_pages / 1024;
2645 if (batch * PAGE_SIZE > 512 * 1024)
2646 batch = (512 * 1024) / PAGE_SIZE;
2647 batch /= 4; /* We effectively *= 4 below */
2648 if (batch < 1)
2649 batch = 1;
2652 * Clamp the batch to a 2^n - 1 value. Having a power
2653 * of 2 value was found to be more likely to have
2654 * suboptimal cache aliasing properties in some cases.
2656 * For example if 2 tasks are alternately allocating
2657 * batches of pages, one task can end up with a lot
2658 * of pages of one half of the possible page colors
2659 * and the other with pages of the other colors.
2661 batch = (1 << (fls(batch + batch/2)-1)) - 1;
2663 return batch;
2666 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
2668 struct per_cpu_pages *pcp;
2670 memset(p, 0, sizeof(*p));
2672 pcp = &p->pcp;
2673 pcp->count = 0;
2674 pcp->high = 6 * batch;
2675 pcp->batch = max(1UL, 1 * batch);
2676 INIT_LIST_HEAD(&pcp->list);
2680 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
2681 * to the value high for the pageset p.
2684 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
2685 unsigned long high)
2687 struct per_cpu_pages *pcp;
2689 pcp = &p->pcp;
2690 pcp->high = high;
2691 pcp->batch = max(1UL, high/4);
2692 if ((high/4) > (PAGE_SHIFT * 8))
2693 pcp->batch = PAGE_SHIFT * 8;
2697 #ifdef CONFIG_NUMA
2699 * Boot pageset table. One per cpu which is going to be used for all
2700 * zones and all nodes. The parameters will be set in such a way
2701 * that an item put on a list will immediately be handed over to
2702 * the buddy list. This is safe since pageset manipulation is done
2703 * with interrupts disabled.
2705 * Some NUMA counter updates may also be caught by the boot pagesets.
2707 * The boot_pagesets must be kept even after bootup is complete for
2708 * unused processors and/or zones. They do play a role for bootstrapping
2709 * hotplugged processors.
2711 * zoneinfo_show() and maybe other functions do
2712 * not check if the processor is online before following the pageset pointer.
2713 * Other parts of the kernel may not check if the zone is available.
2715 static struct per_cpu_pageset boot_pageset[NR_CPUS];
2718 * Dynamically allocate memory for the
2719 * per cpu pageset array in struct zone.
2721 static int __cpuinit process_zones(int cpu)
2723 struct zone *zone, *dzone;
2724 int node = cpu_to_node(cpu);
2726 node_set_state(node, N_CPU); /* this node has a cpu */
2728 for_each_zone(zone) {
2730 if (!populated_zone(zone))
2731 continue;
2733 zone_pcp(zone, cpu) = kmalloc_node(sizeof(struct per_cpu_pageset),
2734 GFP_KERNEL, node);
2735 if (!zone_pcp(zone, cpu))
2736 goto bad;
2738 setup_pageset(zone_pcp(zone, cpu), zone_batchsize(zone));
2740 if (percpu_pagelist_fraction)
2741 setup_pagelist_highmark(zone_pcp(zone, cpu),
2742 (zone->present_pages / percpu_pagelist_fraction));
2745 return 0;
2746 bad:
2747 for_each_zone(dzone) {
2748 if (!populated_zone(dzone))
2749 continue;
2750 if (dzone == zone)
2751 break;
2752 kfree(zone_pcp(dzone, cpu));
2753 zone_pcp(dzone, cpu) = NULL;
2755 return -ENOMEM;
2758 static inline void free_zone_pagesets(int cpu)
2760 struct zone *zone;
2762 for_each_zone(zone) {
2763 struct per_cpu_pageset *pset = zone_pcp(zone, cpu);
2765 /* Free per_cpu_pageset if it is slab allocated */
2766 if (pset != &boot_pageset[cpu])
2767 kfree(pset);
2768 zone_pcp(zone, cpu) = NULL;
2772 static int __cpuinit pageset_cpuup_callback(struct notifier_block *nfb,
2773 unsigned long action,
2774 void *hcpu)
2776 int cpu = (long)hcpu;
2777 int ret = NOTIFY_OK;
2779 switch (action) {
2780 case CPU_UP_PREPARE:
2781 case CPU_UP_PREPARE_FROZEN:
2782 if (process_zones(cpu))
2783 ret = NOTIFY_BAD;
2784 break;
2785 case CPU_UP_CANCELED:
2786 case CPU_UP_CANCELED_FROZEN:
2787 case CPU_DEAD:
2788 case CPU_DEAD_FROZEN:
2789 free_zone_pagesets(cpu);
2790 break;
2791 default:
2792 break;
2794 return ret;
2797 static struct notifier_block __cpuinitdata pageset_notifier =
2798 { &pageset_cpuup_callback, NULL, 0 };
2800 void __init setup_per_cpu_pageset(void)
2802 int err;
2804 /* Initialize per_cpu_pageset for cpu 0.
2805 * A cpuup callback will do this for every cpu
2806 * as it comes online
2808 err = process_zones(smp_processor_id());
2809 BUG_ON(err);
2810 register_cpu_notifier(&pageset_notifier);
2813 #endif
2815 static noinline __init_refok
2816 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
2818 int i;
2819 struct pglist_data *pgdat = zone->zone_pgdat;
2820 size_t alloc_size;
2823 * The per-page waitqueue mechanism uses hashed waitqueues
2824 * per zone.
2826 zone->wait_table_hash_nr_entries =
2827 wait_table_hash_nr_entries(zone_size_pages);
2828 zone->wait_table_bits =
2829 wait_table_bits(zone->wait_table_hash_nr_entries);
2830 alloc_size = zone->wait_table_hash_nr_entries
2831 * sizeof(wait_queue_head_t);
2833 if (!slab_is_available()) {
2834 zone->wait_table = (wait_queue_head_t *)
2835 alloc_bootmem_node(pgdat, alloc_size);
2836 } else {
2838 * This case means that a zone whose size was 0 gets new memory
2839 * via memory hot-add.
2840 * But it may be the case that a new node was hot-added. In
2841 * this case vmalloc() will not be able to use this new node's
2842 * memory - this wait_table must be initialized to use this new
2843 * node itself as well.
2844 * To use this new node's memory, further consideration will be
2845 * necessary.
2847 zone->wait_table = vmalloc(alloc_size);
2849 if (!zone->wait_table)
2850 return -ENOMEM;
2852 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
2853 init_waitqueue_head(zone->wait_table + i);
2855 return 0;
2858 static __meminit void zone_pcp_init(struct zone *zone)
2860 int cpu;
2861 unsigned long batch = zone_batchsize(zone);
2863 for (cpu = 0; cpu < NR_CPUS; cpu++) {
2864 #ifdef CONFIG_NUMA
2865 /* Early boot. Slab allocator not functional yet */
2866 zone_pcp(zone, cpu) = &boot_pageset[cpu];
2867 setup_pageset(&boot_pageset[cpu],0);
2868 #else
2869 setup_pageset(zone_pcp(zone,cpu), batch);
2870 #endif
2872 if (zone->present_pages)
2873 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%lu\n",
2874 zone->name, zone->present_pages, batch);
2877 __meminit int init_currently_empty_zone(struct zone *zone,
2878 unsigned long zone_start_pfn,
2879 unsigned long size,
2880 enum memmap_context context)
2882 struct pglist_data *pgdat = zone->zone_pgdat;
2883 int ret;
2884 ret = zone_wait_table_init(zone, size);
2885 if (ret)
2886 return ret;
2887 pgdat->nr_zones = zone_idx(zone) + 1;
2889 zone->zone_start_pfn = zone_start_pfn;
2891 mminit_dprintk(MMINIT_TRACE, "memmap_init",
2892 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
2893 pgdat->node_id,
2894 (unsigned long)zone_idx(zone),
2895 zone_start_pfn, (zone_start_pfn + size));
2897 zone_init_free_lists(zone);
2899 return 0;
2902 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
2904 * Basic iterator support. Return the first range of PFNs for a node
2905 * Note: nid == MAX_NUMNODES returns first region regardless of node
2907 static int __meminit first_active_region_index_in_nid(int nid)
2909 int i;
2911 for (i = 0; i < nr_nodemap_entries; i++)
2912 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
2913 return i;
2915 return -1;
2919 * Basic iterator support. Return the next active range of PFNs for a node
2920 * Note: nid == MAX_NUMNODES returns next region regardless of node
2922 static int __meminit next_active_region_index_in_nid(int index, int nid)
2924 for (index = index + 1; index < nr_nodemap_entries; index++)
2925 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
2926 return index;
2928 return -1;
2931 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
2933 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
2934 * Architectures may implement their own version but if add_active_range()
2935 * was used and there are no special requirements, this is a convenient
2936 * alternative
2938 int __meminit early_pfn_to_nid(unsigned long pfn)
2940 int i;
2942 for (i = 0; i < nr_nodemap_entries; i++) {
2943 unsigned long start_pfn = early_node_map[i].start_pfn;
2944 unsigned long end_pfn = early_node_map[i].end_pfn;
2946 if (start_pfn <= pfn && pfn < end_pfn)
2947 return early_node_map[i].nid;
2950 return 0;
2952 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
2954 /* Basic iterator support to walk early_node_map[] */
2955 #define for_each_active_range_index_in_nid(i, nid) \
2956 for (i = first_active_region_index_in_nid(nid); i != -1; \
2957 i = next_active_region_index_in_nid(i, nid))
2960 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
2961 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
2962 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
2964 * If an architecture guarantees that all ranges registered with
2965 * add_active_ranges() contain no holes and may be freed, this
2966 * this function may be used instead of calling free_bootmem() manually.
2968 void __init free_bootmem_with_active_regions(int nid,
2969 unsigned long max_low_pfn)
2971 int i;
2973 for_each_active_range_index_in_nid(i, nid) {
2974 unsigned long size_pages = 0;
2975 unsigned long end_pfn = early_node_map[i].end_pfn;
2977 if (early_node_map[i].start_pfn >= max_low_pfn)
2978 continue;
2980 if (end_pfn > max_low_pfn)
2981 end_pfn = max_low_pfn;
2983 size_pages = end_pfn - early_node_map[i].start_pfn;
2984 free_bootmem_node(NODE_DATA(early_node_map[i].nid),
2985 PFN_PHYS(early_node_map[i].start_pfn),
2986 size_pages << PAGE_SHIFT);
2990 void __init work_with_active_regions(int nid, work_fn_t work_fn, void *data)
2992 int i;
2993 int ret;
2995 for_each_active_range_index_in_nid(i, nid) {
2996 ret = work_fn(early_node_map[i].start_pfn,
2997 early_node_map[i].end_pfn, data);
2998 if (ret)
2999 break;
3003 * sparse_memory_present_with_active_regions - Call memory_present for each active range
3004 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
3006 * If an architecture guarantees that all ranges registered with
3007 * add_active_ranges() contain no holes and may be freed, this
3008 * function may be used instead of calling memory_present() manually.
3010 void __init sparse_memory_present_with_active_regions(int nid)
3012 int i;
3014 for_each_active_range_index_in_nid(i, nid)
3015 memory_present(early_node_map[i].nid,
3016 early_node_map[i].start_pfn,
3017 early_node_map[i].end_pfn);
3021 * push_node_boundaries - Push node boundaries to at least the requested boundary
3022 * @nid: The nid of the node to push the boundary for
3023 * @start_pfn: The start pfn of the node
3024 * @end_pfn: The end pfn of the node
3026 * In reserve-based hot-add, mem_map is allocated that is unused until hotadd
3027 * time. Specifically, on x86_64, SRAT will report ranges that can potentially
3028 * be hotplugged even though no physical memory exists. This function allows
3029 * an arch to push out the node boundaries so mem_map is allocated that can
3030 * be used later.
3032 #ifdef CONFIG_MEMORY_HOTPLUG_RESERVE
3033 void __init push_node_boundaries(unsigned int nid,
3034 unsigned long start_pfn, unsigned long end_pfn)
3036 mminit_dprintk(MMINIT_TRACE, "zoneboundary",
3037 "Entering push_node_boundaries(%u, %lu, %lu)\n",
3038 nid, start_pfn, end_pfn);
3040 /* Initialise the boundary for this node if necessary */
3041 if (node_boundary_end_pfn[nid] == 0)
3042 node_boundary_start_pfn[nid] = -1UL;
3044 /* Update the boundaries */
3045 if (node_boundary_start_pfn[nid] > start_pfn)
3046 node_boundary_start_pfn[nid] = start_pfn;
3047 if (node_boundary_end_pfn[nid] < end_pfn)
3048 node_boundary_end_pfn[nid] = end_pfn;
3051 /* If necessary, push the node boundary out for reserve hotadd */
3052 static void __meminit account_node_boundary(unsigned int nid,
3053 unsigned long *start_pfn, unsigned long *end_pfn)
3055 mminit_dprintk(MMINIT_TRACE, "zoneboundary",
3056 "Entering account_node_boundary(%u, %lu, %lu)\n",
3057 nid, *start_pfn, *end_pfn);
3059 /* Return if boundary information has not been provided */
3060 if (node_boundary_end_pfn[nid] == 0)
3061 return;
3063 /* Check the boundaries and update if necessary */
3064 if (node_boundary_start_pfn[nid] < *start_pfn)
3065 *start_pfn = node_boundary_start_pfn[nid];
3066 if (node_boundary_end_pfn[nid] > *end_pfn)
3067 *end_pfn = node_boundary_end_pfn[nid];
3069 #else
3070 void __init push_node_boundaries(unsigned int nid,
3071 unsigned long start_pfn, unsigned long end_pfn) {}
3073 static void __meminit account_node_boundary(unsigned int nid,
3074 unsigned long *start_pfn, unsigned long *end_pfn) {}
3075 #endif
3079 * get_pfn_range_for_nid - Return the start and end page frames for a node
3080 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
3081 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
3082 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
3084 * It returns the start and end page frame of a node based on information
3085 * provided by an arch calling add_active_range(). If called for a node
3086 * with no available memory, a warning is printed and the start and end
3087 * PFNs will be 0.
3089 void __meminit get_pfn_range_for_nid(unsigned int nid,
3090 unsigned long *start_pfn, unsigned long *end_pfn)
3092 int i;
3093 *start_pfn = -1UL;
3094 *end_pfn = 0;
3096 for_each_active_range_index_in_nid(i, nid) {
3097 *start_pfn = min(*start_pfn, early_node_map[i].start_pfn);
3098 *end_pfn = max(*end_pfn, early_node_map[i].end_pfn);
3101 if (*start_pfn == -1UL)
3102 *start_pfn = 0;
3104 /* Push the node boundaries out if requested */
3105 account_node_boundary(nid, start_pfn, end_pfn);
3109 * This finds a zone that can be used for ZONE_MOVABLE pages. The
3110 * assumption is made that zones within a node are ordered in monotonic
3111 * increasing memory addresses so that the "highest" populated zone is used
3113 static void __init find_usable_zone_for_movable(void)
3115 int zone_index;
3116 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
3117 if (zone_index == ZONE_MOVABLE)
3118 continue;
3120 if (arch_zone_highest_possible_pfn[zone_index] >
3121 arch_zone_lowest_possible_pfn[zone_index])
3122 break;
3125 VM_BUG_ON(zone_index == -1);
3126 movable_zone = zone_index;
3130 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
3131 * because it is sized independant of architecture. Unlike the other zones,
3132 * the starting point for ZONE_MOVABLE is not fixed. It may be different
3133 * in each node depending on the size of each node and how evenly kernelcore
3134 * is distributed. This helper function adjusts the zone ranges
3135 * provided by the architecture for a given node by using the end of the
3136 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
3137 * zones within a node are in order of monotonic increases memory addresses
3139 static void __meminit adjust_zone_range_for_zone_movable(int nid,
3140 unsigned long zone_type,
3141 unsigned long node_start_pfn,
3142 unsigned long node_end_pfn,
3143 unsigned long *zone_start_pfn,
3144 unsigned long *zone_end_pfn)
3146 /* Only adjust if ZONE_MOVABLE is on this node */
3147 if (zone_movable_pfn[nid]) {
3148 /* Size ZONE_MOVABLE */
3149 if (zone_type == ZONE_MOVABLE) {
3150 *zone_start_pfn = zone_movable_pfn[nid];
3151 *zone_end_pfn = min(node_end_pfn,
3152 arch_zone_highest_possible_pfn[movable_zone]);
3154 /* Adjust for ZONE_MOVABLE starting within this range */
3155 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
3156 *zone_end_pfn > zone_movable_pfn[nid]) {
3157 *zone_end_pfn = zone_movable_pfn[nid];
3159 /* Check if this whole range is within ZONE_MOVABLE */
3160 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
3161 *zone_start_pfn = *zone_end_pfn;
3166 * Return the number of pages a zone spans in a node, including holes
3167 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
3169 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
3170 unsigned long zone_type,
3171 unsigned long *ignored)
3173 unsigned long node_start_pfn, node_end_pfn;
3174 unsigned long zone_start_pfn, zone_end_pfn;
3176 /* Get the start and end of the node and zone */
3177 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
3178 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
3179 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
3180 adjust_zone_range_for_zone_movable(nid, zone_type,
3181 node_start_pfn, node_end_pfn,
3182 &zone_start_pfn, &zone_end_pfn);
3184 /* Check that this node has pages within the zone's required range */
3185 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
3186 return 0;
3188 /* Move the zone boundaries inside the node if necessary */
3189 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
3190 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
3192 /* Return the spanned pages */
3193 return zone_end_pfn - zone_start_pfn;
3197 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
3198 * then all holes in the requested range will be accounted for.
3200 static unsigned long __meminit __absent_pages_in_range(int nid,
3201 unsigned long range_start_pfn,
3202 unsigned long range_end_pfn)
3204 int i = 0;
3205 unsigned long prev_end_pfn = 0, hole_pages = 0;
3206 unsigned long start_pfn;
3208 /* Find the end_pfn of the first active range of pfns in the node */
3209 i = first_active_region_index_in_nid(nid);
3210 if (i == -1)
3211 return 0;
3213 prev_end_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
3215 /* Account for ranges before physical memory on this node */
3216 if (early_node_map[i].start_pfn > range_start_pfn)
3217 hole_pages = prev_end_pfn - range_start_pfn;
3219 /* Find all holes for the zone within the node */
3220 for (; i != -1; i = next_active_region_index_in_nid(i, nid)) {
3222 /* No need to continue if prev_end_pfn is outside the zone */
3223 if (prev_end_pfn >= range_end_pfn)
3224 break;
3226 /* Make sure the end of the zone is not within the hole */
3227 start_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
3228 prev_end_pfn = max(prev_end_pfn, range_start_pfn);
3230 /* Update the hole size cound and move on */
3231 if (start_pfn > range_start_pfn) {
3232 BUG_ON(prev_end_pfn > start_pfn);
3233 hole_pages += start_pfn - prev_end_pfn;
3235 prev_end_pfn = early_node_map[i].end_pfn;
3238 /* Account for ranges past physical memory on this node */
3239 if (range_end_pfn > prev_end_pfn)
3240 hole_pages += range_end_pfn -
3241 max(range_start_pfn, prev_end_pfn);
3243 return hole_pages;
3247 * absent_pages_in_range - Return number of page frames in holes within a range
3248 * @start_pfn: The start PFN to start searching for holes
3249 * @end_pfn: The end PFN to stop searching for holes
3251 * It returns the number of pages frames in memory holes within a range.
3253 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
3254 unsigned long end_pfn)
3256 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
3259 /* Return the number of page frames in holes in a zone on a node */
3260 static unsigned long __meminit zone_absent_pages_in_node(int nid,
3261 unsigned long zone_type,
3262 unsigned long *ignored)
3264 unsigned long node_start_pfn, node_end_pfn;
3265 unsigned long zone_start_pfn, zone_end_pfn;
3267 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
3268 zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type],
3269 node_start_pfn);
3270 zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type],
3271 node_end_pfn);
3273 adjust_zone_range_for_zone_movable(nid, zone_type,
3274 node_start_pfn, node_end_pfn,
3275 &zone_start_pfn, &zone_end_pfn);
3276 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
3279 #else
3280 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
3281 unsigned long zone_type,
3282 unsigned long *zones_size)
3284 return zones_size[zone_type];
3287 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
3288 unsigned long zone_type,
3289 unsigned long *zholes_size)
3291 if (!zholes_size)
3292 return 0;
3294 return zholes_size[zone_type];
3297 #endif
3299 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
3300 unsigned long *zones_size, unsigned long *zholes_size)
3302 unsigned long realtotalpages, totalpages = 0;
3303 enum zone_type i;
3305 for (i = 0; i < MAX_NR_ZONES; i++)
3306 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
3307 zones_size);
3308 pgdat->node_spanned_pages = totalpages;
3310 realtotalpages = totalpages;
3311 for (i = 0; i < MAX_NR_ZONES; i++)
3312 realtotalpages -=
3313 zone_absent_pages_in_node(pgdat->node_id, i,
3314 zholes_size);
3315 pgdat->node_present_pages = realtotalpages;
3316 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
3317 realtotalpages);
3320 #ifndef CONFIG_SPARSEMEM
3322 * Calculate the size of the zone->blockflags rounded to an unsigned long
3323 * Start by making sure zonesize is a multiple of pageblock_order by rounding
3324 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
3325 * round what is now in bits to nearest long in bits, then return it in
3326 * bytes.
3328 static unsigned long __init usemap_size(unsigned long zonesize)
3330 unsigned long usemapsize;
3332 usemapsize = roundup(zonesize, pageblock_nr_pages);
3333 usemapsize = usemapsize >> pageblock_order;
3334 usemapsize *= NR_PAGEBLOCK_BITS;
3335 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
3337 return usemapsize / 8;
3340 static void __init setup_usemap(struct pglist_data *pgdat,
3341 struct zone *zone, unsigned long zonesize)
3343 unsigned long usemapsize = usemap_size(zonesize);
3344 zone->pageblock_flags = NULL;
3345 if (usemapsize) {
3346 zone->pageblock_flags = alloc_bootmem_node(pgdat, usemapsize);
3347 memset(zone->pageblock_flags, 0, usemapsize);
3350 #else
3351 static void inline setup_usemap(struct pglist_data *pgdat,
3352 struct zone *zone, unsigned long zonesize) {}
3353 #endif /* CONFIG_SPARSEMEM */
3355 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
3357 /* Return a sensible default order for the pageblock size. */
3358 static inline int pageblock_default_order(void)
3360 if (HPAGE_SHIFT > PAGE_SHIFT)
3361 return HUGETLB_PAGE_ORDER;
3363 return MAX_ORDER-1;
3366 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
3367 static inline void __init set_pageblock_order(unsigned int order)
3369 /* Check that pageblock_nr_pages has not already been setup */
3370 if (pageblock_order)
3371 return;
3374 * Assume the largest contiguous order of interest is a huge page.
3375 * This value may be variable depending on boot parameters on IA64
3377 pageblock_order = order;
3379 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
3382 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
3383 * and pageblock_default_order() are unused as pageblock_order is set
3384 * at compile-time. See include/linux/pageblock-flags.h for the values of
3385 * pageblock_order based on the kernel config
3387 static inline int pageblock_default_order(unsigned int order)
3389 return MAX_ORDER-1;
3391 #define set_pageblock_order(x) do {} while (0)
3393 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
3396 * Set up the zone data structures:
3397 * - mark all pages reserved
3398 * - mark all memory queues empty
3399 * - clear the memory bitmaps
3401 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
3402 unsigned long *zones_size, unsigned long *zholes_size)
3404 enum zone_type j;
3405 int nid = pgdat->node_id;
3406 unsigned long zone_start_pfn = pgdat->node_start_pfn;
3407 int ret;
3409 pgdat_resize_init(pgdat);
3410 pgdat->nr_zones = 0;
3411 init_waitqueue_head(&pgdat->kswapd_wait);
3412 pgdat->kswapd_max_order = 0;
3414 for (j = 0; j < MAX_NR_ZONES; j++) {
3415 struct zone *zone = pgdat->node_zones + j;
3416 unsigned long size, realsize, memmap_pages;
3418 size = zone_spanned_pages_in_node(nid, j, zones_size);
3419 realsize = size - zone_absent_pages_in_node(nid, j,
3420 zholes_size);
3423 * Adjust realsize so that it accounts for how much memory
3424 * is used by this zone for memmap. This affects the watermark
3425 * and per-cpu initialisations
3427 memmap_pages =
3428 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT;
3429 if (realsize >= memmap_pages) {
3430 realsize -= memmap_pages;
3431 mminit_dprintk(MMINIT_TRACE, "memmap_init",
3432 "%s zone: %lu pages used for memmap\n",
3433 zone_names[j], memmap_pages);
3434 } else
3435 printk(KERN_WARNING
3436 " %s zone: %lu pages exceeds realsize %lu\n",
3437 zone_names[j], memmap_pages, realsize);
3439 /* Account for reserved pages */
3440 if (j == 0 && realsize > dma_reserve) {
3441 realsize -= dma_reserve;
3442 mminit_dprintk(MMINIT_TRACE, "memmap_init",
3443 "%s zone: %lu pages reserved\n",
3444 zone_names[0], dma_reserve);
3447 if (!is_highmem_idx(j))
3448 nr_kernel_pages += realsize;
3449 nr_all_pages += realsize;
3451 zone->spanned_pages = size;
3452 zone->present_pages = realsize;
3453 #ifdef CONFIG_NUMA
3454 zone->node = nid;
3455 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
3456 / 100;
3457 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
3458 #endif
3459 zone->name = zone_names[j];
3460 spin_lock_init(&zone->lock);
3461 spin_lock_init(&zone->lru_lock);
3462 zone_seqlock_init(zone);
3463 zone->zone_pgdat = pgdat;
3465 zone->prev_priority = DEF_PRIORITY;
3467 zone_pcp_init(zone);
3468 INIT_LIST_HEAD(&zone->active_list);
3469 INIT_LIST_HEAD(&zone->inactive_list);
3470 zone->nr_scan_active = 0;
3471 zone->nr_scan_inactive = 0;
3472 zap_zone_vm_stats(zone);
3473 zone->flags = 0;
3474 if (!size)
3475 continue;
3477 set_pageblock_order(pageblock_default_order());
3478 setup_usemap(pgdat, zone, size);
3479 ret = init_currently_empty_zone(zone, zone_start_pfn,
3480 size, MEMMAP_EARLY);
3481 BUG_ON(ret);
3482 memmap_init(size, nid, j, zone_start_pfn);
3483 zone_start_pfn += size;
3487 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
3489 /* Skip empty nodes */
3490 if (!pgdat->node_spanned_pages)
3491 return;
3493 #ifdef CONFIG_FLAT_NODE_MEM_MAP
3494 /* ia64 gets its own node_mem_map, before this, without bootmem */
3495 if (!pgdat->node_mem_map) {
3496 unsigned long size, start, end;
3497 struct page *map;
3500 * The zone's endpoints aren't required to be MAX_ORDER
3501 * aligned but the node_mem_map endpoints must be in order
3502 * for the buddy allocator to function correctly.
3504 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
3505 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
3506 end = ALIGN(end, MAX_ORDER_NR_PAGES);
3507 size = (end - start) * sizeof(struct page);
3508 map = alloc_remap(pgdat->node_id, size);
3509 if (!map)
3510 map = alloc_bootmem_node(pgdat, size);
3511 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
3513 #ifndef CONFIG_NEED_MULTIPLE_NODES
3515 * With no DISCONTIG, the global mem_map is just set as node 0's
3517 if (pgdat == NODE_DATA(0)) {
3518 mem_map = NODE_DATA(0)->node_mem_map;
3519 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3520 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
3521 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
3522 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
3524 #endif
3525 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
3528 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
3529 unsigned long node_start_pfn, unsigned long *zholes_size)
3531 pg_data_t *pgdat = NODE_DATA(nid);
3533 pgdat->node_id = nid;
3534 pgdat->node_start_pfn = node_start_pfn;
3535 calculate_node_totalpages(pgdat, zones_size, zholes_size);
3537 alloc_node_mem_map(pgdat);
3538 #ifdef CONFIG_FLAT_NODE_MEM_MAP
3539 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
3540 nid, (unsigned long)pgdat,
3541 (unsigned long)pgdat->node_mem_map);
3542 #endif
3544 free_area_init_core(pgdat, zones_size, zholes_size);
3547 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3549 #if MAX_NUMNODES > 1
3551 * Figure out the number of possible node ids.
3553 static void __init setup_nr_node_ids(void)
3555 unsigned int node;
3556 unsigned int highest = 0;
3558 for_each_node_mask(node, node_possible_map)
3559 highest = node;
3560 nr_node_ids = highest + 1;
3562 #else
3563 static inline void setup_nr_node_ids(void)
3566 #endif
3569 * add_active_range - Register a range of PFNs backed by physical memory
3570 * @nid: The node ID the range resides on
3571 * @start_pfn: The start PFN of the available physical memory
3572 * @end_pfn: The end PFN of the available physical memory
3574 * These ranges are stored in an early_node_map[] and later used by
3575 * free_area_init_nodes() to calculate zone sizes and holes. If the
3576 * range spans a memory hole, it is up to the architecture to ensure
3577 * the memory is not freed by the bootmem allocator. If possible
3578 * the range being registered will be merged with existing ranges.
3580 void __init add_active_range(unsigned int nid, unsigned long start_pfn,
3581 unsigned long end_pfn)
3583 int i;
3585 mminit_dprintk(MMINIT_TRACE, "memory_register",
3586 "Entering add_active_range(%d, %#lx, %#lx) "
3587 "%d entries of %d used\n",
3588 nid, start_pfn, end_pfn,
3589 nr_nodemap_entries, MAX_ACTIVE_REGIONS);
3591 mminit_validate_memmodel_limits(&start_pfn, &end_pfn);
3593 /* Merge with existing active regions if possible */
3594 for (i = 0; i < nr_nodemap_entries; i++) {
3595 if (early_node_map[i].nid != nid)
3596 continue;
3598 /* Skip if an existing region covers this new one */
3599 if (start_pfn >= early_node_map[i].start_pfn &&
3600 end_pfn <= early_node_map[i].end_pfn)
3601 return;
3603 /* Merge forward if suitable */
3604 if (start_pfn <= early_node_map[i].end_pfn &&
3605 end_pfn > early_node_map[i].end_pfn) {
3606 early_node_map[i].end_pfn = end_pfn;
3607 return;
3610 /* Merge backward if suitable */
3611 if (start_pfn < early_node_map[i].end_pfn &&
3612 end_pfn >= early_node_map[i].start_pfn) {
3613 early_node_map[i].start_pfn = start_pfn;
3614 return;
3618 /* Check that early_node_map is large enough */
3619 if (i >= MAX_ACTIVE_REGIONS) {
3620 printk(KERN_CRIT "More than %d memory regions, truncating\n",
3621 MAX_ACTIVE_REGIONS);
3622 return;
3625 early_node_map[i].nid = nid;
3626 early_node_map[i].start_pfn = start_pfn;
3627 early_node_map[i].end_pfn = end_pfn;
3628 nr_nodemap_entries = i + 1;
3632 * remove_active_range - Shrink an existing registered range of PFNs
3633 * @nid: The node id the range is on that should be shrunk
3634 * @start_pfn: The new PFN of the range
3635 * @end_pfn: The new PFN of the range
3637 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
3638 * The map is kept near the end physical page range that has already been
3639 * registered. This function allows an arch to shrink an existing registered
3640 * range.
3642 void __init remove_active_range(unsigned int nid, unsigned long start_pfn,
3643 unsigned long end_pfn)
3645 int i, j;
3646 int removed = 0;
3648 printk(KERN_DEBUG "remove_active_range (%d, %lu, %lu)\n",
3649 nid, start_pfn, end_pfn);
3651 /* Find the old active region end and shrink */
3652 for_each_active_range_index_in_nid(i, nid) {
3653 if (early_node_map[i].start_pfn >= start_pfn &&
3654 early_node_map[i].end_pfn <= end_pfn) {
3655 /* clear it */
3656 early_node_map[i].start_pfn = 0;
3657 early_node_map[i].end_pfn = 0;
3658 removed = 1;
3659 continue;
3661 if (early_node_map[i].start_pfn < start_pfn &&
3662 early_node_map[i].end_pfn > start_pfn) {
3663 unsigned long temp_end_pfn = early_node_map[i].end_pfn;
3664 early_node_map[i].end_pfn = start_pfn;
3665 if (temp_end_pfn > end_pfn)
3666 add_active_range(nid, end_pfn, temp_end_pfn);
3667 continue;
3669 if (early_node_map[i].start_pfn >= start_pfn &&
3670 early_node_map[i].end_pfn > end_pfn &&
3671 early_node_map[i].start_pfn < end_pfn) {
3672 early_node_map[i].start_pfn = end_pfn;
3673 continue;
3677 if (!removed)
3678 return;
3680 /* remove the blank ones */
3681 for (i = nr_nodemap_entries - 1; i > 0; i--) {
3682 if (early_node_map[i].nid != nid)
3683 continue;
3684 if (early_node_map[i].end_pfn)
3685 continue;
3686 /* we found it, get rid of it */
3687 for (j = i; j < nr_nodemap_entries - 1; j++)
3688 memcpy(&early_node_map[j], &early_node_map[j+1],
3689 sizeof(early_node_map[j]));
3690 j = nr_nodemap_entries - 1;
3691 memset(&early_node_map[j], 0, sizeof(early_node_map[j]));
3692 nr_nodemap_entries--;
3697 * remove_all_active_ranges - Remove all currently registered regions
3699 * During discovery, it may be found that a table like SRAT is invalid
3700 * and an alternative discovery method must be used. This function removes
3701 * all currently registered regions.
3703 void __init remove_all_active_ranges(void)
3705 memset(early_node_map, 0, sizeof(early_node_map));
3706 nr_nodemap_entries = 0;
3707 #ifdef CONFIG_MEMORY_HOTPLUG_RESERVE
3708 memset(node_boundary_start_pfn, 0, sizeof(node_boundary_start_pfn));
3709 memset(node_boundary_end_pfn, 0, sizeof(node_boundary_end_pfn));
3710 #endif /* CONFIG_MEMORY_HOTPLUG_RESERVE */
3713 /* Compare two active node_active_regions */
3714 static int __init cmp_node_active_region(const void *a, const void *b)
3716 struct node_active_region *arange = (struct node_active_region *)a;
3717 struct node_active_region *brange = (struct node_active_region *)b;
3719 /* Done this way to avoid overflows */
3720 if (arange->start_pfn > brange->start_pfn)
3721 return 1;
3722 if (arange->start_pfn < brange->start_pfn)
3723 return -1;
3725 return 0;
3728 /* sort the node_map by start_pfn */
3729 static void __init sort_node_map(void)
3731 sort(early_node_map, (size_t)nr_nodemap_entries,
3732 sizeof(struct node_active_region),
3733 cmp_node_active_region, NULL);
3736 /* Find the lowest pfn for a node */
3737 static unsigned long __init find_min_pfn_for_node(int nid)
3739 int i;
3740 unsigned long min_pfn = ULONG_MAX;
3742 /* Assuming a sorted map, the first range found has the starting pfn */
3743 for_each_active_range_index_in_nid(i, nid)
3744 min_pfn = min(min_pfn, early_node_map[i].start_pfn);
3746 if (min_pfn == ULONG_MAX) {
3747 printk(KERN_WARNING
3748 "Could not find start_pfn for node %d\n", nid);
3749 return 0;
3752 return min_pfn;
3756 * find_min_pfn_with_active_regions - Find the minimum PFN registered
3758 * It returns the minimum PFN based on information provided via
3759 * add_active_range().
3761 unsigned long __init find_min_pfn_with_active_regions(void)
3763 return find_min_pfn_for_node(MAX_NUMNODES);
3767 * early_calculate_totalpages()
3768 * Sum pages in active regions for movable zone.
3769 * Populate N_HIGH_MEMORY for calculating usable_nodes.
3771 static unsigned long __init early_calculate_totalpages(void)
3773 int i;
3774 unsigned long totalpages = 0;
3776 for (i = 0; i < nr_nodemap_entries; i++) {
3777 unsigned long pages = early_node_map[i].end_pfn -
3778 early_node_map[i].start_pfn;
3779 totalpages += pages;
3780 if (pages)
3781 node_set_state(early_node_map[i].nid, N_HIGH_MEMORY);
3783 return totalpages;
3787 * Find the PFN the Movable zone begins in each node. Kernel memory
3788 * is spread evenly between nodes as long as the nodes have enough
3789 * memory. When they don't, some nodes will have more kernelcore than
3790 * others
3792 static void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn)
3794 int i, nid;
3795 unsigned long usable_startpfn;
3796 unsigned long kernelcore_node, kernelcore_remaining;
3797 unsigned long totalpages = early_calculate_totalpages();
3798 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
3801 * If movablecore was specified, calculate what size of
3802 * kernelcore that corresponds so that memory usable for
3803 * any allocation type is evenly spread. If both kernelcore
3804 * and movablecore are specified, then the value of kernelcore
3805 * will be used for required_kernelcore if it's greater than
3806 * what movablecore would have allowed.
3808 if (required_movablecore) {
3809 unsigned long corepages;
3812 * Round-up so that ZONE_MOVABLE is at least as large as what
3813 * was requested by the user
3815 required_movablecore =
3816 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
3817 corepages = totalpages - required_movablecore;
3819 required_kernelcore = max(required_kernelcore, corepages);
3822 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
3823 if (!required_kernelcore)
3824 return;
3826 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
3827 find_usable_zone_for_movable();
3828 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
3830 restart:
3831 /* Spread kernelcore memory as evenly as possible throughout nodes */
3832 kernelcore_node = required_kernelcore / usable_nodes;
3833 for_each_node_state(nid, N_HIGH_MEMORY) {
3835 * Recalculate kernelcore_node if the division per node
3836 * now exceeds what is necessary to satisfy the requested
3837 * amount of memory for the kernel
3839 if (required_kernelcore < kernelcore_node)
3840 kernelcore_node = required_kernelcore / usable_nodes;
3843 * As the map is walked, we track how much memory is usable
3844 * by the kernel using kernelcore_remaining. When it is
3845 * 0, the rest of the node is usable by ZONE_MOVABLE
3847 kernelcore_remaining = kernelcore_node;
3849 /* Go through each range of PFNs within this node */
3850 for_each_active_range_index_in_nid(i, nid) {
3851 unsigned long start_pfn, end_pfn;
3852 unsigned long size_pages;
3854 start_pfn = max(early_node_map[i].start_pfn,
3855 zone_movable_pfn[nid]);
3856 end_pfn = early_node_map[i].end_pfn;
3857 if (start_pfn >= end_pfn)
3858 continue;
3860 /* Account for what is only usable for kernelcore */
3861 if (start_pfn < usable_startpfn) {
3862 unsigned long kernel_pages;
3863 kernel_pages = min(end_pfn, usable_startpfn)
3864 - start_pfn;
3866 kernelcore_remaining -= min(kernel_pages,
3867 kernelcore_remaining);
3868 required_kernelcore -= min(kernel_pages,
3869 required_kernelcore);
3871 /* Continue if range is now fully accounted */
3872 if (end_pfn <= usable_startpfn) {
3875 * Push zone_movable_pfn to the end so
3876 * that if we have to rebalance
3877 * kernelcore across nodes, we will
3878 * not double account here
3880 zone_movable_pfn[nid] = end_pfn;
3881 continue;
3883 start_pfn = usable_startpfn;
3887 * The usable PFN range for ZONE_MOVABLE is from
3888 * start_pfn->end_pfn. Calculate size_pages as the
3889 * number of pages used as kernelcore
3891 size_pages = end_pfn - start_pfn;
3892 if (size_pages > kernelcore_remaining)
3893 size_pages = kernelcore_remaining;
3894 zone_movable_pfn[nid] = start_pfn + size_pages;
3897 * Some kernelcore has been met, update counts and
3898 * break if the kernelcore for this node has been
3899 * satisified
3901 required_kernelcore -= min(required_kernelcore,
3902 size_pages);
3903 kernelcore_remaining -= size_pages;
3904 if (!kernelcore_remaining)
3905 break;
3910 * If there is still required_kernelcore, we do another pass with one
3911 * less node in the count. This will push zone_movable_pfn[nid] further
3912 * along on the nodes that still have memory until kernelcore is
3913 * satisified
3915 usable_nodes--;
3916 if (usable_nodes && required_kernelcore > usable_nodes)
3917 goto restart;
3919 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
3920 for (nid = 0; nid < MAX_NUMNODES; nid++)
3921 zone_movable_pfn[nid] =
3922 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
3925 /* Any regular memory on that node ? */
3926 static void check_for_regular_memory(pg_data_t *pgdat)
3928 #ifdef CONFIG_HIGHMEM
3929 enum zone_type zone_type;
3931 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) {
3932 struct zone *zone = &pgdat->node_zones[zone_type];
3933 if (zone->present_pages)
3934 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY);
3936 #endif
3940 * free_area_init_nodes - Initialise all pg_data_t and zone data
3941 * @max_zone_pfn: an array of max PFNs for each zone
3943 * This will call free_area_init_node() for each active node in the system.
3944 * Using the page ranges provided by add_active_range(), the size of each
3945 * zone in each node and their holes is calculated. If the maximum PFN
3946 * between two adjacent zones match, it is assumed that the zone is empty.
3947 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
3948 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
3949 * starts where the previous one ended. For example, ZONE_DMA32 starts
3950 * at arch_max_dma_pfn.
3952 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
3954 unsigned long nid;
3955 int i;
3957 /* Sort early_node_map as initialisation assumes it is sorted */
3958 sort_node_map();
3960 /* Record where the zone boundaries are */
3961 memset(arch_zone_lowest_possible_pfn, 0,
3962 sizeof(arch_zone_lowest_possible_pfn));
3963 memset(arch_zone_highest_possible_pfn, 0,
3964 sizeof(arch_zone_highest_possible_pfn));
3965 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
3966 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
3967 for (i = 1; i < MAX_NR_ZONES; i++) {
3968 if (i == ZONE_MOVABLE)
3969 continue;
3970 arch_zone_lowest_possible_pfn[i] =
3971 arch_zone_highest_possible_pfn[i-1];
3972 arch_zone_highest_possible_pfn[i] =
3973 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
3975 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
3976 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
3978 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
3979 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
3980 find_zone_movable_pfns_for_nodes(zone_movable_pfn);
3982 /* Print out the zone ranges */
3983 printk("Zone PFN ranges:\n");
3984 for (i = 0; i < MAX_NR_ZONES; i++) {
3985 if (i == ZONE_MOVABLE)
3986 continue;
3987 printk(" %-8s %0#10lx -> %0#10lx\n",
3988 zone_names[i],
3989 arch_zone_lowest_possible_pfn[i],
3990 arch_zone_highest_possible_pfn[i]);
3993 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
3994 printk("Movable zone start PFN for each node\n");
3995 for (i = 0; i < MAX_NUMNODES; i++) {
3996 if (zone_movable_pfn[i])
3997 printk(" Node %d: %lu\n", i, zone_movable_pfn[i]);
4000 /* Print out the early_node_map[] */
4001 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries);
4002 for (i = 0; i < nr_nodemap_entries; i++)
4003 printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map[i].nid,
4004 early_node_map[i].start_pfn,
4005 early_node_map[i].end_pfn);
4007 /* Initialise every node */
4008 mminit_verify_pageflags_layout();
4009 setup_nr_node_ids();
4010 for_each_online_node(nid) {
4011 pg_data_t *pgdat = NODE_DATA(nid);
4012 free_area_init_node(nid, NULL,
4013 find_min_pfn_for_node(nid), NULL);
4015 /* Any memory on that node */
4016 if (pgdat->node_present_pages)
4017 node_set_state(nid, N_HIGH_MEMORY);
4018 check_for_regular_memory(pgdat);
4022 static int __init cmdline_parse_core(char *p, unsigned long *core)
4024 unsigned long long coremem;
4025 if (!p)
4026 return -EINVAL;
4028 coremem = memparse(p, &p);
4029 *core = coremem >> PAGE_SHIFT;
4031 /* Paranoid check that UL is enough for the coremem value */
4032 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
4034 return 0;
4038 * kernelcore=size sets the amount of memory for use for allocations that
4039 * cannot be reclaimed or migrated.
4041 static int __init cmdline_parse_kernelcore(char *p)
4043 return cmdline_parse_core(p, &required_kernelcore);
4047 * movablecore=size sets the amount of memory for use for allocations that
4048 * can be reclaimed or migrated.
4050 static int __init cmdline_parse_movablecore(char *p)
4052 return cmdline_parse_core(p, &required_movablecore);
4055 early_param("kernelcore", cmdline_parse_kernelcore);
4056 early_param("movablecore", cmdline_parse_movablecore);
4058 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4061 * set_dma_reserve - set the specified number of pages reserved in the first zone
4062 * @new_dma_reserve: The number of pages to mark reserved
4064 * The per-cpu batchsize and zone watermarks are determined by present_pages.
4065 * In the DMA zone, a significant percentage may be consumed by kernel image
4066 * and other unfreeable allocations which can skew the watermarks badly. This
4067 * function may optionally be used to account for unfreeable pages in the
4068 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
4069 * smaller per-cpu batchsize.
4071 void __init set_dma_reserve(unsigned long new_dma_reserve)
4073 dma_reserve = new_dma_reserve;
4076 #ifndef CONFIG_NEED_MULTIPLE_NODES
4077 struct pglist_data __refdata contig_page_data = { .bdata = &bootmem_node_data[0] };
4078 EXPORT_SYMBOL(contig_page_data);
4079 #endif
4081 void __init free_area_init(unsigned long *zones_size)
4083 free_area_init_node(0, zones_size,
4084 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
4087 static int page_alloc_cpu_notify(struct notifier_block *self,
4088 unsigned long action, void *hcpu)
4090 int cpu = (unsigned long)hcpu;
4092 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
4093 drain_pages(cpu);
4096 * Spill the event counters of the dead processor
4097 * into the current processors event counters.
4098 * This artificially elevates the count of the current
4099 * processor.
4101 vm_events_fold_cpu(cpu);
4104 * Zero the differential counters of the dead processor
4105 * so that the vm statistics are consistent.
4107 * This is only okay since the processor is dead and cannot
4108 * race with what we are doing.
4110 refresh_cpu_vm_stats(cpu);
4112 return NOTIFY_OK;
4115 void __init page_alloc_init(void)
4117 hotcpu_notifier(page_alloc_cpu_notify, 0);
4121 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
4122 * or min_free_kbytes changes.
4124 static void calculate_totalreserve_pages(void)
4126 struct pglist_data *pgdat;
4127 unsigned long reserve_pages = 0;
4128 enum zone_type i, j;
4130 for_each_online_pgdat(pgdat) {
4131 for (i = 0; i < MAX_NR_ZONES; i++) {
4132 struct zone *zone = pgdat->node_zones + i;
4133 unsigned long max = 0;
4135 /* Find valid and maximum lowmem_reserve in the zone */
4136 for (j = i; j < MAX_NR_ZONES; j++) {
4137 if (zone->lowmem_reserve[j] > max)
4138 max = zone->lowmem_reserve[j];
4141 /* we treat pages_high as reserved pages. */
4142 max += zone->pages_high;
4144 if (max > zone->present_pages)
4145 max = zone->present_pages;
4146 reserve_pages += max;
4149 totalreserve_pages = reserve_pages;
4153 * setup_per_zone_lowmem_reserve - called whenever
4154 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
4155 * has a correct pages reserved value, so an adequate number of
4156 * pages are left in the zone after a successful __alloc_pages().
4158 static void setup_per_zone_lowmem_reserve(void)
4160 struct pglist_data *pgdat;
4161 enum zone_type j, idx;
4163 for_each_online_pgdat(pgdat) {
4164 for (j = 0; j < MAX_NR_ZONES; j++) {
4165 struct zone *zone = pgdat->node_zones + j;
4166 unsigned long present_pages = zone->present_pages;
4168 zone->lowmem_reserve[j] = 0;
4170 idx = j;
4171 while (idx) {
4172 struct zone *lower_zone;
4174 idx--;
4176 if (sysctl_lowmem_reserve_ratio[idx] < 1)
4177 sysctl_lowmem_reserve_ratio[idx] = 1;
4179 lower_zone = pgdat->node_zones + idx;
4180 lower_zone->lowmem_reserve[j] = present_pages /
4181 sysctl_lowmem_reserve_ratio[idx];
4182 present_pages += lower_zone->present_pages;
4187 /* update totalreserve_pages */
4188 calculate_totalreserve_pages();
4192 * setup_per_zone_pages_min - called when min_free_kbytes changes.
4194 * Ensures that the pages_{min,low,high} values for each zone are set correctly
4195 * with respect to min_free_kbytes.
4197 void setup_per_zone_pages_min(void)
4199 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
4200 unsigned long lowmem_pages = 0;
4201 struct zone *zone;
4202 unsigned long flags;
4204 /* Calculate total number of !ZONE_HIGHMEM pages */
4205 for_each_zone(zone) {
4206 if (!is_highmem(zone))
4207 lowmem_pages += zone->present_pages;
4210 for_each_zone(zone) {
4211 u64 tmp;
4213 spin_lock_irqsave(&zone->lru_lock, flags);
4214 tmp = (u64)pages_min * zone->present_pages;
4215 do_div(tmp, lowmem_pages);
4216 if (is_highmem(zone)) {
4218 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
4219 * need highmem pages, so cap pages_min to a small
4220 * value here.
4222 * The (pages_high-pages_low) and (pages_low-pages_min)
4223 * deltas controls asynch page reclaim, and so should
4224 * not be capped for highmem.
4226 int min_pages;
4228 min_pages = zone->present_pages / 1024;
4229 if (min_pages < SWAP_CLUSTER_MAX)
4230 min_pages = SWAP_CLUSTER_MAX;
4231 if (min_pages > 128)
4232 min_pages = 128;
4233 zone->pages_min = min_pages;
4234 } else {
4236 * If it's a lowmem zone, reserve a number of pages
4237 * proportionate to the zone's size.
4239 zone->pages_min = tmp;
4242 zone->pages_low = zone->pages_min + (tmp >> 2);
4243 zone->pages_high = zone->pages_min + (tmp >> 1);
4244 setup_zone_migrate_reserve(zone);
4245 spin_unlock_irqrestore(&zone->lru_lock, flags);
4248 /* update totalreserve_pages */
4249 calculate_totalreserve_pages();
4253 * Initialise min_free_kbytes.
4255 * For small machines we want it small (128k min). For large machines
4256 * we want it large (64MB max). But it is not linear, because network
4257 * bandwidth does not increase linearly with machine size. We use
4259 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
4260 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
4262 * which yields
4264 * 16MB: 512k
4265 * 32MB: 724k
4266 * 64MB: 1024k
4267 * 128MB: 1448k
4268 * 256MB: 2048k
4269 * 512MB: 2896k
4270 * 1024MB: 4096k
4271 * 2048MB: 5792k
4272 * 4096MB: 8192k
4273 * 8192MB: 11584k
4274 * 16384MB: 16384k
4276 static int __init init_per_zone_pages_min(void)
4278 unsigned long lowmem_kbytes;
4280 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
4282 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
4283 if (min_free_kbytes < 128)
4284 min_free_kbytes = 128;
4285 if (min_free_kbytes > 65536)
4286 min_free_kbytes = 65536;
4287 setup_per_zone_pages_min();
4288 setup_per_zone_lowmem_reserve();
4289 return 0;
4291 module_init(init_per_zone_pages_min)
4294 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
4295 * that we can call two helper functions whenever min_free_kbytes
4296 * changes.
4298 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
4299 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4301 proc_dointvec(table, write, file, buffer, length, ppos);
4302 if (write)
4303 setup_per_zone_pages_min();
4304 return 0;
4307 #ifdef CONFIG_NUMA
4308 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
4309 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4311 struct zone *zone;
4312 int rc;
4314 rc = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
4315 if (rc)
4316 return rc;
4318 for_each_zone(zone)
4319 zone->min_unmapped_pages = (zone->present_pages *
4320 sysctl_min_unmapped_ratio) / 100;
4321 return 0;
4324 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
4325 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4327 struct zone *zone;
4328 int rc;
4330 rc = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
4331 if (rc)
4332 return rc;
4334 for_each_zone(zone)
4335 zone->min_slab_pages = (zone->present_pages *
4336 sysctl_min_slab_ratio) / 100;
4337 return 0;
4339 #endif
4342 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
4343 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
4344 * whenever sysctl_lowmem_reserve_ratio changes.
4346 * The reserve ratio obviously has absolutely no relation with the
4347 * pages_min watermarks. The lowmem reserve ratio can only make sense
4348 * if in function of the boot time zone sizes.
4350 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
4351 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4353 proc_dointvec_minmax(table, write, file, buffer, length, ppos);
4354 setup_per_zone_lowmem_reserve();
4355 return 0;
4359 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
4360 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
4361 * can have before it gets flushed back to buddy allocator.
4364 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
4365 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4367 struct zone *zone;
4368 unsigned int cpu;
4369 int ret;
4371 ret = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
4372 if (!write || (ret == -EINVAL))
4373 return ret;
4374 for_each_zone(zone) {
4375 for_each_online_cpu(cpu) {
4376 unsigned long high;
4377 high = zone->present_pages / percpu_pagelist_fraction;
4378 setup_pagelist_highmark(zone_pcp(zone, cpu), high);
4381 return 0;
4384 int hashdist = HASHDIST_DEFAULT;
4386 #ifdef CONFIG_NUMA
4387 static int __init set_hashdist(char *str)
4389 if (!str)
4390 return 0;
4391 hashdist = simple_strtoul(str, &str, 0);
4392 return 1;
4394 __setup("hashdist=", set_hashdist);
4395 #endif
4398 * allocate a large system hash table from bootmem
4399 * - it is assumed that the hash table must contain an exact power-of-2
4400 * quantity of entries
4401 * - limit is the number of hash buckets, not the total allocation size
4403 void *__init alloc_large_system_hash(const char *tablename,
4404 unsigned long bucketsize,
4405 unsigned long numentries,
4406 int scale,
4407 int flags,
4408 unsigned int *_hash_shift,
4409 unsigned int *_hash_mask,
4410 unsigned long limit)
4412 unsigned long long max = limit;
4413 unsigned long log2qty, size;
4414 void *table = NULL;
4416 /* allow the kernel cmdline to have a say */
4417 if (!numentries) {
4418 /* round applicable memory size up to nearest megabyte */
4419 numentries = nr_kernel_pages;
4420 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
4421 numentries >>= 20 - PAGE_SHIFT;
4422 numentries <<= 20 - PAGE_SHIFT;
4424 /* limit to 1 bucket per 2^scale bytes of low memory */
4425 if (scale > PAGE_SHIFT)
4426 numentries >>= (scale - PAGE_SHIFT);
4427 else
4428 numentries <<= (PAGE_SHIFT - scale);
4430 /* Make sure we've got at least a 0-order allocation.. */
4431 if (unlikely((numentries * bucketsize) < PAGE_SIZE))
4432 numentries = PAGE_SIZE / bucketsize;
4434 numentries = roundup_pow_of_two(numentries);
4436 /* limit allocation size to 1/16 total memory by default */
4437 if (max == 0) {
4438 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
4439 do_div(max, bucketsize);
4442 if (numentries > max)
4443 numentries = max;
4445 log2qty = ilog2(numentries);
4447 do {
4448 size = bucketsize << log2qty;
4449 if (flags & HASH_EARLY)
4450 table = alloc_bootmem_nopanic(size);
4451 else if (hashdist)
4452 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
4453 else {
4454 unsigned long order = get_order(size);
4455 table = (void*) __get_free_pages(GFP_ATOMIC, order);
4457 * If bucketsize is not a power-of-two, we may free
4458 * some pages at the end of hash table.
4460 if (table) {
4461 unsigned long alloc_end = (unsigned long)table +
4462 (PAGE_SIZE << order);
4463 unsigned long used = (unsigned long)table +
4464 PAGE_ALIGN(size);
4465 split_page(virt_to_page(table), order);
4466 while (used < alloc_end) {
4467 free_page(used);
4468 used += PAGE_SIZE;
4472 } while (!table && size > PAGE_SIZE && --log2qty);
4474 if (!table)
4475 panic("Failed to allocate %s hash table\n", tablename);
4477 printk(KERN_INFO "%s hash table entries: %d (order: %d, %lu bytes)\n",
4478 tablename,
4479 (1U << log2qty),
4480 ilog2(size) - PAGE_SHIFT,
4481 size);
4483 if (_hash_shift)
4484 *_hash_shift = log2qty;
4485 if (_hash_mask)
4486 *_hash_mask = (1 << log2qty) - 1;
4488 return table;
4491 #ifdef CONFIG_OUT_OF_LINE_PFN_TO_PAGE
4492 struct page *pfn_to_page(unsigned long pfn)
4494 return __pfn_to_page(pfn);
4496 unsigned long page_to_pfn(struct page *page)
4498 return __page_to_pfn(page);
4500 EXPORT_SYMBOL(pfn_to_page);
4501 EXPORT_SYMBOL(page_to_pfn);
4502 #endif /* CONFIG_OUT_OF_LINE_PFN_TO_PAGE */
4504 /* Return a pointer to the bitmap storing bits affecting a block of pages */
4505 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
4506 unsigned long pfn)
4508 #ifdef CONFIG_SPARSEMEM
4509 return __pfn_to_section(pfn)->pageblock_flags;
4510 #else
4511 return zone->pageblock_flags;
4512 #endif /* CONFIG_SPARSEMEM */
4515 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
4517 #ifdef CONFIG_SPARSEMEM
4518 pfn &= (PAGES_PER_SECTION-1);
4519 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
4520 #else
4521 pfn = pfn - zone->zone_start_pfn;
4522 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
4523 #endif /* CONFIG_SPARSEMEM */
4527 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
4528 * @page: The page within the block of interest
4529 * @start_bitidx: The first bit of interest to retrieve
4530 * @end_bitidx: The last bit of interest
4531 * returns pageblock_bits flags
4533 unsigned long get_pageblock_flags_group(struct page *page,
4534 int start_bitidx, int end_bitidx)
4536 struct zone *zone;
4537 unsigned long *bitmap;
4538 unsigned long pfn, bitidx;
4539 unsigned long flags = 0;
4540 unsigned long value = 1;
4542 zone = page_zone(page);
4543 pfn = page_to_pfn(page);
4544 bitmap = get_pageblock_bitmap(zone, pfn);
4545 bitidx = pfn_to_bitidx(zone, pfn);
4547 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
4548 if (test_bit(bitidx + start_bitidx, bitmap))
4549 flags |= value;
4551 return flags;
4555 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
4556 * @page: The page within the block of interest
4557 * @start_bitidx: The first bit of interest
4558 * @end_bitidx: The last bit of interest
4559 * @flags: The flags to set
4561 void set_pageblock_flags_group(struct page *page, unsigned long flags,
4562 int start_bitidx, int end_bitidx)
4564 struct zone *zone;
4565 unsigned long *bitmap;
4566 unsigned long pfn, bitidx;
4567 unsigned long value = 1;
4569 zone = page_zone(page);
4570 pfn = page_to_pfn(page);
4571 bitmap = get_pageblock_bitmap(zone, pfn);
4572 bitidx = pfn_to_bitidx(zone, pfn);
4573 VM_BUG_ON(pfn < zone->zone_start_pfn);
4574 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages);
4576 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
4577 if (flags & value)
4578 __set_bit(bitidx + start_bitidx, bitmap);
4579 else
4580 __clear_bit(bitidx + start_bitidx, bitmap);
4584 * This is designed as sub function...plz see page_isolation.c also.
4585 * set/clear page block's type to be ISOLATE.
4586 * page allocater never alloc memory from ISOLATE block.
4589 int set_migratetype_isolate(struct page *page)
4591 struct zone *zone;
4592 unsigned long flags;
4593 int ret = -EBUSY;
4595 zone = page_zone(page);
4596 spin_lock_irqsave(&zone->lock, flags);
4598 * In future, more migrate types will be able to be isolation target.
4600 if (get_pageblock_migratetype(page) != MIGRATE_MOVABLE)
4601 goto out;
4602 set_pageblock_migratetype(page, MIGRATE_ISOLATE);
4603 move_freepages_block(zone, page, MIGRATE_ISOLATE);
4604 ret = 0;
4605 out:
4606 spin_unlock_irqrestore(&zone->lock, flags);
4607 if (!ret)
4608 drain_all_pages();
4609 return ret;
4612 void unset_migratetype_isolate(struct page *page)
4614 struct zone *zone;
4615 unsigned long flags;
4616 zone = page_zone(page);
4617 spin_lock_irqsave(&zone->lock, flags);
4618 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
4619 goto out;
4620 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4621 move_freepages_block(zone, page, MIGRATE_MOVABLE);
4622 out:
4623 spin_unlock_irqrestore(&zone->lock, flags);
4626 #ifdef CONFIG_MEMORY_HOTREMOVE
4628 * All pages in the range must be isolated before calling this.
4630 void
4631 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
4633 struct page *page;
4634 struct zone *zone;
4635 int order, i;
4636 unsigned long pfn;
4637 unsigned long flags;
4638 /* find the first valid pfn */
4639 for (pfn = start_pfn; pfn < end_pfn; pfn++)
4640 if (pfn_valid(pfn))
4641 break;
4642 if (pfn == end_pfn)
4643 return;
4644 zone = page_zone(pfn_to_page(pfn));
4645 spin_lock_irqsave(&zone->lock, flags);
4646 pfn = start_pfn;
4647 while (pfn < end_pfn) {
4648 if (!pfn_valid(pfn)) {
4649 pfn++;
4650 continue;
4652 page = pfn_to_page(pfn);
4653 BUG_ON(page_count(page));
4654 BUG_ON(!PageBuddy(page));
4655 order = page_order(page);
4656 #ifdef CONFIG_DEBUG_VM
4657 printk(KERN_INFO "remove from free list %lx %d %lx\n",
4658 pfn, 1 << order, end_pfn);
4659 #endif
4660 list_del(&page->lru);
4661 rmv_page_order(page);
4662 zone->free_area[order].nr_free--;
4663 __mod_zone_page_state(zone, NR_FREE_PAGES,
4664 - (1UL << order));
4665 for (i = 0; i < (1 << order); i++)
4666 SetPageReserved((page+i));
4667 pfn += (1 << order);
4669 spin_unlock_irqrestore(&zone->lock, flags);
4671 #endif