e1000e: commit speed/duplex changes for m88 PHY
[linux-2.6.git] / mm / page_alloc.c
blobd8ac01474563954fed58a1f87e2249c99827c430
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/page_cgroup.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 printk(KERN_EMERG "Bad page state in process '%s'\n" KERN_EMERG
227 "page:%p flags:0x%0*lx mapping:%p mapcount:%d count:%d\n",
228 current->comm, page, (int)(2*sizeof(unsigned long)),
229 (unsigned long)page->flags, page->mapping,
230 page_mapcount(page), page_count(page));
232 printk(KERN_EMERG "Trying to fix it up, but a reboot is needed\n"
233 KERN_EMERG "Backtrace:\n");
234 dump_stack();
235 page->flags &= ~PAGE_FLAGS_CLEAR_WHEN_BAD;
236 set_page_count(page, 0);
237 reset_page_mapcount(page);
238 page->mapping = NULL;
239 add_taint(TAINT_BAD_PAGE);
243 * Higher-order pages are called "compound pages". They are structured thusly:
245 * The first PAGE_SIZE page is called the "head page".
247 * The remaining PAGE_SIZE pages are called "tail pages".
249 * All pages have PG_compound set. All pages have their ->private pointing at
250 * the head page (even the head page has this).
252 * The first tail page's ->lru.next holds the address of the compound page's
253 * put_page() function. Its ->lru.prev holds the order of allocation.
254 * This usage means that zero-order pages may not be compound.
257 static void free_compound_page(struct page *page)
259 __free_pages_ok(page, compound_order(page));
262 void prep_compound_page(struct page *page, unsigned long order)
264 int i;
265 int nr_pages = 1 << order;
267 set_compound_page_dtor(page, free_compound_page);
268 set_compound_order(page, order);
269 __SetPageHead(page);
270 for (i = 1; i < nr_pages; i++) {
271 struct page *p = page + i;
273 __SetPageTail(p);
274 p->first_page = page;
278 #ifdef CONFIG_HUGETLBFS
279 void prep_compound_gigantic_page(struct page *page, unsigned long order)
281 int i;
282 int nr_pages = 1 << order;
283 struct page *p = page + 1;
285 set_compound_page_dtor(page, free_compound_page);
286 set_compound_order(page, order);
287 __SetPageHead(page);
288 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
289 __SetPageTail(p);
290 p->first_page = page;
293 #endif
295 static void destroy_compound_page(struct page *page, unsigned long order)
297 int i;
298 int nr_pages = 1 << order;
300 if (unlikely(compound_order(page) != order))
301 bad_page(page);
303 if (unlikely(!PageHead(page)))
304 bad_page(page);
305 __ClearPageHead(page);
306 for (i = 1; i < nr_pages; i++) {
307 struct page *p = page + i;
309 if (unlikely(!PageTail(p) |
310 (p->first_page != page)))
311 bad_page(page);
312 __ClearPageTail(p);
316 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
318 int i;
321 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
322 * and __GFP_HIGHMEM from hard or soft interrupt context.
324 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
325 for (i = 0; i < (1 << order); i++)
326 clear_highpage(page + i);
329 static inline void set_page_order(struct page *page, int order)
331 set_page_private(page, order);
332 __SetPageBuddy(page);
335 static inline void rmv_page_order(struct page *page)
337 __ClearPageBuddy(page);
338 set_page_private(page, 0);
342 * Locate the struct page for both the matching buddy in our
343 * pair (buddy1) and the combined O(n+1) page they form (page).
345 * 1) Any buddy B1 will have an order O twin B2 which satisfies
346 * the following equation:
347 * B2 = B1 ^ (1 << O)
348 * For example, if the starting buddy (buddy2) is #8 its order
349 * 1 buddy is #10:
350 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
352 * 2) Any buddy B will have an order O+1 parent P which
353 * satisfies the following equation:
354 * P = B & ~(1 << O)
356 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
358 static inline struct page *
359 __page_find_buddy(struct page *page, unsigned long page_idx, unsigned int order)
361 unsigned long buddy_idx = page_idx ^ (1 << order);
363 return page + (buddy_idx - page_idx);
366 static inline unsigned long
367 __find_combined_index(unsigned long page_idx, unsigned int order)
369 return (page_idx & ~(1 << order));
373 * This function checks whether a page is free && is the buddy
374 * we can do coalesce a page and its buddy if
375 * (a) the buddy is not in a hole &&
376 * (b) the buddy is in the buddy system &&
377 * (c) a page and its buddy have the same order &&
378 * (d) a page and its buddy are in the same zone.
380 * For recording whether a page is in the buddy system, we use PG_buddy.
381 * Setting, clearing, and testing PG_buddy is serialized by zone->lock.
383 * For recording page's order, we use page_private(page).
385 static inline int page_is_buddy(struct page *page, struct page *buddy,
386 int order)
388 if (!pfn_valid_within(page_to_pfn(buddy)))
389 return 0;
391 if (page_zone_id(page) != page_zone_id(buddy))
392 return 0;
394 if (PageBuddy(buddy) && page_order(buddy) == order) {
395 BUG_ON(page_count(buddy) != 0);
396 return 1;
398 return 0;
402 * Freeing function for a buddy system allocator.
404 * The concept of a buddy system is to maintain direct-mapped table
405 * (containing bit values) for memory blocks of various "orders".
406 * The bottom level table contains the map for the smallest allocatable
407 * units of memory (here, pages), and each level above it describes
408 * pairs of units from the levels below, hence, "buddies".
409 * At a high level, all that happens here is marking the table entry
410 * at the bottom level available, and propagating the changes upward
411 * as necessary, plus some accounting needed to play nicely with other
412 * parts of the VM system.
413 * At each level, we keep a list of pages, which are heads of continuous
414 * free pages of length of (1 << order) and marked with PG_buddy. Page's
415 * order is recorded in page_private(page) field.
416 * So when we are allocating or freeing one, we can derive the state of the
417 * other. That is, if we allocate a small block, and both were
418 * free, the remainder of the region must be split into blocks.
419 * If a block is freed, and its buddy is also free, then this
420 * triggers coalescing into a block of larger size.
422 * -- wli
425 static inline void __free_one_page(struct page *page,
426 struct zone *zone, unsigned int order)
428 unsigned long page_idx;
429 int order_size = 1 << order;
430 int migratetype = get_pageblock_migratetype(page);
432 if (unlikely(PageCompound(page)))
433 destroy_compound_page(page, order);
435 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
437 VM_BUG_ON(page_idx & (order_size - 1));
438 VM_BUG_ON(bad_range(zone, page));
440 __mod_zone_page_state(zone, NR_FREE_PAGES, order_size);
441 while (order < MAX_ORDER-1) {
442 unsigned long combined_idx;
443 struct page *buddy;
445 buddy = __page_find_buddy(page, page_idx, order);
446 if (!page_is_buddy(page, buddy, order))
447 break;
449 /* Our buddy is free, merge with it and move up one order. */
450 list_del(&buddy->lru);
451 zone->free_area[order].nr_free--;
452 rmv_page_order(buddy);
453 combined_idx = __find_combined_index(page_idx, order);
454 page = page + (combined_idx - page_idx);
455 page_idx = combined_idx;
456 order++;
458 set_page_order(page, order);
459 list_add(&page->lru,
460 &zone->free_area[order].free_list[migratetype]);
461 zone->free_area[order].nr_free++;
464 static inline int free_pages_check(struct page *page)
466 free_page_mlock(page);
467 if (unlikely(page_mapcount(page) |
468 (page->mapping != NULL) |
469 (page_count(page) != 0) |
470 (page->flags & PAGE_FLAGS_CHECK_AT_FREE)))
471 bad_page(page);
472 if (PageDirty(page))
473 __ClearPageDirty(page);
474 if (PageSwapBacked(page))
475 __ClearPageSwapBacked(page);
477 * For now, we report if PG_reserved was found set, but do not
478 * clear it, and do not free the page. But we shall soon need
479 * to do more, for when the ZERO_PAGE count wraps negative.
481 return PageReserved(page);
485 * Frees a list of pages.
486 * Assumes all pages on list are in same zone, and of same order.
487 * count is the number of pages to free.
489 * If the zone was previously in an "all pages pinned" state then look to
490 * see if this freeing clears that state.
492 * And clear the zone's pages_scanned counter, to hold off the "all pages are
493 * pinned" detection logic.
495 static void free_pages_bulk(struct zone *zone, int count,
496 struct list_head *list, int order)
498 spin_lock(&zone->lock);
499 zone_clear_flag(zone, ZONE_ALL_UNRECLAIMABLE);
500 zone->pages_scanned = 0;
501 while (count--) {
502 struct page *page;
504 VM_BUG_ON(list_empty(list));
505 page = list_entry(list->prev, struct page, lru);
506 /* have to delete it as __free_one_page list manipulates */
507 list_del(&page->lru);
508 __free_one_page(page, zone, order);
510 spin_unlock(&zone->lock);
513 static void free_one_page(struct zone *zone, struct page *page, int order)
515 spin_lock(&zone->lock);
516 zone_clear_flag(zone, ZONE_ALL_UNRECLAIMABLE);
517 zone->pages_scanned = 0;
518 __free_one_page(page, zone, order);
519 spin_unlock(&zone->lock);
522 static void __free_pages_ok(struct page *page, unsigned int order)
524 unsigned long flags;
525 int i;
526 int reserved = 0;
528 for (i = 0 ; i < (1 << order) ; ++i)
529 reserved += free_pages_check(page + i);
530 if (reserved)
531 return;
533 if (!PageHighMem(page)) {
534 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
535 debug_check_no_obj_freed(page_address(page),
536 PAGE_SIZE << order);
538 arch_free_page(page, order);
539 kernel_map_pages(page, 1 << order, 0);
541 local_irq_save(flags);
542 __count_vm_events(PGFREE, 1 << order);
543 free_one_page(page_zone(page), page, order);
544 local_irq_restore(flags);
548 * permit the bootmem allocator to evade page validation on high-order frees
550 void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
552 if (order == 0) {
553 __ClearPageReserved(page);
554 set_page_count(page, 0);
555 set_page_refcounted(page);
556 __free_page(page);
557 } else {
558 int loop;
560 prefetchw(page);
561 for (loop = 0; loop < BITS_PER_LONG; loop++) {
562 struct page *p = &page[loop];
564 if (loop + 1 < BITS_PER_LONG)
565 prefetchw(p + 1);
566 __ClearPageReserved(p);
567 set_page_count(p, 0);
570 set_page_refcounted(page);
571 __free_pages(page, order);
577 * The order of subdivision here is critical for the IO subsystem.
578 * Please do not alter this order without good reasons and regression
579 * testing. Specifically, as large blocks of memory are subdivided,
580 * the order in which smaller blocks are delivered depends on the order
581 * they're subdivided in this function. This is the primary factor
582 * influencing the order in which pages are delivered to the IO
583 * subsystem according to empirical testing, and this is also justified
584 * by considering the behavior of a buddy system containing a single
585 * large block of memory acted on by a series of small allocations.
586 * This behavior is a critical factor in sglist merging's success.
588 * -- wli
590 static inline void expand(struct zone *zone, struct page *page,
591 int low, int high, struct free_area *area,
592 int migratetype)
594 unsigned long size = 1 << high;
596 while (high > low) {
597 area--;
598 high--;
599 size >>= 1;
600 VM_BUG_ON(bad_range(zone, &page[size]));
601 list_add(&page[size].lru, &area->free_list[migratetype]);
602 area->nr_free++;
603 set_page_order(&page[size], high);
608 * This page is about to be returned from the page allocator
610 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
612 if (unlikely(page_mapcount(page) |
613 (page->mapping != NULL) |
614 (page_count(page) != 0) |
615 (page->flags & PAGE_FLAGS_CHECK_AT_PREP)))
616 bad_page(page);
619 * For now, we report if PG_reserved was found set, but do not
620 * clear it, and do not allocate the page: as a safety net.
622 if (PageReserved(page))
623 return 1;
625 page->flags &= ~(1 << PG_uptodate | 1 << PG_error | 1 << PG_reclaim |
626 1 << PG_referenced | 1 << PG_arch_1 |
627 1 << PG_owner_priv_1 | 1 << PG_mappedtodisk
628 #ifdef CONFIG_UNEVICTABLE_LRU
629 | 1 << PG_mlocked
630 #endif
632 set_page_private(page, 0);
633 set_page_refcounted(page);
635 arch_alloc_page(page, order);
636 kernel_map_pages(page, 1 << order, 1);
638 if (gfp_flags & __GFP_ZERO)
639 prep_zero_page(page, order, gfp_flags);
641 if (order && (gfp_flags & __GFP_COMP))
642 prep_compound_page(page, order);
644 return 0;
648 * Go through the free lists for the given migratetype and remove
649 * the smallest available page from the freelists
651 static struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
652 int migratetype)
654 unsigned int current_order;
655 struct free_area * area;
656 struct page *page;
658 /* Find a page of the appropriate size in the preferred list */
659 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
660 area = &(zone->free_area[current_order]);
661 if (list_empty(&area->free_list[migratetype]))
662 continue;
664 page = list_entry(area->free_list[migratetype].next,
665 struct page, lru);
666 list_del(&page->lru);
667 rmv_page_order(page);
668 area->nr_free--;
669 __mod_zone_page_state(zone, NR_FREE_PAGES, - (1UL << order));
670 expand(zone, page, order, current_order, area, migratetype);
671 return page;
674 return NULL;
679 * This array describes the order lists are fallen back to when
680 * the free lists for the desirable migrate type are depleted
682 static int fallbacks[MIGRATE_TYPES][MIGRATE_TYPES-1] = {
683 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
684 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
685 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
686 [MIGRATE_RESERVE] = { MIGRATE_RESERVE, MIGRATE_RESERVE, MIGRATE_RESERVE }, /* Never used */
690 * Move the free pages in a range to the free lists of the requested type.
691 * Note that start_page and end_pages are not aligned on a pageblock
692 * boundary. If alignment is required, use move_freepages_block()
694 static int move_freepages(struct zone *zone,
695 struct page *start_page, struct page *end_page,
696 int migratetype)
698 struct page *page;
699 unsigned long order;
700 int pages_moved = 0;
702 #ifndef CONFIG_HOLES_IN_ZONE
704 * page_zone is not safe to call in this context when
705 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
706 * anyway as we check zone boundaries in move_freepages_block().
707 * Remove at a later date when no bug reports exist related to
708 * grouping pages by mobility
710 BUG_ON(page_zone(start_page) != page_zone(end_page));
711 #endif
713 for (page = start_page; page <= end_page;) {
714 /* Make sure we are not inadvertently changing nodes */
715 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
717 if (!pfn_valid_within(page_to_pfn(page))) {
718 page++;
719 continue;
722 if (!PageBuddy(page)) {
723 page++;
724 continue;
727 order = page_order(page);
728 list_del(&page->lru);
729 list_add(&page->lru,
730 &zone->free_area[order].free_list[migratetype]);
731 page += 1 << order;
732 pages_moved += 1 << order;
735 return pages_moved;
738 static int move_freepages_block(struct zone *zone, struct page *page,
739 int migratetype)
741 unsigned long start_pfn, end_pfn;
742 struct page *start_page, *end_page;
744 start_pfn = page_to_pfn(page);
745 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
746 start_page = pfn_to_page(start_pfn);
747 end_page = start_page + pageblock_nr_pages - 1;
748 end_pfn = start_pfn + pageblock_nr_pages - 1;
750 /* Do not cross zone boundaries */
751 if (start_pfn < zone->zone_start_pfn)
752 start_page = page;
753 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
754 return 0;
756 return move_freepages(zone, start_page, end_page, migratetype);
759 /* Remove an element from the buddy allocator from the fallback list */
760 static struct page *__rmqueue_fallback(struct zone *zone, int order,
761 int start_migratetype)
763 struct free_area * area;
764 int current_order;
765 struct page *page;
766 int migratetype, i;
768 /* Find the largest possible block of pages in the other list */
769 for (current_order = MAX_ORDER-1; current_order >= order;
770 --current_order) {
771 for (i = 0; i < MIGRATE_TYPES - 1; i++) {
772 migratetype = fallbacks[start_migratetype][i];
774 /* MIGRATE_RESERVE handled later if necessary */
775 if (migratetype == MIGRATE_RESERVE)
776 continue;
778 area = &(zone->free_area[current_order]);
779 if (list_empty(&area->free_list[migratetype]))
780 continue;
782 page = list_entry(area->free_list[migratetype].next,
783 struct page, lru);
784 area->nr_free--;
787 * If breaking a large block of pages, move all free
788 * pages to the preferred allocation list. If falling
789 * back for a reclaimable kernel allocation, be more
790 * agressive about taking ownership of free pages
792 if (unlikely(current_order >= (pageblock_order >> 1)) ||
793 start_migratetype == MIGRATE_RECLAIMABLE) {
794 unsigned long pages;
795 pages = move_freepages_block(zone, page,
796 start_migratetype);
798 /* Claim the whole block if over half of it is free */
799 if (pages >= (1 << (pageblock_order-1)))
800 set_pageblock_migratetype(page,
801 start_migratetype);
803 migratetype = start_migratetype;
806 /* Remove the page from the freelists */
807 list_del(&page->lru);
808 rmv_page_order(page);
809 __mod_zone_page_state(zone, NR_FREE_PAGES,
810 -(1UL << order));
812 if (current_order == pageblock_order)
813 set_pageblock_migratetype(page,
814 start_migratetype);
816 expand(zone, page, order, current_order, area, migratetype);
817 return page;
821 /* Use MIGRATE_RESERVE rather than fail an allocation */
822 return __rmqueue_smallest(zone, order, MIGRATE_RESERVE);
826 * Do the hard work of removing an element from the buddy allocator.
827 * Call me with the zone->lock already held.
829 static struct page *__rmqueue(struct zone *zone, unsigned int order,
830 int migratetype)
832 struct page *page;
834 page = __rmqueue_smallest(zone, order, migratetype);
836 if (unlikely(!page))
837 page = __rmqueue_fallback(zone, order, migratetype);
839 return page;
843 * Obtain a specified number of elements from the buddy allocator, all under
844 * a single hold of the lock, for efficiency. Add them to the supplied list.
845 * Returns the number of new pages which were placed at *list.
847 static int rmqueue_bulk(struct zone *zone, unsigned int order,
848 unsigned long count, struct list_head *list,
849 int migratetype)
851 int i;
853 spin_lock(&zone->lock);
854 for (i = 0; i < count; ++i) {
855 struct page *page = __rmqueue(zone, order, migratetype);
856 if (unlikely(page == NULL))
857 break;
860 * Split buddy pages returned by expand() are received here
861 * in physical page order. The page is added to the callers and
862 * list and the list head then moves forward. From the callers
863 * perspective, the linked list is ordered by page number in
864 * some conditions. This is useful for IO devices that can
865 * merge IO requests if the physical pages are ordered
866 * properly.
868 list_add(&page->lru, list);
869 set_page_private(page, migratetype);
870 list = &page->lru;
872 spin_unlock(&zone->lock);
873 return i;
876 #ifdef CONFIG_NUMA
878 * Called from the vmstat counter updater to drain pagesets of this
879 * currently executing processor on remote nodes after they have
880 * expired.
882 * Note that this function must be called with the thread pinned to
883 * a single processor.
885 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
887 unsigned long flags;
888 int to_drain;
890 local_irq_save(flags);
891 if (pcp->count >= pcp->batch)
892 to_drain = pcp->batch;
893 else
894 to_drain = pcp->count;
895 free_pages_bulk(zone, to_drain, &pcp->list, 0);
896 pcp->count -= to_drain;
897 local_irq_restore(flags);
899 #endif
902 * Drain pages of the indicated processor.
904 * The processor must either be the current processor and the
905 * thread pinned to the current processor or a processor that
906 * is not online.
908 static void drain_pages(unsigned int cpu)
910 unsigned long flags;
911 struct zone *zone;
913 for_each_zone(zone) {
914 struct per_cpu_pageset *pset;
915 struct per_cpu_pages *pcp;
917 if (!populated_zone(zone))
918 continue;
920 pset = zone_pcp(zone, cpu);
922 pcp = &pset->pcp;
923 local_irq_save(flags);
924 free_pages_bulk(zone, pcp->count, &pcp->list, 0);
925 pcp->count = 0;
926 local_irq_restore(flags);
931 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
933 void drain_local_pages(void *arg)
935 drain_pages(smp_processor_id());
939 * Spill all the per-cpu pages from all CPUs back into the buddy allocator
941 void drain_all_pages(void)
943 on_each_cpu(drain_local_pages, NULL, 1);
946 #ifdef CONFIG_HIBERNATION
948 void mark_free_pages(struct zone *zone)
950 unsigned long pfn, max_zone_pfn;
951 unsigned long flags;
952 int order, t;
953 struct list_head *curr;
955 if (!zone->spanned_pages)
956 return;
958 spin_lock_irqsave(&zone->lock, flags);
960 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
961 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
962 if (pfn_valid(pfn)) {
963 struct page *page = pfn_to_page(pfn);
965 if (!swsusp_page_is_forbidden(page))
966 swsusp_unset_page_free(page);
969 for_each_migratetype_order(order, t) {
970 list_for_each(curr, &zone->free_area[order].free_list[t]) {
971 unsigned long i;
973 pfn = page_to_pfn(list_entry(curr, struct page, lru));
974 for (i = 0; i < (1UL << order); i++)
975 swsusp_set_page_free(pfn_to_page(pfn + i));
978 spin_unlock_irqrestore(&zone->lock, flags);
980 #endif /* CONFIG_PM */
983 * Free a 0-order page
985 static void free_hot_cold_page(struct page *page, int cold)
987 struct zone *zone = page_zone(page);
988 struct per_cpu_pages *pcp;
989 unsigned long flags;
991 if (PageAnon(page))
992 page->mapping = NULL;
993 if (free_pages_check(page))
994 return;
996 if (!PageHighMem(page)) {
997 debug_check_no_locks_freed(page_address(page), PAGE_SIZE);
998 debug_check_no_obj_freed(page_address(page), PAGE_SIZE);
1000 arch_free_page(page, 0);
1001 kernel_map_pages(page, 1, 0);
1003 pcp = &zone_pcp(zone, get_cpu())->pcp;
1004 local_irq_save(flags);
1005 __count_vm_event(PGFREE);
1006 if (cold)
1007 list_add_tail(&page->lru, &pcp->list);
1008 else
1009 list_add(&page->lru, &pcp->list);
1010 set_page_private(page, get_pageblock_migratetype(page));
1011 pcp->count++;
1012 if (pcp->count >= pcp->high) {
1013 free_pages_bulk(zone, pcp->batch, &pcp->list, 0);
1014 pcp->count -= pcp->batch;
1016 local_irq_restore(flags);
1017 put_cpu();
1020 void free_hot_page(struct page *page)
1022 free_hot_cold_page(page, 0);
1025 void free_cold_page(struct page *page)
1027 free_hot_cold_page(page, 1);
1031 * split_page takes a non-compound higher-order page, and splits it into
1032 * n (1<<order) sub-pages: page[0..n]
1033 * Each sub-page must be freed individually.
1035 * Note: this is probably too low level an operation for use in drivers.
1036 * Please consult with lkml before using this in your driver.
1038 void split_page(struct page *page, unsigned int order)
1040 int i;
1042 VM_BUG_ON(PageCompound(page));
1043 VM_BUG_ON(!page_count(page));
1044 for (i = 1; i < (1 << order); i++)
1045 set_page_refcounted(page + i);
1049 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1050 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1051 * or two.
1053 static struct page *buffered_rmqueue(struct zone *preferred_zone,
1054 struct zone *zone, int order, gfp_t gfp_flags)
1056 unsigned long flags;
1057 struct page *page;
1058 int cold = !!(gfp_flags & __GFP_COLD);
1059 int cpu;
1060 int migratetype = allocflags_to_migratetype(gfp_flags);
1062 again:
1063 cpu = get_cpu();
1064 if (likely(order == 0)) {
1065 struct per_cpu_pages *pcp;
1067 pcp = &zone_pcp(zone, cpu)->pcp;
1068 local_irq_save(flags);
1069 if (!pcp->count) {
1070 pcp->count = rmqueue_bulk(zone, 0,
1071 pcp->batch, &pcp->list, migratetype);
1072 if (unlikely(!pcp->count))
1073 goto failed;
1076 /* Find a page of the appropriate migrate type */
1077 if (cold) {
1078 list_for_each_entry_reverse(page, &pcp->list, lru)
1079 if (page_private(page) == migratetype)
1080 break;
1081 } else {
1082 list_for_each_entry(page, &pcp->list, lru)
1083 if (page_private(page) == migratetype)
1084 break;
1087 /* Allocate more to the pcp list if necessary */
1088 if (unlikely(&page->lru == &pcp->list)) {
1089 pcp->count += rmqueue_bulk(zone, 0,
1090 pcp->batch, &pcp->list, migratetype);
1091 page = list_entry(pcp->list.next, struct page, lru);
1094 list_del(&page->lru);
1095 pcp->count--;
1096 } else {
1097 spin_lock_irqsave(&zone->lock, flags);
1098 page = __rmqueue(zone, order, migratetype);
1099 spin_unlock(&zone->lock);
1100 if (!page)
1101 goto failed;
1104 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1105 zone_statistics(preferred_zone, zone);
1106 local_irq_restore(flags);
1107 put_cpu();
1109 VM_BUG_ON(bad_range(zone, page));
1110 if (prep_new_page(page, order, gfp_flags))
1111 goto again;
1112 return page;
1114 failed:
1115 local_irq_restore(flags);
1116 put_cpu();
1117 return NULL;
1120 #define ALLOC_NO_WATERMARKS 0x01 /* don't check watermarks at all */
1121 #define ALLOC_WMARK_MIN 0x02 /* use pages_min watermark */
1122 #define ALLOC_WMARK_LOW 0x04 /* use pages_low watermark */
1123 #define ALLOC_WMARK_HIGH 0x08 /* use pages_high watermark */
1124 #define ALLOC_HARDER 0x10 /* try to alloc harder */
1125 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
1126 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
1128 #ifdef CONFIG_FAIL_PAGE_ALLOC
1130 static struct fail_page_alloc_attr {
1131 struct fault_attr attr;
1133 u32 ignore_gfp_highmem;
1134 u32 ignore_gfp_wait;
1135 u32 min_order;
1137 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1139 struct dentry *ignore_gfp_highmem_file;
1140 struct dentry *ignore_gfp_wait_file;
1141 struct dentry *min_order_file;
1143 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1145 } fail_page_alloc = {
1146 .attr = FAULT_ATTR_INITIALIZER,
1147 .ignore_gfp_wait = 1,
1148 .ignore_gfp_highmem = 1,
1149 .min_order = 1,
1152 static int __init setup_fail_page_alloc(char *str)
1154 return setup_fault_attr(&fail_page_alloc.attr, str);
1156 __setup("fail_page_alloc=", setup_fail_page_alloc);
1158 static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1160 if (order < fail_page_alloc.min_order)
1161 return 0;
1162 if (gfp_mask & __GFP_NOFAIL)
1163 return 0;
1164 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1165 return 0;
1166 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1167 return 0;
1169 return should_fail(&fail_page_alloc.attr, 1 << order);
1172 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1174 static int __init fail_page_alloc_debugfs(void)
1176 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1177 struct dentry *dir;
1178 int err;
1180 err = init_fault_attr_dentries(&fail_page_alloc.attr,
1181 "fail_page_alloc");
1182 if (err)
1183 return err;
1184 dir = fail_page_alloc.attr.dentries.dir;
1186 fail_page_alloc.ignore_gfp_wait_file =
1187 debugfs_create_bool("ignore-gfp-wait", mode, dir,
1188 &fail_page_alloc.ignore_gfp_wait);
1190 fail_page_alloc.ignore_gfp_highmem_file =
1191 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1192 &fail_page_alloc.ignore_gfp_highmem);
1193 fail_page_alloc.min_order_file =
1194 debugfs_create_u32("min-order", mode, dir,
1195 &fail_page_alloc.min_order);
1197 if (!fail_page_alloc.ignore_gfp_wait_file ||
1198 !fail_page_alloc.ignore_gfp_highmem_file ||
1199 !fail_page_alloc.min_order_file) {
1200 err = -ENOMEM;
1201 debugfs_remove(fail_page_alloc.ignore_gfp_wait_file);
1202 debugfs_remove(fail_page_alloc.ignore_gfp_highmem_file);
1203 debugfs_remove(fail_page_alloc.min_order_file);
1204 cleanup_fault_attr_dentries(&fail_page_alloc.attr);
1207 return err;
1210 late_initcall(fail_page_alloc_debugfs);
1212 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1214 #else /* CONFIG_FAIL_PAGE_ALLOC */
1216 static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1218 return 0;
1221 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1224 * Return 1 if free pages are above 'mark'. This takes into account the order
1225 * of the allocation.
1227 int zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1228 int classzone_idx, int alloc_flags)
1230 /* free_pages my go negative - that's OK */
1231 long min = mark;
1232 long free_pages = zone_page_state(z, NR_FREE_PAGES) - (1 << order) + 1;
1233 int o;
1235 if (alloc_flags & ALLOC_HIGH)
1236 min -= min / 2;
1237 if (alloc_flags & ALLOC_HARDER)
1238 min -= min / 4;
1240 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
1241 return 0;
1242 for (o = 0; o < order; o++) {
1243 /* At the next order, this order's pages become unavailable */
1244 free_pages -= z->free_area[o].nr_free << o;
1246 /* Require fewer higher order pages to be free */
1247 min >>= 1;
1249 if (free_pages <= min)
1250 return 0;
1252 return 1;
1255 #ifdef CONFIG_NUMA
1257 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1258 * skip over zones that are not allowed by the cpuset, or that have
1259 * been recently (in last second) found to be nearly full. See further
1260 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1261 * that have to skip over a lot of full or unallowed zones.
1263 * If the zonelist cache is present in the passed in zonelist, then
1264 * returns a pointer to the allowed node mask (either the current
1265 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1267 * If the zonelist cache is not available for this zonelist, does
1268 * nothing and returns NULL.
1270 * If the fullzones BITMAP in the zonelist cache is stale (more than
1271 * a second since last zap'd) then we zap it out (clear its bits.)
1273 * We hold off even calling zlc_setup, until after we've checked the
1274 * first zone in the zonelist, on the theory that most allocations will
1275 * be satisfied from that first zone, so best to examine that zone as
1276 * quickly as we can.
1278 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1280 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1281 nodemask_t *allowednodes; /* zonelist_cache approximation */
1283 zlc = zonelist->zlcache_ptr;
1284 if (!zlc)
1285 return NULL;
1287 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1288 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1289 zlc->last_full_zap = jiffies;
1292 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1293 &cpuset_current_mems_allowed :
1294 &node_states[N_HIGH_MEMORY];
1295 return allowednodes;
1299 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1300 * if it is worth looking at further for free memory:
1301 * 1) Check that the zone isn't thought to be full (doesn't have its
1302 * bit set in the zonelist_cache fullzones BITMAP).
1303 * 2) Check that the zones node (obtained from the zonelist_cache
1304 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1305 * Return true (non-zero) if zone is worth looking at further, or
1306 * else return false (zero) if it is not.
1308 * This check -ignores- the distinction between various watermarks,
1309 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1310 * found to be full for any variation of these watermarks, it will
1311 * be considered full for up to one second by all requests, unless
1312 * we are so low on memory on all allowed nodes that we are forced
1313 * into the second scan of the zonelist.
1315 * In the second scan we ignore this zonelist cache and exactly
1316 * apply the watermarks to all zones, even it is slower to do so.
1317 * We are low on memory in the second scan, and should leave no stone
1318 * unturned looking for a free page.
1320 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1321 nodemask_t *allowednodes)
1323 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1324 int i; /* index of *z in zonelist zones */
1325 int n; /* node that zone *z is on */
1327 zlc = zonelist->zlcache_ptr;
1328 if (!zlc)
1329 return 1;
1331 i = z - zonelist->_zonerefs;
1332 n = zlc->z_to_n[i];
1334 /* This zone is worth trying if it is allowed but not full */
1335 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1339 * Given 'z' scanning a zonelist, set the corresponding bit in
1340 * zlc->fullzones, so that subsequent attempts to allocate a page
1341 * from that zone don't waste time re-examining it.
1343 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1345 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1346 int i; /* index of *z in zonelist zones */
1348 zlc = zonelist->zlcache_ptr;
1349 if (!zlc)
1350 return;
1352 i = z - zonelist->_zonerefs;
1354 set_bit(i, zlc->fullzones);
1357 #else /* CONFIG_NUMA */
1359 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1361 return NULL;
1364 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1365 nodemask_t *allowednodes)
1367 return 1;
1370 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1373 #endif /* CONFIG_NUMA */
1376 * get_page_from_freelist goes through the zonelist trying to allocate
1377 * a page.
1379 static struct page *
1380 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1381 struct zonelist *zonelist, int high_zoneidx, int alloc_flags)
1383 struct zoneref *z;
1384 struct page *page = NULL;
1385 int classzone_idx;
1386 struct zone *zone, *preferred_zone;
1387 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1388 int zlc_active = 0; /* set if using zonelist_cache */
1389 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1391 (void)first_zones_zonelist(zonelist, high_zoneidx, nodemask,
1392 &preferred_zone);
1393 if (!preferred_zone)
1394 return NULL;
1396 classzone_idx = zone_idx(preferred_zone);
1398 zonelist_scan:
1400 * Scan zonelist, looking for a zone with enough free.
1401 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1403 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1404 high_zoneidx, nodemask) {
1405 if (NUMA_BUILD && zlc_active &&
1406 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1407 continue;
1408 if ((alloc_flags & ALLOC_CPUSET) &&
1409 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1410 goto try_next_zone;
1412 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1413 unsigned long mark;
1414 if (alloc_flags & ALLOC_WMARK_MIN)
1415 mark = zone->pages_min;
1416 else if (alloc_flags & ALLOC_WMARK_LOW)
1417 mark = zone->pages_low;
1418 else
1419 mark = zone->pages_high;
1420 if (!zone_watermark_ok(zone, order, mark,
1421 classzone_idx, alloc_flags)) {
1422 if (!zone_reclaim_mode ||
1423 !zone_reclaim(zone, gfp_mask, order))
1424 goto this_zone_full;
1428 page = buffered_rmqueue(preferred_zone, zone, order, gfp_mask);
1429 if (page)
1430 break;
1431 this_zone_full:
1432 if (NUMA_BUILD)
1433 zlc_mark_zone_full(zonelist, z);
1434 try_next_zone:
1435 if (NUMA_BUILD && !did_zlc_setup) {
1436 /* we do zlc_setup after the first zone is tried */
1437 allowednodes = zlc_setup(zonelist, alloc_flags);
1438 zlc_active = 1;
1439 did_zlc_setup = 1;
1443 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) {
1444 /* Disable zlc cache for second zonelist scan */
1445 zlc_active = 0;
1446 goto zonelist_scan;
1448 return page;
1452 * This is the 'heart' of the zoned buddy allocator.
1454 struct page *
1455 __alloc_pages_internal(gfp_t gfp_mask, unsigned int order,
1456 struct zonelist *zonelist, nodemask_t *nodemask)
1458 const gfp_t wait = gfp_mask & __GFP_WAIT;
1459 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
1460 struct zoneref *z;
1461 struct zone *zone;
1462 struct page *page;
1463 struct reclaim_state reclaim_state;
1464 struct task_struct *p = current;
1465 int do_retry;
1466 int alloc_flags;
1467 unsigned long did_some_progress;
1468 unsigned long pages_reclaimed = 0;
1470 might_sleep_if(wait);
1472 if (should_fail_alloc_page(gfp_mask, order))
1473 return NULL;
1475 restart:
1476 z = zonelist->_zonerefs; /* the list of zones suitable for gfp_mask */
1478 if (unlikely(!z->zone)) {
1480 * Happens if we have an empty zonelist as a result of
1481 * GFP_THISNODE being used on a memoryless node
1483 return NULL;
1486 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
1487 zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET);
1488 if (page)
1489 goto got_pg;
1492 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
1493 * __GFP_NOWARN set) should not cause reclaim since the subsystem
1494 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
1495 * using a larger set of nodes after it has established that the
1496 * allowed per node queues are empty and that nodes are
1497 * over allocated.
1499 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
1500 goto nopage;
1502 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
1503 wakeup_kswapd(zone, order);
1506 * OK, we're below the kswapd watermark and have kicked background
1507 * reclaim. Now things get more complex, so set up alloc_flags according
1508 * to how we want to proceed.
1510 * The caller may dip into page reserves a bit more if the caller
1511 * cannot run direct reclaim, or if the caller has realtime scheduling
1512 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
1513 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
1515 alloc_flags = ALLOC_WMARK_MIN;
1516 if ((unlikely(rt_task(p)) && !in_interrupt()) || !wait)
1517 alloc_flags |= ALLOC_HARDER;
1518 if (gfp_mask & __GFP_HIGH)
1519 alloc_flags |= ALLOC_HIGH;
1520 if (wait)
1521 alloc_flags |= ALLOC_CPUSET;
1524 * Go through the zonelist again. Let __GFP_HIGH and allocations
1525 * coming from realtime tasks go deeper into reserves.
1527 * This is the last chance, in general, before the goto nopage.
1528 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
1529 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1531 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
1532 high_zoneidx, alloc_flags);
1533 if (page)
1534 goto got_pg;
1536 /* This allocation should allow future memory freeing. */
1538 rebalance:
1539 if (((p->flags & PF_MEMALLOC) || unlikely(test_thread_flag(TIF_MEMDIE)))
1540 && !in_interrupt()) {
1541 if (!(gfp_mask & __GFP_NOMEMALLOC)) {
1542 nofail_alloc:
1543 /* go through the zonelist yet again, ignoring mins */
1544 page = get_page_from_freelist(gfp_mask, nodemask, order,
1545 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS);
1546 if (page)
1547 goto got_pg;
1548 if (gfp_mask & __GFP_NOFAIL) {
1549 congestion_wait(WRITE, HZ/50);
1550 goto nofail_alloc;
1553 goto nopage;
1556 /* Atomic allocations - we can't balance anything */
1557 if (!wait)
1558 goto nopage;
1560 cond_resched();
1562 /* We now go into synchronous reclaim */
1563 cpuset_memory_pressure_bump();
1565 * The task's cpuset might have expanded its set of allowable nodes
1567 cpuset_update_task_memory_state();
1568 p->flags |= PF_MEMALLOC;
1569 reclaim_state.reclaimed_slab = 0;
1570 p->reclaim_state = &reclaim_state;
1572 did_some_progress = try_to_free_pages(zonelist, order, gfp_mask);
1574 p->reclaim_state = NULL;
1575 p->flags &= ~PF_MEMALLOC;
1577 cond_resched();
1579 if (order != 0)
1580 drain_all_pages();
1582 if (likely(did_some_progress)) {
1583 page = get_page_from_freelist(gfp_mask, nodemask, order,
1584 zonelist, high_zoneidx, alloc_flags);
1585 if (page)
1586 goto got_pg;
1587 } else if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
1588 if (!try_set_zone_oom(zonelist, gfp_mask)) {
1589 schedule_timeout_uninterruptible(1);
1590 goto restart;
1594 * Go through the zonelist yet one more time, keep
1595 * very high watermark here, this is only to catch
1596 * a parallel oom killing, we must fail if we're still
1597 * under heavy pressure.
1599 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
1600 order, zonelist, high_zoneidx,
1601 ALLOC_WMARK_HIGH|ALLOC_CPUSET);
1602 if (page) {
1603 clear_zonelist_oom(zonelist, gfp_mask);
1604 goto got_pg;
1607 /* The OOM killer will not help higher order allocs so fail */
1608 if (order > PAGE_ALLOC_COSTLY_ORDER) {
1609 clear_zonelist_oom(zonelist, gfp_mask);
1610 goto nopage;
1613 out_of_memory(zonelist, gfp_mask, order);
1614 clear_zonelist_oom(zonelist, gfp_mask);
1615 goto restart;
1619 * Don't let big-order allocations loop unless the caller explicitly
1620 * requests that. Wait for some write requests to complete then retry.
1622 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
1623 * means __GFP_NOFAIL, but that may not be true in other
1624 * implementations.
1626 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
1627 * specified, then we retry until we no longer reclaim any pages
1628 * (above), or we've reclaimed an order of pages at least as
1629 * large as the allocation's order. In both cases, if the
1630 * allocation still fails, we stop retrying.
1632 pages_reclaimed += did_some_progress;
1633 do_retry = 0;
1634 if (!(gfp_mask & __GFP_NORETRY)) {
1635 if (order <= PAGE_ALLOC_COSTLY_ORDER) {
1636 do_retry = 1;
1637 } else {
1638 if (gfp_mask & __GFP_REPEAT &&
1639 pages_reclaimed < (1 << order))
1640 do_retry = 1;
1642 if (gfp_mask & __GFP_NOFAIL)
1643 do_retry = 1;
1645 if (do_retry) {
1646 congestion_wait(WRITE, HZ/50);
1647 goto rebalance;
1650 nopage:
1651 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
1652 printk(KERN_WARNING "%s: page allocation failure."
1653 " order:%d, mode:0x%x\n",
1654 p->comm, order, gfp_mask);
1655 dump_stack();
1656 show_mem();
1658 got_pg:
1659 return page;
1661 EXPORT_SYMBOL(__alloc_pages_internal);
1664 * Common helper functions.
1666 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
1668 struct page * page;
1669 page = alloc_pages(gfp_mask, order);
1670 if (!page)
1671 return 0;
1672 return (unsigned long) page_address(page);
1675 EXPORT_SYMBOL(__get_free_pages);
1677 unsigned long get_zeroed_page(gfp_t gfp_mask)
1679 struct page * page;
1682 * get_zeroed_page() returns a 32-bit address, which cannot represent
1683 * a highmem page
1685 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
1687 page = alloc_pages(gfp_mask | __GFP_ZERO, 0);
1688 if (page)
1689 return (unsigned long) page_address(page);
1690 return 0;
1693 EXPORT_SYMBOL(get_zeroed_page);
1695 void __pagevec_free(struct pagevec *pvec)
1697 int i = pagevec_count(pvec);
1699 while (--i >= 0)
1700 free_hot_cold_page(pvec->pages[i], pvec->cold);
1703 void __free_pages(struct page *page, unsigned int order)
1705 if (put_page_testzero(page)) {
1706 if (order == 0)
1707 free_hot_page(page);
1708 else
1709 __free_pages_ok(page, order);
1713 EXPORT_SYMBOL(__free_pages);
1715 void free_pages(unsigned long addr, unsigned int order)
1717 if (addr != 0) {
1718 VM_BUG_ON(!virt_addr_valid((void *)addr));
1719 __free_pages(virt_to_page((void *)addr), order);
1723 EXPORT_SYMBOL(free_pages);
1726 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
1727 * @size: the number of bytes to allocate
1728 * @gfp_mask: GFP flags for the allocation
1730 * This function is similar to alloc_pages(), except that it allocates the
1731 * minimum number of pages to satisfy the request. alloc_pages() can only
1732 * allocate memory in power-of-two pages.
1734 * This function is also limited by MAX_ORDER.
1736 * Memory allocated by this function must be released by free_pages_exact().
1738 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
1740 unsigned int order = get_order(size);
1741 unsigned long addr;
1743 addr = __get_free_pages(gfp_mask, order);
1744 if (addr) {
1745 unsigned long alloc_end = addr + (PAGE_SIZE << order);
1746 unsigned long used = addr + PAGE_ALIGN(size);
1748 split_page(virt_to_page(addr), order);
1749 while (used < alloc_end) {
1750 free_page(used);
1751 used += PAGE_SIZE;
1755 return (void *)addr;
1757 EXPORT_SYMBOL(alloc_pages_exact);
1760 * free_pages_exact - release memory allocated via alloc_pages_exact()
1761 * @virt: the value returned by alloc_pages_exact.
1762 * @size: size of allocation, same value as passed to alloc_pages_exact().
1764 * Release the memory allocated by a previous call to alloc_pages_exact.
1766 void free_pages_exact(void *virt, size_t size)
1768 unsigned long addr = (unsigned long)virt;
1769 unsigned long end = addr + PAGE_ALIGN(size);
1771 while (addr < end) {
1772 free_page(addr);
1773 addr += PAGE_SIZE;
1776 EXPORT_SYMBOL(free_pages_exact);
1778 static unsigned int nr_free_zone_pages(int offset)
1780 struct zoneref *z;
1781 struct zone *zone;
1783 /* Just pick one node, since fallback list is circular */
1784 unsigned int sum = 0;
1786 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
1788 for_each_zone_zonelist(zone, z, zonelist, offset) {
1789 unsigned long size = zone->present_pages;
1790 unsigned long high = zone->pages_high;
1791 if (size > high)
1792 sum += size - high;
1795 return sum;
1799 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
1801 unsigned int nr_free_buffer_pages(void)
1803 return nr_free_zone_pages(gfp_zone(GFP_USER));
1805 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
1808 * Amount of free RAM allocatable within all zones
1810 unsigned int nr_free_pagecache_pages(void)
1812 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
1815 static inline void show_node(struct zone *zone)
1817 if (NUMA_BUILD)
1818 printk("Node %d ", zone_to_nid(zone));
1821 void si_meminfo(struct sysinfo *val)
1823 val->totalram = totalram_pages;
1824 val->sharedram = 0;
1825 val->freeram = global_page_state(NR_FREE_PAGES);
1826 val->bufferram = nr_blockdev_pages();
1827 val->totalhigh = totalhigh_pages;
1828 val->freehigh = nr_free_highpages();
1829 val->mem_unit = PAGE_SIZE;
1832 EXPORT_SYMBOL(si_meminfo);
1834 #ifdef CONFIG_NUMA
1835 void si_meminfo_node(struct sysinfo *val, int nid)
1837 pg_data_t *pgdat = NODE_DATA(nid);
1839 val->totalram = pgdat->node_present_pages;
1840 val->freeram = node_page_state(nid, NR_FREE_PAGES);
1841 #ifdef CONFIG_HIGHMEM
1842 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
1843 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
1844 NR_FREE_PAGES);
1845 #else
1846 val->totalhigh = 0;
1847 val->freehigh = 0;
1848 #endif
1849 val->mem_unit = PAGE_SIZE;
1851 #endif
1853 #define K(x) ((x) << (PAGE_SHIFT-10))
1856 * Show free area list (used inside shift_scroll-lock stuff)
1857 * We also calculate the percentage fragmentation. We do this by counting the
1858 * memory on each free list with the exception of the first item on the list.
1860 void show_free_areas(void)
1862 int cpu;
1863 struct zone *zone;
1865 for_each_zone(zone) {
1866 if (!populated_zone(zone))
1867 continue;
1869 show_node(zone);
1870 printk("%s per-cpu:\n", zone->name);
1872 for_each_online_cpu(cpu) {
1873 struct per_cpu_pageset *pageset;
1875 pageset = zone_pcp(zone, cpu);
1877 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
1878 cpu, pageset->pcp.high,
1879 pageset->pcp.batch, pageset->pcp.count);
1883 printk("Active_anon:%lu active_file:%lu inactive_anon:%lu\n"
1884 " inactive_file:%lu"
1885 //TODO: check/adjust line lengths
1886 #ifdef CONFIG_UNEVICTABLE_LRU
1887 " unevictable:%lu"
1888 #endif
1889 " dirty:%lu writeback:%lu unstable:%lu\n"
1890 " free:%lu slab:%lu mapped:%lu pagetables:%lu bounce:%lu\n",
1891 global_page_state(NR_ACTIVE_ANON),
1892 global_page_state(NR_ACTIVE_FILE),
1893 global_page_state(NR_INACTIVE_ANON),
1894 global_page_state(NR_INACTIVE_FILE),
1895 #ifdef CONFIG_UNEVICTABLE_LRU
1896 global_page_state(NR_UNEVICTABLE),
1897 #endif
1898 global_page_state(NR_FILE_DIRTY),
1899 global_page_state(NR_WRITEBACK),
1900 global_page_state(NR_UNSTABLE_NFS),
1901 global_page_state(NR_FREE_PAGES),
1902 global_page_state(NR_SLAB_RECLAIMABLE) +
1903 global_page_state(NR_SLAB_UNRECLAIMABLE),
1904 global_page_state(NR_FILE_MAPPED),
1905 global_page_state(NR_PAGETABLE),
1906 global_page_state(NR_BOUNCE));
1908 for_each_zone(zone) {
1909 int i;
1911 if (!populated_zone(zone))
1912 continue;
1914 show_node(zone);
1915 printk("%s"
1916 " free:%lukB"
1917 " min:%lukB"
1918 " low:%lukB"
1919 " high:%lukB"
1920 " active_anon:%lukB"
1921 " inactive_anon:%lukB"
1922 " active_file:%lukB"
1923 " inactive_file:%lukB"
1924 #ifdef CONFIG_UNEVICTABLE_LRU
1925 " unevictable:%lukB"
1926 #endif
1927 " present:%lukB"
1928 " pages_scanned:%lu"
1929 " all_unreclaimable? %s"
1930 "\n",
1931 zone->name,
1932 K(zone_page_state(zone, NR_FREE_PAGES)),
1933 K(zone->pages_min),
1934 K(zone->pages_low),
1935 K(zone->pages_high),
1936 K(zone_page_state(zone, NR_ACTIVE_ANON)),
1937 K(zone_page_state(zone, NR_INACTIVE_ANON)),
1938 K(zone_page_state(zone, NR_ACTIVE_FILE)),
1939 K(zone_page_state(zone, NR_INACTIVE_FILE)),
1940 #ifdef CONFIG_UNEVICTABLE_LRU
1941 K(zone_page_state(zone, NR_UNEVICTABLE)),
1942 #endif
1943 K(zone->present_pages),
1944 zone->pages_scanned,
1945 (zone_is_all_unreclaimable(zone) ? "yes" : "no")
1947 printk("lowmem_reserve[]:");
1948 for (i = 0; i < MAX_NR_ZONES; i++)
1949 printk(" %lu", zone->lowmem_reserve[i]);
1950 printk("\n");
1953 for_each_zone(zone) {
1954 unsigned long nr[MAX_ORDER], flags, order, total = 0;
1956 if (!populated_zone(zone))
1957 continue;
1959 show_node(zone);
1960 printk("%s: ", zone->name);
1962 spin_lock_irqsave(&zone->lock, flags);
1963 for (order = 0; order < MAX_ORDER; order++) {
1964 nr[order] = zone->free_area[order].nr_free;
1965 total += nr[order] << order;
1967 spin_unlock_irqrestore(&zone->lock, flags);
1968 for (order = 0; order < MAX_ORDER; order++)
1969 printk("%lu*%lukB ", nr[order], K(1UL) << order);
1970 printk("= %lukB\n", K(total));
1973 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
1975 show_swap_cache_info();
1978 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
1980 zoneref->zone = zone;
1981 zoneref->zone_idx = zone_idx(zone);
1985 * Builds allocation fallback zone lists.
1987 * Add all populated zones of a node to the zonelist.
1989 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
1990 int nr_zones, enum zone_type zone_type)
1992 struct zone *zone;
1994 BUG_ON(zone_type >= MAX_NR_ZONES);
1995 zone_type++;
1997 do {
1998 zone_type--;
1999 zone = pgdat->node_zones + zone_type;
2000 if (populated_zone(zone)) {
2001 zoneref_set_zone(zone,
2002 &zonelist->_zonerefs[nr_zones++]);
2003 check_highest_zone(zone_type);
2006 } while (zone_type);
2007 return nr_zones;
2012 * zonelist_order:
2013 * 0 = automatic detection of better ordering.
2014 * 1 = order by ([node] distance, -zonetype)
2015 * 2 = order by (-zonetype, [node] distance)
2017 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
2018 * the same zonelist. So only NUMA can configure this param.
2020 #define ZONELIST_ORDER_DEFAULT 0
2021 #define ZONELIST_ORDER_NODE 1
2022 #define ZONELIST_ORDER_ZONE 2
2024 /* zonelist order in the kernel.
2025 * set_zonelist_order() will set this to NODE or ZONE.
2027 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
2028 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
2031 #ifdef CONFIG_NUMA
2032 /* The value user specified ....changed by config */
2033 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2034 /* string for sysctl */
2035 #define NUMA_ZONELIST_ORDER_LEN 16
2036 char numa_zonelist_order[16] = "default";
2039 * interface for configure zonelist ordering.
2040 * command line option "numa_zonelist_order"
2041 * = "[dD]efault - default, automatic configuration.
2042 * = "[nN]ode - order by node locality, then by zone within node
2043 * = "[zZ]one - order by zone, then by locality within zone
2046 static int __parse_numa_zonelist_order(char *s)
2048 if (*s == 'd' || *s == 'D') {
2049 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2050 } else if (*s == 'n' || *s == 'N') {
2051 user_zonelist_order = ZONELIST_ORDER_NODE;
2052 } else if (*s == 'z' || *s == 'Z') {
2053 user_zonelist_order = ZONELIST_ORDER_ZONE;
2054 } else {
2055 printk(KERN_WARNING
2056 "Ignoring invalid numa_zonelist_order value: "
2057 "%s\n", s);
2058 return -EINVAL;
2060 return 0;
2063 static __init int setup_numa_zonelist_order(char *s)
2065 if (s)
2066 return __parse_numa_zonelist_order(s);
2067 return 0;
2069 early_param("numa_zonelist_order", setup_numa_zonelist_order);
2072 * sysctl handler for numa_zonelist_order
2074 int numa_zonelist_order_handler(ctl_table *table, int write,
2075 struct file *file, void __user *buffer, size_t *length,
2076 loff_t *ppos)
2078 char saved_string[NUMA_ZONELIST_ORDER_LEN];
2079 int ret;
2081 if (write)
2082 strncpy(saved_string, (char*)table->data,
2083 NUMA_ZONELIST_ORDER_LEN);
2084 ret = proc_dostring(table, write, file, buffer, length, ppos);
2085 if (ret)
2086 return ret;
2087 if (write) {
2088 int oldval = user_zonelist_order;
2089 if (__parse_numa_zonelist_order((char*)table->data)) {
2091 * bogus value. restore saved string
2093 strncpy((char*)table->data, saved_string,
2094 NUMA_ZONELIST_ORDER_LEN);
2095 user_zonelist_order = oldval;
2096 } else if (oldval != user_zonelist_order)
2097 build_all_zonelists();
2099 return 0;
2103 #define MAX_NODE_LOAD (num_online_nodes())
2104 static int node_load[MAX_NUMNODES];
2107 * find_next_best_node - find the next node that should appear in a given node's fallback list
2108 * @node: node whose fallback list we're appending
2109 * @used_node_mask: nodemask_t of already used nodes
2111 * We use a number of factors to determine which is the next node that should
2112 * appear on a given node's fallback list. The node should not have appeared
2113 * already in @node's fallback list, and it should be the next closest node
2114 * according to the distance array (which contains arbitrary distance values
2115 * from each node to each node in the system), and should also prefer nodes
2116 * with no CPUs, since presumably they'll have very little allocation pressure
2117 * on them otherwise.
2118 * It returns -1 if no node is found.
2120 static int find_next_best_node(int node, nodemask_t *used_node_mask)
2122 int n, val;
2123 int min_val = INT_MAX;
2124 int best_node = -1;
2125 node_to_cpumask_ptr(tmp, 0);
2127 /* Use the local node if we haven't already */
2128 if (!node_isset(node, *used_node_mask)) {
2129 node_set(node, *used_node_mask);
2130 return node;
2133 for_each_node_state(n, N_HIGH_MEMORY) {
2135 /* Don't want a node to appear more than once */
2136 if (node_isset(n, *used_node_mask))
2137 continue;
2139 /* Use the distance array to find the distance */
2140 val = node_distance(node, n);
2142 /* Penalize nodes under us ("prefer the next node") */
2143 val += (n < node);
2145 /* Give preference to headless and unused nodes */
2146 node_to_cpumask_ptr_next(tmp, n);
2147 if (!cpus_empty(*tmp))
2148 val += PENALTY_FOR_NODE_WITH_CPUS;
2150 /* Slight preference for less loaded node */
2151 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
2152 val += node_load[n];
2154 if (val < min_val) {
2155 min_val = val;
2156 best_node = n;
2160 if (best_node >= 0)
2161 node_set(best_node, *used_node_mask);
2163 return best_node;
2168 * Build zonelists ordered by node and zones within node.
2169 * This results in maximum locality--normal zone overflows into local
2170 * DMA zone, if any--but risks exhausting DMA zone.
2172 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
2174 int j;
2175 struct zonelist *zonelist;
2177 zonelist = &pgdat->node_zonelists[0];
2178 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
2180 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2181 MAX_NR_ZONES - 1);
2182 zonelist->_zonerefs[j].zone = NULL;
2183 zonelist->_zonerefs[j].zone_idx = 0;
2187 * Build gfp_thisnode zonelists
2189 static void build_thisnode_zonelists(pg_data_t *pgdat)
2191 int j;
2192 struct zonelist *zonelist;
2194 zonelist = &pgdat->node_zonelists[1];
2195 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2196 zonelist->_zonerefs[j].zone = NULL;
2197 zonelist->_zonerefs[j].zone_idx = 0;
2201 * Build zonelists ordered by zone and nodes within zones.
2202 * This results in conserving DMA zone[s] until all Normal memory is
2203 * exhausted, but results in overflowing to remote node while memory
2204 * may still exist in local DMA zone.
2206 static int node_order[MAX_NUMNODES];
2208 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
2210 int pos, j, node;
2211 int zone_type; /* needs to be signed */
2212 struct zone *z;
2213 struct zonelist *zonelist;
2215 zonelist = &pgdat->node_zonelists[0];
2216 pos = 0;
2217 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
2218 for (j = 0; j < nr_nodes; j++) {
2219 node = node_order[j];
2220 z = &NODE_DATA(node)->node_zones[zone_type];
2221 if (populated_zone(z)) {
2222 zoneref_set_zone(z,
2223 &zonelist->_zonerefs[pos++]);
2224 check_highest_zone(zone_type);
2228 zonelist->_zonerefs[pos].zone = NULL;
2229 zonelist->_zonerefs[pos].zone_idx = 0;
2232 static int default_zonelist_order(void)
2234 int nid, zone_type;
2235 unsigned long low_kmem_size,total_size;
2236 struct zone *z;
2237 int average_size;
2239 * ZONE_DMA and ZONE_DMA32 can be very small area in the sytem.
2240 * If they are really small and used heavily, the system can fall
2241 * into OOM very easily.
2242 * This function detect ZONE_DMA/DMA32 size and confgigures zone order.
2244 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
2245 low_kmem_size = 0;
2246 total_size = 0;
2247 for_each_online_node(nid) {
2248 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2249 z = &NODE_DATA(nid)->node_zones[zone_type];
2250 if (populated_zone(z)) {
2251 if (zone_type < ZONE_NORMAL)
2252 low_kmem_size += z->present_pages;
2253 total_size += z->present_pages;
2257 if (!low_kmem_size || /* there are no DMA area. */
2258 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
2259 return ZONELIST_ORDER_NODE;
2261 * look into each node's config.
2262 * If there is a node whose DMA/DMA32 memory is very big area on
2263 * local memory, NODE_ORDER may be suitable.
2265 average_size = total_size /
2266 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1);
2267 for_each_online_node(nid) {
2268 low_kmem_size = 0;
2269 total_size = 0;
2270 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2271 z = &NODE_DATA(nid)->node_zones[zone_type];
2272 if (populated_zone(z)) {
2273 if (zone_type < ZONE_NORMAL)
2274 low_kmem_size += z->present_pages;
2275 total_size += z->present_pages;
2278 if (low_kmem_size &&
2279 total_size > average_size && /* ignore small node */
2280 low_kmem_size > total_size * 70/100)
2281 return ZONELIST_ORDER_NODE;
2283 return ZONELIST_ORDER_ZONE;
2286 static void set_zonelist_order(void)
2288 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
2289 current_zonelist_order = default_zonelist_order();
2290 else
2291 current_zonelist_order = user_zonelist_order;
2294 static void build_zonelists(pg_data_t *pgdat)
2296 int j, node, load;
2297 enum zone_type i;
2298 nodemask_t used_mask;
2299 int local_node, prev_node;
2300 struct zonelist *zonelist;
2301 int order = current_zonelist_order;
2303 /* initialize zonelists */
2304 for (i = 0; i < MAX_ZONELISTS; i++) {
2305 zonelist = pgdat->node_zonelists + i;
2306 zonelist->_zonerefs[0].zone = NULL;
2307 zonelist->_zonerefs[0].zone_idx = 0;
2310 /* NUMA-aware ordering of nodes */
2311 local_node = pgdat->node_id;
2312 load = num_online_nodes();
2313 prev_node = local_node;
2314 nodes_clear(used_mask);
2316 memset(node_load, 0, sizeof(node_load));
2317 memset(node_order, 0, sizeof(node_order));
2318 j = 0;
2320 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
2321 int distance = node_distance(local_node, node);
2324 * If another node is sufficiently far away then it is better
2325 * to reclaim pages in a zone before going off node.
2327 if (distance > RECLAIM_DISTANCE)
2328 zone_reclaim_mode = 1;
2331 * We don't want to pressure a particular node.
2332 * So adding penalty to the first node in same
2333 * distance group to make it round-robin.
2335 if (distance != node_distance(local_node, prev_node))
2336 node_load[node] = load;
2338 prev_node = node;
2339 load--;
2340 if (order == ZONELIST_ORDER_NODE)
2341 build_zonelists_in_node_order(pgdat, node);
2342 else
2343 node_order[j++] = node; /* remember order */
2346 if (order == ZONELIST_ORDER_ZONE) {
2347 /* calculate node order -- i.e., DMA last! */
2348 build_zonelists_in_zone_order(pgdat, j);
2351 build_thisnode_zonelists(pgdat);
2354 /* Construct the zonelist performance cache - see further mmzone.h */
2355 static void build_zonelist_cache(pg_data_t *pgdat)
2357 struct zonelist *zonelist;
2358 struct zonelist_cache *zlc;
2359 struct zoneref *z;
2361 zonelist = &pgdat->node_zonelists[0];
2362 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
2363 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
2364 for (z = zonelist->_zonerefs; z->zone; z++)
2365 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
2369 #else /* CONFIG_NUMA */
2371 static void set_zonelist_order(void)
2373 current_zonelist_order = ZONELIST_ORDER_ZONE;
2376 static void build_zonelists(pg_data_t *pgdat)
2378 int node, local_node;
2379 enum zone_type j;
2380 struct zonelist *zonelist;
2382 local_node = pgdat->node_id;
2384 zonelist = &pgdat->node_zonelists[0];
2385 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2388 * Now we build the zonelist so that it contains the zones
2389 * of all the other nodes.
2390 * We don't want to pressure a particular node, so when
2391 * building the zones for node N, we make sure that the
2392 * zones coming right after the local ones are those from
2393 * node N+1 (modulo N)
2395 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
2396 if (!node_online(node))
2397 continue;
2398 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2399 MAX_NR_ZONES - 1);
2401 for (node = 0; node < local_node; node++) {
2402 if (!node_online(node))
2403 continue;
2404 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2405 MAX_NR_ZONES - 1);
2408 zonelist->_zonerefs[j].zone = NULL;
2409 zonelist->_zonerefs[j].zone_idx = 0;
2412 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
2413 static void build_zonelist_cache(pg_data_t *pgdat)
2415 pgdat->node_zonelists[0].zlcache_ptr = NULL;
2418 #endif /* CONFIG_NUMA */
2420 /* return values int ....just for stop_machine() */
2421 static int __build_all_zonelists(void *dummy)
2423 int nid;
2425 for_each_online_node(nid) {
2426 pg_data_t *pgdat = NODE_DATA(nid);
2428 build_zonelists(pgdat);
2429 build_zonelist_cache(pgdat);
2431 return 0;
2434 void build_all_zonelists(void)
2436 set_zonelist_order();
2438 if (system_state == SYSTEM_BOOTING) {
2439 __build_all_zonelists(NULL);
2440 mminit_verify_zonelist();
2441 cpuset_init_current_mems_allowed();
2442 } else {
2443 /* we have to stop all cpus to guarantee there is no user
2444 of zonelist */
2445 stop_machine(__build_all_zonelists, NULL, NULL);
2446 /* cpuset refresh routine should be here */
2448 vm_total_pages = nr_free_pagecache_pages();
2450 * Disable grouping by mobility if the number of pages in the
2451 * system is too low to allow the mechanism to work. It would be
2452 * more accurate, but expensive to check per-zone. This check is
2453 * made on memory-hotadd so a system can start with mobility
2454 * disabled and enable it later
2456 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
2457 page_group_by_mobility_disabled = 1;
2458 else
2459 page_group_by_mobility_disabled = 0;
2461 printk("Built %i zonelists in %s order, mobility grouping %s. "
2462 "Total pages: %ld\n",
2463 num_online_nodes(),
2464 zonelist_order_name[current_zonelist_order],
2465 page_group_by_mobility_disabled ? "off" : "on",
2466 vm_total_pages);
2467 #ifdef CONFIG_NUMA
2468 printk("Policy zone: %s\n", zone_names[policy_zone]);
2469 #endif
2473 * Helper functions to size the waitqueue hash table.
2474 * Essentially these want to choose hash table sizes sufficiently
2475 * large so that collisions trying to wait on pages are rare.
2476 * But in fact, the number of active page waitqueues on typical
2477 * systems is ridiculously low, less than 200. So this is even
2478 * conservative, even though it seems large.
2480 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
2481 * waitqueues, i.e. the size of the waitq table given the number of pages.
2483 #define PAGES_PER_WAITQUEUE 256
2485 #ifndef CONFIG_MEMORY_HOTPLUG
2486 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
2488 unsigned long size = 1;
2490 pages /= PAGES_PER_WAITQUEUE;
2492 while (size < pages)
2493 size <<= 1;
2496 * Once we have dozens or even hundreds of threads sleeping
2497 * on IO we've got bigger problems than wait queue collision.
2498 * Limit the size of the wait table to a reasonable size.
2500 size = min(size, 4096UL);
2502 return max(size, 4UL);
2504 #else
2506 * A zone's size might be changed by hot-add, so it is not possible to determine
2507 * a suitable size for its wait_table. So we use the maximum size now.
2509 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
2511 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
2512 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
2513 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
2515 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
2516 * or more by the traditional way. (See above). It equals:
2518 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
2519 * ia64(16K page size) : = ( 8G + 4M)byte.
2520 * powerpc (64K page size) : = (32G +16M)byte.
2522 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
2524 return 4096UL;
2526 #endif
2529 * This is an integer logarithm so that shifts can be used later
2530 * to extract the more random high bits from the multiplicative
2531 * hash function before the remainder is taken.
2533 static inline unsigned long wait_table_bits(unsigned long size)
2535 return ffz(~size);
2538 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
2541 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
2542 * of blocks reserved is based on zone->pages_min. The memory within the
2543 * reserve will tend to store contiguous free pages. Setting min_free_kbytes
2544 * higher will lead to a bigger reserve which will get freed as contiguous
2545 * blocks as reclaim kicks in
2547 static void setup_zone_migrate_reserve(struct zone *zone)
2549 unsigned long start_pfn, pfn, end_pfn;
2550 struct page *page;
2551 unsigned long reserve, block_migratetype;
2553 /* Get the start pfn, end pfn and the number of blocks to reserve */
2554 start_pfn = zone->zone_start_pfn;
2555 end_pfn = start_pfn + zone->spanned_pages;
2556 reserve = roundup(zone->pages_min, pageblock_nr_pages) >>
2557 pageblock_order;
2559 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
2560 if (!pfn_valid(pfn))
2561 continue;
2562 page = pfn_to_page(pfn);
2564 /* Watch out for overlapping nodes */
2565 if (page_to_nid(page) != zone_to_nid(zone))
2566 continue;
2568 /* Blocks with reserved pages will never free, skip them. */
2569 if (PageReserved(page))
2570 continue;
2572 block_migratetype = get_pageblock_migratetype(page);
2574 /* If this block is reserved, account for it */
2575 if (reserve > 0 && block_migratetype == MIGRATE_RESERVE) {
2576 reserve--;
2577 continue;
2580 /* Suitable for reserving if this block is movable */
2581 if (reserve > 0 && block_migratetype == MIGRATE_MOVABLE) {
2582 set_pageblock_migratetype(page, MIGRATE_RESERVE);
2583 move_freepages_block(zone, page, MIGRATE_RESERVE);
2584 reserve--;
2585 continue;
2589 * If the reserve is met and this is a previous reserved block,
2590 * take it back
2592 if (block_migratetype == MIGRATE_RESERVE) {
2593 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
2594 move_freepages_block(zone, page, MIGRATE_MOVABLE);
2600 * Initially all pages are reserved - free ones are freed
2601 * up by free_all_bootmem() once the early boot process is
2602 * done. Non-atomic initialization, single-pass.
2604 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
2605 unsigned long start_pfn, enum memmap_context context)
2607 struct page *page;
2608 unsigned long end_pfn = start_pfn + size;
2609 unsigned long pfn;
2610 struct zone *z;
2612 z = &NODE_DATA(nid)->node_zones[zone];
2613 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
2615 * There can be holes in boot-time mem_map[]s
2616 * handed to this function. They do not
2617 * exist on hotplugged memory.
2619 if (context == MEMMAP_EARLY) {
2620 if (!early_pfn_valid(pfn))
2621 continue;
2622 if (!early_pfn_in_nid(pfn, nid))
2623 continue;
2625 page = pfn_to_page(pfn);
2626 set_page_links(page, zone, nid, pfn);
2627 mminit_verify_page_links(page, zone, nid, pfn);
2628 init_page_count(page);
2629 reset_page_mapcount(page);
2630 SetPageReserved(page);
2632 * Mark the block movable so that blocks are reserved for
2633 * movable at startup. This will force kernel allocations
2634 * to reserve their blocks rather than leaking throughout
2635 * the address space during boot when many long-lived
2636 * kernel allocations are made. Later some blocks near
2637 * the start are marked MIGRATE_RESERVE by
2638 * setup_zone_migrate_reserve()
2640 * bitmap is created for zone's valid pfn range. but memmap
2641 * can be created for invalid pages (for alignment)
2642 * check here not to call set_pageblock_migratetype() against
2643 * pfn out of zone.
2645 if ((z->zone_start_pfn <= pfn)
2646 && (pfn < z->zone_start_pfn + z->spanned_pages)
2647 && !(pfn & (pageblock_nr_pages - 1)))
2648 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
2650 INIT_LIST_HEAD(&page->lru);
2651 #ifdef WANT_PAGE_VIRTUAL
2652 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
2653 if (!is_highmem_idx(zone))
2654 set_page_address(page, __va(pfn << PAGE_SHIFT));
2655 #endif
2659 static void __meminit zone_init_free_lists(struct zone *zone)
2661 int order, t;
2662 for_each_migratetype_order(order, t) {
2663 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
2664 zone->free_area[order].nr_free = 0;
2668 #ifndef __HAVE_ARCH_MEMMAP_INIT
2669 #define memmap_init(size, nid, zone, start_pfn) \
2670 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
2671 #endif
2673 static int zone_batchsize(struct zone *zone)
2675 int batch;
2678 * The per-cpu-pages pools are set to around 1000th of the
2679 * size of the zone. But no more than 1/2 of a meg.
2681 * OK, so we don't know how big the cache is. So guess.
2683 batch = zone->present_pages / 1024;
2684 if (batch * PAGE_SIZE > 512 * 1024)
2685 batch = (512 * 1024) / PAGE_SIZE;
2686 batch /= 4; /* We effectively *= 4 below */
2687 if (batch < 1)
2688 batch = 1;
2691 * Clamp the batch to a 2^n - 1 value. Having a power
2692 * of 2 value was found to be more likely to have
2693 * suboptimal cache aliasing properties in some cases.
2695 * For example if 2 tasks are alternately allocating
2696 * batches of pages, one task can end up with a lot
2697 * of pages of one half of the possible page colors
2698 * and the other with pages of the other colors.
2700 batch = (1 << (fls(batch + batch/2)-1)) - 1;
2702 return batch;
2705 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
2707 struct per_cpu_pages *pcp;
2709 memset(p, 0, sizeof(*p));
2711 pcp = &p->pcp;
2712 pcp->count = 0;
2713 pcp->high = 6 * batch;
2714 pcp->batch = max(1UL, 1 * batch);
2715 INIT_LIST_HEAD(&pcp->list);
2719 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
2720 * to the value high for the pageset p.
2723 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
2724 unsigned long high)
2726 struct per_cpu_pages *pcp;
2728 pcp = &p->pcp;
2729 pcp->high = high;
2730 pcp->batch = max(1UL, high/4);
2731 if ((high/4) > (PAGE_SHIFT * 8))
2732 pcp->batch = PAGE_SHIFT * 8;
2736 #ifdef CONFIG_NUMA
2738 * Boot pageset table. One per cpu which is going to be used for all
2739 * zones and all nodes. The parameters will be set in such a way
2740 * that an item put on a list will immediately be handed over to
2741 * the buddy list. This is safe since pageset manipulation is done
2742 * with interrupts disabled.
2744 * Some NUMA counter updates may also be caught by the boot pagesets.
2746 * The boot_pagesets must be kept even after bootup is complete for
2747 * unused processors and/or zones. They do play a role for bootstrapping
2748 * hotplugged processors.
2750 * zoneinfo_show() and maybe other functions do
2751 * not check if the processor is online before following the pageset pointer.
2752 * Other parts of the kernel may not check if the zone is available.
2754 static struct per_cpu_pageset boot_pageset[NR_CPUS];
2757 * Dynamically allocate memory for the
2758 * per cpu pageset array in struct zone.
2760 static int __cpuinit process_zones(int cpu)
2762 struct zone *zone, *dzone;
2763 int node = cpu_to_node(cpu);
2765 node_set_state(node, N_CPU); /* this node has a cpu */
2767 for_each_zone(zone) {
2769 if (!populated_zone(zone))
2770 continue;
2772 zone_pcp(zone, cpu) = kmalloc_node(sizeof(struct per_cpu_pageset),
2773 GFP_KERNEL, node);
2774 if (!zone_pcp(zone, cpu))
2775 goto bad;
2777 setup_pageset(zone_pcp(zone, cpu), zone_batchsize(zone));
2779 if (percpu_pagelist_fraction)
2780 setup_pagelist_highmark(zone_pcp(zone, cpu),
2781 (zone->present_pages / percpu_pagelist_fraction));
2784 return 0;
2785 bad:
2786 for_each_zone(dzone) {
2787 if (!populated_zone(dzone))
2788 continue;
2789 if (dzone == zone)
2790 break;
2791 kfree(zone_pcp(dzone, cpu));
2792 zone_pcp(dzone, cpu) = NULL;
2794 return -ENOMEM;
2797 static inline void free_zone_pagesets(int cpu)
2799 struct zone *zone;
2801 for_each_zone(zone) {
2802 struct per_cpu_pageset *pset = zone_pcp(zone, cpu);
2804 /* Free per_cpu_pageset if it is slab allocated */
2805 if (pset != &boot_pageset[cpu])
2806 kfree(pset);
2807 zone_pcp(zone, cpu) = NULL;
2811 static int __cpuinit pageset_cpuup_callback(struct notifier_block *nfb,
2812 unsigned long action,
2813 void *hcpu)
2815 int cpu = (long)hcpu;
2816 int ret = NOTIFY_OK;
2818 switch (action) {
2819 case CPU_UP_PREPARE:
2820 case CPU_UP_PREPARE_FROZEN:
2821 if (process_zones(cpu))
2822 ret = NOTIFY_BAD;
2823 break;
2824 case CPU_UP_CANCELED:
2825 case CPU_UP_CANCELED_FROZEN:
2826 case CPU_DEAD:
2827 case CPU_DEAD_FROZEN:
2828 free_zone_pagesets(cpu);
2829 break;
2830 default:
2831 break;
2833 return ret;
2836 static struct notifier_block __cpuinitdata pageset_notifier =
2837 { &pageset_cpuup_callback, NULL, 0 };
2839 void __init setup_per_cpu_pageset(void)
2841 int err;
2843 /* Initialize per_cpu_pageset for cpu 0.
2844 * A cpuup callback will do this for every cpu
2845 * as it comes online
2847 err = process_zones(smp_processor_id());
2848 BUG_ON(err);
2849 register_cpu_notifier(&pageset_notifier);
2852 #endif
2854 static noinline __init_refok
2855 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
2857 int i;
2858 struct pglist_data *pgdat = zone->zone_pgdat;
2859 size_t alloc_size;
2862 * The per-page waitqueue mechanism uses hashed waitqueues
2863 * per zone.
2865 zone->wait_table_hash_nr_entries =
2866 wait_table_hash_nr_entries(zone_size_pages);
2867 zone->wait_table_bits =
2868 wait_table_bits(zone->wait_table_hash_nr_entries);
2869 alloc_size = zone->wait_table_hash_nr_entries
2870 * sizeof(wait_queue_head_t);
2872 if (!slab_is_available()) {
2873 zone->wait_table = (wait_queue_head_t *)
2874 alloc_bootmem_node(pgdat, alloc_size);
2875 } else {
2877 * This case means that a zone whose size was 0 gets new memory
2878 * via memory hot-add.
2879 * But it may be the case that a new node was hot-added. In
2880 * this case vmalloc() will not be able to use this new node's
2881 * memory - this wait_table must be initialized to use this new
2882 * node itself as well.
2883 * To use this new node's memory, further consideration will be
2884 * necessary.
2886 zone->wait_table = vmalloc(alloc_size);
2888 if (!zone->wait_table)
2889 return -ENOMEM;
2891 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
2892 init_waitqueue_head(zone->wait_table + i);
2894 return 0;
2897 static __meminit void zone_pcp_init(struct zone *zone)
2899 int cpu;
2900 unsigned long batch = zone_batchsize(zone);
2902 for (cpu = 0; cpu < NR_CPUS; cpu++) {
2903 #ifdef CONFIG_NUMA
2904 /* Early boot. Slab allocator not functional yet */
2905 zone_pcp(zone, cpu) = &boot_pageset[cpu];
2906 setup_pageset(&boot_pageset[cpu],0);
2907 #else
2908 setup_pageset(zone_pcp(zone,cpu), batch);
2909 #endif
2911 if (zone->present_pages)
2912 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%lu\n",
2913 zone->name, zone->present_pages, batch);
2916 __meminit int init_currently_empty_zone(struct zone *zone,
2917 unsigned long zone_start_pfn,
2918 unsigned long size,
2919 enum memmap_context context)
2921 struct pglist_data *pgdat = zone->zone_pgdat;
2922 int ret;
2923 ret = zone_wait_table_init(zone, size);
2924 if (ret)
2925 return ret;
2926 pgdat->nr_zones = zone_idx(zone) + 1;
2928 zone->zone_start_pfn = zone_start_pfn;
2930 mminit_dprintk(MMINIT_TRACE, "memmap_init",
2931 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
2932 pgdat->node_id,
2933 (unsigned long)zone_idx(zone),
2934 zone_start_pfn, (zone_start_pfn + size));
2936 zone_init_free_lists(zone);
2938 return 0;
2941 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
2943 * Basic iterator support. Return the first range of PFNs for a node
2944 * Note: nid == MAX_NUMNODES returns first region regardless of node
2946 static int __meminit first_active_region_index_in_nid(int nid)
2948 int i;
2950 for (i = 0; i < nr_nodemap_entries; i++)
2951 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
2952 return i;
2954 return -1;
2958 * Basic iterator support. Return the next active range of PFNs for a node
2959 * Note: nid == MAX_NUMNODES returns next region regardless of node
2961 static int __meminit next_active_region_index_in_nid(int index, int nid)
2963 for (index = index + 1; index < nr_nodemap_entries; index++)
2964 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
2965 return index;
2967 return -1;
2970 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
2972 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
2973 * Architectures may implement their own version but if add_active_range()
2974 * was used and there are no special requirements, this is a convenient
2975 * alternative
2977 int __meminit early_pfn_to_nid(unsigned long pfn)
2979 int i;
2981 for (i = 0; i < nr_nodemap_entries; i++) {
2982 unsigned long start_pfn = early_node_map[i].start_pfn;
2983 unsigned long end_pfn = early_node_map[i].end_pfn;
2985 if (start_pfn <= pfn && pfn < end_pfn)
2986 return early_node_map[i].nid;
2989 return 0;
2991 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
2993 /* Basic iterator support to walk early_node_map[] */
2994 #define for_each_active_range_index_in_nid(i, nid) \
2995 for (i = first_active_region_index_in_nid(nid); i != -1; \
2996 i = next_active_region_index_in_nid(i, nid))
2999 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
3000 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
3001 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
3003 * If an architecture guarantees that all ranges registered with
3004 * add_active_ranges() contain no holes and may be freed, this
3005 * this function may be used instead of calling free_bootmem() manually.
3007 void __init free_bootmem_with_active_regions(int nid,
3008 unsigned long max_low_pfn)
3010 int i;
3012 for_each_active_range_index_in_nid(i, nid) {
3013 unsigned long size_pages = 0;
3014 unsigned long end_pfn = early_node_map[i].end_pfn;
3016 if (early_node_map[i].start_pfn >= max_low_pfn)
3017 continue;
3019 if (end_pfn > max_low_pfn)
3020 end_pfn = max_low_pfn;
3022 size_pages = end_pfn - early_node_map[i].start_pfn;
3023 free_bootmem_node(NODE_DATA(early_node_map[i].nid),
3024 PFN_PHYS(early_node_map[i].start_pfn),
3025 size_pages << PAGE_SHIFT);
3029 void __init work_with_active_regions(int nid, work_fn_t work_fn, void *data)
3031 int i;
3032 int ret;
3034 for_each_active_range_index_in_nid(i, nid) {
3035 ret = work_fn(early_node_map[i].start_pfn,
3036 early_node_map[i].end_pfn, data);
3037 if (ret)
3038 break;
3042 * sparse_memory_present_with_active_regions - Call memory_present for each active range
3043 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
3045 * If an architecture guarantees that all ranges registered with
3046 * add_active_ranges() contain no holes and may be freed, this
3047 * function may be used instead of calling memory_present() manually.
3049 void __init sparse_memory_present_with_active_regions(int nid)
3051 int i;
3053 for_each_active_range_index_in_nid(i, nid)
3054 memory_present(early_node_map[i].nid,
3055 early_node_map[i].start_pfn,
3056 early_node_map[i].end_pfn);
3060 * push_node_boundaries - Push node boundaries to at least the requested boundary
3061 * @nid: The nid of the node to push the boundary for
3062 * @start_pfn: The start pfn of the node
3063 * @end_pfn: The end pfn of the node
3065 * In reserve-based hot-add, mem_map is allocated that is unused until hotadd
3066 * time. Specifically, on x86_64, SRAT will report ranges that can potentially
3067 * be hotplugged even though no physical memory exists. This function allows
3068 * an arch to push out the node boundaries so mem_map is allocated that can
3069 * be used later.
3071 #ifdef CONFIG_MEMORY_HOTPLUG_RESERVE
3072 void __init push_node_boundaries(unsigned int nid,
3073 unsigned long start_pfn, unsigned long end_pfn)
3075 mminit_dprintk(MMINIT_TRACE, "zoneboundary",
3076 "Entering push_node_boundaries(%u, %lu, %lu)\n",
3077 nid, start_pfn, end_pfn);
3079 /* Initialise the boundary for this node if necessary */
3080 if (node_boundary_end_pfn[nid] == 0)
3081 node_boundary_start_pfn[nid] = -1UL;
3083 /* Update the boundaries */
3084 if (node_boundary_start_pfn[nid] > start_pfn)
3085 node_boundary_start_pfn[nid] = start_pfn;
3086 if (node_boundary_end_pfn[nid] < end_pfn)
3087 node_boundary_end_pfn[nid] = end_pfn;
3090 /* If necessary, push the node boundary out for reserve hotadd */
3091 static void __meminit account_node_boundary(unsigned int nid,
3092 unsigned long *start_pfn, unsigned long *end_pfn)
3094 mminit_dprintk(MMINIT_TRACE, "zoneboundary",
3095 "Entering account_node_boundary(%u, %lu, %lu)\n",
3096 nid, *start_pfn, *end_pfn);
3098 /* Return if boundary information has not been provided */
3099 if (node_boundary_end_pfn[nid] == 0)
3100 return;
3102 /* Check the boundaries and update if necessary */
3103 if (node_boundary_start_pfn[nid] < *start_pfn)
3104 *start_pfn = node_boundary_start_pfn[nid];
3105 if (node_boundary_end_pfn[nid] > *end_pfn)
3106 *end_pfn = node_boundary_end_pfn[nid];
3108 #else
3109 void __init push_node_boundaries(unsigned int nid,
3110 unsigned long start_pfn, unsigned long end_pfn) {}
3112 static void __meminit account_node_boundary(unsigned int nid,
3113 unsigned long *start_pfn, unsigned long *end_pfn) {}
3114 #endif
3118 * get_pfn_range_for_nid - Return the start and end page frames for a node
3119 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
3120 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
3121 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
3123 * It returns the start and end page frame of a node based on information
3124 * provided by an arch calling add_active_range(). If called for a node
3125 * with no available memory, a warning is printed and the start and end
3126 * PFNs will be 0.
3128 void __meminit get_pfn_range_for_nid(unsigned int nid,
3129 unsigned long *start_pfn, unsigned long *end_pfn)
3131 int i;
3132 *start_pfn = -1UL;
3133 *end_pfn = 0;
3135 for_each_active_range_index_in_nid(i, nid) {
3136 *start_pfn = min(*start_pfn, early_node_map[i].start_pfn);
3137 *end_pfn = max(*end_pfn, early_node_map[i].end_pfn);
3140 if (*start_pfn == -1UL)
3141 *start_pfn = 0;
3143 /* Push the node boundaries out if requested */
3144 account_node_boundary(nid, start_pfn, end_pfn);
3148 * This finds a zone that can be used for ZONE_MOVABLE pages. The
3149 * assumption is made that zones within a node are ordered in monotonic
3150 * increasing memory addresses so that the "highest" populated zone is used
3152 static void __init find_usable_zone_for_movable(void)
3154 int zone_index;
3155 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
3156 if (zone_index == ZONE_MOVABLE)
3157 continue;
3159 if (arch_zone_highest_possible_pfn[zone_index] >
3160 arch_zone_lowest_possible_pfn[zone_index])
3161 break;
3164 VM_BUG_ON(zone_index == -1);
3165 movable_zone = zone_index;
3169 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
3170 * because it is sized independant of architecture. Unlike the other zones,
3171 * the starting point for ZONE_MOVABLE is not fixed. It may be different
3172 * in each node depending on the size of each node and how evenly kernelcore
3173 * is distributed. This helper function adjusts the zone ranges
3174 * provided by the architecture for a given node by using the end of the
3175 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
3176 * zones within a node are in order of monotonic increases memory addresses
3178 static void __meminit adjust_zone_range_for_zone_movable(int nid,
3179 unsigned long zone_type,
3180 unsigned long node_start_pfn,
3181 unsigned long node_end_pfn,
3182 unsigned long *zone_start_pfn,
3183 unsigned long *zone_end_pfn)
3185 /* Only adjust if ZONE_MOVABLE is on this node */
3186 if (zone_movable_pfn[nid]) {
3187 /* Size ZONE_MOVABLE */
3188 if (zone_type == ZONE_MOVABLE) {
3189 *zone_start_pfn = zone_movable_pfn[nid];
3190 *zone_end_pfn = min(node_end_pfn,
3191 arch_zone_highest_possible_pfn[movable_zone]);
3193 /* Adjust for ZONE_MOVABLE starting within this range */
3194 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
3195 *zone_end_pfn > zone_movable_pfn[nid]) {
3196 *zone_end_pfn = zone_movable_pfn[nid];
3198 /* Check if this whole range is within ZONE_MOVABLE */
3199 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
3200 *zone_start_pfn = *zone_end_pfn;
3205 * Return the number of pages a zone spans in a node, including holes
3206 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
3208 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
3209 unsigned long zone_type,
3210 unsigned long *ignored)
3212 unsigned long node_start_pfn, node_end_pfn;
3213 unsigned long zone_start_pfn, zone_end_pfn;
3215 /* Get the start and end of the node and zone */
3216 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
3217 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
3218 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
3219 adjust_zone_range_for_zone_movable(nid, zone_type,
3220 node_start_pfn, node_end_pfn,
3221 &zone_start_pfn, &zone_end_pfn);
3223 /* Check that this node has pages within the zone's required range */
3224 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
3225 return 0;
3227 /* Move the zone boundaries inside the node if necessary */
3228 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
3229 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
3231 /* Return the spanned pages */
3232 return zone_end_pfn - zone_start_pfn;
3236 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
3237 * then all holes in the requested range will be accounted for.
3239 static unsigned long __meminit __absent_pages_in_range(int nid,
3240 unsigned long range_start_pfn,
3241 unsigned long range_end_pfn)
3243 int i = 0;
3244 unsigned long prev_end_pfn = 0, hole_pages = 0;
3245 unsigned long start_pfn;
3247 /* Find the end_pfn of the first active range of pfns in the node */
3248 i = first_active_region_index_in_nid(nid);
3249 if (i == -1)
3250 return 0;
3252 prev_end_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
3254 /* Account for ranges before physical memory on this node */
3255 if (early_node_map[i].start_pfn > range_start_pfn)
3256 hole_pages = prev_end_pfn - range_start_pfn;
3258 /* Find all holes for the zone within the node */
3259 for (; i != -1; i = next_active_region_index_in_nid(i, nid)) {
3261 /* No need to continue if prev_end_pfn is outside the zone */
3262 if (prev_end_pfn >= range_end_pfn)
3263 break;
3265 /* Make sure the end of the zone is not within the hole */
3266 start_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
3267 prev_end_pfn = max(prev_end_pfn, range_start_pfn);
3269 /* Update the hole size cound and move on */
3270 if (start_pfn > range_start_pfn) {
3271 BUG_ON(prev_end_pfn > start_pfn);
3272 hole_pages += start_pfn - prev_end_pfn;
3274 prev_end_pfn = early_node_map[i].end_pfn;
3277 /* Account for ranges past physical memory on this node */
3278 if (range_end_pfn > prev_end_pfn)
3279 hole_pages += range_end_pfn -
3280 max(range_start_pfn, prev_end_pfn);
3282 return hole_pages;
3286 * absent_pages_in_range - Return number of page frames in holes within a range
3287 * @start_pfn: The start PFN to start searching for holes
3288 * @end_pfn: The end PFN to stop searching for holes
3290 * It returns the number of pages frames in memory holes within a range.
3292 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
3293 unsigned long end_pfn)
3295 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
3298 /* Return the number of page frames in holes in a zone on a node */
3299 static unsigned long __meminit zone_absent_pages_in_node(int nid,
3300 unsigned long zone_type,
3301 unsigned long *ignored)
3303 unsigned long node_start_pfn, node_end_pfn;
3304 unsigned long zone_start_pfn, zone_end_pfn;
3306 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
3307 zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type],
3308 node_start_pfn);
3309 zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type],
3310 node_end_pfn);
3312 adjust_zone_range_for_zone_movable(nid, zone_type,
3313 node_start_pfn, node_end_pfn,
3314 &zone_start_pfn, &zone_end_pfn);
3315 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
3318 #else
3319 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
3320 unsigned long zone_type,
3321 unsigned long *zones_size)
3323 return zones_size[zone_type];
3326 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
3327 unsigned long zone_type,
3328 unsigned long *zholes_size)
3330 if (!zholes_size)
3331 return 0;
3333 return zholes_size[zone_type];
3336 #endif
3338 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
3339 unsigned long *zones_size, unsigned long *zholes_size)
3341 unsigned long realtotalpages, totalpages = 0;
3342 enum zone_type i;
3344 for (i = 0; i < MAX_NR_ZONES; i++)
3345 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
3346 zones_size);
3347 pgdat->node_spanned_pages = totalpages;
3349 realtotalpages = totalpages;
3350 for (i = 0; i < MAX_NR_ZONES; i++)
3351 realtotalpages -=
3352 zone_absent_pages_in_node(pgdat->node_id, i,
3353 zholes_size);
3354 pgdat->node_present_pages = realtotalpages;
3355 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
3356 realtotalpages);
3359 #ifndef CONFIG_SPARSEMEM
3361 * Calculate the size of the zone->blockflags rounded to an unsigned long
3362 * Start by making sure zonesize is a multiple of pageblock_order by rounding
3363 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
3364 * round what is now in bits to nearest long in bits, then return it in
3365 * bytes.
3367 static unsigned long __init usemap_size(unsigned long zonesize)
3369 unsigned long usemapsize;
3371 usemapsize = roundup(zonesize, pageblock_nr_pages);
3372 usemapsize = usemapsize >> pageblock_order;
3373 usemapsize *= NR_PAGEBLOCK_BITS;
3374 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
3376 return usemapsize / 8;
3379 static void __init setup_usemap(struct pglist_data *pgdat,
3380 struct zone *zone, unsigned long zonesize)
3382 unsigned long usemapsize = usemap_size(zonesize);
3383 zone->pageblock_flags = NULL;
3384 if (usemapsize) {
3385 zone->pageblock_flags = alloc_bootmem_node(pgdat, usemapsize);
3386 memset(zone->pageblock_flags, 0, usemapsize);
3389 #else
3390 static void inline setup_usemap(struct pglist_data *pgdat,
3391 struct zone *zone, unsigned long zonesize) {}
3392 #endif /* CONFIG_SPARSEMEM */
3394 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
3396 /* Return a sensible default order for the pageblock size. */
3397 static inline int pageblock_default_order(void)
3399 if (HPAGE_SHIFT > PAGE_SHIFT)
3400 return HUGETLB_PAGE_ORDER;
3402 return MAX_ORDER-1;
3405 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
3406 static inline void __init set_pageblock_order(unsigned int order)
3408 /* Check that pageblock_nr_pages has not already been setup */
3409 if (pageblock_order)
3410 return;
3413 * Assume the largest contiguous order of interest is a huge page.
3414 * This value may be variable depending on boot parameters on IA64
3416 pageblock_order = order;
3418 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
3421 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
3422 * and pageblock_default_order() are unused as pageblock_order is set
3423 * at compile-time. See include/linux/pageblock-flags.h for the values of
3424 * pageblock_order based on the kernel config
3426 static inline int pageblock_default_order(unsigned int order)
3428 return MAX_ORDER-1;
3430 #define set_pageblock_order(x) do {} while (0)
3432 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
3435 * Set up the zone data structures:
3436 * - mark all pages reserved
3437 * - mark all memory queues empty
3438 * - clear the memory bitmaps
3440 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
3441 unsigned long *zones_size, unsigned long *zholes_size)
3443 enum zone_type j;
3444 int nid = pgdat->node_id;
3445 unsigned long zone_start_pfn = pgdat->node_start_pfn;
3446 int ret;
3448 pgdat_resize_init(pgdat);
3449 pgdat->nr_zones = 0;
3450 init_waitqueue_head(&pgdat->kswapd_wait);
3451 pgdat->kswapd_max_order = 0;
3452 pgdat_page_cgroup_init(pgdat);
3454 for (j = 0; j < MAX_NR_ZONES; j++) {
3455 struct zone *zone = pgdat->node_zones + j;
3456 unsigned long size, realsize, memmap_pages;
3457 enum lru_list l;
3459 size = zone_spanned_pages_in_node(nid, j, zones_size);
3460 realsize = size - zone_absent_pages_in_node(nid, j,
3461 zholes_size);
3464 * Adjust realsize so that it accounts for how much memory
3465 * is used by this zone for memmap. This affects the watermark
3466 * and per-cpu initialisations
3468 memmap_pages =
3469 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT;
3470 if (realsize >= memmap_pages) {
3471 realsize -= memmap_pages;
3472 printk(KERN_DEBUG
3473 " %s zone: %lu pages used for memmap\n",
3474 zone_names[j], memmap_pages);
3475 } else
3476 printk(KERN_WARNING
3477 " %s zone: %lu pages exceeds realsize %lu\n",
3478 zone_names[j], memmap_pages, realsize);
3480 /* Account for reserved pages */
3481 if (j == 0 && realsize > dma_reserve) {
3482 realsize -= dma_reserve;
3483 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
3484 zone_names[0], dma_reserve);
3487 if (!is_highmem_idx(j))
3488 nr_kernel_pages += realsize;
3489 nr_all_pages += realsize;
3491 zone->spanned_pages = size;
3492 zone->present_pages = realsize;
3493 #ifdef CONFIG_NUMA
3494 zone->node = nid;
3495 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
3496 / 100;
3497 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
3498 #endif
3499 zone->name = zone_names[j];
3500 spin_lock_init(&zone->lock);
3501 spin_lock_init(&zone->lru_lock);
3502 zone_seqlock_init(zone);
3503 zone->zone_pgdat = pgdat;
3505 zone->prev_priority = DEF_PRIORITY;
3507 zone_pcp_init(zone);
3508 for_each_lru(l) {
3509 INIT_LIST_HEAD(&zone->lru[l].list);
3510 zone->lru[l].nr_scan = 0;
3512 zone->recent_rotated[0] = 0;
3513 zone->recent_rotated[1] = 0;
3514 zone->recent_scanned[0] = 0;
3515 zone->recent_scanned[1] = 0;
3516 zap_zone_vm_stats(zone);
3517 zone->flags = 0;
3518 if (!size)
3519 continue;
3521 set_pageblock_order(pageblock_default_order());
3522 setup_usemap(pgdat, zone, size);
3523 ret = init_currently_empty_zone(zone, zone_start_pfn,
3524 size, MEMMAP_EARLY);
3525 BUG_ON(ret);
3526 memmap_init(size, nid, j, zone_start_pfn);
3527 zone_start_pfn += size;
3531 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
3533 /* Skip empty nodes */
3534 if (!pgdat->node_spanned_pages)
3535 return;
3537 #ifdef CONFIG_FLAT_NODE_MEM_MAP
3538 /* ia64 gets its own node_mem_map, before this, without bootmem */
3539 if (!pgdat->node_mem_map) {
3540 unsigned long size, start, end;
3541 struct page *map;
3544 * The zone's endpoints aren't required to be MAX_ORDER
3545 * aligned but the node_mem_map endpoints must be in order
3546 * for the buddy allocator to function correctly.
3548 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
3549 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
3550 end = ALIGN(end, MAX_ORDER_NR_PAGES);
3551 size = (end - start) * sizeof(struct page);
3552 map = alloc_remap(pgdat->node_id, size);
3553 if (!map)
3554 map = alloc_bootmem_node(pgdat, size);
3555 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
3557 #ifndef CONFIG_NEED_MULTIPLE_NODES
3559 * With no DISCONTIG, the global mem_map is just set as node 0's
3561 if (pgdat == NODE_DATA(0)) {
3562 mem_map = NODE_DATA(0)->node_mem_map;
3563 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3564 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
3565 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
3566 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
3568 #endif
3569 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
3572 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
3573 unsigned long node_start_pfn, unsigned long *zholes_size)
3575 pg_data_t *pgdat = NODE_DATA(nid);
3577 pgdat->node_id = nid;
3578 pgdat->node_start_pfn = node_start_pfn;
3579 calculate_node_totalpages(pgdat, zones_size, zholes_size);
3581 alloc_node_mem_map(pgdat);
3582 #ifdef CONFIG_FLAT_NODE_MEM_MAP
3583 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
3584 nid, (unsigned long)pgdat,
3585 (unsigned long)pgdat->node_mem_map);
3586 #endif
3588 free_area_init_core(pgdat, zones_size, zholes_size);
3591 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3593 #if MAX_NUMNODES > 1
3595 * Figure out the number of possible node ids.
3597 static void __init setup_nr_node_ids(void)
3599 unsigned int node;
3600 unsigned int highest = 0;
3602 for_each_node_mask(node, node_possible_map)
3603 highest = node;
3604 nr_node_ids = highest + 1;
3606 #else
3607 static inline void setup_nr_node_ids(void)
3610 #endif
3613 * add_active_range - Register a range of PFNs backed by physical memory
3614 * @nid: The node ID the range resides on
3615 * @start_pfn: The start PFN of the available physical memory
3616 * @end_pfn: The end PFN of the available physical memory
3618 * These ranges are stored in an early_node_map[] and later used by
3619 * free_area_init_nodes() to calculate zone sizes and holes. If the
3620 * range spans a memory hole, it is up to the architecture to ensure
3621 * the memory is not freed by the bootmem allocator. If possible
3622 * the range being registered will be merged with existing ranges.
3624 void __init add_active_range(unsigned int nid, unsigned long start_pfn,
3625 unsigned long end_pfn)
3627 int i;
3629 mminit_dprintk(MMINIT_TRACE, "memory_register",
3630 "Entering add_active_range(%d, %#lx, %#lx) "
3631 "%d entries of %d used\n",
3632 nid, start_pfn, end_pfn,
3633 nr_nodemap_entries, MAX_ACTIVE_REGIONS);
3635 mminit_validate_memmodel_limits(&start_pfn, &end_pfn);
3637 /* Merge with existing active regions if possible */
3638 for (i = 0; i < nr_nodemap_entries; i++) {
3639 if (early_node_map[i].nid != nid)
3640 continue;
3642 /* Skip if an existing region covers this new one */
3643 if (start_pfn >= early_node_map[i].start_pfn &&
3644 end_pfn <= early_node_map[i].end_pfn)
3645 return;
3647 /* Merge forward if suitable */
3648 if (start_pfn <= early_node_map[i].end_pfn &&
3649 end_pfn > early_node_map[i].end_pfn) {
3650 early_node_map[i].end_pfn = end_pfn;
3651 return;
3654 /* Merge backward if suitable */
3655 if (start_pfn < early_node_map[i].end_pfn &&
3656 end_pfn >= early_node_map[i].start_pfn) {
3657 early_node_map[i].start_pfn = start_pfn;
3658 return;
3662 /* Check that early_node_map is large enough */
3663 if (i >= MAX_ACTIVE_REGIONS) {
3664 printk(KERN_CRIT "More than %d memory regions, truncating\n",
3665 MAX_ACTIVE_REGIONS);
3666 return;
3669 early_node_map[i].nid = nid;
3670 early_node_map[i].start_pfn = start_pfn;
3671 early_node_map[i].end_pfn = end_pfn;
3672 nr_nodemap_entries = i + 1;
3676 * remove_active_range - Shrink an existing registered range of PFNs
3677 * @nid: The node id the range is on that should be shrunk
3678 * @start_pfn: The new PFN of the range
3679 * @end_pfn: The new PFN of the range
3681 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
3682 * The map is kept near the end physical page range that has already been
3683 * registered. This function allows an arch to shrink an existing registered
3684 * range.
3686 void __init remove_active_range(unsigned int nid, unsigned long start_pfn,
3687 unsigned long end_pfn)
3689 int i, j;
3690 int removed = 0;
3692 printk(KERN_DEBUG "remove_active_range (%d, %lu, %lu)\n",
3693 nid, start_pfn, end_pfn);
3695 /* Find the old active region end and shrink */
3696 for_each_active_range_index_in_nid(i, nid) {
3697 if (early_node_map[i].start_pfn >= start_pfn &&
3698 early_node_map[i].end_pfn <= end_pfn) {
3699 /* clear it */
3700 early_node_map[i].start_pfn = 0;
3701 early_node_map[i].end_pfn = 0;
3702 removed = 1;
3703 continue;
3705 if (early_node_map[i].start_pfn < start_pfn &&
3706 early_node_map[i].end_pfn > start_pfn) {
3707 unsigned long temp_end_pfn = early_node_map[i].end_pfn;
3708 early_node_map[i].end_pfn = start_pfn;
3709 if (temp_end_pfn > end_pfn)
3710 add_active_range(nid, end_pfn, temp_end_pfn);
3711 continue;
3713 if (early_node_map[i].start_pfn >= start_pfn &&
3714 early_node_map[i].end_pfn > end_pfn &&
3715 early_node_map[i].start_pfn < end_pfn) {
3716 early_node_map[i].start_pfn = end_pfn;
3717 continue;
3721 if (!removed)
3722 return;
3724 /* remove the blank ones */
3725 for (i = nr_nodemap_entries - 1; i > 0; i--) {
3726 if (early_node_map[i].nid != nid)
3727 continue;
3728 if (early_node_map[i].end_pfn)
3729 continue;
3730 /* we found it, get rid of it */
3731 for (j = i; j < nr_nodemap_entries - 1; j++)
3732 memcpy(&early_node_map[j], &early_node_map[j+1],
3733 sizeof(early_node_map[j]));
3734 j = nr_nodemap_entries - 1;
3735 memset(&early_node_map[j], 0, sizeof(early_node_map[j]));
3736 nr_nodemap_entries--;
3741 * remove_all_active_ranges - Remove all currently registered regions
3743 * During discovery, it may be found that a table like SRAT is invalid
3744 * and an alternative discovery method must be used. This function removes
3745 * all currently registered regions.
3747 void __init remove_all_active_ranges(void)
3749 memset(early_node_map, 0, sizeof(early_node_map));
3750 nr_nodemap_entries = 0;
3751 #ifdef CONFIG_MEMORY_HOTPLUG_RESERVE
3752 memset(node_boundary_start_pfn, 0, sizeof(node_boundary_start_pfn));
3753 memset(node_boundary_end_pfn, 0, sizeof(node_boundary_end_pfn));
3754 #endif /* CONFIG_MEMORY_HOTPLUG_RESERVE */
3757 /* Compare two active node_active_regions */
3758 static int __init cmp_node_active_region(const void *a, const void *b)
3760 struct node_active_region *arange = (struct node_active_region *)a;
3761 struct node_active_region *brange = (struct node_active_region *)b;
3763 /* Done this way to avoid overflows */
3764 if (arange->start_pfn > brange->start_pfn)
3765 return 1;
3766 if (arange->start_pfn < brange->start_pfn)
3767 return -1;
3769 return 0;
3772 /* sort the node_map by start_pfn */
3773 static void __init sort_node_map(void)
3775 sort(early_node_map, (size_t)nr_nodemap_entries,
3776 sizeof(struct node_active_region),
3777 cmp_node_active_region, NULL);
3780 /* Find the lowest pfn for a node */
3781 static unsigned long __init find_min_pfn_for_node(int nid)
3783 int i;
3784 unsigned long min_pfn = ULONG_MAX;
3786 /* Assuming a sorted map, the first range found has the starting pfn */
3787 for_each_active_range_index_in_nid(i, nid)
3788 min_pfn = min(min_pfn, early_node_map[i].start_pfn);
3790 if (min_pfn == ULONG_MAX) {
3791 printk(KERN_WARNING
3792 "Could not find start_pfn for node %d\n", nid);
3793 return 0;
3796 return min_pfn;
3800 * find_min_pfn_with_active_regions - Find the minimum PFN registered
3802 * It returns the minimum PFN based on information provided via
3803 * add_active_range().
3805 unsigned long __init find_min_pfn_with_active_regions(void)
3807 return find_min_pfn_for_node(MAX_NUMNODES);
3811 * early_calculate_totalpages()
3812 * Sum pages in active regions for movable zone.
3813 * Populate N_HIGH_MEMORY for calculating usable_nodes.
3815 static unsigned long __init early_calculate_totalpages(void)
3817 int i;
3818 unsigned long totalpages = 0;
3820 for (i = 0; i < nr_nodemap_entries; i++) {
3821 unsigned long pages = early_node_map[i].end_pfn -
3822 early_node_map[i].start_pfn;
3823 totalpages += pages;
3824 if (pages)
3825 node_set_state(early_node_map[i].nid, N_HIGH_MEMORY);
3827 return totalpages;
3831 * Find the PFN the Movable zone begins in each node. Kernel memory
3832 * is spread evenly between nodes as long as the nodes have enough
3833 * memory. When they don't, some nodes will have more kernelcore than
3834 * others
3836 static void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn)
3838 int i, nid;
3839 unsigned long usable_startpfn;
3840 unsigned long kernelcore_node, kernelcore_remaining;
3841 unsigned long totalpages = early_calculate_totalpages();
3842 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
3845 * If movablecore was specified, calculate what size of
3846 * kernelcore that corresponds so that memory usable for
3847 * any allocation type is evenly spread. If both kernelcore
3848 * and movablecore are specified, then the value of kernelcore
3849 * will be used for required_kernelcore if it's greater than
3850 * what movablecore would have allowed.
3852 if (required_movablecore) {
3853 unsigned long corepages;
3856 * Round-up so that ZONE_MOVABLE is at least as large as what
3857 * was requested by the user
3859 required_movablecore =
3860 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
3861 corepages = totalpages - required_movablecore;
3863 required_kernelcore = max(required_kernelcore, corepages);
3866 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
3867 if (!required_kernelcore)
3868 return;
3870 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
3871 find_usable_zone_for_movable();
3872 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
3874 restart:
3875 /* Spread kernelcore memory as evenly as possible throughout nodes */
3876 kernelcore_node = required_kernelcore / usable_nodes;
3877 for_each_node_state(nid, N_HIGH_MEMORY) {
3879 * Recalculate kernelcore_node if the division per node
3880 * now exceeds what is necessary to satisfy the requested
3881 * amount of memory for the kernel
3883 if (required_kernelcore < kernelcore_node)
3884 kernelcore_node = required_kernelcore / usable_nodes;
3887 * As the map is walked, we track how much memory is usable
3888 * by the kernel using kernelcore_remaining. When it is
3889 * 0, the rest of the node is usable by ZONE_MOVABLE
3891 kernelcore_remaining = kernelcore_node;
3893 /* Go through each range of PFNs within this node */
3894 for_each_active_range_index_in_nid(i, nid) {
3895 unsigned long start_pfn, end_pfn;
3896 unsigned long size_pages;
3898 start_pfn = max(early_node_map[i].start_pfn,
3899 zone_movable_pfn[nid]);
3900 end_pfn = early_node_map[i].end_pfn;
3901 if (start_pfn >= end_pfn)
3902 continue;
3904 /* Account for what is only usable for kernelcore */
3905 if (start_pfn < usable_startpfn) {
3906 unsigned long kernel_pages;
3907 kernel_pages = min(end_pfn, usable_startpfn)
3908 - start_pfn;
3910 kernelcore_remaining -= min(kernel_pages,
3911 kernelcore_remaining);
3912 required_kernelcore -= min(kernel_pages,
3913 required_kernelcore);
3915 /* Continue if range is now fully accounted */
3916 if (end_pfn <= usable_startpfn) {
3919 * Push zone_movable_pfn to the end so
3920 * that if we have to rebalance
3921 * kernelcore across nodes, we will
3922 * not double account here
3924 zone_movable_pfn[nid] = end_pfn;
3925 continue;
3927 start_pfn = usable_startpfn;
3931 * The usable PFN range for ZONE_MOVABLE is from
3932 * start_pfn->end_pfn. Calculate size_pages as the
3933 * number of pages used as kernelcore
3935 size_pages = end_pfn - start_pfn;
3936 if (size_pages > kernelcore_remaining)
3937 size_pages = kernelcore_remaining;
3938 zone_movable_pfn[nid] = start_pfn + size_pages;
3941 * Some kernelcore has been met, update counts and
3942 * break if the kernelcore for this node has been
3943 * satisified
3945 required_kernelcore -= min(required_kernelcore,
3946 size_pages);
3947 kernelcore_remaining -= size_pages;
3948 if (!kernelcore_remaining)
3949 break;
3954 * If there is still required_kernelcore, we do another pass with one
3955 * less node in the count. This will push zone_movable_pfn[nid] further
3956 * along on the nodes that still have memory until kernelcore is
3957 * satisified
3959 usable_nodes--;
3960 if (usable_nodes && required_kernelcore > usable_nodes)
3961 goto restart;
3963 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
3964 for (nid = 0; nid < MAX_NUMNODES; nid++)
3965 zone_movable_pfn[nid] =
3966 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
3969 /* Any regular memory on that node ? */
3970 static void check_for_regular_memory(pg_data_t *pgdat)
3972 #ifdef CONFIG_HIGHMEM
3973 enum zone_type zone_type;
3975 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) {
3976 struct zone *zone = &pgdat->node_zones[zone_type];
3977 if (zone->present_pages)
3978 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY);
3980 #endif
3984 * free_area_init_nodes - Initialise all pg_data_t and zone data
3985 * @max_zone_pfn: an array of max PFNs for each zone
3987 * This will call free_area_init_node() for each active node in the system.
3988 * Using the page ranges provided by add_active_range(), the size of each
3989 * zone in each node and their holes is calculated. If the maximum PFN
3990 * between two adjacent zones match, it is assumed that the zone is empty.
3991 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
3992 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
3993 * starts where the previous one ended. For example, ZONE_DMA32 starts
3994 * at arch_max_dma_pfn.
3996 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
3998 unsigned long nid;
3999 int i;
4001 /* Sort early_node_map as initialisation assumes it is sorted */
4002 sort_node_map();
4004 /* Record where the zone boundaries are */
4005 memset(arch_zone_lowest_possible_pfn, 0,
4006 sizeof(arch_zone_lowest_possible_pfn));
4007 memset(arch_zone_highest_possible_pfn, 0,
4008 sizeof(arch_zone_highest_possible_pfn));
4009 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
4010 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
4011 for (i = 1; i < MAX_NR_ZONES; i++) {
4012 if (i == ZONE_MOVABLE)
4013 continue;
4014 arch_zone_lowest_possible_pfn[i] =
4015 arch_zone_highest_possible_pfn[i-1];
4016 arch_zone_highest_possible_pfn[i] =
4017 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
4019 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
4020 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
4022 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4023 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
4024 find_zone_movable_pfns_for_nodes(zone_movable_pfn);
4026 /* Print out the zone ranges */
4027 printk("Zone PFN ranges:\n");
4028 for (i = 0; i < MAX_NR_ZONES; i++) {
4029 if (i == ZONE_MOVABLE)
4030 continue;
4031 printk(" %-8s %0#10lx -> %0#10lx\n",
4032 zone_names[i],
4033 arch_zone_lowest_possible_pfn[i],
4034 arch_zone_highest_possible_pfn[i]);
4037 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4038 printk("Movable zone start PFN for each node\n");
4039 for (i = 0; i < MAX_NUMNODES; i++) {
4040 if (zone_movable_pfn[i])
4041 printk(" Node %d: %lu\n", i, zone_movable_pfn[i]);
4044 /* Print out the early_node_map[] */
4045 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries);
4046 for (i = 0; i < nr_nodemap_entries; i++)
4047 printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map[i].nid,
4048 early_node_map[i].start_pfn,
4049 early_node_map[i].end_pfn);
4051 /* Initialise every node */
4052 mminit_verify_pageflags_layout();
4053 setup_nr_node_ids();
4054 for_each_online_node(nid) {
4055 pg_data_t *pgdat = NODE_DATA(nid);
4056 free_area_init_node(nid, NULL,
4057 find_min_pfn_for_node(nid), NULL);
4059 /* Any memory on that node */
4060 if (pgdat->node_present_pages)
4061 node_set_state(nid, N_HIGH_MEMORY);
4062 check_for_regular_memory(pgdat);
4066 static int __init cmdline_parse_core(char *p, unsigned long *core)
4068 unsigned long long coremem;
4069 if (!p)
4070 return -EINVAL;
4072 coremem = memparse(p, &p);
4073 *core = coremem >> PAGE_SHIFT;
4075 /* Paranoid check that UL is enough for the coremem value */
4076 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
4078 return 0;
4082 * kernelcore=size sets the amount of memory for use for allocations that
4083 * cannot be reclaimed or migrated.
4085 static int __init cmdline_parse_kernelcore(char *p)
4087 return cmdline_parse_core(p, &required_kernelcore);
4091 * movablecore=size sets the amount of memory for use for allocations that
4092 * can be reclaimed or migrated.
4094 static int __init cmdline_parse_movablecore(char *p)
4096 return cmdline_parse_core(p, &required_movablecore);
4099 early_param("kernelcore", cmdline_parse_kernelcore);
4100 early_param("movablecore", cmdline_parse_movablecore);
4102 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4105 * set_dma_reserve - set the specified number of pages reserved in the first zone
4106 * @new_dma_reserve: The number of pages to mark reserved
4108 * The per-cpu batchsize and zone watermarks are determined by present_pages.
4109 * In the DMA zone, a significant percentage may be consumed by kernel image
4110 * and other unfreeable allocations which can skew the watermarks badly. This
4111 * function may optionally be used to account for unfreeable pages in the
4112 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
4113 * smaller per-cpu batchsize.
4115 void __init set_dma_reserve(unsigned long new_dma_reserve)
4117 dma_reserve = new_dma_reserve;
4120 #ifndef CONFIG_NEED_MULTIPLE_NODES
4121 struct pglist_data __refdata contig_page_data = { .bdata = &bootmem_node_data[0] };
4122 EXPORT_SYMBOL(contig_page_data);
4123 #endif
4125 void __init free_area_init(unsigned long *zones_size)
4127 free_area_init_node(0, zones_size,
4128 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
4131 static int page_alloc_cpu_notify(struct notifier_block *self,
4132 unsigned long action, void *hcpu)
4134 int cpu = (unsigned long)hcpu;
4136 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
4137 drain_pages(cpu);
4140 * Spill the event counters of the dead processor
4141 * into the current processors event counters.
4142 * This artificially elevates the count of the current
4143 * processor.
4145 vm_events_fold_cpu(cpu);
4148 * Zero the differential counters of the dead processor
4149 * so that the vm statistics are consistent.
4151 * This is only okay since the processor is dead and cannot
4152 * race with what we are doing.
4154 refresh_cpu_vm_stats(cpu);
4156 return NOTIFY_OK;
4159 void __init page_alloc_init(void)
4161 hotcpu_notifier(page_alloc_cpu_notify, 0);
4165 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
4166 * or min_free_kbytes changes.
4168 static void calculate_totalreserve_pages(void)
4170 struct pglist_data *pgdat;
4171 unsigned long reserve_pages = 0;
4172 enum zone_type i, j;
4174 for_each_online_pgdat(pgdat) {
4175 for (i = 0; i < MAX_NR_ZONES; i++) {
4176 struct zone *zone = pgdat->node_zones + i;
4177 unsigned long max = 0;
4179 /* Find valid and maximum lowmem_reserve in the zone */
4180 for (j = i; j < MAX_NR_ZONES; j++) {
4181 if (zone->lowmem_reserve[j] > max)
4182 max = zone->lowmem_reserve[j];
4185 /* we treat pages_high as reserved pages. */
4186 max += zone->pages_high;
4188 if (max > zone->present_pages)
4189 max = zone->present_pages;
4190 reserve_pages += max;
4193 totalreserve_pages = reserve_pages;
4197 * setup_per_zone_lowmem_reserve - called whenever
4198 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
4199 * has a correct pages reserved value, so an adequate number of
4200 * pages are left in the zone after a successful __alloc_pages().
4202 static void setup_per_zone_lowmem_reserve(void)
4204 struct pglist_data *pgdat;
4205 enum zone_type j, idx;
4207 for_each_online_pgdat(pgdat) {
4208 for (j = 0; j < MAX_NR_ZONES; j++) {
4209 struct zone *zone = pgdat->node_zones + j;
4210 unsigned long present_pages = zone->present_pages;
4212 zone->lowmem_reserve[j] = 0;
4214 idx = j;
4215 while (idx) {
4216 struct zone *lower_zone;
4218 idx--;
4220 if (sysctl_lowmem_reserve_ratio[idx] < 1)
4221 sysctl_lowmem_reserve_ratio[idx] = 1;
4223 lower_zone = pgdat->node_zones + idx;
4224 lower_zone->lowmem_reserve[j] = present_pages /
4225 sysctl_lowmem_reserve_ratio[idx];
4226 present_pages += lower_zone->present_pages;
4231 /* update totalreserve_pages */
4232 calculate_totalreserve_pages();
4236 * setup_per_zone_pages_min - called when min_free_kbytes changes.
4238 * Ensures that the pages_{min,low,high} values for each zone are set correctly
4239 * with respect to min_free_kbytes.
4241 void setup_per_zone_pages_min(void)
4243 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
4244 unsigned long lowmem_pages = 0;
4245 struct zone *zone;
4246 unsigned long flags;
4248 /* Calculate total number of !ZONE_HIGHMEM pages */
4249 for_each_zone(zone) {
4250 if (!is_highmem(zone))
4251 lowmem_pages += zone->present_pages;
4254 for_each_zone(zone) {
4255 u64 tmp;
4257 spin_lock_irqsave(&zone->lock, flags);
4258 tmp = (u64)pages_min * zone->present_pages;
4259 do_div(tmp, lowmem_pages);
4260 if (is_highmem(zone)) {
4262 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
4263 * need highmem pages, so cap pages_min to a small
4264 * value here.
4266 * The (pages_high-pages_low) and (pages_low-pages_min)
4267 * deltas controls asynch page reclaim, and so should
4268 * not be capped for highmem.
4270 int min_pages;
4272 min_pages = zone->present_pages / 1024;
4273 if (min_pages < SWAP_CLUSTER_MAX)
4274 min_pages = SWAP_CLUSTER_MAX;
4275 if (min_pages > 128)
4276 min_pages = 128;
4277 zone->pages_min = min_pages;
4278 } else {
4280 * If it's a lowmem zone, reserve a number of pages
4281 * proportionate to the zone's size.
4283 zone->pages_min = tmp;
4286 zone->pages_low = zone->pages_min + (tmp >> 2);
4287 zone->pages_high = zone->pages_min + (tmp >> 1);
4288 setup_zone_migrate_reserve(zone);
4289 spin_unlock_irqrestore(&zone->lock, flags);
4292 /* update totalreserve_pages */
4293 calculate_totalreserve_pages();
4297 * setup_per_zone_inactive_ratio - called when min_free_kbytes changes.
4299 * The inactive anon list should be small enough that the VM never has to
4300 * do too much work, but large enough that each inactive page has a chance
4301 * to be referenced again before it is swapped out.
4303 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
4304 * INACTIVE_ANON pages on this zone's LRU, maintained by the
4305 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
4306 * the anonymous pages are kept on the inactive list.
4308 * total target max
4309 * memory ratio inactive anon
4310 * -------------------------------------
4311 * 10MB 1 5MB
4312 * 100MB 1 50MB
4313 * 1GB 3 250MB
4314 * 10GB 10 0.9GB
4315 * 100GB 31 3GB
4316 * 1TB 101 10GB
4317 * 10TB 320 32GB
4319 void setup_per_zone_inactive_ratio(void)
4321 struct zone *zone;
4323 for_each_zone(zone) {
4324 unsigned int gb, ratio;
4326 /* Zone size in gigabytes */
4327 gb = zone->present_pages >> (30 - PAGE_SHIFT);
4328 ratio = int_sqrt(10 * gb);
4329 if (!ratio)
4330 ratio = 1;
4332 zone->inactive_ratio = ratio;
4337 * Initialise min_free_kbytes.
4339 * For small machines we want it small (128k min). For large machines
4340 * we want it large (64MB max). But it is not linear, because network
4341 * bandwidth does not increase linearly with machine size. We use
4343 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
4344 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
4346 * which yields
4348 * 16MB: 512k
4349 * 32MB: 724k
4350 * 64MB: 1024k
4351 * 128MB: 1448k
4352 * 256MB: 2048k
4353 * 512MB: 2896k
4354 * 1024MB: 4096k
4355 * 2048MB: 5792k
4356 * 4096MB: 8192k
4357 * 8192MB: 11584k
4358 * 16384MB: 16384k
4360 static int __init init_per_zone_pages_min(void)
4362 unsigned long lowmem_kbytes;
4364 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
4366 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
4367 if (min_free_kbytes < 128)
4368 min_free_kbytes = 128;
4369 if (min_free_kbytes > 65536)
4370 min_free_kbytes = 65536;
4371 setup_per_zone_pages_min();
4372 setup_per_zone_lowmem_reserve();
4373 setup_per_zone_inactive_ratio();
4374 return 0;
4376 module_init(init_per_zone_pages_min)
4379 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
4380 * that we can call two helper functions whenever min_free_kbytes
4381 * changes.
4383 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
4384 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4386 proc_dointvec(table, write, file, buffer, length, ppos);
4387 if (write)
4388 setup_per_zone_pages_min();
4389 return 0;
4392 #ifdef CONFIG_NUMA
4393 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
4394 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4396 struct zone *zone;
4397 int rc;
4399 rc = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
4400 if (rc)
4401 return rc;
4403 for_each_zone(zone)
4404 zone->min_unmapped_pages = (zone->present_pages *
4405 sysctl_min_unmapped_ratio) / 100;
4406 return 0;
4409 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
4410 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4412 struct zone *zone;
4413 int rc;
4415 rc = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
4416 if (rc)
4417 return rc;
4419 for_each_zone(zone)
4420 zone->min_slab_pages = (zone->present_pages *
4421 sysctl_min_slab_ratio) / 100;
4422 return 0;
4424 #endif
4427 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
4428 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
4429 * whenever sysctl_lowmem_reserve_ratio changes.
4431 * The reserve ratio obviously has absolutely no relation with the
4432 * pages_min watermarks. The lowmem reserve ratio can only make sense
4433 * if in function of the boot time zone sizes.
4435 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
4436 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4438 proc_dointvec_minmax(table, write, file, buffer, length, ppos);
4439 setup_per_zone_lowmem_reserve();
4440 return 0;
4444 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
4445 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
4446 * can have before it gets flushed back to buddy allocator.
4449 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
4450 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4452 struct zone *zone;
4453 unsigned int cpu;
4454 int ret;
4456 ret = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
4457 if (!write || (ret == -EINVAL))
4458 return ret;
4459 for_each_zone(zone) {
4460 for_each_online_cpu(cpu) {
4461 unsigned long high;
4462 high = zone->present_pages / percpu_pagelist_fraction;
4463 setup_pagelist_highmark(zone_pcp(zone, cpu), high);
4466 return 0;
4469 int hashdist = HASHDIST_DEFAULT;
4471 #ifdef CONFIG_NUMA
4472 static int __init set_hashdist(char *str)
4474 if (!str)
4475 return 0;
4476 hashdist = simple_strtoul(str, &str, 0);
4477 return 1;
4479 __setup("hashdist=", set_hashdist);
4480 #endif
4483 * allocate a large system hash table from bootmem
4484 * - it is assumed that the hash table must contain an exact power-of-2
4485 * quantity of entries
4486 * - limit is the number of hash buckets, not the total allocation size
4488 void *__init alloc_large_system_hash(const char *tablename,
4489 unsigned long bucketsize,
4490 unsigned long numentries,
4491 int scale,
4492 int flags,
4493 unsigned int *_hash_shift,
4494 unsigned int *_hash_mask,
4495 unsigned long limit)
4497 unsigned long long max = limit;
4498 unsigned long log2qty, size;
4499 void *table = NULL;
4501 /* allow the kernel cmdline to have a say */
4502 if (!numentries) {
4503 /* round applicable memory size up to nearest megabyte */
4504 numentries = nr_kernel_pages;
4505 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
4506 numentries >>= 20 - PAGE_SHIFT;
4507 numentries <<= 20 - PAGE_SHIFT;
4509 /* limit to 1 bucket per 2^scale bytes of low memory */
4510 if (scale > PAGE_SHIFT)
4511 numentries >>= (scale - PAGE_SHIFT);
4512 else
4513 numentries <<= (PAGE_SHIFT - scale);
4515 /* Make sure we've got at least a 0-order allocation.. */
4516 if (unlikely((numentries * bucketsize) < PAGE_SIZE))
4517 numentries = PAGE_SIZE / bucketsize;
4519 numentries = roundup_pow_of_two(numentries);
4521 /* limit allocation size to 1/16 total memory by default */
4522 if (max == 0) {
4523 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
4524 do_div(max, bucketsize);
4527 if (numentries > max)
4528 numentries = max;
4530 log2qty = ilog2(numentries);
4532 do {
4533 size = bucketsize << log2qty;
4534 if (flags & HASH_EARLY)
4535 table = alloc_bootmem_nopanic(size);
4536 else if (hashdist)
4537 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
4538 else {
4539 unsigned long order = get_order(size);
4540 table = (void*) __get_free_pages(GFP_ATOMIC, order);
4542 * If bucketsize is not a power-of-two, we may free
4543 * some pages at the end of hash table.
4545 if (table) {
4546 unsigned long alloc_end = (unsigned long)table +
4547 (PAGE_SIZE << order);
4548 unsigned long used = (unsigned long)table +
4549 PAGE_ALIGN(size);
4550 split_page(virt_to_page(table), order);
4551 while (used < alloc_end) {
4552 free_page(used);
4553 used += PAGE_SIZE;
4557 } while (!table && size > PAGE_SIZE && --log2qty);
4559 if (!table)
4560 panic("Failed to allocate %s hash table\n", tablename);
4562 printk(KERN_INFO "%s hash table entries: %d (order: %d, %lu bytes)\n",
4563 tablename,
4564 (1U << log2qty),
4565 ilog2(size) - PAGE_SHIFT,
4566 size);
4568 if (_hash_shift)
4569 *_hash_shift = log2qty;
4570 if (_hash_mask)
4571 *_hash_mask = (1 << log2qty) - 1;
4573 return table;
4576 #ifdef CONFIG_OUT_OF_LINE_PFN_TO_PAGE
4577 struct page *pfn_to_page(unsigned long pfn)
4579 return __pfn_to_page(pfn);
4581 unsigned long page_to_pfn(struct page *page)
4583 return __page_to_pfn(page);
4585 EXPORT_SYMBOL(pfn_to_page);
4586 EXPORT_SYMBOL(page_to_pfn);
4587 #endif /* CONFIG_OUT_OF_LINE_PFN_TO_PAGE */
4589 /* Return a pointer to the bitmap storing bits affecting a block of pages */
4590 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
4591 unsigned long pfn)
4593 #ifdef CONFIG_SPARSEMEM
4594 return __pfn_to_section(pfn)->pageblock_flags;
4595 #else
4596 return zone->pageblock_flags;
4597 #endif /* CONFIG_SPARSEMEM */
4600 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
4602 #ifdef CONFIG_SPARSEMEM
4603 pfn &= (PAGES_PER_SECTION-1);
4604 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
4605 #else
4606 pfn = pfn - zone->zone_start_pfn;
4607 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
4608 #endif /* CONFIG_SPARSEMEM */
4612 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
4613 * @page: The page within the block of interest
4614 * @start_bitidx: The first bit of interest to retrieve
4615 * @end_bitidx: The last bit of interest
4616 * returns pageblock_bits flags
4618 unsigned long get_pageblock_flags_group(struct page *page,
4619 int start_bitidx, int end_bitidx)
4621 struct zone *zone;
4622 unsigned long *bitmap;
4623 unsigned long pfn, bitidx;
4624 unsigned long flags = 0;
4625 unsigned long value = 1;
4627 zone = page_zone(page);
4628 pfn = page_to_pfn(page);
4629 bitmap = get_pageblock_bitmap(zone, pfn);
4630 bitidx = pfn_to_bitidx(zone, pfn);
4632 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
4633 if (test_bit(bitidx + start_bitidx, bitmap))
4634 flags |= value;
4636 return flags;
4640 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
4641 * @page: The page within the block of interest
4642 * @start_bitidx: The first bit of interest
4643 * @end_bitidx: The last bit of interest
4644 * @flags: The flags to set
4646 void set_pageblock_flags_group(struct page *page, unsigned long flags,
4647 int start_bitidx, int end_bitidx)
4649 struct zone *zone;
4650 unsigned long *bitmap;
4651 unsigned long pfn, bitidx;
4652 unsigned long value = 1;
4654 zone = page_zone(page);
4655 pfn = page_to_pfn(page);
4656 bitmap = get_pageblock_bitmap(zone, pfn);
4657 bitidx = pfn_to_bitidx(zone, pfn);
4658 VM_BUG_ON(pfn < zone->zone_start_pfn);
4659 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages);
4661 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
4662 if (flags & value)
4663 __set_bit(bitidx + start_bitidx, bitmap);
4664 else
4665 __clear_bit(bitidx + start_bitidx, bitmap);
4669 * This is designed as sub function...plz see page_isolation.c also.
4670 * set/clear page block's type to be ISOLATE.
4671 * page allocater never alloc memory from ISOLATE block.
4674 int set_migratetype_isolate(struct page *page)
4676 struct zone *zone;
4677 unsigned long flags;
4678 int ret = -EBUSY;
4680 zone = page_zone(page);
4681 spin_lock_irqsave(&zone->lock, flags);
4683 * In future, more migrate types will be able to be isolation target.
4685 if (get_pageblock_migratetype(page) != MIGRATE_MOVABLE)
4686 goto out;
4687 set_pageblock_migratetype(page, MIGRATE_ISOLATE);
4688 move_freepages_block(zone, page, MIGRATE_ISOLATE);
4689 ret = 0;
4690 out:
4691 spin_unlock_irqrestore(&zone->lock, flags);
4692 if (!ret)
4693 drain_all_pages();
4694 return ret;
4697 void unset_migratetype_isolate(struct page *page)
4699 struct zone *zone;
4700 unsigned long flags;
4701 zone = page_zone(page);
4702 spin_lock_irqsave(&zone->lock, flags);
4703 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
4704 goto out;
4705 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4706 move_freepages_block(zone, page, MIGRATE_MOVABLE);
4707 out:
4708 spin_unlock_irqrestore(&zone->lock, flags);
4711 #ifdef CONFIG_MEMORY_HOTREMOVE
4713 * All pages in the range must be isolated before calling this.
4715 void
4716 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
4718 struct page *page;
4719 struct zone *zone;
4720 int order, i;
4721 unsigned long pfn;
4722 unsigned long flags;
4723 /* find the first valid pfn */
4724 for (pfn = start_pfn; pfn < end_pfn; pfn++)
4725 if (pfn_valid(pfn))
4726 break;
4727 if (pfn == end_pfn)
4728 return;
4729 zone = page_zone(pfn_to_page(pfn));
4730 spin_lock_irqsave(&zone->lock, flags);
4731 pfn = start_pfn;
4732 while (pfn < end_pfn) {
4733 if (!pfn_valid(pfn)) {
4734 pfn++;
4735 continue;
4737 page = pfn_to_page(pfn);
4738 BUG_ON(page_count(page));
4739 BUG_ON(!PageBuddy(page));
4740 order = page_order(page);
4741 #ifdef CONFIG_DEBUG_VM
4742 printk(KERN_INFO "remove from free list %lx %d %lx\n",
4743 pfn, 1 << order, end_pfn);
4744 #endif
4745 list_del(&page->lru);
4746 rmv_page_order(page);
4747 zone->free_area[order].nr_free--;
4748 __mod_zone_page_state(zone, NR_FREE_PAGES,
4749 - (1UL << order));
4750 for (i = 0; i < (1 << order); i++)
4751 SetPageReserved((page+i));
4752 pfn += (1 << order);
4754 spin_unlock_irqrestore(&zone->lock, flags);
4756 #endif