mempolicy: use MPOL_PREFERRED for system-wide default policy
[linux-2.6/mini2440.git] / mm / page_alloc.c
blobe0fc3baba843f2976a0248d22202431e932d1085
1 /*
2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
18 #include <linux/mm.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/compiler.h>
25 #include <linux/kernel.h>
26 #include <linux/module.h>
27 #include <linux/suspend.h>
28 #include <linux/pagevec.h>
29 #include <linux/blkdev.h>
30 #include <linux/slab.h>
31 #include <linux/oom.h>
32 #include <linux/notifier.h>
33 #include <linux/topology.h>
34 #include <linux/sysctl.h>
35 #include <linux/cpu.h>
36 #include <linux/cpuset.h>
37 #include <linux/memory_hotplug.h>
38 #include <linux/nodemask.h>
39 #include <linux/vmalloc.h>
40 #include <linux/mempolicy.h>
41 #include <linux/stop_machine.h>
42 #include <linux/sort.h>
43 #include <linux/pfn.h>
44 #include <linux/backing-dev.h>
45 #include <linux/fault-inject.h>
46 #include <linux/page-isolation.h>
47 #include <linux/memcontrol.h>
49 #include <asm/tlbflush.h>
50 #include <asm/div64.h>
51 #include "internal.h"
54 * Array of node states.
56 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
57 [N_POSSIBLE] = NODE_MASK_ALL,
58 [N_ONLINE] = { { [0] = 1UL } },
59 #ifndef CONFIG_NUMA
60 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
61 #ifdef CONFIG_HIGHMEM
62 [N_HIGH_MEMORY] = { { [0] = 1UL } },
63 #endif
64 [N_CPU] = { { [0] = 1UL } },
65 #endif /* NUMA */
67 EXPORT_SYMBOL(node_states);
69 unsigned long totalram_pages __read_mostly;
70 unsigned long totalreserve_pages __read_mostly;
71 long nr_swap_pages;
72 int percpu_pagelist_fraction;
74 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
75 int pageblock_order __read_mostly;
76 #endif
78 static void __free_pages_ok(struct page *page, unsigned int order);
81 * results with 256, 32 in the lowmem_reserve sysctl:
82 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
83 * 1G machine -> (16M dma, 784M normal, 224M high)
84 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
85 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
86 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
88 * TBD: should special case ZONE_DMA32 machines here - in those we normally
89 * don't need any ZONE_NORMAL reservation
91 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
92 #ifdef CONFIG_ZONE_DMA
93 256,
94 #endif
95 #ifdef CONFIG_ZONE_DMA32
96 256,
97 #endif
98 #ifdef CONFIG_HIGHMEM
99 32,
100 #endif
104 EXPORT_SYMBOL(totalram_pages);
106 static char * const zone_names[MAX_NR_ZONES] = {
107 #ifdef CONFIG_ZONE_DMA
108 "DMA",
109 #endif
110 #ifdef CONFIG_ZONE_DMA32
111 "DMA32",
112 #endif
113 "Normal",
114 #ifdef CONFIG_HIGHMEM
115 "HighMem",
116 #endif
117 "Movable",
120 int min_free_kbytes = 1024;
122 unsigned long __meminitdata nr_kernel_pages;
123 unsigned long __meminitdata nr_all_pages;
124 static unsigned long __meminitdata dma_reserve;
126 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
128 * MAX_ACTIVE_REGIONS determines the maximum number of distinct
129 * ranges of memory (RAM) that may be registered with add_active_range().
130 * Ranges passed to add_active_range() will be merged if possible
131 * so the number of times add_active_range() can be called is
132 * related to the number of nodes and the number of holes
134 #ifdef CONFIG_MAX_ACTIVE_REGIONS
135 /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */
136 #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS
137 #else
138 #if MAX_NUMNODES >= 32
139 /* If there can be many nodes, allow up to 50 holes per node */
140 #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50)
141 #else
142 /* By default, allow up to 256 distinct regions */
143 #define MAX_ACTIVE_REGIONS 256
144 #endif
145 #endif
147 static struct node_active_region __meminitdata early_node_map[MAX_ACTIVE_REGIONS];
148 static int __meminitdata nr_nodemap_entries;
149 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
150 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
151 #ifdef CONFIG_MEMORY_HOTPLUG_RESERVE
152 static unsigned long __meminitdata node_boundary_start_pfn[MAX_NUMNODES];
153 static unsigned long __meminitdata node_boundary_end_pfn[MAX_NUMNODES];
154 #endif /* CONFIG_MEMORY_HOTPLUG_RESERVE */
155 unsigned long __initdata required_kernelcore;
156 static unsigned long __initdata required_movablecore;
157 unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
159 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
160 int movable_zone;
161 EXPORT_SYMBOL(movable_zone);
162 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
164 #if MAX_NUMNODES > 1
165 int nr_node_ids __read_mostly = MAX_NUMNODES;
166 EXPORT_SYMBOL(nr_node_ids);
167 #endif
169 int page_group_by_mobility_disabled __read_mostly;
171 static void set_pageblock_migratetype(struct page *page, int migratetype)
173 set_pageblock_flags_group(page, (unsigned long)migratetype,
174 PB_migrate, PB_migrate_end);
177 #ifdef CONFIG_DEBUG_VM
178 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
180 int ret = 0;
181 unsigned seq;
182 unsigned long pfn = page_to_pfn(page);
184 do {
185 seq = zone_span_seqbegin(zone);
186 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
187 ret = 1;
188 else if (pfn < zone->zone_start_pfn)
189 ret = 1;
190 } while (zone_span_seqretry(zone, seq));
192 return ret;
195 static int page_is_consistent(struct zone *zone, struct page *page)
197 if (!pfn_valid_within(page_to_pfn(page)))
198 return 0;
199 if (zone != page_zone(page))
200 return 0;
202 return 1;
205 * Temporary debugging check for pages not lying within a given zone.
207 static int bad_range(struct zone *zone, struct page *page)
209 if (page_outside_zone_boundaries(zone, page))
210 return 1;
211 if (!page_is_consistent(zone, page))
212 return 1;
214 return 0;
216 #else
217 static inline int bad_range(struct zone *zone, struct page *page)
219 return 0;
221 #endif
223 static void bad_page(struct page *page)
225 void *pc = page_get_page_cgroup(page);
227 printk(KERN_EMERG "Bad page state in process '%s'\n" KERN_EMERG
228 "page:%p flags:0x%0*lx mapping:%p mapcount:%d count:%d\n",
229 current->comm, page, (int)(2*sizeof(unsigned long)),
230 (unsigned long)page->flags, page->mapping,
231 page_mapcount(page), page_count(page));
232 if (pc) {
233 printk(KERN_EMERG "cgroup:%p\n", pc);
234 page_reset_bad_cgroup(page);
236 printk(KERN_EMERG "Trying to fix it up, but a reboot is needed\n"
237 KERN_EMERG "Backtrace:\n");
238 dump_stack();
239 page->flags &= ~(1 << PG_lru |
240 1 << PG_private |
241 1 << PG_locked |
242 1 << PG_active |
243 1 << PG_dirty |
244 1 << PG_reclaim |
245 1 << PG_slab |
246 1 << PG_swapcache |
247 1 << PG_writeback |
248 1 << PG_buddy );
249 set_page_count(page, 0);
250 reset_page_mapcount(page);
251 page->mapping = NULL;
252 add_taint(TAINT_BAD_PAGE);
256 * Higher-order pages are called "compound pages". They are structured thusly:
258 * The first PAGE_SIZE page is called the "head page".
260 * The remaining PAGE_SIZE pages are called "tail pages".
262 * All pages have PG_compound set. All pages have their ->private pointing at
263 * the head page (even the head page has this).
265 * The first tail page's ->lru.next holds the address of the compound page's
266 * put_page() function. Its ->lru.prev holds the order of allocation.
267 * This usage means that zero-order pages may not be compound.
270 static void free_compound_page(struct page *page)
272 __free_pages_ok(page, compound_order(page));
275 static void prep_compound_page(struct page *page, unsigned long order)
277 int i;
278 int nr_pages = 1 << order;
280 set_compound_page_dtor(page, free_compound_page);
281 set_compound_order(page, order);
282 __SetPageHead(page);
283 for (i = 1; i < nr_pages; i++) {
284 struct page *p = page + i;
286 __SetPageTail(p);
287 p->first_page = page;
291 static void destroy_compound_page(struct page *page, unsigned long order)
293 int i;
294 int nr_pages = 1 << order;
296 if (unlikely(compound_order(page) != order))
297 bad_page(page);
299 if (unlikely(!PageHead(page)))
300 bad_page(page);
301 __ClearPageHead(page);
302 for (i = 1; i < nr_pages; i++) {
303 struct page *p = page + i;
305 if (unlikely(!PageTail(p) |
306 (p->first_page != page)))
307 bad_page(page);
308 __ClearPageTail(p);
312 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
314 int i;
317 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
318 * and __GFP_HIGHMEM from hard or soft interrupt context.
320 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
321 for (i = 0; i < (1 << order); i++)
322 clear_highpage(page + i);
325 static inline void set_page_order(struct page *page, int order)
327 set_page_private(page, order);
328 __SetPageBuddy(page);
331 static inline void rmv_page_order(struct page *page)
333 __ClearPageBuddy(page);
334 set_page_private(page, 0);
338 * Locate the struct page for both the matching buddy in our
339 * pair (buddy1) and the combined O(n+1) page they form (page).
341 * 1) Any buddy B1 will have an order O twin B2 which satisfies
342 * the following equation:
343 * B2 = B1 ^ (1 << O)
344 * For example, if the starting buddy (buddy2) is #8 its order
345 * 1 buddy is #10:
346 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
348 * 2) Any buddy B will have an order O+1 parent P which
349 * satisfies the following equation:
350 * P = B & ~(1 << O)
352 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
354 static inline struct page *
355 __page_find_buddy(struct page *page, unsigned long page_idx, unsigned int order)
357 unsigned long buddy_idx = page_idx ^ (1 << order);
359 return page + (buddy_idx - page_idx);
362 static inline unsigned long
363 __find_combined_index(unsigned long page_idx, unsigned int order)
365 return (page_idx & ~(1 << order));
369 * This function checks whether a page is free && is the buddy
370 * we can do coalesce a page and its buddy if
371 * (a) the buddy is not in a hole &&
372 * (b) the buddy is in the buddy system &&
373 * (c) a page and its buddy have the same order &&
374 * (d) a page and its buddy are in the same zone.
376 * For recording whether a page is in the buddy system, we use PG_buddy.
377 * Setting, clearing, and testing PG_buddy is serialized by zone->lock.
379 * For recording page's order, we use page_private(page).
381 static inline int page_is_buddy(struct page *page, struct page *buddy,
382 int order)
384 if (!pfn_valid_within(page_to_pfn(buddy)))
385 return 0;
387 if (page_zone_id(page) != page_zone_id(buddy))
388 return 0;
390 if (PageBuddy(buddy) && page_order(buddy) == order) {
391 BUG_ON(page_count(buddy) != 0);
392 return 1;
394 return 0;
398 * Freeing function for a buddy system allocator.
400 * The concept of a buddy system is to maintain direct-mapped table
401 * (containing bit values) for memory blocks of various "orders".
402 * The bottom level table contains the map for the smallest allocatable
403 * units of memory (here, pages), and each level above it describes
404 * pairs of units from the levels below, hence, "buddies".
405 * At a high level, all that happens here is marking the table entry
406 * at the bottom level available, and propagating the changes upward
407 * as necessary, plus some accounting needed to play nicely with other
408 * parts of the VM system.
409 * At each level, we keep a list of pages, which are heads of continuous
410 * free pages of length of (1 << order) and marked with PG_buddy. Page's
411 * order is recorded in page_private(page) field.
412 * So when we are allocating or freeing one, we can derive the state of the
413 * other. That is, if we allocate a small block, and both were
414 * free, the remainder of the region must be split into blocks.
415 * If a block is freed, and its buddy is also free, then this
416 * triggers coalescing into a block of larger size.
418 * -- wli
421 static inline void __free_one_page(struct page *page,
422 struct zone *zone, unsigned int order)
424 unsigned long page_idx;
425 int order_size = 1 << order;
426 int migratetype = get_pageblock_migratetype(page);
428 if (unlikely(PageCompound(page)))
429 destroy_compound_page(page, order);
431 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
433 VM_BUG_ON(page_idx & (order_size - 1));
434 VM_BUG_ON(bad_range(zone, page));
436 __mod_zone_page_state(zone, NR_FREE_PAGES, order_size);
437 while (order < MAX_ORDER-1) {
438 unsigned long combined_idx;
439 struct page *buddy;
441 buddy = __page_find_buddy(page, page_idx, order);
442 if (!page_is_buddy(page, buddy, order))
443 break; /* Move the buddy up one level. */
445 list_del(&buddy->lru);
446 zone->free_area[order].nr_free--;
447 rmv_page_order(buddy);
448 combined_idx = __find_combined_index(page_idx, order);
449 page = page + (combined_idx - page_idx);
450 page_idx = combined_idx;
451 order++;
453 set_page_order(page, order);
454 list_add(&page->lru,
455 &zone->free_area[order].free_list[migratetype]);
456 zone->free_area[order].nr_free++;
459 static inline int free_pages_check(struct page *page)
461 if (unlikely(page_mapcount(page) |
462 (page->mapping != NULL) |
463 (page_get_page_cgroup(page) != NULL) |
464 (page_count(page) != 0) |
465 (page->flags & (
466 1 << PG_lru |
467 1 << PG_private |
468 1 << PG_locked |
469 1 << PG_active |
470 1 << PG_slab |
471 1 << PG_swapcache |
472 1 << PG_writeback |
473 1 << PG_reserved |
474 1 << PG_buddy ))))
475 bad_page(page);
476 if (PageDirty(page))
477 __ClearPageDirty(page);
479 * For now, we report if PG_reserved was found set, but do not
480 * clear it, and do not free the page. But we shall soon need
481 * to do more, for when the ZERO_PAGE count wraps negative.
483 return PageReserved(page);
487 * Frees a list of pages.
488 * Assumes all pages on list are in same zone, and of same order.
489 * count is the number of pages to free.
491 * If the zone was previously in an "all pages pinned" state then look to
492 * see if this freeing clears that state.
494 * And clear the zone's pages_scanned counter, to hold off the "all pages are
495 * pinned" detection logic.
497 static void free_pages_bulk(struct zone *zone, int count,
498 struct list_head *list, int order)
500 spin_lock(&zone->lock);
501 zone_clear_flag(zone, ZONE_ALL_UNRECLAIMABLE);
502 zone->pages_scanned = 0;
503 while (count--) {
504 struct page *page;
506 VM_BUG_ON(list_empty(list));
507 page = list_entry(list->prev, struct page, lru);
508 /* have to delete it as __free_one_page list manipulates */
509 list_del(&page->lru);
510 __free_one_page(page, zone, order);
512 spin_unlock(&zone->lock);
515 static void free_one_page(struct zone *zone, struct page *page, int order)
517 spin_lock(&zone->lock);
518 zone_clear_flag(zone, ZONE_ALL_UNRECLAIMABLE);
519 zone->pages_scanned = 0;
520 __free_one_page(page, zone, order);
521 spin_unlock(&zone->lock);
524 static void __free_pages_ok(struct page *page, unsigned int order)
526 unsigned long flags;
527 int i;
528 int reserved = 0;
530 for (i = 0 ; i < (1 << order) ; ++i)
531 reserved += free_pages_check(page + i);
532 if (reserved)
533 return;
535 if (!PageHighMem(page))
536 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
537 arch_free_page(page, order);
538 kernel_map_pages(page, 1 << order, 0);
540 local_irq_save(flags);
541 __count_vm_events(PGFREE, 1 << order);
542 free_one_page(page_zone(page), page, order);
543 local_irq_restore(flags);
547 * permit the bootmem allocator to evade page validation on high-order frees
549 void __init __free_pages_bootmem(struct page *page, unsigned int order)
551 if (order == 0) {
552 __ClearPageReserved(page);
553 set_page_count(page, 0);
554 set_page_refcounted(page);
555 __free_page(page);
556 } else {
557 int loop;
559 prefetchw(page);
560 for (loop = 0; loop < BITS_PER_LONG; loop++) {
561 struct page *p = &page[loop];
563 if (loop + 1 < BITS_PER_LONG)
564 prefetchw(p + 1);
565 __ClearPageReserved(p);
566 set_page_count(p, 0);
569 set_page_refcounted(page);
570 __free_pages(page, order);
576 * The order of subdivision here is critical for the IO subsystem.
577 * Please do not alter this order without good reasons and regression
578 * testing. Specifically, as large blocks of memory are subdivided,
579 * the order in which smaller blocks are delivered depends on the order
580 * they're subdivided in this function. This is the primary factor
581 * influencing the order in which pages are delivered to the IO
582 * subsystem according to empirical testing, and this is also justified
583 * by considering the behavior of a buddy system containing a single
584 * large block of memory acted on by a series of small allocations.
585 * This behavior is a critical factor in sglist merging's success.
587 * -- wli
589 static inline void expand(struct zone *zone, struct page *page,
590 int low, int high, struct free_area *area,
591 int migratetype)
593 unsigned long size = 1 << high;
595 while (high > low) {
596 area--;
597 high--;
598 size >>= 1;
599 VM_BUG_ON(bad_range(zone, &page[size]));
600 list_add(&page[size].lru, &area->free_list[migratetype]);
601 area->nr_free++;
602 set_page_order(&page[size], high);
607 * This page is about to be returned from the page allocator
609 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
611 if (unlikely(page_mapcount(page) |
612 (page->mapping != NULL) |
613 (page_get_page_cgroup(page) != NULL) |
614 (page_count(page) != 0) |
615 (page->flags & (
616 1 << PG_lru |
617 1 << PG_private |
618 1 << PG_locked |
619 1 << PG_active |
620 1 << PG_dirty |
621 1 << PG_slab |
622 1 << PG_swapcache |
623 1 << PG_writeback |
624 1 << PG_reserved |
625 1 << PG_buddy ))))
626 bad_page(page);
629 * For now, we report if PG_reserved was found set, but do not
630 * clear it, and do not allocate the page: as a safety net.
632 if (PageReserved(page))
633 return 1;
635 page->flags &= ~(1 << PG_uptodate | 1 << PG_error | 1 << PG_reclaim |
636 1 << PG_referenced | 1 << PG_arch_1 |
637 1 << PG_owner_priv_1 | 1 << PG_mappedtodisk);
638 set_page_private(page, 0);
639 set_page_refcounted(page);
641 arch_alloc_page(page, order);
642 kernel_map_pages(page, 1 << order, 1);
644 if (gfp_flags & __GFP_ZERO)
645 prep_zero_page(page, order, gfp_flags);
647 if (order && (gfp_flags & __GFP_COMP))
648 prep_compound_page(page, order);
650 return 0;
654 * Go through the free lists for the given migratetype and remove
655 * the smallest available page from the freelists
657 static struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
658 int migratetype)
660 unsigned int current_order;
661 struct free_area * area;
662 struct page *page;
664 /* Find a page of the appropriate size in the preferred list */
665 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
666 area = &(zone->free_area[current_order]);
667 if (list_empty(&area->free_list[migratetype]))
668 continue;
670 page = list_entry(area->free_list[migratetype].next,
671 struct page, lru);
672 list_del(&page->lru);
673 rmv_page_order(page);
674 area->nr_free--;
675 __mod_zone_page_state(zone, NR_FREE_PAGES, - (1UL << order));
676 expand(zone, page, order, current_order, area, migratetype);
677 return page;
680 return NULL;
685 * This array describes the order lists are fallen back to when
686 * the free lists for the desirable migrate type are depleted
688 static int fallbacks[MIGRATE_TYPES][MIGRATE_TYPES-1] = {
689 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
690 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
691 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
692 [MIGRATE_RESERVE] = { MIGRATE_RESERVE, MIGRATE_RESERVE, MIGRATE_RESERVE }, /* Never used */
696 * Move the free pages in a range to the free lists of the requested type.
697 * Note that start_page and end_pages are not aligned on a pageblock
698 * boundary. If alignment is required, use move_freepages_block()
700 int move_freepages(struct zone *zone,
701 struct page *start_page, struct page *end_page,
702 int migratetype)
704 struct page *page;
705 unsigned long order;
706 int pages_moved = 0;
708 #ifndef CONFIG_HOLES_IN_ZONE
710 * page_zone is not safe to call in this context when
711 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
712 * anyway as we check zone boundaries in move_freepages_block().
713 * Remove at a later date when no bug reports exist related to
714 * grouping pages by mobility
716 BUG_ON(page_zone(start_page) != page_zone(end_page));
717 #endif
719 for (page = start_page; page <= end_page;) {
720 if (!pfn_valid_within(page_to_pfn(page))) {
721 page++;
722 continue;
725 if (!PageBuddy(page)) {
726 page++;
727 continue;
730 order = page_order(page);
731 list_del(&page->lru);
732 list_add(&page->lru,
733 &zone->free_area[order].free_list[migratetype]);
734 page += 1 << order;
735 pages_moved += 1 << order;
738 return pages_moved;
741 int move_freepages_block(struct zone *zone, struct page *page, int migratetype)
743 unsigned long start_pfn, end_pfn;
744 struct page *start_page, *end_page;
746 start_pfn = page_to_pfn(page);
747 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
748 start_page = pfn_to_page(start_pfn);
749 end_page = start_page + pageblock_nr_pages - 1;
750 end_pfn = start_pfn + pageblock_nr_pages - 1;
752 /* Do not cross zone boundaries */
753 if (start_pfn < zone->zone_start_pfn)
754 start_page = page;
755 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
756 return 0;
758 return move_freepages(zone, start_page, end_page, migratetype);
761 /* Remove an element from the buddy allocator from the fallback list */
762 static struct page *__rmqueue_fallback(struct zone *zone, int order,
763 int start_migratetype)
765 struct free_area * area;
766 int current_order;
767 struct page *page;
768 int migratetype, i;
770 /* Find the largest possible block of pages in the other list */
771 for (current_order = MAX_ORDER-1; current_order >= order;
772 --current_order) {
773 for (i = 0; i < MIGRATE_TYPES - 1; i++) {
774 migratetype = fallbacks[start_migratetype][i];
776 /* MIGRATE_RESERVE handled later if necessary */
777 if (migratetype == MIGRATE_RESERVE)
778 continue;
780 area = &(zone->free_area[current_order]);
781 if (list_empty(&area->free_list[migratetype]))
782 continue;
784 page = list_entry(area->free_list[migratetype].next,
785 struct page, lru);
786 area->nr_free--;
789 * If breaking a large block of pages, move all free
790 * pages to the preferred allocation list. If falling
791 * back for a reclaimable kernel allocation, be more
792 * agressive about taking ownership of free pages
794 if (unlikely(current_order >= (pageblock_order >> 1)) ||
795 start_migratetype == MIGRATE_RECLAIMABLE) {
796 unsigned long pages;
797 pages = move_freepages_block(zone, page,
798 start_migratetype);
800 /* Claim the whole block if over half of it is free */
801 if (pages >= (1 << (pageblock_order-1)))
802 set_pageblock_migratetype(page,
803 start_migratetype);
805 migratetype = start_migratetype;
808 /* Remove the page from the freelists */
809 list_del(&page->lru);
810 rmv_page_order(page);
811 __mod_zone_page_state(zone, NR_FREE_PAGES,
812 -(1UL << order));
814 if (current_order == pageblock_order)
815 set_pageblock_migratetype(page,
816 start_migratetype);
818 expand(zone, page, order, current_order, area, migratetype);
819 return page;
823 /* Use MIGRATE_RESERVE rather than fail an allocation */
824 return __rmqueue_smallest(zone, order, MIGRATE_RESERVE);
828 * Do the hard work of removing an element from the buddy allocator.
829 * Call me with the zone->lock already held.
831 static struct page *__rmqueue(struct zone *zone, unsigned int order,
832 int migratetype)
834 struct page *page;
836 page = __rmqueue_smallest(zone, order, migratetype);
838 if (unlikely(!page))
839 page = __rmqueue_fallback(zone, order, migratetype);
841 return page;
845 * Obtain a specified number of elements from the buddy allocator, all under
846 * a single hold of the lock, for efficiency. Add them to the supplied list.
847 * Returns the number of new pages which were placed at *list.
849 static int rmqueue_bulk(struct zone *zone, unsigned int order,
850 unsigned long count, struct list_head *list,
851 int migratetype)
853 int i;
855 spin_lock(&zone->lock);
856 for (i = 0; i < count; ++i) {
857 struct page *page = __rmqueue(zone, order, migratetype);
858 if (unlikely(page == NULL))
859 break;
862 * Split buddy pages returned by expand() are received here
863 * in physical page order. The page is added to the callers and
864 * list and the list head then moves forward. From the callers
865 * perspective, the linked list is ordered by page number in
866 * some conditions. This is useful for IO devices that can
867 * merge IO requests if the physical pages are ordered
868 * properly.
870 list_add(&page->lru, list);
871 set_page_private(page, migratetype);
872 list = &page->lru;
874 spin_unlock(&zone->lock);
875 return i;
878 #ifdef CONFIG_NUMA
880 * Called from the vmstat counter updater to drain pagesets of this
881 * currently executing processor on remote nodes after they have
882 * expired.
884 * Note that this function must be called with the thread pinned to
885 * a single processor.
887 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
889 unsigned long flags;
890 int to_drain;
892 local_irq_save(flags);
893 if (pcp->count >= pcp->batch)
894 to_drain = pcp->batch;
895 else
896 to_drain = pcp->count;
897 free_pages_bulk(zone, to_drain, &pcp->list, 0);
898 pcp->count -= to_drain;
899 local_irq_restore(flags);
901 #endif
904 * Drain pages of the indicated processor.
906 * The processor must either be the current processor and the
907 * thread pinned to the current processor or a processor that
908 * is not online.
910 static void drain_pages(unsigned int cpu)
912 unsigned long flags;
913 struct zone *zone;
915 for_each_zone(zone) {
916 struct per_cpu_pageset *pset;
917 struct per_cpu_pages *pcp;
919 if (!populated_zone(zone))
920 continue;
922 pset = zone_pcp(zone, cpu);
924 pcp = &pset->pcp;
925 local_irq_save(flags);
926 free_pages_bulk(zone, pcp->count, &pcp->list, 0);
927 pcp->count = 0;
928 local_irq_restore(flags);
933 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
935 void drain_local_pages(void *arg)
937 drain_pages(smp_processor_id());
941 * Spill all the per-cpu pages from all CPUs back into the buddy allocator
943 void drain_all_pages(void)
945 on_each_cpu(drain_local_pages, NULL, 0, 1);
948 #ifdef CONFIG_HIBERNATION
950 void mark_free_pages(struct zone *zone)
952 unsigned long pfn, max_zone_pfn;
953 unsigned long flags;
954 int order, t;
955 struct list_head *curr;
957 if (!zone->spanned_pages)
958 return;
960 spin_lock_irqsave(&zone->lock, flags);
962 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
963 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
964 if (pfn_valid(pfn)) {
965 struct page *page = pfn_to_page(pfn);
967 if (!swsusp_page_is_forbidden(page))
968 swsusp_unset_page_free(page);
971 for_each_migratetype_order(order, t) {
972 list_for_each(curr, &zone->free_area[order].free_list[t]) {
973 unsigned long i;
975 pfn = page_to_pfn(list_entry(curr, struct page, lru));
976 for (i = 0; i < (1UL << order); i++)
977 swsusp_set_page_free(pfn_to_page(pfn + i));
980 spin_unlock_irqrestore(&zone->lock, flags);
982 #endif /* CONFIG_PM */
985 * Free a 0-order page
987 static void free_hot_cold_page(struct page *page, int cold)
989 struct zone *zone = page_zone(page);
990 struct per_cpu_pages *pcp;
991 unsigned long flags;
993 if (PageAnon(page))
994 page->mapping = NULL;
995 if (free_pages_check(page))
996 return;
998 if (!PageHighMem(page))
999 debug_check_no_locks_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 classzone_idx = zone_idx(preferred_zone);
1395 zonelist_scan:
1397 * Scan zonelist, looking for a zone with enough free.
1398 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1400 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1401 high_zoneidx, nodemask) {
1402 if (NUMA_BUILD && zlc_active &&
1403 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1404 continue;
1405 if ((alloc_flags & ALLOC_CPUSET) &&
1406 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1407 goto try_next_zone;
1409 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1410 unsigned long mark;
1411 if (alloc_flags & ALLOC_WMARK_MIN)
1412 mark = zone->pages_min;
1413 else if (alloc_flags & ALLOC_WMARK_LOW)
1414 mark = zone->pages_low;
1415 else
1416 mark = zone->pages_high;
1417 if (!zone_watermark_ok(zone, order, mark,
1418 classzone_idx, alloc_flags)) {
1419 if (!zone_reclaim_mode ||
1420 !zone_reclaim(zone, gfp_mask, order))
1421 goto this_zone_full;
1425 page = buffered_rmqueue(preferred_zone, zone, order, gfp_mask);
1426 if (page)
1427 break;
1428 this_zone_full:
1429 if (NUMA_BUILD)
1430 zlc_mark_zone_full(zonelist, z);
1431 try_next_zone:
1432 if (NUMA_BUILD && !did_zlc_setup) {
1433 /* we do zlc_setup after the first zone is tried */
1434 allowednodes = zlc_setup(zonelist, alloc_flags);
1435 zlc_active = 1;
1436 did_zlc_setup = 1;
1440 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) {
1441 /* Disable zlc cache for second zonelist scan */
1442 zlc_active = 0;
1443 goto zonelist_scan;
1445 return page;
1449 * This is the 'heart' of the zoned buddy allocator.
1451 static struct page *
1452 __alloc_pages_internal(gfp_t gfp_mask, unsigned int order,
1453 struct zonelist *zonelist, nodemask_t *nodemask)
1455 const gfp_t wait = gfp_mask & __GFP_WAIT;
1456 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
1457 struct zoneref *z;
1458 struct zone *zone;
1459 struct page *page;
1460 struct reclaim_state reclaim_state;
1461 struct task_struct *p = current;
1462 int do_retry;
1463 int alloc_flags;
1464 int did_some_progress;
1466 might_sleep_if(wait);
1468 if (should_fail_alloc_page(gfp_mask, order))
1469 return NULL;
1471 restart:
1472 z = zonelist->_zonerefs; /* the list of zones suitable for gfp_mask */
1474 if (unlikely(!z->zone)) {
1476 * Happens if we have an empty zonelist as a result of
1477 * GFP_THISNODE being used on a memoryless node
1479 return NULL;
1482 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
1483 zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET);
1484 if (page)
1485 goto got_pg;
1488 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
1489 * __GFP_NOWARN set) should not cause reclaim since the subsystem
1490 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
1491 * using a larger set of nodes after it has established that the
1492 * allowed per node queues are empty and that nodes are
1493 * over allocated.
1495 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
1496 goto nopage;
1498 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
1499 wakeup_kswapd(zone, order);
1502 * OK, we're below the kswapd watermark and have kicked background
1503 * reclaim. Now things get more complex, so set up alloc_flags according
1504 * to how we want to proceed.
1506 * The caller may dip into page reserves a bit more if the caller
1507 * cannot run direct reclaim, or if the caller has realtime scheduling
1508 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
1509 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
1511 alloc_flags = ALLOC_WMARK_MIN;
1512 if ((unlikely(rt_task(p)) && !in_interrupt()) || !wait)
1513 alloc_flags |= ALLOC_HARDER;
1514 if (gfp_mask & __GFP_HIGH)
1515 alloc_flags |= ALLOC_HIGH;
1516 if (wait)
1517 alloc_flags |= ALLOC_CPUSET;
1520 * Go through the zonelist again. Let __GFP_HIGH and allocations
1521 * coming from realtime tasks go deeper into reserves.
1523 * This is the last chance, in general, before the goto nopage.
1524 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
1525 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1527 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
1528 high_zoneidx, alloc_flags);
1529 if (page)
1530 goto got_pg;
1532 /* This allocation should allow future memory freeing. */
1534 rebalance:
1535 if (((p->flags & PF_MEMALLOC) || unlikely(test_thread_flag(TIF_MEMDIE)))
1536 && !in_interrupt()) {
1537 if (!(gfp_mask & __GFP_NOMEMALLOC)) {
1538 nofail_alloc:
1539 /* go through the zonelist yet again, ignoring mins */
1540 page = get_page_from_freelist(gfp_mask, nodemask, order,
1541 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS);
1542 if (page)
1543 goto got_pg;
1544 if (gfp_mask & __GFP_NOFAIL) {
1545 congestion_wait(WRITE, HZ/50);
1546 goto nofail_alloc;
1549 goto nopage;
1552 /* Atomic allocations - we can't balance anything */
1553 if (!wait)
1554 goto nopage;
1556 cond_resched();
1558 /* We now go into synchronous reclaim */
1559 cpuset_memory_pressure_bump();
1560 p->flags |= PF_MEMALLOC;
1561 reclaim_state.reclaimed_slab = 0;
1562 p->reclaim_state = &reclaim_state;
1564 did_some_progress = try_to_free_pages(zonelist, order, gfp_mask);
1566 p->reclaim_state = NULL;
1567 p->flags &= ~PF_MEMALLOC;
1569 cond_resched();
1571 if (order != 0)
1572 drain_all_pages();
1574 if (likely(did_some_progress)) {
1575 page = get_page_from_freelist(gfp_mask, nodemask, order,
1576 zonelist, high_zoneidx, alloc_flags);
1577 if (page)
1578 goto got_pg;
1579 } else if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
1580 if (!try_set_zone_oom(zonelist, gfp_mask)) {
1581 schedule_timeout_uninterruptible(1);
1582 goto restart;
1586 * Go through the zonelist yet one more time, keep
1587 * very high watermark here, this is only to catch
1588 * a parallel oom killing, we must fail if we're still
1589 * under heavy pressure.
1591 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
1592 order, zonelist, high_zoneidx,
1593 ALLOC_WMARK_HIGH|ALLOC_CPUSET);
1594 if (page) {
1595 clear_zonelist_oom(zonelist, gfp_mask);
1596 goto got_pg;
1599 /* The OOM killer will not help higher order allocs so fail */
1600 if (order > PAGE_ALLOC_COSTLY_ORDER) {
1601 clear_zonelist_oom(zonelist, gfp_mask);
1602 goto nopage;
1605 out_of_memory(zonelist, gfp_mask, order);
1606 clear_zonelist_oom(zonelist, gfp_mask);
1607 goto restart;
1611 * Don't let big-order allocations loop unless the caller explicitly
1612 * requests that. Wait for some write requests to complete then retry.
1614 * In this implementation, __GFP_REPEAT means __GFP_NOFAIL for order
1615 * <= 3, but that may not be true in other implementations.
1617 do_retry = 0;
1618 if (!(gfp_mask & __GFP_NORETRY)) {
1619 if ((order <= PAGE_ALLOC_COSTLY_ORDER) ||
1620 (gfp_mask & __GFP_REPEAT))
1621 do_retry = 1;
1622 if (gfp_mask & __GFP_NOFAIL)
1623 do_retry = 1;
1625 if (do_retry) {
1626 congestion_wait(WRITE, HZ/50);
1627 goto rebalance;
1630 nopage:
1631 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
1632 printk(KERN_WARNING "%s: page allocation failure."
1633 " order:%d, mode:0x%x\n",
1634 p->comm, order, gfp_mask);
1635 dump_stack();
1636 show_mem();
1638 got_pg:
1639 return page;
1642 struct page *
1643 __alloc_pages(gfp_t gfp_mask, unsigned int order,
1644 struct zonelist *zonelist)
1646 return __alloc_pages_internal(gfp_mask, order, zonelist, NULL);
1649 struct page *
1650 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
1651 struct zonelist *zonelist, nodemask_t *nodemask)
1653 return __alloc_pages_internal(gfp_mask, order, zonelist, nodemask);
1656 EXPORT_SYMBOL(__alloc_pages);
1659 * Common helper functions.
1661 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
1663 struct page * page;
1664 page = alloc_pages(gfp_mask, order);
1665 if (!page)
1666 return 0;
1667 return (unsigned long) page_address(page);
1670 EXPORT_SYMBOL(__get_free_pages);
1672 unsigned long get_zeroed_page(gfp_t gfp_mask)
1674 struct page * page;
1677 * get_zeroed_page() returns a 32-bit address, which cannot represent
1678 * a highmem page
1680 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
1682 page = alloc_pages(gfp_mask | __GFP_ZERO, 0);
1683 if (page)
1684 return (unsigned long) page_address(page);
1685 return 0;
1688 EXPORT_SYMBOL(get_zeroed_page);
1690 void __pagevec_free(struct pagevec *pvec)
1692 int i = pagevec_count(pvec);
1694 while (--i >= 0)
1695 free_hot_cold_page(pvec->pages[i], pvec->cold);
1698 void __free_pages(struct page *page, unsigned int order)
1700 if (put_page_testzero(page)) {
1701 if (order == 0)
1702 free_hot_page(page);
1703 else
1704 __free_pages_ok(page, order);
1708 EXPORT_SYMBOL(__free_pages);
1710 void free_pages(unsigned long addr, unsigned int order)
1712 if (addr != 0) {
1713 VM_BUG_ON(!virt_addr_valid((void *)addr));
1714 __free_pages(virt_to_page((void *)addr), order);
1718 EXPORT_SYMBOL(free_pages);
1720 static unsigned int nr_free_zone_pages(int offset)
1722 struct zoneref *z;
1723 struct zone *zone;
1725 /* Just pick one node, since fallback list is circular */
1726 unsigned int sum = 0;
1728 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
1730 for_each_zone_zonelist(zone, z, zonelist, offset) {
1731 unsigned long size = zone->present_pages;
1732 unsigned long high = zone->pages_high;
1733 if (size > high)
1734 sum += size - high;
1737 return sum;
1741 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
1743 unsigned int nr_free_buffer_pages(void)
1745 return nr_free_zone_pages(gfp_zone(GFP_USER));
1747 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
1750 * Amount of free RAM allocatable within all zones
1752 unsigned int nr_free_pagecache_pages(void)
1754 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
1757 static inline void show_node(struct zone *zone)
1759 if (NUMA_BUILD)
1760 printk("Node %d ", zone_to_nid(zone));
1763 void si_meminfo(struct sysinfo *val)
1765 val->totalram = totalram_pages;
1766 val->sharedram = 0;
1767 val->freeram = global_page_state(NR_FREE_PAGES);
1768 val->bufferram = nr_blockdev_pages();
1769 val->totalhigh = totalhigh_pages;
1770 val->freehigh = nr_free_highpages();
1771 val->mem_unit = PAGE_SIZE;
1774 EXPORT_SYMBOL(si_meminfo);
1776 #ifdef CONFIG_NUMA
1777 void si_meminfo_node(struct sysinfo *val, int nid)
1779 pg_data_t *pgdat = NODE_DATA(nid);
1781 val->totalram = pgdat->node_present_pages;
1782 val->freeram = node_page_state(nid, NR_FREE_PAGES);
1783 #ifdef CONFIG_HIGHMEM
1784 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
1785 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
1786 NR_FREE_PAGES);
1787 #else
1788 val->totalhigh = 0;
1789 val->freehigh = 0;
1790 #endif
1791 val->mem_unit = PAGE_SIZE;
1793 #endif
1795 #define K(x) ((x) << (PAGE_SHIFT-10))
1798 * Show free area list (used inside shift_scroll-lock stuff)
1799 * We also calculate the percentage fragmentation. We do this by counting the
1800 * memory on each free list with the exception of the first item on the list.
1802 void show_free_areas(void)
1804 int cpu;
1805 struct zone *zone;
1807 for_each_zone(zone) {
1808 if (!populated_zone(zone))
1809 continue;
1811 show_node(zone);
1812 printk("%s per-cpu:\n", zone->name);
1814 for_each_online_cpu(cpu) {
1815 struct per_cpu_pageset *pageset;
1817 pageset = zone_pcp(zone, cpu);
1819 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
1820 cpu, pageset->pcp.high,
1821 pageset->pcp.batch, pageset->pcp.count);
1825 printk("Active:%lu inactive:%lu dirty:%lu writeback:%lu unstable:%lu\n"
1826 " free:%lu slab:%lu mapped:%lu pagetables:%lu bounce:%lu\n",
1827 global_page_state(NR_ACTIVE),
1828 global_page_state(NR_INACTIVE),
1829 global_page_state(NR_FILE_DIRTY),
1830 global_page_state(NR_WRITEBACK),
1831 global_page_state(NR_UNSTABLE_NFS),
1832 global_page_state(NR_FREE_PAGES),
1833 global_page_state(NR_SLAB_RECLAIMABLE) +
1834 global_page_state(NR_SLAB_UNRECLAIMABLE),
1835 global_page_state(NR_FILE_MAPPED),
1836 global_page_state(NR_PAGETABLE),
1837 global_page_state(NR_BOUNCE));
1839 for_each_zone(zone) {
1840 int i;
1842 if (!populated_zone(zone))
1843 continue;
1845 show_node(zone);
1846 printk("%s"
1847 " free:%lukB"
1848 " min:%lukB"
1849 " low:%lukB"
1850 " high:%lukB"
1851 " active:%lukB"
1852 " inactive:%lukB"
1853 " present:%lukB"
1854 " pages_scanned:%lu"
1855 " all_unreclaimable? %s"
1856 "\n",
1857 zone->name,
1858 K(zone_page_state(zone, NR_FREE_PAGES)),
1859 K(zone->pages_min),
1860 K(zone->pages_low),
1861 K(zone->pages_high),
1862 K(zone_page_state(zone, NR_ACTIVE)),
1863 K(zone_page_state(zone, NR_INACTIVE)),
1864 K(zone->present_pages),
1865 zone->pages_scanned,
1866 (zone_is_all_unreclaimable(zone) ? "yes" : "no")
1868 printk("lowmem_reserve[]:");
1869 for (i = 0; i < MAX_NR_ZONES; i++)
1870 printk(" %lu", zone->lowmem_reserve[i]);
1871 printk("\n");
1874 for_each_zone(zone) {
1875 unsigned long nr[MAX_ORDER], flags, order, total = 0;
1877 if (!populated_zone(zone))
1878 continue;
1880 show_node(zone);
1881 printk("%s: ", zone->name);
1883 spin_lock_irqsave(&zone->lock, flags);
1884 for (order = 0; order < MAX_ORDER; order++) {
1885 nr[order] = zone->free_area[order].nr_free;
1886 total += nr[order] << order;
1888 spin_unlock_irqrestore(&zone->lock, flags);
1889 for (order = 0; order < MAX_ORDER; order++)
1890 printk("%lu*%lukB ", nr[order], K(1UL) << order);
1891 printk("= %lukB\n", K(total));
1894 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
1896 show_swap_cache_info();
1899 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
1901 zoneref->zone = zone;
1902 zoneref->zone_idx = zone_idx(zone);
1906 * Builds allocation fallback zone lists.
1908 * Add all populated zones of a node to the zonelist.
1910 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
1911 int nr_zones, enum zone_type zone_type)
1913 struct zone *zone;
1915 BUG_ON(zone_type >= MAX_NR_ZONES);
1916 zone_type++;
1918 do {
1919 zone_type--;
1920 zone = pgdat->node_zones + zone_type;
1921 if (populated_zone(zone)) {
1922 zoneref_set_zone(zone,
1923 &zonelist->_zonerefs[nr_zones++]);
1924 check_highest_zone(zone_type);
1927 } while (zone_type);
1928 return nr_zones;
1933 * zonelist_order:
1934 * 0 = automatic detection of better ordering.
1935 * 1 = order by ([node] distance, -zonetype)
1936 * 2 = order by (-zonetype, [node] distance)
1938 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
1939 * the same zonelist. So only NUMA can configure this param.
1941 #define ZONELIST_ORDER_DEFAULT 0
1942 #define ZONELIST_ORDER_NODE 1
1943 #define ZONELIST_ORDER_ZONE 2
1945 /* zonelist order in the kernel.
1946 * set_zonelist_order() will set this to NODE or ZONE.
1948 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
1949 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
1952 #ifdef CONFIG_NUMA
1953 /* The value user specified ....changed by config */
1954 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
1955 /* string for sysctl */
1956 #define NUMA_ZONELIST_ORDER_LEN 16
1957 char numa_zonelist_order[16] = "default";
1960 * interface for configure zonelist ordering.
1961 * command line option "numa_zonelist_order"
1962 * = "[dD]efault - default, automatic configuration.
1963 * = "[nN]ode - order by node locality, then by zone within node
1964 * = "[zZ]one - order by zone, then by locality within zone
1967 static int __parse_numa_zonelist_order(char *s)
1969 if (*s == 'd' || *s == 'D') {
1970 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
1971 } else if (*s == 'n' || *s == 'N') {
1972 user_zonelist_order = ZONELIST_ORDER_NODE;
1973 } else if (*s == 'z' || *s == 'Z') {
1974 user_zonelist_order = ZONELIST_ORDER_ZONE;
1975 } else {
1976 printk(KERN_WARNING
1977 "Ignoring invalid numa_zonelist_order value: "
1978 "%s\n", s);
1979 return -EINVAL;
1981 return 0;
1984 static __init int setup_numa_zonelist_order(char *s)
1986 if (s)
1987 return __parse_numa_zonelist_order(s);
1988 return 0;
1990 early_param("numa_zonelist_order", setup_numa_zonelist_order);
1993 * sysctl handler for numa_zonelist_order
1995 int numa_zonelist_order_handler(ctl_table *table, int write,
1996 struct file *file, void __user *buffer, size_t *length,
1997 loff_t *ppos)
1999 char saved_string[NUMA_ZONELIST_ORDER_LEN];
2000 int ret;
2002 if (write)
2003 strncpy(saved_string, (char*)table->data,
2004 NUMA_ZONELIST_ORDER_LEN);
2005 ret = proc_dostring(table, write, file, buffer, length, ppos);
2006 if (ret)
2007 return ret;
2008 if (write) {
2009 int oldval = user_zonelist_order;
2010 if (__parse_numa_zonelist_order((char*)table->data)) {
2012 * bogus value. restore saved string
2014 strncpy((char*)table->data, saved_string,
2015 NUMA_ZONELIST_ORDER_LEN);
2016 user_zonelist_order = oldval;
2017 } else if (oldval != user_zonelist_order)
2018 build_all_zonelists();
2020 return 0;
2024 #define MAX_NODE_LOAD (num_online_nodes())
2025 static int node_load[MAX_NUMNODES];
2028 * find_next_best_node - find the next node that should appear in a given node's fallback list
2029 * @node: node whose fallback list we're appending
2030 * @used_node_mask: nodemask_t of already used nodes
2032 * We use a number of factors to determine which is the next node that should
2033 * appear on a given node's fallback list. The node should not have appeared
2034 * already in @node's fallback list, and it should be the next closest node
2035 * according to the distance array (which contains arbitrary distance values
2036 * from each node to each node in the system), and should also prefer nodes
2037 * with no CPUs, since presumably they'll have very little allocation pressure
2038 * on them otherwise.
2039 * It returns -1 if no node is found.
2041 static int find_next_best_node(int node, nodemask_t *used_node_mask)
2043 int n, val;
2044 int min_val = INT_MAX;
2045 int best_node = -1;
2046 node_to_cpumask_ptr(tmp, 0);
2048 /* Use the local node if we haven't already */
2049 if (!node_isset(node, *used_node_mask)) {
2050 node_set(node, *used_node_mask);
2051 return node;
2054 for_each_node_state(n, N_HIGH_MEMORY) {
2056 /* Don't want a node to appear more than once */
2057 if (node_isset(n, *used_node_mask))
2058 continue;
2060 /* Use the distance array to find the distance */
2061 val = node_distance(node, n);
2063 /* Penalize nodes under us ("prefer the next node") */
2064 val += (n < node);
2066 /* Give preference to headless and unused nodes */
2067 node_to_cpumask_ptr_next(tmp, n);
2068 if (!cpus_empty(*tmp))
2069 val += PENALTY_FOR_NODE_WITH_CPUS;
2071 /* Slight preference for less loaded node */
2072 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
2073 val += node_load[n];
2075 if (val < min_val) {
2076 min_val = val;
2077 best_node = n;
2081 if (best_node >= 0)
2082 node_set(best_node, *used_node_mask);
2084 return best_node;
2089 * Build zonelists ordered by node and zones within node.
2090 * This results in maximum locality--normal zone overflows into local
2091 * DMA zone, if any--but risks exhausting DMA zone.
2093 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
2095 int j;
2096 struct zonelist *zonelist;
2098 zonelist = &pgdat->node_zonelists[0];
2099 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
2101 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2102 MAX_NR_ZONES - 1);
2103 zonelist->_zonerefs[j].zone = NULL;
2104 zonelist->_zonerefs[j].zone_idx = 0;
2108 * Build gfp_thisnode zonelists
2110 static void build_thisnode_zonelists(pg_data_t *pgdat)
2112 int j;
2113 struct zonelist *zonelist;
2115 zonelist = &pgdat->node_zonelists[1];
2116 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2117 zonelist->_zonerefs[j].zone = NULL;
2118 zonelist->_zonerefs[j].zone_idx = 0;
2122 * Build zonelists ordered by zone and nodes within zones.
2123 * This results in conserving DMA zone[s] until all Normal memory is
2124 * exhausted, but results in overflowing to remote node while memory
2125 * may still exist in local DMA zone.
2127 static int node_order[MAX_NUMNODES];
2129 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
2131 int pos, j, node;
2132 int zone_type; /* needs to be signed */
2133 struct zone *z;
2134 struct zonelist *zonelist;
2136 zonelist = &pgdat->node_zonelists[0];
2137 pos = 0;
2138 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
2139 for (j = 0; j < nr_nodes; j++) {
2140 node = node_order[j];
2141 z = &NODE_DATA(node)->node_zones[zone_type];
2142 if (populated_zone(z)) {
2143 zoneref_set_zone(z,
2144 &zonelist->_zonerefs[pos++]);
2145 check_highest_zone(zone_type);
2149 zonelist->_zonerefs[pos].zone = NULL;
2150 zonelist->_zonerefs[pos].zone_idx = 0;
2153 static int default_zonelist_order(void)
2155 int nid, zone_type;
2156 unsigned long low_kmem_size,total_size;
2157 struct zone *z;
2158 int average_size;
2160 * ZONE_DMA and ZONE_DMA32 can be very small area in the sytem.
2161 * If they are really small and used heavily, the system can fall
2162 * into OOM very easily.
2163 * This function detect ZONE_DMA/DMA32 size and confgigures zone order.
2165 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
2166 low_kmem_size = 0;
2167 total_size = 0;
2168 for_each_online_node(nid) {
2169 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2170 z = &NODE_DATA(nid)->node_zones[zone_type];
2171 if (populated_zone(z)) {
2172 if (zone_type < ZONE_NORMAL)
2173 low_kmem_size += z->present_pages;
2174 total_size += z->present_pages;
2178 if (!low_kmem_size || /* there are no DMA area. */
2179 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
2180 return ZONELIST_ORDER_NODE;
2182 * look into each node's config.
2183 * If there is a node whose DMA/DMA32 memory is very big area on
2184 * local memory, NODE_ORDER may be suitable.
2186 average_size = total_size /
2187 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1);
2188 for_each_online_node(nid) {
2189 low_kmem_size = 0;
2190 total_size = 0;
2191 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2192 z = &NODE_DATA(nid)->node_zones[zone_type];
2193 if (populated_zone(z)) {
2194 if (zone_type < ZONE_NORMAL)
2195 low_kmem_size += z->present_pages;
2196 total_size += z->present_pages;
2199 if (low_kmem_size &&
2200 total_size > average_size && /* ignore small node */
2201 low_kmem_size > total_size * 70/100)
2202 return ZONELIST_ORDER_NODE;
2204 return ZONELIST_ORDER_ZONE;
2207 static void set_zonelist_order(void)
2209 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
2210 current_zonelist_order = default_zonelist_order();
2211 else
2212 current_zonelist_order = user_zonelist_order;
2215 static void build_zonelists(pg_data_t *pgdat)
2217 int j, node, load;
2218 enum zone_type i;
2219 nodemask_t used_mask;
2220 int local_node, prev_node;
2221 struct zonelist *zonelist;
2222 int order = current_zonelist_order;
2224 /* initialize zonelists */
2225 for (i = 0; i < MAX_ZONELISTS; i++) {
2226 zonelist = pgdat->node_zonelists + i;
2227 zonelist->_zonerefs[0].zone = NULL;
2228 zonelist->_zonerefs[0].zone_idx = 0;
2231 /* NUMA-aware ordering of nodes */
2232 local_node = pgdat->node_id;
2233 load = num_online_nodes();
2234 prev_node = local_node;
2235 nodes_clear(used_mask);
2237 memset(node_load, 0, sizeof(node_load));
2238 memset(node_order, 0, sizeof(node_order));
2239 j = 0;
2241 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
2242 int distance = node_distance(local_node, node);
2245 * If another node is sufficiently far away then it is better
2246 * to reclaim pages in a zone before going off node.
2248 if (distance > RECLAIM_DISTANCE)
2249 zone_reclaim_mode = 1;
2252 * We don't want to pressure a particular node.
2253 * So adding penalty to the first node in same
2254 * distance group to make it round-robin.
2256 if (distance != node_distance(local_node, prev_node))
2257 node_load[node] = load;
2259 prev_node = node;
2260 load--;
2261 if (order == ZONELIST_ORDER_NODE)
2262 build_zonelists_in_node_order(pgdat, node);
2263 else
2264 node_order[j++] = node; /* remember order */
2267 if (order == ZONELIST_ORDER_ZONE) {
2268 /* calculate node order -- i.e., DMA last! */
2269 build_zonelists_in_zone_order(pgdat, j);
2272 build_thisnode_zonelists(pgdat);
2275 /* Construct the zonelist performance cache - see further mmzone.h */
2276 static void build_zonelist_cache(pg_data_t *pgdat)
2278 struct zonelist *zonelist;
2279 struct zonelist_cache *zlc;
2280 struct zoneref *z;
2282 zonelist = &pgdat->node_zonelists[0];
2283 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
2284 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
2285 for (z = zonelist->_zonerefs; z->zone; z++)
2286 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
2290 #else /* CONFIG_NUMA */
2292 static void set_zonelist_order(void)
2294 current_zonelist_order = ZONELIST_ORDER_ZONE;
2297 static void build_zonelists(pg_data_t *pgdat)
2299 int node, local_node;
2300 enum zone_type j;
2301 struct zonelist *zonelist;
2303 local_node = pgdat->node_id;
2305 zonelist = &pgdat->node_zonelists[0];
2306 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2309 * Now we build the zonelist so that it contains the zones
2310 * of all the other nodes.
2311 * We don't want to pressure a particular node, so when
2312 * building the zones for node N, we make sure that the
2313 * zones coming right after the local ones are those from
2314 * node N+1 (modulo N)
2316 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
2317 if (!node_online(node))
2318 continue;
2319 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2320 MAX_NR_ZONES - 1);
2322 for (node = 0; node < local_node; node++) {
2323 if (!node_online(node))
2324 continue;
2325 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2326 MAX_NR_ZONES - 1);
2329 zonelist->_zonerefs[j].zone = NULL;
2330 zonelist->_zonerefs[j].zone_idx = 0;
2333 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
2334 static void build_zonelist_cache(pg_data_t *pgdat)
2336 pgdat->node_zonelists[0].zlcache_ptr = NULL;
2337 pgdat->node_zonelists[1].zlcache_ptr = NULL;
2340 #endif /* CONFIG_NUMA */
2342 /* return values int ....just for stop_machine_run() */
2343 static int __build_all_zonelists(void *dummy)
2345 int nid;
2347 for_each_online_node(nid) {
2348 pg_data_t *pgdat = NODE_DATA(nid);
2350 build_zonelists(pgdat);
2351 build_zonelist_cache(pgdat);
2353 return 0;
2356 void build_all_zonelists(void)
2358 set_zonelist_order();
2360 if (system_state == SYSTEM_BOOTING) {
2361 __build_all_zonelists(NULL);
2362 cpuset_init_current_mems_allowed();
2363 } else {
2364 /* we have to stop all cpus to guarantee there is no user
2365 of zonelist */
2366 stop_machine_run(__build_all_zonelists, NULL, NR_CPUS);
2367 /* cpuset refresh routine should be here */
2369 vm_total_pages = nr_free_pagecache_pages();
2371 * Disable grouping by mobility if the number of pages in the
2372 * system is too low to allow the mechanism to work. It would be
2373 * more accurate, but expensive to check per-zone. This check is
2374 * made on memory-hotadd so a system can start with mobility
2375 * disabled and enable it later
2377 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
2378 page_group_by_mobility_disabled = 1;
2379 else
2380 page_group_by_mobility_disabled = 0;
2382 printk("Built %i zonelists in %s order, mobility grouping %s. "
2383 "Total pages: %ld\n",
2384 num_online_nodes(),
2385 zonelist_order_name[current_zonelist_order],
2386 page_group_by_mobility_disabled ? "off" : "on",
2387 vm_total_pages);
2388 #ifdef CONFIG_NUMA
2389 printk("Policy zone: %s\n", zone_names[policy_zone]);
2390 #endif
2394 * Helper functions to size the waitqueue hash table.
2395 * Essentially these want to choose hash table sizes sufficiently
2396 * large so that collisions trying to wait on pages are rare.
2397 * But in fact, the number of active page waitqueues on typical
2398 * systems is ridiculously low, less than 200. So this is even
2399 * conservative, even though it seems large.
2401 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
2402 * waitqueues, i.e. the size of the waitq table given the number of pages.
2404 #define PAGES_PER_WAITQUEUE 256
2406 #ifndef CONFIG_MEMORY_HOTPLUG
2407 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
2409 unsigned long size = 1;
2411 pages /= PAGES_PER_WAITQUEUE;
2413 while (size < pages)
2414 size <<= 1;
2417 * Once we have dozens or even hundreds of threads sleeping
2418 * on IO we've got bigger problems than wait queue collision.
2419 * Limit the size of the wait table to a reasonable size.
2421 size = min(size, 4096UL);
2423 return max(size, 4UL);
2425 #else
2427 * A zone's size might be changed by hot-add, so it is not possible to determine
2428 * a suitable size for its wait_table. So we use the maximum size now.
2430 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
2432 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
2433 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
2434 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
2436 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
2437 * or more by the traditional way. (See above). It equals:
2439 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
2440 * ia64(16K page size) : = ( 8G + 4M)byte.
2441 * powerpc (64K page size) : = (32G +16M)byte.
2443 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
2445 return 4096UL;
2447 #endif
2450 * This is an integer logarithm so that shifts can be used later
2451 * to extract the more random high bits from the multiplicative
2452 * hash function before the remainder is taken.
2454 static inline unsigned long wait_table_bits(unsigned long size)
2456 return ffz(~size);
2459 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
2462 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
2463 * of blocks reserved is based on zone->pages_min. The memory within the
2464 * reserve will tend to store contiguous free pages. Setting min_free_kbytes
2465 * higher will lead to a bigger reserve which will get freed as contiguous
2466 * blocks as reclaim kicks in
2468 static void setup_zone_migrate_reserve(struct zone *zone)
2470 unsigned long start_pfn, pfn, end_pfn;
2471 struct page *page;
2472 unsigned long reserve, block_migratetype;
2474 /* Get the start pfn, end pfn and the number of blocks to reserve */
2475 start_pfn = zone->zone_start_pfn;
2476 end_pfn = start_pfn + zone->spanned_pages;
2477 reserve = roundup(zone->pages_min, pageblock_nr_pages) >>
2478 pageblock_order;
2480 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
2481 if (!pfn_valid(pfn))
2482 continue;
2483 page = pfn_to_page(pfn);
2485 /* Blocks with reserved pages will never free, skip them. */
2486 if (PageReserved(page))
2487 continue;
2489 block_migratetype = get_pageblock_migratetype(page);
2491 /* If this block is reserved, account for it */
2492 if (reserve > 0 && block_migratetype == MIGRATE_RESERVE) {
2493 reserve--;
2494 continue;
2497 /* Suitable for reserving if this block is movable */
2498 if (reserve > 0 && block_migratetype == MIGRATE_MOVABLE) {
2499 set_pageblock_migratetype(page, MIGRATE_RESERVE);
2500 move_freepages_block(zone, page, MIGRATE_RESERVE);
2501 reserve--;
2502 continue;
2506 * If the reserve is met and this is a previous reserved block,
2507 * take it back
2509 if (block_migratetype == MIGRATE_RESERVE) {
2510 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
2511 move_freepages_block(zone, page, MIGRATE_MOVABLE);
2517 * Initially all pages are reserved - free ones are freed
2518 * up by free_all_bootmem() once the early boot process is
2519 * done. Non-atomic initialization, single-pass.
2521 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
2522 unsigned long start_pfn, enum memmap_context context)
2524 struct page *page;
2525 unsigned long end_pfn = start_pfn + size;
2526 unsigned long pfn;
2528 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
2530 * There can be holes in boot-time mem_map[]s
2531 * handed to this function. They do not
2532 * exist on hotplugged memory.
2534 if (context == MEMMAP_EARLY) {
2535 if (!early_pfn_valid(pfn))
2536 continue;
2537 if (!early_pfn_in_nid(pfn, nid))
2538 continue;
2540 page = pfn_to_page(pfn);
2541 set_page_links(page, zone, nid, pfn);
2542 init_page_count(page);
2543 reset_page_mapcount(page);
2544 SetPageReserved(page);
2547 * Mark the block movable so that blocks are reserved for
2548 * movable at startup. This will force kernel allocations
2549 * to reserve their blocks rather than leaking throughout
2550 * the address space during boot when many long-lived
2551 * kernel allocations are made. Later some blocks near
2552 * the start are marked MIGRATE_RESERVE by
2553 * setup_zone_migrate_reserve()
2555 if ((pfn & (pageblock_nr_pages-1)))
2556 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
2558 INIT_LIST_HEAD(&page->lru);
2559 #ifdef WANT_PAGE_VIRTUAL
2560 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
2561 if (!is_highmem_idx(zone))
2562 set_page_address(page, __va(pfn << PAGE_SHIFT));
2563 #endif
2567 static void __meminit zone_init_free_lists(struct zone *zone)
2569 int order, t;
2570 for_each_migratetype_order(order, t) {
2571 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
2572 zone->free_area[order].nr_free = 0;
2576 #ifndef __HAVE_ARCH_MEMMAP_INIT
2577 #define memmap_init(size, nid, zone, start_pfn) \
2578 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
2579 #endif
2581 static int zone_batchsize(struct zone *zone)
2583 int batch;
2586 * The per-cpu-pages pools are set to around 1000th of the
2587 * size of the zone. But no more than 1/2 of a meg.
2589 * OK, so we don't know how big the cache is. So guess.
2591 batch = zone->present_pages / 1024;
2592 if (batch * PAGE_SIZE > 512 * 1024)
2593 batch = (512 * 1024) / PAGE_SIZE;
2594 batch /= 4; /* We effectively *= 4 below */
2595 if (batch < 1)
2596 batch = 1;
2599 * Clamp the batch to a 2^n - 1 value. Having a power
2600 * of 2 value was found to be more likely to have
2601 * suboptimal cache aliasing properties in some cases.
2603 * For example if 2 tasks are alternately allocating
2604 * batches of pages, one task can end up with a lot
2605 * of pages of one half of the possible page colors
2606 * and the other with pages of the other colors.
2608 batch = (1 << (fls(batch + batch/2)-1)) - 1;
2610 return batch;
2613 inline void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
2615 struct per_cpu_pages *pcp;
2617 memset(p, 0, sizeof(*p));
2619 pcp = &p->pcp;
2620 pcp->count = 0;
2621 pcp->high = 6 * batch;
2622 pcp->batch = max(1UL, 1 * batch);
2623 INIT_LIST_HEAD(&pcp->list);
2627 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
2628 * to the value high for the pageset p.
2631 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
2632 unsigned long high)
2634 struct per_cpu_pages *pcp;
2636 pcp = &p->pcp;
2637 pcp->high = high;
2638 pcp->batch = max(1UL, high/4);
2639 if ((high/4) > (PAGE_SHIFT * 8))
2640 pcp->batch = PAGE_SHIFT * 8;
2644 #ifdef CONFIG_NUMA
2646 * Boot pageset table. One per cpu which is going to be used for all
2647 * zones and all nodes. The parameters will be set in such a way
2648 * that an item put on a list will immediately be handed over to
2649 * the buddy list. This is safe since pageset manipulation is done
2650 * with interrupts disabled.
2652 * Some NUMA counter updates may also be caught by the boot pagesets.
2654 * The boot_pagesets must be kept even after bootup is complete for
2655 * unused processors and/or zones. They do play a role for bootstrapping
2656 * hotplugged processors.
2658 * zoneinfo_show() and maybe other functions do
2659 * not check if the processor is online before following the pageset pointer.
2660 * Other parts of the kernel may not check if the zone is available.
2662 static struct per_cpu_pageset boot_pageset[NR_CPUS];
2665 * Dynamically allocate memory for the
2666 * per cpu pageset array in struct zone.
2668 static int __cpuinit process_zones(int cpu)
2670 struct zone *zone, *dzone;
2671 int node = cpu_to_node(cpu);
2673 node_set_state(node, N_CPU); /* this node has a cpu */
2675 for_each_zone(zone) {
2677 if (!populated_zone(zone))
2678 continue;
2680 zone_pcp(zone, cpu) = kmalloc_node(sizeof(struct per_cpu_pageset),
2681 GFP_KERNEL, node);
2682 if (!zone_pcp(zone, cpu))
2683 goto bad;
2685 setup_pageset(zone_pcp(zone, cpu), zone_batchsize(zone));
2687 if (percpu_pagelist_fraction)
2688 setup_pagelist_highmark(zone_pcp(zone, cpu),
2689 (zone->present_pages / percpu_pagelist_fraction));
2692 return 0;
2693 bad:
2694 for_each_zone(dzone) {
2695 if (!populated_zone(dzone))
2696 continue;
2697 if (dzone == zone)
2698 break;
2699 kfree(zone_pcp(dzone, cpu));
2700 zone_pcp(dzone, cpu) = NULL;
2702 return -ENOMEM;
2705 static inline void free_zone_pagesets(int cpu)
2707 struct zone *zone;
2709 for_each_zone(zone) {
2710 struct per_cpu_pageset *pset = zone_pcp(zone, cpu);
2712 /* Free per_cpu_pageset if it is slab allocated */
2713 if (pset != &boot_pageset[cpu])
2714 kfree(pset);
2715 zone_pcp(zone, cpu) = NULL;
2719 static int __cpuinit pageset_cpuup_callback(struct notifier_block *nfb,
2720 unsigned long action,
2721 void *hcpu)
2723 int cpu = (long)hcpu;
2724 int ret = NOTIFY_OK;
2726 switch (action) {
2727 case CPU_UP_PREPARE:
2728 case CPU_UP_PREPARE_FROZEN:
2729 if (process_zones(cpu))
2730 ret = NOTIFY_BAD;
2731 break;
2732 case CPU_UP_CANCELED:
2733 case CPU_UP_CANCELED_FROZEN:
2734 case CPU_DEAD:
2735 case CPU_DEAD_FROZEN:
2736 free_zone_pagesets(cpu);
2737 break;
2738 default:
2739 break;
2741 return ret;
2744 static struct notifier_block __cpuinitdata pageset_notifier =
2745 { &pageset_cpuup_callback, NULL, 0 };
2747 void __init setup_per_cpu_pageset(void)
2749 int err;
2751 /* Initialize per_cpu_pageset for cpu 0.
2752 * A cpuup callback will do this for every cpu
2753 * as it comes online
2755 err = process_zones(smp_processor_id());
2756 BUG_ON(err);
2757 register_cpu_notifier(&pageset_notifier);
2760 #endif
2762 static noinline __init_refok
2763 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
2765 int i;
2766 struct pglist_data *pgdat = zone->zone_pgdat;
2767 size_t alloc_size;
2770 * The per-page waitqueue mechanism uses hashed waitqueues
2771 * per zone.
2773 zone->wait_table_hash_nr_entries =
2774 wait_table_hash_nr_entries(zone_size_pages);
2775 zone->wait_table_bits =
2776 wait_table_bits(zone->wait_table_hash_nr_entries);
2777 alloc_size = zone->wait_table_hash_nr_entries
2778 * sizeof(wait_queue_head_t);
2780 if (system_state == SYSTEM_BOOTING) {
2781 zone->wait_table = (wait_queue_head_t *)
2782 alloc_bootmem_node(pgdat, alloc_size);
2783 } else {
2785 * This case means that a zone whose size was 0 gets new memory
2786 * via memory hot-add.
2787 * But it may be the case that a new node was hot-added. In
2788 * this case vmalloc() will not be able to use this new node's
2789 * memory - this wait_table must be initialized to use this new
2790 * node itself as well.
2791 * To use this new node's memory, further consideration will be
2792 * necessary.
2794 zone->wait_table = vmalloc(alloc_size);
2796 if (!zone->wait_table)
2797 return -ENOMEM;
2799 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
2800 init_waitqueue_head(zone->wait_table + i);
2802 return 0;
2805 static __meminit void zone_pcp_init(struct zone *zone)
2807 int cpu;
2808 unsigned long batch = zone_batchsize(zone);
2810 for (cpu = 0; cpu < NR_CPUS; cpu++) {
2811 #ifdef CONFIG_NUMA
2812 /* Early boot. Slab allocator not functional yet */
2813 zone_pcp(zone, cpu) = &boot_pageset[cpu];
2814 setup_pageset(&boot_pageset[cpu],0);
2815 #else
2816 setup_pageset(zone_pcp(zone,cpu), batch);
2817 #endif
2819 if (zone->present_pages)
2820 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%lu\n",
2821 zone->name, zone->present_pages, batch);
2824 __meminit int init_currently_empty_zone(struct zone *zone,
2825 unsigned long zone_start_pfn,
2826 unsigned long size,
2827 enum memmap_context context)
2829 struct pglist_data *pgdat = zone->zone_pgdat;
2830 int ret;
2831 ret = zone_wait_table_init(zone, size);
2832 if (ret)
2833 return ret;
2834 pgdat->nr_zones = zone_idx(zone) + 1;
2836 zone->zone_start_pfn = zone_start_pfn;
2838 memmap_init(size, pgdat->node_id, zone_idx(zone), zone_start_pfn);
2840 zone_init_free_lists(zone);
2842 return 0;
2845 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
2847 * Basic iterator support. Return the first range of PFNs for a node
2848 * Note: nid == MAX_NUMNODES returns first region regardless of node
2850 static int __meminit first_active_region_index_in_nid(int nid)
2852 int i;
2854 for (i = 0; i < nr_nodemap_entries; i++)
2855 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
2856 return i;
2858 return -1;
2862 * Basic iterator support. Return the next active range of PFNs for a node
2863 * Note: nid == MAX_NUMNODES returns next region regardless of node
2865 static int __meminit next_active_region_index_in_nid(int index, int nid)
2867 for (index = index + 1; index < nr_nodemap_entries; index++)
2868 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
2869 return index;
2871 return -1;
2874 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
2876 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
2877 * Architectures may implement their own version but if add_active_range()
2878 * was used and there are no special requirements, this is a convenient
2879 * alternative
2881 int __meminit early_pfn_to_nid(unsigned long pfn)
2883 int i;
2885 for (i = 0; i < nr_nodemap_entries; i++) {
2886 unsigned long start_pfn = early_node_map[i].start_pfn;
2887 unsigned long end_pfn = early_node_map[i].end_pfn;
2889 if (start_pfn <= pfn && pfn < end_pfn)
2890 return early_node_map[i].nid;
2893 return 0;
2895 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
2897 /* Basic iterator support to walk early_node_map[] */
2898 #define for_each_active_range_index_in_nid(i, nid) \
2899 for (i = first_active_region_index_in_nid(nid); i != -1; \
2900 i = next_active_region_index_in_nid(i, nid))
2903 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
2904 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
2905 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
2907 * If an architecture guarantees that all ranges registered with
2908 * add_active_ranges() contain no holes and may be freed, this
2909 * this function may be used instead of calling free_bootmem() manually.
2911 void __init free_bootmem_with_active_regions(int nid,
2912 unsigned long max_low_pfn)
2914 int i;
2916 for_each_active_range_index_in_nid(i, nid) {
2917 unsigned long size_pages = 0;
2918 unsigned long end_pfn = early_node_map[i].end_pfn;
2920 if (early_node_map[i].start_pfn >= max_low_pfn)
2921 continue;
2923 if (end_pfn > max_low_pfn)
2924 end_pfn = max_low_pfn;
2926 size_pages = end_pfn - early_node_map[i].start_pfn;
2927 free_bootmem_node(NODE_DATA(early_node_map[i].nid),
2928 PFN_PHYS(early_node_map[i].start_pfn),
2929 size_pages << PAGE_SHIFT);
2934 * sparse_memory_present_with_active_regions - Call memory_present for each active range
2935 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
2937 * If an architecture guarantees that all ranges registered with
2938 * add_active_ranges() contain no holes and may be freed, this
2939 * function may be used instead of calling memory_present() manually.
2941 void __init sparse_memory_present_with_active_regions(int nid)
2943 int i;
2945 for_each_active_range_index_in_nid(i, nid)
2946 memory_present(early_node_map[i].nid,
2947 early_node_map[i].start_pfn,
2948 early_node_map[i].end_pfn);
2952 * push_node_boundaries - Push node boundaries to at least the requested boundary
2953 * @nid: The nid of the node to push the boundary for
2954 * @start_pfn: The start pfn of the node
2955 * @end_pfn: The end pfn of the node
2957 * In reserve-based hot-add, mem_map is allocated that is unused until hotadd
2958 * time. Specifically, on x86_64, SRAT will report ranges that can potentially
2959 * be hotplugged even though no physical memory exists. This function allows
2960 * an arch to push out the node boundaries so mem_map is allocated that can
2961 * be used later.
2963 #ifdef CONFIG_MEMORY_HOTPLUG_RESERVE
2964 void __init push_node_boundaries(unsigned int nid,
2965 unsigned long start_pfn, unsigned long end_pfn)
2967 printk(KERN_DEBUG "Entering push_node_boundaries(%u, %lu, %lu)\n",
2968 nid, start_pfn, end_pfn);
2970 /* Initialise the boundary for this node if necessary */
2971 if (node_boundary_end_pfn[nid] == 0)
2972 node_boundary_start_pfn[nid] = -1UL;
2974 /* Update the boundaries */
2975 if (node_boundary_start_pfn[nid] > start_pfn)
2976 node_boundary_start_pfn[nid] = start_pfn;
2977 if (node_boundary_end_pfn[nid] < end_pfn)
2978 node_boundary_end_pfn[nid] = end_pfn;
2981 /* If necessary, push the node boundary out for reserve hotadd */
2982 static void __meminit account_node_boundary(unsigned int nid,
2983 unsigned long *start_pfn, unsigned long *end_pfn)
2985 printk(KERN_DEBUG "Entering account_node_boundary(%u, %lu, %lu)\n",
2986 nid, *start_pfn, *end_pfn);
2988 /* Return if boundary information has not been provided */
2989 if (node_boundary_end_pfn[nid] == 0)
2990 return;
2992 /* Check the boundaries and update if necessary */
2993 if (node_boundary_start_pfn[nid] < *start_pfn)
2994 *start_pfn = node_boundary_start_pfn[nid];
2995 if (node_boundary_end_pfn[nid] > *end_pfn)
2996 *end_pfn = node_boundary_end_pfn[nid];
2998 #else
2999 void __init push_node_boundaries(unsigned int nid,
3000 unsigned long start_pfn, unsigned long end_pfn) {}
3002 static void __meminit account_node_boundary(unsigned int nid,
3003 unsigned long *start_pfn, unsigned long *end_pfn) {}
3004 #endif
3008 * get_pfn_range_for_nid - Return the start and end page frames for a node
3009 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
3010 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
3011 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
3013 * It returns the start and end page frame of a node based on information
3014 * provided by an arch calling add_active_range(). If called for a node
3015 * with no available memory, a warning is printed and the start and end
3016 * PFNs will be 0.
3018 void __meminit get_pfn_range_for_nid(unsigned int nid,
3019 unsigned long *start_pfn, unsigned long *end_pfn)
3021 int i;
3022 *start_pfn = -1UL;
3023 *end_pfn = 0;
3025 for_each_active_range_index_in_nid(i, nid) {
3026 *start_pfn = min(*start_pfn, early_node_map[i].start_pfn);
3027 *end_pfn = max(*end_pfn, early_node_map[i].end_pfn);
3030 if (*start_pfn == -1UL)
3031 *start_pfn = 0;
3033 /* Push the node boundaries out if requested */
3034 account_node_boundary(nid, start_pfn, end_pfn);
3038 * This finds a zone that can be used for ZONE_MOVABLE pages. The
3039 * assumption is made that zones within a node are ordered in monotonic
3040 * increasing memory addresses so that the "highest" populated zone is used
3042 void __init find_usable_zone_for_movable(void)
3044 int zone_index;
3045 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
3046 if (zone_index == ZONE_MOVABLE)
3047 continue;
3049 if (arch_zone_highest_possible_pfn[zone_index] >
3050 arch_zone_lowest_possible_pfn[zone_index])
3051 break;
3054 VM_BUG_ON(zone_index == -1);
3055 movable_zone = zone_index;
3059 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
3060 * because it is sized independant of architecture. Unlike the other zones,
3061 * the starting point for ZONE_MOVABLE is not fixed. It may be different
3062 * in each node depending on the size of each node and how evenly kernelcore
3063 * is distributed. This helper function adjusts the zone ranges
3064 * provided by the architecture for a given node by using the end of the
3065 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
3066 * zones within a node are in order of monotonic increases memory addresses
3068 void __meminit adjust_zone_range_for_zone_movable(int nid,
3069 unsigned long zone_type,
3070 unsigned long node_start_pfn,
3071 unsigned long node_end_pfn,
3072 unsigned long *zone_start_pfn,
3073 unsigned long *zone_end_pfn)
3075 /* Only adjust if ZONE_MOVABLE is on this node */
3076 if (zone_movable_pfn[nid]) {
3077 /* Size ZONE_MOVABLE */
3078 if (zone_type == ZONE_MOVABLE) {
3079 *zone_start_pfn = zone_movable_pfn[nid];
3080 *zone_end_pfn = min(node_end_pfn,
3081 arch_zone_highest_possible_pfn[movable_zone]);
3083 /* Adjust for ZONE_MOVABLE starting within this range */
3084 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
3085 *zone_end_pfn > zone_movable_pfn[nid]) {
3086 *zone_end_pfn = zone_movable_pfn[nid];
3088 /* Check if this whole range is within ZONE_MOVABLE */
3089 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
3090 *zone_start_pfn = *zone_end_pfn;
3095 * Return the number of pages a zone spans in a node, including holes
3096 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
3098 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
3099 unsigned long zone_type,
3100 unsigned long *ignored)
3102 unsigned long node_start_pfn, node_end_pfn;
3103 unsigned long zone_start_pfn, zone_end_pfn;
3105 /* Get the start and end of the node and zone */
3106 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
3107 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
3108 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
3109 adjust_zone_range_for_zone_movable(nid, zone_type,
3110 node_start_pfn, node_end_pfn,
3111 &zone_start_pfn, &zone_end_pfn);
3113 /* Check that this node has pages within the zone's required range */
3114 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
3115 return 0;
3117 /* Move the zone boundaries inside the node if necessary */
3118 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
3119 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
3121 /* Return the spanned pages */
3122 return zone_end_pfn - zone_start_pfn;
3126 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
3127 * then all holes in the requested range will be accounted for.
3129 unsigned long __meminit __absent_pages_in_range(int nid,
3130 unsigned long range_start_pfn,
3131 unsigned long range_end_pfn)
3133 int i = 0;
3134 unsigned long prev_end_pfn = 0, hole_pages = 0;
3135 unsigned long start_pfn;
3137 /* Find the end_pfn of the first active range of pfns in the node */
3138 i = first_active_region_index_in_nid(nid);
3139 if (i == -1)
3140 return 0;
3142 prev_end_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
3144 /* Account for ranges before physical memory on this node */
3145 if (early_node_map[i].start_pfn > range_start_pfn)
3146 hole_pages = prev_end_pfn - range_start_pfn;
3148 /* Find all holes for the zone within the node */
3149 for (; i != -1; i = next_active_region_index_in_nid(i, nid)) {
3151 /* No need to continue if prev_end_pfn is outside the zone */
3152 if (prev_end_pfn >= range_end_pfn)
3153 break;
3155 /* Make sure the end of the zone is not within the hole */
3156 start_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
3157 prev_end_pfn = max(prev_end_pfn, range_start_pfn);
3159 /* Update the hole size cound and move on */
3160 if (start_pfn > range_start_pfn) {
3161 BUG_ON(prev_end_pfn > start_pfn);
3162 hole_pages += start_pfn - prev_end_pfn;
3164 prev_end_pfn = early_node_map[i].end_pfn;
3167 /* Account for ranges past physical memory on this node */
3168 if (range_end_pfn > prev_end_pfn)
3169 hole_pages += range_end_pfn -
3170 max(range_start_pfn, prev_end_pfn);
3172 return hole_pages;
3176 * absent_pages_in_range - Return number of page frames in holes within a range
3177 * @start_pfn: The start PFN to start searching for holes
3178 * @end_pfn: The end PFN to stop searching for holes
3180 * It returns the number of pages frames in memory holes within a range.
3182 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
3183 unsigned long end_pfn)
3185 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
3188 /* Return the number of page frames in holes in a zone on a node */
3189 static unsigned long __meminit zone_absent_pages_in_node(int nid,
3190 unsigned long zone_type,
3191 unsigned long *ignored)
3193 unsigned long node_start_pfn, node_end_pfn;
3194 unsigned long zone_start_pfn, zone_end_pfn;
3196 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
3197 zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type],
3198 node_start_pfn);
3199 zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type],
3200 node_end_pfn);
3202 adjust_zone_range_for_zone_movable(nid, zone_type,
3203 node_start_pfn, node_end_pfn,
3204 &zone_start_pfn, &zone_end_pfn);
3205 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
3208 #else
3209 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
3210 unsigned long zone_type,
3211 unsigned long *zones_size)
3213 return zones_size[zone_type];
3216 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
3217 unsigned long zone_type,
3218 unsigned long *zholes_size)
3220 if (!zholes_size)
3221 return 0;
3223 return zholes_size[zone_type];
3226 #endif
3228 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
3229 unsigned long *zones_size, unsigned long *zholes_size)
3231 unsigned long realtotalpages, totalpages = 0;
3232 enum zone_type i;
3234 for (i = 0; i < MAX_NR_ZONES; i++)
3235 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
3236 zones_size);
3237 pgdat->node_spanned_pages = totalpages;
3239 realtotalpages = totalpages;
3240 for (i = 0; i < MAX_NR_ZONES; i++)
3241 realtotalpages -=
3242 zone_absent_pages_in_node(pgdat->node_id, i,
3243 zholes_size);
3244 pgdat->node_present_pages = realtotalpages;
3245 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
3246 realtotalpages);
3249 #ifndef CONFIG_SPARSEMEM
3251 * Calculate the size of the zone->blockflags rounded to an unsigned long
3252 * Start by making sure zonesize is a multiple of pageblock_order by rounding
3253 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
3254 * round what is now in bits to nearest long in bits, then return it in
3255 * bytes.
3257 static unsigned long __init usemap_size(unsigned long zonesize)
3259 unsigned long usemapsize;
3261 usemapsize = roundup(zonesize, pageblock_nr_pages);
3262 usemapsize = usemapsize >> pageblock_order;
3263 usemapsize *= NR_PAGEBLOCK_BITS;
3264 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
3266 return usemapsize / 8;
3269 static void __init setup_usemap(struct pglist_data *pgdat,
3270 struct zone *zone, unsigned long zonesize)
3272 unsigned long usemapsize = usemap_size(zonesize);
3273 zone->pageblock_flags = NULL;
3274 if (usemapsize) {
3275 zone->pageblock_flags = alloc_bootmem_node(pgdat, usemapsize);
3276 memset(zone->pageblock_flags, 0, usemapsize);
3279 #else
3280 static void inline setup_usemap(struct pglist_data *pgdat,
3281 struct zone *zone, unsigned long zonesize) {}
3282 #endif /* CONFIG_SPARSEMEM */
3284 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
3286 /* Return a sensible default order for the pageblock size. */
3287 static inline int pageblock_default_order(void)
3289 if (HPAGE_SHIFT > PAGE_SHIFT)
3290 return HUGETLB_PAGE_ORDER;
3292 return MAX_ORDER-1;
3295 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
3296 static inline void __init set_pageblock_order(unsigned int order)
3298 /* Check that pageblock_nr_pages has not already been setup */
3299 if (pageblock_order)
3300 return;
3303 * Assume the largest contiguous order of interest is a huge page.
3304 * This value may be variable depending on boot parameters on IA64
3306 pageblock_order = order;
3308 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
3311 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
3312 * and pageblock_default_order() are unused as pageblock_order is set
3313 * at compile-time. See include/linux/pageblock-flags.h for the values of
3314 * pageblock_order based on the kernel config
3316 static inline int pageblock_default_order(unsigned int order)
3318 return MAX_ORDER-1;
3320 #define set_pageblock_order(x) do {} while (0)
3322 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
3325 * Set up the zone data structures:
3326 * - mark all pages reserved
3327 * - mark all memory queues empty
3328 * - clear the memory bitmaps
3330 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
3331 unsigned long *zones_size, unsigned long *zholes_size)
3333 enum zone_type j;
3334 int nid = pgdat->node_id;
3335 unsigned long zone_start_pfn = pgdat->node_start_pfn;
3336 int ret;
3338 pgdat_resize_init(pgdat);
3339 pgdat->nr_zones = 0;
3340 init_waitqueue_head(&pgdat->kswapd_wait);
3341 pgdat->kswapd_max_order = 0;
3343 for (j = 0; j < MAX_NR_ZONES; j++) {
3344 struct zone *zone = pgdat->node_zones + j;
3345 unsigned long size, realsize, memmap_pages;
3347 size = zone_spanned_pages_in_node(nid, j, zones_size);
3348 realsize = size - zone_absent_pages_in_node(nid, j,
3349 zholes_size);
3352 * Adjust realsize so that it accounts for how much memory
3353 * is used by this zone for memmap. This affects the watermark
3354 * and per-cpu initialisations
3356 memmap_pages = (size * sizeof(struct page)) >> PAGE_SHIFT;
3357 if (realsize >= memmap_pages) {
3358 realsize -= memmap_pages;
3359 printk(KERN_DEBUG
3360 " %s zone: %lu pages used for memmap\n",
3361 zone_names[j], memmap_pages);
3362 } else
3363 printk(KERN_WARNING
3364 " %s zone: %lu pages exceeds realsize %lu\n",
3365 zone_names[j], memmap_pages, realsize);
3367 /* Account for reserved pages */
3368 if (j == 0 && realsize > dma_reserve) {
3369 realsize -= dma_reserve;
3370 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
3371 zone_names[0], dma_reserve);
3374 if (!is_highmem_idx(j))
3375 nr_kernel_pages += realsize;
3376 nr_all_pages += realsize;
3378 zone->spanned_pages = size;
3379 zone->present_pages = realsize;
3380 #ifdef CONFIG_NUMA
3381 zone->node = nid;
3382 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
3383 / 100;
3384 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
3385 #endif
3386 zone->name = zone_names[j];
3387 spin_lock_init(&zone->lock);
3388 spin_lock_init(&zone->lru_lock);
3389 zone_seqlock_init(zone);
3390 zone->zone_pgdat = pgdat;
3392 zone->prev_priority = DEF_PRIORITY;
3394 zone_pcp_init(zone);
3395 INIT_LIST_HEAD(&zone->active_list);
3396 INIT_LIST_HEAD(&zone->inactive_list);
3397 zone->nr_scan_active = 0;
3398 zone->nr_scan_inactive = 0;
3399 zap_zone_vm_stats(zone);
3400 zone->flags = 0;
3401 if (!size)
3402 continue;
3404 set_pageblock_order(pageblock_default_order());
3405 setup_usemap(pgdat, zone, size);
3406 ret = init_currently_empty_zone(zone, zone_start_pfn,
3407 size, MEMMAP_EARLY);
3408 BUG_ON(ret);
3409 zone_start_pfn += size;
3413 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
3415 /* Skip empty nodes */
3416 if (!pgdat->node_spanned_pages)
3417 return;
3419 #ifdef CONFIG_FLAT_NODE_MEM_MAP
3420 /* ia64 gets its own node_mem_map, before this, without bootmem */
3421 if (!pgdat->node_mem_map) {
3422 unsigned long size, start, end;
3423 struct page *map;
3426 * The zone's endpoints aren't required to be MAX_ORDER
3427 * aligned but the node_mem_map endpoints must be in order
3428 * for the buddy allocator to function correctly.
3430 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
3431 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
3432 end = ALIGN(end, MAX_ORDER_NR_PAGES);
3433 size = (end - start) * sizeof(struct page);
3434 map = alloc_remap(pgdat->node_id, size);
3435 if (!map)
3436 map = alloc_bootmem_node(pgdat, size);
3437 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
3439 #ifndef CONFIG_NEED_MULTIPLE_NODES
3441 * With no DISCONTIG, the global mem_map is just set as node 0's
3443 if (pgdat == NODE_DATA(0)) {
3444 mem_map = NODE_DATA(0)->node_mem_map;
3445 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3446 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
3447 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
3448 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
3450 #endif
3451 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
3454 void __paginginit free_area_init_node(int nid, struct pglist_data *pgdat,
3455 unsigned long *zones_size, unsigned long node_start_pfn,
3456 unsigned long *zholes_size)
3458 pgdat->node_id = nid;
3459 pgdat->node_start_pfn = node_start_pfn;
3460 calculate_node_totalpages(pgdat, zones_size, zholes_size);
3462 alloc_node_mem_map(pgdat);
3464 free_area_init_core(pgdat, zones_size, zholes_size);
3467 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3469 #if MAX_NUMNODES > 1
3471 * Figure out the number of possible node ids.
3473 static void __init setup_nr_node_ids(void)
3475 unsigned int node;
3476 unsigned int highest = 0;
3478 for_each_node_mask(node, node_possible_map)
3479 highest = node;
3480 nr_node_ids = highest + 1;
3482 #else
3483 static inline void setup_nr_node_ids(void)
3486 #endif
3489 * add_active_range - Register a range of PFNs backed by physical memory
3490 * @nid: The node ID the range resides on
3491 * @start_pfn: The start PFN of the available physical memory
3492 * @end_pfn: The end PFN of the available physical memory
3494 * These ranges are stored in an early_node_map[] and later used by
3495 * free_area_init_nodes() to calculate zone sizes and holes. If the
3496 * range spans a memory hole, it is up to the architecture to ensure
3497 * the memory is not freed by the bootmem allocator. If possible
3498 * the range being registered will be merged with existing ranges.
3500 void __init add_active_range(unsigned int nid, unsigned long start_pfn,
3501 unsigned long end_pfn)
3503 int i;
3505 printk(KERN_DEBUG "Entering add_active_range(%d, %lu, %lu) "
3506 "%d entries of %d used\n",
3507 nid, start_pfn, end_pfn,
3508 nr_nodemap_entries, MAX_ACTIVE_REGIONS);
3510 /* Merge with existing active regions if possible */
3511 for (i = 0; i < nr_nodemap_entries; i++) {
3512 if (early_node_map[i].nid != nid)
3513 continue;
3515 /* Skip if an existing region covers this new one */
3516 if (start_pfn >= early_node_map[i].start_pfn &&
3517 end_pfn <= early_node_map[i].end_pfn)
3518 return;
3520 /* Merge forward if suitable */
3521 if (start_pfn <= early_node_map[i].end_pfn &&
3522 end_pfn > early_node_map[i].end_pfn) {
3523 early_node_map[i].end_pfn = end_pfn;
3524 return;
3527 /* Merge backward if suitable */
3528 if (start_pfn < early_node_map[i].end_pfn &&
3529 end_pfn >= early_node_map[i].start_pfn) {
3530 early_node_map[i].start_pfn = start_pfn;
3531 return;
3535 /* Check that early_node_map is large enough */
3536 if (i >= MAX_ACTIVE_REGIONS) {
3537 printk(KERN_CRIT "More than %d memory regions, truncating\n",
3538 MAX_ACTIVE_REGIONS);
3539 return;
3542 early_node_map[i].nid = nid;
3543 early_node_map[i].start_pfn = start_pfn;
3544 early_node_map[i].end_pfn = end_pfn;
3545 nr_nodemap_entries = i + 1;
3549 * shrink_active_range - Shrink an existing registered range of PFNs
3550 * @nid: The node id the range is on that should be shrunk
3551 * @old_end_pfn: The old end PFN of the range
3552 * @new_end_pfn: The new PFN of the range
3554 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
3555 * The map is kept at the end physical page range that has already been
3556 * registered with add_active_range(). This function allows an arch to shrink
3557 * an existing registered range.
3559 void __init shrink_active_range(unsigned int nid, unsigned long old_end_pfn,
3560 unsigned long new_end_pfn)
3562 int i;
3564 /* Find the old active region end and shrink */
3565 for_each_active_range_index_in_nid(i, nid)
3566 if (early_node_map[i].end_pfn == old_end_pfn) {
3567 early_node_map[i].end_pfn = new_end_pfn;
3568 break;
3573 * remove_all_active_ranges - Remove all currently registered regions
3575 * During discovery, it may be found that a table like SRAT is invalid
3576 * and an alternative discovery method must be used. This function removes
3577 * all currently registered regions.
3579 void __init remove_all_active_ranges(void)
3581 memset(early_node_map, 0, sizeof(early_node_map));
3582 nr_nodemap_entries = 0;
3583 #ifdef CONFIG_MEMORY_HOTPLUG_RESERVE
3584 memset(node_boundary_start_pfn, 0, sizeof(node_boundary_start_pfn));
3585 memset(node_boundary_end_pfn, 0, sizeof(node_boundary_end_pfn));
3586 #endif /* CONFIG_MEMORY_HOTPLUG_RESERVE */
3589 /* Compare two active node_active_regions */
3590 static int __init cmp_node_active_region(const void *a, const void *b)
3592 struct node_active_region *arange = (struct node_active_region *)a;
3593 struct node_active_region *brange = (struct node_active_region *)b;
3595 /* Done this way to avoid overflows */
3596 if (arange->start_pfn > brange->start_pfn)
3597 return 1;
3598 if (arange->start_pfn < brange->start_pfn)
3599 return -1;
3601 return 0;
3604 /* sort the node_map by start_pfn */
3605 static void __init sort_node_map(void)
3607 sort(early_node_map, (size_t)nr_nodemap_entries,
3608 sizeof(struct node_active_region),
3609 cmp_node_active_region, NULL);
3612 /* Find the lowest pfn for a node */
3613 unsigned long __init find_min_pfn_for_node(unsigned long nid)
3615 int i;
3616 unsigned long min_pfn = ULONG_MAX;
3618 /* Assuming a sorted map, the first range found has the starting pfn */
3619 for_each_active_range_index_in_nid(i, nid)
3620 min_pfn = min(min_pfn, early_node_map[i].start_pfn);
3622 if (min_pfn == ULONG_MAX) {
3623 printk(KERN_WARNING
3624 "Could not find start_pfn for node %lu\n", nid);
3625 return 0;
3628 return min_pfn;
3632 * find_min_pfn_with_active_regions - Find the minimum PFN registered
3634 * It returns the minimum PFN based on information provided via
3635 * add_active_range().
3637 unsigned long __init find_min_pfn_with_active_regions(void)
3639 return find_min_pfn_for_node(MAX_NUMNODES);
3643 * find_max_pfn_with_active_regions - Find the maximum PFN registered
3645 * It returns the maximum PFN based on information provided via
3646 * add_active_range().
3648 unsigned long __init find_max_pfn_with_active_regions(void)
3650 int i;
3651 unsigned long max_pfn = 0;
3653 for (i = 0; i < nr_nodemap_entries; i++)
3654 max_pfn = max(max_pfn, early_node_map[i].end_pfn);
3656 return max_pfn;
3660 * early_calculate_totalpages()
3661 * Sum pages in active regions for movable zone.
3662 * Populate N_HIGH_MEMORY for calculating usable_nodes.
3664 static unsigned long __init early_calculate_totalpages(void)
3666 int i;
3667 unsigned long totalpages = 0;
3669 for (i = 0; i < nr_nodemap_entries; i++) {
3670 unsigned long pages = early_node_map[i].end_pfn -
3671 early_node_map[i].start_pfn;
3672 totalpages += pages;
3673 if (pages)
3674 node_set_state(early_node_map[i].nid, N_HIGH_MEMORY);
3676 return totalpages;
3680 * Find the PFN the Movable zone begins in each node. Kernel memory
3681 * is spread evenly between nodes as long as the nodes have enough
3682 * memory. When they don't, some nodes will have more kernelcore than
3683 * others
3685 void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn)
3687 int i, nid;
3688 unsigned long usable_startpfn;
3689 unsigned long kernelcore_node, kernelcore_remaining;
3690 unsigned long totalpages = early_calculate_totalpages();
3691 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
3694 * If movablecore was specified, calculate what size of
3695 * kernelcore that corresponds so that memory usable for
3696 * any allocation type is evenly spread. If both kernelcore
3697 * and movablecore are specified, then the value of kernelcore
3698 * will be used for required_kernelcore if it's greater than
3699 * what movablecore would have allowed.
3701 if (required_movablecore) {
3702 unsigned long corepages;
3705 * Round-up so that ZONE_MOVABLE is at least as large as what
3706 * was requested by the user
3708 required_movablecore =
3709 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
3710 corepages = totalpages - required_movablecore;
3712 required_kernelcore = max(required_kernelcore, corepages);
3715 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
3716 if (!required_kernelcore)
3717 return;
3719 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
3720 find_usable_zone_for_movable();
3721 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
3723 restart:
3724 /* Spread kernelcore memory as evenly as possible throughout nodes */
3725 kernelcore_node = required_kernelcore / usable_nodes;
3726 for_each_node_state(nid, N_HIGH_MEMORY) {
3728 * Recalculate kernelcore_node if the division per node
3729 * now exceeds what is necessary to satisfy the requested
3730 * amount of memory for the kernel
3732 if (required_kernelcore < kernelcore_node)
3733 kernelcore_node = required_kernelcore / usable_nodes;
3736 * As the map is walked, we track how much memory is usable
3737 * by the kernel using kernelcore_remaining. When it is
3738 * 0, the rest of the node is usable by ZONE_MOVABLE
3740 kernelcore_remaining = kernelcore_node;
3742 /* Go through each range of PFNs within this node */
3743 for_each_active_range_index_in_nid(i, nid) {
3744 unsigned long start_pfn, end_pfn;
3745 unsigned long size_pages;
3747 start_pfn = max(early_node_map[i].start_pfn,
3748 zone_movable_pfn[nid]);
3749 end_pfn = early_node_map[i].end_pfn;
3750 if (start_pfn >= end_pfn)
3751 continue;
3753 /* Account for what is only usable for kernelcore */
3754 if (start_pfn < usable_startpfn) {
3755 unsigned long kernel_pages;
3756 kernel_pages = min(end_pfn, usable_startpfn)
3757 - start_pfn;
3759 kernelcore_remaining -= min(kernel_pages,
3760 kernelcore_remaining);
3761 required_kernelcore -= min(kernel_pages,
3762 required_kernelcore);
3764 /* Continue if range is now fully accounted */
3765 if (end_pfn <= usable_startpfn) {
3768 * Push zone_movable_pfn to the end so
3769 * that if we have to rebalance
3770 * kernelcore across nodes, we will
3771 * not double account here
3773 zone_movable_pfn[nid] = end_pfn;
3774 continue;
3776 start_pfn = usable_startpfn;
3780 * The usable PFN range for ZONE_MOVABLE is from
3781 * start_pfn->end_pfn. Calculate size_pages as the
3782 * number of pages used as kernelcore
3784 size_pages = end_pfn - start_pfn;
3785 if (size_pages > kernelcore_remaining)
3786 size_pages = kernelcore_remaining;
3787 zone_movable_pfn[nid] = start_pfn + size_pages;
3790 * Some kernelcore has been met, update counts and
3791 * break if the kernelcore for this node has been
3792 * satisified
3794 required_kernelcore -= min(required_kernelcore,
3795 size_pages);
3796 kernelcore_remaining -= size_pages;
3797 if (!kernelcore_remaining)
3798 break;
3803 * If there is still required_kernelcore, we do another pass with one
3804 * less node in the count. This will push zone_movable_pfn[nid] further
3805 * along on the nodes that still have memory until kernelcore is
3806 * satisified
3808 usable_nodes--;
3809 if (usable_nodes && required_kernelcore > usable_nodes)
3810 goto restart;
3812 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
3813 for (nid = 0; nid < MAX_NUMNODES; nid++)
3814 zone_movable_pfn[nid] =
3815 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
3818 /* Any regular memory on that node ? */
3819 static void check_for_regular_memory(pg_data_t *pgdat)
3821 #ifdef CONFIG_HIGHMEM
3822 enum zone_type zone_type;
3824 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) {
3825 struct zone *zone = &pgdat->node_zones[zone_type];
3826 if (zone->present_pages)
3827 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY);
3829 #endif
3833 * free_area_init_nodes - Initialise all pg_data_t and zone data
3834 * @max_zone_pfn: an array of max PFNs for each zone
3836 * This will call free_area_init_node() for each active node in the system.
3837 * Using the page ranges provided by add_active_range(), the size of each
3838 * zone in each node and their holes is calculated. If the maximum PFN
3839 * between two adjacent zones match, it is assumed that the zone is empty.
3840 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
3841 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
3842 * starts where the previous one ended. For example, ZONE_DMA32 starts
3843 * at arch_max_dma_pfn.
3845 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
3847 unsigned long nid;
3848 enum zone_type i;
3850 /* Sort early_node_map as initialisation assumes it is sorted */
3851 sort_node_map();
3853 /* Record where the zone boundaries are */
3854 memset(arch_zone_lowest_possible_pfn, 0,
3855 sizeof(arch_zone_lowest_possible_pfn));
3856 memset(arch_zone_highest_possible_pfn, 0,
3857 sizeof(arch_zone_highest_possible_pfn));
3858 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
3859 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
3860 for (i = 1; i < MAX_NR_ZONES; i++) {
3861 if (i == ZONE_MOVABLE)
3862 continue;
3863 arch_zone_lowest_possible_pfn[i] =
3864 arch_zone_highest_possible_pfn[i-1];
3865 arch_zone_highest_possible_pfn[i] =
3866 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
3868 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
3869 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
3871 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
3872 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
3873 find_zone_movable_pfns_for_nodes(zone_movable_pfn);
3875 /* Print out the zone ranges */
3876 printk("Zone PFN ranges:\n");
3877 for (i = 0; i < MAX_NR_ZONES; i++) {
3878 if (i == ZONE_MOVABLE)
3879 continue;
3880 printk(" %-8s %8lu -> %8lu\n",
3881 zone_names[i],
3882 arch_zone_lowest_possible_pfn[i],
3883 arch_zone_highest_possible_pfn[i]);
3886 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
3887 printk("Movable zone start PFN for each node\n");
3888 for (i = 0; i < MAX_NUMNODES; i++) {
3889 if (zone_movable_pfn[i])
3890 printk(" Node %d: %lu\n", i, zone_movable_pfn[i]);
3893 /* Print out the early_node_map[] */
3894 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries);
3895 for (i = 0; i < nr_nodemap_entries; i++)
3896 printk(" %3d: %8lu -> %8lu\n", early_node_map[i].nid,
3897 early_node_map[i].start_pfn,
3898 early_node_map[i].end_pfn);
3900 /* Initialise every node */
3901 setup_nr_node_ids();
3902 for_each_online_node(nid) {
3903 pg_data_t *pgdat = NODE_DATA(nid);
3904 free_area_init_node(nid, pgdat, NULL,
3905 find_min_pfn_for_node(nid), NULL);
3907 /* Any memory on that node */
3908 if (pgdat->node_present_pages)
3909 node_set_state(nid, N_HIGH_MEMORY);
3910 check_for_regular_memory(pgdat);
3914 static int __init cmdline_parse_core(char *p, unsigned long *core)
3916 unsigned long long coremem;
3917 if (!p)
3918 return -EINVAL;
3920 coremem = memparse(p, &p);
3921 *core = coremem >> PAGE_SHIFT;
3923 /* Paranoid check that UL is enough for the coremem value */
3924 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
3926 return 0;
3930 * kernelcore=size sets the amount of memory for use for allocations that
3931 * cannot be reclaimed or migrated.
3933 static int __init cmdline_parse_kernelcore(char *p)
3935 return cmdline_parse_core(p, &required_kernelcore);
3939 * movablecore=size sets the amount of memory for use for allocations that
3940 * can be reclaimed or migrated.
3942 static int __init cmdline_parse_movablecore(char *p)
3944 return cmdline_parse_core(p, &required_movablecore);
3947 early_param("kernelcore", cmdline_parse_kernelcore);
3948 early_param("movablecore", cmdline_parse_movablecore);
3950 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
3953 * set_dma_reserve - set the specified number of pages reserved in the first zone
3954 * @new_dma_reserve: The number of pages to mark reserved
3956 * The per-cpu batchsize and zone watermarks are determined by present_pages.
3957 * In the DMA zone, a significant percentage may be consumed by kernel image
3958 * and other unfreeable allocations which can skew the watermarks badly. This
3959 * function may optionally be used to account for unfreeable pages in the
3960 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
3961 * smaller per-cpu batchsize.
3963 void __init set_dma_reserve(unsigned long new_dma_reserve)
3965 dma_reserve = new_dma_reserve;
3968 #ifndef CONFIG_NEED_MULTIPLE_NODES
3969 static bootmem_data_t contig_bootmem_data;
3970 struct pglist_data contig_page_data = { .bdata = &contig_bootmem_data };
3972 EXPORT_SYMBOL(contig_page_data);
3973 #endif
3975 void __init free_area_init(unsigned long *zones_size)
3977 free_area_init_node(0, NODE_DATA(0), zones_size,
3978 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
3981 static int page_alloc_cpu_notify(struct notifier_block *self,
3982 unsigned long action, void *hcpu)
3984 int cpu = (unsigned long)hcpu;
3986 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
3987 drain_pages(cpu);
3990 * Spill the event counters of the dead processor
3991 * into the current processors event counters.
3992 * This artificially elevates the count of the current
3993 * processor.
3995 vm_events_fold_cpu(cpu);
3998 * Zero the differential counters of the dead processor
3999 * so that the vm statistics are consistent.
4001 * This is only okay since the processor is dead and cannot
4002 * race with what we are doing.
4004 refresh_cpu_vm_stats(cpu);
4006 return NOTIFY_OK;
4009 void __init page_alloc_init(void)
4011 hotcpu_notifier(page_alloc_cpu_notify, 0);
4015 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
4016 * or min_free_kbytes changes.
4018 static void calculate_totalreserve_pages(void)
4020 struct pglist_data *pgdat;
4021 unsigned long reserve_pages = 0;
4022 enum zone_type i, j;
4024 for_each_online_pgdat(pgdat) {
4025 for (i = 0; i < MAX_NR_ZONES; i++) {
4026 struct zone *zone = pgdat->node_zones + i;
4027 unsigned long max = 0;
4029 /* Find valid and maximum lowmem_reserve in the zone */
4030 for (j = i; j < MAX_NR_ZONES; j++) {
4031 if (zone->lowmem_reserve[j] > max)
4032 max = zone->lowmem_reserve[j];
4035 /* we treat pages_high as reserved pages. */
4036 max += zone->pages_high;
4038 if (max > zone->present_pages)
4039 max = zone->present_pages;
4040 reserve_pages += max;
4043 totalreserve_pages = reserve_pages;
4047 * setup_per_zone_lowmem_reserve - called whenever
4048 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
4049 * has a correct pages reserved value, so an adequate number of
4050 * pages are left in the zone after a successful __alloc_pages().
4052 static void setup_per_zone_lowmem_reserve(void)
4054 struct pglist_data *pgdat;
4055 enum zone_type j, idx;
4057 for_each_online_pgdat(pgdat) {
4058 for (j = 0; j < MAX_NR_ZONES; j++) {
4059 struct zone *zone = pgdat->node_zones + j;
4060 unsigned long present_pages = zone->present_pages;
4062 zone->lowmem_reserve[j] = 0;
4064 idx = j;
4065 while (idx) {
4066 struct zone *lower_zone;
4068 idx--;
4070 if (sysctl_lowmem_reserve_ratio[idx] < 1)
4071 sysctl_lowmem_reserve_ratio[idx] = 1;
4073 lower_zone = pgdat->node_zones + idx;
4074 lower_zone->lowmem_reserve[j] = present_pages /
4075 sysctl_lowmem_reserve_ratio[idx];
4076 present_pages += lower_zone->present_pages;
4081 /* update totalreserve_pages */
4082 calculate_totalreserve_pages();
4086 * setup_per_zone_pages_min - called when min_free_kbytes changes.
4088 * Ensures that the pages_{min,low,high} values for each zone are set correctly
4089 * with respect to min_free_kbytes.
4091 void setup_per_zone_pages_min(void)
4093 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
4094 unsigned long lowmem_pages = 0;
4095 struct zone *zone;
4096 unsigned long flags;
4098 /* Calculate total number of !ZONE_HIGHMEM pages */
4099 for_each_zone(zone) {
4100 if (!is_highmem(zone))
4101 lowmem_pages += zone->present_pages;
4104 for_each_zone(zone) {
4105 u64 tmp;
4107 spin_lock_irqsave(&zone->lru_lock, flags);
4108 tmp = (u64)pages_min * zone->present_pages;
4109 do_div(tmp, lowmem_pages);
4110 if (is_highmem(zone)) {
4112 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
4113 * need highmem pages, so cap pages_min to a small
4114 * value here.
4116 * The (pages_high-pages_low) and (pages_low-pages_min)
4117 * deltas controls asynch page reclaim, and so should
4118 * not be capped for highmem.
4120 int min_pages;
4122 min_pages = zone->present_pages / 1024;
4123 if (min_pages < SWAP_CLUSTER_MAX)
4124 min_pages = SWAP_CLUSTER_MAX;
4125 if (min_pages > 128)
4126 min_pages = 128;
4127 zone->pages_min = min_pages;
4128 } else {
4130 * If it's a lowmem zone, reserve a number of pages
4131 * proportionate to the zone's size.
4133 zone->pages_min = tmp;
4136 zone->pages_low = zone->pages_min + (tmp >> 2);
4137 zone->pages_high = zone->pages_min + (tmp >> 1);
4138 setup_zone_migrate_reserve(zone);
4139 spin_unlock_irqrestore(&zone->lru_lock, flags);
4142 /* update totalreserve_pages */
4143 calculate_totalreserve_pages();
4147 * Initialise min_free_kbytes.
4149 * For small machines we want it small (128k min). For large machines
4150 * we want it large (64MB max). But it is not linear, because network
4151 * bandwidth does not increase linearly with machine size. We use
4153 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
4154 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
4156 * which yields
4158 * 16MB: 512k
4159 * 32MB: 724k
4160 * 64MB: 1024k
4161 * 128MB: 1448k
4162 * 256MB: 2048k
4163 * 512MB: 2896k
4164 * 1024MB: 4096k
4165 * 2048MB: 5792k
4166 * 4096MB: 8192k
4167 * 8192MB: 11584k
4168 * 16384MB: 16384k
4170 static int __init init_per_zone_pages_min(void)
4172 unsigned long lowmem_kbytes;
4174 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
4176 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
4177 if (min_free_kbytes < 128)
4178 min_free_kbytes = 128;
4179 if (min_free_kbytes > 65536)
4180 min_free_kbytes = 65536;
4181 setup_per_zone_pages_min();
4182 setup_per_zone_lowmem_reserve();
4183 return 0;
4185 module_init(init_per_zone_pages_min)
4188 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
4189 * that we can call two helper functions whenever min_free_kbytes
4190 * changes.
4192 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
4193 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4195 proc_dointvec(table, write, file, buffer, length, ppos);
4196 if (write)
4197 setup_per_zone_pages_min();
4198 return 0;
4201 #ifdef CONFIG_NUMA
4202 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
4203 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4205 struct zone *zone;
4206 int rc;
4208 rc = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
4209 if (rc)
4210 return rc;
4212 for_each_zone(zone)
4213 zone->min_unmapped_pages = (zone->present_pages *
4214 sysctl_min_unmapped_ratio) / 100;
4215 return 0;
4218 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
4219 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4221 struct zone *zone;
4222 int rc;
4224 rc = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
4225 if (rc)
4226 return rc;
4228 for_each_zone(zone)
4229 zone->min_slab_pages = (zone->present_pages *
4230 sysctl_min_slab_ratio) / 100;
4231 return 0;
4233 #endif
4236 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
4237 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
4238 * whenever sysctl_lowmem_reserve_ratio changes.
4240 * The reserve ratio obviously has absolutely no relation with the
4241 * pages_min watermarks. The lowmem reserve ratio can only make sense
4242 * if in function of the boot time zone sizes.
4244 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
4245 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4247 proc_dointvec_minmax(table, write, file, buffer, length, ppos);
4248 setup_per_zone_lowmem_reserve();
4249 return 0;
4253 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
4254 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
4255 * can have before it gets flushed back to buddy allocator.
4258 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
4259 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4261 struct zone *zone;
4262 unsigned int cpu;
4263 int ret;
4265 ret = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
4266 if (!write || (ret == -EINVAL))
4267 return ret;
4268 for_each_zone(zone) {
4269 for_each_online_cpu(cpu) {
4270 unsigned long high;
4271 high = zone->present_pages / percpu_pagelist_fraction;
4272 setup_pagelist_highmark(zone_pcp(zone, cpu), high);
4275 return 0;
4278 int hashdist = HASHDIST_DEFAULT;
4280 #ifdef CONFIG_NUMA
4281 static int __init set_hashdist(char *str)
4283 if (!str)
4284 return 0;
4285 hashdist = simple_strtoul(str, &str, 0);
4286 return 1;
4288 __setup("hashdist=", set_hashdist);
4289 #endif
4292 * allocate a large system hash table from bootmem
4293 * - it is assumed that the hash table must contain an exact power-of-2
4294 * quantity of entries
4295 * - limit is the number of hash buckets, not the total allocation size
4297 void *__init alloc_large_system_hash(const char *tablename,
4298 unsigned long bucketsize,
4299 unsigned long numentries,
4300 int scale,
4301 int flags,
4302 unsigned int *_hash_shift,
4303 unsigned int *_hash_mask,
4304 unsigned long limit)
4306 unsigned long long max = limit;
4307 unsigned long log2qty, size;
4308 void *table = NULL;
4310 /* allow the kernel cmdline to have a say */
4311 if (!numentries) {
4312 /* round applicable memory size up to nearest megabyte */
4313 numentries = nr_kernel_pages;
4314 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
4315 numentries >>= 20 - PAGE_SHIFT;
4316 numentries <<= 20 - PAGE_SHIFT;
4318 /* limit to 1 bucket per 2^scale bytes of low memory */
4319 if (scale > PAGE_SHIFT)
4320 numentries >>= (scale - PAGE_SHIFT);
4321 else
4322 numentries <<= (PAGE_SHIFT - scale);
4324 /* Make sure we've got at least a 0-order allocation.. */
4325 if (unlikely((numentries * bucketsize) < PAGE_SIZE))
4326 numentries = PAGE_SIZE / bucketsize;
4328 numentries = roundup_pow_of_two(numentries);
4330 /* limit allocation size to 1/16 total memory by default */
4331 if (max == 0) {
4332 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
4333 do_div(max, bucketsize);
4336 if (numentries > max)
4337 numentries = max;
4339 log2qty = ilog2(numentries);
4341 do {
4342 size = bucketsize << log2qty;
4343 if (flags & HASH_EARLY)
4344 table = alloc_bootmem(size);
4345 else if (hashdist)
4346 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
4347 else {
4348 unsigned long order;
4349 for (order = 0; ((1UL << order) << PAGE_SHIFT) < size; order++)
4351 table = (void*) __get_free_pages(GFP_ATOMIC, order);
4353 * If bucketsize is not a power-of-two, we may free
4354 * some pages at the end of hash table.
4356 if (table) {
4357 unsigned long alloc_end = (unsigned long)table +
4358 (PAGE_SIZE << order);
4359 unsigned long used = (unsigned long)table +
4360 PAGE_ALIGN(size);
4361 split_page(virt_to_page(table), order);
4362 while (used < alloc_end) {
4363 free_page(used);
4364 used += PAGE_SIZE;
4368 } while (!table && size > PAGE_SIZE && --log2qty);
4370 if (!table)
4371 panic("Failed to allocate %s hash table\n", tablename);
4373 printk(KERN_INFO "%s hash table entries: %d (order: %d, %lu bytes)\n",
4374 tablename,
4375 (1U << log2qty),
4376 ilog2(size) - PAGE_SHIFT,
4377 size);
4379 if (_hash_shift)
4380 *_hash_shift = log2qty;
4381 if (_hash_mask)
4382 *_hash_mask = (1 << log2qty) - 1;
4384 return table;
4387 #ifdef CONFIG_OUT_OF_LINE_PFN_TO_PAGE
4388 struct page *pfn_to_page(unsigned long pfn)
4390 return __pfn_to_page(pfn);
4392 unsigned long page_to_pfn(struct page *page)
4394 return __page_to_pfn(page);
4396 EXPORT_SYMBOL(pfn_to_page);
4397 EXPORT_SYMBOL(page_to_pfn);
4398 #endif /* CONFIG_OUT_OF_LINE_PFN_TO_PAGE */
4400 /* Return a pointer to the bitmap storing bits affecting a block of pages */
4401 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
4402 unsigned long pfn)
4404 #ifdef CONFIG_SPARSEMEM
4405 return __pfn_to_section(pfn)->pageblock_flags;
4406 #else
4407 return zone->pageblock_flags;
4408 #endif /* CONFIG_SPARSEMEM */
4411 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
4413 #ifdef CONFIG_SPARSEMEM
4414 pfn &= (PAGES_PER_SECTION-1);
4415 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
4416 #else
4417 pfn = pfn - zone->zone_start_pfn;
4418 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
4419 #endif /* CONFIG_SPARSEMEM */
4423 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
4424 * @page: The page within the block of interest
4425 * @start_bitidx: The first bit of interest to retrieve
4426 * @end_bitidx: The last bit of interest
4427 * returns pageblock_bits flags
4429 unsigned long get_pageblock_flags_group(struct page *page,
4430 int start_bitidx, int end_bitidx)
4432 struct zone *zone;
4433 unsigned long *bitmap;
4434 unsigned long pfn, bitidx;
4435 unsigned long flags = 0;
4436 unsigned long value = 1;
4438 zone = page_zone(page);
4439 pfn = page_to_pfn(page);
4440 bitmap = get_pageblock_bitmap(zone, pfn);
4441 bitidx = pfn_to_bitidx(zone, pfn);
4443 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
4444 if (test_bit(bitidx + start_bitidx, bitmap))
4445 flags |= value;
4447 return flags;
4451 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
4452 * @page: The page within the block of interest
4453 * @start_bitidx: The first bit of interest
4454 * @end_bitidx: The last bit of interest
4455 * @flags: The flags to set
4457 void set_pageblock_flags_group(struct page *page, unsigned long flags,
4458 int start_bitidx, int end_bitidx)
4460 struct zone *zone;
4461 unsigned long *bitmap;
4462 unsigned long pfn, bitidx;
4463 unsigned long value = 1;
4465 zone = page_zone(page);
4466 pfn = page_to_pfn(page);
4467 bitmap = get_pageblock_bitmap(zone, pfn);
4468 bitidx = pfn_to_bitidx(zone, pfn);
4470 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
4471 if (flags & value)
4472 __set_bit(bitidx + start_bitidx, bitmap);
4473 else
4474 __clear_bit(bitidx + start_bitidx, bitmap);
4478 * This is designed as sub function...plz see page_isolation.c also.
4479 * set/clear page block's type to be ISOLATE.
4480 * page allocater never alloc memory from ISOLATE block.
4483 int set_migratetype_isolate(struct page *page)
4485 struct zone *zone;
4486 unsigned long flags;
4487 int ret = -EBUSY;
4489 zone = page_zone(page);
4490 spin_lock_irqsave(&zone->lock, flags);
4492 * In future, more migrate types will be able to be isolation target.
4494 if (get_pageblock_migratetype(page) != MIGRATE_MOVABLE)
4495 goto out;
4496 set_pageblock_migratetype(page, MIGRATE_ISOLATE);
4497 move_freepages_block(zone, page, MIGRATE_ISOLATE);
4498 ret = 0;
4499 out:
4500 spin_unlock_irqrestore(&zone->lock, flags);
4501 if (!ret)
4502 drain_all_pages();
4503 return ret;
4506 void unset_migratetype_isolate(struct page *page)
4508 struct zone *zone;
4509 unsigned long flags;
4510 zone = page_zone(page);
4511 spin_lock_irqsave(&zone->lock, flags);
4512 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
4513 goto out;
4514 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4515 move_freepages_block(zone, page, MIGRATE_MOVABLE);
4516 out:
4517 spin_unlock_irqrestore(&zone->lock, flags);
4520 #ifdef CONFIG_MEMORY_HOTREMOVE
4522 * All pages in the range must be isolated before calling this.
4524 void
4525 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
4527 struct page *page;
4528 struct zone *zone;
4529 int order, i;
4530 unsigned long pfn;
4531 unsigned long flags;
4532 /* find the first valid pfn */
4533 for (pfn = start_pfn; pfn < end_pfn; pfn++)
4534 if (pfn_valid(pfn))
4535 break;
4536 if (pfn == end_pfn)
4537 return;
4538 zone = page_zone(pfn_to_page(pfn));
4539 spin_lock_irqsave(&zone->lock, flags);
4540 pfn = start_pfn;
4541 while (pfn < end_pfn) {
4542 if (!pfn_valid(pfn)) {
4543 pfn++;
4544 continue;
4546 page = pfn_to_page(pfn);
4547 BUG_ON(page_count(page));
4548 BUG_ON(!PageBuddy(page));
4549 order = page_order(page);
4550 #ifdef CONFIG_DEBUG_VM
4551 printk(KERN_INFO "remove from free list %lx %d %lx\n",
4552 pfn, 1 << order, end_pfn);
4553 #endif
4554 list_del(&page->lru);
4555 rmv_page_order(page);
4556 zone->free_area[order].nr_free--;
4557 __mod_zone_page_state(zone, NR_FREE_PAGES,
4558 - (1UL << order));
4559 for (i = 0; i < (1 << order); i++)
4560 SetPageReserved((page+i));
4561 pfn += (1 << order);
4563 spin_unlock_irqrestore(&zone->lock, flags);
4565 #endif