page_alloc: fix kernel-doc warning
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / mm / page_alloc.c
blob91ae36c303306d8aa5de4e1b83cb10cfcee5d280
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/kmemcheck.h>
27 #include <linux/module.h>
28 #include <linux/suspend.h>
29 #include <linux/pagevec.h>
30 #include <linux/blkdev.h>
31 #include <linux/slab.h>
32 #include <linux/oom.h>
33 #include <linux/notifier.h>
34 #include <linux/topology.h>
35 #include <linux/sysctl.h>
36 #include <linux/cpu.h>
37 #include <linux/cpuset.h>
38 #include <linux/memory_hotplug.h>
39 #include <linux/nodemask.h>
40 #include <linux/vmalloc.h>
41 #include <linux/mempolicy.h>
42 #include <linux/stop_machine.h>
43 #include <linux/sort.h>
44 #include <linux/pfn.h>
45 #include <linux/backing-dev.h>
46 #include <linux/fault-inject.h>
47 #include <linux/page-isolation.h>
48 #include <linux/page_cgroup.h>
49 #include <linux/debugobjects.h>
50 #include <linux/kmemleak.h>
52 #include <asm/tlbflush.h>
53 #include <asm/div64.h>
54 #include "internal.h"
57 * Array of node states.
59 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
60 [N_POSSIBLE] = NODE_MASK_ALL,
61 [N_ONLINE] = { { [0] = 1UL } },
62 #ifndef CONFIG_NUMA
63 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
64 #ifdef CONFIG_HIGHMEM
65 [N_HIGH_MEMORY] = { { [0] = 1UL } },
66 #endif
67 [N_CPU] = { { [0] = 1UL } },
68 #endif /* NUMA */
70 EXPORT_SYMBOL(node_states);
72 unsigned long totalram_pages __read_mostly;
73 unsigned long totalreserve_pages __read_mostly;
74 unsigned long highest_memmap_pfn __read_mostly;
75 int percpu_pagelist_fraction;
76 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
78 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
79 int pageblock_order __read_mostly;
80 #endif
82 static void __free_pages_ok(struct page *page, unsigned int order);
85 * results with 256, 32 in the lowmem_reserve sysctl:
86 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
87 * 1G machine -> (16M dma, 784M normal, 224M high)
88 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
89 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
90 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
92 * TBD: should special case ZONE_DMA32 machines here - in those we normally
93 * don't need any ZONE_NORMAL reservation
95 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
96 #ifdef CONFIG_ZONE_DMA
97 256,
98 #endif
99 #ifdef CONFIG_ZONE_DMA32
100 256,
101 #endif
102 #ifdef CONFIG_HIGHMEM
104 #endif
108 EXPORT_SYMBOL(totalram_pages);
110 static char * const zone_names[MAX_NR_ZONES] = {
111 #ifdef CONFIG_ZONE_DMA
112 "DMA",
113 #endif
114 #ifdef CONFIG_ZONE_DMA32
115 "DMA32",
116 #endif
117 "Normal",
118 #ifdef CONFIG_HIGHMEM
119 "HighMem",
120 #endif
121 "Movable",
124 int min_free_kbytes = 1024;
126 unsigned long __meminitdata nr_kernel_pages;
127 unsigned long __meminitdata nr_all_pages;
128 static unsigned long __meminitdata dma_reserve;
130 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
132 * MAX_ACTIVE_REGIONS determines the maximum number of distinct
133 * ranges of memory (RAM) that may be registered with add_active_range().
134 * Ranges passed to add_active_range() will be merged if possible
135 * so the number of times add_active_range() can be called is
136 * related to the number of nodes and the number of holes
138 #ifdef CONFIG_MAX_ACTIVE_REGIONS
139 /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */
140 #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS
141 #else
142 #if MAX_NUMNODES >= 32
143 /* If there can be many nodes, allow up to 50 holes per node */
144 #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50)
145 #else
146 /* By default, allow up to 256 distinct regions */
147 #define MAX_ACTIVE_REGIONS 256
148 #endif
149 #endif
151 static struct node_active_region __meminitdata early_node_map[MAX_ACTIVE_REGIONS];
152 static int __meminitdata nr_nodemap_entries;
153 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
154 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
155 static unsigned long __initdata required_kernelcore;
156 static unsigned long __initdata required_movablecore;
157 static 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 int nr_online_nodes __read_mostly = 1;
167 EXPORT_SYMBOL(nr_node_ids);
168 EXPORT_SYMBOL(nr_online_nodes);
169 #endif
171 int page_group_by_mobility_disabled __read_mostly;
173 static void set_pageblock_migratetype(struct page *page, int migratetype)
176 if (unlikely(page_group_by_mobility_disabled))
177 migratetype = MIGRATE_UNMOVABLE;
179 set_pageblock_flags_group(page, (unsigned long)migratetype,
180 PB_migrate, PB_migrate_end);
183 bool oom_killer_disabled __read_mostly;
185 #ifdef CONFIG_DEBUG_VM
186 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
188 int ret = 0;
189 unsigned seq;
190 unsigned long pfn = page_to_pfn(page);
192 do {
193 seq = zone_span_seqbegin(zone);
194 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
195 ret = 1;
196 else if (pfn < zone->zone_start_pfn)
197 ret = 1;
198 } while (zone_span_seqretry(zone, seq));
200 return ret;
203 static int page_is_consistent(struct zone *zone, struct page *page)
205 if (!pfn_valid_within(page_to_pfn(page)))
206 return 0;
207 if (zone != page_zone(page))
208 return 0;
210 return 1;
213 * Temporary debugging check for pages not lying within a given zone.
215 static int bad_range(struct zone *zone, struct page *page)
217 if (page_outside_zone_boundaries(zone, page))
218 return 1;
219 if (!page_is_consistent(zone, page))
220 return 1;
222 return 0;
224 #else
225 static inline int bad_range(struct zone *zone, struct page *page)
227 return 0;
229 #endif
231 static void bad_page(struct page *page)
233 static unsigned long resume;
234 static unsigned long nr_shown;
235 static unsigned long nr_unshown;
238 * Allow a burst of 60 reports, then keep quiet for that minute;
239 * or allow a steady drip of one report per second.
241 if (nr_shown == 60) {
242 if (time_before(jiffies, resume)) {
243 nr_unshown++;
244 goto out;
246 if (nr_unshown) {
247 printk(KERN_ALERT
248 "BUG: Bad page state: %lu messages suppressed\n",
249 nr_unshown);
250 nr_unshown = 0;
252 nr_shown = 0;
254 if (nr_shown++ == 0)
255 resume = jiffies + 60 * HZ;
257 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
258 current->comm, page_to_pfn(page));
259 printk(KERN_ALERT
260 "page:%p flags:%p count:%d mapcount:%d mapping:%p index:%lx\n",
261 page, (void *)page->flags, page_count(page),
262 page_mapcount(page), page->mapping, page->index);
264 dump_stack();
265 out:
266 /* Leave bad fields for debug, except PageBuddy could make trouble */
267 __ClearPageBuddy(page);
268 add_taint(TAINT_BAD_PAGE);
272 * Higher-order pages are called "compound pages". They are structured thusly:
274 * The first PAGE_SIZE page is called the "head page".
276 * The remaining PAGE_SIZE pages are called "tail pages".
278 * All pages have PG_compound set. All pages have their ->private pointing at
279 * the head page (even the head page has this).
281 * The first tail page's ->lru.next holds the address of the compound page's
282 * put_page() function. Its ->lru.prev holds the order of allocation.
283 * This usage means that zero-order pages may not be compound.
286 static void free_compound_page(struct page *page)
288 __free_pages_ok(page, compound_order(page));
291 void prep_compound_page(struct page *page, unsigned long order)
293 int i;
294 int nr_pages = 1 << order;
296 set_compound_page_dtor(page, free_compound_page);
297 set_compound_order(page, order);
298 __SetPageHead(page);
299 for (i = 1; i < nr_pages; i++) {
300 struct page *p = page + i;
302 __SetPageTail(p);
303 p->first_page = page;
307 static int destroy_compound_page(struct page *page, unsigned long order)
309 int i;
310 int nr_pages = 1 << order;
311 int bad = 0;
313 if (unlikely(compound_order(page) != order) ||
314 unlikely(!PageHead(page))) {
315 bad_page(page);
316 bad++;
319 __ClearPageHead(page);
321 for (i = 1; i < nr_pages; i++) {
322 struct page *p = page + i;
324 if (unlikely(!PageTail(p) || (p->first_page != page))) {
325 bad_page(page);
326 bad++;
328 __ClearPageTail(p);
331 return bad;
334 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
336 int i;
339 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
340 * and __GFP_HIGHMEM from hard or soft interrupt context.
342 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
343 for (i = 0; i < (1 << order); i++)
344 clear_highpage(page + i);
347 static inline void set_page_order(struct page *page, int order)
349 set_page_private(page, order);
350 __SetPageBuddy(page);
353 static inline void rmv_page_order(struct page *page)
355 __ClearPageBuddy(page);
356 set_page_private(page, 0);
360 * Locate the struct page for both the matching buddy in our
361 * pair (buddy1) and the combined O(n+1) page they form (page).
363 * 1) Any buddy B1 will have an order O twin B2 which satisfies
364 * the following equation:
365 * B2 = B1 ^ (1 << O)
366 * For example, if the starting buddy (buddy2) is #8 its order
367 * 1 buddy is #10:
368 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
370 * 2) Any buddy B will have an order O+1 parent P which
371 * satisfies the following equation:
372 * P = B & ~(1 << O)
374 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
376 static inline struct page *
377 __page_find_buddy(struct page *page, unsigned long page_idx, unsigned int order)
379 unsigned long buddy_idx = page_idx ^ (1 << order);
381 return page + (buddy_idx - page_idx);
384 static inline unsigned long
385 __find_combined_index(unsigned long page_idx, unsigned int order)
387 return (page_idx & ~(1 << order));
391 * This function checks whether a page is free && is the buddy
392 * we can do coalesce a page and its buddy if
393 * (a) the buddy is not in a hole &&
394 * (b) the buddy is in the buddy system &&
395 * (c) a page and its buddy have the same order &&
396 * (d) a page and its buddy are in the same zone.
398 * For recording whether a page is in the buddy system, we use PG_buddy.
399 * Setting, clearing, and testing PG_buddy is serialized by zone->lock.
401 * For recording page's order, we use page_private(page).
403 static inline int page_is_buddy(struct page *page, struct page *buddy,
404 int order)
406 if (!pfn_valid_within(page_to_pfn(buddy)))
407 return 0;
409 if (page_zone_id(page) != page_zone_id(buddy))
410 return 0;
412 if (PageBuddy(buddy) && page_order(buddy) == order) {
413 VM_BUG_ON(page_count(buddy) != 0);
414 return 1;
416 return 0;
420 * Freeing function for a buddy system allocator.
422 * The concept of a buddy system is to maintain direct-mapped table
423 * (containing bit values) for memory blocks of various "orders".
424 * The bottom level table contains the map for the smallest allocatable
425 * units of memory (here, pages), and each level above it describes
426 * pairs of units from the levels below, hence, "buddies".
427 * At a high level, all that happens here is marking the table entry
428 * at the bottom level available, and propagating the changes upward
429 * as necessary, plus some accounting needed to play nicely with other
430 * parts of the VM system.
431 * At each level, we keep a list of pages, which are heads of continuous
432 * free pages of length of (1 << order) and marked with PG_buddy. Page's
433 * order is recorded in page_private(page) field.
434 * So when we are allocating or freeing one, we can derive the state of the
435 * other. That is, if we allocate a small block, and both were
436 * free, the remainder of the region must be split into blocks.
437 * If a block is freed, and its buddy is also free, then this
438 * triggers coalescing into a block of larger size.
440 * -- wli
443 static inline void __free_one_page(struct page *page,
444 struct zone *zone, unsigned int order,
445 int migratetype)
447 unsigned long page_idx;
449 if (unlikely(PageCompound(page)))
450 if (unlikely(destroy_compound_page(page, order)))
451 return;
453 VM_BUG_ON(migratetype == -1);
455 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
457 VM_BUG_ON(page_idx & ((1 << order) - 1));
458 VM_BUG_ON(bad_range(zone, page));
460 while (order < MAX_ORDER-1) {
461 unsigned long combined_idx;
462 struct page *buddy;
464 buddy = __page_find_buddy(page, page_idx, order);
465 if (!page_is_buddy(page, buddy, order))
466 break;
468 /* Our buddy is free, merge with it and move up one order. */
469 list_del(&buddy->lru);
470 zone->free_area[order].nr_free--;
471 rmv_page_order(buddy);
472 combined_idx = __find_combined_index(page_idx, order);
473 page = page + (combined_idx - page_idx);
474 page_idx = combined_idx;
475 order++;
477 set_page_order(page, order);
478 list_add(&page->lru,
479 &zone->free_area[order].free_list[migratetype]);
480 zone->free_area[order].nr_free++;
483 #ifdef CONFIG_HAVE_MLOCKED_PAGE_BIT
485 * free_page_mlock() -- clean up attempts to free and mlocked() page.
486 * Page should not be on lru, so no need to fix that up.
487 * free_pages_check() will verify...
489 static inline void free_page_mlock(struct page *page)
491 __dec_zone_page_state(page, NR_MLOCK);
492 __count_vm_event(UNEVICTABLE_MLOCKFREED);
494 #else
495 static void free_page_mlock(struct page *page) { }
496 #endif
498 static inline int free_pages_check(struct page *page)
500 if (unlikely(page_mapcount(page) |
501 (page->mapping != NULL) |
502 (atomic_read(&page->_count) != 0) |
503 (page->flags & PAGE_FLAGS_CHECK_AT_FREE))) {
504 bad_page(page);
505 return 1;
507 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
508 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
509 return 0;
513 * Frees a list of pages.
514 * Assumes all pages on list are in same zone, and of same order.
515 * count is the number of pages to free.
517 * If the zone was previously in an "all pages pinned" state then look to
518 * see if this freeing clears that state.
520 * And clear the zone's pages_scanned counter, to hold off the "all pages are
521 * pinned" detection logic.
523 static void free_pages_bulk(struct zone *zone, int count,
524 struct list_head *list, int order)
526 spin_lock(&zone->lock);
527 zone_clear_flag(zone, ZONE_ALL_UNRECLAIMABLE);
528 zone->pages_scanned = 0;
530 __mod_zone_page_state(zone, NR_FREE_PAGES, count << order);
531 while (count--) {
532 struct page *page;
534 VM_BUG_ON(list_empty(list));
535 page = list_entry(list->prev, struct page, lru);
536 /* have to delete it as __free_one_page list manipulates */
537 list_del(&page->lru);
538 __free_one_page(page, zone, order, page_private(page));
540 spin_unlock(&zone->lock);
543 static void free_one_page(struct zone *zone, struct page *page, int order,
544 int migratetype)
546 spin_lock(&zone->lock);
547 zone_clear_flag(zone, ZONE_ALL_UNRECLAIMABLE);
548 zone->pages_scanned = 0;
550 __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order);
551 __free_one_page(page, zone, order, migratetype);
552 spin_unlock(&zone->lock);
555 static void __free_pages_ok(struct page *page, unsigned int order)
557 unsigned long flags;
558 int i;
559 int bad = 0;
560 int wasMlocked = TestClearPageMlocked(page);
562 kmemcheck_free_shadow(page, order);
564 for (i = 0 ; i < (1 << order) ; ++i)
565 bad += free_pages_check(page + i);
566 if (bad)
567 return;
569 if (!PageHighMem(page)) {
570 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
571 debug_check_no_obj_freed(page_address(page),
572 PAGE_SIZE << order);
574 arch_free_page(page, order);
575 kernel_map_pages(page, 1 << order, 0);
577 local_irq_save(flags);
578 if (unlikely(wasMlocked))
579 free_page_mlock(page);
580 __count_vm_events(PGFREE, 1 << order);
581 free_one_page(page_zone(page), page, order,
582 get_pageblock_migratetype(page));
583 local_irq_restore(flags);
587 * permit the bootmem allocator to evade page validation on high-order frees
589 void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
591 if (order == 0) {
592 __ClearPageReserved(page);
593 set_page_count(page, 0);
594 set_page_refcounted(page);
595 __free_page(page);
596 } else {
597 int loop;
599 prefetchw(page);
600 for (loop = 0; loop < BITS_PER_LONG; loop++) {
601 struct page *p = &page[loop];
603 if (loop + 1 < BITS_PER_LONG)
604 prefetchw(p + 1);
605 __ClearPageReserved(p);
606 set_page_count(p, 0);
609 set_page_refcounted(page);
610 __free_pages(page, order);
616 * The order of subdivision here is critical for the IO subsystem.
617 * Please do not alter this order without good reasons and regression
618 * testing. Specifically, as large blocks of memory are subdivided,
619 * the order in which smaller blocks are delivered depends on the order
620 * they're subdivided in this function. This is the primary factor
621 * influencing the order in which pages are delivered to the IO
622 * subsystem according to empirical testing, and this is also justified
623 * by considering the behavior of a buddy system containing a single
624 * large block of memory acted on by a series of small allocations.
625 * This behavior is a critical factor in sglist merging's success.
627 * -- wli
629 static inline void expand(struct zone *zone, struct page *page,
630 int low, int high, struct free_area *area,
631 int migratetype)
633 unsigned long size = 1 << high;
635 while (high > low) {
636 area--;
637 high--;
638 size >>= 1;
639 VM_BUG_ON(bad_range(zone, &page[size]));
640 list_add(&page[size].lru, &area->free_list[migratetype]);
641 area->nr_free++;
642 set_page_order(&page[size], high);
647 * This page is about to be returned from the page allocator
649 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
651 if (unlikely(page_mapcount(page) |
652 (page->mapping != NULL) |
653 (atomic_read(&page->_count) != 0) |
654 (page->flags & PAGE_FLAGS_CHECK_AT_PREP))) {
655 bad_page(page);
656 return 1;
659 set_page_private(page, 0);
660 set_page_refcounted(page);
662 arch_alloc_page(page, order);
663 kernel_map_pages(page, 1 << order, 1);
665 if (gfp_flags & __GFP_ZERO)
666 prep_zero_page(page, order, gfp_flags);
668 if (order && (gfp_flags & __GFP_COMP))
669 prep_compound_page(page, order);
671 return 0;
675 * Go through the free lists for the given migratetype and remove
676 * the smallest available page from the freelists
678 static inline
679 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
680 int migratetype)
682 unsigned int current_order;
683 struct free_area * area;
684 struct page *page;
686 /* Find a page of the appropriate size in the preferred list */
687 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
688 area = &(zone->free_area[current_order]);
689 if (list_empty(&area->free_list[migratetype]))
690 continue;
692 page = list_entry(area->free_list[migratetype].next,
693 struct page, lru);
694 list_del(&page->lru);
695 rmv_page_order(page);
696 area->nr_free--;
697 expand(zone, page, order, current_order, area, migratetype);
698 return page;
701 return NULL;
706 * This array describes the order lists are fallen back to when
707 * the free lists for the desirable migrate type are depleted
709 static int fallbacks[MIGRATE_TYPES][MIGRATE_TYPES-1] = {
710 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
711 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
712 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
713 [MIGRATE_RESERVE] = { MIGRATE_RESERVE, MIGRATE_RESERVE, MIGRATE_RESERVE }, /* Never used */
717 * Move the free pages in a range to the free lists of the requested type.
718 * Note that start_page and end_pages are not aligned on a pageblock
719 * boundary. If alignment is required, use move_freepages_block()
721 static int move_freepages(struct zone *zone,
722 struct page *start_page, struct page *end_page,
723 int migratetype)
725 struct page *page;
726 unsigned long order;
727 int pages_moved = 0;
729 #ifndef CONFIG_HOLES_IN_ZONE
731 * page_zone is not safe to call in this context when
732 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
733 * anyway as we check zone boundaries in move_freepages_block().
734 * Remove at a later date when no bug reports exist related to
735 * grouping pages by mobility
737 BUG_ON(page_zone(start_page) != page_zone(end_page));
738 #endif
740 for (page = start_page; page <= end_page;) {
741 /* Make sure we are not inadvertently changing nodes */
742 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
744 if (!pfn_valid_within(page_to_pfn(page))) {
745 page++;
746 continue;
749 if (!PageBuddy(page)) {
750 page++;
751 continue;
754 order = page_order(page);
755 list_del(&page->lru);
756 list_add(&page->lru,
757 &zone->free_area[order].free_list[migratetype]);
758 page += 1 << order;
759 pages_moved += 1 << order;
762 return pages_moved;
765 static int move_freepages_block(struct zone *zone, struct page *page,
766 int migratetype)
768 unsigned long start_pfn, end_pfn;
769 struct page *start_page, *end_page;
771 start_pfn = page_to_pfn(page);
772 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
773 start_page = pfn_to_page(start_pfn);
774 end_page = start_page + pageblock_nr_pages - 1;
775 end_pfn = start_pfn + pageblock_nr_pages - 1;
777 /* Do not cross zone boundaries */
778 if (start_pfn < zone->zone_start_pfn)
779 start_page = page;
780 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
781 return 0;
783 return move_freepages(zone, start_page, end_page, migratetype);
786 /* Remove an element from the buddy allocator from the fallback list */
787 static inline struct page *
788 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
790 struct free_area * area;
791 int current_order;
792 struct page *page;
793 int migratetype, i;
795 /* Find the largest possible block of pages in the other list */
796 for (current_order = MAX_ORDER-1; current_order >= order;
797 --current_order) {
798 for (i = 0; i < MIGRATE_TYPES - 1; i++) {
799 migratetype = fallbacks[start_migratetype][i];
801 /* MIGRATE_RESERVE handled later if necessary */
802 if (migratetype == MIGRATE_RESERVE)
803 continue;
805 area = &(zone->free_area[current_order]);
806 if (list_empty(&area->free_list[migratetype]))
807 continue;
809 page = list_entry(area->free_list[migratetype].next,
810 struct page, lru);
811 area->nr_free--;
814 * If breaking a large block of pages, move all free
815 * pages to the preferred allocation list. If falling
816 * back for a reclaimable kernel allocation, be more
817 * agressive about taking ownership of free pages
819 if (unlikely(current_order >= (pageblock_order >> 1)) ||
820 start_migratetype == MIGRATE_RECLAIMABLE ||
821 page_group_by_mobility_disabled) {
822 unsigned long pages;
823 pages = move_freepages_block(zone, page,
824 start_migratetype);
826 /* Claim the whole block if over half of it is free */
827 if (pages >= (1 << (pageblock_order-1)) ||
828 page_group_by_mobility_disabled)
829 set_pageblock_migratetype(page,
830 start_migratetype);
832 migratetype = start_migratetype;
835 /* Remove the page from the freelists */
836 list_del(&page->lru);
837 rmv_page_order(page);
839 if (current_order == pageblock_order)
840 set_pageblock_migratetype(page,
841 start_migratetype);
843 expand(zone, page, order, current_order, area, migratetype);
844 return page;
848 return NULL;
852 * Do the hard work of removing an element from the buddy allocator.
853 * Call me with the zone->lock already held.
855 static struct page *__rmqueue(struct zone *zone, unsigned int order,
856 int migratetype)
858 struct page *page;
860 retry_reserve:
861 page = __rmqueue_smallest(zone, order, migratetype);
863 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
864 page = __rmqueue_fallback(zone, order, migratetype);
867 * Use MIGRATE_RESERVE rather than fail an allocation. goto
868 * is used because __rmqueue_smallest is an inline function
869 * and we want just one call site
871 if (!page) {
872 migratetype = MIGRATE_RESERVE;
873 goto retry_reserve;
877 return page;
881 * Obtain a specified number of elements from the buddy allocator, all under
882 * a single hold of the lock, for efficiency. Add them to the supplied list.
883 * Returns the number of new pages which were placed at *list.
885 static int rmqueue_bulk(struct zone *zone, unsigned int order,
886 unsigned long count, struct list_head *list,
887 int migratetype, int cold)
889 int i;
891 spin_lock(&zone->lock);
892 for (i = 0; i < count; ++i) {
893 struct page *page = __rmqueue(zone, order, migratetype);
894 if (unlikely(page == NULL))
895 break;
898 * Split buddy pages returned by expand() are received here
899 * in physical page order. The page is added to the callers and
900 * list and the list head then moves forward. From the callers
901 * perspective, the linked list is ordered by page number in
902 * some conditions. This is useful for IO devices that can
903 * merge IO requests if the physical pages are ordered
904 * properly.
906 if (likely(cold == 0))
907 list_add(&page->lru, list);
908 else
909 list_add_tail(&page->lru, list);
910 set_page_private(page, migratetype);
911 list = &page->lru;
913 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
914 spin_unlock(&zone->lock);
915 return i;
918 #ifdef CONFIG_NUMA
920 * Called from the vmstat counter updater to drain pagesets of this
921 * currently executing processor on remote nodes after they have
922 * expired.
924 * Note that this function must be called with the thread pinned to
925 * a single processor.
927 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
929 unsigned long flags;
930 int to_drain;
932 local_irq_save(flags);
933 if (pcp->count >= pcp->batch)
934 to_drain = pcp->batch;
935 else
936 to_drain = pcp->count;
937 free_pages_bulk(zone, to_drain, &pcp->list, 0);
938 pcp->count -= to_drain;
939 local_irq_restore(flags);
941 #endif
944 * Drain pages of the indicated processor.
946 * The processor must either be the current processor and the
947 * thread pinned to the current processor or a processor that
948 * is not online.
950 static void drain_pages(unsigned int cpu)
952 unsigned long flags;
953 struct zone *zone;
955 for_each_populated_zone(zone) {
956 struct per_cpu_pageset *pset;
957 struct per_cpu_pages *pcp;
959 pset = zone_pcp(zone, cpu);
961 pcp = &pset->pcp;
962 local_irq_save(flags);
963 free_pages_bulk(zone, pcp->count, &pcp->list, 0);
964 pcp->count = 0;
965 local_irq_restore(flags);
970 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
972 void drain_local_pages(void *arg)
974 drain_pages(smp_processor_id());
978 * Spill all the per-cpu pages from all CPUs back into the buddy allocator
980 void drain_all_pages(void)
982 on_each_cpu(drain_local_pages, NULL, 1);
985 #ifdef CONFIG_HIBERNATION
987 void mark_free_pages(struct zone *zone)
989 unsigned long pfn, max_zone_pfn;
990 unsigned long flags;
991 int order, t;
992 struct list_head *curr;
994 if (!zone->spanned_pages)
995 return;
997 spin_lock_irqsave(&zone->lock, flags);
999 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
1000 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1001 if (pfn_valid(pfn)) {
1002 struct page *page = pfn_to_page(pfn);
1004 if (!swsusp_page_is_forbidden(page))
1005 swsusp_unset_page_free(page);
1008 for_each_migratetype_order(order, t) {
1009 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1010 unsigned long i;
1012 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1013 for (i = 0; i < (1UL << order); i++)
1014 swsusp_set_page_free(pfn_to_page(pfn + i));
1017 spin_unlock_irqrestore(&zone->lock, flags);
1019 #endif /* CONFIG_PM */
1022 * Free a 0-order page
1024 static void free_hot_cold_page(struct page *page, int cold)
1026 struct zone *zone = page_zone(page);
1027 struct per_cpu_pages *pcp;
1028 unsigned long flags;
1029 int wasMlocked = TestClearPageMlocked(page);
1031 kmemcheck_free_shadow(page, 0);
1033 if (PageAnon(page))
1034 page->mapping = NULL;
1035 if (free_pages_check(page))
1036 return;
1038 if (!PageHighMem(page)) {
1039 debug_check_no_locks_freed(page_address(page), PAGE_SIZE);
1040 debug_check_no_obj_freed(page_address(page), PAGE_SIZE);
1042 arch_free_page(page, 0);
1043 kernel_map_pages(page, 1, 0);
1045 pcp = &zone_pcp(zone, get_cpu())->pcp;
1046 set_page_private(page, get_pageblock_migratetype(page));
1047 local_irq_save(flags);
1048 if (unlikely(wasMlocked))
1049 free_page_mlock(page);
1050 __count_vm_event(PGFREE);
1052 if (cold)
1053 list_add_tail(&page->lru, &pcp->list);
1054 else
1055 list_add(&page->lru, &pcp->list);
1056 pcp->count++;
1057 if (pcp->count >= pcp->high) {
1058 free_pages_bulk(zone, pcp->batch, &pcp->list, 0);
1059 pcp->count -= pcp->batch;
1061 local_irq_restore(flags);
1062 put_cpu();
1065 void free_hot_page(struct page *page)
1067 free_hot_cold_page(page, 0);
1070 void free_cold_page(struct page *page)
1072 free_hot_cold_page(page, 1);
1076 * split_page takes a non-compound higher-order page, and splits it into
1077 * n (1<<order) sub-pages: page[0..n]
1078 * Each sub-page must be freed individually.
1080 * Note: this is probably too low level an operation for use in drivers.
1081 * Please consult with lkml before using this in your driver.
1083 void split_page(struct page *page, unsigned int order)
1085 int i;
1087 VM_BUG_ON(PageCompound(page));
1088 VM_BUG_ON(!page_count(page));
1090 #ifdef CONFIG_KMEMCHECK
1092 * Split shadow pages too, because free(page[0]) would
1093 * otherwise free the whole shadow.
1095 if (kmemcheck_page_is_tracked(page))
1096 split_page(virt_to_page(page[0].shadow), order);
1097 #endif
1099 for (i = 1; i < (1 << order); i++)
1100 set_page_refcounted(page + i);
1104 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1105 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1106 * or two.
1108 static inline
1109 struct page *buffered_rmqueue(struct zone *preferred_zone,
1110 struct zone *zone, int order, gfp_t gfp_flags,
1111 int migratetype)
1113 unsigned long flags;
1114 struct page *page;
1115 int cold = !!(gfp_flags & __GFP_COLD);
1116 int cpu;
1118 again:
1119 cpu = get_cpu();
1120 if (likely(order == 0)) {
1121 struct per_cpu_pages *pcp;
1123 pcp = &zone_pcp(zone, cpu)->pcp;
1124 local_irq_save(flags);
1125 if (!pcp->count) {
1126 pcp->count = rmqueue_bulk(zone, 0,
1127 pcp->batch, &pcp->list,
1128 migratetype, cold);
1129 if (unlikely(!pcp->count))
1130 goto failed;
1133 /* Find a page of the appropriate migrate type */
1134 if (cold) {
1135 list_for_each_entry_reverse(page, &pcp->list, lru)
1136 if (page_private(page) == migratetype)
1137 break;
1138 } else {
1139 list_for_each_entry(page, &pcp->list, lru)
1140 if (page_private(page) == migratetype)
1141 break;
1144 /* Allocate more to the pcp list if necessary */
1145 if (unlikely(&page->lru == &pcp->list)) {
1146 int get_one_page = 0;
1148 pcp->count += rmqueue_bulk(zone, 0,
1149 pcp->batch, &pcp->list,
1150 migratetype, cold);
1151 list_for_each_entry(page, &pcp->list, lru) {
1152 if (get_pageblock_migratetype(page) !=
1153 MIGRATE_ISOLATE) {
1154 get_one_page = 1;
1155 break;
1158 if (!get_one_page)
1159 goto failed;
1162 list_del(&page->lru);
1163 pcp->count--;
1164 } else {
1165 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1167 * __GFP_NOFAIL is not to be used in new code.
1169 * All __GFP_NOFAIL callers should be fixed so that they
1170 * properly detect and handle allocation failures.
1172 * We most definitely don't want callers attempting to
1173 * allocate greater than order-1 page units with
1174 * __GFP_NOFAIL.
1176 WARN_ON_ONCE(order > 1);
1178 spin_lock_irqsave(&zone->lock, flags);
1179 page = __rmqueue(zone, order, migratetype);
1180 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << order));
1181 spin_unlock(&zone->lock);
1182 if (!page)
1183 goto failed;
1186 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1187 zone_statistics(preferred_zone, zone);
1188 local_irq_restore(flags);
1189 put_cpu();
1191 VM_BUG_ON(bad_range(zone, page));
1192 if (prep_new_page(page, order, gfp_flags))
1193 goto again;
1194 return page;
1196 failed:
1197 local_irq_restore(flags);
1198 put_cpu();
1199 return NULL;
1202 /* The ALLOC_WMARK bits are used as an index to zone->watermark */
1203 #define ALLOC_WMARK_MIN WMARK_MIN
1204 #define ALLOC_WMARK_LOW WMARK_LOW
1205 #define ALLOC_WMARK_HIGH WMARK_HIGH
1206 #define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */
1208 /* Mask to get the watermark bits */
1209 #define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1)
1211 #define ALLOC_HARDER 0x10 /* try to alloc harder */
1212 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
1213 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
1215 #ifdef CONFIG_FAIL_PAGE_ALLOC
1217 static struct fail_page_alloc_attr {
1218 struct fault_attr attr;
1220 u32 ignore_gfp_highmem;
1221 u32 ignore_gfp_wait;
1222 u32 min_order;
1224 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1226 struct dentry *ignore_gfp_highmem_file;
1227 struct dentry *ignore_gfp_wait_file;
1228 struct dentry *min_order_file;
1230 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1232 } fail_page_alloc = {
1233 .attr = FAULT_ATTR_INITIALIZER,
1234 .ignore_gfp_wait = 1,
1235 .ignore_gfp_highmem = 1,
1236 .min_order = 1,
1239 static int __init setup_fail_page_alloc(char *str)
1241 return setup_fault_attr(&fail_page_alloc.attr, str);
1243 __setup("fail_page_alloc=", setup_fail_page_alloc);
1245 static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1247 if (order < fail_page_alloc.min_order)
1248 return 0;
1249 if (gfp_mask & __GFP_NOFAIL)
1250 return 0;
1251 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1252 return 0;
1253 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1254 return 0;
1256 return should_fail(&fail_page_alloc.attr, 1 << order);
1259 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1261 static int __init fail_page_alloc_debugfs(void)
1263 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1264 struct dentry *dir;
1265 int err;
1267 err = init_fault_attr_dentries(&fail_page_alloc.attr,
1268 "fail_page_alloc");
1269 if (err)
1270 return err;
1271 dir = fail_page_alloc.attr.dentries.dir;
1273 fail_page_alloc.ignore_gfp_wait_file =
1274 debugfs_create_bool("ignore-gfp-wait", mode, dir,
1275 &fail_page_alloc.ignore_gfp_wait);
1277 fail_page_alloc.ignore_gfp_highmem_file =
1278 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1279 &fail_page_alloc.ignore_gfp_highmem);
1280 fail_page_alloc.min_order_file =
1281 debugfs_create_u32("min-order", mode, dir,
1282 &fail_page_alloc.min_order);
1284 if (!fail_page_alloc.ignore_gfp_wait_file ||
1285 !fail_page_alloc.ignore_gfp_highmem_file ||
1286 !fail_page_alloc.min_order_file) {
1287 err = -ENOMEM;
1288 debugfs_remove(fail_page_alloc.ignore_gfp_wait_file);
1289 debugfs_remove(fail_page_alloc.ignore_gfp_highmem_file);
1290 debugfs_remove(fail_page_alloc.min_order_file);
1291 cleanup_fault_attr_dentries(&fail_page_alloc.attr);
1294 return err;
1297 late_initcall(fail_page_alloc_debugfs);
1299 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1301 #else /* CONFIG_FAIL_PAGE_ALLOC */
1303 static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1305 return 0;
1308 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1311 * Return 1 if free pages are above 'mark'. This takes into account the order
1312 * of the allocation.
1314 int zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1315 int classzone_idx, int alloc_flags)
1317 /* free_pages my go negative - that's OK */
1318 long min = mark;
1319 long free_pages = zone_page_state(z, NR_FREE_PAGES) - (1 << order) + 1;
1320 int o;
1322 if (alloc_flags & ALLOC_HIGH)
1323 min -= min / 2;
1324 if (alloc_flags & ALLOC_HARDER)
1325 min -= min / 4;
1327 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
1328 return 0;
1329 for (o = 0; o < order; o++) {
1330 /* At the next order, this order's pages become unavailable */
1331 free_pages -= z->free_area[o].nr_free << o;
1333 /* Require fewer higher order pages to be free */
1334 min >>= 1;
1336 if (free_pages <= min)
1337 return 0;
1339 return 1;
1342 #ifdef CONFIG_NUMA
1344 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1345 * skip over zones that are not allowed by the cpuset, or that have
1346 * been recently (in last second) found to be nearly full. See further
1347 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1348 * that have to skip over a lot of full or unallowed zones.
1350 * If the zonelist cache is present in the passed in zonelist, then
1351 * returns a pointer to the allowed node mask (either the current
1352 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1354 * If the zonelist cache is not available for this zonelist, does
1355 * nothing and returns NULL.
1357 * If the fullzones BITMAP in the zonelist cache is stale (more than
1358 * a second since last zap'd) then we zap it out (clear its bits.)
1360 * We hold off even calling zlc_setup, until after we've checked the
1361 * first zone in the zonelist, on the theory that most allocations will
1362 * be satisfied from that first zone, so best to examine that zone as
1363 * quickly as we can.
1365 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1367 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1368 nodemask_t *allowednodes; /* zonelist_cache approximation */
1370 zlc = zonelist->zlcache_ptr;
1371 if (!zlc)
1372 return NULL;
1374 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1375 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1376 zlc->last_full_zap = jiffies;
1379 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1380 &cpuset_current_mems_allowed :
1381 &node_states[N_HIGH_MEMORY];
1382 return allowednodes;
1386 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1387 * if it is worth looking at further for free memory:
1388 * 1) Check that the zone isn't thought to be full (doesn't have its
1389 * bit set in the zonelist_cache fullzones BITMAP).
1390 * 2) Check that the zones node (obtained from the zonelist_cache
1391 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1392 * Return true (non-zero) if zone is worth looking at further, or
1393 * else return false (zero) if it is not.
1395 * This check -ignores- the distinction between various watermarks,
1396 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1397 * found to be full for any variation of these watermarks, it will
1398 * be considered full for up to one second by all requests, unless
1399 * we are so low on memory on all allowed nodes that we are forced
1400 * into the second scan of the zonelist.
1402 * In the second scan we ignore this zonelist cache and exactly
1403 * apply the watermarks to all zones, even it is slower to do so.
1404 * We are low on memory in the second scan, and should leave no stone
1405 * unturned looking for a free page.
1407 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1408 nodemask_t *allowednodes)
1410 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1411 int i; /* index of *z in zonelist zones */
1412 int n; /* node that zone *z is on */
1414 zlc = zonelist->zlcache_ptr;
1415 if (!zlc)
1416 return 1;
1418 i = z - zonelist->_zonerefs;
1419 n = zlc->z_to_n[i];
1421 /* This zone is worth trying if it is allowed but not full */
1422 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1426 * Given 'z' scanning a zonelist, set the corresponding bit in
1427 * zlc->fullzones, so that subsequent attempts to allocate a page
1428 * from that zone don't waste time re-examining it.
1430 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1432 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1433 int i; /* index of *z in zonelist zones */
1435 zlc = zonelist->zlcache_ptr;
1436 if (!zlc)
1437 return;
1439 i = z - zonelist->_zonerefs;
1441 set_bit(i, zlc->fullzones);
1444 #else /* CONFIG_NUMA */
1446 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1448 return NULL;
1451 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1452 nodemask_t *allowednodes)
1454 return 1;
1457 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1460 #endif /* CONFIG_NUMA */
1463 * get_page_from_freelist goes through the zonelist trying to allocate
1464 * a page.
1466 static struct page *
1467 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1468 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1469 struct zone *preferred_zone, int migratetype)
1471 struct zoneref *z;
1472 struct page *page = NULL;
1473 int classzone_idx;
1474 struct zone *zone;
1475 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1476 int zlc_active = 0; /* set if using zonelist_cache */
1477 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1479 classzone_idx = zone_idx(preferred_zone);
1480 zonelist_scan:
1482 * Scan zonelist, looking for a zone with enough free.
1483 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1485 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1486 high_zoneidx, nodemask) {
1487 if (NUMA_BUILD && zlc_active &&
1488 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1489 continue;
1490 if ((alloc_flags & ALLOC_CPUSET) &&
1491 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1492 goto try_next_zone;
1494 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1495 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1496 unsigned long mark;
1497 int ret;
1499 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1500 if (zone_watermark_ok(zone, order, mark,
1501 classzone_idx, alloc_flags))
1502 goto try_this_zone;
1504 if (zone_reclaim_mode == 0)
1505 goto this_zone_full;
1507 ret = zone_reclaim(zone, gfp_mask, order);
1508 switch (ret) {
1509 case ZONE_RECLAIM_NOSCAN:
1510 /* did not scan */
1511 goto try_next_zone;
1512 case ZONE_RECLAIM_FULL:
1513 /* scanned but unreclaimable */
1514 goto this_zone_full;
1515 default:
1516 /* did we reclaim enough */
1517 if (!zone_watermark_ok(zone, order, mark,
1518 classzone_idx, alloc_flags))
1519 goto this_zone_full;
1523 try_this_zone:
1524 page = buffered_rmqueue(preferred_zone, zone, order,
1525 gfp_mask, migratetype);
1526 if (page)
1527 break;
1528 this_zone_full:
1529 if (NUMA_BUILD)
1530 zlc_mark_zone_full(zonelist, z);
1531 try_next_zone:
1532 if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) {
1534 * we do zlc_setup after the first zone is tried but only
1535 * if there are multiple nodes make it worthwhile
1537 allowednodes = zlc_setup(zonelist, alloc_flags);
1538 zlc_active = 1;
1539 did_zlc_setup = 1;
1543 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) {
1544 /* Disable zlc cache for second zonelist scan */
1545 zlc_active = 0;
1546 goto zonelist_scan;
1548 return page;
1551 static inline int
1552 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
1553 unsigned long pages_reclaimed)
1555 /* Do not loop if specifically requested */
1556 if (gfp_mask & __GFP_NORETRY)
1557 return 0;
1560 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
1561 * means __GFP_NOFAIL, but that may not be true in other
1562 * implementations.
1564 if (order <= PAGE_ALLOC_COSTLY_ORDER)
1565 return 1;
1568 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
1569 * specified, then we retry until we no longer reclaim any pages
1570 * (above), or we've reclaimed an order of pages at least as
1571 * large as the allocation's order. In both cases, if the
1572 * allocation still fails, we stop retrying.
1574 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
1575 return 1;
1578 * Don't let big-order allocations loop unless the caller
1579 * explicitly requests that.
1581 if (gfp_mask & __GFP_NOFAIL)
1582 return 1;
1584 return 0;
1587 static inline struct page *
1588 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
1589 struct zonelist *zonelist, enum zone_type high_zoneidx,
1590 nodemask_t *nodemask, struct zone *preferred_zone,
1591 int migratetype)
1593 struct page *page;
1595 /* Acquire the OOM killer lock for the zones in zonelist */
1596 if (!try_set_zone_oom(zonelist, gfp_mask)) {
1597 schedule_timeout_uninterruptible(1);
1598 return NULL;
1602 * Go through the zonelist yet one more time, keep very high watermark
1603 * here, this is only to catch a parallel oom killing, we must fail if
1604 * we're still under heavy pressure.
1606 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
1607 order, zonelist, high_zoneidx,
1608 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
1609 preferred_zone, migratetype);
1610 if (page)
1611 goto out;
1613 /* The OOM killer will not help higher order allocs */
1614 if (order > PAGE_ALLOC_COSTLY_ORDER && !(gfp_mask & __GFP_NOFAIL))
1615 goto out;
1617 /* Exhausted what can be done so it's blamo time */
1618 out_of_memory(zonelist, gfp_mask, order);
1620 out:
1621 clear_zonelist_oom(zonelist, gfp_mask);
1622 return page;
1625 /* The really slow allocator path where we enter direct reclaim */
1626 static inline struct page *
1627 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
1628 struct zonelist *zonelist, enum zone_type high_zoneidx,
1629 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1630 int migratetype, unsigned long *did_some_progress)
1632 struct page *page = NULL;
1633 struct reclaim_state reclaim_state;
1634 struct task_struct *p = current;
1636 cond_resched();
1638 /* We now go into synchronous reclaim */
1639 cpuset_memory_pressure_bump();
1640 p->flags |= PF_MEMALLOC;
1641 lockdep_set_current_reclaim_state(gfp_mask);
1642 reclaim_state.reclaimed_slab = 0;
1643 p->reclaim_state = &reclaim_state;
1645 *did_some_progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
1647 p->reclaim_state = NULL;
1648 lockdep_clear_current_reclaim_state();
1649 p->flags &= ~PF_MEMALLOC;
1651 cond_resched();
1653 if (order != 0)
1654 drain_all_pages();
1656 if (likely(*did_some_progress))
1657 page = get_page_from_freelist(gfp_mask, nodemask, order,
1658 zonelist, high_zoneidx,
1659 alloc_flags, preferred_zone,
1660 migratetype);
1661 return page;
1665 * This is called in the allocator slow-path if the allocation request is of
1666 * sufficient urgency to ignore watermarks and take other desperate measures
1668 static inline struct page *
1669 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
1670 struct zonelist *zonelist, enum zone_type high_zoneidx,
1671 nodemask_t *nodemask, struct zone *preferred_zone,
1672 int migratetype)
1674 struct page *page;
1676 do {
1677 page = get_page_from_freelist(gfp_mask, nodemask, order,
1678 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
1679 preferred_zone, migratetype);
1681 if (!page && gfp_mask & __GFP_NOFAIL)
1682 congestion_wait(BLK_RW_ASYNC, HZ/50);
1683 } while (!page && (gfp_mask & __GFP_NOFAIL));
1685 return page;
1688 static inline
1689 void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
1690 enum zone_type high_zoneidx)
1692 struct zoneref *z;
1693 struct zone *zone;
1695 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
1696 wakeup_kswapd(zone, order);
1699 static inline int
1700 gfp_to_alloc_flags(gfp_t gfp_mask)
1702 struct task_struct *p = current;
1703 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
1704 const gfp_t wait = gfp_mask & __GFP_WAIT;
1706 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
1707 BUILD_BUG_ON(__GFP_HIGH != ALLOC_HIGH);
1710 * The caller may dip into page reserves a bit more if the caller
1711 * cannot run direct reclaim, or if the caller has realtime scheduling
1712 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
1713 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
1715 alloc_flags |= (gfp_mask & __GFP_HIGH);
1717 if (!wait) {
1718 alloc_flags |= ALLOC_HARDER;
1720 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
1721 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1723 alloc_flags &= ~ALLOC_CPUSET;
1724 } else if (unlikely(rt_task(p)))
1725 alloc_flags |= ALLOC_HARDER;
1727 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
1728 if (!in_interrupt() &&
1729 ((p->flags & PF_MEMALLOC) ||
1730 unlikely(test_thread_flag(TIF_MEMDIE))))
1731 alloc_flags |= ALLOC_NO_WATERMARKS;
1734 return alloc_flags;
1737 static inline struct page *
1738 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
1739 struct zonelist *zonelist, enum zone_type high_zoneidx,
1740 nodemask_t *nodemask, struct zone *preferred_zone,
1741 int migratetype)
1743 const gfp_t wait = gfp_mask & __GFP_WAIT;
1744 struct page *page = NULL;
1745 int alloc_flags;
1746 unsigned long pages_reclaimed = 0;
1747 unsigned long did_some_progress;
1748 struct task_struct *p = current;
1751 * In the slowpath, we sanity check order to avoid ever trying to
1752 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
1753 * be using allocators in order of preference for an area that is
1754 * too large.
1756 if (order >= MAX_ORDER) {
1757 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
1758 return NULL;
1762 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
1763 * __GFP_NOWARN set) should not cause reclaim since the subsystem
1764 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
1765 * using a larger set of nodes after it has established that the
1766 * allowed per node queues are empty and that nodes are
1767 * over allocated.
1769 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
1770 goto nopage;
1772 wake_all_kswapd(order, zonelist, high_zoneidx);
1775 * OK, we're below the kswapd watermark and have kicked background
1776 * reclaim. Now things get more complex, so set up alloc_flags according
1777 * to how we want to proceed.
1779 alloc_flags = gfp_to_alloc_flags(gfp_mask);
1781 restart:
1782 /* This is the last chance, in general, before the goto nopage. */
1783 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
1784 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
1785 preferred_zone, migratetype);
1786 if (page)
1787 goto got_pg;
1789 rebalance:
1790 /* Allocate without watermarks if the context allows */
1791 if (alloc_flags & ALLOC_NO_WATERMARKS) {
1792 page = __alloc_pages_high_priority(gfp_mask, order,
1793 zonelist, high_zoneidx, nodemask,
1794 preferred_zone, migratetype);
1795 if (page)
1796 goto got_pg;
1799 /* Atomic allocations - we can't balance anything */
1800 if (!wait)
1801 goto nopage;
1803 /* Avoid recursion of direct reclaim */
1804 if (p->flags & PF_MEMALLOC)
1805 goto nopage;
1807 /* Avoid allocations with no watermarks from looping endlessly */
1808 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
1809 goto nopage;
1811 /* Try direct reclaim and then allocating */
1812 page = __alloc_pages_direct_reclaim(gfp_mask, order,
1813 zonelist, high_zoneidx,
1814 nodemask,
1815 alloc_flags, preferred_zone,
1816 migratetype, &did_some_progress);
1817 if (page)
1818 goto got_pg;
1821 * If we failed to make any progress reclaiming, then we are
1822 * running out of options and have to consider going OOM
1824 if (!did_some_progress) {
1825 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
1826 if (oom_killer_disabled)
1827 goto nopage;
1828 page = __alloc_pages_may_oom(gfp_mask, order,
1829 zonelist, high_zoneidx,
1830 nodemask, preferred_zone,
1831 migratetype);
1832 if (page)
1833 goto got_pg;
1836 * The OOM killer does not trigger for high-order
1837 * ~__GFP_NOFAIL allocations so if no progress is being
1838 * made, there are no other options and retrying is
1839 * unlikely to help.
1841 if (order > PAGE_ALLOC_COSTLY_ORDER &&
1842 !(gfp_mask & __GFP_NOFAIL))
1843 goto nopage;
1845 goto restart;
1849 /* Check if we should retry the allocation */
1850 pages_reclaimed += did_some_progress;
1851 if (should_alloc_retry(gfp_mask, order, pages_reclaimed)) {
1852 /* Wait for some write requests to complete then retry */
1853 congestion_wait(BLK_RW_ASYNC, HZ/50);
1854 goto rebalance;
1857 nopage:
1858 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
1859 printk(KERN_WARNING "%s: page allocation failure."
1860 " order:%d, mode:0x%x\n",
1861 p->comm, order, gfp_mask);
1862 dump_stack();
1863 show_mem();
1865 return page;
1866 got_pg:
1867 if (kmemcheck_enabled)
1868 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
1869 return page;
1874 * This is the 'heart' of the zoned buddy allocator.
1876 struct page *
1877 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
1878 struct zonelist *zonelist, nodemask_t *nodemask)
1880 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
1881 struct zone *preferred_zone;
1882 struct page *page;
1883 int migratetype = allocflags_to_migratetype(gfp_mask);
1885 gfp_mask &= gfp_allowed_mask;
1887 lockdep_trace_alloc(gfp_mask);
1889 might_sleep_if(gfp_mask & __GFP_WAIT);
1891 if (should_fail_alloc_page(gfp_mask, order))
1892 return NULL;
1895 * Check the zones suitable for the gfp_mask contain at least one
1896 * valid zone. It's possible to have an empty zonelist as a result
1897 * of GFP_THISNODE and a memoryless node
1899 if (unlikely(!zonelist->_zonerefs->zone))
1900 return NULL;
1902 /* The preferred zone is used for statistics later */
1903 first_zones_zonelist(zonelist, high_zoneidx, nodemask, &preferred_zone);
1904 if (!preferred_zone)
1905 return NULL;
1907 /* First allocation attempt */
1908 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
1909 zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET,
1910 preferred_zone, migratetype);
1911 if (unlikely(!page))
1912 page = __alloc_pages_slowpath(gfp_mask, order,
1913 zonelist, high_zoneidx, nodemask,
1914 preferred_zone, migratetype);
1916 return page;
1918 EXPORT_SYMBOL(__alloc_pages_nodemask);
1921 * Common helper functions.
1923 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
1925 struct page * page;
1926 page = alloc_pages(gfp_mask, order);
1927 if (!page)
1928 return 0;
1929 return (unsigned long) page_address(page);
1932 EXPORT_SYMBOL(__get_free_pages);
1934 unsigned long get_zeroed_page(gfp_t gfp_mask)
1936 struct page * page;
1939 * get_zeroed_page() returns a 32-bit address, which cannot represent
1940 * a highmem page
1942 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
1944 page = alloc_pages(gfp_mask | __GFP_ZERO, 0);
1945 if (page)
1946 return (unsigned long) page_address(page);
1947 return 0;
1950 EXPORT_SYMBOL(get_zeroed_page);
1952 void __pagevec_free(struct pagevec *pvec)
1954 int i = pagevec_count(pvec);
1956 while (--i >= 0)
1957 free_hot_cold_page(pvec->pages[i], pvec->cold);
1960 void __free_pages(struct page *page, unsigned int order)
1962 if (put_page_testzero(page)) {
1963 if (order == 0)
1964 free_hot_page(page);
1965 else
1966 __free_pages_ok(page, order);
1970 EXPORT_SYMBOL(__free_pages);
1972 void free_pages(unsigned long addr, unsigned int order)
1974 if (addr != 0) {
1975 VM_BUG_ON(!virt_addr_valid((void *)addr));
1976 __free_pages(virt_to_page((void *)addr), order);
1980 EXPORT_SYMBOL(free_pages);
1983 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
1984 * @size: the number of bytes to allocate
1985 * @gfp_mask: GFP flags for the allocation
1987 * This function is similar to alloc_pages(), except that it allocates the
1988 * minimum number of pages to satisfy the request. alloc_pages() can only
1989 * allocate memory in power-of-two pages.
1991 * This function is also limited by MAX_ORDER.
1993 * Memory allocated by this function must be released by free_pages_exact().
1995 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
1997 unsigned int order = get_order(size);
1998 unsigned long addr;
2000 addr = __get_free_pages(gfp_mask, order);
2001 if (addr) {
2002 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2003 unsigned long used = addr + PAGE_ALIGN(size);
2005 split_page(virt_to_page((void *)addr), order);
2006 while (used < alloc_end) {
2007 free_page(used);
2008 used += PAGE_SIZE;
2012 return (void *)addr;
2014 EXPORT_SYMBOL(alloc_pages_exact);
2017 * free_pages_exact - release memory allocated via alloc_pages_exact()
2018 * @virt: the value returned by alloc_pages_exact.
2019 * @size: size of allocation, same value as passed to alloc_pages_exact().
2021 * Release the memory allocated by a previous call to alloc_pages_exact.
2023 void free_pages_exact(void *virt, size_t size)
2025 unsigned long addr = (unsigned long)virt;
2026 unsigned long end = addr + PAGE_ALIGN(size);
2028 while (addr < end) {
2029 free_page(addr);
2030 addr += PAGE_SIZE;
2033 EXPORT_SYMBOL(free_pages_exact);
2035 static unsigned int nr_free_zone_pages(int offset)
2037 struct zoneref *z;
2038 struct zone *zone;
2040 /* Just pick one node, since fallback list is circular */
2041 unsigned int sum = 0;
2043 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2045 for_each_zone_zonelist(zone, z, zonelist, offset) {
2046 unsigned long size = zone->present_pages;
2047 unsigned long high = high_wmark_pages(zone);
2048 if (size > high)
2049 sum += size - high;
2052 return sum;
2056 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2058 unsigned int nr_free_buffer_pages(void)
2060 return nr_free_zone_pages(gfp_zone(GFP_USER));
2062 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2065 * Amount of free RAM allocatable within all zones
2067 unsigned int nr_free_pagecache_pages(void)
2069 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2072 static inline void show_node(struct zone *zone)
2074 if (NUMA_BUILD)
2075 printk("Node %d ", zone_to_nid(zone));
2078 void si_meminfo(struct sysinfo *val)
2080 val->totalram = totalram_pages;
2081 val->sharedram = 0;
2082 val->freeram = global_page_state(NR_FREE_PAGES);
2083 val->bufferram = nr_blockdev_pages();
2084 val->totalhigh = totalhigh_pages;
2085 val->freehigh = nr_free_highpages();
2086 val->mem_unit = PAGE_SIZE;
2089 EXPORT_SYMBOL(si_meminfo);
2091 #ifdef CONFIG_NUMA
2092 void si_meminfo_node(struct sysinfo *val, int nid)
2094 pg_data_t *pgdat = NODE_DATA(nid);
2096 val->totalram = pgdat->node_present_pages;
2097 val->freeram = node_page_state(nid, NR_FREE_PAGES);
2098 #ifdef CONFIG_HIGHMEM
2099 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
2100 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2101 NR_FREE_PAGES);
2102 #else
2103 val->totalhigh = 0;
2104 val->freehigh = 0;
2105 #endif
2106 val->mem_unit = PAGE_SIZE;
2108 #endif
2110 #define K(x) ((x) << (PAGE_SHIFT-10))
2113 * Show free area list (used inside shift_scroll-lock stuff)
2114 * We also calculate the percentage fragmentation. We do this by counting the
2115 * memory on each free list with the exception of the first item on the list.
2117 void show_free_areas(void)
2119 int cpu;
2120 struct zone *zone;
2122 for_each_populated_zone(zone) {
2123 show_node(zone);
2124 printk("%s per-cpu:\n", zone->name);
2126 for_each_online_cpu(cpu) {
2127 struct per_cpu_pageset *pageset;
2129 pageset = zone_pcp(zone, cpu);
2131 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2132 cpu, pageset->pcp.high,
2133 pageset->pcp.batch, pageset->pcp.count);
2137 printk("Active_anon:%lu active_file:%lu inactive_anon:%lu\n"
2138 " inactive_file:%lu"
2139 " unevictable:%lu"
2140 " dirty:%lu writeback:%lu unstable:%lu\n"
2141 " free:%lu slab:%lu mapped:%lu pagetables:%lu bounce:%lu\n",
2142 global_page_state(NR_ACTIVE_ANON),
2143 global_page_state(NR_ACTIVE_FILE),
2144 global_page_state(NR_INACTIVE_ANON),
2145 global_page_state(NR_INACTIVE_FILE),
2146 global_page_state(NR_UNEVICTABLE),
2147 global_page_state(NR_FILE_DIRTY),
2148 global_page_state(NR_WRITEBACK),
2149 global_page_state(NR_UNSTABLE_NFS),
2150 global_page_state(NR_FREE_PAGES),
2151 global_page_state(NR_SLAB_RECLAIMABLE) +
2152 global_page_state(NR_SLAB_UNRECLAIMABLE),
2153 global_page_state(NR_FILE_MAPPED),
2154 global_page_state(NR_PAGETABLE),
2155 global_page_state(NR_BOUNCE));
2157 for_each_populated_zone(zone) {
2158 int i;
2160 show_node(zone);
2161 printk("%s"
2162 " free:%lukB"
2163 " min:%lukB"
2164 " low:%lukB"
2165 " high:%lukB"
2166 " active_anon:%lukB"
2167 " inactive_anon:%lukB"
2168 " active_file:%lukB"
2169 " inactive_file:%lukB"
2170 " unevictable:%lukB"
2171 " present:%lukB"
2172 " pages_scanned:%lu"
2173 " all_unreclaimable? %s"
2174 "\n",
2175 zone->name,
2176 K(zone_page_state(zone, NR_FREE_PAGES)),
2177 K(min_wmark_pages(zone)),
2178 K(low_wmark_pages(zone)),
2179 K(high_wmark_pages(zone)),
2180 K(zone_page_state(zone, NR_ACTIVE_ANON)),
2181 K(zone_page_state(zone, NR_INACTIVE_ANON)),
2182 K(zone_page_state(zone, NR_ACTIVE_FILE)),
2183 K(zone_page_state(zone, NR_INACTIVE_FILE)),
2184 K(zone_page_state(zone, NR_UNEVICTABLE)),
2185 K(zone->present_pages),
2186 zone->pages_scanned,
2187 (zone_is_all_unreclaimable(zone) ? "yes" : "no")
2189 printk("lowmem_reserve[]:");
2190 for (i = 0; i < MAX_NR_ZONES; i++)
2191 printk(" %lu", zone->lowmem_reserve[i]);
2192 printk("\n");
2195 for_each_populated_zone(zone) {
2196 unsigned long nr[MAX_ORDER], flags, order, total = 0;
2198 show_node(zone);
2199 printk("%s: ", zone->name);
2201 spin_lock_irqsave(&zone->lock, flags);
2202 for (order = 0; order < MAX_ORDER; order++) {
2203 nr[order] = zone->free_area[order].nr_free;
2204 total += nr[order] << order;
2206 spin_unlock_irqrestore(&zone->lock, flags);
2207 for (order = 0; order < MAX_ORDER; order++)
2208 printk("%lu*%lukB ", nr[order], K(1UL) << order);
2209 printk("= %lukB\n", K(total));
2212 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
2214 show_swap_cache_info();
2217 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
2219 zoneref->zone = zone;
2220 zoneref->zone_idx = zone_idx(zone);
2224 * Builds allocation fallback zone lists.
2226 * Add all populated zones of a node to the zonelist.
2228 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
2229 int nr_zones, enum zone_type zone_type)
2231 struct zone *zone;
2233 BUG_ON(zone_type >= MAX_NR_ZONES);
2234 zone_type++;
2236 do {
2237 zone_type--;
2238 zone = pgdat->node_zones + zone_type;
2239 if (populated_zone(zone)) {
2240 zoneref_set_zone(zone,
2241 &zonelist->_zonerefs[nr_zones++]);
2242 check_highest_zone(zone_type);
2245 } while (zone_type);
2246 return nr_zones;
2251 * zonelist_order:
2252 * 0 = automatic detection of better ordering.
2253 * 1 = order by ([node] distance, -zonetype)
2254 * 2 = order by (-zonetype, [node] distance)
2256 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
2257 * the same zonelist. So only NUMA can configure this param.
2259 #define ZONELIST_ORDER_DEFAULT 0
2260 #define ZONELIST_ORDER_NODE 1
2261 #define ZONELIST_ORDER_ZONE 2
2263 /* zonelist order in the kernel.
2264 * set_zonelist_order() will set this to NODE or ZONE.
2266 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
2267 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
2270 #ifdef CONFIG_NUMA
2271 /* The value user specified ....changed by config */
2272 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2273 /* string for sysctl */
2274 #define NUMA_ZONELIST_ORDER_LEN 16
2275 char numa_zonelist_order[16] = "default";
2278 * interface for configure zonelist ordering.
2279 * command line option "numa_zonelist_order"
2280 * = "[dD]efault - default, automatic configuration.
2281 * = "[nN]ode - order by node locality, then by zone within node
2282 * = "[zZ]one - order by zone, then by locality within zone
2285 static int __parse_numa_zonelist_order(char *s)
2287 if (*s == 'd' || *s == 'D') {
2288 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2289 } else if (*s == 'n' || *s == 'N') {
2290 user_zonelist_order = ZONELIST_ORDER_NODE;
2291 } else if (*s == 'z' || *s == 'Z') {
2292 user_zonelist_order = ZONELIST_ORDER_ZONE;
2293 } else {
2294 printk(KERN_WARNING
2295 "Ignoring invalid numa_zonelist_order value: "
2296 "%s\n", s);
2297 return -EINVAL;
2299 return 0;
2302 static __init int setup_numa_zonelist_order(char *s)
2304 if (s)
2305 return __parse_numa_zonelist_order(s);
2306 return 0;
2308 early_param("numa_zonelist_order", setup_numa_zonelist_order);
2311 * sysctl handler for numa_zonelist_order
2313 int numa_zonelist_order_handler(ctl_table *table, int write,
2314 struct file *file, void __user *buffer, size_t *length,
2315 loff_t *ppos)
2317 char saved_string[NUMA_ZONELIST_ORDER_LEN];
2318 int ret;
2320 if (write)
2321 strncpy(saved_string, (char*)table->data,
2322 NUMA_ZONELIST_ORDER_LEN);
2323 ret = proc_dostring(table, write, file, buffer, length, ppos);
2324 if (ret)
2325 return ret;
2326 if (write) {
2327 int oldval = user_zonelist_order;
2328 if (__parse_numa_zonelist_order((char*)table->data)) {
2330 * bogus value. restore saved string
2332 strncpy((char*)table->data, saved_string,
2333 NUMA_ZONELIST_ORDER_LEN);
2334 user_zonelist_order = oldval;
2335 } else if (oldval != user_zonelist_order)
2336 build_all_zonelists();
2338 return 0;
2342 #define MAX_NODE_LOAD (nr_online_nodes)
2343 static int node_load[MAX_NUMNODES];
2346 * find_next_best_node - find the next node that should appear in a given node's fallback list
2347 * @node: node whose fallback list we're appending
2348 * @used_node_mask: nodemask_t of already used nodes
2350 * We use a number of factors to determine which is the next node that should
2351 * appear on a given node's fallback list. The node should not have appeared
2352 * already in @node's fallback list, and it should be the next closest node
2353 * according to the distance array (which contains arbitrary distance values
2354 * from each node to each node in the system), and should also prefer nodes
2355 * with no CPUs, since presumably they'll have very little allocation pressure
2356 * on them otherwise.
2357 * It returns -1 if no node is found.
2359 static int find_next_best_node(int node, nodemask_t *used_node_mask)
2361 int n, val;
2362 int min_val = INT_MAX;
2363 int best_node = -1;
2364 const struct cpumask *tmp = cpumask_of_node(0);
2366 /* Use the local node if we haven't already */
2367 if (!node_isset(node, *used_node_mask)) {
2368 node_set(node, *used_node_mask);
2369 return node;
2372 for_each_node_state(n, N_HIGH_MEMORY) {
2374 /* Don't want a node to appear more than once */
2375 if (node_isset(n, *used_node_mask))
2376 continue;
2378 /* Use the distance array to find the distance */
2379 val = node_distance(node, n);
2381 /* Penalize nodes under us ("prefer the next node") */
2382 val += (n < node);
2384 /* Give preference to headless and unused nodes */
2385 tmp = cpumask_of_node(n);
2386 if (!cpumask_empty(tmp))
2387 val += PENALTY_FOR_NODE_WITH_CPUS;
2389 /* Slight preference for less loaded node */
2390 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
2391 val += node_load[n];
2393 if (val < min_val) {
2394 min_val = val;
2395 best_node = n;
2399 if (best_node >= 0)
2400 node_set(best_node, *used_node_mask);
2402 return best_node;
2407 * Build zonelists ordered by node and zones within node.
2408 * This results in maximum locality--normal zone overflows into local
2409 * DMA zone, if any--but risks exhausting DMA zone.
2411 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
2413 int j;
2414 struct zonelist *zonelist;
2416 zonelist = &pgdat->node_zonelists[0];
2417 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
2419 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2420 MAX_NR_ZONES - 1);
2421 zonelist->_zonerefs[j].zone = NULL;
2422 zonelist->_zonerefs[j].zone_idx = 0;
2426 * Build gfp_thisnode zonelists
2428 static void build_thisnode_zonelists(pg_data_t *pgdat)
2430 int j;
2431 struct zonelist *zonelist;
2433 zonelist = &pgdat->node_zonelists[1];
2434 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2435 zonelist->_zonerefs[j].zone = NULL;
2436 zonelist->_zonerefs[j].zone_idx = 0;
2440 * Build zonelists ordered by zone and nodes within zones.
2441 * This results in conserving DMA zone[s] until all Normal memory is
2442 * exhausted, but results in overflowing to remote node while memory
2443 * may still exist in local DMA zone.
2445 static int node_order[MAX_NUMNODES];
2447 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
2449 int pos, j, node;
2450 int zone_type; /* needs to be signed */
2451 struct zone *z;
2452 struct zonelist *zonelist;
2454 zonelist = &pgdat->node_zonelists[0];
2455 pos = 0;
2456 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
2457 for (j = 0; j < nr_nodes; j++) {
2458 node = node_order[j];
2459 z = &NODE_DATA(node)->node_zones[zone_type];
2460 if (populated_zone(z)) {
2461 zoneref_set_zone(z,
2462 &zonelist->_zonerefs[pos++]);
2463 check_highest_zone(zone_type);
2467 zonelist->_zonerefs[pos].zone = NULL;
2468 zonelist->_zonerefs[pos].zone_idx = 0;
2471 static int default_zonelist_order(void)
2473 int nid, zone_type;
2474 unsigned long low_kmem_size,total_size;
2475 struct zone *z;
2476 int average_size;
2478 * ZONE_DMA and ZONE_DMA32 can be very small area in the sytem.
2479 * If they are really small and used heavily, the system can fall
2480 * into OOM very easily.
2481 * This function detect ZONE_DMA/DMA32 size and confgigures zone order.
2483 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
2484 low_kmem_size = 0;
2485 total_size = 0;
2486 for_each_online_node(nid) {
2487 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2488 z = &NODE_DATA(nid)->node_zones[zone_type];
2489 if (populated_zone(z)) {
2490 if (zone_type < ZONE_NORMAL)
2491 low_kmem_size += z->present_pages;
2492 total_size += z->present_pages;
2496 if (!low_kmem_size || /* there are no DMA area. */
2497 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
2498 return ZONELIST_ORDER_NODE;
2500 * look into each node's config.
2501 * If there is a node whose DMA/DMA32 memory is very big area on
2502 * local memory, NODE_ORDER may be suitable.
2504 average_size = total_size /
2505 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1);
2506 for_each_online_node(nid) {
2507 low_kmem_size = 0;
2508 total_size = 0;
2509 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2510 z = &NODE_DATA(nid)->node_zones[zone_type];
2511 if (populated_zone(z)) {
2512 if (zone_type < ZONE_NORMAL)
2513 low_kmem_size += z->present_pages;
2514 total_size += z->present_pages;
2517 if (low_kmem_size &&
2518 total_size > average_size && /* ignore small node */
2519 low_kmem_size > total_size * 70/100)
2520 return ZONELIST_ORDER_NODE;
2522 return ZONELIST_ORDER_ZONE;
2525 static void set_zonelist_order(void)
2527 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
2528 current_zonelist_order = default_zonelist_order();
2529 else
2530 current_zonelist_order = user_zonelist_order;
2533 static void build_zonelists(pg_data_t *pgdat)
2535 int j, node, load;
2536 enum zone_type i;
2537 nodemask_t used_mask;
2538 int local_node, prev_node;
2539 struct zonelist *zonelist;
2540 int order = current_zonelist_order;
2542 /* initialize zonelists */
2543 for (i = 0; i < MAX_ZONELISTS; i++) {
2544 zonelist = pgdat->node_zonelists + i;
2545 zonelist->_zonerefs[0].zone = NULL;
2546 zonelist->_zonerefs[0].zone_idx = 0;
2549 /* NUMA-aware ordering of nodes */
2550 local_node = pgdat->node_id;
2551 load = nr_online_nodes;
2552 prev_node = local_node;
2553 nodes_clear(used_mask);
2555 memset(node_order, 0, sizeof(node_order));
2556 j = 0;
2558 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
2559 int distance = node_distance(local_node, node);
2562 * If another node is sufficiently far away then it is better
2563 * to reclaim pages in a zone before going off node.
2565 if (distance > RECLAIM_DISTANCE)
2566 zone_reclaim_mode = 1;
2569 * We don't want to pressure a particular node.
2570 * So adding penalty to the first node in same
2571 * distance group to make it round-robin.
2573 if (distance != node_distance(local_node, prev_node))
2574 node_load[node] = load;
2576 prev_node = node;
2577 load--;
2578 if (order == ZONELIST_ORDER_NODE)
2579 build_zonelists_in_node_order(pgdat, node);
2580 else
2581 node_order[j++] = node; /* remember order */
2584 if (order == ZONELIST_ORDER_ZONE) {
2585 /* calculate node order -- i.e., DMA last! */
2586 build_zonelists_in_zone_order(pgdat, j);
2589 build_thisnode_zonelists(pgdat);
2592 /* Construct the zonelist performance cache - see further mmzone.h */
2593 static void build_zonelist_cache(pg_data_t *pgdat)
2595 struct zonelist *zonelist;
2596 struct zonelist_cache *zlc;
2597 struct zoneref *z;
2599 zonelist = &pgdat->node_zonelists[0];
2600 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
2601 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
2602 for (z = zonelist->_zonerefs; z->zone; z++)
2603 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
2607 #else /* CONFIG_NUMA */
2609 static void set_zonelist_order(void)
2611 current_zonelist_order = ZONELIST_ORDER_ZONE;
2614 static void build_zonelists(pg_data_t *pgdat)
2616 int node, local_node;
2617 enum zone_type j;
2618 struct zonelist *zonelist;
2620 local_node = pgdat->node_id;
2622 zonelist = &pgdat->node_zonelists[0];
2623 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2626 * Now we build the zonelist so that it contains the zones
2627 * of all the other nodes.
2628 * We don't want to pressure a particular node, so when
2629 * building the zones for node N, we make sure that the
2630 * zones coming right after the local ones are those from
2631 * node N+1 (modulo N)
2633 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
2634 if (!node_online(node))
2635 continue;
2636 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2637 MAX_NR_ZONES - 1);
2639 for (node = 0; node < local_node; node++) {
2640 if (!node_online(node))
2641 continue;
2642 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2643 MAX_NR_ZONES - 1);
2646 zonelist->_zonerefs[j].zone = NULL;
2647 zonelist->_zonerefs[j].zone_idx = 0;
2650 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
2651 static void build_zonelist_cache(pg_data_t *pgdat)
2653 pgdat->node_zonelists[0].zlcache_ptr = NULL;
2656 #endif /* CONFIG_NUMA */
2658 /* return values int ....just for stop_machine() */
2659 static int __build_all_zonelists(void *dummy)
2661 int nid;
2663 #ifdef CONFIG_NUMA
2664 memset(node_load, 0, sizeof(node_load));
2665 #endif
2666 for_each_online_node(nid) {
2667 pg_data_t *pgdat = NODE_DATA(nid);
2669 build_zonelists(pgdat);
2670 build_zonelist_cache(pgdat);
2672 return 0;
2675 void build_all_zonelists(void)
2677 set_zonelist_order();
2679 if (system_state == SYSTEM_BOOTING) {
2680 __build_all_zonelists(NULL);
2681 mminit_verify_zonelist();
2682 cpuset_init_current_mems_allowed();
2683 } else {
2684 /* we have to stop all cpus to guarantee there is no user
2685 of zonelist */
2686 stop_machine(__build_all_zonelists, NULL, NULL);
2687 /* cpuset refresh routine should be here */
2689 vm_total_pages = nr_free_pagecache_pages();
2691 * Disable grouping by mobility if the number of pages in the
2692 * system is too low to allow the mechanism to work. It would be
2693 * more accurate, but expensive to check per-zone. This check is
2694 * made on memory-hotadd so a system can start with mobility
2695 * disabled and enable it later
2697 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
2698 page_group_by_mobility_disabled = 1;
2699 else
2700 page_group_by_mobility_disabled = 0;
2702 printk("Built %i zonelists in %s order, mobility grouping %s. "
2703 "Total pages: %ld\n",
2704 nr_online_nodes,
2705 zonelist_order_name[current_zonelist_order],
2706 page_group_by_mobility_disabled ? "off" : "on",
2707 vm_total_pages);
2708 #ifdef CONFIG_NUMA
2709 printk("Policy zone: %s\n", zone_names[policy_zone]);
2710 #endif
2714 * Helper functions to size the waitqueue hash table.
2715 * Essentially these want to choose hash table sizes sufficiently
2716 * large so that collisions trying to wait on pages are rare.
2717 * But in fact, the number of active page waitqueues on typical
2718 * systems is ridiculously low, less than 200. So this is even
2719 * conservative, even though it seems large.
2721 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
2722 * waitqueues, i.e. the size of the waitq table given the number of pages.
2724 #define PAGES_PER_WAITQUEUE 256
2726 #ifndef CONFIG_MEMORY_HOTPLUG
2727 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
2729 unsigned long size = 1;
2731 pages /= PAGES_PER_WAITQUEUE;
2733 while (size < pages)
2734 size <<= 1;
2737 * Once we have dozens or even hundreds of threads sleeping
2738 * on IO we've got bigger problems than wait queue collision.
2739 * Limit the size of the wait table to a reasonable size.
2741 size = min(size, 4096UL);
2743 return max(size, 4UL);
2745 #else
2747 * A zone's size might be changed by hot-add, so it is not possible to determine
2748 * a suitable size for its wait_table. So we use the maximum size now.
2750 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
2752 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
2753 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
2754 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
2756 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
2757 * or more by the traditional way. (See above). It equals:
2759 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
2760 * ia64(16K page size) : = ( 8G + 4M)byte.
2761 * powerpc (64K page size) : = (32G +16M)byte.
2763 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
2765 return 4096UL;
2767 #endif
2770 * This is an integer logarithm so that shifts can be used later
2771 * to extract the more random high bits from the multiplicative
2772 * hash function before the remainder is taken.
2774 static inline unsigned long wait_table_bits(unsigned long size)
2776 return ffz(~size);
2779 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
2782 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
2783 * of blocks reserved is based on min_wmark_pages(zone). The memory within
2784 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
2785 * higher will lead to a bigger reserve which will get freed as contiguous
2786 * blocks as reclaim kicks in
2788 static void setup_zone_migrate_reserve(struct zone *zone)
2790 unsigned long start_pfn, pfn, end_pfn;
2791 struct page *page;
2792 unsigned long reserve, block_migratetype;
2794 /* Get the start pfn, end pfn and the number of blocks to reserve */
2795 start_pfn = zone->zone_start_pfn;
2796 end_pfn = start_pfn + zone->spanned_pages;
2797 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
2798 pageblock_order;
2800 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
2801 if (!pfn_valid(pfn))
2802 continue;
2803 page = pfn_to_page(pfn);
2805 /* Watch out for overlapping nodes */
2806 if (page_to_nid(page) != zone_to_nid(zone))
2807 continue;
2809 /* Blocks with reserved pages will never free, skip them. */
2810 if (PageReserved(page))
2811 continue;
2813 block_migratetype = get_pageblock_migratetype(page);
2815 /* If this block is reserved, account for it */
2816 if (reserve > 0 && block_migratetype == MIGRATE_RESERVE) {
2817 reserve--;
2818 continue;
2821 /* Suitable for reserving if this block is movable */
2822 if (reserve > 0 && block_migratetype == MIGRATE_MOVABLE) {
2823 set_pageblock_migratetype(page, MIGRATE_RESERVE);
2824 move_freepages_block(zone, page, MIGRATE_RESERVE);
2825 reserve--;
2826 continue;
2830 * If the reserve is met and this is a previous reserved block,
2831 * take it back
2833 if (block_migratetype == MIGRATE_RESERVE) {
2834 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
2835 move_freepages_block(zone, page, MIGRATE_MOVABLE);
2841 * Initially all pages are reserved - free ones are freed
2842 * up by free_all_bootmem() once the early boot process is
2843 * done. Non-atomic initialization, single-pass.
2845 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
2846 unsigned long start_pfn, enum memmap_context context)
2848 struct page *page;
2849 unsigned long end_pfn = start_pfn + size;
2850 unsigned long pfn;
2851 struct zone *z;
2853 if (highest_memmap_pfn < end_pfn - 1)
2854 highest_memmap_pfn = end_pfn - 1;
2856 z = &NODE_DATA(nid)->node_zones[zone];
2857 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
2859 * There can be holes in boot-time mem_map[]s
2860 * handed to this function. They do not
2861 * exist on hotplugged memory.
2863 if (context == MEMMAP_EARLY) {
2864 if (!early_pfn_valid(pfn))
2865 continue;
2866 if (!early_pfn_in_nid(pfn, nid))
2867 continue;
2869 page = pfn_to_page(pfn);
2870 set_page_links(page, zone, nid, pfn);
2871 mminit_verify_page_links(page, zone, nid, pfn);
2872 init_page_count(page);
2873 reset_page_mapcount(page);
2874 SetPageReserved(page);
2876 * Mark the block movable so that blocks are reserved for
2877 * movable at startup. This will force kernel allocations
2878 * to reserve their blocks rather than leaking throughout
2879 * the address space during boot when many long-lived
2880 * kernel allocations are made. Later some blocks near
2881 * the start are marked MIGRATE_RESERVE by
2882 * setup_zone_migrate_reserve()
2884 * bitmap is created for zone's valid pfn range. but memmap
2885 * can be created for invalid pages (for alignment)
2886 * check here not to call set_pageblock_migratetype() against
2887 * pfn out of zone.
2889 if ((z->zone_start_pfn <= pfn)
2890 && (pfn < z->zone_start_pfn + z->spanned_pages)
2891 && !(pfn & (pageblock_nr_pages - 1)))
2892 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
2894 INIT_LIST_HEAD(&page->lru);
2895 #ifdef WANT_PAGE_VIRTUAL
2896 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
2897 if (!is_highmem_idx(zone))
2898 set_page_address(page, __va(pfn << PAGE_SHIFT));
2899 #endif
2903 static void __meminit zone_init_free_lists(struct zone *zone)
2905 int order, t;
2906 for_each_migratetype_order(order, t) {
2907 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
2908 zone->free_area[order].nr_free = 0;
2912 #ifndef __HAVE_ARCH_MEMMAP_INIT
2913 #define memmap_init(size, nid, zone, start_pfn) \
2914 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
2915 #endif
2917 static int zone_batchsize(struct zone *zone)
2919 #ifdef CONFIG_MMU
2920 int batch;
2923 * The per-cpu-pages pools are set to around 1000th of the
2924 * size of the zone. But no more than 1/2 of a meg.
2926 * OK, so we don't know how big the cache is. So guess.
2928 batch = zone->present_pages / 1024;
2929 if (batch * PAGE_SIZE > 512 * 1024)
2930 batch = (512 * 1024) / PAGE_SIZE;
2931 batch /= 4; /* We effectively *= 4 below */
2932 if (batch < 1)
2933 batch = 1;
2936 * Clamp the batch to a 2^n - 1 value. Having a power
2937 * of 2 value was found to be more likely to have
2938 * suboptimal cache aliasing properties in some cases.
2940 * For example if 2 tasks are alternately allocating
2941 * batches of pages, one task can end up with a lot
2942 * of pages of one half of the possible page colors
2943 * and the other with pages of the other colors.
2945 batch = rounddown_pow_of_two(batch + batch/2) - 1;
2947 return batch;
2949 #else
2950 /* The deferral and batching of frees should be suppressed under NOMMU
2951 * conditions.
2953 * The problem is that NOMMU needs to be able to allocate large chunks
2954 * of contiguous memory as there's no hardware page translation to
2955 * assemble apparent contiguous memory from discontiguous pages.
2957 * Queueing large contiguous runs of pages for batching, however,
2958 * causes the pages to actually be freed in smaller chunks. As there
2959 * can be a significant delay between the individual batches being
2960 * recycled, this leads to the once large chunks of space being
2961 * fragmented and becoming unavailable for high-order allocations.
2963 return 0;
2964 #endif
2967 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
2969 struct per_cpu_pages *pcp;
2971 memset(p, 0, sizeof(*p));
2973 pcp = &p->pcp;
2974 pcp->count = 0;
2975 pcp->high = 6 * batch;
2976 pcp->batch = max(1UL, 1 * batch);
2977 INIT_LIST_HEAD(&pcp->list);
2981 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
2982 * to the value high for the pageset p.
2985 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
2986 unsigned long high)
2988 struct per_cpu_pages *pcp;
2990 pcp = &p->pcp;
2991 pcp->high = high;
2992 pcp->batch = max(1UL, high/4);
2993 if ((high/4) > (PAGE_SHIFT * 8))
2994 pcp->batch = PAGE_SHIFT * 8;
2998 #ifdef CONFIG_NUMA
3000 * Boot pageset table. One per cpu which is going to be used for all
3001 * zones and all nodes. The parameters will be set in such a way
3002 * that an item put on a list will immediately be handed over to
3003 * the buddy list. This is safe since pageset manipulation is done
3004 * with interrupts disabled.
3006 * Some NUMA counter updates may also be caught by the boot pagesets.
3008 * The boot_pagesets must be kept even after bootup is complete for
3009 * unused processors and/or zones. They do play a role for bootstrapping
3010 * hotplugged processors.
3012 * zoneinfo_show() and maybe other functions do
3013 * not check if the processor is online before following the pageset pointer.
3014 * Other parts of the kernel may not check if the zone is available.
3016 static struct per_cpu_pageset boot_pageset[NR_CPUS];
3019 * Dynamically allocate memory for the
3020 * per cpu pageset array in struct zone.
3022 static int __cpuinit process_zones(int cpu)
3024 struct zone *zone, *dzone;
3025 int node = cpu_to_node(cpu);
3027 node_set_state(node, N_CPU); /* this node has a cpu */
3029 for_each_populated_zone(zone) {
3030 zone_pcp(zone, cpu) = kmalloc_node(sizeof(struct per_cpu_pageset),
3031 GFP_KERNEL, node);
3032 if (!zone_pcp(zone, cpu))
3033 goto bad;
3035 setup_pageset(zone_pcp(zone, cpu), zone_batchsize(zone));
3037 if (percpu_pagelist_fraction)
3038 setup_pagelist_highmark(zone_pcp(zone, cpu),
3039 (zone->present_pages / percpu_pagelist_fraction));
3042 return 0;
3043 bad:
3044 for_each_zone(dzone) {
3045 if (!populated_zone(dzone))
3046 continue;
3047 if (dzone == zone)
3048 break;
3049 kfree(zone_pcp(dzone, cpu));
3050 zone_pcp(dzone, cpu) = &boot_pageset[cpu];
3052 return -ENOMEM;
3055 static inline void free_zone_pagesets(int cpu)
3057 struct zone *zone;
3059 for_each_zone(zone) {
3060 struct per_cpu_pageset *pset = zone_pcp(zone, cpu);
3062 /* Free per_cpu_pageset if it is slab allocated */
3063 if (pset != &boot_pageset[cpu])
3064 kfree(pset);
3065 zone_pcp(zone, cpu) = &boot_pageset[cpu];
3069 static int __cpuinit pageset_cpuup_callback(struct notifier_block *nfb,
3070 unsigned long action,
3071 void *hcpu)
3073 int cpu = (long)hcpu;
3074 int ret = NOTIFY_OK;
3076 switch (action) {
3077 case CPU_UP_PREPARE:
3078 case CPU_UP_PREPARE_FROZEN:
3079 if (process_zones(cpu))
3080 ret = NOTIFY_BAD;
3081 break;
3082 case CPU_UP_CANCELED:
3083 case CPU_UP_CANCELED_FROZEN:
3084 case CPU_DEAD:
3085 case CPU_DEAD_FROZEN:
3086 free_zone_pagesets(cpu);
3087 break;
3088 default:
3089 break;
3091 return ret;
3094 static struct notifier_block __cpuinitdata pageset_notifier =
3095 { &pageset_cpuup_callback, NULL, 0 };
3097 void __init setup_per_cpu_pageset(void)
3099 int err;
3101 /* Initialize per_cpu_pageset for cpu 0.
3102 * A cpuup callback will do this for every cpu
3103 * as it comes online
3105 err = process_zones(smp_processor_id());
3106 BUG_ON(err);
3107 register_cpu_notifier(&pageset_notifier);
3110 #endif
3112 static noinline __init_refok
3113 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
3115 int i;
3116 struct pglist_data *pgdat = zone->zone_pgdat;
3117 size_t alloc_size;
3120 * The per-page waitqueue mechanism uses hashed waitqueues
3121 * per zone.
3123 zone->wait_table_hash_nr_entries =
3124 wait_table_hash_nr_entries(zone_size_pages);
3125 zone->wait_table_bits =
3126 wait_table_bits(zone->wait_table_hash_nr_entries);
3127 alloc_size = zone->wait_table_hash_nr_entries
3128 * sizeof(wait_queue_head_t);
3130 if (!slab_is_available()) {
3131 zone->wait_table = (wait_queue_head_t *)
3132 alloc_bootmem_node(pgdat, alloc_size);
3133 } else {
3135 * This case means that a zone whose size was 0 gets new memory
3136 * via memory hot-add.
3137 * But it may be the case that a new node was hot-added. In
3138 * this case vmalloc() will not be able to use this new node's
3139 * memory - this wait_table must be initialized to use this new
3140 * node itself as well.
3141 * To use this new node's memory, further consideration will be
3142 * necessary.
3144 zone->wait_table = vmalloc(alloc_size);
3146 if (!zone->wait_table)
3147 return -ENOMEM;
3149 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
3150 init_waitqueue_head(zone->wait_table + i);
3152 return 0;
3155 static int __zone_pcp_update(void *data)
3157 struct zone *zone = data;
3158 int cpu;
3159 unsigned long batch = zone_batchsize(zone), flags;
3161 for (cpu = 0; cpu < NR_CPUS; cpu++) {
3162 struct per_cpu_pageset *pset;
3163 struct per_cpu_pages *pcp;
3165 pset = zone_pcp(zone, cpu);
3166 pcp = &pset->pcp;
3168 local_irq_save(flags);
3169 free_pages_bulk(zone, pcp->count, &pcp->list, 0);
3170 setup_pageset(pset, batch);
3171 local_irq_restore(flags);
3173 return 0;
3176 void zone_pcp_update(struct zone *zone)
3178 stop_machine(__zone_pcp_update, zone, NULL);
3181 static __meminit void zone_pcp_init(struct zone *zone)
3183 int cpu;
3184 unsigned long batch = zone_batchsize(zone);
3186 for (cpu = 0; cpu < NR_CPUS; cpu++) {
3187 #ifdef CONFIG_NUMA
3188 /* Early boot. Slab allocator not functional yet */
3189 zone_pcp(zone, cpu) = &boot_pageset[cpu];
3190 setup_pageset(&boot_pageset[cpu],0);
3191 #else
3192 setup_pageset(zone_pcp(zone,cpu), batch);
3193 #endif
3195 if (zone->present_pages)
3196 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%lu\n",
3197 zone->name, zone->present_pages, batch);
3200 __meminit int init_currently_empty_zone(struct zone *zone,
3201 unsigned long zone_start_pfn,
3202 unsigned long size,
3203 enum memmap_context context)
3205 struct pglist_data *pgdat = zone->zone_pgdat;
3206 int ret;
3207 ret = zone_wait_table_init(zone, size);
3208 if (ret)
3209 return ret;
3210 pgdat->nr_zones = zone_idx(zone) + 1;
3212 zone->zone_start_pfn = zone_start_pfn;
3214 mminit_dprintk(MMINIT_TRACE, "memmap_init",
3215 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
3216 pgdat->node_id,
3217 (unsigned long)zone_idx(zone),
3218 zone_start_pfn, (zone_start_pfn + size));
3220 zone_init_free_lists(zone);
3222 return 0;
3225 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3227 * Basic iterator support. Return the first range of PFNs for a node
3228 * Note: nid == MAX_NUMNODES returns first region regardless of node
3230 static int __meminit first_active_region_index_in_nid(int nid)
3232 int i;
3234 for (i = 0; i < nr_nodemap_entries; i++)
3235 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
3236 return i;
3238 return -1;
3242 * Basic iterator support. Return the next active range of PFNs for a node
3243 * Note: nid == MAX_NUMNODES returns next region regardless of node
3245 static int __meminit next_active_region_index_in_nid(int index, int nid)
3247 for (index = index + 1; index < nr_nodemap_entries; index++)
3248 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
3249 return index;
3251 return -1;
3254 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
3256 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
3257 * Architectures may implement their own version but if add_active_range()
3258 * was used and there are no special requirements, this is a convenient
3259 * alternative
3261 int __meminit __early_pfn_to_nid(unsigned long pfn)
3263 int i;
3265 for (i = 0; i < nr_nodemap_entries; i++) {
3266 unsigned long start_pfn = early_node_map[i].start_pfn;
3267 unsigned long end_pfn = early_node_map[i].end_pfn;
3269 if (start_pfn <= pfn && pfn < end_pfn)
3270 return early_node_map[i].nid;
3272 /* This is a memory hole */
3273 return -1;
3275 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
3277 int __meminit early_pfn_to_nid(unsigned long pfn)
3279 int nid;
3281 nid = __early_pfn_to_nid(pfn);
3282 if (nid >= 0)
3283 return nid;
3284 /* just returns 0 */
3285 return 0;
3288 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
3289 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
3291 int nid;
3293 nid = __early_pfn_to_nid(pfn);
3294 if (nid >= 0 && nid != node)
3295 return false;
3296 return true;
3298 #endif
3300 /* Basic iterator support to walk early_node_map[] */
3301 #define for_each_active_range_index_in_nid(i, nid) \
3302 for (i = first_active_region_index_in_nid(nid); i != -1; \
3303 i = next_active_region_index_in_nid(i, nid))
3306 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
3307 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
3308 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
3310 * If an architecture guarantees that all ranges registered with
3311 * add_active_ranges() contain no holes and may be freed, this
3312 * this function may be used instead of calling free_bootmem() manually.
3314 void __init free_bootmem_with_active_regions(int nid,
3315 unsigned long max_low_pfn)
3317 int i;
3319 for_each_active_range_index_in_nid(i, nid) {
3320 unsigned long size_pages = 0;
3321 unsigned long end_pfn = early_node_map[i].end_pfn;
3323 if (early_node_map[i].start_pfn >= max_low_pfn)
3324 continue;
3326 if (end_pfn > max_low_pfn)
3327 end_pfn = max_low_pfn;
3329 size_pages = end_pfn - early_node_map[i].start_pfn;
3330 free_bootmem_node(NODE_DATA(early_node_map[i].nid),
3331 PFN_PHYS(early_node_map[i].start_pfn),
3332 size_pages << PAGE_SHIFT);
3336 void __init work_with_active_regions(int nid, work_fn_t work_fn, void *data)
3338 int i;
3339 int ret;
3341 for_each_active_range_index_in_nid(i, nid) {
3342 ret = work_fn(early_node_map[i].start_pfn,
3343 early_node_map[i].end_pfn, data);
3344 if (ret)
3345 break;
3349 * sparse_memory_present_with_active_regions - Call memory_present for each active range
3350 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
3352 * If an architecture guarantees that all ranges registered with
3353 * add_active_ranges() contain no holes and may be freed, this
3354 * function may be used instead of calling memory_present() manually.
3356 void __init sparse_memory_present_with_active_regions(int nid)
3358 int i;
3360 for_each_active_range_index_in_nid(i, nid)
3361 memory_present(early_node_map[i].nid,
3362 early_node_map[i].start_pfn,
3363 early_node_map[i].end_pfn);
3367 * get_pfn_range_for_nid - Return the start and end page frames for a node
3368 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
3369 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
3370 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
3372 * It returns the start and end page frame of a node based on information
3373 * provided by an arch calling add_active_range(). If called for a node
3374 * with no available memory, a warning is printed and the start and end
3375 * PFNs will be 0.
3377 void __meminit get_pfn_range_for_nid(unsigned int nid,
3378 unsigned long *start_pfn, unsigned long *end_pfn)
3380 int i;
3381 *start_pfn = -1UL;
3382 *end_pfn = 0;
3384 for_each_active_range_index_in_nid(i, nid) {
3385 *start_pfn = min(*start_pfn, early_node_map[i].start_pfn);
3386 *end_pfn = max(*end_pfn, early_node_map[i].end_pfn);
3389 if (*start_pfn == -1UL)
3390 *start_pfn = 0;
3394 * This finds a zone that can be used for ZONE_MOVABLE pages. The
3395 * assumption is made that zones within a node are ordered in monotonic
3396 * increasing memory addresses so that the "highest" populated zone is used
3398 static void __init find_usable_zone_for_movable(void)
3400 int zone_index;
3401 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
3402 if (zone_index == ZONE_MOVABLE)
3403 continue;
3405 if (arch_zone_highest_possible_pfn[zone_index] >
3406 arch_zone_lowest_possible_pfn[zone_index])
3407 break;
3410 VM_BUG_ON(zone_index == -1);
3411 movable_zone = zone_index;
3415 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
3416 * because it is sized independant of architecture. Unlike the other zones,
3417 * the starting point for ZONE_MOVABLE is not fixed. It may be different
3418 * in each node depending on the size of each node and how evenly kernelcore
3419 * is distributed. This helper function adjusts the zone ranges
3420 * provided by the architecture for a given node by using the end of the
3421 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
3422 * zones within a node are in order of monotonic increases memory addresses
3424 static void __meminit adjust_zone_range_for_zone_movable(int nid,
3425 unsigned long zone_type,
3426 unsigned long node_start_pfn,
3427 unsigned long node_end_pfn,
3428 unsigned long *zone_start_pfn,
3429 unsigned long *zone_end_pfn)
3431 /* Only adjust if ZONE_MOVABLE is on this node */
3432 if (zone_movable_pfn[nid]) {
3433 /* Size ZONE_MOVABLE */
3434 if (zone_type == ZONE_MOVABLE) {
3435 *zone_start_pfn = zone_movable_pfn[nid];
3436 *zone_end_pfn = min(node_end_pfn,
3437 arch_zone_highest_possible_pfn[movable_zone]);
3439 /* Adjust for ZONE_MOVABLE starting within this range */
3440 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
3441 *zone_end_pfn > zone_movable_pfn[nid]) {
3442 *zone_end_pfn = zone_movable_pfn[nid];
3444 /* Check if this whole range is within ZONE_MOVABLE */
3445 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
3446 *zone_start_pfn = *zone_end_pfn;
3451 * Return the number of pages a zone spans in a node, including holes
3452 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
3454 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
3455 unsigned long zone_type,
3456 unsigned long *ignored)
3458 unsigned long node_start_pfn, node_end_pfn;
3459 unsigned long zone_start_pfn, zone_end_pfn;
3461 /* Get the start and end of the node and zone */
3462 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
3463 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
3464 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
3465 adjust_zone_range_for_zone_movable(nid, zone_type,
3466 node_start_pfn, node_end_pfn,
3467 &zone_start_pfn, &zone_end_pfn);
3469 /* Check that this node has pages within the zone's required range */
3470 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
3471 return 0;
3473 /* Move the zone boundaries inside the node if necessary */
3474 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
3475 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
3477 /* Return the spanned pages */
3478 return zone_end_pfn - zone_start_pfn;
3482 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
3483 * then all holes in the requested range will be accounted for.
3485 static unsigned long __meminit __absent_pages_in_range(int nid,
3486 unsigned long range_start_pfn,
3487 unsigned long range_end_pfn)
3489 int i = 0;
3490 unsigned long prev_end_pfn = 0, hole_pages = 0;
3491 unsigned long start_pfn;
3493 /* Find the end_pfn of the first active range of pfns in the node */
3494 i = first_active_region_index_in_nid(nid);
3495 if (i == -1)
3496 return 0;
3498 prev_end_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
3500 /* Account for ranges before physical memory on this node */
3501 if (early_node_map[i].start_pfn > range_start_pfn)
3502 hole_pages = prev_end_pfn - range_start_pfn;
3504 /* Find all holes for the zone within the node */
3505 for (; i != -1; i = next_active_region_index_in_nid(i, nid)) {
3507 /* No need to continue if prev_end_pfn is outside the zone */
3508 if (prev_end_pfn >= range_end_pfn)
3509 break;
3511 /* Make sure the end of the zone is not within the hole */
3512 start_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
3513 prev_end_pfn = max(prev_end_pfn, range_start_pfn);
3515 /* Update the hole size cound and move on */
3516 if (start_pfn > range_start_pfn) {
3517 BUG_ON(prev_end_pfn > start_pfn);
3518 hole_pages += start_pfn - prev_end_pfn;
3520 prev_end_pfn = early_node_map[i].end_pfn;
3523 /* Account for ranges past physical memory on this node */
3524 if (range_end_pfn > prev_end_pfn)
3525 hole_pages += range_end_pfn -
3526 max(range_start_pfn, prev_end_pfn);
3528 return hole_pages;
3532 * absent_pages_in_range - Return number of page frames in holes within a range
3533 * @start_pfn: The start PFN to start searching for holes
3534 * @end_pfn: The end PFN to stop searching for holes
3536 * It returns the number of pages frames in memory holes within a range.
3538 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
3539 unsigned long end_pfn)
3541 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
3544 /* Return the number of page frames in holes in a zone on a node */
3545 static unsigned long __meminit zone_absent_pages_in_node(int nid,
3546 unsigned long zone_type,
3547 unsigned long *ignored)
3549 unsigned long node_start_pfn, node_end_pfn;
3550 unsigned long zone_start_pfn, zone_end_pfn;
3552 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
3553 zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type],
3554 node_start_pfn);
3555 zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type],
3556 node_end_pfn);
3558 adjust_zone_range_for_zone_movable(nid, zone_type,
3559 node_start_pfn, node_end_pfn,
3560 &zone_start_pfn, &zone_end_pfn);
3561 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
3564 #else
3565 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
3566 unsigned long zone_type,
3567 unsigned long *zones_size)
3569 return zones_size[zone_type];
3572 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
3573 unsigned long zone_type,
3574 unsigned long *zholes_size)
3576 if (!zholes_size)
3577 return 0;
3579 return zholes_size[zone_type];
3582 #endif
3584 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
3585 unsigned long *zones_size, unsigned long *zholes_size)
3587 unsigned long realtotalpages, totalpages = 0;
3588 enum zone_type i;
3590 for (i = 0; i < MAX_NR_ZONES; i++)
3591 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
3592 zones_size);
3593 pgdat->node_spanned_pages = totalpages;
3595 realtotalpages = totalpages;
3596 for (i = 0; i < MAX_NR_ZONES; i++)
3597 realtotalpages -=
3598 zone_absent_pages_in_node(pgdat->node_id, i,
3599 zholes_size);
3600 pgdat->node_present_pages = realtotalpages;
3601 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
3602 realtotalpages);
3605 #ifndef CONFIG_SPARSEMEM
3607 * Calculate the size of the zone->blockflags rounded to an unsigned long
3608 * Start by making sure zonesize is a multiple of pageblock_order by rounding
3609 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
3610 * round what is now in bits to nearest long in bits, then return it in
3611 * bytes.
3613 static unsigned long __init usemap_size(unsigned long zonesize)
3615 unsigned long usemapsize;
3617 usemapsize = roundup(zonesize, pageblock_nr_pages);
3618 usemapsize = usemapsize >> pageblock_order;
3619 usemapsize *= NR_PAGEBLOCK_BITS;
3620 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
3622 return usemapsize / 8;
3625 static void __init setup_usemap(struct pglist_data *pgdat,
3626 struct zone *zone, unsigned long zonesize)
3628 unsigned long usemapsize = usemap_size(zonesize);
3629 zone->pageblock_flags = NULL;
3630 if (usemapsize)
3631 zone->pageblock_flags = alloc_bootmem_node(pgdat, usemapsize);
3633 #else
3634 static void inline setup_usemap(struct pglist_data *pgdat,
3635 struct zone *zone, unsigned long zonesize) {}
3636 #endif /* CONFIG_SPARSEMEM */
3638 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
3640 /* Return a sensible default order for the pageblock size. */
3641 static inline int pageblock_default_order(void)
3643 if (HPAGE_SHIFT > PAGE_SHIFT)
3644 return HUGETLB_PAGE_ORDER;
3646 return MAX_ORDER-1;
3649 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
3650 static inline void __init set_pageblock_order(unsigned int order)
3652 /* Check that pageblock_nr_pages has not already been setup */
3653 if (pageblock_order)
3654 return;
3657 * Assume the largest contiguous order of interest is a huge page.
3658 * This value may be variable depending on boot parameters on IA64
3660 pageblock_order = order;
3662 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
3665 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
3666 * and pageblock_default_order() are unused as pageblock_order is set
3667 * at compile-time. See include/linux/pageblock-flags.h for the values of
3668 * pageblock_order based on the kernel config
3670 static inline int pageblock_default_order(unsigned int order)
3672 return MAX_ORDER-1;
3674 #define set_pageblock_order(x) do {} while (0)
3676 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
3679 * Set up the zone data structures:
3680 * - mark all pages reserved
3681 * - mark all memory queues empty
3682 * - clear the memory bitmaps
3684 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
3685 unsigned long *zones_size, unsigned long *zholes_size)
3687 enum zone_type j;
3688 int nid = pgdat->node_id;
3689 unsigned long zone_start_pfn = pgdat->node_start_pfn;
3690 int ret;
3692 pgdat_resize_init(pgdat);
3693 pgdat->nr_zones = 0;
3694 init_waitqueue_head(&pgdat->kswapd_wait);
3695 pgdat->kswapd_max_order = 0;
3696 pgdat_page_cgroup_init(pgdat);
3698 for (j = 0; j < MAX_NR_ZONES; j++) {
3699 struct zone *zone = pgdat->node_zones + j;
3700 unsigned long size, realsize, memmap_pages;
3701 enum lru_list l;
3703 size = zone_spanned_pages_in_node(nid, j, zones_size);
3704 realsize = size - zone_absent_pages_in_node(nid, j,
3705 zholes_size);
3708 * Adjust realsize so that it accounts for how much memory
3709 * is used by this zone for memmap. This affects the watermark
3710 * and per-cpu initialisations
3712 memmap_pages =
3713 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT;
3714 if (realsize >= memmap_pages) {
3715 realsize -= memmap_pages;
3716 if (memmap_pages)
3717 printk(KERN_DEBUG
3718 " %s zone: %lu pages used for memmap\n",
3719 zone_names[j], memmap_pages);
3720 } else
3721 printk(KERN_WARNING
3722 " %s zone: %lu pages exceeds realsize %lu\n",
3723 zone_names[j], memmap_pages, realsize);
3725 /* Account for reserved pages */
3726 if (j == 0 && realsize > dma_reserve) {
3727 realsize -= dma_reserve;
3728 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
3729 zone_names[0], dma_reserve);
3732 if (!is_highmem_idx(j))
3733 nr_kernel_pages += realsize;
3734 nr_all_pages += realsize;
3736 zone->spanned_pages = size;
3737 zone->present_pages = realsize;
3738 #ifdef CONFIG_NUMA
3739 zone->node = nid;
3740 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
3741 / 100;
3742 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
3743 #endif
3744 zone->name = zone_names[j];
3745 spin_lock_init(&zone->lock);
3746 spin_lock_init(&zone->lru_lock);
3747 zone_seqlock_init(zone);
3748 zone->zone_pgdat = pgdat;
3750 zone->prev_priority = DEF_PRIORITY;
3752 zone_pcp_init(zone);
3753 for_each_lru(l) {
3754 INIT_LIST_HEAD(&zone->lru[l].list);
3755 zone->lru[l].nr_saved_scan = 0;
3757 zone->reclaim_stat.recent_rotated[0] = 0;
3758 zone->reclaim_stat.recent_rotated[1] = 0;
3759 zone->reclaim_stat.recent_scanned[0] = 0;
3760 zone->reclaim_stat.recent_scanned[1] = 0;
3761 zap_zone_vm_stats(zone);
3762 zone->flags = 0;
3763 if (!size)
3764 continue;
3766 set_pageblock_order(pageblock_default_order());
3767 setup_usemap(pgdat, zone, size);
3768 ret = init_currently_empty_zone(zone, zone_start_pfn,
3769 size, MEMMAP_EARLY);
3770 BUG_ON(ret);
3771 memmap_init(size, nid, j, zone_start_pfn);
3772 zone_start_pfn += size;
3776 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
3778 /* Skip empty nodes */
3779 if (!pgdat->node_spanned_pages)
3780 return;
3782 #ifdef CONFIG_FLAT_NODE_MEM_MAP
3783 /* ia64 gets its own node_mem_map, before this, without bootmem */
3784 if (!pgdat->node_mem_map) {
3785 unsigned long size, start, end;
3786 struct page *map;
3789 * The zone's endpoints aren't required to be MAX_ORDER
3790 * aligned but the node_mem_map endpoints must be in order
3791 * for the buddy allocator to function correctly.
3793 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
3794 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
3795 end = ALIGN(end, MAX_ORDER_NR_PAGES);
3796 size = (end - start) * sizeof(struct page);
3797 map = alloc_remap(pgdat->node_id, size);
3798 if (!map)
3799 map = alloc_bootmem_node(pgdat, size);
3800 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
3802 #ifndef CONFIG_NEED_MULTIPLE_NODES
3804 * With no DISCONTIG, the global mem_map is just set as node 0's
3806 if (pgdat == NODE_DATA(0)) {
3807 mem_map = NODE_DATA(0)->node_mem_map;
3808 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3809 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
3810 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
3811 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
3813 #endif
3814 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
3817 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
3818 unsigned long node_start_pfn, unsigned long *zholes_size)
3820 pg_data_t *pgdat = NODE_DATA(nid);
3822 pgdat->node_id = nid;
3823 pgdat->node_start_pfn = node_start_pfn;
3824 calculate_node_totalpages(pgdat, zones_size, zholes_size);
3826 alloc_node_mem_map(pgdat);
3827 #ifdef CONFIG_FLAT_NODE_MEM_MAP
3828 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
3829 nid, (unsigned long)pgdat,
3830 (unsigned long)pgdat->node_mem_map);
3831 #endif
3833 free_area_init_core(pgdat, zones_size, zholes_size);
3836 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3838 #if MAX_NUMNODES > 1
3840 * Figure out the number of possible node ids.
3842 static void __init setup_nr_node_ids(void)
3844 unsigned int node;
3845 unsigned int highest = 0;
3847 for_each_node_mask(node, node_possible_map)
3848 highest = node;
3849 nr_node_ids = highest + 1;
3851 #else
3852 static inline void setup_nr_node_ids(void)
3855 #endif
3858 * add_active_range - Register a range of PFNs backed by physical memory
3859 * @nid: The node ID the range resides on
3860 * @start_pfn: The start PFN of the available physical memory
3861 * @end_pfn: The end PFN of the available physical memory
3863 * These ranges are stored in an early_node_map[] and later used by
3864 * free_area_init_nodes() to calculate zone sizes and holes. If the
3865 * range spans a memory hole, it is up to the architecture to ensure
3866 * the memory is not freed by the bootmem allocator. If possible
3867 * the range being registered will be merged with existing ranges.
3869 void __init add_active_range(unsigned int nid, unsigned long start_pfn,
3870 unsigned long end_pfn)
3872 int i;
3874 mminit_dprintk(MMINIT_TRACE, "memory_register",
3875 "Entering add_active_range(%d, %#lx, %#lx) "
3876 "%d entries of %d used\n",
3877 nid, start_pfn, end_pfn,
3878 nr_nodemap_entries, MAX_ACTIVE_REGIONS);
3880 mminit_validate_memmodel_limits(&start_pfn, &end_pfn);
3882 /* Merge with existing active regions if possible */
3883 for (i = 0; i < nr_nodemap_entries; i++) {
3884 if (early_node_map[i].nid != nid)
3885 continue;
3887 /* Skip if an existing region covers this new one */
3888 if (start_pfn >= early_node_map[i].start_pfn &&
3889 end_pfn <= early_node_map[i].end_pfn)
3890 return;
3892 /* Merge forward if suitable */
3893 if (start_pfn <= early_node_map[i].end_pfn &&
3894 end_pfn > early_node_map[i].end_pfn) {
3895 early_node_map[i].end_pfn = end_pfn;
3896 return;
3899 /* Merge backward if suitable */
3900 if (start_pfn < early_node_map[i].end_pfn &&
3901 end_pfn >= early_node_map[i].start_pfn) {
3902 early_node_map[i].start_pfn = start_pfn;
3903 return;
3907 /* Check that early_node_map is large enough */
3908 if (i >= MAX_ACTIVE_REGIONS) {
3909 printk(KERN_CRIT "More than %d memory regions, truncating\n",
3910 MAX_ACTIVE_REGIONS);
3911 return;
3914 early_node_map[i].nid = nid;
3915 early_node_map[i].start_pfn = start_pfn;
3916 early_node_map[i].end_pfn = end_pfn;
3917 nr_nodemap_entries = i + 1;
3921 * remove_active_range - Shrink an existing registered range of PFNs
3922 * @nid: The node id the range is on that should be shrunk
3923 * @start_pfn: The new PFN of the range
3924 * @end_pfn: The new PFN of the range
3926 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
3927 * The map is kept near the end physical page range that has already been
3928 * registered. This function allows an arch to shrink an existing registered
3929 * range.
3931 void __init remove_active_range(unsigned int nid, unsigned long start_pfn,
3932 unsigned long end_pfn)
3934 int i, j;
3935 int removed = 0;
3937 printk(KERN_DEBUG "remove_active_range (%d, %lu, %lu)\n",
3938 nid, start_pfn, end_pfn);
3940 /* Find the old active region end and shrink */
3941 for_each_active_range_index_in_nid(i, nid) {
3942 if (early_node_map[i].start_pfn >= start_pfn &&
3943 early_node_map[i].end_pfn <= end_pfn) {
3944 /* clear it */
3945 early_node_map[i].start_pfn = 0;
3946 early_node_map[i].end_pfn = 0;
3947 removed = 1;
3948 continue;
3950 if (early_node_map[i].start_pfn < start_pfn &&
3951 early_node_map[i].end_pfn > start_pfn) {
3952 unsigned long temp_end_pfn = early_node_map[i].end_pfn;
3953 early_node_map[i].end_pfn = start_pfn;
3954 if (temp_end_pfn > end_pfn)
3955 add_active_range(nid, end_pfn, temp_end_pfn);
3956 continue;
3958 if (early_node_map[i].start_pfn >= start_pfn &&
3959 early_node_map[i].end_pfn > end_pfn &&
3960 early_node_map[i].start_pfn < end_pfn) {
3961 early_node_map[i].start_pfn = end_pfn;
3962 continue;
3966 if (!removed)
3967 return;
3969 /* remove the blank ones */
3970 for (i = nr_nodemap_entries - 1; i > 0; i--) {
3971 if (early_node_map[i].nid != nid)
3972 continue;
3973 if (early_node_map[i].end_pfn)
3974 continue;
3975 /* we found it, get rid of it */
3976 for (j = i; j < nr_nodemap_entries - 1; j++)
3977 memcpy(&early_node_map[j], &early_node_map[j+1],
3978 sizeof(early_node_map[j]));
3979 j = nr_nodemap_entries - 1;
3980 memset(&early_node_map[j], 0, sizeof(early_node_map[j]));
3981 nr_nodemap_entries--;
3986 * remove_all_active_ranges - Remove all currently registered regions
3988 * During discovery, it may be found that a table like SRAT is invalid
3989 * and an alternative discovery method must be used. This function removes
3990 * all currently registered regions.
3992 void __init remove_all_active_ranges(void)
3994 memset(early_node_map, 0, sizeof(early_node_map));
3995 nr_nodemap_entries = 0;
3998 /* Compare two active node_active_regions */
3999 static int __init cmp_node_active_region(const void *a, const void *b)
4001 struct node_active_region *arange = (struct node_active_region *)a;
4002 struct node_active_region *brange = (struct node_active_region *)b;
4004 /* Done this way to avoid overflows */
4005 if (arange->start_pfn > brange->start_pfn)
4006 return 1;
4007 if (arange->start_pfn < brange->start_pfn)
4008 return -1;
4010 return 0;
4013 /* sort the node_map by start_pfn */
4014 static void __init sort_node_map(void)
4016 sort(early_node_map, (size_t)nr_nodemap_entries,
4017 sizeof(struct node_active_region),
4018 cmp_node_active_region, NULL);
4021 /* Find the lowest pfn for a node */
4022 static unsigned long __init find_min_pfn_for_node(int nid)
4024 int i;
4025 unsigned long min_pfn = ULONG_MAX;
4027 /* Assuming a sorted map, the first range found has the starting pfn */
4028 for_each_active_range_index_in_nid(i, nid)
4029 min_pfn = min(min_pfn, early_node_map[i].start_pfn);
4031 if (min_pfn == ULONG_MAX) {
4032 printk(KERN_WARNING
4033 "Could not find start_pfn for node %d\n", nid);
4034 return 0;
4037 return min_pfn;
4041 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4043 * It returns the minimum PFN based on information provided via
4044 * add_active_range().
4046 unsigned long __init find_min_pfn_with_active_regions(void)
4048 return find_min_pfn_for_node(MAX_NUMNODES);
4052 * early_calculate_totalpages()
4053 * Sum pages in active regions for movable zone.
4054 * Populate N_HIGH_MEMORY for calculating usable_nodes.
4056 static unsigned long __init early_calculate_totalpages(void)
4058 int i;
4059 unsigned long totalpages = 0;
4061 for (i = 0; i < nr_nodemap_entries; i++) {
4062 unsigned long pages = early_node_map[i].end_pfn -
4063 early_node_map[i].start_pfn;
4064 totalpages += pages;
4065 if (pages)
4066 node_set_state(early_node_map[i].nid, N_HIGH_MEMORY);
4068 return totalpages;
4072 * Find the PFN the Movable zone begins in each node. Kernel memory
4073 * is spread evenly between nodes as long as the nodes have enough
4074 * memory. When they don't, some nodes will have more kernelcore than
4075 * others
4077 static void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn)
4079 int i, nid;
4080 unsigned long usable_startpfn;
4081 unsigned long kernelcore_node, kernelcore_remaining;
4082 /* save the state before borrow the nodemask */
4083 nodemask_t saved_node_state = node_states[N_HIGH_MEMORY];
4084 unsigned long totalpages = early_calculate_totalpages();
4085 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
4088 * If movablecore was specified, calculate what size of
4089 * kernelcore that corresponds so that memory usable for
4090 * any allocation type is evenly spread. If both kernelcore
4091 * and movablecore are specified, then the value of kernelcore
4092 * will be used for required_kernelcore if it's greater than
4093 * what movablecore would have allowed.
4095 if (required_movablecore) {
4096 unsigned long corepages;
4099 * Round-up so that ZONE_MOVABLE is at least as large as what
4100 * was requested by the user
4102 required_movablecore =
4103 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
4104 corepages = totalpages - required_movablecore;
4106 required_kernelcore = max(required_kernelcore, corepages);
4109 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4110 if (!required_kernelcore)
4111 goto out;
4113 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4114 find_usable_zone_for_movable();
4115 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
4117 restart:
4118 /* Spread kernelcore memory as evenly as possible throughout nodes */
4119 kernelcore_node = required_kernelcore / usable_nodes;
4120 for_each_node_state(nid, N_HIGH_MEMORY) {
4122 * Recalculate kernelcore_node if the division per node
4123 * now exceeds what is necessary to satisfy the requested
4124 * amount of memory for the kernel
4126 if (required_kernelcore < kernelcore_node)
4127 kernelcore_node = required_kernelcore / usable_nodes;
4130 * As the map is walked, we track how much memory is usable
4131 * by the kernel using kernelcore_remaining. When it is
4132 * 0, the rest of the node is usable by ZONE_MOVABLE
4134 kernelcore_remaining = kernelcore_node;
4136 /* Go through each range of PFNs within this node */
4137 for_each_active_range_index_in_nid(i, nid) {
4138 unsigned long start_pfn, end_pfn;
4139 unsigned long size_pages;
4141 start_pfn = max(early_node_map[i].start_pfn,
4142 zone_movable_pfn[nid]);
4143 end_pfn = early_node_map[i].end_pfn;
4144 if (start_pfn >= end_pfn)
4145 continue;
4147 /* Account for what is only usable for kernelcore */
4148 if (start_pfn < usable_startpfn) {
4149 unsigned long kernel_pages;
4150 kernel_pages = min(end_pfn, usable_startpfn)
4151 - start_pfn;
4153 kernelcore_remaining -= min(kernel_pages,
4154 kernelcore_remaining);
4155 required_kernelcore -= min(kernel_pages,
4156 required_kernelcore);
4158 /* Continue if range is now fully accounted */
4159 if (end_pfn <= usable_startpfn) {
4162 * Push zone_movable_pfn to the end so
4163 * that if we have to rebalance
4164 * kernelcore across nodes, we will
4165 * not double account here
4167 zone_movable_pfn[nid] = end_pfn;
4168 continue;
4170 start_pfn = usable_startpfn;
4174 * The usable PFN range for ZONE_MOVABLE is from
4175 * start_pfn->end_pfn. Calculate size_pages as the
4176 * number of pages used as kernelcore
4178 size_pages = end_pfn - start_pfn;
4179 if (size_pages > kernelcore_remaining)
4180 size_pages = kernelcore_remaining;
4181 zone_movable_pfn[nid] = start_pfn + size_pages;
4184 * Some kernelcore has been met, update counts and
4185 * break if the kernelcore for this node has been
4186 * satisified
4188 required_kernelcore -= min(required_kernelcore,
4189 size_pages);
4190 kernelcore_remaining -= size_pages;
4191 if (!kernelcore_remaining)
4192 break;
4197 * If there is still required_kernelcore, we do another pass with one
4198 * less node in the count. This will push zone_movable_pfn[nid] further
4199 * along on the nodes that still have memory until kernelcore is
4200 * satisified
4202 usable_nodes--;
4203 if (usable_nodes && required_kernelcore > usable_nodes)
4204 goto restart;
4206 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4207 for (nid = 0; nid < MAX_NUMNODES; nid++)
4208 zone_movable_pfn[nid] =
4209 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
4211 out:
4212 /* restore the node_state */
4213 node_states[N_HIGH_MEMORY] = saved_node_state;
4216 /* Any regular memory on that node ? */
4217 static void check_for_regular_memory(pg_data_t *pgdat)
4219 #ifdef CONFIG_HIGHMEM
4220 enum zone_type zone_type;
4222 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) {
4223 struct zone *zone = &pgdat->node_zones[zone_type];
4224 if (zone->present_pages)
4225 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY);
4227 #endif
4231 * free_area_init_nodes - Initialise all pg_data_t and zone data
4232 * @max_zone_pfn: an array of max PFNs for each zone
4234 * This will call free_area_init_node() for each active node in the system.
4235 * Using the page ranges provided by add_active_range(), the size of each
4236 * zone in each node and their holes is calculated. If the maximum PFN
4237 * between two adjacent zones match, it is assumed that the zone is empty.
4238 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4239 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4240 * starts where the previous one ended. For example, ZONE_DMA32 starts
4241 * at arch_max_dma_pfn.
4243 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
4245 unsigned long nid;
4246 int i;
4248 /* Sort early_node_map as initialisation assumes it is sorted */
4249 sort_node_map();
4251 /* Record where the zone boundaries are */
4252 memset(arch_zone_lowest_possible_pfn, 0,
4253 sizeof(arch_zone_lowest_possible_pfn));
4254 memset(arch_zone_highest_possible_pfn, 0,
4255 sizeof(arch_zone_highest_possible_pfn));
4256 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
4257 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
4258 for (i = 1; i < MAX_NR_ZONES; i++) {
4259 if (i == ZONE_MOVABLE)
4260 continue;
4261 arch_zone_lowest_possible_pfn[i] =
4262 arch_zone_highest_possible_pfn[i-1];
4263 arch_zone_highest_possible_pfn[i] =
4264 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
4266 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
4267 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
4269 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4270 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
4271 find_zone_movable_pfns_for_nodes(zone_movable_pfn);
4273 /* Print out the zone ranges */
4274 printk("Zone PFN ranges:\n");
4275 for (i = 0; i < MAX_NR_ZONES; i++) {
4276 if (i == ZONE_MOVABLE)
4277 continue;
4278 printk(" %-8s %0#10lx -> %0#10lx\n",
4279 zone_names[i],
4280 arch_zone_lowest_possible_pfn[i],
4281 arch_zone_highest_possible_pfn[i]);
4284 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4285 printk("Movable zone start PFN for each node\n");
4286 for (i = 0; i < MAX_NUMNODES; i++) {
4287 if (zone_movable_pfn[i])
4288 printk(" Node %d: %lu\n", i, zone_movable_pfn[i]);
4291 /* Print out the early_node_map[] */
4292 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries);
4293 for (i = 0; i < nr_nodemap_entries; i++)
4294 printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map[i].nid,
4295 early_node_map[i].start_pfn,
4296 early_node_map[i].end_pfn);
4298 /* Initialise every node */
4299 mminit_verify_pageflags_layout();
4300 setup_nr_node_ids();
4301 for_each_online_node(nid) {
4302 pg_data_t *pgdat = NODE_DATA(nid);
4303 free_area_init_node(nid, NULL,
4304 find_min_pfn_for_node(nid), NULL);
4306 /* Any memory on that node */
4307 if (pgdat->node_present_pages)
4308 node_set_state(nid, N_HIGH_MEMORY);
4309 check_for_regular_memory(pgdat);
4313 static int __init cmdline_parse_core(char *p, unsigned long *core)
4315 unsigned long long coremem;
4316 if (!p)
4317 return -EINVAL;
4319 coremem = memparse(p, &p);
4320 *core = coremem >> PAGE_SHIFT;
4322 /* Paranoid check that UL is enough for the coremem value */
4323 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
4325 return 0;
4329 * kernelcore=size sets the amount of memory for use for allocations that
4330 * cannot be reclaimed or migrated.
4332 static int __init cmdline_parse_kernelcore(char *p)
4334 return cmdline_parse_core(p, &required_kernelcore);
4338 * movablecore=size sets the amount of memory for use for allocations that
4339 * can be reclaimed or migrated.
4341 static int __init cmdline_parse_movablecore(char *p)
4343 return cmdline_parse_core(p, &required_movablecore);
4346 early_param("kernelcore", cmdline_parse_kernelcore);
4347 early_param("movablecore", cmdline_parse_movablecore);
4349 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4352 * set_dma_reserve - set the specified number of pages reserved in the first zone
4353 * @new_dma_reserve: The number of pages to mark reserved
4355 * The per-cpu batchsize and zone watermarks are determined by present_pages.
4356 * In the DMA zone, a significant percentage may be consumed by kernel image
4357 * and other unfreeable allocations which can skew the watermarks badly. This
4358 * function may optionally be used to account for unfreeable pages in the
4359 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
4360 * smaller per-cpu batchsize.
4362 void __init set_dma_reserve(unsigned long new_dma_reserve)
4364 dma_reserve = new_dma_reserve;
4367 #ifndef CONFIG_NEED_MULTIPLE_NODES
4368 struct pglist_data __refdata contig_page_data = { .bdata = &bootmem_node_data[0] };
4369 EXPORT_SYMBOL(contig_page_data);
4370 #endif
4372 void __init free_area_init(unsigned long *zones_size)
4374 free_area_init_node(0, zones_size,
4375 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
4378 static int page_alloc_cpu_notify(struct notifier_block *self,
4379 unsigned long action, void *hcpu)
4381 int cpu = (unsigned long)hcpu;
4383 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
4384 drain_pages(cpu);
4387 * Spill the event counters of the dead processor
4388 * into the current processors event counters.
4389 * This artificially elevates the count of the current
4390 * processor.
4392 vm_events_fold_cpu(cpu);
4395 * Zero the differential counters of the dead processor
4396 * so that the vm statistics are consistent.
4398 * This is only okay since the processor is dead and cannot
4399 * race with what we are doing.
4401 refresh_cpu_vm_stats(cpu);
4403 return NOTIFY_OK;
4406 void __init page_alloc_init(void)
4408 hotcpu_notifier(page_alloc_cpu_notify, 0);
4412 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
4413 * or min_free_kbytes changes.
4415 static void calculate_totalreserve_pages(void)
4417 struct pglist_data *pgdat;
4418 unsigned long reserve_pages = 0;
4419 enum zone_type i, j;
4421 for_each_online_pgdat(pgdat) {
4422 for (i = 0; i < MAX_NR_ZONES; i++) {
4423 struct zone *zone = pgdat->node_zones + i;
4424 unsigned long max = 0;
4426 /* Find valid and maximum lowmem_reserve in the zone */
4427 for (j = i; j < MAX_NR_ZONES; j++) {
4428 if (zone->lowmem_reserve[j] > max)
4429 max = zone->lowmem_reserve[j];
4432 /* we treat the high watermark as reserved pages. */
4433 max += high_wmark_pages(zone);
4435 if (max > zone->present_pages)
4436 max = zone->present_pages;
4437 reserve_pages += max;
4440 totalreserve_pages = reserve_pages;
4444 * setup_per_zone_lowmem_reserve - called whenever
4445 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
4446 * has a correct pages reserved value, so an adequate number of
4447 * pages are left in the zone after a successful __alloc_pages().
4449 static void setup_per_zone_lowmem_reserve(void)
4451 struct pglist_data *pgdat;
4452 enum zone_type j, idx;
4454 for_each_online_pgdat(pgdat) {
4455 for (j = 0; j < MAX_NR_ZONES; j++) {
4456 struct zone *zone = pgdat->node_zones + j;
4457 unsigned long present_pages = zone->present_pages;
4459 zone->lowmem_reserve[j] = 0;
4461 idx = j;
4462 while (idx) {
4463 struct zone *lower_zone;
4465 idx--;
4467 if (sysctl_lowmem_reserve_ratio[idx] < 1)
4468 sysctl_lowmem_reserve_ratio[idx] = 1;
4470 lower_zone = pgdat->node_zones + idx;
4471 lower_zone->lowmem_reserve[j] = present_pages /
4472 sysctl_lowmem_reserve_ratio[idx];
4473 present_pages += lower_zone->present_pages;
4478 /* update totalreserve_pages */
4479 calculate_totalreserve_pages();
4483 * setup_per_zone_wmarks - called when min_free_kbytes changes
4484 * or when memory is hot-{added|removed}
4486 * Ensures that the watermark[min,low,high] values for each zone are set
4487 * correctly with respect to min_free_kbytes.
4489 void setup_per_zone_wmarks(void)
4491 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
4492 unsigned long lowmem_pages = 0;
4493 struct zone *zone;
4494 unsigned long flags;
4496 /* Calculate total number of !ZONE_HIGHMEM pages */
4497 for_each_zone(zone) {
4498 if (!is_highmem(zone))
4499 lowmem_pages += zone->present_pages;
4502 for_each_zone(zone) {
4503 u64 tmp;
4505 spin_lock_irqsave(&zone->lock, flags);
4506 tmp = (u64)pages_min * zone->present_pages;
4507 do_div(tmp, lowmem_pages);
4508 if (is_highmem(zone)) {
4510 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
4511 * need highmem pages, so cap pages_min to a small
4512 * value here.
4514 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
4515 * deltas controls asynch page reclaim, and so should
4516 * not be capped for highmem.
4518 int min_pages;
4520 min_pages = zone->present_pages / 1024;
4521 if (min_pages < SWAP_CLUSTER_MAX)
4522 min_pages = SWAP_CLUSTER_MAX;
4523 if (min_pages > 128)
4524 min_pages = 128;
4525 zone->watermark[WMARK_MIN] = min_pages;
4526 } else {
4528 * If it's a lowmem zone, reserve a number of pages
4529 * proportionate to the zone's size.
4531 zone->watermark[WMARK_MIN] = tmp;
4534 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
4535 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
4536 setup_zone_migrate_reserve(zone);
4537 spin_unlock_irqrestore(&zone->lock, flags);
4540 /* update totalreserve_pages */
4541 calculate_totalreserve_pages();
4545 * The inactive anon list should be small enough that the VM never has to
4546 * do too much work, but large enough that each inactive page has a chance
4547 * to be referenced again before it is swapped out.
4549 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
4550 * INACTIVE_ANON pages on this zone's LRU, maintained by the
4551 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
4552 * the anonymous pages are kept on the inactive list.
4554 * total target max
4555 * memory ratio inactive anon
4556 * -------------------------------------
4557 * 10MB 1 5MB
4558 * 100MB 1 50MB
4559 * 1GB 3 250MB
4560 * 10GB 10 0.9GB
4561 * 100GB 31 3GB
4562 * 1TB 101 10GB
4563 * 10TB 320 32GB
4565 void calculate_zone_inactive_ratio(struct zone *zone)
4567 unsigned int gb, ratio;
4569 /* Zone size in gigabytes */
4570 gb = zone->present_pages >> (30 - PAGE_SHIFT);
4571 if (gb)
4572 ratio = int_sqrt(10 * gb);
4573 else
4574 ratio = 1;
4576 zone->inactive_ratio = ratio;
4579 static void __init setup_per_zone_inactive_ratio(void)
4581 struct zone *zone;
4583 for_each_zone(zone)
4584 calculate_zone_inactive_ratio(zone);
4588 * Initialise min_free_kbytes.
4590 * For small machines we want it small (128k min). For large machines
4591 * we want it large (64MB max). But it is not linear, because network
4592 * bandwidth does not increase linearly with machine size. We use
4594 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
4595 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
4597 * which yields
4599 * 16MB: 512k
4600 * 32MB: 724k
4601 * 64MB: 1024k
4602 * 128MB: 1448k
4603 * 256MB: 2048k
4604 * 512MB: 2896k
4605 * 1024MB: 4096k
4606 * 2048MB: 5792k
4607 * 4096MB: 8192k
4608 * 8192MB: 11584k
4609 * 16384MB: 16384k
4611 static int __init init_per_zone_wmark_min(void)
4613 unsigned long lowmem_kbytes;
4615 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
4617 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
4618 if (min_free_kbytes < 128)
4619 min_free_kbytes = 128;
4620 if (min_free_kbytes > 65536)
4621 min_free_kbytes = 65536;
4622 setup_per_zone_wmarks();
4623 setup_per_zone_lowmem_reserve();
4624 setup_per_zone_inactive_ratio();
4625 return 0;
4627 module_init(init_per_zone_wmark_min)
4630 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
4631 * that we can call two helper functions whenever min_free_kbytes
4632 * changes.
4634 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
4635 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4637 proc_dointvec(table, write, file, buffer, length, ppos);
4638 if (write)
4639 setup_per_zone_wmarks();
4640 return 0;
4643 #ifdef CONFIG_NUMA
4644 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
4645 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4647 struct zone *zone;
4648 int rc;
4650 rc = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
4651 if (rc)
4652 return rc;
4654 for_each_zone(zone)
4655 zone->min_unmapped_pages = (zone->present_pages *
4656 sysctl_min_unmapped_ratio) / 100;
4657 return 0;
4660 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
4661 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4663 struct zone *zone;
4664 int rc;
4666 rc = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
4667 if (rc)
4668 return rc;
4670 for_each_zone(zone)
4671 zone->min_slab_pages = (zone->present_pages *
4672 sysctl_min_slab_ratio) / 100;
4673 return 0;
4675 #endif
4678 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
4679 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
4680 * whenever sysctl_lowmem_reserve_ratio changes.
4682 * The reserve ratio obviously has absolutely no relation with the
4683 * minimum watermarks. The lowmem reserve ratio can only make sense
4684 * if in function of the boot time zone sizes.
4686 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
4687 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4689 proc_dointvec_minmax(table, write, file, buffer, length, ppos);
4690 setup_per_zone_lowmem_reserve();
4691 return 0;
4695 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
4696 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
4697 * can have before it gets flushed back to buddy allocator.
4700 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
4701 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4703 struct zone *zone;
4704 unsigned int cpu;
4705 int ret;
4707 ret = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
4708 if (!write || (ret == -EINVAL))
4709 return ret;
4710 for_each_populated_zone(zone) {
4711 for_each_online_cpu(cpu) {
4712 unsigned long high;
4713 high = zone->present_pages / percpu_pagelist_fraction;
4714 setup_pagelist_highmark(zone_pcp(zone, cpu), high);
4717 return 0;
4720 int hashdist = HASHDIST_DEFAULT;
4722 #ifdef CONFIG_NUMA
4723 static int __init set_hashdist(char *str)
4725 if (!str)
4726 return 0;
4727 hashdist = simple_strtoul(str, &str, 0);
4728 return 1;
4730 __setup("hashdist=", set_hashdist);
4731 #endif
4734 * allocate a large system hash table from bootmem
4735 * - it is assumed that the hash table must contain an exact power-of-2
4736 * quantity of entries
4737 * - limit is the number of hash buckets, not the total allocation size
4739 void *__init alloc_large_system_hash(const char *tablename,
4740 unsigned long bucketsize,
4741 unsigned long numentries,
4742 int scale,
4743 int flags,
4744 unsigned int *_hash_shift,
4745 unsigned int *_hash_mask,
4746 unsigned long limit)
4748 unsigned long long max = limit;
4749 unsigned long log2qty, size;
4750 void *table = NULL;
4752 /* allow the kernel cmdline to have a say */
4753 if (!numentries) {
4754 /* round applicable memory size up to nearest megabyte */
4755 numentries = nr_kernel_pages;
4756 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
4757 numentries >>= 20 - PAGE_SHIFT;
4758 numentries <<= 20 - PAGE_SHIFT;
4760 /* limit to 1 bucket per 2^scale bytes of low memory */
4761 if (scale > PAGE_SHIFT)
4762 numentries >>= (scale - PAGE_SHIFT);
4763 else
4764 numentries <<= (PAGE_SHIFT - scale);
4766 /* Make sure we've got at least a 0-order allocation.. */
4767 if (unlikely((numentries * bucketsize) < PAGE_SIZE))
4768 numentries = PAGE_SIZE / bucketsize;
4770 numentries = roundup_pow_of_two(numentries);
4772 /* limit allocation size to 1/16 total memory by default */
4773 if (max == 0) {
4774 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
4775 do_div(max, bucketsize);
4778 if (numentries > max)
4779 numentries = max;
4781 log2qty = ilog2(numentries);
4783 do {
4784 size = bucketsize << log2qty;
4785 if (flags & HASH_EARLY)
4786 table = alloc_bootmem_nopanic(size);
4787 else if (hashdist)
4788 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
4789 else {
4791 * If bucketsize is not a power-of-two, we may free
4792 * some pages at the end of hash table which
4793 * alloc_pages_exact() automatically does
4795 if (get_order(size) < MAX_ORDER) {
4796 table = alloc_pages_exact(size, GFP_ATOMIC);
4797 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
4800 } while (!table && size > PAGE_SIZE && --log2qty);
4802 if (!table)
4803 panic("Failed to allocate %s hash table\n", tablename);
4805 printk(KERN_INFO "%s hash table entries: %d (order: %d, %lu bytes)\n",
4806 tablename,
4807 (1U << log2qty),
4808 ilog2(size) - PAGE_SHIFT,
4809 size);
4811 if (_hash_shift)
4812 *_hash_shift = log2qty;
4813 if (_hash_mask)
4814 *_hash_mask = (1 << log2qty) - 1;
4816 return table;
4819 /* Return a pointer to the bitmap storing bits affecting a block of pages */
4820 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
4821 unsigned long pfn)
4823 #ifdef CONFIG_SPARSEMEM
4824 return __pfn_to_section(pfn)->pageblock_flags;
4825 #else
4826 return zone->pageblock_flags;
4827 #endif /* CONFIG_SPARSEMEM */
4830 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
4832 #ifdef CONFIG_SPARSEMEM
4833 pfn &= (PAGES_PER_SECTION-1);
4834 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
4835 #else
4836 pfn = pfn - zone->zone_start_pfn;
4837 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
4838 #endif /* CONFIG_SPARSEMEM */
4842 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
4843 * @page: The page within the block of interest
4844 * @start_bitidx: The first bit of interest to retrieve
4845 * @end_bitidx: The last bit of interest
4846 * returns pageblock_bits flags
4848 unsigned long get_pageblock_flags_group(struct page *page,
4849 int start_bitidx, int end_bitidx)
4851 struct zone *zone;
4852 unsigned long *bitmap;
4853 unsigned long pfn, bitidx;
4854 unsigned long flags = 0;
4855 unsigned long value = 1;
4857 zone = page_zone(page);
4858 pfn = page_to_pfn(page);
4859 bitmap = get_pageblock_bitmap(zone, pfn);
4860 bitidx = pfn_to_bitidx(zone, pfn);
4862 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
4863 if (test_bit(bitidx + start_bitidx, bitmap))
4864 flags |= value;
4866 return flags;
4870 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
4871 * @page: The page within the block of interest
4872 * @start_bitidx: The first bit of interest
4873 * @end_bitidx: The last bit of interest
4874 * @flags: The flags to set
4876 void set_pageblock_flags_group(struct page *page, unsigned long flags,
4877 int start_bitidx, int end_bitidx)
4879 struct zone *zone;
4880 unsigned long *bitmap;
4881 unsigned long pfn, bitidx;
4882 unsigned long value = 1;
4884 zone = page_zone(page);
4885 pfn = page_to_pfn(page);
4886 bitmap = get_pageblock_bitmap(zone, pfn);
4887 bitidx = pfn_to_bitidx(zone, pfn);
4888 VM_BUG_ON(pfn < zone->zone_start_pfn);
4889 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages);
4891 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
4892 if (flags & value)
4893 __set_bit(bitidx + start_bitidx, bitmap);
4894 else
4895 __clear_bit(bitidx + start_bitidx, bitmap);
4899 * This is designed as sub function...plz see page_isolation.c also.
4900 * set/clear page block's type to be ISOLATE.
4901 * page allocater never alloc memory from ISOLATE block.
4904 int set_migratetype_isolate(struct page *page)
4906 struct zone *zone;
4907 unsigned long flags;
4908 int ret = -EBUSY;
4909 int zone_idx;
4911 zone = page_zone(page);
4912 zone_idx = zone_idx(zone);
4913 spin_lock_irqsave(&zone->lock, flags);
4915 * In future, more migrate types will be able to be isolation target.
4917 if (get_pageblock_migratetype(page) != MIGRATE_MOVABLE &&
4918 zone_idx != ZONE_MOVABLE)
4919 goto out;
4920 set_pageblock_migratetype(page, MIGRATE_ISOLATE);
4921 move_freepages_block(zone, page, MIGRATE_ISOLATE);
4922 ret = 0;
4923 out:
4924 spin_unlock_irqrestore(&zone->lock, flags);
4925 if (!ret)
4926 drain_all_pages();
4927 return ret;
4930 void unset_migratetype_isolate(struct page *page)
4932 struct zone *zone;
4933 unsigned long flags;
4934 zone = page_zone(page);
4935 spin_lock_irqsave(&zone->lock, flags);
4936 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
4937 goto out;
4938 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4939 move_freepages_block(zone, page, MIGRATE_MOVABLE);
4940 out:
4941 spin_unlock_irqrestore(&zone->lock, flags);
4944 #ifdef CONFIG_MEMORY_HOTREMOVE
4946 * All pages in the range must be isolated before calling this.
4948 void
4949 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
4951 struct page *page;
4952 struct zone *zone;
4953 int order, i;
4954 unsigned long pfn;
4955 unsigned long flags;
4956 /* find the first valid pfn */
4957 for (pfn = start_pfn; pfn < end_pfn; pfn++)
4958 if (pfn_valid(pfn))
4959 break;
4960 if (pfn == end_pfn)
4961 return;
4962 zone = page_zone(pfn_to_page(pfn));
4963 spin_lock_irqsave(&zone->lock, flags);
4964 pfn = start_pfn;
4965 while (pfn < end_pfn) {
4966 if (!pfn_valid(pfn)) {
4967 pfn++;
4968 continue;
4970 page = pfn_to_page(pfn);
4971 BUG_ON(page_count(page));
4972 BUG_ON(!PageBuddy(page));
4973 order = page_order(page);
4974 #ifdef CONFIG_DEBUG_VM
4975 printk(KERN_INFO "remove from free list %lx %d %lx\n",
4976 pfn, 1 << order, end_pfn);
4977 #endif
4978 list_del(&page->lru);
4979 rmv_page_order(page);
4980 zone->free_area[order].nr_free--;
4981 __mod_zone_page_state(zone, NR_FREE_PAGES,
4982 - (1UL << order));
4983 for (i = 0; i < (1 << order); i++)
4984 SetPageReserved((page+i));
4985 pfn += (1 << order);
4987 spin_unlock_irqrestore(&zone->lock, flags);
4989 #endif