KVM: x86 emulator: convert group 4 to new style
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / mm / page_alloc.c
blobf12ad1836abe115b1b8e3bf4b9187c01249e8a30
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>
51 #include <linux/memory.h>
52 #include <linux/compaction.h>
53 #include <trace/events/kmem.h>
54 #include <linux/ftrace_event.h>
56 #include <asm/tlbflush.h>
57 #include <asm/div64.h>
58 #include "internal.h"
60 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
61 DEFINE_PER_CPU(int, numa_node);
62 EXPORT_PER_CPU_SYMBOL(numa_node);
63 #endif
65 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
67 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
68 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
69 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
70 * defined in <linux/topology.h>.
72 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
73 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
74 #endif
77 * Array of node states.
79 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
80 [N_POSSIBLE] = NODE_MASK_ALL,
81 [N_ONLINE] = { { [0] = 1UL } },
82 #ifndef CONFIG_NUMA
83 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
84 #ifdef CONFIG_HIGHMEM
85 [N_HIGH_MEMORY] = { { [0] = 1UL } },
86 #endif
87 [N_CPU] = { { [0] = 1UL } },
88 #endif /* NUMA */
90 EXPORT_SYMBOL(node_states);
92 unsigned long totalram_pages __read_mostly;
93 unsigned long totalreserve_pages __read_mostly;
94 int percpu_pagelist_fraction;
95 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
97 #ifdef CONFIG_PM_SLEEP
99 * The following functions are used by the suspend/hibernate code to temporarily
100 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
101 * while devices are suspended. To avoid races with the suspend/hibernate code,
102 * they should always be called with pm_mutex held (gfp_allowed_mask also should
103 * only be modified with pm_mutex held, unless the suspend/hibernate code is
104 * guaranteed not to run in parallel with that modification).
106 void set_gfp_allowed_mask(gfp_t mask)
108 WARN_ON(!mutex_is_locked(&pm_mutex));
109 gfp_allowed_mask = mask;
112 gfp_t clear_gfp_allowed_mask(gfp_t mask)
114 gfp_t ret = gfp_allowed_mask;
116 WARN_ON(!mutex_is_locked(&pm_mutex));
117 gfp_allowed_mask &= ~mask;
118 return ret;
120 #endif /* CONFIG_PM_SLEEP */
122 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
123 int pageblock_order __read_mostly;
124 #endif
126 static void __free_pages_ok(struct page *page, unsigned int order);
129 * results with 256, 32 in the lowmem_reserve sysctl:
130 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
131 * 1G machine -> (16M dma, 784M normal, 224M high)
132 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
133 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
134 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
136 * TBD: should special case ZONE_DMA32 machines here - in those we normally
137 * don't need any ZONE_NORMAL reservation
139 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
140 #ifdef CONFIG_ZONE_DMA
141 256,
142 #endif
143 #ifdef CONFIG_ZONE_DMA32
144 256,
145 #endif
146 #ifdef CONFIG_HIGHMEM
148 #endif
152 EXPORT_SYMBOL(totalram_pages);
154 static char * const zone_names[MAX_NR_ZONES] = {
155 #ifdef CONFIG_ZONE_DMA
156 "DMA",
157 #endif
158 #ifdef CONFIG_ZONE_DMA32
159 "DMA32",
160 #endif
161 "Normal",
162 #ifdef CONFIG_HIGHMEM
163 "HighMem",
164 #endif
165 "Movable",
168 int min_free_kbytes = 1024;
170 static unsigned long __meminitdata nr_kernel_pages;
171 static unsigned long __meminitdata nr_all_pages;
172 static unsigned long __meminitdata dma_reserve;
174 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
176 * MAX_ACTIVE_REGIONS determines the maximum number of distinct
177 * ranges of memory (RAM) that may be registered with add_active_range().
178 * Ranges passed to add_active_range() will be merged if possible
179 * so the number of times add_active_range() can be called is
180 * related to the number of nodes and the number of holes
182 #ifdef CONFIG_MAX_ACTIVE_REGIONS
183 /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */
184 #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS
185 #else
186 #if MAX_NUMNODES >= 32
187 /* If there can be many nodes, allow up to 50 holes per node */
188 #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50)
189 #else
190 /* By default, allow up to 256 distinct regions */
191 #define MAX_ACTIVE_REGIONS 256
192 #endif
193 #endif
195 static struct node_active_region __meminitdata early_node_map[MAX_ACTIVE_REGIONS];
196 static int __meminitdata nr_nodemap_entries;
197 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
198 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
199 static unsigned long __initdata required_kernelcore;
200 static unsigned long __initdata required_movablecore;
201 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
203 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
204 int movable_zone;
205 EXPORT_SYMBOL(movable_zone);
206 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
208 #if MAX_NUMNODES > 1
209 int nr_node_ids __read_mostly = MAX_NUMNODES;
210 int nr_online_nodes __read_mostly = 1;
211 EXPORT_SYMBOL(nr_node_ids);
212 EXPORT_SYMBOL(nr_online_nodes);
213 #endif
215 int page_group_by_mobility_disabled __read_mostly;
217 static void set_pageblock_migratetype(struct page *page, int migratetype)
220 if (unlikely(page_group_by_mobility_disabled))
221 migratetype = MIGRATE_UNMOVABLE;
223 set_pageblock_flags_group(page, (unsigned long)migratetype,
224 PB_migrate, PB_migrate_end);
227 bool oom_killer_disabled __read_mostly;
229 #ifdef CONFIG_DEBUG_VM
230 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
232 int ret = 0;
233 unsigned seq;
234 unsigned long pfn = page_to_pfn(page);
236 do {
237 seq = zone_span_seqbegin(zone);
238 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
239 ret = 1;
240 else if (pfn < zone->zone_start_pfn)
241 ret = 1;
242 } while (zone_span_seqretry(zone, seq));
244 return ret;
247 static int page_is_consistent(struct zone *zone, struct page *page)
249 if (!pfn_valid_within(page_to_pfn(page)))
250 return 0;
251 if (zone != page_zone(page))
252 return 0;
254 return 1;
257 * Temporary debugging check for pages not lying within a given zone.
259 static int bad_range(struct zone *zone, struct page *page)
261 if (page_outside_zone_boundaries(zone, page))
262 return 1;
263 if (!page_is_consistent(zone, page))
264 return 1;
266 return 0;
268 #else
269 static inline int bad_range(struct zone *zone, struct page *page)
271 return 0;
273 #endif
275 static void bad_page(struct page *page)
277 static unsigned long resume;
278 static unsigned long nr_shown;
279 static unsigned long nr_unshown;
281 /* Don't complain about poisoned pages */
282 if (PageHWPoison(page)) {
283 __ClearPageBuddy(page);
284 return;
288 * Allow a burst of 60 reports, then keep quiet for that minute;
289 * or allow a steady drip of one report per second.
291 if (nr_shown == 60) {
292 if (time_before(jiffies, resume)) {
293 nr_unshown++;
294 goto out;
296 if (nr_unshown) {
297 printk(KERN_ALERT
298 "BUG: Bad page state: %lu messages suppressed\n",
299 nr_unshown);
300 nr_unshown = 0;
302 nr_shown = 0;
304 if (nr_shown++ == 0)
305 resume = jiffies + 60 * HZ;
307 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
308 current->comm, page_to_pfn(page));
309 dump_page(page);
311 dump_stack();
312 out:
313 /* Leave bad fields for debug, except PageBuddy could make trouble */
314 __ClearPageBuddy(page);
315 add_taint(TAINT_BAD_PAGE);
319 * Higher-order pages are called "compound pages". They are structured thusly:
321 * The first PAGE_SIZE page is called the "head page".
323 * The remaining PAGE_SIZE pages are called "tail pages".
325 * All pages have PG_compound set. All pages have their ->private pointing at
326 * the head page (even the head page has this).
328 * The first tail page's ->lru.next holds the address of the compound page's
329 * put_page() function. Its ->lru.prev holds the order of allocation.
330 * This usage means that zero-order pages may not be compound.
333 static void free_compound_page(struct page *page)
335 __free_pages_ok(page, compound_order(page));
338 void prep_compound_page(struct page *page, unsigned long order)
340 int i;
341 int nr_pages = 1 << order;
343 set_compound_page_dtor(page, free_compound_page);
344 set_compound_order(page, order);
345 __SetPageHead(page);
346 for (i = 1; i < nr_pages; i++) {
347 struct page *p = page + i;
349 __SetPageTail(p);
350 p->first_page = page;
354 static int destroy_compound_page(struct page *page, unsigned long order)
356 int i;
357 int nr_pages = 1 << order;
358 int bad = 0;
360 if (unlikely(compound_order(page) != order) ||
361 unlikely(!PageHead(page))) {
362 bad_page(page);
363 bad++;
366 __ClearPageHead(page);
368 for (i = 1; i < nr_pages; i++) {
369 struct page *p = page + i;
371 if (unlikely(!PageTail(p) || (p->first_page != page))) {
372 bad_page(page);
373 bad++;
375 __ClearPageTail(p);
378 return bad;
381 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
383 int i;
386 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
387 * and __GFP_HIGHMEM from hard or soft interrupt context.
389 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
390 for (i = 0; i < (1 << order); i++)
391 clear_highpage(page + i);
394 static inline void set_page_order(struct page *page, int order)
396 set_page_private(page, order);
397 __SetPageBuddy(page);
400 static inline void rmv_page_order(struct page *page)
402 __ClearPageBuddy(page);
403 set_page_private(page, 0);
407 * Locate the struct page for both the matching buddy in our
408 * pair (buddy1) and the combined O(n+1) page they form (page).
410 * 1) Any buddy B1 will have an order O twin B2 which satisfies
411 * the following equation:
412 * B2 = B1 ^ (1 << O)
413 * For example, if the starting buddy (buddy2) is #8 its order
414 * 1 buddy is #10:
415 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
417 * 2) Any buddy B will have an order O+1 parent P which
418 * satisfies the following equation:
419 * P = B & ~(1 << O)
421 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
423 static inline struct page *
424 __page_find_buddy(struct page *page, unsigned long page_idx, unsigned int order)
426 unsigned long buddy_idx = page_idx ^ (1 << order);
428 return page + (buddy_idx - page_idx);
431 static inline unsigned long
432 __find_combined_index(unsigned long page_idx, unsigned int order)
434 return (page_idx & ~(1 << order));
438 * This function checks whether a page is free && is the buddy
439 * we can do coalesce a page and its buddy if
440 * (a) the buddy is not in a hole &&
441 * (b) the buddy is in the buddy system &&
442 * (c) a page and its buddy have the same order &&
443 * (d) a page and its buddy are in the same zone.
445 * For recording whether a page is in the buddy system, we use PG_buddy.
446 * Setting, clearing, and testing PG_buddy is serialized by zone->lock.
448 * For recording page's order, we use page_private(page).
450 static inline int page_is_buddy(struct page *page, struct page *buddy,
451 int order)
453 if (!pfn_valid_within(page_to_pfn(buddy)))
454 return 0;
456 if (page_zone_id(page) != page_zone_id(buddy))
457 return 0;
459 if (PageBuddy(buddy) && page_order(buddy) == order) {
460 VM_BUG_ON(page_count(buddy) != 0);
461 return 1;
463 return 0;
467 * Freeing function for a buddy system allocator.
469 * The concept of a buddy system is to maintain direct-mapped table
470 * (containing bit values) for memory blocks of various "orders".
471 * The bottom level table contains the map for the smallest allocatable
472 * units of memory (here, pages), and each level above it describes
473 * pairs of units from the levels below, hence, "buddies".
474 * At a high level, all that happens here is marking the table entry
475 * at the bottom level available, and propagating the changes upward
476 * as necessary, plus some accounting needed to play nicely with other
477 * parts of the VM system.
478 * At each level, we keep a list of pages, which are heads of continuous
479 * free pages of length of (1 << order) and marked with PG_buddy. Page's
480 * order is recorded in page_private(page) field.
481 * So when we are allocating or freeing one, we can derive the state of the
482 * other. That is, if we allocate a small block, and both were
483 * free, the remainder of the region must be split into blocks.
484 * If a block is freed, and its buddy is also free, then this
485 * triggers coalescing into a block of larger size.
487 * -- wli
490 static inline void __free_one_page(struct page *page,
491 struct zone *zone, unsigned int order,
492 int migratetype)
494 unsigned long page_idx;
495 unsigned long combined_idx;
496 struct page *buddy;
498 if (unlikely(PageCompound(page)))
499 if (unlikely(destroy_compound_page(page, order)))
500 return;
502 VM_BUG_ON(migratetype == -1);
504 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
506 VM_BUG_ON(page_idx & ((1 << order) - 1));
507 VM_BUG_ON(bad_range(zone, page));
509 while (order < MAX_ORDER-1) {
510 buddy = __page_find_buddy(page, page_idx, order);
511 if (!page_is_buddy(page, buddy, order))
512 break;
514 /* Our buddy is free, merge with it and move up one order. */
515 list_del(&buddy->lru);
516 zone->free_area[order].nr_free--;
517 rmv_page_order(buddy);
518 combined_idx = __find_combined_index(page_idx, order);
519 page = page + (combined_idx - page_idx);
520 page_idx = combined_idx;
521 order++;
523 set_page_order(page, order);
526 * If this is not the largest possible page, check if the buddy
527 * of the next-highest order is free. If it is, it's possible
528 * that pages are being freed that will coalesce soon. In case,
529 * that is happening, add the free page to the tail of the list
530 * so it's less likely to be used soon and more likely to be merged
531 * as a higher order page
533 if ((order < MAX_ORDER-1) && pfn_valid_within(page_to_pfn(buddy))) {
534 struct page *higher_page, *higher_buddy;
535 combined_idx = __find_combined_index(page_idx, order);
536 higher_page = page + combined_idx - page_idx;
537 higher_buddy = __page_find_buddy(higher_page, combined_idx, order + 1);
538 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
539 list_add_tail(&page->lru,
540 &zone->free_area[order].free_list[migratetype]);
541 goto out;
545 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
546 out:
547 zone->free_area[order].nr_free++;
551 * free_page_mlock() -- clean up attempts to free and mlocked() page.
552 * Page should not be on lru, so no need to fix that up.
553 * free_pages_check() will verify...
555 static inline void free_page_mlock(struct page *page)
557 __dec_zone_page_state(page, NR_MLOCK);
558 __count_vm_event(UNEVICTABLE_MLOCKFREED);
561 static inline int free_pages_check(struct page *page)
563 if (unlikely(page_mapcount(page) |
564 (page->mapping != NULL) |
565 (atomic_read(&page->_count) != 0) |
566 (page->flags & PAGE_FLAGS_CHECK_AT_FREE))) {
567 bad_page(page);
568 return 1;
570 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
571 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
572 return 0;
576 * Frees a number of pages from the PCP lists
577 * Assumes all pages on list are in same zone, and of same order.
578 * count is the number of pages to free.
580 * If the zone was previously in an "all pages pinned" state then look to
581 * see if this freeing clears that state.
583 * And clear the zone's pages_scanned counter, to hold off the "all pages are
584 * pinned" detection logic.
586 static void free_pcppages_bulk(struct zone *zone, int count,
587 struct per_cpu_pages *pcp)
589 int migratetype = 0;
590 int batch_free = 0;
591 int to_free = count;
593 spin_lock(&zone->lock);
594 zone->all_unreclaimable = 0;
595 zone->pages_scanned = 0;
597 while (to_free) {
598 struct page *page;
599 struct list_head *list;
602 * Remove pages from lists in a round-robin fashion. A
603 * batch_free count is maintained that is incremented when an
604 * empty list is encountered. This is so more pages are freed
605 * off fuller lists instead of spinning excessively around empty
606 * lists
608 do {
609 batch_free++;
610 if (++migratetype == MIGRATE_PCPTYPES)
611 migratetype = 0;
612 list = &pcp->lists[migratetype];
613 } while (list_empty(list));
615 do {
616 page = list_entry(list->prev, struct page, lru);
617 /* must delete as __free_one_page list manipulates */
618 list_del(&page->lru);
619 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
620 __free_one_page(page, zone, 0, page_private(page));
621 trace_mm_page_pcpu_drain(page, 0, page_private(page));
622 } while (--to_free && --batch_free && !list_empty(list));
624 __mod_zone_page_state(zone, NR_FREE_PAGES, count);
625 spin_unlock(&zone->lock);
628 static void free_one_page(struct zone *zone, struct page *page, int order,
629 int migratetype)
631 spin_lock(&zone->lock);
632 zone->all_unreclaimable = 0;
633 zone->pages_scanned = 0;
635 __free_one_page(page, zone, order, migratetype);
636 __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order);
637 spin_unlock(&zone->lock);
640 static bool free_pages_prepare(struct page *page, unsigned int order)
642 int i;
643 int bad = 0;
645 trace_mm_page_free_direct(page, order);
646 kmemcheck_free_shadow(page, order);
648 for (i = 0; i < (1 << order); i++) {
649 struct page *pg = page + i;
651 if (PageAnon(pg))
652 pg->mapping = NULL;
653 bad += free_pages_check(pg);
655 if (bad)
656 return false;
658 if (!PageHighMem(page)) {
659 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
660 debug_check_no_obj_freed(page_address(page),
661 PAGE_SIZE << order);
663 arch_free_page(page, order);
664 kernel_map_pages(page, 1 << order, 0);
666 return true;
669 static void __free_pages_ok(struct page *page, unsigned int order)
671 unsigned long flags;
672 int wasMlocked = __TestClearPageMlocked(page);
674 if (!free_pages_prepare(page, order))
675 return;
677 local_irq_save(flags);
678 if (unlikely(wasMlocked))
679 free_page_mlock(page);
680 __count_vm_events(PGFREE, 1 << order);
681 free_one_page(page_zone(page), page, order,
682 get_pageblock_migratetype(page));
683 local_irq_restore(flags);
687 * permit the bootmem allocator to evade page validation on high-order frees
689 void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
691 if (order == 0) {
692 __ClearPageReserved(page);
693 set_page_count(page, 0);
694 set_page_refcounted(page);
695 __free_page(page);
696 } else {
697 int loop;
699 prefetchw(page);
700 for (loop = 0; loop < BITS_PER_LONG; loop++) {
701 struct page *p = &page[loop];
703 if (loop + 1 < BITS_PER_LONG)
704 prefetchw(p + 1);
705 __ClearPageReserved(p);
706 set_page_count(p, 0);
709 set_page_refcounted(page);
710 __free_pages(page, order);
716 * The order of subdivision here is critical for the IO subsystem.
717 * Please do not alter this order without good reasons and regression
718 * testing. Specifically, as large blocks of memory are subdivided,
719 * the order in which smaller blocks are delivered depends on the order
720 * they're subdivided in this function. This is the primary factor
721 * influencing the order in which pages are delivered to the IO
722 * subsystem according to empirical testing, and this is also justified
723 * by considering the behavior of a buddy system containing a single
724 * large block of memory acted on by a series of small allocations.
725 * This behavior is a critical factor in sglist merging's success.
727 * -- wli
729 static inline void expand(struct zone *zone, struct page *page,
730 int low, int high, struct free_area *area,
731 int migratetype)
733 unsigned long size = 1 << high;
735 while (high > low) {
736 area--;
737 high--;
738 size >>= 1;
739 VM_BUG_ON(bad_range(zone, &page[size]));
740 list_add(&page[size].lru, &area->free_list[migratetype]);
741 area->nr_free++;
742 set_page_order(&page[size], high);
747 * This page is about to be returned from the page allocator
749 static inline int check_new_page(struct page *page)
751 if (unlikely(page_mapcount(page) |
752 (page->mapping != NULL) |
753 (atomic_read(&page->_count) != 0) |
754 (page->flags & PAGE_FLAGS_CHECK_AT_PREP))) {
755 bad_page(page);
756 return 1;
758 return 0;
761 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
763 int i;
765 for (i = 0; i < (1 << order); i++) {
766 struct page *p = page + i;
767 if (unlikely(check_new_page(p)))
768 return 1;
771 set_page_private(page, 0);
772 set_page_refcounted(page);
774 arch_alloc_page(page, order);
775 kernel_map_pages(page, 1 << order, 1);
777 if (gfp_flags & __GFP_ZERO)
778 prep_zero_page(page, order, gfp_flags);
780 if (order && (gfp_flags & __GFP_COMP))
781 prep_compound_page(page, order);
783 return 0;
787 * Go through the free lists for the given migratetype and remove
788 * the smallest available page from the freelists
790 static inline
791 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
792 int migratetype)
794 unsigned int current_order;
795 struct free_area * area;
796 struct page *page;
798 /* Find a page of the appropriate size in the preferred list */
799 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
800 area = &(zone->free_area[current_order]);
801 if (list_empty(&area->free_list[migratetype]))
802 continue;
804 page = list_entry(area->free_list[migratetype].next,
805 struct page, lru);
806 list_del(&page->lru);
807 rmv_page_order(page);
808 area->nr_free--;
809 expand(zone, page, order, current_order, area, migratetype);
810 return page;
813 return NULL;
818 * This array describes the order lists are fallen back to when
819 * the free lists for the desirable migrate type are depleted
821 static int fallbacks[MIGRATE_TYPES][MIGRATE_TYPES-1] = {
822 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
823 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
824 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
825 [MIGRATE_RESERVE] = { MIGRATE_RESERVE, MIGRATE_RESERVE, MIGRATE_RESERVE }, /* Never used */
829 * Move the free pages in a range to the free lists of the requested type.
830 * Note that start_page and end_pages are not aligned on a pageblock
831 * boundary. If alignment is required, use move_freepages_block()
833 static int move_freepages(struct zone *zone,
834 struct page *start_page, struct page *end_page,
835 int migratetype)
837 struct page *page;
838 unsigned long order;
839 int pages_moved = 0;
841 #ifndef CONFIG_HOLES_IN_ZONE
843 * page_zone is not safe to call in this context when
844 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
845 * anyway as we check zone boundaries in move_freepages_block().
846 * Remove at a later date when no bug reports exist related to
847 * grouping pages by mobility
849 BUG_ON(page_zone(start_page) != page_zone(end_page));
850 #endif
852 for (page = start_page; page <= end_page;) {
853 /* Make sure we are not inadvertently changing nodes */
854 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
856 if (!pfn_valid_within(page_to_pfn(page))) {
857 page++;
858 continue;
861 if (!PageBuddy(page)) {
862 page++;
863 continue;
866 order = page_order(page);
867 list_del(&page->lru);
868 list_add(&page->lru,
869 &zone->free_area[order].free_list[migratetype]);
870 page += 1 << order;
871 pages_moved += 1 << order;
874 return pages_moved;
877 static int move_freepages_block(struct zone *zone, struct page *page,
878 int migratetype)
880 unsigned long start_pfn, end_pfn;
881 struct page *start_page, *end_page;
883 start_pfn = page_to_pfn(page);
884 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
885 start_page = pfn_to_page(start_pfn);
886 end_page = start_page + pageblock_nr_pages - 1;
887 end_pfn = start_pfn + pageblock_nr_pages - 1;
889 /* Do not cross zone boundaries */
890 if (start_pfn < zone->zone_start_pfn)
891 start_page = page;
892 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
893 return 0;
895 return move_freepages(zone, start_page, end_page, migratetype);
898 static void change_pageblock_range(struct page *pageblock_page,
899 int start_order, int migratetype)
901 int nr_pageblocks = 1 << (start_order - pageblock_order);
903 while (nr_pageblocks--) {
904 set_pageblock_migratetype(pageblock_page, migratetype);
905 pageblock_page += pageblock_nr_pages;
909 /* Remove an element from the buddy allocator from the fallback list */
910 static inline struct page *
911 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
913 struct free_area * area;
914 int current_order;
915 struct page *page;
916 int migratetype, i;
918 /* Find the largest possible block of pages in the other list */
919 for (current_order = MAX_ORDER-1; current_order >= order;
920 --current_order) {
921 for (i = 0; i < MIGRATE_TYPES - 1; i++) {
922 migratetype = fallbacks[start_migratetype][i];
924 /* MIGRATE_RESERVE handled later if necessary */
925 if (migratetype == MIGRATE_RESERVE)
926 continue;
928 area = &(zone->free_area[current_order]);
929 if (list_empty(&area->free_list[migratetype]))
930 continue;
932 page = list_entry(area->free_list[migratetype].next,
933 struct page, lru);
934 area->nr_free--;
937 * If breaking a large block of pages, move all free
938 * pages to the preferred allocation list. If falling
939 * back for a reclaimable kernel allocation, be more
940 * agressive about taking ownership of free pages
942 if (unlikely(current_order >= (pageblock_order >> 1)) ||
943 start_migratetype == MIGRATE_RECLAIMABLE ||
944 page_group_by_mobility_disabled) {
945 unsigned long pages;
946 pages = move_freepages_block(zone, page,
947 start_migratetype);
949 /* Claim the whole block if over half of it is free */
950 if (pages >= (1 << (pageblock_order-1)) ||
951 page_group_by_mobility_disabled)
952 set_pageblock_migratetype(page,
953 start_migratetype);
955 migratetype = start_migratetype;
958 /* Remove the page from the freelists */
959 list_del(&page->lru);
960 rmv_page_order(page);
962 /* Take ownership for orders >= pageblock_order */
963 if (current_order >= pageblock_order)
964 change_pageblock_range(page, current_order,
965 start_migratetype);
967 expand(zone, page, order, current_order, area, migratetype);
969 trace_mm_page_alloc_extfrag(page, order, current_order,
970 start_migratetype, migratetype);
972 return page;
976 return NULL;
980 * Do the hard work of removing an element from the buddy allocator.
981 * Call me with the zone->lock already held.
983 static struct page *__rmqueue(struct zone *zone, unsigned int order,
984 int migratetype)
986 struct page *page;
988 retry_reserve:
989 page = __rmqueue_smallest(zone, order, migratetype);
991 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
992 page = __rmqueue_fallback(zone, order, migratetype);
995 * Use MIGRATE_RESERVE rather than fail an allocation. goto
996 * is used because __rmqueue_smallest is an inline function
997 * and we want just one call site
999 if (!page) {
1000 migratetype = MIGRATE_RESERVE;
1001 goto retry_reserve;
1005 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1006 return page;
1010 * Obtain a specified number of elements from the buddy allocator, all under
1011 * a single hold of the lock, for efficiency. Add them to the supplied list.
1012 * Returns the number of new pages which were placed at *list.
1014 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1015 unsigned long count, struct list_head *list,
1016 int migratetype, int cold)
1018 int i;
1020 spin_lock(&zone->lock);
1021 for (i = 0; i < count; ++i) {
1022 struct page *page = __rmqueue(zone, order, migratetype);
1023 if (unlikely(page == NULL))
1024 break;
1027 * Split buddy pages returned by expand() are received here
1028 * in physical page order. The page is added to the callers and
1029 * list and the list head then moves forward. From the callers
1030 * perspective, the linked list is ordered by page number in
1031 * some conditions. This is useful for IO devices that can
1032 * merge IO requests if the physical pages are ordered
1033 * properly.
1035 if (likely(cold == 0))
1036 list_add(&page->lru, list);
1037 else
1038 list_add_tail(&page->lru, list);
1039 set_page_private(page, migratetype);
1040 list = &page->lru;
1042 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1043 spin_unlock(&zone->lock);
1044 return i;
1047 #ifdef CONFIG_NUMA
1049 * Called from the vmstat counter updater to drain pagesets of this
1050 * currently executing processor on remote nodes after they have
1051 * expired.
1053 * Note that this function must be called with the thread pinned to
1054 * a single processor.
1056 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1058 unsigned long flags;
1059 int to_drain;
1061 local_irq_save(flags);
1062 if (pcp->count >= pcp->batch)
1063 to_drain = pcp->batch;
1064 else
1065 to_drain = pcp->count;
1066 free_pcppages_bulk(zone, to_drain, pcp);
1067 pcp->count -= to_drain;
1068 local_irq_restore(flags);
1070 #endif
1073 * Drain pages of the indicated processor.
1075 * The processor must either be the current processor and the
1076 * thread pinned to the current processor or a processor that
1077 * is not online.
1079 static void drain_pages(unsigned int cpu)
1081 unsigned long flags;
1082 struct zone *zone;
1084 for_each_populated_zone(zone) {
1085 struct per_cpu_pageset *pset;
1086 struct per_cpu_pages *pcp;
1088 local_irq_save(flags);
1089 pset = per_cpu_ptr(zone->pageset, cpu);
1091 pcp = &pset->pcp;
1092 free_pcppages_bulk(zone, pcp->count, pcp);
1093 pcp->count = 0;
1094 local_irq_restore(flags);
1099 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1101 void drain_local_pages(void *arg)
1103 drain_pages(smp_processor_id());
1107 * Spill all the per-cpu pages from all CPUs back into the buddy allocator
1109 void drain_all_pages(void)
1111 on_each_cpu(drain_local_pages, NULL, 1);
1114 #ifdef CONFIG_HIBERNATION
1116 void mark_free_pages(struct zone *zone)
1118 unsigned long pfn, max_zone_pfn;
1119 unsigned long flags;
1120 int order, t;
1121 struct list_head *curr;
1123 if (!zone->spanned_pages)
1124 return;
1126 spin_lock_irqsave(&zone->lock, flags);
1128 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
1129 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1130 if (pfn_valid(pfn)) {
1131 struct page *page = pfn_to_page(pfn);
1133 if (!swsusp_page_is_forbidden(page))
1134 swsusp_unset_page_free(page);
1137 for_each_migratetype_order(order, t) {
1138 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1139 unsigned long i;
1141 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1142 for (i = 0; i < (1UL << order); i++)
1143 swsusp_set_page_free(pfn_to_page(pfn + i));
1146 spin_unlock_irqrestore(&zone->lock, flags);
1148 #endif /* CONFIG_PM */
1151 * Free a 0-order page
1152 * cold == 1 ? free a cold page : free a hot page
1154 void free_hot_cold_page(struct page *page, int cold)
1156 struct zone *zone = page_zone(page);
1157 struct per_cpu_pages *pcp;
1158 unsigned long flags;
1159 int migratetype;
1160 int wasMlocked = __TestClearPageMlocked(page);
1162 if (!free_pages_prepare(page, 0))
1163 return;
1165 migratetype = get_pageblock_migratetype(page);
1166 set_page_private(page, migratetype);
1167 local_irq_save(flags);
1168 if (unlikely(wasMlocked))
1169 free_page_mlock(page);
1170 __count_vm_event(PGFREE);
1173 * We only track unmovable, reclaimable and movable on pcp lists.
1174 * Free ISOLATE pages back to the allocator because they are being
1175 * offlined but treat RESERVE as movable pages so we can get those
1176 * areas back if necessary. Otherwise, we may have to free
1177 * excessively into the page allocator
1179 if (migratetype >= MIGRATE_PCPTYPES) {
1180 if (unlikely(migratetype == MIGRATE_ISOLATE)) {
1181 free_one_page(zone, page, 0, migratetype);
1182 goto out;
1184 migratetype = MIGRATE_MOVABLE;
1187 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1188 if (cold)
1189 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1190 else
1191 list_add(&page->lru, &pcp->lists[migratetype]);
1192 pcp->count++;
1193 if (pcp->count >= pcp->high) {
1194 free_pcppages_bulk(zone, pcp->batch, pcp);
1195 pcp->count -= pcp->batch;
1198 out:
1199 local_irq_restore(flags);
1203 * split_page takes a non-compound higher-order page, and splits it into
1204 * n (1<<order) sub-pages: page[0..n]
1205 * Each sub-page must be freed individually.
1207 * Note: this is probably too low level an operation for use in drivers.
1208 * Please consult with lkml before using this in your driver.
1210 void split_page(struct page *page, unsigned int order)
1212 int i;
1214 VM_BUG_ON(PageCompound(page));
1215 VM_BUG_ON(!page_count(page));
1217 #ifdef CONFIG_KMEMCHECK
1219 * Split shadow pages too, because free(page[0]) would
1220 * otherwise free the whole shadow.
1222 if (kmemcheck_page_is_tracked(page))
1223 split_page(virt_to_page(page[0].shadow), order);
1224 #endif
1226 for (i = 1; i < (1 << order); i++)
1227 set_page_refcounted(page + i);
1231 * Similar to split_page except the page is already free. As this is only
1232 * being used for migration, the migratetype of the block also changes.
1233 * As this is called with interrupts disabled, the caller is responsible
1234 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1235 * are enabled.
1237 * Note: this is probably too low level an operation for use in drivers.
1238 * Please consult with lkml before using this in your driver.
1240 int split_free_page(struct page *page)
1242 unsigned int order;
1243 unsigned long watermark;
1244 struct zone *zone;
1246 BUG_ON(!PageBuddy(page));
1248 zone = page_zone(page);
1249 order = page_order(page);
1251 /* Obey watermarks as if the page was being allocated */
1252 watermark = low_wmark_pages(zone) + (1 << order);
1253 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1254 return 0;
1256 /* Remove page from free list */
1257 list_del(&page->lru);
1258 zone->free_area[order].nr_free--;
1259 rmv_page_order(page);
1260 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1UL << order));
1262 /* Split into individual pages */
1263 set_page_refcounted(page);
1264 split_page(page, order);
1266 if (order >= pageblock_order - 1) {
1267 struct page *endpage = page + (1 << order) - 1;
1268 for (; page < endpage; page += pageblock_nr_pages)
1269 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1272 return 1 << order;
1276 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1277 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1278 * or two.
1280 static inline
1281 struct page *buffered_rmqueue(struct zone *preferred_zone,
1282 struct zone *zone, int order, gfp_t gfp_flags,
1283 int migratetype)
1285 unsigned long flags;
1286 struct page *page;
1287 int cold = !!(gfp_flags & __GFP_COLD);
1289 again:
1290 if (likely(order == 0)) {
1291 struct per_cpu_pages *pcp;
1292 struct list_head *list;
1294 local_irq_save(flags);
1295 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1296 list = &pcp->lists[migratetype];
1297 if (list_empty(list)) {
1298 pcp->count += rmqueue_bulk(zone, 0,
1299 pcp->batch, list,
1300 migratetype, cold);
1301 if (unlikely(list_empty(list)))
1302 goto failed;
1305 if (cold)
1306 page = list_entry(list->prev, struct page, lru);
1307 else
1308 page = list_entry(list->next, struct page, lru);
1310 list_del(&page->lru);
1311 pcp->count--;
1312 } else {
1313 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1315 * __GFP_NOFAIL is not to be used in new code.
1317 * All __GFP_NOFAIL callers should be fixed so that they
1318 * properly detect and handle allocation failures.
1320 * We most definitely don't want callers attempting to
1321 * allocate greater than order-1 page units with
1322 * __GFP_NOFAIL.
1324 WARN_ON_ONCE(order > 1);
1326 spin_lock_irqsave(&zone->lock, flags);
1327 page = __rmqueue(zone, order, migratetype);
1328 spin_unlock(&zone->lock);
1329 if (!page)
1330 goto failed;
1331 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << order));
1334 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1335 zone_statistics(preferred_zone, zone);
1336 local_irq_restore(flags);
1338 VM_BUG_ON(bad_range(zone, page));
1339 if (prep_new_page(page, order, gfp_flags))
1340 goto again;
1341 return page;
1343 failed:
1344 local_irq_restore(flags);
1345 return NULL;
1348 /* The ALLOC_WMARK bits are used as an index to zone->watermark */
1349 #define ALLOC_WMARK_MIN WMARK_MIN
1350 #define ALLOC_WMARK_LOW WMARK_LOW
1351 #define ALLOC_WMARK_HIGH WMARK_HIGH
1352 #define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */
1354 /* Mask to get the watermark bits */
1355 #define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1)
1357 #define ALLOC_HARDER 0x10 /* try to alloc harder */
1358 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
1359 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
1361 #ifdef CONFIG_FAIL_PAGE_ALLOC
1363 static struct fail_page_alloc_attr {
1364 struct fault_attr attr;
1366 u32 ignore_gfp_highmem;
1367 u32 ignore_gfp_wait;
1368 u32 min_order;
1370 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1372 struct dentry *ignore_gfp_highmem_file;
1373 struct dentry *ignore_gfp_wait_file;
1374 struct dentry *min_order_file;
1376 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1378 } fail_page_alloc = {
1379 .attr = FAULT_ATTR_INITIALIZER,
1380 .ignore_gfp_wait = 1,
1381 .ignore_gfp_highmem = 1,
1382 .min_order = 1,
1385 static int __init setup_fail_page_alloc(char *str)
1387 return setup_fault_attr(&fail_page_alloc.attr, str);
1389 __setup("fail_page_alloc=", setup_fail_page_alloc);
1391 static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1393 if (order < fail_page_alloc.min_order)
1394 return 0;
1395 if (gfp_mask & __GFP_NOFAIL)
1396 return 0;
1397 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1398 return 0;
1399 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1400 return 0;
1402 return should_fail(&fail_page_alloc.attr, 1 << order);
1405 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1407 static int __init fail_page_alloc_debugfs(void)
1409 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1410 struct dentry *dir;
1411 int err;
1413 err = init_fault_attr_dentries(&fail_page_alloc.attr,
1414 "fail_page_alloc");
1415 if (err)
1416 return err;
1417 dir = fail_page_alloc.attr.dentries.dir;
1419 fail_page_alloc.ignore_gfp_wait_file =
1420 debugfs_create_bool("ignore-gfp-wait", mode, dir,
1421 &fail_page_alloc.ignore_gfp_wait);
1423 fail_page_alloc.ignore_gfp_highmem_file =
1424 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1425 &fail_page_alloc.ignore_gfp_highmem);
1426 fail_page_alloc.min_order_file =
1427 debugfs_create_u32("min-order", mode, dir,
1428 &fail_page_alloc.min_order);
1430 if (!fail_page_alloc.ignore_gfp_wait_file ||
1431 !fail_page_alloc.ignore_gfp_highmem_file ||
1432 !fail_page_alloc.min_order_file) {
1433 err = -ENOMEM;
1434 debugfs_remove(fail_page_alloc.ignore_gfp_wait_file);
1435 debugfs_remove(fail_page_alloc.ignore_gfp_highmem_file);
1436 debugfs_remove(fail_page_alloc.min_order_file);
1437 cleanup_fault_attr_dentries(&fail_page_alloc.attr);
1440 return err;
1443 late_initcall(fail_page_alloc_debugfs);
1445 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1447 #else /* CONFIG_FAIL_PAGE_ALLOC */
1449 static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1451 return 0;
1454 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1457 * Return 1 if free pages are above 'mark'. This takes into account the order
1458 * of the allocation.
1460 int zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1461 int classzone_idx, int alloc_flags)
1463 /* free_pages my go negative - that's OK */
1464 long min = mark;
1465 long free_pages = zone_nr_free_pages(z) - (1 << order) + 1;
1466 int o;
1468 if (alloc_flags & ALLOC_HIGH)
1469 min -= min / 2;
1470 if (alloc_flags & ALLOC_HARDER)
1471 min -= min / 4;
1473 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
1474 return 0;
1475 for (o = 0; o < order; o++) {
1476 /* At the next order, this order's pages become unavailable */
1477 free_pages -= z->free_area[o].nr_free << o;
1479 /* Require fewer higher order pages to be free */
1480 min >>= 1;
1482 if (free_pages <= min)
1483 return 0;
1485 return 1;
1488 #ifdef CONFIG_NUMA
1490 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1491 * skip over zones that are not allowed by the cpuset, or that have
1492 * been recently (in last second) found to be nearly full. See further
1493 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1494 * that have to skip over a lot of full or unallowed zones.
1496 * If the zonelist cache is present in the passed in zonelist, then
1497 * returns a pointer to the allowed node mask (either the current
1498 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1500 * If the zonelist cache is not available for this zonelist, does
1501 * nothing and returns NULL.
1503 * If the fullzones BITMAP in the zonelist cache is stale (more than
1504 * a second since last zap'd) then we zap it out (clear its bits.)
1506 * We hold off even calling zlc_setup, until after we've checked the
1507 * first zone in the zonelist, on the theory that most allocations will
1508 * be satisfied from that first zone, so best to examine that zone as
1509 * quickly as we can.
1511 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1513 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1514 nodemask_t *allowednodes; /* zonelist_cache approximation */
1516 zlc = zonelist->zlcache_ptr;
1517 if (!zlc)
1518 return NULL;
1520 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1521 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1522 zlc->last_full_zap = jiffies;
1525 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1526 &cpuset_current_mems_allowed :
1527 &node_states[N_HIGH_MEMORY];
1528 return allowednodes;
1532 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1533 * if it is worth looking at further for free memory:
1534 * 1) Check that the zone isn't thought to be full (doesn't have its
1535 * bit set in the zonelist_cache fullzones BITMAP).
1536 * 2) Check that the zones node (obtained from the zonelist_cache
1537 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1538 * Return true (non-zero) if zone is worth looking at further, or
1539 * else return false (zero) if it is not.
1541 * This check -ignores- the distinction between various watermarks,
1542 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1543 * found to be full for any variation of these watermarks, it will
1544 * be considered full for up to one second by all requests, unless
1545 * we are so low on memory on all allowed nodes that we are forced
1546 * into the second scan of the zonelist.
1548 * In the second scan we ignore this zonelist cache and exactly
1549 * apply the watermarks to all zones, even it is slower to do so.
1550 * We are low on memory in the second scan, and should leave no stone
1551 * unturned looking for a free page.
1553 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1554 nodemask_t *allowednodes)
1556 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1557 int i; /* index of *z in zonelist zones */
1558 int n; /* node that zone *z is on */
1560 zlc = zonelist->zlcache_ptr;
1561 if (!zlc)
1562 return 1;
1564 i = z - zonelist->_zonerefs;
1565 n = zlc->z_to_n[i];
1567 /* This zone is worth trying if it is allowed but not full */
1568 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1572 * Given 'z' scanning a zonelist, set the corresponding bit in
1573 * zlc->fullzones, so that subsequent attempts to allocate a page
1574 * from that zone don't waste time re-examining it.
1576 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1578 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1579 int i; /* index of *z in zonelist zones */
1581 zlc = zonelist->zlcache_ptr;
1582 if (!zlc)
1583 return;
1585 i = z - zonelist->_zonerefs;
1587 set_bit(i, zlc->fullzones);
1590 #else /* CONFIG_NUMA */
1592 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1594 return NULL;
1597 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1598 nodemask_t *allowednodes)
1600 return 1;
1603 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1606 #endif /* CONFIG_NUMA */
1609 * get_page_from_freelist goes through the zonelist trying to allocate
1610 * a page.
1612 static struct page *
1613 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1614 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1615 struct zone *preferred_zone, int migratetype)
1617 struct zoneref *z;
1618 struct page *page = NULL;
1619 int classzone_idx;
1620 struct zone *zone;
1621 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1622 int zlc_active = 0; /* set if using zonelist_cache */
1623 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1625 classzone_idx = zone_idx(preferred_zone);
1626 zonelist_scan:
1628 * Scan zonelist, looking for a zone with enough free.
1629 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1631 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1632 high_zoneidx, nodemask) {
1633 if (NUMA_BUILD && zlc_active &&
1634 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1635 continue;
1636 if ((alloc_flags & ALLOC_CPUSET) &&
1637 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1638 goto try_next_zone;
1640 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1641 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1642 unsigned long mark;
1643 int ret;
1645 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1646 if (zone_watermark_ok(zone, order, mark,
1647 classzone_idx, alloc_flags))
1648 goto try_this_zone;
1650 if (zone_reclaim_mode == 0)
1651 goto this_zone_full;
1653 ret = zone_reclaim(zone, gfp_mask, order);
1654 switch (ret) {
1655 case ZONE_RECLAIM_NOSCAN:
1656 /* did not scan */
1657 goto try_next_zone;
1658 case ZONE_RECLAIM_FULL:
1659 /* scanned but unreclaimable */
1660 goto this_zone_full;
1661 default:
1662 /* did we reclaim enough */
1663 if (!zone_watermark_ok(zone, order, mark,
1664 classzone_idx, alloc_flags))
1665 goto this_zone_full;
1669 try_this_zone:
1670 page = buffered_rmqueue(preferred_zone, zone, order,
1671 gfp_mask, migratetype);
1672 if (page)
1673 break;
1674 this_zone_full:
1675 if (NUMA_BUILD)
1676 zlc_mark_zone_full(zonelist, z);
1677 try_next_zone:
1678 if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) {
1680 * we do zlc_setup after the first zone is tried but only
1681 * if there are multiple nodes make it worthwhile
1683 allowednodes = zlc_setup(zonelist, alloc_flags);
1684 zlc_active = 1;
1685 did_zlc_setup = 1;
1689 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) {
1690 /* Disable zlc cache for second zonelist scan */
1691 zlc_active = 0;
1692 goto zonelist_scan;
1694 return page;
1697 static inline int
1698 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
1699 unsigned long pages_reclaimed)
1701 /* Do not loop if specifically requested */
1702 if (gfp_mask & __GFP_NORETRY)
1703 return 0;
1706 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
1707 * means __GFP_NOFAIL, but that may not be true in other
1708 * implementations.
1710 if (order <= PAGE_ALLOC_COSTLY_ORDER)
1711 return 1;
1714 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
1715 * specified, then we retry until we no longer reclaim any pages
1716 * (above), or we've reclaimed an order of pages at least as
1717 * large as the allocation's order. In both cases, if the
1718 * allocation still fails, we stop retrying.
1720 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
1721 return 1;
1724 * Don't let big-order allocations loop unless the caller
1725 * explicitly requests that.
1727 if (gfp_mask & __GFP_NOFAIL)
1728 return 1;
1730 return 0;
1733 static inline struct page *
1734 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
1735 struct zonelist *zonelist, enum zone_type high_zoneidx,
1736 nodemask_t *nodemask, struct zone *preferred_zone,
1737 int migratetype)
1739 struct page *page;
1741 /* Acquire the OOM killer lock for the zones in zonelist */
1742 if (!try_set_zonelist_oom(zonelist, gfp_mask)) {
1743 schedule_timeout_uninterruptible(1);
1744 return NULL;
1748 * Go through the zonelist yet one more time, keep very high watermark
1749 * here, this is only to catch a parallel oom killing, we must fail if
1750 * we're still under heavy pressure.
1752 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
1753 order, zonelist, high_zoneidx,
1754 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
1755 preferred_zone, migratetype);
1756 if (page)
1757 goto out;
1759 if (!(gfp_mask & __GFP_NOFAIL)) {
1760 /* The OOM killer will not help higher order allocs */
1761 if (order > PAGE_ALLOC_COSTLY_ORDER)
1762 goto out;
1763 /* The OOM killer does not needlessly kill tasks for lowmem */
1764 if (high_zoneidx < ZONE_NORMAL)
1765 goto out;
1767 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
1768 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
1769 * The caller should handle page allocation failure by itself if
1770 * it specifies __GFP_THISNODE.
1771 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
1773 if (gfp_mask & __GFP_THISNODE)
1774 goto out;
1776 /* Exhausted what can be done so it's blamo time */
1777 out_of_memory(zonelist, gfp_mask, order, nodemask);
1779 out:
1780 clear_zonelist_oom(zonelist, gfp_mask);
1781 return page;
1784 #ifdef CONFIG_COMPACTION
1785 /* Try memory compaction for high-order allocations before reclaim */
1786 static struct page *
1787 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
1788 struct zonelist *zonelist, enum zone_type high_zoneidx,
1789 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1790 int migratetype, unsigned long *did_some_progress)
1792 struct page *page;
1794 if (!order || compaction_deferred(preferred_zone))
1795 return NULL;
1797 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
1798 nodemask);
1799 if (*did_some_progress != COMPACT_SKIPPED) {
1801 /* Page migration frees to the PCP lists but we want merging */
1802 drain_pages(get_cpu());
1803 put_cpu();
1805 page = get_page_from_freelist(gfp_mask, nodemask,
1806 order, zonelist, high_zoneidx,
1807 alloc_flags, preferred_zone,
1808 migratetype);
1809 if (page) {
1810 preferred_zone->compact_considered = 0;
1811 preferred_zone->compact_defer_shift = 0;
1812 count_vm_event(COMPACTSUCCESS);
1813 return page;
1817 * It's bad if compaction run occurs and fails.
1818 * The most likely reason is that pages exist,
1819 * but not enough to satisfy watermarks.
1821 count_vm_event(COMPACTFAIL);
1822 defer_compaction(preferred_zone);
1824 cond_resched();
1827 return NULL;
1829 #else
1830 static inline struct page *
1831 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
1832 struct zonelist *zonelist, enum zone_type high_zoneidx,
1833 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1834 int migratetype, unsigned long *did_some_progress)
1836 return NULL;
1838 #endif /* CONFIG_COMPACTION */
1840 /* The really slow allocator path where we enter direct reclaim */
1841 static inline struct page *
1842 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
1843 struct zonelist *zonelist, enum zone_type high_zoneidx,
1844 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1845 int migratetype, unsigned long *did_some_progress)
1847 struct page *page = NULL;
1848 struct reclaim_state reclaim_state;
1849 struct task_struct *p = current;
1850 bool drained = false;
1852 cond_resched();
1854 /* We now go into synchronous reclaim */
1855 cpuset_memory_pressure_bump();
1856 p->flags |= PF_MEMALLOC;
1857 lockdep_set_current_reclaim_state(gfp_mask);
1858 reclaim_state.reclaimed_slab = 0;
1859 p->reclaim_state = &reclaim_state;
1861 *did_some_progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
1863 p->reclaim_state = NULL;
1864 lockdep_clear_current_reclaim_state();
1865 p->flags &= ~PF_MEMALLOC;
1867 cond_resched();
1869 if (unlikely(!(*did_some_progress)))
1870 return NULL;
1872 retry:
1873 page = get_page_from_freelist(gfp_mask, nodemask, order,
1874 zonelist, high_zoneidx,
1875 alloc_flags, preferred_zone,
1876 migratetype);
1879 * If an allocation failed after direct reclaim, it could be because
1880 * pages are pinned on the per-cpu lists. Drain them and try again
1882 if (!page && !drained) {
1883 drain_all_pages();
1884 drained = true;
1885 goto retry;
1888 return page;
1892 * This is called in the allocator slow-path if the allocation request is of
1893 * sufficient urgency to ignore watermarks and take other desperate measures
1895 static inline struct page *
1896 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
1897 struct zonelist *zonelist, enum zone_type high_zoneidx,
1898 nodemask_t *nodemask, struct zone *preferred_zone,
1899 int migratetype)
1901 struct page *page;
1903 do {
1904 page = get_page_from_freelist(gfp_mask, nodemask, order,
1905 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
1906 preferred_zone, migratetype);
1908 if (!page && gfp_mask & __GFP_NOFAIL)
1909 congestion_wait(BLK_RW_ASYNC, HZ/50);
1910 } while (!page && (gfp_mask & __GFP_NOFAIL));
1912 return page;
1915 static inline
1916 void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
1917 enum zone_type high_zoneidx)
1919 struct zoneref *z;
1920 struct zone *zone;
1922 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
1923 wakeup_kswapd(zone, order);
1926 static inline int
1927 gfp_to_alloc_flags(gfp_t gfp_mask)
1929 struct task_struct *p = current;
1930 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
1931 const gfp_t wait = gfp_mask & __GFP_WAIT;
1933 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
1934 BUILD_BUG_ON(__GFP_HIGH != ALLOC_HIGH);
1937 * The caller may dip into page reserves a bit more if the caller
1938 * cannot run direct reclaim, or if the caller has realtime scheduling
1939 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
1940 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
1942 alloc_flags |= (gfp_mask & __GFP_HIGH);
1944 if (!wait) {
1945 alloc_flags |= ALLOC_HARDER;
1947 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
1948 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1950 alloc_flags &= ~ALLOC_CPUSET;
1951 } else if (unlikely(rt_task(p)) && !in_interrupt())
1952 alloc_flags |= ALLOC_HARDER;
1954 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
1955 if (!in_interrupt() &&
1956 ((p->flags & PF_MEMALLOC) ||
1957 unlikely(test_thread_flag(TIF_MEMDIE))))
1958 alloc_flags |= ALLOC_NO_WATERMARKS;
1961 return alloc_flags;
1964 static inline struct page *
1965 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
1966 struct zonelist *zonelist, enum zone_type high_zoneidx,
1967 nodemask_t *nodemask, struct zone *preferred_zone,
1968 int migratetype)
1970 const gfp_t wait = gfp_mask & __GFP_WAIT;
1971 struct page *page = NULL;
1972 int alloc_flags;
1973 unsigned long pages_reclaimed = 0;
1974 unsigned long did_some_progress;
1975 struct task_struct *p = current;
1978 * In the slowpath, we sanity check order to avoid ever trying to
1979 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
1980 * be using allocators in order of preference for an area that is
1981 * too large.
1983 if (order >= MAX_ORDER) {
1984 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
1985 return NULL;
1989 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
1990 * __GFP_NOWARN set) should not cause reclaim since the subsystem
1991 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
1992 * using a larger set of nodes after it has established that the
1993 * allowed per node queues are empty and that nodes are
1994 * over allocated.
1996 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
1997 goto nopage;
1999 restart:
2000 wake_all_kswapd(order, zonelist, high_zoneidx);
2003 * OK, we're below the kswapd watermark and have kicked background
2004 * reclaim. Now things get more complex, so set up alloc_flags according
2005 * to how we want to proceed.
2007 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2009 /* This is the last chance, in general, before the goto nopage. */
2010 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2011 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2012 preferred_zone, migratetype);
2013 if (page)
2014 goto got_pg;
2016 rebalance:
2017 /* Allocate without watermarks if the context allows */
2018 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2019 page = __alloc_pages_high_priority(gfp_mask, order,
2020 zonelist, high_zoneidx, nodemask,
2021 preferred_zone, migratetype);
2022 if (page)
2023 goto got_pg;
2026 /* Atomic allocations - we can't balance anything */
2027 if (!wait)
2028 goto nopage;
2030 /* Avoid recursion of direct reclaim */
2031 if (p->flags & PF_MEMALLOC)
2032 goto nopage;
2034 /* Avoid allocations with no watermarks from looping endlessly */
2035 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2036 goto nopage;
2038 /* Try direct compaction */
2039 page = __alloc_pages_direct_compact(gfp_mask, order,
2040 zonelist, high_zoneidx,
2041 nodemask,
2042 alloc_flags, preferred_zone,
2043 migratetype, &did_some_progress);
2044 if (page)
2045 goto got_pg;
2047 /* Try direct reclaim and then allocating */
2048 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2049 zonelist, high_zoneidx,
2050 nodemask,
2051 alloc_flags, preferred_zone,
2052 migratetype, &did_some_progress);
2053 if (page)
2054 goto got_pg;
2057 * If we failed to make any progress reclaiming, then we are
2058 * running out of options and have to consider going OOM
2060 if (!did_some_progress) {
2061 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
2062 if (oom_killer_disabled)
2063 goto nopage;
2064 page = __alloc_pages_may_oom(gfp_mask, order,
2065 zonelist, high_zoneidx,
2066 nodemask, preferred_zone,
2067 migratetype);
2068 if (page)
2069 goto got_pg;
2071 if (!(gfp_mask & __GFP_NOFAIL)) {
2073 * The oom killer is not called for high-order
2074 * allocations that may fail, so if no progress
2075 * is being made, there are no other options and
2076 * retrying is unlikely to help.
2078 if (order > PAGE_ALLOC_COSTLY_ORDER)
2079 goto nopage;
2081 * The oom killer is not called for lowmem
2082 * allocations to prevent needlessly killing
2083 * innocent tasks.
2085 if (high_zoneidx < ZONE_NORMAL)
2086 goto nopage;
2089 goto restart;
2093 /* Check if we should retry the allocation */
2094 pages_reclaimed += did_some_progress;
2095 if (should_alloc_retry(gfp_mask, order, pages_reclaimed)) {
2096 /* Wait for some write requests to complete then retry */
2097 congestion_wait(BLK_RW_ASYNC, HZ/50);
2098 goto rebalance;
2101 nopage:
2102 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
2103 printk(KERN_WARNING "%s: page allocation failure."
2104 " order:%d, mode:0x%x\n",
2105 p->comm, order, gfp_mask);
2106 dump_stack();
2107 show_mem();
2109 return page;
2110 got_pg:
2111 if (kmemcheck_enabled)
2112 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2113 return page;
2118 * This is the 'heart' of the zoned buddy allocator.
2120 struct page *
2121 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2122 struct zonelist *zonelist, nodemask_t *nodemask)
2124 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2125 struct zone *preferred_zone;
2126 struct page *page;
2127 int migratetype = allocflags_to_migratetype(gfp_mask);
2129 gfp_mask &= gfp_allowed_mask;
2131 lockdep_trace_alloc(gfp_mask);
2133 might_sleep_if(gfp_mask & __GFP_WAIT);
2135 if (should_fail_alloc_page(gfp_mask, order))
2136 return NULL;
2139 * Check the zones suitable for the gfp_mask contain at least one
2140 * valid zone. It's possible to have an empty zonelist as a result
2141 * of GFP_THISNODE and a memoryless node
2143 if (unlikely(!zonelist->_zonerefs->zone))
2144 return NULL;
2146 get_mems_allowed();
2147 /* The preferred zone is used for statistics later */
2148 first_zones_zonelist(zonelist, high_zoneidx, nodemask, &preferred_zone);
2149 if (!preferred_zone) {
2150 put_mems_allowed();
2151 return NULL;
2154 /* First allocation attempt */
2155 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2156 zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET,
2157 preferred_zone, migratetype);
2158 if (unlikely(!page))
2159 page = __alloc_pages_slowpath(gfp_mask, order,
2160 zonelist, high_zoneidx, nodemask,
2161 preferred_zone, migratetype);
2162 put_mems_allowed();
2164 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2165 return page;
2167 EXPORT_SYMBOL(__alloc_pages_nodemask);
2170 * Common helper functions.
2172 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2174 struct page *page;
2177 * __get_free_pages() returns a 32-bit address, which cannot represent
2178 * a highmem page
2180 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2182 page = alloc_pages(gfp_mask, order);
2183 if (!page)
2184 return 0;
2185 return (unsigned long) page_address(page);
2187 EXPORT_SYMBOL(__get_free_pages);
2189 unsigned long get_zeroed_page(gfp_t gfp_mask)
2191 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2193 EXPORT_SYMBOL(get_zeroed_page);
2195 void __pagevec_free(struct pagevec *pvec)
2197 int i = pagevec_count(pvec);
2199 while (--i >= 0) {
2200 trace_mm_pagevec_free(pvec->pages[i], pvec->cold);
2201 free_hot_cold_page(pvec->pages[i], pvec->cold);
2205 void __free_pages(struct page *page, unsigned int order)
2207 if (put_page_testzero(page)) {
2208 if (order == 0)
2209 free_hot_cold_page(page, 0);
2210 else
2211 __free_pages_ok(page, order);
2215 EXPORT_SYMBOL(__free_pages);
2217 void free_pages(unsigned long addr, unsigned int order)
2219 if (addr != 0) {
2220 VM_BUG_ON(!virt_addr_valid((void *)addr));
2221 __free_pages(virt_to_page((void *)addr), order);
2225 EXPORT_SYMBOL(free_pages);
2228 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2229 * @size: the number of bytes to allocate
2230 * @gfp_mask: GFP flags for the allocation
2232 * This function is similar to alloc_pages(), except that it allocates the
2233 * minimum number of pages to satisfy the request. alloc_pages() can only
2234 * allocate memory in power-of-two pages.
2236 * This function is also limited by MAX_ORDER.
2238 * Memory allocated by this function must be released by free_pages_exact().
2240 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2242 unsigned int order = get_order(size);
2243 unsigned long addr;
2245 addr = __get_free_pages(gfp_mask, order);
2246 if (addr) {
2247 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2248 unsigned long used = addr + PAGE_ALIGN(size);
2250 split_page(virt_to_page((void *)addr), order);
2251 while (used < alloc_end) {
2252 free_page(used);
2253 used += PAGE_SIZE;
2257 return (void *)addr;
2259 EXPORT_SYMBOL(alloc_pages_exact);
2262 * free_pages_exact - release memory allocated via alloc_pages_exact()
2263 * @virt: the value returned by alloc_pages_exact.
2264 * @size: size of allocation, same value as passed to alloc_pages_exact().
2266 * Release the memory allocated by a previous call to alloc_pages_exact.
2268 void free_pages_exact(void *virt, size_t size)
2270 unsigned long addr = (unsigned long)virt;
2271 unsigned long end = addr + PAGE_ALIGN(size);
2273 while (addr < end) {
2274 free_page(addr);
2275 addr += PAGE_SIZE;
2278 EXPORT_SYMBOL(free_pages_exact);
2280 static unsigned int nr_free_zone_pages(int offset)
2282 struct zoneref *z;
2283 struct zone *zone;
2285 /* Just pick one node, since fallback list is circular */
2286 unsigned int sum = 0;
2288 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2290 for_each_zone_zonelist(zone, z, zonelist, offset) {
2291 unsigned long size = zone->present_pages;
2292 unsigned long high = high_wmark_pages(zone);
2293 if (size > high)
2294 sum += size - high;
2297 return sum;
2301 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2303 unsigned int nr_free_buffer_pages(void)
2305 return nr_free_zone_pages(gfp_zone(GFP_USER));
2307 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2310 * Amount of free RAM allocatable within all zones
2312 unsigned int nr_free_pagecache_pages(void)
2314 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2317 static inline void show_node(struct zone *zone)
2319 if (NUMA_BUILD)
2320 printk("Node %d ", zone_to_nid(zone));
2323 void si_meminfo(struct sysinfo *val)
2325 val->totalram = totalram_pages;
2326 val->sharedram = 0;
2327 val->freeram = global_page_state(NR_FREE_PAGES);
2328 val->bufferram = nr_blockdev_pages();
2329 val->totalhigh = totalhigh_pages;
2330 val->freehigh = nr_free_highpages();
2331 val->mem_unit = PAGE_SIZE;
2334 EXPORT_SYMBOL(si_meminfo);
2336 #ifdef CONFIG_NUMA
2337 void si_meminfo_node(struct sysinfo *val, int nid)
2339 pg_data_t *pgdat = NODE_DATA(nid);
2341 val->totalram = pgdat->node_present_pages;
2342 val->freeram = node_page_state(nid, NR_FREE_PAGES);
2343 #ifdef CONFIG_HIGHMEM
2344 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
2345 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2346 NR_FREE_PAGES);
2347 #else
2348 val->totalhigh = 0;
2349 val->freehigh = 0;
2350 #endif
2351 val->mem_unit = PAGE_SIZE;
2353 #endif
2355 #define K(x) ((x) << (PAGE_SHIFT-10))
2358 * Show free area list (used inside shift_scroll-lock stuff)
2359 * We also calculate the percentage fragmentation. We do this by counting the
2360 * memory on each free list with the exception of the first item on the list.
2362 void show_free_areas(void)
2364 int cpu;
2365 struct zone *zone;
2367 for_each_populated_zone(zone) {
2368 show_node(zone);
2369 printk("%s per-cpu:\n", zone->name);
2371 for_each_online_cpu(cpu) {
2372 struct per_cpu_pageset *pageset;
2374 pageset = per_cpu_ptr(zone->pageset, cpu);
2376 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2377 cpu, pageset->pcp.high,
2378 pageset->pcp.batch, pageset->pcp.count);
2382 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2383 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2384 " unevictable:%lu"
2385 " dirty:%lu writeback:%lu unstable:%lu\n"
2386 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2387 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n",
2388 global_page_state(NR_ACTIVE_ANON),
2389 global_page_state(NR_INACTIVE_ANON),
2390 global_page_state(NR_ISOLATED_ANON),
2391 global_page_state(NR_ACTIVE_FILE),
2392 global_page_state(NR_INACTIVE_FILE),
2393 global_page_state(NR_ISOLATED_FILE),
2394 global_page_state(NR_UNEVICTABLE),
2395 global_page_state(NR_FILE_DIRTY),
2396 global_page_state(NR_WRITEBACK),
2397 global_page_state(NR_UNSTABLE_NFS),
2398 global_page_state(NR_FREE_PAGES),
2399 global_page_state(NR_SLAB_RECLAIMABLE),
2400 global_page_state(NR_SLAB_UNRECLAIMABLE),
2401 global_page_state(NR_FILE_MAPPED),
2402 global_page_state(NR_SHMEM),
2403 global_page_state(NR_PAGETABLE),
2404 global_page_state(NR_BOUNCE));
2406 for_each_populated_zone(zone) {
2407 int i;
2409 show_node(zone);
2410 printk("%s"
2411 " free:%lukB"
2412 " min:%lukB"
2413 " low:%lukB"
2414 " high:%lukB"
2415 " active_anon:%lukB"
2416 " inactive_anon:%lukB"
2417 " active_file:%lukB"
2418 " inactive_file:%lukB"
2419 " unevictable:%lukB"
2420 " isolated(anon):%lukB"
2421 " isolated(file):%lukB"
2422 " present:%lukB"
2423 " mlocked:%lukB"
2424 " dirty:%lukB"
2425 " writeback:%lukB"
2426 " mapped:%lukB"
2427 " shmem:%lukB"
2428 " slab_reclaimable:%lukB"
2429 " slab_unreclaimable:%lukB"
2430 " kernel_stack:%lukB"
2431 " pagetables:%lukB"
2432 " unstable:%lukB"
2433 " bounce:%lukB"
2434 " writeback_tmp:%lukB"
2435 " pages_scanned:%lu"
2436 " all_unreclaimable? %s"
2437 "\n",
2438 zone->name,
2439 K(zone_nr_free_pages(zone)),
2440 K(min_wmark_pages(zone)),
2441 K(low_wmark_pages(zone)),
2442 K(high_wmark_pages(zone)),
2443 K(zone_page_state(zone, NR_ACTIVE_ANON)),
2444 K(zone_page_state(zone, NR_INACTIVE_ANON)),
2445 K(zone_page_state(zone, NR_ACTIVE_FILE)),
2446 K(zone_page_state(zone, NR_INACTIVE_FILE)),
2447 K(zone_page_state(zone, NR_UNEVICTABLE)),
2448 K(zone_page_state(zone, NR_ISOLATED_ANON)),
2449 K(zone_page_state(zone, NR_ISOLATED_FILE)),
2450 K(zone->present_pages),
2451 K(zone_page_state(zone, NR_MLOCK)),
2452 K(zone_page_state(zone, NR_FILE_DIRTY)),
2453 K(zone_page_state(zone, NR_WRITEBACK)),
2454 K(zone_page_state(zone, NR_FILE_MAPPED)),
2455 K(zone_page_state(zone, NR_SHMEM)),
2456 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
2457 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
2458 zone_page_state(zone, NR_KERNEL_STACK) *
2459 THREAD_SIZE / 1024,
2460 K(zone_page_state(zone, NR_PAGETABLE)),
2461 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
2462 K(zone_page_state(zone, NR_BOUNCE)),
2463 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
2464 zone->pages_scanned,
2465 (zone->all_unreclaimable ? "yes" : "no")
2467 printk("lowmem_reserve[]:");
2468 for (i = 0; i < MAX_NR_ZONES; i++)
2469 printk(" %lu", zone->lowmem_reserve[i]);
2470 printk("\n");
2473 for_each_populated_zone(zone) {
2474 unsigned long nr[MAX_ORDER], flags, order, total = 0;
2476 show_node(zone);
2477 printk("%s: ", zone->name);
2479 spin_lock_irqsave(&zone->lock, flags);
2480 for (order = 0; order < MAX_ORDER; order++) {
2481 nr[order] = zone->free_area[order].nr_free;
2482 total += nr[order] << order;
2484 spin_unlock_irqrestore(&zone->lock, flags);
2485 for (order = 0; order < MAX_ORDER; order++)
2486 printk("%lu*%lukB ", nr[order], K(1UL) << order);
2487 printk("= %lukB\n", K(total));
2490 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
2492 show_swap_cache_info();
2495 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
2497 zoneref->zone = zone;
2498 zoneref->zone_idx = zone_idx(zone);
2502 * Builds allocation fallback zone lists.
2504 * Add all populated zones of a node to the zonelist.
2506 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
2507 int nr_zones, enum zone_type zone_type)
2509 struct zone *zone;
2511 BUG_ON(zone_type >= MAX_NR_ZONES);
2512 zone_type++;
2514 do {
2515 zone_type--;
2516 zone = pgdat->node_zones + zone_type;
2517 if (populated_zone(zone)) {
2518 zoneref_set_zone(zone,
2519 &zonelist->_zonerefs[nr_zones++]);
2520 check_highest_zone(zone_type);
2523 } while (zone_type);
2524 return nr_zones;
2529 * zonelist_order:
2530 * 0 = automatic detection of better ordering.
2531 * 1 = order by ([node] distance, -zonetype)
2532 * 2 = order by (-zonetype, [node] distance)
2534 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
2535 * the same zonelist. So only NUMA can configure this param.
2537 #define ZONELIST_ORDER_DEFAULT 0
2538 #define ZONELIST_ORDER_NODE 1
2539 #define ZONELIST_ORDER_ZONE 2
2541 /* zonelist order in the kernel.
2542 * set_zonelist_order() will set this to NODE or ZONE.
2544 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
2545 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
2548 #ifdef CONFIG_NUMA
2549 /* The value user specified ....changed by config */
2550 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2551 /* string for sysctl */
2552 #define NUMA_ZONELIST_ORDER_LEN 16
2553 char numa_zonelist_order[16] = "default";
2556 * interface for configure zonelist ordering.
2557 * command line option "numa_zonelist_order"
2558 * = "[dD]efault - default, automatic configuration.
2559 * = "[nN]ode - order by node locality, then by zone within node
2560 * = "[zZ]one - order by zone, then by locality within zone
2563 static int __parse_numa_zonelist_order(char *s)
2565 if (*s == 'd' || *s == 'D') {
2566 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2567 } else if (*s == 'n' || *s == 'N') {
2568 user_zonelist_order = ZONELIST_ORDER_NODE;
2569 } else if (*s == 'z' || *s == 'Z') {
2570 user_zonelist_order = ZONELIST_ORDER_ZONE;
2571 } else {
2572 printk(KERN_WARNING
2573 "Ignoring invalid numa_zonelist_order value: "
2574 "%s\n", s);
2575 return -EINVAL;
2577 return 0;
2580 static __init int setup_numa_zonelist_order(char *s)
2582 if (s)
2583 return __parse_numa_zonelist_order(s);
2584 return 0;
2586 early_param("numa_zonelist_order", setup_numa_zonelist_order);
2589 * sysctl handler for numa_zonelist_order
2591 int numa_zonelist_order_handler(ctl_table *table, int write,
2592 void __user *buffer, size_t *length,
2593 loff_t *ppos)
2595 char saved_string[NUMA_ZONELIST_ORDER_LEN];
2596 int ret;
2597 static DEFINE_MUTEX(zl_order_mutex);
2599 mutex_lock(&zl_order_mutex);
2600 if (write)
2601 strcpy(saved_string, (char*)table->data);
2602 ret = proc_dostring(table, write, buffer, length, ppos);
2603 if (ret)
2604 goto out;
2605 if (write) {
2606 int oldval = user_zonelist_order;
2607 if (__parse_numa_zonelist_order((char*)table->data)) {
2609 * bogus value. restore saved string
2611 strncpy((char*)table->data, saved_string,
2612 NUMA_ZONELIST_ORDER_LEN);
2613 user_zonelist_order = oldval;
2614 } else if (oldval != user_zonelist_order) {
2615 mutex_lock(&zonelists_mutex);
2616 build_all_zonelists(NULL);
2617 mutex_unlock(&zonelists_mutex);
2620 out:
2621 mutex_unlock(&zl_order_mutex);
2622 return ret;
2626 #define MAX_NODE_LOAD (nr_online_nodes)
2627 static int node_load[MAX_NUMNODES];
2630 * find_next_best_node - find the next node that should appear in a given node's fallback list
2631 * @node: node whose fallback list we're appending
2632 * @used_node_mask: nodemask_t of already used nodes
2634 * We use a number of factors to determine which is the next node that should
2635 * appear on a given node's fallback list. The node should not have appeared
2636 * already in @node's fallback list, and it should be the next closest node
2637 * according to the distance array (which contains arbitrary distance values
2638 * from each node to each node in the system), and should also prefer nodes
2639 * with no CPUs, since presumably they'll have very little allocation pressure
2640 * on them otherwise.
2641 * It returns -1 if no node is found.
2643 static int find_next_best_node(int node, nodemask_t *used_node_mask)
2645 int n, val;
2646 int min_val = INT_MAX;
2647 int best_node = -1;
2648 const struct cpumask *tmp = cpumask_of_node(0);
2650 /* Use the local node if we haven't already */
2651 if (!node_isset(node, *used_node_mask)) {
2652 node_set(node, *used_node_mask);
2653 return node;
2656 for_each_node_state(n, N_HIGH_MEMORY) {
2658 /* Don't want a node to appear more than once */
2659 if (node_isset(n, *used_node_mask))
2660 continue;
2662 /* Use the distance array to find the distance */
2663 val = node_distance(node, n);
2665 /* Penalize nodes under us ("prefer the next node") */
2666 val += (n < node);
2668 /* Give preference to headless and unused nodes */
2669 tmp = cpumask_of_node(n);
2670 if (!cpumask_empty(tmp))
2671 val += PENALTY_FOR_NODE_WITH_CPUS;
2673 /* Slight preference for less loaded node */
2674 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
2675 val += node_load[n];
2677 if (val < min_val) {
2678 min_val = val;
2679 best_node = n;
2683 if (best_node >= 0)
2684 node_set(best_node, *used_node_mask);
2686 return best_node;
2691 * Build zonelists ordered by node and zones within node.
2692 * This results in maximum locality--normal zone overflows into local
2693 * DMA zone, if any--but risks exhausting DMA zone.
2695 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
2697 int j;
2698 struct zonelist *zonelist;
2700 zonelist = &pgdat->node_zonelists[0];
2701 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
2703 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2704 MAX_NR_ZONES - 1);
2705 zonelist->_zonerefs[j].zone = NULL;
2706 zonelist->_zonerefs[j].zone_idx = 0;
2710 * Build gfp_thisnode zonelists
2712 static void build_thisnode_zonelists(pg_data_t *pgdat)
2714 int j;
2715 struct zonelist *zonelist;
2717 zonelist = &pgdat->node_zonelists[1];
2718 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2719 zonelist->_zonerefs[j].zone = NULL;
2720 zonelist->_zonerefs[j].zone_idx = 0;
2724 * Build zonelists ordered by zone and nodes within zones.
2725 * This results in conserving DMA zone[s] until all Normal memory is
2726 * exhausted, but results in overflowing to remote node while memory
2727 * may still exist in local DMA zone.
2729 static int node_order[MAX_NUMNODES];
2731 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
2733 int pos, j, node;
2734 int zone_type; /* needs to be signed */
2735 struct zone *z;
2736 struct zonelist *zonelist;
2738 zonelist = &pgdat->node_zonelists[0];
2739 pos = 0;
2740 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
2741 for (j = 0; j < nr_nodes; j++) {
2742 node = node_order[j];
2743 z = &NODE_DATA(node)->node_zones[zone_type];
2744 if (populated_zone(z)) {
2745 zoneref_set_zone(z,
2746 &zonelist->_zonerefs[pos++]);
2747 check_highest_zone(zone_type);
2751 zonelist->_zonerefs[pos].zone = NULL;
2752 zonelist->_zonerefs[pos].zone_idx = 0;
2755 static int default_zonelist_order(void)
2757 int nid, zone_type;
2758 unsigned long low_kmem_size,total_size;
2759 struct zone *z;
2760 int average_size;
2762 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
2763 * If they are really small and used heavily, the system can fall
2764 * into OOM very easily.
2765 * This function detect ZONE_DMA/DMA32 size and configures zone order.
2767 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
2768 low_kmem_size = 0;
2769 total_size = 0;
2770 for_each_online_node(nid) {
2771 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2772 z = &NODE_DATA(nid)->node_zones[zone_type];
2773 if (populated_zone(z)) {
2774 if (zone_type < ZONE_NORMAL)
2775 low_kmem_size += z->present_pages;
2776 total_size += z->present_pages;
2777 } else if (zone_type == ZONE_NORMAL) {
2779 * If any node has only lowmem, then node order
2780 * is preferred to allow kernel allocations
2781 * locally; otherwise, they can easily infringe
2782 * on other nodes when there is an abundance of
2783 * lowmem available to allocate from.
2785 return ZONELIST_ORDER_NODE;
2789 if (!low_kmem_size || /* there are no DMA area. */
2790 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
2791 return ZONELIST_ORDER_NODE;
2793 * look into each node's config.
2794 * If there is a node whose DMA/DMA32 memory is very big area on
2795 * local memory, NODE_ORDER may be suitable.
2797 average_size = total_size /
2798 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1);
2799 for_each_online_node(nid) {
2800 low_kmem_size = 0;
2801 total_size = 0;
2802 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2803 z = &NODE_DATA(nid)->node_zones[zone_type];
2804 if (populated_zone(z)) {
2805 if (zone_type < ZONE_NORMAL)
2806 low_kmem_size += z->present_pages;
2807 total_size += z->present_pages;
2810 if (low_kmem_size &&
2811 total_size > average_size && /* ignore small node */
2812 low_kmem_size > total_size * 70/100)
2813 return ZONELIST_ORDER_NODE;
2815 return ZONELIST_ORDER_ZONE;
2818 static void set_zonelist_order(void)
2820 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
2821 current_zonelist_order = default_zonelist_order();
2822 else
2823 current_zonelist_order = user_zonelist_order;
2826 static void build_zonelists(pg_data_t *pgdat)
2828 int j, node, load;
2829 enum zone_type i;
2830 nodemask_t used_mask;
2831 int local_node, prev_node;
2832 struct zonelist *zonelist;
2833 int order = current_zonelist_order;
2835 /* initialize zonelists */
2836 for (i = 0; i < MAX_ZONELISTS; i++) {
2837 zonelist = pgdat->node_zonelists + i;
2838 zonelist->_zonerefs[0].zone = NULL;
2839 zonelist->_zonerefs[0].zone_idx = 0;
2842 /* NUMA-aware ordering of nodes */
2843 local_node = pgdat->node_id;
2844 load = nr_online_nodes;
2845 prev_node = local_node;
2846 nodes_clear(used_mask);
2848 memset(node_order, 0, sizeof(node_order));
2849 j = 0;
2851 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
2852 int distance = node_distance(local_node, node);
2855 * If another node is sufficiently far away then it is better
2856 * to reclaim pages in a zone before going off node.
2858 if (distance > RECLAIM_DISTANCE)
2859 zone_reclaim_mode = 1;
2862 * We don't want to pressure a particular node.
2863 * So adding penalty to the first node in same
2864 * distance group to make it round-robin.
2866 if (distance != node_distance(local_node, prev_node))
2867 node_load[node] = load;
2869 prev_node = node;
2870 load--;
2871 if (order == ZONELIST_ORDER_NODE)
2872 build_zonelists_in_node_order(pgdat, node);
2873 else
2874 node_order[j++] = node; /* remember order */
2877 if (order == ZONELIST_ORDER_ZONE) {
2878 /* calculate node order -- i.e., DMA last! */
2879 build_zonelists_in_zone_order(pgdat, j);
2882 build_thisnode_zonelists(pgdat);
2885 /* Construct the zonelist performance cache - see further mmzone.h */
2886 static void build_zonelist_cache(pg_data_t *pgdat)
2888 struct zonelist *zonelist;
2889 struct zonelist_cache *zlc;
2890 struct zoneref *z;
2892 zonelist = &pgdat->node_zonelists[0];
2893 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
2894 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
2895 for (z = zonelist->_zonerefs; z->zone; z++)
2896 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
2899 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
2901 * Return node id of node used for "local" allocations.
2902 * I.e., first node id of first zone in arg node's generic zonelist.
2903 * Used for initializing percpu 'numa_mem', which is used primarily
2904 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
2906 int local_memory_node(int node)
2908 struct zone *zone;
2910 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
2911 gfp_zone(GFP_KERNEL),
2912 NULL,
2913 &zone);
2914 return zone->node;
2916 #endif
2918 #else /* CONFIG_NUMA */
2920 static void set_zonelist_order(void)
2922 current_zonelist_order = ZONELIST_ORDER_ZONE;
2925 static void build_zonelists(pg_data_t *pgdat)
2927 int node, local_node;
2928 enum zone_type j;
2929 struct zonelist *zonelist;
2931 local_node = pgdat->node_id;
2933 zonelist = &pgdat->node_zonelists[0];
2934 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2937 * Now we build the zonelist so that it contains the zones
2938 * of all the other nodes.
2939 * We don't want to pressure a particular node, so when
2940 * building the zones for node N, we make sure that the
2941 * zones coming right after the local ones are those from
2942 * node N+1 (modulo N)
2944 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
2945 if (!node_online(node))
2946 continue;
2947 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2948 MAX_NR_ZONES - 1);
2950 for (node = 0; node < local_node; node++) {
2951 if (!node_online(node))
2952 continue;
2953 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2954 MAX_NR_ZONES - 1);
2957 zonelist->_zonerefs[j].zone = NULL;
2958 zonelist->_zonerefs[j].zone_idx = 0;
2961 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
2962 static void build_zonelist_cache(pg_data_t *pgdat)
2964 pgdat->node_zonelists[0].zlcache_ptr = NULL;
2967 #endif /* CONFIG_NUMA */
2970 * Boot pageset table. One per cpu which is going to be used for all
2971 * zones and all nodes. The parameters will be set in such a way
2972 * that an item put on a list will immediately be handed over to
2973 * the buddy list. This is safe since pageset manipulation is done
2974 * with interrupts disabled.
2976 * The boot_pagesets must be kept even after bootup is complete for
2977 * unused processors and/or zones. They do play a role for bootstrapping
2978 * hotplugged processors.
2980 * zoneinfo_show() and maybe other functions do
2981 * not check if the processor is online before following the pageset pointer.
2982 * Other parts of the kernel may not check if the zone is available.
2984 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
2985 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
2986 static void setup_zone_pageset(struct zone *zone);
2989 * Global mutex to protect against size modification of zonelists
2990 * as well as to serialize pageset setup for the new populated zone.
2992 DEFINE_MUTEX(zonelists_mutex);
2994 /* return values int ....just for stop_machine() */
2995 static __init_refok int __build_all_zonelists(void *data)
2997 int nid;
2998 int cpu;
3000 #ifdef CONFIG_NUMA
3001 memset(node_load, 0, sizeof(node_load));
3002 #endif
3003 for_each_online_node(nid) {
3004 pg_data_t *pgdat = NODE_DATA(nid);
3006 build_zonelists(pgdat);
3007 build_zonelist_cache(pgdat);
3010 #ifdef CONFIG_MEMORY_HOTPLUG
3011 /* Setup real pagesets for the new zone */
3012 if (data) {
3013 struct zone *zone = data;
3014 setup_zone_pageset(zone);
3016 #endif
3019 * Initialize the boot_pagesets that are going to be used
3020 * for bootstrapping processors. The real pagesets for
3021 * each zone will be allocated later when the per cpu
3022 * allocator is available.
3024 * boot_pagesets are used also for bootstrapping offline
3025 * cpus if the system is already booted because the pagesets
3026 * are needed to initialize allocators on a specific cpu too.
3027 * F.e. the percpu allocator needs the page allocator which
3028 * needs the percpu allocator in order to allocate its pagesets
3029 * (a chicken-egg dilemma).
3031 for_each_possible_cpu(cpu) {
3032 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3034 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3036 * We now know the "local memory node" for each node--
3037 * i.e., the node of the first zone in the generic zonelist.
3038 * Set up numa_mem percpu variable for on-line cpus. During
3039 * boot, only the boot cpu should be on-line; we'll init the
3040 * secondary cpus' numa_mem as they come on-line. During
3041 * node/memory hotplug, we'll fixup all on-line cpus.
3043 if (cpu_online(cpu))
3044 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3045 #endif
3048 return 0;
3052 * Called with zonelists_mutex held always
3053 * unless system_state == SYSTEM_BOOTING.
3055 void build_all_zonelists(void *data)
3057 set_zonelist_order();
3059 if (system_state == SYSTEM_BOOTING) {
3060 __build_all_zonelists(NULL);
3061 mminit_verify_zonelist();
3062 cpuset_init_current_mems_allowed();
3063 } else {
3064 /* we have to stop all cpus to guarantee there is no user
3065 of zonelist */
3066 stop_machine(__build_all_zonelists, data, NULL);
3067 /* cpuset refresh routine should be here */
3069 vm_total_pages = nr_free_pagecache_pages();
3071 * Disable grouping by mobility if the number of pages in the
3072 * system is too low to allow the mechanism to work. It would be
3073 * more accurate, but expensive to check per-zone. This check is
3074 * made on memory-hotadd so a system can start with mobility
3075 * disabled and enable it later
3077 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3078 page_group_by_mobility_disabled = 1;
3079 else
3080 page_group_by_mobility_disabled = 0;
3082 printk("Built %i zonelists in %s order, mobility grouping %s. "
3083 "Total pages: %ld\n",
3084 nr_online_nodes,
3085 zonelist_order_name[current_zonelist_order],
3086 page_group_by_mobility_disabled ? "off" : "on",
3087 vm_total_pages);
3088 #ifdef CONFIG_NUMA
3089 printk("Policy zone: %s\n", zone_names[policy_zone]);
3090 #endif
3094 * Helper functions to size the waitqueue hash table.
3095 * Essentially these want to choose hash table sizes sufficiently
3096 * large so that collisions trying to wait on pages are rare.
3097 * But in fact, the number of active page waitqueues on typical
3098 * systems is ridiculously low, less than 200. So this is even
3099 * conservative, even though it seems large.
3101 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3102 * waitqueues, i.e. the size of the waitq table given the number of pages.
3104 #define PAGES_PER_WAITQUEUE 256
3106 #ifndef CONFIG_MEMORY_HOTPLUG
3107 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3109 unsigned long size = 1;
3111 pages /= PAGES_PER_WAITQUEUE;
3113 while (size < pages)
3114 size <<= 1;
3117 * Once we have dozens or even hundreds of threads sleeping
3118 * on IO we've got bigger problems than wait queue collision.
3119 * Limit the size of the wait table to a reasonable size.
3121 size = min(size, 4096UL);
3123 return max(size, 4UL);
3125 #else
3127 * A zone's size might be changed by hot-add, so it is not possible to determine
3128 * a suitable size for its wait_table. So we use the maximum size now.
3130 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3132 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3133 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3134 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3136 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3137 * or more by the traditional way. (See above). It equals:
3139 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3140 * ia64(16K page size) : = ( 8G + 4M)byte.
3141 * powerpc (64K page size) : = (32G +16M)byte.
3143 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3145 return 4096UL;
3147 #endif
3150 * This is an integer logarithm so that shifts can be used later
3151 * to extract the more random high bits from the multiplicative
3152 * hash function before the remainder is taken.
3154 static inline unsigned long wait_table_bits(unsigned long size)
3156 return ffz(~size);
3159 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3162 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3163 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3164 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3165 * higher will lead to a bigger reserve which will get freed as contiguous
3166 * blocks as reclaim kicks in
3168 static void setup_zone_migrate_reserve(struct zone *zone)
3170 unsigned long start_pfn, pfn, end_pfn;
3171 struct page *page;
3172 unsigned long block_migratetype;
3173 int reserve;
3175 /* Get the start pfn, end pfn and the number of blocks to reserve */
3176 start_pfn = zone->zone_start_pfn;
3177 end_pfn = start_pfn + zone->spanned_pages;
3178 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3179 pageblock_order;
3182 * Reserve blocks are generally in place to help high-order atomic
3183 * allocations that are short-lived. A min_free_kbytes value that
3184 * would result in more than 2 reserve blocks for atomic allocations
3185 * is assumed to be in place to help anti-fragmentation for the
3186 * future allocation of hugepages at runtime.
3188 reserve = min(2, reserve);
3190 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
3191 if (!pfn_valid(pfn))
3192 continue;
3193 page = pfn_to_page(pfn);
3195 /* Watch out for overlapping nodes */
3196 if (page_to_nid(page) != zone_to_nid(zone))
3197 continue;
3199 /* Blocks with reserved pages will never free, skip them. */
3200 if (PageReserved(page))
3201 continue;
3203 block_migratetype = get_pageblock_migratetype(page);
3205 /* If this block is reserved, account for it */
3206 if (reserve > 0 && block_migratetype == MIGRATE_RESERVE) {
3207 reserve--;
3208 continue;
3211 /* Suitable for reserving if this block is movable */
3212 if (reserve > 0 && block_migratetype == MIGRATE_MOVABLE) {
3213 set_pageblock_migratetype(page, MIGRATE_RESERVE);
3214 move_freepages_block(zone, page, MIGRATE_RESERVE);
3215 reserve--;
3216 continue;
3220 * If the reserve is met and this is a previous reserved block,
3221 * take it back
3223 if (block_migratetype == MIGRATE_RESERVE) {
3224 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3225 move_freepages_block(zone, page, MIGRATE_MOVABLE);
3231 * Initially all pages are reserved - free ones are freed
3232 * up by free_all_bootmem() once the early boot process is
3233 * done. Non-atomic initialization, single-pass.
3235 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
3236 unsigned long start_pfn, enum memmap_context context)
3238 struct page *page;
3239 unsigned long end_pfn = start_pfn + size;
3240 unsigned long pfn;
3241 struct zone *z;
3243 if (highest_memmap_pfn < end_pfn - 1)
3244 highest_memmap_pfn = end_pfn - 1;
3246 z = &NODE_DATA(nid)->node_zones[zone];
3247 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3249 * There can be holes in boot-time mem_map[]s
3250 * handed to this function. They do not
3251 * exist on hotplugged memory.
3253 if (context == MEMMAP_EARLY) {
3254 if (!early_pfn_valid(pfn))
3255 continue;
3256 if (!early_pfn_in_nid(pfn, nid))
3257 continue;
3259 page = pfn_to_page(pfn);
3260 set_page_links(page, zone, nid, pfn);
3261 mminit_verify_page_links(page, zone, nid, pfn);
3262 init_page_count(page);
3263 reset_page_mapcount(page);
3264 SetPageReserved(page);
3266 * Mark the block movable so that blocks are reserved for
3267 * movable at startup. This will force kernel allocations
3268 * to reserve their blocks rather than leaking throughout
3269 * the address space during boot when many long-lived
3270 * kernel allocations are made. Later some blocks near
3271 * the start are marked MIGRATE_RESERVE by
3272 * setup_zone_migrate_reserve()
3274 * bitmap is created for zone's valid pfn range. but memmap
3275 * can be created for invalid pages (for alignment)
3276 * check here not to call set_pageblock_migratetype() against
3277 * pfn out of zone.
3279 if ((z->zone_start_pfn <= pfn)
3280 && (pfn < z->zone_start_pfn + z->spanned_pages)
3281 && !(pfn & (pageblock_nr_pages - 1)))
3282 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3284 INIT_LIST_HEAD(&page->lru);
3285 #ifdef WANT_PAGE_VIRTUAL
3286 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3287 if (!is_highmem_idx(zone))
3288 set_page_address(page, __va(pfn << PAGE_SHIFT));
3289 #endif
3293 static void __meminit zone_init_free_lists(struct zone *zone)
3295 int order, t;
3296 for_each_migratetype_order(order, t) {
3297 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
3298 zone->free_area[order].nr_free = 0;
3302 #ifndef __HAVE_ARCH_MEMMAP_INIT
3303 #define memmap_init(size, nid, zone, start_pfn) \
3304 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3305 #endif
3307 static int zone_batchsize(struct zone *zone)
3309 #ifdef CONFIG_MMU
3310 int batch;
3313 * The per-cpu-pages pools are set to around 1000th of the
3314 * size of the zone. But no more than 1/2 of a meg.
3316 * OK, so we don't know how big the cache is. So guess.
3318 batch = zone->present_pages / 1024;
3319 if (batch * PAGE_SIZE > 512 * 1024)
3320 batch = (512 * 1024) / PAGE_SIZE;
3321 batch /= 4; /* We effectively *= 4 below */
3322 if (batch < 1)
3323 batch = 1;
3326 * Clamp the batch to a 2^n - 1 value. Having a power
3327 * of 2 value was found to be more likely to have
3328 * suboptimal cache aliasing properties in some cases.
3330 * For example if 2 tasks are alternately allocating
3331 * batches of pages, one task can end up with a lot
3332 * of pages of one half of the possible page colors
3333 * and the other with pages of the other colors.
3335 batch = rounddown_pow_of_two(batch + batch/2) - 1;
3337 return batch;
3339 #else
3340 /* The deferral and batching of frees should be suppressed under NOMMU
3341 * conditions.
3343 * The problem is that NOMMU needs to be able to allocate large chunks
3344 * of contiguous memory as there's no hardware page translation to
3345 * assemble apparent contiguous memory from discontiguous pages.
3347 * Queueing large contiguous runs of pages for batching, however,
3348 * causes the pages to actually be freed in smaller chunks. As there
3349 * can be a significant delay between the individual batches being
3350 * recycled, this leads to the once large chunks of space being
3351 * fragmented and becoming unavailable for high-order allocations.
3353 return 0;
3354 #endif
3357 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
3359 struct per_cpu_pages *pcp;
3360 int migratetype;
3362 memset(p, 0, sizeof(*p));
3364 pcp = &p->pcp;
3365 pcp->count = 0;
3366 pcp->high = 6 * batch;
3367 pcp->batch = max(1UL, 1 * batch);
3368 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
3369 INIT_LIST_HEAD(&pcp->lists[migratetype]);
3373 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3374 * to the value high for the pageset p.
3377 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
3378 unsigned long high)
3380 struct per_cpu_pages *pcp;
3382 pcp = &p->pcp;
3383 pcp->high = high;
3384 pcp->batch = max(1UL, high/4);
3385 if ((high/4) > (PAGE_SHIFT * 8))
3386 pcp->batch = PAGE_SHIFT * 8;
3389 static __meminit void setup_zone_pageset(struct zone *zone)
3391 int cpu;
3393 zone->pageset = alloc_percpu(struct per_cpu_pageset);
3395 for_each_possible_cpu(cpu) {
3396 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
3398 setup_pageset(pcp, zone_batchsize(zone));
3400 if (percpu_pagelist_fraction)
3401 setup_pagelist_highmark(pcp,
3402 (zone->present_pages /
3403 percpu_pagelist_fraction));
3408 * Allocate per cpu pagesets and initialize them.
3409 * Before this call only boot pagesets were available.
3411 void __init setup_per_cpu_pageset(void)
3413 struct zone *zone;
3415 for_each_populated_zone(zone)
3416 setup_zone_pageset(zone);
3419 static noinline __init_refok
3420 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
3422 int i;
3423 struct pglist_data *pgdat = zone->zone_pgdat;
3424 size_t alloc_size;
3427 * The per-page waitqueue mechanism uses hashed waitqueues
3428 * per zone.
3430 zone->wait_table_hash_nr_entries =
3431 wait_table_hash_nr_entries(zone_size_pages);
3432 zone->wait_table_bits =
3433 wait_table_bits(zone->wait_table_hash_nr_entries);
3434 alloc_size = zone->wait_table_hash_nr_entries
3435 * sizeof(wait_queue_head_t);
3437 if (!slab_is_available()) {
3438 zone->wait_table = (wait_queue_head_t *)
3439 alloc_bootmem_node(pgdat, alloc_size);
3440 } else {
3442 * This case means that a zone whose size was 0 gets new memory
3443 * via memory hot-add.
3444 * But it may be the case that a new node was hot-added. In
3445 * this case vmalloc() will not be able to use this new node's
3446 * memory - this wait_table must be initialized to use this new
3447 * node itself as well.
3448 * To use this new node's memory, further consideration will be
3449 * necessary.
3451 zone->wait_table = vmalloc(alloc_size);
3453 if (!zone->wait_table)
3454 return -ENOMEM;
3456 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
3457 init_waitqueue_head(zone->wait_table + i);
3459 return 0;
3462 static int __zone_pcp_update(void *data)
3464 struct zone *zone = data;
3465 int cpu;
3466 unsigned long batch = zone_batchsize(zone), flags;
3468 for_each_possible_cpu(cpu) {
3469 struct per_cpu_pageset *pset;
3470 struct per_cpu_pages *pcp;
3472 pset = per_cpu_ptr(zone->pageset, cpu);
3473 pcp = &pset->pcp;
3475 local_irq_save(flags);
3476 free_pcppages_bulk(zone, pcp->count, pcp);
3477 setup_pageset(pset, batch);
3478 local_irq_restore(flags);
3480 return 0;
3483 void zone_pcp_update(struct zone *zone)
3485 stop_machine(__zone_pcp_update, zone, NULL);
3488 static __meminit void zone_pcp_init(struct zone *zone)
3491 * per cpu subsystem is not up at this point. The following code
3492 * relies on the ability of the linker to provide the
3493 * offset of a (static) per cpu variable into the per cpu area.
3495 zone->pageset = &boot_pageset;
3497 if (zone->present_pages)
3498 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
3499 zone->name, zone->present_pages,
3500 zone_batchsize(zone));
3503 __meminit int init_currently_empty_zone(struct zone *zone,
3504 unsigned long zone_start_pfn,
3505 unsigned long size,
3506 enum memmap_context context)
3508 struct pglist_data *pgdat = zone->zone_pgdat;
3509 int ret;
3510 ret = zone_wait_table_init(zone, size);
3511 if (ret)
3512 return ret;
3513 pgdat->nr_zones = zone_idx(zone) + 1;
3515 zone->zone_start_pfn = zone_start_pfn;
3517 mminit_dprintk(MMINIT_TRACE, "memmap_init",
3518 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
3519 pgdat->node_id,
3520 (unsigned long)zone_idx(zone),
3521 zone_start_pfn, (zone_start_pfn + size));
3523 zone_init_free_lists(zone);
3525 return 0;
3528 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3530 * Basic iterator support. Return the first range of PFNs for a node
3531 * Note: nid == MAX_NUMNODES returns first region regardless of node
3533 static int __meminit first_active_region_index_in_nid(int nid)
3535 int i;
3537 for (i = 0; i < nr_nodemap_entries; i++)
3538 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
3539 return i;
3541 return -1;
3545 * Basic iterator support. Return the next active range of PFNs for a node
3546 * Note: nid == MAX_NUMNODES returns next region regardless of node
3548 static int __meminit next_active_region_index_in_nid(int index, int nid)
3550 for (index = index + 1; index < nr_nodemap_entries; index++)
3551 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
3552 return index;
3554 return -1;
3557 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
3559 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
3560 * Architectures may implement their own version but if add_active_range()
3561 * was used and there are no special requirements, this is a convenient
3562 * alternative
3564 int __meminit __early_pfn_to_nid(unsigned long pfn)
3566 int i;
3568 for (i = 0; i < nr_nodemap_entries; i++) {
3569 unsigned long start_pfn = early_node_map[i].start_pfn;
3570 unsigned long end_pfn = early_node_map[i].end_pfn;
3572 if (start_pfn <= pfn && pfn < end_pfn)
3573 return early_node_map[i].nid;
3575 /* This is a memory hole */
3576 return -1;
3578 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
3580 int __meminit early_pfn_to_nid(unsigned long pfn)
3582 int nid;
3584 nid = __early_pfn_to_nid(pfn);
3585 if (nid >= 0)
3586 return nid;
3587 /* just returns 0 */
3588 return 0;
3591 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
3592 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
3594 int nid;
3596 nid = __early_pfn_to_nid(pfn);
3597 if (nid >= 0 && nid != node)
3598 return false;
3599 return true;
3601 #endif
3603 /* Basic iterator support to walk early_node_map[] */
3604 #define for_each_active_range_index_in_nid(i, nid) \
3605 for (i = first_active_region_index_in_nid(nid); i != -1; \
3606 i = next_active_region_index_in_nid(i, nid))
3609 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
3610 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
3611 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
3613 * If an architecture guarantees that all ranges registered with
3614 * add_active_ranges() contain no holes and may be freed, this
3615 * this function may be used instead of calling free_bootmem() manually.
3617 void __init free_bootmem_with_active_regions(int nid,
3618 unsigned long max_low_pfn)
3620 int i;
3622 for_each_active_range_index_in_nid(i, nid) {
3623 unsigned long size_pages = 0;
3624 unsigned long end_pfn = early_node_map[i].end_pfn;
3626 if (early_node_map[i].start_pfn >= max_low_pfn)
3627 continue;
3629 if (end_pfn > max_low_pfn)
3630 end_pfn = max_low_pfn;
3632 size_pages = end_pfn - early_node_map[i].start_pfn;
3633 free_bootmem_node(NODE_DATA(early_node_map[i].nid),
3634 PFN_PHYS(early_node_map[i].start_pfn),
3635 size_pages << PAGE_SHIFT);
3639 int __init add_from_early_node_map(struct range *range, int az,
3640 int nr_range, int nid)
3642 int i;
3643 u64 start, end;
3645 /* need to go over early_node_map to find out good range for node */
3646 for_each_active_range_index_in_nid(i, nid) {
3647 start = early_node_map[i].start_pfn;
3648 end = early_node_map[i].end_pfn;
3649 nr_range = add_range(range, az, nr_range, start, end);
3651 return nr_range;
3654 #ifdef CONFIG_NO_BOOTMEM
3655 void * __init __alloc_memory_core_early(int nid, u64 size, u64 align,
3656 u64 goal, u64 limit)
3658 int i;
3659 void *ptr;
3661 if (limit > get_max_mapped())
3662 limit = get_max_mapped();
3664 /* need to go over early_node_map to find out good range for node */
3665 for_each_active_range_index_in_nid(i, nid) {
3666 u64 addr;
3667 u64 ei_start, ei_last;
3669 ei_last = early_node_map[i].end_pfn;
3670 ei_last <<= PAGE_SHIFT;
3671 ei_start = early_node_map[i].start_pfn;
3672 ei_start <<= PAGE_SHIFT;
3673 addr = find_early_area(ei_start, ei_last,
3674 goal, limit, size, align);
3676 if (addr == -1ULL)
3677 continue;
3679 #if 0
3680 printk(KERN_DEBUG "alloc (nid=%d %llx - %llx) (%llx - %llx) %llx %llx => %llx\n",
3681 nid,
3682 ei_start, ei_last, goal, limit, size,
3683 align, addr);
3684 #endif
3686 ptr = phys_to_virt(addr);
3687 memset(ptr, 0, size);
3688 reserve_early_without_check(addr, addr + size, "BOOTMEM");
3690 * The min_count is set to 0 so that bootmem allocated blocks
3691 * are never reported as leaks.
3693 kmemleak_alloc(ptr, size, 0, 0);
3694 return ptr;
3697 return NULL;
3699 #endif
3702 void __init work_with_active_regions(int nid, work_fn_t work_fn, void *data)
3704 int i;
3705 int ret;
3707 for_each_active_range_index_in_nid(i, nid) {
3708 ret = work_fn(early_node_map[i].start_pfn,
3709 early_node_map[i].end_pfn, data);
3710 if (ret)
3711 break;
3715 * sparse_memory_present_with_active_regions - Call memory_present for each active range
3716 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
3718 * If an architecture guarantees that all ranges registered with
3719 * add_active_ranges() contain no holes and may be freed, this
3720 * function may be used instead of calling memory_present() manually.
3722 void __init sparse_memory_present_with_active_regions(int nid)
3724 int i;
3726 for_each_active_range_index_in_nid(i, nid)
3727 memory_present(early_node_map[i].nid,
3728 early_node_map[i].start_pfn,
3729 early_node_map[i].end_pfn);
3733 * get_pfn_range_for_nid - Return the start and end page frames for a node
3734 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
3735 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
3736 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
3738 * It returns the start and end page frame of a node based on information
3739 * provided by an arch calling add_active_range(). If called for a node
3740 * with no available memory, a warning is printed and the start and end
3741 * PFNs will be 0.
3743 void __meminit get_pfn_range_for_nid(unsigned int nid,
3744 unsigned long *start_pfn, unsigned long *end_pfn)
3746 int i;
3747 *start_pfn = -1UL;
3748 *end_pfn = 0;
3750 for_each_active_range_index_in_nid(i, nid) {
3751 *start_pfn = min(*start_pfn, early_node_map[i].start_pfn);
3752 *end_pfn = max(*end_pfn, early_node_map[i].end_pfn);
3755 if (*start_pfn == -1UL)
3756 *start_pfn = 0;
3760 * This finds a zone that can be used for ZONE_MOVABLE pages. The
3761 * assumption is made that zones within a node are ordered in monotonic
3762 * increasing memory addresses so that the "highest" populated zone is used
3764 static void __init find_usable_zone_for_movable(void)
3766 int zone_index;
3767 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
3768 if (zone_index == ZONE_MOVABLE)
3769 continue;
3771 if (arch_zone_highest_possible_pfn[zone_index] >
3772 arch_zone_lowest_possible_pfn[zone_index])
3773 break;
3776 VM_BUG_ON(zone_index == -1);
3777 movable_zone = zone_index;
3781 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
3782 * because it is sized independant of architecture. Unlike the other zones,
3783 * the starting point for ZONE_MOVABLE is not fixed. It may be different
3784 * in each node depending on the size of each node and how evenly kernelcore
3785 * is distributed. This helper function adjusts the zone ranges
3786 * provided by the architecture for a given node by using the end of the
3787 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
3788 * zones within a node are in order of monotonic increases memory addresses
3790 static void __meminit adjust_zone_range_for_zone_movable(int nid,
3791 unsigned long zone_type,
3792 unsigned long node_start_pfn,
3793 unsigned long node_end_pfn,
3794 unsigned long *zone_start_pfn,
3795 unsigned long *zone_end_pfn)
3797 /* Only adjust if ZONE_MOVABLE is on this node */
3798 if (zone_movable_pfn[nid]) {
3799 /* Size ZONE_MOVABLE */
3800 if (zone_type == ZONE_MOVABLE) {
3801 *zone_start_pfn = zone_movable_pfn[nid];
3802 *zone_end_pfn = min(node_end_pfn,
3803 arch_zone_highest_possible_pfn[movable_zone]);
3805 /* Adjust for ZONE_MOVABLE starting within this range */
3806 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
3807 *zone_end_pfn > zone_movable_pfn[nid]) {
3808 *zone_end_pfn = zone_movable_pfn[nid];
3810 /* Check if this whole range is within ZONE_MOVABLE */
3811 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
3812 *zone_start_pfn = *zone_end_pfn;
3817 * Return the number of pages a zone spans in a node, including holes
3818 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
3820 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
3821 unsigned long zone_type,
3822 unsigned long *ignored)
3824 unsigned long node_start_pfn, node_end_pfn;
3825 unsigned long zone_start_pfn, zone_end_pfn;
3827 /* Get the start and end of the node and zone */
3828 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
3829 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
3830 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
3831 adjust_zone_range_for_zone_movable(nid, zone_type,
3832 node_start_pfn, node_end_pfn,
3833 &zone_start_pfn, &zone_end_pfn);
3835 /* Check that this node has pages within the zone's required range */
3836 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
3837 return 0;
3839 /* Move the zone boundaries inside the node if necessary */
3840 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
3841 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
3843 /* Return the spanned pages */
3844 return zone_end_pfn - zone_start_pfn;
3848 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
3849 * then all holes in the requested range will be accounted for.
3851 unsigned long __meminit __absent_pages_in_range(int nid,
3852 unsigned long range_start_pfn,
3853 unsigned long range_end_pfn)
3855 int i = 0;
3856 unsigned long prev_end_pfn = 0, hole_pages = 0;
3857 unsigned long start_pfn;
3859 /* Find the end_pfn of the first active range of pfns in the node */
3860 i = first_active_region_index_in_nid(nid);
3861 if (i == -1)
3862 return 0;
3864 prev_end_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
3866 /* Account for ranges before physical memory on this node */
3867 if (early_node_map[i].start_pfn > range_start_pfn)
3868 hole_pages = prev_end_pfn - range_start_pfn;
3870 /* Find all holes for the zone within the node */
3871 for (; i != -1; i = next_active_region_index_in_nid(i, nid)) {
3873 /* No need to continue if prev_end_pfn is outside the zone */
3874 if (prev_end_pfn >= range_end_pfn)
3875 break;
3877 /* Make sure the end of the zone is not within the hole */
3878 start_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
3879 prev_end_pfn = max(prev_end_pfn, range_start_pfn);
3881 /* Update the hole size cound and move on */
3882 if (start_pfn > range_start_pfn) {
3883 BUG_ON(prev_end_pfn > start_pfn);
3884 hole_pages += start_pfn - prev_end_pfn;
3886 prev_end_pfn = early_node_map[i].end_pfn;
3889 /* Account for ranges past physical memory on this node */
3890 if (range_end_pfn > prev_end_pfn)
3891 hole_pages += range_end_pfn -
3892 max(range_start_pfn, prev_end_pfn);
3894 return hole_pages;
3898 * absent_pages_in_range - Return number of page frames in holes within a range
3899 * @start_pfn: The start PFN to start searching for holes
3900 * @end_pfn: The end PFN to stop searching for holes
3902 * It returns the number of pages frames in memory holes within a range.
3904 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
3905 unsigned long end_pfn)
3907 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
3910 /* Return the number of page frames in holes in a zone on a node */
3911 static unsigned long __meminit zone_absent_pages_in_node(int nid,
3912 unsigned long zone_type,
3913 unsigned long *ignored)
3915 unsigned long node_start_pfn, node_end_pfn;
3916 unsigned long zone_start_pfn, zone_end_pfn;
3918 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
3919 zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type],
3920 node_start_pfn);
3921 zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type],
3922 node_end_pfn);
3924 adjust_zone_range_for_zone_movable(nid, zone_type,
3925 node_start_pfn, node_end_pfn,
3926 &zone_start_pfn, &zone_end_pfn);
3927 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
3930 #else
3931 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
3932 unsigned long zone_type,
3933 unsigned long *zones_size)
3935 return zones_size[zone_type];
3938 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
3939 unsigned long zone_type,
3940 unsigned long *zholes_size)
3942 if (!zholes_size)
3943 return 0;
3945 return zholes_size[zone_type];
3948 #endif
3950 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
3951 unsigned long *zones_size, unsigned long *zholes_size)
3953 unsigned long realtotalpages, totalpages = 0;
3954 enum zone_type i;
3956 for (i = 0; i < MAX_NR_ZONES; i++)
3957 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
3958 zones_size);
3959 pgdat->node_spanned_pages = totalpages;
3961 realtotalpages = totalpages;
3962 for (i = 0; i < MAX_NR_ZONES; i++)
3963 realtotalpages -=
3964 zone_absent_pages_in_node(pgdat->node_id, i,
3965 zholes_size);
3966 pgdat->node_present_pages = realtotalpages;
3967 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
3968 realtotalpages);
3971 #ifndef CONFIG_SPARSEMEM
3973 * Calculate the size of the zone->blockflags rounded to an unsigned long
3974 * Start by making sure zonesize is a multiple of pageblock_order by rounding
3975 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
3976 * round what is now in bits to nearest long in bits, then return it in
3977 * bytes.
3979 static unsigned long __init usemap_size(unsigned long zonesize)
3981 unsigned long usemapsize;
3983 usemapsize = roundup(zonesize, pageblock_nr_pages);
3984 usemapsize = usemapsize >> pageblock_order;
3985 usemapsize *= NR_PAGEBLOCK_BITS;
3986 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
3988 return usemapsize / 8;
3991 static void __init setup_usemap(struct pglist_data *pgdat,
3992 struct zone *zone, unsigned long zonesize)
3994 unsigned long usemapsize = usemap_size(zonesize);
3995 zone->pageblock_flags = NULL;
3996 if (usemapsize)
3997 zone->pageblock_flags = alloc_bootmem_node(pgdat, usemapsize);
3999 #else
4000 static void inline setup_usemap(struct pglist_data *pgdat,
4001 struct zone *zone, unsigned long zonesize) {}
4002 #endif /* CONFIG_SPARSEMEM */
4004 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4006 /* Return a sensible default order for the pageblock size. */
4007 static inline int pageblock_default_order(void)
4009 if (HPAGE_SHIFT > PAGE_SHIFT)
4010 return HUGETLB_PAGE_ORDER;
4012 return MAX_ORDER-1;
4015 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4016 static inline void __init set_pageblock_order(unsigned int order)
4018 /* Check that pageblock_nr_pages has not already been setup */
4019 if (pageblock_order)
4020 return;
4023 * Assume the largest contiguous order of interest is a huge page.
4024 * This value may be variable depending on boot parameters on IA64
4026 pageblock_order = order;
4028 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4031 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4032 * and pageblock_default_order() are unused as pageblock_order is set
4033 * at compile-time. See include/linux/pageblock-flags.h for the values of
4034 * pageblock_order based on the kernel config
4036 static inline int pageblock_default_order(unsigned int order)
4038 return MAX_ORDER-1;
4040 #define set_pageblock_order(x) do {} while (0)
4042 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4045 * Set up the zone data structures:
4046 * - mark all pages reserved
4047 * - mark all memory queues empty
4048 * - clear the memory bitmaps
4050 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4051 unsigned long *zones_size, unsigned long *zholes_size)
4053 enum zone_type j;
4054 int nid = pgdat->node_id;
4055 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4056 int ret;
4058 pgdat_resize_init(pgdat);
4059 pgdat->nr_zones = 0;
4060 init_waitqueue_head(&pgdat->kswapd_wait);
4061 pgdat->kswapd_max_order = 0;
4062 pgdat_page_cgroup_init(pgdat);
4064 for (j = 0; j < MAX_NR_ZONES; j++) {
4065 struct zone *zone = pgdat->node_zones + j;
4066 unsigned long size, realsize, memmap_pages;
4067 enum lru_list l;
4069 size = zone_spanned_pages_in_node(nid, j, zones_size);
4070 realsize = size - zone_absent_pages_in_node(nid, j,
4071 zholes_size);
4074 * Adjust realsize so that it accounts for how much memory
4075 * is used by this zone for memmap. This affects the watermark
4076 * and per-cpu initialisations
4078 memmap_pages =
4079 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT;
4080 if (realsize >= memmap_pages) {
4081 realsize -= memmap_pages;
4082 if (memmap_pages)
4083 printk(KERN_DEBUG
4084 " %s zone: %lu pages used for memmap\n",
4085 zone_names[j], memmap_pages);
4086 } else
4087 printk(KERN_WARNING
4088 " %s zone: %lu pages exceeds realsize %lu\n",
4089 zone_names[j], memmap_pages, realsize);
4091 /* Account for reserved pages */
4092 if (j == 0 && realsize > dma_reserve) {
4093 realsize -= dma_reserve;
4094 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4095 zone_names[0], dma_reserve);
4098 if (!is_highmem_idx(j))
4099 nr_kernel_pages += realsize;
4100 nr_all_pages += realsize;
4102 zone->spanned_pages = size;
4103 zone->present_pages = realsize;
4104 #ifdef CONFIG_NUMA
4105 zone->node = nid;
4106 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
4107 / 100;
4108 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
4109 #endif
4110 zone->name = zone_names[j];
4111 spin_lock_init(&zone->lock);
4112 spin_lock_init(&zone->lru_lock);
4113 zone_seqlock_init(zone);
4114 zone->zone_pgdat = pgdat;
4116 zone_pcp_init(zone);
4117 for_each_lru(l) {
4118 INIT_LIST_HEAD(&zone->lru[l].list);
4119 zone->reclaim_stat.nr_saved_scan[l] = 0;
4121 zone->reclaim_stat.recent_rotated[0] = 0;
4122 zone->reclaim_stat.recent_rotated[1] = 0;
4123 zone->reclaim_stat.recent_scanned[0] = 0;
4124 zone->reclaim_stat.recent_scanned[1] = 0;
4125 zap_zone_vm_stats(zone);
4126 zone->flags = 0;
4127 if (!size)
4128 continue;
4130 set_pageblock_order(pageblock_default_order());
4131 setup_usemap(pgdat, zone, size);
4132 ret = init_currently_empty_zone(zone, zone_start_pfn,
4133 size, MEMMAP_EARLY);
4134 BUG_ON(ret);
4135 memmap_init(size, nid, j, zone_start_pfn);
4136 zone_start_pfn += size;
4140 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4142 /* Skip empty nodes */
4143 if (!pgdat->node_spanned_pages)
4144 return;
4146 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4147 /* ia64 gets its own node_mem_map, before this, without bootmem */
4148 if (!pgdat->node_mem_map) {
4149 unsigned long size, start, end;
4150 struct page *map;
4153 * The zone's endpoints aren't required to be MAX_ORDER
4154 * aligned but the node_mem_map endpoints must be in order
4155 * for the buddy allocator to function correctly.
4157 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4158 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
4159 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4160 size = (end - start) * sizeof(struct page);
4161 map = alloc_remap(pgdat->node_id, size);
4162 if (!map)
4163 map = alloc_bootmem_node(pgdat, size);
4164 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4166 #ifndef CONFIG_NEED_MULTIPLE_NODES
4168 * With no DISCONTIG, the global mem_map is just set as node 0's
4170 if (pgdat == NODE_DATA(0)) {
4171 mem_map = NODE_DATA(0)->node_mem_map;
4172 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
4173 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4174 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4175 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4177 #endif
4178 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4181 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4182 unsigned long node_start_pfn, unsigned long *zholes_size)
4184 pg_data_t *pgdat = NODE_DATA(nid);
4186 pgdat->node_id = nid;
4187 pgdat->node_start_pfn = node_start_pfn;
4188 calculate_node_totalpages(pgdat, zones_size, zholes_size);
4190 alloc_node_mem_map(pgdat);
4191 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4192 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4193 nid, (unsigned long)pgdat,
4194 (unsigned long)pgdat->node_mem_map);
4195 #endif
4197 free_area_init_core(pgdat, zones_size, zholes_size);
4200 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
4202 #if MAX_NUMNODES > 1
4204 * Figure out the number of possible node ids.
4206 static void __init setup_nr_node_ids(void)
4208 unsigned int node;
4209 unsigned int highest = 0;
4211 for_each_node_mask(node, node_possible_map)
4212 highest = node;
4213 nr_node_ids = highest + 1;
4215 #else
4216 static inline void setup_nr_node_ids(void)
4219 #endif
4222 * add_active_range - Register a range of PFNs backed by physical memory
4223 * @nid: The node ID the range resides on
4224 * @start_pfn: The start PFN of the available physical memory
4225 * @end_pfn: The end PFN of the available physical memory
4227 * These ranges are stored in an early_node_map[] and later used by
4228 * free_area_init_nodes() to calculate zone sizes and holes. If the
4229 * range spans a memory hole, it is up to the architecture to ensure
4230 * the memory is not freed by the bootmem allocator. If possible
4231 * the range being registered will be merged with existing ranges.
4233 void __init add_active_range(unsigned int nid, unsigned long start_pfn,
4234 unsigned long end_pfn)
4236 int i;
4238 mminit_dprintk(MMINIT_TRACE, "memory_register",
4239 "Entering add_active_range(%d, %#lx, %#lx) "
4240 "%d entries of %d used\n",
4241 nid, start_pfn, end_pfn,
4242 nr_nodemap_entries, MAX_ACTIVE_REGIONS);
4244 mminit_validate_memmodel_limits(&start_pfn, &end_pfn);
4246 /* Merge with existing active regions if possible */
4247 for (i = 0; i < nr_nodemap_entries; i++) {
4248 if (early_node_map[i].nid != nid)
4249 continue;
4251 /* Skip if an existing region covers this new one */
4252 if (start_pfn >= early_node_map[i].start_pfn &&
4253 end_pfn <= early_node_map[i].end_pfn)
4254 return;
4256 /* Merge forward if suitable */
4257 if (start_pfn <= early_node_map[i].end_pfn &&
4258 end_pfn > early_node_map[i].end_pfn) {
4259 early_node_map[i].end_pfn = end_pfn;
4260 return;
4263 /* Merge backward if suitable */
4264 if (start_pfn < early_node_map[i].start_pfn &&
4265 end_pfn >= early_node_map[i].start_pfn) {
4266 early_node_map[i].start_pfn = start_pfn;
4267 return;
4271 /* Check that early_node_map is large enough */
4272 if (i >= MAX_ACTIVE_REGIONS) {
4273 printk(KERN_CRIT "More than %d memory regions, truncating\n",
4274 MAX_ACTIVE_REGIONS);
4275 return;
4278 early_node_map[i].nid = nid;
4279 early_node_map[i].start_pfn = start_pfn;
4280 early_node_map[i].end_pfn = end_pfn;
4281 nr_nodemap_entries = i + 1;
4285 * remove_active_range - Shrink an existing registered range of PFNs
4286 * @nid: The node id the range is on that should be shrunk
4287 * @start_pfn: The new PFN of the range
4288 * @end_pfn: The new PFN of the range
4290 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
4291 * The map is kept near the end physical page range that has already been
4292 * registered. This function allows an arch to shrink an existing registered
4293 * range.
4295 void __init remove_active_range(unsigned int nid, unsigned long start_pfn,
4296 unsigned long end_pfn)
4298 int i, j;
4299 int removed = 0;
4301 printk(KERN_DEBUG "remove_active_range (%d, %lu, %lu)\n",
4302 nid, start_pfn, end_pfn);
4304 /* Find the old active region end and shrink */
4305 for_each_active_range_index_in_nid(i, nid) {
4306 if (early_node_map[i].start_pfn >= start_pfn &&
4307 early_node_map[i].end_pfn <= end_pfn) {
4308 /* clear it */
4309 early_node_map[i].start_pfn = 0;
4310 early_node_map[i].end_pfn = 0;
4311 removed = 1;
4312 continue;
4314 if (early_node_map[i].start_pfn < start_pfn &&
4315 early_node_map[i].end_pfn > start_pfn) {
4316 unsigned long temp_end_pfn = early_node_map[i].end_pfn;
4317 early_node_map[i].end_pfn = start_pfn;
4318 if (temp_end_pfn > end_pfn)
4319 add_active_range(nid, end_pfn, temp_end_pfn);
4320 continue;
4322 if (early_node_map[i].start_pfn >= start_pfn &&
4323 early_node_map[i].end_pfn > end_pfn &&
4324 early_node_map[i].start_pfn < end_pfn) {
4325 early_node_map[i].start_pfn = end_pfn;
4326 continue;
4330 if (!removed)
4331 return;
4333 /* remove the blank ones */
4334 for (i = nr_nodemap_entries - 1; i > 0; i--) {
4335 if (early_node_map[i].nid != nid)
4336 continue;
4337 if (early_node_map[i].end_pfn)
4338 continue;
4339 /* we found it, get rid of it */
4340 for (j = i; j < nr_nodemap_entries - 1; j++)
4341 memcpy(&early_node_map[j], &early_node_map[j+1],
4342 sizeof(early_node_map[j]));
4343 j = nr_nodemap_entries - 1;
4344 memset(&early_node_map[j], 0, sizeof(early_node_map[j]));
4345 nr_nodemap_entries--;
4350 * remove_all_active_ranges - Remove all currently registered regions
4352 * During discovery, it may be found that a table like SRAT is invalid
4353 * and an alternative discovery method must be used. This function removes
4354 * all currently registered regions.
4356 void __init remove_all_active_ranges(void)
4358 memset(early_node_map, 0, sizeof(early_node_map));
4359 nr_nodemap_entries = 0;
4362 /* Compare two active node_active_regions */
4363 static int __init cmp_node_active_region(const void *a, const void *b)
4365 struct node_active_region *arange = (struct node_active_region *)a;
4366 struct node_active_region *brange = (struct node_active_region *)b;
4368 /* Done this way to avoid overflows */
4369 if (arange->start_pfn > brange->start_pfn)
4370 return 1;
4371 if (arange->start_pfn < brange->start_pfn)
4372 return -1;
4374 return 0;
4377 /* sort the node_map by start_pfn */
4378 void __init sort_node_map(void)
4380 sort(early_node_map, (size_t)nr_nodemap_entries,
4381 sizeof(struct node_active_region),
4382 cmp_node_active_region, NULL);
4385 /* Find the lowest pfn for a node */
4386 static unsigned long __init find_min_pfn_for_node(int nid)
4388 int i;
4389 unsigned long min_pfn = ULONG_MAX;
4391 /* Assuming a sorted map, the first range found has the starting pfn */
4392 for_each_active_range_index_in_nid(i, nid)
4393 min_pfn = min(min_pfn, early_node_map[i].start_pfn);
4395 if (min_pfn == ULONG_MAX) {
4396 printk(KERN_WARNING
4397 "Could not find start_pfn for node %d\n", nid);
4398 return 0;
4401 return min_pfn;
4405 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4407 * It returns the minimum PFN based on information provided via
4408 * add_active_range().
4410 unsigned long __init find_min_pfn_with_active_regions(void)
4412 return find_min_pfn_for_node(MAX_NUMNODES);
4416 * early_calculate_totalpages()
4417 * Sum pages in active regions for movable zone.
4418 * Populate N_HIGH_MEMORY for calculating usable_nodes.
4420 static unsigned long __init early_calculate_totalpages(void)
4422 int i;
4423 unsigned long totalpages = 0;
4425 for (i = 0; i < nr_nodemap_entries; i++) {
4426 unsigned long pages = early_node_map[i].end_pfn -
4427 early_node_map[i].start_pfn;
4428 totalpages += pages;
4429 if (pages)
4430 node_set_state(early_node_map[i].nid, N_HIGH_MEMORY);
4432 return totalpages;
4436 * Find the PFN the Movable zone begins in each node. Kernel memory
4437 * is spread evenly between nodes as long as the nodes have enough
4438 * memory. When they don't, some nodes will have more kernelcore than
4439 * others
4441 static void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn)
4443 int i, nid;
4444 unsigned long usable_startpfn;
4445 unsigned long kernelcore_node, kernelcore_remaining;
4446 /* save the state before borrow the nodemask */
4447 nodemask_t saved_node_state = node_states[N_HIGH_MEMORY];
4448 unsigned long totalpages = early_calculate_totalpages();
4449 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
4452 * If movablecore was specified, calculate what size of
4453 * kernelcore that corresponds so that memory usable for
4454 * any allocation type is evenly spread. If both kernelcore
4455 * and movablecore are specified, then the value of kernelcore
4456 * will be used for required_kernelcore if it's greater than
4457 * what movablecore would have allowed.
4459 if (required_movablecore) {
4460 unsigned long corepages;
4463 * Round-up so that ZONE_MOVABLE is at least as large as what
4464 * was requested by the user
4466 required_movablecore =
4467 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
4468 corepages = totalpages - required_movablecore;
4470 required_kernelcore = max(required_kernelcore, corepages);
4473 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4474 if (!required_kernelcore)
4475 goto out;
4477 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4478 find_usable_zone_for_movable();
4479 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
4481 restart:
4482 /* Spread kernelcore memory as evenly as possible throughout nodes */
4483 kernelcore_node = required_kernelcore / usable_nodes;
4484 for_each_node_state(nid, N_HIGH_MEMORY) {
4486 * Recalculate kernelcore_node if the division per node
4487 * now exceeds what is necessary to satisfy the requested
4488 * amount of memory for the kernel
4490 if (required_kernelcore < kernelcore_node)
4491 kernelcore_node = required_kernelcore / usable_nodes;
4494 * As the map is walked, we track how much memory is usable
4495 * by the kernel using kernelcore_remaining. When it is
4496 * 0, the rest of the node is usable by ZONE_MOVABLE
4498 kernelcore_remaining = kernelcore_node;
4500 /* Go through each range of PFNs within this node */
4501 for_each_active_range_index_in_nid(i, nid) {
4502 unsigned long start_pfn, end_pfn;
4503 unsigned long size_pages;
4505 start_pfn = max(early_node_map[i].start_pfn,
4506 zone_movable_pfn[nid]);
4507 end_pfn = early_node_map[i].end_pfn;
4508 if (start_pfn >= end_pfn)
4509 continue;
4511 /* Account for what is only usable for kernelcore */
4512 if (start_pfn < usable_startpfn) {
4513 unsigned long kernel_pages;
4514 kernel_pages = min(end_pfn, usable_startpfn)
4515 - start_pfn;
4517 kernelcore_remaining -= min(kernel_pages,
4518 kernelcore_remaining);
4519 required_kernelcore -= min(kernel_pages,
4520 required_kernelcore);
4522 /* Continue if range is now fully accounted */
4523 if (end_pfn <= usable_startpfn) {
4526 * Push zone_movable_pfn to the end so
4527 * that if we have to rebalance
4528 * kernelcore across nodes, we will
4529 * not double account here
4531 zone_movable_pfn[nid] = end_pfn;
4532 continue;
4534 start_pfn = usable_startpfn;
4538 * The usable PFN range for ZONE_MOVABLE is from
4539 * start_pfn->end_pfn. Calculate size_pages as the
4540 * number of pages used as kernelcore
4542 size_pages = end_pfn - start_pfn;
4543 if (size_pages > kernelcore_remaining)
4544 size_pages = kernelcore_remaining;
4545 zone_movable_pfn[nid] = start_pfn + size_pages;
4548 * Some kernelcore has been met, update counts and
4549 * break if the kernelcore for this node has been
4550 * satisified
4552 required_kernelcore -= min(required_kernelcore,
4553 size_pages);
4554 kernelcore_remaining -= size_pages;
4555 if (!kernelcore_remaining)
4556 break;
4561 * If there is still required_kernelcore, we do another pass with one
4562 * less node in the count. This will push zone_movable_pfn[nid] further
4563 * along on the nodes that still have memory until kernelcore is
4564 * satisified
4566 usable_nodes--;
4567 if (usable_nodes && required_kernelcore > usable_nodes)
4568 goto restart;
4570 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4571 for (nid = 0; nid < MAX_NUMNODES; nid++)
4572 zone_movable_pfn[nid] =
4573 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
4575 out:
4576 /* restore the node_state */
4577 node_states[N_HIGH_MEMORY] = saved_node_state;
4580 /* Any regular memory on that node ? */
4581 static void check_for_regular_memory(pg_data_t *pgdat)
4583 #ifdef CONFIG_HIGHMEM
4584 enum zone_type zone_type;
4586 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) {
4587 struct zone *zone = &pgdat->node_zones[zone_type];
4588 if (zone->present_pages)
4589 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY);
4591 #endif
4595 * free_area_init_nodes - Initialise all pg_data_t and zone data
4596 * @max_zone_pfn: an array of max PFNs for each zone
4598 * This will call free_area_init_node() for each active node in the system.
4599 * Using the page ranges provided by add_active_range(), the size of each
4600 * zone in each node and their holes is calculated. If the maximum PFN
4601 * between two adjacent zones match, it is assumed that the zone is empty.
4602 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4603 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4604 * starts where the previous one ended. For example, ZONE_DMA32 starts
4605 * at arch_max_dma_pfn.
4607 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
4609 unsigned long nid;
4610 int i;
4612 /* Sort early_node_map as initialisation assumes it is sorted */
4613 sort_node_map();
4615 /* Record where the zone boundaries are */
4616 memset(arch_zone_lowest_possible_pfn, 0,
4617 sizeof(arch_zone_lowest_possible_pfn));
4618 memset(arch_zone_highest_possible_pfn, 0,
4619 sizeof(arch_zone_highest_possible_pfn));
4620 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
4621 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
4622 for (i = 1; i < MAX_NR_ZONES; i++) {
4623 if (i == ZONE_MOVABLE)
4624 continue;
4625 arch_zone_lowest_possible_pfn[i] =
4626 arch_zone_highest_possible_pfn[i-1];
4627 arch_zone_highest_possible_pfn[i] =
4628 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
4630 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
4631 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
4633 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4634 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
4635 find_zone_movable_pfns_for_nodes(zone_movable_pfn);
4637 /* Print out the zone ranges */
4638 printk("Zone PFN ranges:\n");
4639 for (i = 0; i < MAX_NR_ZONES; i++) {
4640 if (i == ZONE_MOVABLE)
4641 continue;
4642 printk(" %-8s ", zone_names[i]);
4643 if (arch_zone_lowest_possible_pfn[i] ==
4644 arch_zone_highest_possible_pfn[i])
4645 printk("empty\n");
4646 else
4647 printk("%0#10lx -> %0#10lx\n",
4648 arch_zone_lowest_possible_pfn[i],
4649 arch_zone_highest_possible_pfn[i]);
4652 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4653 printk("Movable zone start PFN for each node\n");
4654 for (i = 0; i < MAX_NUMNODES; i++) {
4655 if (zone_movable_pfn[i])
4656 printk(" Node %d: %lu\n", i, zone_movable_pfn[i]);
4659 /* Print out the early_node_map[] */
4660 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries);
4661 for (i = 0; i < nr_nodemap_entries; i++)
4662 printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map[i].nid,
4663 early_node_map[i].start_pfn,
4664 early_node_map[i].end_pfn);
4666 /* Initialise every node */
4667 mminit_verify_pageflags_layout();
4668 setup_nr_node_ids();
4669 for_each_online_node(nid) {
4670 pg_data_t *pgdat = NODE_DATA(nid);
4671 free_area_init_node(nid, NULL,
4672 find_min_pfn_for_node(nid), NULL);
4674 /* Any memory on that node */
4675 if (pgdat->node_present_pages)
4676 node_set_state(nid, N_HIGH_MEMORY);
4677 check_for_regular_memory(pgdat);
4681 static int __init cmdline_parse_core(char *p, unsigned long *core)
4683 unsigned long long coremem;
4684 if (!p)
4685 return -EINVAL;
4687 coremem = memparse(p, &p);
4688 *core = coremem >> PAGE_SHIFT;
4690 /* Paranoid check that UL is enough for the coremem value */
4691 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
4693 return 0;
4697 * kernelcore=size sets the amount of memory for use for allocations that
4698 * cannot be reclaimed or migrated.
4700 static int __init cmdline_parse_kernelcore(char *p)
4702 return cmdline_parse_core(p, &required_kernelcore);
4706 * movablecore=size sets the amount of memory for use for allocations that
4707 * can be reclaimed or migrated.
4709 static int __init cmdline_parse_movablecore(char *p)
4711 return cmdline_parse_core(p, &required_movablecore);
4714 early_param("kernelcore", cmdline_parse_kernelcore);
4715 early_param("movablecore", cmdline_parse_movablecore);
4717 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4720 * set_dma_reserve - set the specified number of pages reserved in the first zone
4721 * @new_dma_reserve: The number of pages to mark reserved
4723 * The per-cpu batchsize and zone watermarks are determined by present_pages.
4724 * In the DMA zone, a significant percentage may be consumed by kernel image
4725 * and other unfreeable allocations which can skew the watermarks badly. This
4726 * function may optionally be used to account for unfreeable pages in the
4727 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
4728 * smaller per-cpu batchsize.
4730 void __init set_dma_reserve(unsigned long new_dma_reserve)
4732 dma_reserve = new_dma_reserve;
4735 #ifndef CONFIG_NEED_MULTIPLE_NODES
4736 struct pglist_data __refdata contig_page_data = {
4737 #ifndef CONFIG_NO_BOOTMEM
4738 .bdata = &bootmem_node_data[0]
4739 #endif
4741 EXPORT_SYMBOL(contig_page_data);
4742 #endif
4744 void __init free_area_init(unsigned long *zones_size)
4746 free_area_init_node(0, zones_size,
4747 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
4750 static int page_alloc_cpu_notify(struct notifier_block *self,
4751 unsigned long action, void *hcpu)
4753 int cpu = (unsigned long)hcpu;
4755 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
4756 drain_pages(cpu);
4759 * Spill the event counters of the dead processor
4760 * into the current processors event counters.
4761 * This artificially elevates the count of the current
4762 * processor.
4764 vm_events_fold_cpu(cpu);
4767 * Zero the differential counters of the dead processor
4768 * so that the vm statistics are consistent.
4770 * This is only okay since the processor is dead and cannot
4771 * race with what we are doing.
4773 refresh_cpu_vm_stats(cpu);
4775 return NOTIFY_OK;
4778 void __init page_alloc_init(void)
4780 hotcpu_notifier(page_alloc_cpu_notify, 0);
4784 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
4785 * or min_free_kbytes changes.
4787 static void calculate_totalreserve_pages(void)
4789 struct pglist_data *pgdat;
4790 unsigned long reserve_pages = 0;
4791 enum zone_type i, j;
4793 for_each_online_pgdat(pgdat) {
4794 for (i = 0; i < MAX_NR_ZONES; i++) {
4795 struct zone *zone = pgdat->node_zones + i;
4796 unsigned long max = 0;
4798 /* Find valid and maximum lowmem_reserve in the zone */
4799 for (j = i; j < MAX_NR_ZONES; j++) {
4800 if (zone->lowmem_reserve[j] > max)
4801 max = zone->lowmem_reserve[j];
4804 /* we treat the high watermark as reserved pages. */
4805 max += high_wmark_pages(zone);
4807 if (max > zone->present_pages)
4808 max = zone->present_pages;
4809 reserve_pages += max;
4812 totalreserve_pages = reserve_pages;
4816 * setup_per_zone_lowmem_reserve - called whenever
4817 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
4818 * has a correct pages reserved value, so an adequate number of
4819 * pages are left in the zone after a successful __alloc_pages().
4821 static void setup_per_zone_lowmem_reserve(void)
4823 struct pglist_data *pgdat;
4824 enum zone_type j, idx;
4826 for_each_online_pgdat(pgdat) {
4827 for (j = 0; j < MAX_NR_ZONES; j++) {
4828 struct zone *zone = pgdat->node_zones + j;
4829 unsigned long present_pages = zone->present_pages;
4831 zone->lowmem_reserve[j] = 0;
4833 idx = j;
4834 while (idx) {
4835 struct zone *lower_zone;
4837 idx--;
4839 if (sysctl_lowmem_reserve_ratio[idx] < 1)
4840 sysctl_lowmem_reserve_ratio[idx] = 1;
4842 lower_zone = pgdat->node_zones + idx;
4843 lower_zone->lowmem_reserve[j] = present_pages /
4844 sysctl_lowmem_reserve_ratio[idx];
4845 present_pages += lower_zone->present_pages;
4850 /* update totalreserve_pages */
4851 calculate_totalreserve_pages();
4855 * setup_per_zone_wmarks - called when min_free_kbytes changes
4856 * or when memory is hot-{added|removed}
4858 * Ensures that the watermark[min,low,high] values for each zone are set
4859 * correctly with respect to min_free_kbytes.
4861 void setup_per_zone_wmarks(void)
4863 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
4864 unsigned long lowmem_pages = 0;
4865 struct zone *zone;
4866 unsigned long flags;
4868 /* Calculate total number of !ZONE_HIGHMEM pages */
4869 for_each_zone(zone) {
4870 if (!is_highmem(zone))
4871 lowmem_pages += zone->present_pages;
4874 for_each_zone(zone) {
4875 u64 tmp;
4877 spin_lock_irqsave(&zone->lock, flags);
4878 tmp = (u64)pages_min * zone->present_pages;
4879 do_div(tmp, lowmem_pages);
4880 if (is_highmem(zone)) {
4882 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
4883 * need highmem pages, so cap pages_min to a small
4884 * value here.
4886 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
4887 * deltas controls asynch page reclaim, and so should
4888 * not be capped for highmem.
4890 int min_pages;
4892 min_pages = zone->present_pages / 1024;
4893 if (min_pages < SWAP_CLUSTER_MAX)
4894 min_pages = SWAP_CLUSTER_MAX;
4895 if (min_pages > 128)
4896 min_pages = 128;
4897 zone->watermark[WMARK_MIN] = min_pages;
4898 } else {
4900 * If it's a lowmem zone, reserve a number of pages
4901 * proportionate to the zone's size.
4903 zone->watermark[WMARK_MIN] = tmp;
4906 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
4907 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
4908 setup_zone_migrate_reserve(zone);
4909 spin_unlock_irqrestore(&zone->lock, flags);
4912 /* update totalreserve_pages */
4913 calculate_totalreserve_pages();
4917 * The inactive anon list should be small enough that the VM never has to
4918 * do too much work, but large enough that each inactive page has a chance
4919 * to be referenced again before it is swapped out.
4921 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
4922 * INACTIVE_ANON pages on this zone's LRU, maintained by the
4923 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
4924 * the anonymous pages are kept on the inactive list.
4926 * total target max
4927 * memory ratio inactive anon
4928 * -------------------------------------
4929 * 10MB 1 5MB
4930 * 100MB 1 50MB
4931 * 1GB 3 250MB
4932 * 10GB 10 0.9GB
4933 * 100GB 31 3GB
4934 * 1TB 101 10GB
4935 * 10TB 320 32GB
4937 void calculate_zone_inactive_ratio(struct zone *zone)
4939 unsigned int gb, ratio;
4941 /* Zone size in gigabytes */
4942 gb = zone->present_pages >> (30 - PAGE_SHIFT);
4943 if (gb)
4944 ratio = int_sqrt(10 * gb);
4945 else
4946 ratio = 1;
4948 zone->inactive_ratio = ratio;
4951 static void __init setup_per_zone_inactive_ratio(void)
4953 struct zone *zone;
4955 for_each_zone(zone)
4956 calculate_zone_inactive_ratio(zone);
4960 * Initialise min_free_kbytes.
4962 * For small machines we want it small (128k min). For large machines
4963 * we want it large (64MB max). But it is not linear, because network
4964 * bandwidth does not increase linearly with machine size. We use
4966 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
4967 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
4969 * which yields
4971 * 16MB: 512k
4972 * 32MB: 724k
4973 * 64MB: 1024k
4974 * 128MB: 1448k
4975 * 256MB: 2048k
4976 * 512MB: 2896k
4977 * 1024MB: 4096k
4978 * 2048MB: 5792k
4979 * 4096MB: 8192k
4980 * 8192MB: 11584k
4981 * 16384MB: 16384k
4983 static int __init init_per_zone_wmark_min(void)
4985 unsigned long lowmem_kbytes;
4987 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
4989 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
4990 if (min_free_kbytes < 128)
4991 min_free_kbytes = 128;
4992 if (min_free_kbytes > 65536)
4993 min_free_kbytes = 65536;
4994 setup_per_zone_wmarks();
4995 setup_per_zone_lowmem_reserve();
4996 setup_per_zone_inactive_ratio();
4997 return 0;
4999 module_init(init_per_zone_wmark_min)
5002 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5003 * that we can call two helper functions whenever min_free_kbytes
5004 * changes.
5006 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
5007 void __user *buffer, size_t *length, loff_t *ppos)
5009 proc_dointvec(table, write, buffer, length, ppos);
5010 if (write)
5011 setup_per_zone_wmarks();
5012 return 0;
5015 #ifdef CONFIG_NUMA
5016 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
5017 void __user *buffer, size_t *length, loff_t *ppos)
5019 struct zone *zone;
5020 int rc;
5022 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5023 if (rc)
5024 return rc;
5026 for_each_zone(zone)
5027 zone->min_unmapped_pages = (zone->present_pages *
5028 sysctl_min_unmapped_ratio) / 100;
5029 return 0;
5032 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
5033 void __user *buffer, size_t *length, loff_t *ppos)
5035 struct zone *zone;
5036 int rc;
5038 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5039 if (rc)
5040 return rc;
5042 for_each_zone(zone)
5043 zone->min_slab_pages = (zone->present_pages *
5044 sysctl_min_slab_ratio) / 100;
5045 return 0;
5047 #endif
5050 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5051 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5052 * whenever sysctl_lowmem_reserve_ratio changes.
5054 * The reserve ratio obviously has absolutely no relation with the
5055 * minimum watermarks. The lowmem reserve ratio can only make sense
5056 * if in function of the boot time zone sizes.
5058 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
5059 void __user *buffer, size_t *length, loff_t *ppos)
5061 proc_dointvec_minmax(table, write, buffer, length, ppos);
5062 setup_per_zone_lowmem_reserve();
5063 return 0;
5067 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5068 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
5069 * can have before it gets flushed back to buddy allocator.
5072 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
5073 void __user *buffer, size_t *length, loff_t *ppos)
5075 struct zone *zone;
5076 unsigned int cpu;
5077 int ret;
5079 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5080 if (!write || (ret == -EINVAL))
5081 return ret;
5082 for_each_populated_zone(zone) {
5083 for_each_possible_cpu(cpu) {
5084 unsigned long high;
5085 high = zone->present_pages / percpu_pagelist_fraction;
5086 setup_pagelist_highmark(
5087 per_cpu_ptr(zone->pageset, cpu), high);
5090 return 0;
5093 int hashdist = HASHDIST_DEFAULT;
5095 #ifdef CONFIG_NUMA
5096 static int __init set_hashdist(char *str)
5098 if (!str)
5099 return 0;
5100 hashdist = simple_strtoul(str, &str, 0);
5101 return 1;
5103 __setup("hashdist=", set_hashdist);
5104 #endif
5107 * allocate a large system hash table from bootmem
5108 * - it is assumed that the hash table must contain an exact power-of-2
5109 * quantity of entries
5110 * - limit is the number of hash buckets, not the total allocation size
5112 void *__init alloc_large_system_hash(const char *tablename,
5113 unsigned long bucketsize,
5114 unsigned long numentries,
5115 int scale,
5116 int flags,
5117 unsigned int *_hash_shift,
5118 unsigned int *_hash_mask,
5119 unsigned long limit)
5121 unsigned long long max = limit;
5122 unsigned long log2qty, size;
5123 void *table = NULL;
5125 /* allow the kernel cmdline to have a say */
5126 if (!numentries) {
5127 /* round applicable memory size up to nearest megabyte */
5128 numentries = nr_kernel_pages;
5129 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
5130 numentries >>= 20 - PAGE_SHIFT;
5131 numentries <<= 20 - PAGE_SHIFT;
5133 /* limit to 1 bucket per 2^scale bytes of low memory */
5134 if (scale > PAGE_SHIFT)
5135 numentries >>= (scale - PAGE_SHIFT);
5136 else
5137 numentries <<= (PAGE_SHIFT - scale);
5139 /* Make sure we've got at least a 0-order allocation.. */
5140 if (unlikely(flags & HASH_SMALL)) {
5141 /* Makes no sense without HASH_EARLY */
5142 WARN_ON(!(flags & HASH_EARLY));
5143 if (!(numentries >> *_hash_shift)) {
5144 numentries = 1UL << *_hash_shift;
5145 BUG_ON(!numentries);
5147 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
5148 numentries = PAGE_SIZE / bucketsize;
5150 numentries = roundup_pow_of_two(numentries);
5152 /* limit allocation size to 1/16 total memory by default */
5153 if (max == 0) {
5154 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5155 do_div(max, bucketsize);
5158 if (numentries > max)
5159 numentries = max;
5161 log2qty = ilog2(numentries);
5163 do {
5164 size = bucketsize << log2qty;
5165 if (flags & HASH_EARLY)
5166 table = alloc_bootmem_nopanic(size);
5167 else if (hashdist)
5168 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5169 else {
5171 * If bucketsize is not a power-of-two, we may free
5172 * some pages at the end of hash table which
5173 * alloc_pages_exact() automatically does
5175 if (get_order(size) < MAX_ORDER) {
5176 table = alloc_pages_exact(size, GFP_ATOMIC);
5177 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5180 } while (!table && size > PAGE_SIZE && --log2qty);
5182 if (!table)
5183 panic("Failed to allocate %s hash table\n", tablename);
5185 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
5186 tablename,
5187 (1UL << log2qty),
5188 ilog2(size) - PAGE_SHIFT,
5189 size);
5191 if (_hash_shift)
5192 *_hash_shift = log2qty;
5193 if (_hash_mask)
5194 *_hash_mask = (1 << log2qty) - 1;
5196 return table;
5199 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5200 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
5201 unsigned long pfn)
5203 #ifdef CONFIG_SPARSEMEM
5204 return __pfn_to_section(pfn)->pageblock_flags;
5205 #else
5206 return zone->pageblock_flags;
5207 #endif /* CONFIG_SPARSEMEM */
5210 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
5212 #ifdef CONFIG_SPARSEMEM
5213 pfn &= (PAGES_PER_SECTION-1);
5214 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5215 #else
5216 pfn = pfn - zone->zone_start_pfn;
5217 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5218 #endif /* CONFIG_SPARSEMEM */
5222 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5223 * @page: The page within the block of interest
5224 * @start_bitidx: The first bit of interest to retrieve
5225 * @end_bitidx: The last bit of interest
5226 * returns pageblock_bits flags
5228 unsigned long get_pageblock_flags_group(struct page *page,
5229 int start_bitidx, int end_bitidx)
5231 struct zone *zone;
5232 unsigned long *bitmap;
5233 unsigned long pfn, bitidx;
5234 unsigned long flags = 0;
5235 unsigned long value = 1;
5237 zone = page_zone(page);
5238 pfn = page_to_pfn(page);
5239 bitmap = get_pageblock_bitmap(zone, pfn);
5240 bitidx = pfn_to_bitidx(zone, pfn);
5242 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5243 if (test_bit(bitidx + start_bitidx, bitmap))
5244 flags |= value;
5246 return flags;
5250 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5251 * @page: The page within the block of interest
5252 * @start_bitidx: The first bit of interest
5253 * @end_bitidx: The last bit of interest
5254 * @flags: The flags to set
5256 void set_pageblock_flags_group(struct page *page, unsigned long flags,
5257 int start_bitidx, int end_bitidx)
5259 struct zone *zone;
5260 unsigned long *bitmap;
5261 unsigned long pfn, bitidx;
5262 unsigned long value = 1;
5264 zone = page_zone(page);
5265 pfn = page_to_pfn(page);
5266 bitmap = get_pageblock_bitmap(zone, pfn);
5267 bitidx = pfn_to_bitidx(zone, pfn);
5268 VM_BUG_ON(pfn < zone->zone_start_pfn);
5269 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages);
5271 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5272 if (flags & value)
5273 __set_bit(bitidx + start_bitidx, bitmap);
5274 else
5275 __clear_bit(bitidx + start_bitidx, bitmap);
5279 * This is designed as sub function...plz see page_isolation.c also.
5280 * set/clear page block's type to be ISOLATE.
5281 * page allocater never alloc memory from ISOLATE block.
5284 int set_migratetype_isolate(struct page *page)
5286 struct zone *zone;
5287 struct page *curr_page;
5288 unsigned long flags, pfn, iter;
5289 unsigned long immobile = 0;
5290 struct memory_isolate_notify arg;
5291 int notifier_ret;
5292 int ret = -EBUSY;
5293 int zone_idx;
5295 zone = page_zone(page);
5296 zone_idx = zone_idx(zone);
5298 spin_lock_irqsave(&zone->lock, flags);
5299 if (get_pageblock_migratetype(page) == MIGRATE_MOVABLE ||
5300 zone_idx == ZONE_MOVABLE) {
5301 ret = 0;
5302 goto out;
5305 pfn = page_to_pfn(page);
5306 arg.start_pfn = pfn;
5307 arg.nr_pages = pageblock_nr_pages;
5308 arg.pages_found = 0;
5311 * It may be possible to isolate a pageblock even if the
5312 * migratetype is not MIGRATE_MOVABLE. The memory isolation
5313 * notifier chain is used by balloon drivers to return the
5314 * number of pages in a range that are held by the balloon
5315 * driver to shrink memory. If all the pages are accounted for
5316 * by balloons, are free, or on the LRU, isolation can continue.
5317 * Later, for example, when memory hotplug notifier runs, these
5318 * pages reported as "can be isolated" should be isolated(freed)
5319 * by the balloon driver through the memory notifier chain.
5321 notifier_ret = memory_isolate_notify(MEM_ISOLATE_COUNT, &arg);
5322 notifier_ret = notifier_to_errno(notifier_ret);
5323 if (notifier_ret || !arg.pages_found)
5324 goto out;
5326 for (iter = pfn; iter < (pfn + pageblock_nr_pages); iter++) {
5327 if (!pfn_valid_within(pfn))
5328 continue;
5330 curr_page = pfn_to_page(iter);
5331 if (!page_count(curr_page) || PageLRU(curr_page))
5332 continue;
5334 immobile++;
5337 if (arg.pages_found == immobile)
5338 ret = 0;
5340 out:
5341 if (!ret) {
5342 set_pageblock_migratetype(page, MIGRATE_ISOLATE);
5343 move_freepages_block(zone, page, MIGRATE_ISOLATE);
5346 spin_unlock_irqrestore(&zone->lock, flags);
5347 if (!ret)
5348 drain_all_pages();
5349 return ret;
5352 void unset_migratetype_isolate(struct page *page)
5354 struct zone *zone;
5355 unsigned long flags;
5356 zone = page_zone(page);
5357 spin_lock_irqsave(&zone->lock, flags);
5358 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
5359 goto out;
5360 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5361 move_freepages_block(zone, page, MIGRATE_MOVABLE);
5362 out:
5363 spin_unlock_irqrestore(&zone->lock, flags);
5366 #ifdef CONFIG_MEMORY_HOTREMOVE
5368 * All pages in the range must be isolated before calling this.
5370 void
5371 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
5373 struct page *page;
5374 struct zone *zone;
5375 int order, i;
5376 unsigned long pfn;
5377 unsigned long flags;
5378 /* find the first valid pfn */
5379 for (pfn = start_pfn; pfn < end_pfn; pfn++)
5380 if (pfn_valid(pfn))
5381 break;
5382 if (pfn == end_pfn)
5383 return;
5384 zone = page_zone(pfn_to_page(pfn));
5385 spin_lock_irqsave(&zone->lock, flags);
5386 pfn = start_pfn;
5387 while (pfn < end_pfn) {
5388 if (!pfn_valid(pfn)) {
5389 pfn++;
5390 continue;
5392 page = pfn_to_page(pfn);
5393 BUG_ON(page_count(page));
5394 BUG_ON(!PageBuddy(page));
5395 order = page_order(page);
5396 #ifdef CONFIG_DEBUG_VM
5397 printk(KERN_INFO "remove from free list %lx %d %lx\n",
5398 pfn, 1 << order, end_pfn);
5399 #endif
5400 list_del(&page->lru);
5401 rmv_page_order(page);
5402 zone->free_area[order].nr_free--;
5403 __mod_zone_page_state(zone, NR_FREE_PAGES,
5404 - (1UL << order));
5405 for (i = 0; i < (1 << order); i++)
5406 SetPageReserved((page+i));
5407 pfn += (1 << order);
5409 spin_unlock_irqrestore(&zone->lock, flags);
5411 #endif
5413 #ifdef CONFIG_MEMORY_FAILURE
5414 bool is_free_buddy_page(struct page *page)
5416 struct zone *zone = page_zone(page);
5417 unsigned long pfn = page_to_pfn(page);
5418 unsigned long flags;
5419 int order;
5421 spin_lock_irqsave(&zone->lock, flags);
5422 for (order = 0; order < MAX_ORDER; order++) {
5423 struct page *page_head = page - (pfn & ((1 << order) - 1));
5425 if (PageBuddy(page_head) && page_order(page_head) >= order)
5426 break;
5428 spin_unlock_irqrestore(&zone->lock, flags);
5430 return order < MAX_ORDER;
5432 #endif
5434 static struct trace_print_flags pageflag_names[] = {
5435 {1UL << PG_locked, "locked" },
5436 {1UL << PG_error, "error" },
5437 {1UL << PG_referenced, "referenced" },
5438 {1UL << PG_uptodate, "uptodate" },
5439 {1UL << PG_dirty, "dirty" },
5440 {1UL << PG_lru, "lru" },
5441 {1UL << PG_active, "active" },
5442 {1UL << PG_slab, "slab" },
5443 {1UL << PG_owner_priv_1, "owner_priv_1" },
5444 {1UL << PG_arch_1, "arch_1" },
5445 {1UL << PG_reserved, "reserved" },
5446 {1UL << PG_private, "private" },
5447 {1UL << PG_private_2, "private_2" },
5448 {1UL << PG_writeback, "writeback" },
5449 #ifdef CONFIG_PAGEFLAGS_EXTENDED
5450 {1UL << PG_head, "head" },
5451 {1UL << PG_tail, "tail" },
5452 #else
5453 {1UL << PG_compound, "compound" },
5454 #endif
5455 {1UL << PG_swapcache, "swapcache" },
5456 {1UL << PG_mappedtodisk, "mappedtodisk" },
5457 {1UL << PG_reclaim, "reclaim" },
5458 {1UL << PG_buddy, "buddy" },
5459 {1UL << PG_swapbacked, "swapbacked" },
5460 {1UL << PG_unevictable, "unevictable" },
5461 #ifdef CONFIG_MMU
5462 {1UL << PG_mlocked, "mlocked" },
5463 #endif
5464 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
5465 {1UL << PG_uncached, "uncached" },
5466 #endif
5467 #ifdef CONFIG_MEMORY_FAILURE
5468 {1UL << PG_hwpoison, "hwpoison" },
5469 #endif
5470 {-1UL, NULL },
5473 static void dump_page_flags(unsigned long flags)
5475 const char *delim = "";
5476 unsigned long mask;
5477 int i;
5479 printk(KERN_ALERT "page flags: %#lx(", flags);
5481 /* remove zone id */
5482 flags &= (1UL << NR_PAGEFLAGS) - 1;
5484 for (i = 0; pageflag_names[i].name && flags; i++) {
5486 mask = pageflag_names[i].mask;
5487 if ((flags & mask) != mask)
5488 continue;
5490 flags &= ~mask;
5491 printk("%s%s", delim, pageflag_names[i].name);
5492 delim = "|";
5495 /* check for left over flags */
5496 if (flags)
5497 printk("%s%#lx", delim, flags);
5499 printk(")\n");
5502 void dump_page(struct page *page)
5504 printk(KERN_ALERT
5505 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
5506 page, page_count(page), page_mapcount(page),
5507 page->mapping, page->index);
5508 dump_page_flags(page->flags);