[PATCH] x86_64: Remove obsolete eat_key prototype
[linux-2.6.git] / mm / page_alloc.c
1 /*
2  *  linux/mm/page_alloc.c
3  *
4  *  Manages the free list, the system allocates free pages here.
5  *  Note that kmalloc() lives in slab.c
6  *
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)
15  */
16
17 #include <linux/config.h>
18 #include <linux/stddef.h>
19 #include <linux/mm.h>
20 #include <linux/swap.h>
21 #include <linux/interrupt.h>
22 #include <linux/pagemap.h>
23 #include <linux/bootmem.h>
24 #include <linux/compiler.h>
25 #include <linux/module.h>
26 #include <linux/suspend.h>
27 #include <linux/pagevec.h>
28 #include <linux/blkdev.h>
29 #include <linux/slab.h>
30 #include <linux/notifier.h>
31 #include <linux/topology.h>
32 #include <linux/sysctl.h>
33 #include <linux/cpu.h>
34 #include <linux/cpuset.h>
35 #include <linux/nodemask.h>
36 #include <linux/vmalloc.h>
37
38 #include <asm/tlbflush.h>
39 #include "internal.h"
40
41 /*
42  * MCD - HACK: Find somewhere to initialize this EARLY, or make this
43  * initializer cleaner
44  */
45 nodemask_t node_online_map = { { [0] = 1UL } };
46 EXPORT_SYMBOL(node_online_map);
47 nodemask_t node_possible_map = NODE_MASK_ALL;
48 EXPORT_SYMBOL(node_possible_map);
49 struct pglist_data *pgdat_list;
50 unsigned long totalram_pages;
51 unsigned long totalhigh_pages;
52 long nr_swap_pages;
53
54 /*
55  * results with 256, 32 in the lowmem_reserve sysctl:
56  *      1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
57  *      1G machine -> (16M dma, 784M normal, 224M high)
58  *      NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
59  *      HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
60  *      HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
61  */
62 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = { 256, 32 };
63
64 EXPORT_SYMBOL(totalram_pages);
65 EXPORT_SYMBOL(nr_swap_pages);
66
67 /*
68  * Used by page_zone() to look up the address of the struct zone whose
69  * id is encoded in the upper bits of page->flags
70  */
71 struct zone *zone_table[1 << ZONETABLE_SHIFT];
72 EXPORT_SYMBOL(zone_table);
73
74 static char *zone_names[MAX_NR_ZONES] = { "DMA", "Normal", "HighMem" };
75 int min_free_kbytes = 1024;
76
77 unsigned long __initdata nr_kernel_pages;
78 unsigned long __initdata nr_all_pages;
79
80 /*
81  * Temporary debugging check for pages not lying within a given zone.
82  */
83 static int bad_range(struct zone *zone, struct page *page)
84 {
85         if (page_to_pfn(page) >= zone->zone_start_pfn + zone->spanned_pages)
86                 return 1;
87         if (page_to_pfn(page) < zone->zone_start_pfn)
88                 return 1;
89 #ifdef CONFIG_HOLES_IN_ZONE
90         if (!pfn_valid(page_to_pfn(page)))
91                 return 1;
92 #endif
93         if (zone != page_zone(page))
94                 return 1;
95         return 0;
96 }
97
98 static void bad_page(const char *function, struct page *page)
99 {
100         printk(KERN_EMERG "Bad page state at %s (in process '%s', page %p)\n",
101                 function, current->comm, page);
102         printk(KERN_EMERG "flags:0x%0*lx mapping:%p mapcount:%d count:%d\n",
103                 (int)(2*sizeof(page_flags_t)), (unsigned long)page->flags,
104                 page->mapping, page_mapcount(page), page_count(page));
105         printk(KERN_EMERG "Backtrace:\n");
106         dump_stack();
107         printk(KERN_EMERG "Trying to fix it up, but a reboot is needed\n");
108         page->flags &= ~(1 << PG_lru    |
109                         1 << PG_private |
110                         1 << PG_locked  |
111                         1 << PG_active  |
112                         1 << PG_dirty   |
113                         1 << PG_reclaim |
114                         1 << PG_slab    |
115                         1 << PG_swapcache |
116                         1 << PG_writeback);
117         set_page_count(page, 0);
118         reset_page_mapcount(page);
119         page->mapping = NULL;
120         tainted |= TAINT_BAD_PAGE;
121 }
122
123 #ifndef CONFIG_HUGETLB_PAGE
124 #define prep_compound_page(page, order) do { } while (0)
125 #define destroy_compound_page(page, order) do { } while (0)
126 #else
127 /*
128  * Higher-order pages are called "compound pages".  They are structured thusly:
129  *
130  * The first PAGE_SIZE page is called the "head page".
131  *
132  * The remaining PAGE_SIZE pages are called "tail pages".
133  *
134  * All pages have PG_compound set.  All pages have their ->private pointing at
135  * the head page (even the head page has this).
136  *
137  * The first tail page's ->mapping, if non-zero, holds the address of the
138  * compound page's put_page() function.
139  *
140  * The order of the allocation is stored in the first tail page's ->index
141  * This is only for debug at present.  This usage means that zero-order pages
142  * may not be compound.
143  */
144 static void prep_compound_page(struct page *page, unsigned long order)
145 {
146         int i;
147         int nr_pages = 1 << order;
148
149         page[1].mapping = NULL;
150         page[1].index = order;
151         for (i = 0; i < nr_pages; i++) {
152                 struct page *p = page + i;
153
154                 SetPageCompound(p);
155                 p->private = (unsigned long)page;
156         }
157 }
158
159 static void destroy_compound_page(struct page *page, unsigned long order)
160 {
161         int i;
162         int nr_pages = 1 << order;
163
164         if (!PageCompound(page))
165                 return;
166
167         if (page[1].index != order)
168                 bad_page(__FUNCTION__, page);
169
170         for (i = 0; i < nr_pages; i++) {
171                 struct page *p = page + i;
172
173                 if (!PageCompound(p))
174                         bad_page(__FUNCTION__, page);
175                 if (p->private != (unsigned long)page)
176                         bad_page(__FUNCTION__, page);
177                 ClearPageCompound(p);
178         }
179 }
180 #endif          /* CONFIG_HUGETLB_PAGE */
181
182 /*
183  * function for dealing with page's order in buddy system.
184  * zone->lock is already acquired when we use these.
185  * So, we don't need atomic page->flags operations here.
186  */
187 static inline unsigned long page_order(struct page *page) {
188         return page->private;
189 }
190
191 static inline void set_page_order(struct page *page, int order) {
192         page->private = order;
193         __SetPagePrivate(page);
194 }
195
196 static inline void rmv_page_order(struct page *page)
197 {
198         __ClearPagePrivate(page);
199         page->private = 0;
200 }
201
202 /*
203  * Locate the struct page for both the matching buddy in our
204  * pair (buddy1) and the combined O(n+1) page they form (page).
205  *
206  * 1) Any buddy B1 will have an order O twin B2 which satisfies
207  * the following equation:
208  *     B2 = B1 ^ (1 << O)
209  * For example, if the starting buddy (buddy2) is #8 its order
210  * 1 buddy is #10:
211  *     B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
212  *
213  * 2) Any buddy B will have an order O+1 parent P which
214  * satisfies the following equation:
215  *     P = B & ~(1 << O)
216  *
217  * Assumption: *_mem_map is contigious at least up to MAX_ORDER
218  */
219 static inline struct page *
220 __page_find_buddy(struct page *page, unsigned long page_idx, unsigned int order)
221 {
222         unsigned long buddy_idx = page_idx ^ (1 << order);
223
224         return page + (buddy_idx - page_idx);
225 }
226
227 static inline unsigned long
228 __find_combined_index(unsigned long page_idx, unsigned int order)
229 {
230         return (page_idx & ~(1 << order));
231 }
232
233 /*
234  * This function checks whether a page is free && is the buddy
235  * we can do coalesce a page and its buddy if
236  * (a) the buddy is free &&
237  * (b) the buddy is on the buddy system &&
238  * (c) a page and its buddy have the same order.
239  * for recording page's order, we use page->private and PG_private.
240  *
241  */
242 static inline int page_is_buddy(struct page *page, int order)
243 {
244        if (PagePrivate(page)           &&
245            (page_order(page) == order) &&
246            !PageReserved(page)         &&
247             page_count(page) == 0)
248                return 1;
249        return 0;
250 }
251
252 /*
253  * Freeing function for a buddy system allocator.
254  *
255  * The concept of a buddy system is to maintain direct-mapped table
256  * (containing bit values) for memory blocks of various "orders".
257  * The bottom level table contains the map for the smallest allocatable
258  * units of memory (here, pages), and each level above it describes
259  * pairs of units from the levels below, hence, "buddies".
260  * At a high level, all that happens here is marking the table entry
261  * at the bottom level available, and propagating the changes upward
262  * as necessary, plus some accounting needed to play nicely with other
263  * parts of the VM system.
264  * At each level, we keep a list of pages, which are heads of continuous
265  * free pages of length of (1 << order) and marked with PG_Private.Page's
266  * order is recorded in page->private field.
267  * So when we are allocating or freeing one, we can derive the state of the
268  * other.  That is, if we allocate a small block, and both were   
269  * free, the remainder of the region must be split into blocks.   
270  * If a block is freed, and its buddy is also free, then this
271  * triggers coalescing into a block of larger size.            
272  *
273  * -- wli
274  */
275
276 static inline void __free_pages_bulk (struct page *page,
277                 struct zone *zone, unsigned int order)
278 {
279         unsigned long page_idx;
280         int order_size = 1 << order;
281
282         if (unlikely(order))
283                 destroy_compound_page(page, order);
284
285         page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
286
287         BUG_ON(page_idx & (order_size - 1));
288         BUG_ON(bad_range(zone, page));
289
290         zone->free_pages += order_size;
291         while (order < MAX_ORDER-1) {
292                 unsigned long combined_idx;
293                 struct free_area *area;
294                 struct page *buddy;
295
296                 combined_idx = __find_combined_index(page_idx, order);
297                 buddy = __page_find_buddy(page, page_idx, order);
298
299                 if (bad_range(zone, buddy))
300                         break;
301                 if (!page_is_buddy(buddy, order))
302                         break;          /* Move the buddy up one level. */
303                 list_del(&buddy->lru);
304                 area = zone->free_area + order;
305                 area->nr_free--;
306                 rmv_page_order(buddy);
307                 page = page + (combined_idx - page_idx);
308                 page_idx = combined_idx;
309                 order++;
310         }
311         set_page_order(page, order);
312         list_add(&page->lru, &zone->free_area[order].free_list);
313         zone->free_area[order].nr_free++;
314 }
315
316 static inline void free_pages_check(const char *function, struct page *page)
317 {
318         if (    page_mapcount(page) ||
319                 page->mapping != NULL ||
320                 page_count(page) != 0 ||
321                 (page->flags & (
322                         1 << PG_lru     |
323                         1 << PG_private |
324                         1 << PG_locked  |
325                         1 << PG_active  |
326                         1 << PG_reclaim |
327                         1 << PG_slab    |
328                         1 << PG_swapcache |
329                         1 << PG_writeback )))
330                 bad_page(function, page);
331         if (PageDirty(page))
332                 ClearPageDirty(page);
333 }
334
335 /*
336  * Frees a list of pages. 
337  * Assumes all pages on list are in same zone, and of same order.
338  * count is the number of pages to free, or 0 for all on the list.
339  *
340  * If the zone was previously in an "all pages pinned" state then look to
341  * see if this freeing clears that state.
342  *
343  * And clear the zone's pages_scanned counter, to hold off the "all pages are
344  * pinned" detection logic.
345  */
346 static int
347 free_pages_bulk(struct zone *zone, int count,
348                 struct list_head *list, unsigned int order)
349 {
350         unsigned long flags;
351         struct page *page = NULL;
352         int ret = 0;
353
354         spin_lock_irqsave(&zone->lock, flags);
355         zone->all_unreclaimable = 0;
356         zone->pages_scanned = 0;
357         while (!list_empty(list) && count--) {
358                 page = list_entry(list->prev, struct page, lru);
359                 /* have to delete it as __free_pages_bulk list manipulates */
360                 list_del(&page->lru);
361                 __free_pages_bulk(page, zone, order);
362                 ret++;
363         }
364         spin_unlock_irqrestore(&zone->lock, flags);
365         return ret;
366 }
367
368 void __free_pages_ok(struct page *page, unsigned int order)
369 {
370         LIST_HEAD(list);
371         int i;
372
373         arch_free_page(page, order);
374
375         mod_page_state(pgfree, 1 << order);
376
377 #ifndef CONFIG_MMU
378         if (order > 0)
379                 for (i = 1 ; i < (1 << order) ; ++i)
380                         __put_page(page + i);
381 #endif
382
383         for (i = 0 ; i < (1 << order) ; ++i)
384                 free_pages_check(__FUNCTION__, page + i);
385         list_add(&page->lru, &list);
386         kernel_map_pages(page, 1<<order, 0);
387         free_pages_bulk(page_zone(page), 1, &list, order);
388 }
389
390
391 /*
392  * The order of subdivision here is critical for the IO subsystem.
393  * Please do not alter this order without good reasons and regression
394  * testing. Specifically, as large blocks of memory are subdivided,
395  * the order in which smaller blocks are delivered depends on the order
396  * they're subdivided in this function. This is the primary factor
397  * influencing the order in which pages are delivered to the IO
398  * subsystem according to empirical testing, and this is also justified
399  * by considering the behavior of a buddy system containing a single
400  * large block of memory acted on by a series of small allocations.
401  * This behavior is a critical factor in sglist merging's success.
402  *
403  * -- wli
404  */
405 static inline struct page *
406 expand(struct zone *zone, struct page *page,
407         int low, int high, struct free_area *area)
408 {
409         unsigned long size = 1 << high;
410
411         while (high > low) {
412                 area--;
413                 high--;
414                 size >>= 1;
415                 BUG_ON(bad_range(zone, &page[size]));
416                 list_add(&page[size].lru, &area->free_list);
417                 area->nr_free++;
418                 set_page_order(&page[size], high);
419         }
420         return page;
421 }
422
423 void set_page_refs(struct page *page, int order)
424 {
425 #ifdef CONFIG_MMU
426         set_page_count(page, 1);
427 #else
428         int i;
429
430         /*
431          * We need to reference all the pages for this order, otherwise if
432          * anyone accesses one of the pages with (get/put) it will be freed.
433          * - eg: access_process_vm()
434          */
435         for (i = 0; i < (1 << order); i++)
436                 set_page_count(page + i, 1);
437 #endif /* CONFIG_MMU */
438 }
439
440 /*
441  * This page is about to be returned from the page allocator
442  */
443 static void prep_new_page(struct page *page, int order)
444 {
445         if (    page_mapcount(page) ||
446                 page->mapping != NULL ||
447                 page_count(page) != 0 ||
448                 (page->flags & (
449                         1 << PG_lru     |
450                         1 << PG_private |
451                         1 << PG_locked  |
452                         1 << PG_active  |
453                         1 << PG_dirty   |
454                         1 << PG_reclaim |
455                         1 << PG_slab    |
456                         1 << PG_swapcache |
457                         1 << PG_writeback )))
458                 bad_page(__FUNCTION__, page);
459
460         page->flags &= ~(1 << PG_uptodate | 1 << PG_error |
461                         1 << PG_referenced | 1 << PG_arch_1 |
462                         1 << PG_checked | 1 << PG_mappedtodisk);
463         page->private = 0;
464         set_page_refs(page, order);
465         kernel_map_pages(page, 1 << order, 1);
466 }
467
468 /* 
469  * Do the hard work of removing an element from the buddy allocator.
470  * Call me with the zone->lock already held.
471  */
472 static struct page *__rmqueue(struct zone *zone, unsigned int order)
473 {
474         struct free_area * area;
475         unsigned int current_order;
476         struct page *page;
477
478         for (current_order = order; current_order < MAX_ORDER; ++current_order) {
479                 area = zone->free_area + current_order;
480                 if (list_empty(&area->free_list))
481                         continue;
482
483                 page = list_entry(area->free_list.next, struct page, lru);
484                 list_del(&page->lru);
485                 rmv_page_order(page);
486                 area->nr_free--;
487                 zone->free_pages -= 1UL << order;
488                 return expand(zone, page, order, current_order, area);
489         }
490
491         return NULL;
492 }
493
494 /* 
495  * Obtain a specified number of elements from the buddy allocator, all under
496  * a single hold of the lock, for efficiency.  Add them to the supplied list.
497  * Returns the number of new pages which were placed at *list.
498  */
499 static int rmqueue_bulk(struct zone *zone, unsigned int order, 
500                         unsigned long count, struct list_head *list)
501 {
502         unsigned long flags;
503         int i;
504         int allocated = 0;
505         struct page *page;
506         
507         spin_lock_irqsave(&zone->lock, flags);
508         for (i = 0; i < count; ++i) {
509                 page = __rmqueue(zone, order);
510                 if (page == NULL)
511                         break;
512                 allocated++;
513                 list_add_tail(&page->lru, list);
514         }
515         spin_unlock_irqrestore(&zone->lock, flags);
516         return allocated;
517 }
518
519 #ifdef CONFIG_NUMA
520 /* Called from the slab reaper to drain remote pagesets */
521 void drain_remote_pages(void)
522 {
523         struct zone *zone;
524         int i;
525         unsigned long flags;
526
527         local_irq_save(flags);
528         for_each_zone(zone) {
529                 struct per_cpu_pageset *pset;
530
531                 /* Do not drain local pagesets */
532                 if (zone->zone_pgdat->node_id == numa_node_id())
533                         continue;
534
535                 pset = zone->pageset[smp_processor_id()];
536                 for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
537                         struct per_cpu_pages *pcp;
538
539                         pcp = &pset->pcp[i];
540                         if (pcp->count)
541                                 pcp->count -= free_pages_bulk(zone, pcp->count,
542                                                 &pcp->list, 0);
543                 }
544         }
545         local_irq_restore(flags);
546 }
547 #endif
548
549 #if defined(CONFIG_PM) || defined(CONFIG_HOTPLUG_CPU)
550 static void __drain_pages(unsigned int cpu)
551 {
552         struct zone *zone;
553         int i;
554
555         for_each_zone(zone) {
556                 struct per_cpu_pageset *pset;
557
558                 pset = zone_pcp(zone, cpu);
559                 for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
560                         struct per_cpu_pages *pcp;
561
562                         pcp = &pset->pcp[i];
563                         pcp->count -= free_pages_bulk(zone, pcp->count,
564                                                 &pcp->list, 0);
565                 }
566         }
567 }
568 #endif /* CONFIG_PM || CONFIG_HOTPLUG_CPU */
569
570 #ifdef CONFIG_PM
571
572 void mark_free_pages(struct zone *zone)
573 {
574         unsigned long zone_pfn, flags;
575         int order;
576         struct list_head *curr;
577
578         if (!zone->spanned_pages)
579                 return;
580
581         spin_lock_irqsave(&zone->lock, flags);
582         for (zone_pfn = 0; zone_pfn < zone->spanned_pages; ++zone_pfn)
583                 ClearPageNosaveFree(pfn_to_page(zone_pfn + zone->zone_start_pfn));
584
585         for (order = MAX_ORDER - 1; order >= 0; --order)
586                 list_for_each(curr, &zone->free_area[order].free_list) {
587                         unsigned long start_pfn, i;
588
589                         start_pfn = page_to_pfn(list_entry(curr, struct page, lru));
590
591                         for (i=0; i < (1<<order); i++)
592                                 SetPageNosaveFree(pfn_to_page(start_pfn+i));
593         }
594         spin_unlock_irqrestore(&zone->lock, flags);
595 }
596
597 /*
598  * Spill all of this CPU's per-cpu pages back into the buddy allocator.
599  */
600 void drain_local_pages(void)
601 {
602         unsigned long flags;
603
604         local_irq_save(flags);  
605         __drain_pages(smp_processor_id());
606         local_irq_restore(flags);       
607 }
608 #endif /* CONFIG_PM */
609
610 static void zone_statistics(struct zonelist *zonelist, struct zone *z)
611 {
612 #ifdef CONFIG_NUMA
613         unsigned long flags;
614         int cpu;
615         pg_data_t *pg = z->zone_pgdat;
616         pg_data_t *orig = zonelist->zones[0]->zone_pgdat;
617         struct per_cpu_pageset *p;
618
619         local_irq_save(flags);
620         cpu = smp_processor_id();
621         p = zone_pcp(z,cpu);
622         if (pg == orig) {
623                 p->numa_hit++;
624         } else {
625                 p->numa_miss++;
626                 zone_pcp(zonelist->zones[0], cpu)->numa_foreign++;
627         }
628         if (pg == NODE_DATA(numa_node_id()))
629                 p->local_node++;
630         else
631                 p->other_node++;
632         local_irq_restore(flags);
633 #endif
634 }
635
636 /*
637  * Free a 0-order page
638  */
639 static void FASTCALL(free_hot_cold_page(struct page *page, int cold));
640 static void fastcall free_hot_cold_page(struct page *page, int cold)
641 {
642         struct zone *zone = page_zone(page);
643         struct per_cpu_pages *pcp;
644         unsigned long flags;
645
646         arch_free_page(page, 0);
647
648         kernel_map_pages(page, 1, 0);
649         inc_page_state(pgfree);
650         if (PageAnon(page))
651                 page->mapping = NULL;
652         free_pages_check(__FUNCTION__, page);
653         pcp = &zone_pcp(zone, get_cpu())->pcp[cold];
654         local_irq_save(flags);
655         list_add(&page->lru, &pcp->list);
656         pcp->count++;
657         if (pcp->count >= pcp->high)
658                 pcp->count -= free_pages_bulk(zone, pcp->batch, &pcp->list, 0);
659         local_irq_restore(flags);
660         put_cpu();
661 }
662
663 void fastcall free_hot_page(struct page *page)
664 {
665         free_hot_cold_page(page, 0);
666 }
667         
668 void fastcall free_cold_page(struct page *page)
669 {
670         free_hot_cold_page(page, 1);
671 }
672
673 static inline void prep_zero_page(struct page *page, int order, unsigned int __nocast gfp_flags)
674 {
675         int i;
676
677         BUG_ON((gfp_flags & (__GFP_WAIT | __GFP_HIGHMEM)) == __GFP_HIGHMEM);
678         for(i = 0; i < (1 << order); i++)
679                 clear_highpage(page + i);
680 }
681
682 /*
683  * Really, prep_compound_page() should be called from __rmqueue_bulk().  But
684  * we cheat by calling it from here, in the order > 0 path.  Saves a branch
685  * or two.
686  */
687 static struct page *
688 buffered_rmqueue(struct zone *zone, int order, unsigned int __nocast gfp_flags)
689 {
690         unsigned long flags;
691         struct page *page = NULL;
692         int cold = !!(gfp_flags & __GFP_COLD);
693
694         if (order == 0) {
695                 struct per_cpu_pages *pcp;
696
697                 pcp = &zone_pcp(zone, get_cpu())->pcp[cold];
698                 local_irq_save(flags);
699                 if (pcp->count <= pcp->low)
700                         pcp->count += rmqueue_bulk(zone, 0,
701                                                 pcp->batch, &pcp->list);
702                 if (pcp->count) {
703                         page = list_entry(pcp->list.next, struct page, lru);
704                         list_del(&page->lru);
705                         pcp->count--;
706                 }
707                 local_irq_restore(flags);
708                 put_cpu();
709         }
710
711         if (page == NULL) {
712                 spin_lock_irqsave(&zone->lock, flags);
713                 page = __rmqueue(zone, order);
714                 spin_unlock_irqrestore(&zone->lock, flags);
715         }
716
717         if (page != NULL) {
718                 BUG_ON(bad_range(zone, page));
719                 mod_page_state_zone(zone, pgalloc, 1 << order);
720                 prep_new_page(page, order);
721
722                 if (gfp_flags & __GFP_ZERO)
723                         prep_zero_page(page, order, gfp_flags);
724
725                 if (order && (gfp_flags & __GFP_COMP))
726                         prep_compound_page(page, order);
727         }
728         return page;
729 }
730
731 /*
732  * Return 1 if free pages are above 'mark'. This takes into account the order
733  * of the allocation.
734  */
735 int zone_watermark_ok(struct zone *z, int order, unsigned long mark,
736                       int classzone_idx, int can_try_harder, int gfp_high)
737 {
738         /* free_pages my go negative - that's OK */
739         long min = mark, free_pages = z->free_pages - (1 << order) + 1;
740         int o;
741
742         if (gfp_high)
743                 min -= min / 2;
744         if (can_try_harder)
745                 min -= min / 4;
746
747         if (free_pages <= min + z->lowmem_reserve[classzone_idx])
748                 return 0;
749         for (o = 0; o < order; o++) {
750                 /* At the next order, this order's pages become unavailable */
751                 free_pages -= z->free_area[o].nr_free << o;
752
753                 /* Require fewer higher order pages to be free */
754                 min >>= 1;
755
756                 if (free_pages <= min)
757                         return 0;
758         }
759         return 1;
760 }
761
762 static inline int
763 should_reclaim_zone(struct zone *z, unsigned int gfp_mask)
764 {
765         if (!z->reclaim_pages)
766                 return 0;
767         if (gfp_mask & __GFP_NORECLAIM)
768                 return 0;
769         return 1;
770 }
771
772 /*
773  * This is the 'heart' of the zoned buddy allocator.
774  */
775 struct page * fastcall
776 __alloc_pages(unsigned int __nocast gfp_mask, unsigned int order,
777                 struct zonelist *zonelist)
778 {
779         const int wait = gfp_mask & __GFP_WAIT;
780         struct zone **zones, *z;
781         struct page *page;
782         struct reclaim_state reclaim_state;
783         struct task_struct *p = current;
784         int i;
785         int classzone_idx;
786         int do_retry;
787         int can_try_harder;
788         int did_some_progress;
789
790         might_sleep_if(wait);
791
792         /*
793          * The caller may dip into page reserves a bit more if the caller
794          * cannot run direct reclaim, or is the caller has realtime scheduling
795          * policy
796          */
797         can_try_harder = (unlikely(rt_task(p)) && !in_interrupt()) || !wait;
798
799         zones = zonelist->zones;  /* the list of zones suitable for gfp_mask */
800
801         if (unlikely(zones[0] == NULL)) {
802                 /* Should this ever happen?? */
803                 return NULL;
804         }
805
806         classzone_idx = zone_idx(zones[0]);
807
808 restart:
809         /* Go through the zonelist once, looking for a zone with enough free */
810         for (i = 0; (z = zones[i]) != NULL; i++) {
811                 int do_reclaim = should_reclaim_zone(z, gfp_mask);
812
813                 if (!cpuset_zone_allowed(z))
814                         continue;
815
816                 /*
817                  * If the zone is to attempt early page reclaim then this loop
818                  * will try to reclaim pages and check the watermark a second
819                  * time before giving up and falling back to the next zone.
820                  */
821 zone_reclaim_retry:
822                 if (!zone_watermark_ok(z, order, z->pages_low,
823                                        classzone_idx, 0, 0)) {
824                         if (!do_reclaim)
825                                 continue;
826                         else {
827                                 zone_reclaim(z, gfp_mask, order);
828                                 /* Only try reclaim once */
829                                 do_reclaim = 0;
830                                 goto zone_reclaim_retry;
831                         }
832                 }
833
834                 page = buffered_rmqueue(z, order, gfp_mask);
835                 if (page)
836                         goto got_pg;
837         }
838
839         for (i = 0; (z = zones[i]) != NULL; i++)
840                 wakeup_kswapd(z, order);
841
842         /*
843          * Go through the zonelist again. Let __GFP_HIGH and allocations
844          * coming from realtime tasks to go deeper into reserves
845          *
846          * This is the last chance, in general, before the goto nopage.
847          * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
848          */
849         for (i = 0; (z = zones[i]) != NULL; i++) {
850                 if (!zone_watermark_ok(z, order, z->pages_min,
851                                        classzone_idx, can_try_harder,
852                                        gfp_mask & __GFP_HIGH))
853                         continue;
854
855                 if (wait && !cpuset_zone_allowed(z))
856                         continue;
857
858                 page = buffered_rmqueue(z, order, gfp_mask);
859                 if (page)
860                         goto got_pg;
861         }
862
863         /* This allocation should allow future memory freeing. */
864
865         if (((p->flags & PF_MEMALLOC) || unlikely(test_thread_flag(TIF_MEMDIE)))
866                         && !in_interrupt()) {
867                 if (!(gfp_mask & __GFP_NOMEMALLOC)) {
868                         /* go through the zonelist yet again, ignoring mins */
869                         for (i = 0; (z = zones[i]) != NULL; i++) {
870                                 if (!cpuset_zone_allowed(z))
871                                         continue;
872                                 page = buffered_rmqueue(z, order, gfp_mask);
873                                 if (page)
874                                         goto got_pg;
875                         }
876                 }
877                 goto nopage;
878         }
879
880         /* Atomic allocations - we can't balance anything */
881         if (!wait)
882                 goto nopage;
883
884 rebalance:
885         cond_resched();
886
887         /* We now go into synchronous reclaim */
888         p->flags |= PF_MEMALLOC;
889         reclaim_state.reclaimed_slab = 0;
890         p->reclaim_state = &reclaim_state;
891
892         did_some_progress = try_to_free_pages(zones, gfp_mask);
893
894         p->reclaim_state = NULL;
895         p->flags &= ~PF_MEMALLOC;
896
897         cond_resched();
898
899         if (likely(did_some_progress)) {
900                 for (i = 0; (z = zones[i]) != NULL; i++) {
901                         if (!zone_watermark_ok(z, order, z->pages_min,
902                                                classzone_idx, can_try_harder,
903                                                gfp_mask & __GFP_HIGH))
904                                 continue;
905
906                         if (!cpuset_zone_allowed(z))
907                                 continue;
908
909                         page = buffered_rmqueue(z, order, gfp_mask);
910                         if (page)
911                                 goto got_pg;
912                 }
913         } else if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
914                 /*
915                  * Go through the zonelist yet one more time, keep
916                  * very high watermark here, this is only to catch
917                  * a parallel oom killing, we must fail if we're still
918                  * under heavy pressure.
919                  */
920                 for (i = 0; (z = zones[i]) != NULL; i++) {
921                         if (!zone_watermark_ok(z, order, z->pages_high,
922                                                classzone_idx, 0, 0))
923                                 continue;
924
925                         if (!cpuset_zone_allowed(z))
926                                 continue;
927
928                         page = buffered_rmqueue(z, order, gfp_mask);
929                         if (page)
930                                 goto got_pg;
931                 }
932
933                 out_of_memory(gfp_mask, order);
934                 goto restart;
935         }
936
937         /*
938          * Don't let big-order allocations loop unless the caller explicitly
939          * requests that.  Wait for some write requests to complete then retry.
940          *
941          * In this implementation, __GFP_REPEAT means __GFP_NOFAIL for order
942          * <= 3, but that may not be true in other implementations.
943          */
944         do_retry = 0;
945         if (!(gfp_mask & __GFP_NORETRY)) {
946                 if ((order <= 3) || (gfp_mask & __GFP_REPEAT))
947                         do_retry = 1;
948                 if (gfp_mask & __GFP_NOFAIL)
949                         do_retry = 1;
950         }
951         if (do_retry) {
952                 blk_congestion_wait(WRITE, HZ/50);
953                 goto rebalance;
954         }
955
956 nopage:
957         if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
958                 printk(KERN_WARNING "%s: page allocation failure."
959                         " order:%d, mode:0x%x\n",
960                         p->comm, order, gfp_mask);
961                 dump_stack();
962                 show_mem();
963         }
964         return NULL;
965 got_pg:
966         zone_statistics(zonelist, z);
967         return page;
968 }
969
970 EXPORT_SYMBOL(__alloc_pages);
971
972 /*
973  * Common helper functions.
974  */
975 fastcall unsigned long __get_free_pages(unsigned int __nocast gfp_mask, unsigned int order)
976 {
977         struct page * page;
978         page = alloc_pages(gfp_mask, order);
979         if (!page)
980                 return 0;
981         return (unsigned long) page_address(page);
982 }
983
984 EXPORT_SYMBOL(__get_free_pages);
985
986 fastcall unsigned long get_zeroed_page(unsigned int __nocast gfp_mask)
987 {
988         struct page * page;
989
990         /*
991          * get_zeroed_page() returns a 32-bit address, which cannot represent
992          * a highmem page
993          */
994         BUG_ON(gfp_mask & __GFP_HIGHMEM);
995
996         page = alloc_pages(gfp_mask | __GFP_ZERO, 0);
997         if (page)
998                 return (unsigned long) page_address(page);
999         return 0;
1000 }
1001
1002 EXPORT_SYMBOL(get_zeroed_page);
1003
1004 void __pagevec_free(struct pagevec *pvec)
1005 {
1006         int i = pagevec_count(pvec);
1007
1008         while (--i >= 0)
1009                 free_hot_cold_page(pvec->pages[i], pvec->cold);
1010 }
1011
1012 fastcall void __free_pages(struct page *page, unsigned int order)
1013 {
1014         if (!PageReserved(page) && put_page_testzero(page)) {
1015                 if (order == 0)
1016                         free_hot_page(page);
1017                 else
1018                         __free_pages_ok(page, order);
1019         }
1020 }
1021
1022 EXPORT_SYMBOL(__free_pages);
1023
1024 fastcall void free_pages(unsigned long addr, unsigned int order)
1025 {
1026         if (addr != 0) {
1027                 BUG_ON(!virt_addr_valid((void *)addr));
1028                 __free_pages(virt_to_page((void *)addr), order);
1029         }
1030 }
1031
1032 EXPORT_SYMBOL(free_pages);
1033
1034 /*
1035  * Total amount of free (allocatable) RAM:
1036  */
1037 unsigned int nr_free_pages(void)
1038 {
1039         unsigned int sum = 0;
1040         struct zone *zone;
1041
1042         for_each_zone(zone)
1043                 sum += zone->free_pages;
1044
1045         return sum;
1046 }
1047
1048 EXPORT_SYMBOL(nr_free_pages);
1049
1050 #ifdef CONFIG_NUMA
1051 unsigned int nr_free_pages_pgdat(pg_data_t *pgdat)
1052 {
1053         unsigned int i, sum = 0;
1054
1055         for (i = 0; i < MAX_NR_ZONES; i++)
1056                 sum += pgdat->node_zones[i].free_pages;
1057
1058         return sum;
1059 }
1060 #endif
1061
1062 static unsigned int nr_free_zone_pages(int offset)
1063 {
1064         pg_data_t *pgdat;
1065         unsigned int sum = 0;
1066
1067         for_each_pgdat(pgdat) {
1068                 struct zonelist *zonelist = pgdat->node_zonelists + offset;
1069                 struct zone **zonep = zonelist->zones;
1070                 struct zone *zone;
1071
1072                 for (zone = *zonep++; zone; zone = *zonep++) {
1073                         unsigned long size = zone->present_pages;
1074                         unsigned long high = zone->pages_high;
1075                         if (size > high)
1076                                 sum += size - high;
1077                 }
1078         }
1079
1080         return sum;
1081 }
1082
1083 /*
1084  * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
1085  */
1086 unsigned int nr_free_buffer_pages(void)
1087 {
1088         return nr_free_zone_pages(GFP_USER & GFP_ZONEMASK);
1089 }
1090
1091 /*
1092  * Amount of free RAM allocatable within all zones
1093  */
1094 unsigned int nr_free_pagecache_pages(void)
1095 {
1096         return nr_free_zone_pages(GFP_HIGHUSER & GFP_ZONEMASK);
1097 }
1098
1099 #ifdef CONFIG_HIGHMEM
1100 unsigned int nr_free_highpages (void)
1101 {
1102         pg_data_t *pgdat;
1103         unsigned int pages = 0;
1104
1105         for_each_pgdat(pgdat)
1106                 pages += pgdat->node_zones[ZONE_HIGHMEM].free_pages;
1107
1108         return pages;
1109 }
1110 #endif
1111
1112 #ifdef CONFIG_NUMA
1113 static void show_node(struct zone *zone)
1114 {
1115         printk("Node %d ", zone->zone_pgdat->node_id);
1116 }
1117 #else
1118 #define show_node(zone) do { } while (0)
1119 #endif
1120
1121 /*
1122  * Accumulate the page_state information across all CPUs.
1123  * The result is unavoidably approximate - it can change
1124  * during and after execution of this function.
1125  */
1126 static DEFINE_PER_CPU(struct page_state, page_states) = {0};
1127
1128 atomic_t nr_pagecache = ATOMIC_INIT(0);
1129 EXPORT_SYMBOL(nr_pagecache);
1130 #ifdef CONFIG_SMP
1131 DEFINE_PER_CPU(long, nr_pagecache_local) = 0;
1132 #endif
1133
1134 void __get_page_state(struct page_state *ret, int nr)
1135 {
1136         int cpu = 0;
1137
1138         memset(ret, 0, sizeof(*ret));
1139
1140         cpu = first_cpu(cpu_online_map);
1141         while (cpu < NR_CPUS) {
1142                 unsigned long *in, *out, off;
1143
1144                 in = (unsigned long *)&per_cpu(page_states, cpu);
1145
1146                 cpu = next_cpu(cpu, cpu_online_map);
1147
1148                 if (cpu < NR_CPUS)
1149                         prefetch(&per_cpu(page_states, cpu));
1150
1151                 out = (unsigned long *)ret;
1152                 for (off = 0; off < nr; off++)
1153                         *out++ += *in++;
1154         }
1155 }
1156
1157 void get_page_state(struct page_state *ret)
1158 {
1159         int nr;
1160
1161         nr = offsetof(struct page_state, GET_PAGE_STATE_LAST);
1162         nr /= sizeof(unsigned long);
1163
1164         __get_page_state(ret, nr + 1);
1165 }
1166
1167 void get_full_page_state(struct page_state *ret)
1168 {
1169         __get_page_state(ret, sizeof(*ret) / sizeof(unsigned long));
1170 }
1171
1172 unsigned long __read_page_state(unsigned long offset)
1173 {
1174         unsigned long ret = 0;
1175         int cpu;
1176
1177         for_each_online_cpu(cpu) {
1178                 unsigned long in;
1179
1180                 in = (unsigned long)&per_cpu(page_states, cpu) + offset;
1181                 ret += *((unsigned long *)in);
1182         }
1183         return ret;
1184 }
1185
1186 void __mod_page_state(unsigned long offset, unsigned long delta)
1187 {
1188         unsigned long flags;
1189         void* ptr;
1190
1191         local_irq_save(flags);
1192         ptr = &__get_cpu_var(page_states);
1193         *(unsigned long*)(ptr + offset) += delta;
1194         local_irq_restore(flags);
1195 }
1196
1197 EXPORT_SYMBOL(__mod_page_state);
1198
1199 void __get_zone_counts(unsigned long *active, unsigned long *inactive,
1200                         unsigned long *free, struct pglist_data *pgdat)
1201 {
1202         struct zone *zones = pgdat->node_zones;
1203         int i;
1204
1205         *active = 0;
1206         *inactive = 0;
1207         *free = 0;
1208         for (i = 0; i < MAX_NR_ZONES; i++) {
1209                 *active += zones[i].nr_active;
1210                 *inactive += zones[i].nr_inactive;
1211                 *free += zones[i].free_pages;
1212         }
1213 }
1214
1215 void get_zone_counts(unsigned long *active,
1216                 unsigned long *inactive, unsigned long *free)
1217 {
1218         struct pglist_data *pgdat;
1219
1220         *active = 0;
1221         *inactive = 0;
1222         *free = 0;
1223         for_each_pgdat(pgdat) {
1224                 unsigned long l, m, n;
1225                 __get_zone_counts(&l, &m, &n, pgdat);
1226                 *active += l;
1227                 *inactive += m;
1228                 *free += n;
1229         }
1230 }
1231
1232 void si_meminfo(struct sysinfo *val)
1233 {
1234         val->totalram = totalram_pages;
1235         val->sharedram = 0;
1236         val->freeram = nr_free_pages();
1237         val->bufferram = nr_blockdev_pages();
1238 #ifdef CONFIG_HIGHMEM
1239         val->totalhigh = totalhigh_pages;
1240         val->freehigh = nr_free_highpages();
1241 #else
1242         val->totalhigh = 0;
1243         val->freehigh = 0;
1244 #endif
1245         val->mem_unit = PAGE_SIZE;
1246 }
1247
1248 EXPORT_SYMBOL(si_meminfo);
1249
1250 #ifdef CONFIG_NUMA
1251 void si_meminfo_node(struct sysinfo *val, int nid)
1252 {
1253         pg_data_t *pgdat = NODE_DATA(nid);
1254
1255         val->totalram = pgdat->node_present_pages;
1256         val->freeram = nr_free_pages_pgdat(pgdat);
1257         val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
1258         val->freehigh = pgdat->node_zones[ZONE_HIGHMEM].free_pages;
1259         val->mem_unit = PAGE_SIZE;
1260 }
1261 #endif
1262
1263 #define K(x) ((x) << (PAGE_SHIFT-10))
1264
1265 /*
1266  * Show free area list (used inside shift_scroll-lock stuff)
1267  * We also calculate the percentage fragmentation. We do this by counting the
1268  * memory on each free list with the exception of the first item on the list.
1269  */
1270 void show_free_areas(void)
1271 {
1272         struct page_state ps;
1273         int cpu, temperature;
1274         unsigned long active;
1275         unsigned long inactive;
1276         unsigned long free;
1277         struct zone *zone;
1278
1279         for_each_zone(zone) {
1280                 show_node(zone);
1281                 printk("%s per-cpu:", zone->name);
1282
1283                 if (!zone->present_pages) {
1284                         printk(" empty\n");
1285                         continue;
1286                 } else
1287                         printk("\n");
1288
1289                 for (cpu = 0; cpu < NR_CPUS; ++cpu) {
1290                         struct per_cpu_pageset *pageset;
1291
1292                         if (!cpu_possible(cpu))
1293                                 continue;
1294
1295                         pageset = zone_pcp(zone, cpu);
1296
1297                         for (temperature = 0; temperature < 2; temperature++)
1298                                 printk("cpu %d %s: low %d, high %d, batch %d used:%d\n",
1299                                         cpu,
1300                                         temperature ? "cold" : "hot",
1301                                         pageset->pcp[temperature].low,
1302                                         pageset->pcp[temperature].high,
1303                                         pageset->pcp[temperature].batch,
1304                                         pageset->pcp[temperature].count);
1305                 }
1306         }
1307
1308         get_page_state(&ps);
1309         get_zone_counts(&active, &inactive, &free);
1310
1311         printk("Free pages: %11ukB (%ukB HighMem)\n",
1312                 K(nr_free_pages()),
1313                 K(nr_free_highpages()));
1314
1315         printk("Active:%lu inactive:%lu dirty:%lu writeback:%lu "
1316                 "unstable:%lu free:%u slab:%lu mapped:%lu pagetables:%lu\n",
1317                 active,
1318                 inactive,
1319                 ps.nr_dirty,
1320                 ps.nr_writeback,
1321                 ps.nr_unstable,
1322                 nr_free_pages(),
1323                 ps.nr_slab,
1324                 ps.nr_mapped,
1325                 ps.nr_page_table_pages);
1326
1327         for_each_zone(zone) {
1328                 int i;
1329
1330                 show_node(zone);
1331                 printk("%s"
1332                         " free:%lukB"
1333                         " min:%lukB"
1334                         " low:%lukB"
1335                         " high:%lukB"
1336                         " active:%lukB"
1337                         " inactive:%lukB"
1338                         " present:%lukB"
1339                         " pages_scanned:%lu"
1340                         " all_unreclaimable? %s"
1341                         "\n",
1342                         zone->name,
1343                         K(zone->free_pages),
1344                         K(zone->pages_min),
1345                         K(zone->pages_low),
1346                         K(zone->pages_high),
1347                         K(zone->nr_active),
1348                         K(zone->nr_inactive),
1349                         K(zone->present_pages),
1350                         zone->pages_scanned,
1351                         (zone->all_unreclaimable ? "yes" : "no")
1352                         );
1353                 printk("lowmem_reserve[]:");
1354                 for (i = 0; i < MAX_NR_ZONES; i++)
1355                         printk(" %lu", zone->lowmem_reserve[i]);
1356                 printk("\n");
1357         }
1358
1359         for_each_zone(zone) {
1360                 unsigned long nr, flags, order, total = 0;
1361
1362                 show_node(zone);
1363                 printk("%s: ", zone->name);
1364                 if (!zone->present_pages) {
1365                         printk("empty\n");
1366                         continue;
1367                 }
1368
1369                 spin_lock_irqsave(&zone->lock, flags);
1370                 for (order = 0; order < MAX_ORDER; order++) {
1371                         nr = zone->free_area[order].nr_free;
1372                         total += nr << order;
1373                         printk("%lu*%lukB ", nr, K(1UL) << order);
1374                 }
1375                 spin_unlock_irqrestore(&zone->lock, flags);
1376                 printk("= %lukB\n", K(total));
1377         }
1378
1379         show_swap_cache_info();
1380 }
1381
1382 /*
1383  * Builds allocation fallback zone lists.
1384  */
1385 static int __init build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist, int j, int k)
1386 {
1387         switch (k) {
1388                 struct zone *zone;
1389         default:
1390                 BUG();
1391         case ZONE_HIGHMEM:
1392                 zone = pgdat->node_zones + ZONE_HIGHMEM;
1393                 if (zone->present_pages) {
1394 #ifndef CONFIG_HIGHMEM
1395                         BUG();
1396 #endif
1397                         zonelist->zones[j++] = zone;
1398                 }
1399         case ZONE_NORMAL:
1400                 zone = pgdat->node_zones + ZONE_NORMAL;
1401                 if (zone->present_pages)
1402                         zonelist->zones[j++] = zone;
1403         case ZONE_DMA:
1404                 zone = pgdat->node_zones + ZONE_DMA;
1405                 if (zone->present_pages)
1406                         zonelist->zones[j++] = zone;
1407         }
1408
1409         return j;
1410 }
1411
1412 #ifdef CONFIG_NUMA
1413 #define MAX_NODE_LOAD (num_online_nodes())
1414 static int __initdata node_load[MAX_NUMNODES];
1415 /**
1416  * find_next_best_node - find the next node that should appear in a given node's fallback list
1417  * @node: node whose fallback list we're appending
1418  * @used_node_mask: nodemask_t of already used nodes
1419  *
1420  * We use a number of factors to determine which is the next node that should
1421  * appear on a given node's fallback list.  The node should not have appeared
1422  * already in @node's fallback list, and it should be the next closest node
1423  * according to the distance array (which contains arbitrary distance values
1424  * from each node to each node in the system), and should also prefer nodes
1425  * with no CPUs, since presumably they'll have very little allocation pressure
1426  * on them otherwise.
1427  * It returns -1 if no node is found.
1428  */
1429 static int __init find_next_best_node(int node, nodemask_t *used_node_mask)
1430 {
1431         int i, n, val;
1432         int min_val = INT_MAX;
1433         int best_node = -1;
1434
1435         for_each_online_node(i) {
1436                 cpumask_t tmp;
1437
1438                 /* Start from local node */
1439                 n = (node+i) % num_online_nodes();
1440
1441                 /* Don't want a node to appear more than once */
1442                 if (node_isset(n, *used_node_mask))
1443                         continue;
1444
1445                 /* Use the local node if we haven't already */
1446                 if (!node_isset(node, *used_node_mask)) {
1447                         best_node = node;
1448                         break;
1449                 }
1450
1451                 /* Use the distance array to find the distance */
1452                 val = node_distance(node, n);
1453
1454                 /* Give preference to headless and unused nodes */
1455                 tmp = node_to_cpumask(n);
1456                 if (!cpus_empty(tmp))
1457                         val += PENALTY_FOR_NODE_WITH_CPUS;
1458
1459                 /* Slight preference for less loaded node */
1460                 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
1461                 val += node_load[n];
1462
1463                 if (val < min_val) {
1464                         min_val = val;
1465                         best_node = n;
1466                 }
1467         }
1468
1469         if (best_node >= 0)
1470                 node_set(best_node, *used_node_mask);
1471
1472         return best_node;
1473 }
1474
1475 static void __init build_zonelists(pg_data_t *pgdat)
1476 {
1477         int i, j, k, node, local_node;
1478         int prev_node, load;
1479         struct zonelist *zonelist;
1480         nodemask_t used_mask;
1481
1482         /* initialize zonelists */
1483         for (i = 0; i < GFP_ZONETYPES; i++) {
1484                 zonelist = pgdat->node_zonelists + i;
1485                 zonelist->zones[0] = NULL;
1486         }
1487
1488         /* NUMA-aware ordering of nodes */
1489         local_node = pgdat->node_id;
1490         load = num_online_nodes();
1491         prev_node = local_node;
1492         nodes_clear(used_mask);
1493         while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
1494                 /*
1495                  * We don't want to pressure a particular node.
1496                  * So adding penalty to the first node in same
1497                  * distance group to make it round-robin.
1498                  */
1499                 if (node_distance(local_node, node) !=
1500                                 node_distance(local_node, prev_node))
1501                         node_load[node] += load;
1502                 prev_node = node;
1503                 load--;
1504                 for (i = 0; i < GFP_ZONETYPES; i++) {
1505                         zonelist = pgdat->node_zonelists + i;
1506                         for (j = 0; zonelist->zones[j] != NULL; j++);
1507
1508                         k = ZONE_NORMAL;
1509                         if (i & __GFP_HIGHMEM)
1510                                 k = ZONE_HIGHMEM;
1511                         if (i & __GFP_DMA)
1512                                 k = ZONE_DMA;
1513
1514                         j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1515                         zonelist->zones[j] = NULL;
1516                 }
1517         }
1518 }
1519
1520 #else   /* CONFIG_NUMA */
1521
1522 static void __init build_zonelists(pg_data_t *pgdat)
1523 {
1524         int i, j, k, node, local_node;
1525
1526         local_node = pgdat->node_id;
1527         for (i = 0; i < GFP_ZONETYPES; i++) {
1528                 struct zonelist *zonelist;
1529
1530                 zonelist = pgdat->node_zonelists + i;
1531
1532                 j = 0;
1533                 k = ZONE_NORMAL;
1534                 if (i & __GFP_HIGHMEM)
1535                         k = ZONE_HIGHMEM;
1536                 if (i & __GFP_DMA)
1537                         k = ZONE_DMA;
1538
1539                 j = build_zonelists_node(pgdat, zonelist, j, k);
1540                 /*
1541                  * Now we build the zonelist so that it contains the zones
1542                  * of all the other nodes.
1543                  * We don't want to pressure a particular node, so when
1544                  * building the zones for node N, we make sure that the
1545                  * zones coming right after the local ones are those from
1546                  * node N+1 (modulo N)
1547                  */
1548                 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
1549                         if (!node_online(node))
1550                                 continue;
1551                         j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1552                 }
1553                 for (node = 0; node < local_node; node++) {
1554                         if (!node_online(node))
1555                                 continue;
1556                         j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1557                 }
1558
1559                 zonelist->zones[j] = NULL;
1560         }
1561 }
1562
1563 #endif  /* CONFIG_NUMA */
1564
1565 void __init build_all_zonelists(void)
1566 {
1567         int i;
1568
1569         for_each_online_node(i)
1570                 build_zonelists(NODE_DATA(i));
1571         printk("Built %i zonelists\n", num_online_nodes());
1572         cpuset_init_current_mems_allowed();
1573 }
1574
1575 /*
1576  * Helper functions to size the waitqueue hash table.
1577  * Essentially these want to choose hash table sizes sufficiently
1578  * large so that collisions trying to wait on pages are rare.
1579  * But in fact, the number of active page waitqueues on typical
1580  * systems is ridiculously low, less than 200. So this is even
1581  * conservative, even though it seems large.
1582  *
1583  * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
1584  * waitqueues, i.e. the size of the waitq table given the number of pages.
1585  */
1586 #define PAGES_PER_WAITQUEUE     256
1587
1588 static inline unsigned long wait_table_size(unsigned long pages)
1589 {
1590         unsigned long size = 1;
1591
1592         pages /= PAGES_PER_WAITQUEUE;
1593
1594         while (size < pages)
1595                 size <<= 1;
1596
1597         /*
1598          * Once we have dozens or even hundreds of threads sleeping
1599          * on IO we've got bigger problems than wait queue collision.
1600          * Limit the size of the wait table to a reasonable size.
1601          */
1602         size = min(size, 4096UL);
1603
1604         return max(size, 4UL);
1605 }
1606
1607 /*
1608  * This is an integer logarithm so that shifts can be used later
1609  * to extract the more random high bits from the multiplicative
1610  * hash function before the remainder is taken.
1611  */
1612 static inline unsigned long wait_table_bits(unsigned long size)
1613 {
1614         return ffz(~size);
1615 }
1616
1617 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
1618
1619 static void __init calculate_zone_totalpages(struct pglist_data *pgdat,
1620                 unsigned long *zones_size, unsigned long *zholes_size)
1621 {
1622         unsigned long realtotalpages, totalpages = 0;
1623         int i;
1624
1625         for (i = 0; i < MAX_NR_ZONES; i++)
1626                 totalpages += zones_size[i];
1627         pgdat->node_spanned_pages = totalpages;
1628
1629         realtotalpages = totalpages;
1630         if (zholes_size)
1631                 for (i = 0; i < MAX_NR_ZONES; i++)
1632                         realtotalpages -= zholes_size[i];
1633         pgdat->node_present_pages = realtotalpages;
1634         printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id, realtotalpages);
1635 }
1636
1637
1638 /*
1639  * Initially all pages are reserved - free ones are freed
1640  * up by free_all_bootmem() once the early boot process is
1641  * done. Non-atomic initialization, single-pass.
1642  */
1643 void __init memmap_init_zone(unsigned long size, int nid, unsigned long zone,
1644                 unsigned long start_pfn)
1645 {
1646         struct page *page;
1647         unsigned long end_pfn = start_pfn + size;
1648         unsigned long pfn;
1649
1650         for (pfn = start_pfn; pfn < end_pfn; pfn++, page++) {
1651                 if (!early_pfn_valid(pfn))
1652                         continue;
1653                 if (!early_pfn_in_nid(pfn, nid))
1654                         continue;
1655                 page = pfn_to_page(pfn);
1656                 set_page_links(page, zone, nid, pfn);
1657                 set_page_count(page, 0);
1658                 reset_page_mapcount(page);
1659                 SetPageReserved(page);
1660                 INIT_LIST_HEAD(&page->lru);
1661 #ifdef WANT_PAGE_VIRTUAL
1662                 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1663                 if (!is_highmem_idx(zone))
1664                         set_page_address(page, __va(pfn << PAGE_SHIFT));
1665 #endif
1666         }
1667 }
1668
1669 void zone_init_free_lists(struct pglist_data *pgdat, struct zone *zone,
1670                                 unsigned long size)
1671 {
1672         int order;
1673         for (order = 0; order < MAX_ORDER ; order++) {
1674                 INIT_LIST_HEAD(&zone->free_area[order].free_list);
1675                 zone->free_area[order].nr_free = 0;
1676         }
1677 }
1678
1679 #define ZONETABLE_INDEX(x, zone_nr)     ((x << ZONES_SHIFT) | zone_nr)
1680 void zonetable_add(struct zone *zone, int nid, int zid, unsigned long pfn,
1681                 unsigned long size)
1682 {
1683         unsigned long snum = pfn_to_section_nr(pfn);
1684         unsigned long end = pfn_to_section_nr(pfn + size);
1685
1686         if (FLAGS_HAS_NODE)
1687                 zone_table[ZONETABLE_INDEX(nid, zid)] = zone;
1688         else
1689                 for (; snum <= end; snum++)
1690                         zone_table[ZONETABLE_INDEX(snum, zid)] = zone;
1691 }
1692
1693 #ifndef __HAVE_ARCH_MEMMAP_INIT
1694 #define memmap_init(size, nid, zone, start_pfn) \
1695         memmap_init_zone((size), (nid), (zone), (start_pfn))
1696 #endif
1697
1698 static int __devinit zone_batchsize(struct zone *zone)
1699 {
1700         int batch;
1701
1702         /*
1703          * The per-cpu-pages pools are set to around 1000th of the
1704          * size of the zone.  But no more than 1/4 of a meg - there's
1705          * no point in going beyond the size of L2 cache.
1706          *
1707          * OK, so we don't know how big the cache is.  So guess.
1708          */
1709         batch = zone->present_pages / 1024;
1710         if (batch * PAGE_SIZE > 256 * 1024)
1711                 batch = (256 * 1024) / PAGE_SIZE;
1712         batch /= 4;             /* We effectively *= 4 below */
1713         if (batch < 1)
1714                 batch = 1;
1715
1716         /*
1717          * Clamp the batch to a 2^n - 1 value. Having a power
1718          * of 2 value was found to be more likely to have
1719          * suboptimal cache aliasing properties in some cases.
1720          *
1721          * For example if 2 tasks are alternately allocating
1722          * batches of pages, one task can end up with a lot
1723          * of pages of one half of the possible page colors
1724          * and the other with pages of the other colors.
1725          */
1726         batch = (1 << fls(batch + batch/2)) - 1;
1727         return batch;
1728 }
1729
1730 inline void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
1731 {
1732         struct per_cpu_pages *pcp;
1733
1734         pcp = &p->pcp[0];               /* hot */
1735         pcp->count = 0;
1736         pcp->low = 2 * batch;
1737         pcp->high = 6 * batch;
1738         pcp->batch = max(1UL, 1 * batch);
1739         INIT_LIST_HEAD(&pcp->list);
1740
1741         pcp = &p->pcp[1];               /* cold*/
1742         pcp->count = 0;
1743         pcp->low = 0;
1744         pcp->high = 2 * batch;
1745         pcp->batch = max(1UL, 1 * batch);
1746         INIT_LIST_HEAD(&pcp->list);
1747 }
1748
1749 #ifdef CONFIG_NUMA
1750 /*
1751  * Boot pageset table. One per cpu which is going to be used for all
1752  * zones and all nodes. The parameters will be set in such a way
1753  * that an item put on a list will immediately be handed over to
1754  * the buddy list. This is safe since pageset manipulation is done
1755  * with interrupts disabled.
1756  *
1757  * Some NUMA counter updates may also be caught by the boot pagesets.
1758  *
1759  * The boot_pagesets must be kept even after bootup is complete for
1760  * unused processors and/or zones. They do play a role for bootstrapping
1761  * hotplugged processors.
1762  *
1763  * zoneinfo_show() and maybe other functions do
1764  * not check if the processor is online before following the pageset pointer.
1765  * Other parts of the kernel may not check if the zone is available.
1766  */
1767 static struct per_cpu_pageset
1768         boot_pageset[NR_CPUS];
1769
1770 /*
1771  * Dynamically allocate memory for the
1772  * per cpu pageset array in struct zone.
1773  */
1774 static int __devinit process_zones(int cpu)
1775 {
1776         struct zone *zone, *dzone;
1777
1778         for_each_zone(zone) {
1779
1780                 zone->pageset[cpu] = kmalloc_node(sizeof(struct per_cpu_pageset),
1781                                          GFP_KERNEL, cpu_to_node(cpu));
1782                 if (!zone->pageset[cpu])
1783                         goto bad;
1784
1785                 setup_pageset(zone->pageset[cpu], zone_batchsize(zone));
1786         }
1787
1788         return 0;
1789 bad:
1790         for_each_zone(dzone) {
1791                 if (dzone == zone)
1792                         break;
1793                 kfree(dzone->pageset[cpu]);
1794                 dzone->pageset[cpu] = NULL;
1795         }
1796         return -ENOMEM;
1797 }
1798
1799 static inline void free_zone_pagesets(int cpu)
1800 {
1801 #ifdef CONFIG_NUMA
1802         struct zone *zone;
1803
1804         for_each_zone(zone) {
1805                 struct per_cpu_pageset *pset = zone_pcp(zone, cpu);
1806
1807                 zone_pcp(zone, cpu) = NULL;
1808                 kfree(pset);
1809         }
1810 #endif
1811 }
1812
1813 static int __devinit pageset_cpuup_callback(struct notifier_block *nfb,
1814                 unsigned long action,
1815                 void *hcpu)
1816 {
1817         int cpu = (long)hcpu;
1818         int ret = NOTIFY_OK;
1819
1820         switch (action) {
1821                 case CPU_UP_PREPARE:
1822                         if (process_zones(cpu))
1823                                 ret = NOTIFY_BAD;
1824                         break;
1825 #ifdef CONFIG_HOTPLUG_CPU
1826                 case CPU_DEAD:
1827                         free_zone_pagesets(cpu);
1828                         break;
1829 #endif
1830                 default:
1831                         break;
1832         }
1833         return ret;
1834 }
1835
1836 static struct notifier_block pageset_notifier =
1837         { &pageset_cpuup_callback, NULL, 0 };
1838
1839 void __init setup_per_cpu_pageset()
1840 {
1841         int err;
1842
1843         /* Initialize per_cpu_pageset for cpu 0.
1844          * A cpuup callback will do this for every cpu
1845          * as it comes online
1846          */
1847         err = process_zones(smp_processor_id());
1848         BUG_ON(err);
1849         register_cpu_notifier(&pageset_notifier);
1850 }
1851
1852 #endif
1853
1854 /*
1855  * Set up the zone data structures:
1856  *   - mark all pages reserved
1857  *   - mark all memory queues empty
1858  *   - clear the memory bitmaps
1859  */
1860 static void __init free_area_init_core(struct pglist_data *pgdat,
1861                 unsigned long *zones_size, unsigned long *zholes_size)
1862 {
1863         unsigned long i, j;
1864         int cpu, nid = pgdat->node_id;
1865         unsigned long zone_start_pfn = pgdat->node_start_pfn;
1866
1867         pgdat->nr_zones = 0;
1868         init_waitqueue_head(&pgdat->kswapd_wait);
1869         pgdat->kswapd_max_order = 0;
1870         
1871         for (j = 0; j < MAX_NR_ZONES; j++) {
1872                 struct zone *zone = pgdat->node_zones + j;
1873                 unsigned long size, realsize;
1874                 unsigned long batch;
1875
1876                 realsize = size = zones_size[j];
1877                 if (zholes_size)
1878                         realsize -= zholes_size[j];
1879
1880                 if (j == ZONE_DMA || j == ZONE_NORMAL)
1881                         nr_kernel_pages += realsize;
1882                 nr_all_pages += realsize;
1883
1884                 zone->spanned_pages = size;
1885                 zone->present_pages = realsize;
1886                 zone->name = zone_names[j];
1887                 spin_lock_init(&zone->lock);
1888                 spin_lock_init(&zone->lru_lock);
1889                 zone->zone_pgdat = pgdat;
1890                 zone->free_pages = 0;
1891
1892                 zone->temp_priority = zone->prev_priority = DEF_PRIORITY;
1893
1894                 batch = zone_batchsize(zone);
1895
1896                 for (cpu = 0; cpu < NR_CPUS; cpu++) {
1897 #ifdef CONFIG_NUMA
1898                         /* Early boot. Slab allocator not functional yet */
1899                         zone->pageset[cpu] = &boot_pageset[cpu];
1900                         setup_pageset(&boot_pageset[cpu],0);
1901 #else
1902                         setup_pageset(zone_pcp(zone,cpu), batch);
1903 #endif
1904                 }
1905                 printk(KERN_DEBUG "  %s zone: %lu pages, LIFO batch:%lu\n",
1906                                 zone_names[j], realsize, batch);
1907                 INIT_LIST_HEAD(&zone->active_list);
1908                 INIT_LIST_HEAD(&zone->inactive_list);
1909                 zone->nr_scan_active = 0;
1910                 zone->nr_scan_inactive = 0;
1911                 zone->nr_active = 0;
1912                 zone->nr_inactive = 0;
1913                 atomic_set(&zone->reclaim_in_progress, -1);
1914                 if (!size)
1915                         continue;
1916
1917                 /*
1918                  * The per-page waitqueue mechanism uses hashed waitqueues
1919                  * per zone.
1920                  */
1921                 zone->wait_table_size = wait_table_size(size);
1922                 zone->wait_table_bits =
1923                         wait_table_bits(zone->wait_table_size);
1924                 zone->wait_table = (wait_queue_head_t *)
1925                         alloc_bootmem_node(pgdat, zone->wait_table_size
1926                                                 * sizeof(wait_queue_head_t));
1927
1928                 for(i = 0; i < zone->wait_table_size; ++i)
1929                         init_waitqueue_head(zone->wait_table + i);
1930
1931                 pgdat->nr_zones = j+1;
1932
1933                 zone->zone_mem_map = pfn_to_page(zone_start_pfn);
1934                 zone->zone_start_pfn = zone_start_pfn;
1935
1936                 memmap_init(size, nid, j, zone_start_pfn);
1937
1938                 zonetable_add(zone, nid, j, zone_start_pfn, size);
1939
1940                 zone_start_pfn += size;
1941
1942                 zone_init_free_lists(pgdat, zone, zone->spanned_pages);
1943         }
1944 }
1945
1946 static void __init alloc_node_mem_map(struct pglist_data *pgdat)
1947 {
1948         /* Skip empty nodes */
1949         if (!pgdat->node_spanned_pages)
1950                 return;
1951
1952 #ifdef CONFIG_FLAT_NODE_MEM_MAP
1953         /* ia64 gets its own node_mem_map, before this, without bootmem */
1954         if (!pgdat->node_mem_map) {
1955                 unsigned long size;
1956                 struct page *map;
1957
1958                 size = (pgdat->node_spanned_pages + 1) * sizeof(struct page);
1959                 map = alloc_remap(pgdat->node_id, size);
1960                 if (!map)
1961                         map = alloc_bootmem_node(pgdat, size);
1962                 pgdat->node_mem_map = map;
1963         }
1964 #ifdef CONFIG_FLATMEM
1965         /*
1966          * With no DISCONTIG, the global mem_map is just set as node 0's
1967          */
1968         if (pgdat == NODE_DATA(0))
1969                 mem_map = NODE_DATA(0)->node_mem_map;
1970 #endif
1971 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
1972 }
1973
1974 void __init free_area_init_node(int nid, struct pglist_data *pgdat,
1975                 unsigned long *zones_size, unsigned long node_start_pfn,
1976                 unsigned long *zholes_size)
1977 {
1978         pgdat->node_id = nid;
1979         pgdat->node_start_pfn = node_start_pfn;
1980         calculate_zone_totalpages(pgdat, zones_size, zholes_size);
1981
1982         alloc_node_mem_map(pgdat);
1983
1984         free_area_init_core(pgdat, zones_size, zholes_size);
1985 }
1986
1987 #ifndef CONFIG_NEED_MULTIPLE_NODES
1988 static bootmem_data_t contig_bootmem_data;
1989 struct pglist_data contig_page_data = { .bdata = &contig_bootmem_data };
1990
1991 EXPORT_SYMBOL(contig_page_data);
1992 #endif
1993
1994 void __init free_area_init(unsigned long *zones_size)
1995 {
1996         free_area_init_node(0, NODE_DATA(0), zones_size,
1997                         __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
1998 }
1999
2000 #ifdef CONFIG_PROC_FS
2001
2002 #include <linux/seq_file.h>
2003
2004 static void *frag_start(struct seq_file *m, loff_t *pos)
2005 {
2006         pg_data_t *pgdat;
2007         loff_t node = *pos;
2008
2009         for (pgdat = pgdat_list; pgdat && node; pgdat = pgdat->pgdat_next)
2010                 --node;
2011
2012         return pgdat;
2013 }
2014
2015 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
2016 {
2017         pg_data_t *pgdat = (pg_data_t *)arg;
2018
2019         (*pos)++;
2020         return pgdat->pgdat_next;
2021 }
2022
2023 static void frag_stop(struct seq_file *m, void *arg)
2024 {
2025 }
2026
2027 /* 
2028  * This walks the free areas for each zone.
2029  */
2030 static int frag_show(struct seq_file *m, void *arg)
2031 {
2032         pg_data_t *pgdat = (pg_data_t *)arg;
2033         struct zone *zone;
2034         struct zone *node_zones = pgdat->node_zones;
2035         unsigned long flags;
2036         int order;
2037
2038         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
2039                 if (!zone->present_pages)
2040                         continue;
2041
2042                 spin_lock_irqsave(&zone->lock, flags);
2043                 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
2044                 for (order = 0; order < MAX_ORDER; ++order)
2045                         seq_printf(m, "%6lu ", zone->free_area[order].nr_free);
2046                 spin_unlock_irqrestore(&zone->lock, flags);
2047                 seq_putc(m, '\n');
2048         }
2049         return 0;
2050 }
2051
2052 struct seq_operations fragmentation_op = {
2053         .start  = frag_start,
2054         .next   = frag_next,
2055         .stop   = frag_stop,
2056         .show   = frag_show,
2057 };
2058
2059 /*
2060  * Output information about zones in @pgdat.
2061  */
2062 static int zoneinfo_show(struct seq_file *m, void *arg)
2063 {
2064         pg_data_t *pgdat = arg;
2065         struct zone *zone;
2066         struct zone *node_zones = pgdat->node_zones;
2067         unsigned long flags;
2068
2069         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; zone++) {
2070                 int i;
2071
2072                 if (!zone->present_pages)
2073                         continue;
2074
2075                 spin_lock_irqsave(&zone->lock, flags);
2076                 seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
2077                 seq_printf(m,
2078                            "\n  pages free     %lu"
2079                            "\n        min      %lu"
2080                            "\n        low      %lu"
2081                            "\n        high     %lu"
2082                            "\n        active   %lu"
2083                            "\n        inactive %lu"
2084                            "\n        scanned  %lu (a: %lu i: %lu)"
2085                            "\n        spanned  %lu"
2086                            "\n        present  %lu",
2087                            zone->free_pages,
2088                            zone->pages_min,
2089                            zone->pages_low,
2090                            zone->pages_high,
2091                            zone->nr_active,
2092                            zone->nr_inactive,
2093                            zone->pages_scanned,
2094                            zone->nr_scan_active, zone->nr_scan_inactive,
2095                            zone->spanned_pages,
2096                            zone->present_pages);
2097                 seq_printf(m,
2098                            "\n        protection: (%lu",
2099                            zone->lowmem_reserve[0]);
2100                 for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
2101                         seq_printf(m, ", %lu", zone->lowmem_reserve[i]);
2102                 seq_printf(m,
2103                            ")"
2104                            "\n  pagesets");
2105                 for (i = 0; i < ARRAY_SIZE(zone->pageset); i++) {
2106                         struct per_cpu_pageset *pageset;
2107                         int j;
2108
2109                         pageset = zone_pcp(zone, i);
2110                         for (j = 0; j < ARRAY_SIZE(pageset->pcp); j++) {
2111                                 if (pageset->pcp[j].count)
2112                                         break;
2113                         }
2114                         if (j == ARRAY_SIZE(pageset->pcp))
2115                                 continue;
2116                         for (j = 0; j < ARRAY_SIZE(pageset->pcp); j++) {
2117                                 seq_printf(m,
2118                                            "\n    cpu: %i pcp: %i"
2119                                            "\n              count: %i"
2120                                            "\n              low:   %i"
2121                                            "\n              high:  %i"
2122                                            "\n              batch: %i",
2123                                            i, j,
2124                                            pageset->pcp[j].count,
2125                                            pageset->pcp[j].low,
2126                                            pageset->pcp[j].high,
2127                                            pageset->pcp[j].batch);
2128                         }
2129 #ifdef CONFIG_NUMA
2130                         seq_printf(m,
2131                                    "\n            numa_hit:       %lu"
2132                                    "\n            numa_miss:      %lu"
2133                                    "\n            numa_foreign:   %lu"
2134                                    "\n            interleave_hit: %lu"
2135                                    "\n            local_node:     %lu"
2136                                    "\n            other_node:     %lu",
2137                                    pageset->numa_hit,
2138                                    pageset->numa_miss,
2139                                    pageset->numa_foreign,
2140                                    pageset->interleave_hit,
2141                                    pageset->local_node,
2142                                    pageset->other_node);
2143 #endif
2144                 }
2145                 seq_printf(m,
2146                            "\n  all_unreclaimable: %u"
2147                            "\n  prev_priority:     %i"
2148                            "\n  temp_priority:     %i"
2149                            "\n  start_pfn:         %lu",
2150                            zone->all_unreclaimable,
2151                            zone->prev_priority,
2152                            zone->temp_priority,
2153                            zone->zone_start_pfn);
2154                 spin_unlock_irqrestore(&zone->lock, flags);
2155                 seq_putc(m, '\n');
2156         }
2157         return 0;
2158 }
2159
2160 struct seq_operations zoneinfo_op = {
2161         .start  = frag_start, /* iterate over all zones. The same as in
2162                                * fragmentation. */
2163         .next   = frag_next,
2164         .stop   = frag_stop,
2165         .show   = zoneinfo_show,
2166 };
2167
2168 static char *vmstat_text[] = {
2169         "nr_dirty",
2170         "nr_writeback",
2171         "nr_unstable",
2172         "nr_page_table_pages",
2173         "nr_mapped",
2174         "nr_slab",
2175
2176         "pgpgin",
2177         "pgpgout",
2178         "pswpin",
2179         "pswpout",
2180         "pgalloc_high",
2181
2182         "pgalloc_normal",
2183         "pgalloc_dma",
2184         "pgfree",
2185         "pgactivate",
2186         "pgdeactivate",
2187
2188         "pgfault",
2189         "pgmajfault",
2190         "pgrefill_high",
2191         "pgrefill_normal",
2192         "pgrefill_dma",
2193
2194         "pgsteal_high",
2195         "pgsteal_normal",
2196         "pgsteal_dma",
2197         "pgscan_kswapd_high",
2198         "pgscan_kswapd_normal",
2199
2200         "pgscan_kswapd_dma",
2201         "pgscan_direct_high",
2202         "pgscan_direct_normal",
2203         "pgscan_direct_dma",
2204         "pginodesteal",
2205
2206         "slabs_scanned",
2207         "kswapd_steal",
2208         "kswapd_inodesteal",
2209         "pageoutrun",
2210         "allocstall",
2211
2212         "pgrotated",
2213         "nr_bounce",
2214 };
2215
2216 static void *vmstat_start(struct seq_file *m, loff_t *pos)
2217 {
2218         struct page_state *ps;
2219
2220         if (*pos >= ARRAY_SIZE(vmstat_text))
2221                 return NULL;
2222
2223         ps = kmalloc(sizeof(*ps), GFP_KERNEL);
2224         m->private = ps;
2225         if (!ps)
2226                 return ERR_PTR(-ENOMEM);
2227         get_full_page_state(ps);
2228         ps->pgpgin /= 2;                /* sectors -> kbytes */
2229         ps->pgpgout /= 2;
2230         return (unsigned long *)ps + *pos;
2231 }
2232
2233 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
2234 {
2235         (*pos)++;
2236         if (*pos >= ARRAY_SIZE(vmstat_text))
2237                 return NULL;
2238         return (unsigned long *)m->private + *pos;
2239 }
2240
2241 static int vmstat_show(struct seq_file *m, void *arg)
2242 {
2243         unsigned long *l = arg;
2244         unsigned long off = l - (unsigned long *)m->private;
2245
2246         seq_printf(m, "%s %lu\n", vmstat_text[off], *l);
2247         return 0;
2248 }
2249
2250 static void vmstat_stop(struct seq_file *m, void *arg)
2251 {
2252         kfree(m->private);
2253         m->private = NULL;
2254 }
2255
2256 struct seq_operations vmstat_op = {
2257         .start  = vmstat_start,
2258         .next   = vmstat_next,
2259         .stop   = vmstat_stop,
2260         .show   = vmstat_show,
2261 };
2262
2263 #endif /* CONFIG_PROC_FS */
2264
2265 #ifdef CONFIG_HOTPLUG_CPU
2266 static int page_alloc_cpu_notify(struct notifier_block *self,
2267                                  unsigned long action, void *hcpu)
2268 {
2269         int cpu = (unsigned long)hcpu;
2270         long *count;
2271         unsigned long *src, *dest;
2272
2273         if (action == CPU_DEAD) {
2274                 int i;
2275
2276                 /* Drain local pagecache count. */
2277                 count = &per_cpu(nr_pagecache_local, cpu);
2278                 atomic_add(*count, &nr_pagecache);
2279                 *count = 0;
2280                 local_irq_disable();
2281                 __drain_pages(cpu);
2282
2283                 /* Add dead cpu's page_states to our own. */
2284                 dest = (unsigned long *)&__get_cpu_var(page_states);
2285                 src = (unsigned long *)&per_cpu(page_states, cpu);
2286
2287                 for (i = 0; i < sizeof(struct page_state)/sizeof(unsigned long);
2288                                 i++) {
2289                         dest[i] += src[i];
2290                         src[i] = 0;
2291                 }
2292
2293                 local_irq_enable();
2294         }
2295         return NOTIFY_OK;
2296 }
2297 #endif /* CONFIG_HOTPLUG_CPU */
2298
2299 void __init page_alloc_init(void)
2300 {
2301         hotcpu_notifier(page_alloc_cpu_notify, 0);
2302 }
2303
2304 /*
2305  * setup_per_zone_lowmem_reserve - called whenever
2306  *      sysctl_lower_zone_reserve_ratio changes.  Ensures that each zone
2307  *      has a correct pages reserved value, so an adequate number of
2308  *      pages are left in the zone after a successful __alloc_pages().
2309  */
2310 static void setup_per_zone_lowmem_reserve(void)
2311 {
2312         struct pglist_data *pgdat;
2313         int j, idx;
2314
2315         for_each_pgdat(pgdat) {
2316                 for (j = 0; j < MAX_NR_ZONES; j++) {
2317                         struct zone *zone = pgdat->node_zones + j;
2318                         unsigned long present_pages = zone->present_pages;
2319
2320                         zone->lowmem_reserve[j] = 0;
2321
2322                         for (idx = j-1; idx >= 0; idx--) {
2323                                 struct zone *lower_zone;
2324
2325                                 if (sysctl_lowmem_reserve_ratio[idx] < 1)
2326                                         sysctl_lowmem_reserve_ratio[idx] = 1;
2327
2328                                 lower_zone = pgdat->node_zones + idx;
2329                                 lower_zone->lowmem_reserve[j] = present_pages /
2330                                         sysctl_lowmem_reserve_ratio[idx];
2331                                 present_pages += lower_zone->present_pages;
2332                         }
2333                 }
2334         }
2335 }
2336
2337 /*
2338  * setup_per_zone_pages_min - called when min_free_kbytes changes.  Ensures 
2339  *      that the pages_{min,low,high} values for each zone are set correctly 
2340  *      with respect to min_free_kbytes.
2341  */
2342 static void setup_per_zone_pages_min(void)
2343 {
2344         unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
2345         unsigned long lowmem_pages = 0;
2346         struct zone *zone;
2347         unsigned long flags;
2348
2349         /* Calculate total number of !ZONE_HIGHMEM pages */
2350         for_each_zone(zone) {
2351                 if (!is_highmem(zone))
2352                         lowmem_pages += zone->present_pages;
2353         }
2354
2355         for_each_zone(zone) {
2356                 spin_lock_irqsave(&zone->lru_lock, flags);
2357                 if (is_highmem(zone)) {
2358                         /*
2359                          * Often, highmem doesn't need to reserve any pages.
2360                          * But the pages_min/low/high values are also used for
2361                          * batching up page reclaim activity so we need a
2362                          * decent value here.
2363                          */
2364                         int min_pages;
2365
2366                         min_pages = zone->present_pages / 1024;
2367                         if (min_pages < SWAP_CLUSTER_MAX)
2368                                 min_pages = SWAP_CLUSTER_MAX;
2369                         if (min_pages > 128)
2370                                 min_pages = 128;
2371                         zone->pages_min = min_pages;
2372                 } else {
2373                         /* if it's a lowmem zone, reserve a number of pages
2374                          * proportionate to the zone's size.
2375                          */
2376                         zone->pages_min = (pages_min * zone->present_pages) /
2377                                            lowmem_pages;
2378                 }
2379
2380                 /*
2381                  * When interpreting these watermarks, just keep in mind that:
2382                  * zone->pages_min == (zone->pages_min * 4) / 4;
2383                  */
2384                 zone->pages_low   = (zone->pages_min * 5) / 4;
2385                 zone->pages_high  = (zone->pages_min * 6) / 4;
2386                 spin_unlock_irqrestore(&zone->lru_lock, flags);
2387         }
2388 }
2389
2390 /*
2391  * Initialise min_free_kbytes.
2392  *
2393  * For small machines we want it small (128k min).  For large machines
2394  * we want it large (64MB max).  But it is not linear, because network
2395  * bandwidth does not increase linearly with machine size.  We use
2396  *
2397  *      min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
2398  *      min_free_kbytes = sqrt(lowmem_kbytes * 16)
2399  *
2400  * which yields
2401  *
2402  * 16MB:        512k
2403  * 32MB:        724k
2404  * 64MB:        1024k
2405  * 128MB:       1448k
2406  * 256MB:       2048k
2407  * 512MB:       2896k
2408  * 1024MB:      4096k
2409  * 2048MB:      5792k
2410  * 4096MB:      8192k
2411  * 8192MB:      11584k
2412  * 16384MB:     16384k
2413  */
2414 static int __init init_per_zone_pages_min(void)
2415 {
2416         unsigned long lowmem_kbytes;
2417
2418         lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
2419
2420         min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
2421         if (min_free_kbytes < 128)
2422                 min_free_kbytes = 128;
2423         if (min_free_kbytes > 65536)
2424                 min_free_kbytes = 65536;
2425         setup_per_zone_pages_min();
2426         setup_per_zone_lowmem_reserve();
2427         return 0;
2428 }
2429 module_init(init_per_zone_pages_min)
2430
2431 /*
2432  * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so 
2433  *      that we can call two helper functions whenever min_free_kbytes
2434  *      changes.
2435  */
2436 int min_free_kbytes_sysctl_handler(ctl_table *table, int write, 
2437         struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2438 {
2439         proc_dointvec(table, write, file, buffer, length, ppos);
2440         setup_per_zone_pages_min();
2441         return 0;
2442 }
2443
2444 /*
2445  * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
2446  *      proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
2447  *      whenever sysctl_lowmem_reserve_ratio changes.
2448  *
2449  * The reserve ratio obviously has absolutely no relation with the
2450  * pages_min watermarks. The lowmem reserve ratio can only make sense
2451  * if in function of the boot time zone sizes.
2452  */
2453 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
2454         struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2455 {
2456         proc_dointvec_minmax(table, write, file, buffer, length, ppos);
2457         setup_per_zone_lowmem_reserve();
2458         return 0;
2459 }
2460
2461 __initdata int hashdist = HASHDIST_DEFAULT;
2462
2463 #ifdef CONFIG_NUMA
2464 static int __init set_hashdist(char *str)
2465 {
2466         if (!str)
2467                 return 0;
2468         hashdist = simple_strtoul(str, &str, 0);
2469         return 1;
2470 }
2471 __setup("hashdist=", set_hashdist);
2472 #endif
2473
2474 /*
2475  * allocate a large system hash table from bootmem
2476  * - it is assumed that the hash table must contain an exact power-of-2
2477  *   quantity of entries
2478  * - limit is the number of hash buckets, not the total allocation size
2479  */
2480 void *__init alloc_large_system_hash(const char *tablename,
2481                                      unsigned long bucketsize,
2482                                      unsigned long numentries,
2483                                      int scale,
2484                                      int flags,
2485                                      unsigned int *_hash_shift,
2486                                      unsigned int *_hash_mask,
2487                                      unsigned long limit)
2488 {
2489         unsigned long long max = limit;
2490         unsigned long log2qty, size;
2491         void *table = NULL;
2492
2493         /* allow the kernel cmdline to have a say */
2494         if (!numentries) {
2495                 /* round applicable memory size up to nearest megabyte */
2496                 numentries = (flags & HASH_HIGHMEM) ? nr_all_pages : nr_kernel_pages;
2497                 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
2498                 numentries >>= 20 - PAGE_SHIFT;
2499                 numentries <<= 20 - PAGE_SHIFT;
2500
2501                 /* limit to 1 bucket per 2^scale bytes of low memory */
2502                 if (scale > PAGE_SHIFT)
2503                         numentries >>= (scale - PAGE_SHIFT);
2504                 else
2505                         numentries <<= (PAGE_SHIFT - scale);
2506         }
2507         /* rounded up to nearest power of 2 in size */
2508         numentries = 1UL << (long_log2(numentries) + 1);
2509
2510         /* limit allocation size to 1/16 total memory by default */
2511         if (max == 0) {
2512                 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
2513                 do_div(max, bucketsize);
2514         }
2515
2516         if (numentries > max)
2517                 numentries = max;
2518
2519         log2qty = long_log2(numentries);
2520
2521         do {
2522                 size = bucketsize << log2qty;
2523                 if (flags & HASH_EARLY)
2524                         table = alloc_bootmem(size);
2525                 else if (hashdist)
2526                         table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
2527                 else {
2528                         unsigned long order;
2529                         for (order = 0; ((1UL << order) << PAGE_SHIFT) < size; order++)
2530                                 ;
2531                         table = (void*) __get_free_pages(GFP_ATOMIC, order);
2532                 }
2533         } while (!table && size > PAGE_SIZE && --log2qty);
2534
2535         if (!table)
2536                 panic("Failed to allocate %s hash table\n", tablename);
2537
2538         printk("%s hash table entries: %d (order: %d, %lu bytes)\n",
2539                tablename,
2540                (1U << log2qty),
2541                long_log2(size) - PAGE_SHIFT,
2542                size);
2543
2544         if (_hash_shift)
2545                 *_hash_shift = log2qty;
2546         if (_hash_mask)
2547                 *_hash_mask = (1 << log2qty) - 1;
2548
2549         return table;
2550 }