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